Created
October 14, 2022 14:16
-
-
Save pashu123/f7d4d2a2d2c91bd46423d0006d23c030 to your computer and use it in GitHub Desktop.
This file has been truncated, but you can view the full file.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| #map0 = affine_map<(d0) -> (0)> | |
| #map1 = affine_map<(d0) -> (d0)> | |
| #map2 = affine_map<(d0) -> ()> | |
| #map3 = affine_map<() -> ()> | |
| #map4 = affine_map<(d0, d1) -> ()> | |
| #map5 = affine_map<(d0, d1) -> (d0, d1)> | |
| #map6 = affine_map<(d0, d1) -> (d0, 0)> | |
| #map7 = affine_map<(d0, d1) -> (0, d1)> | |
| #map8 = affine_map<(d0, d1) -> (d1, d0)> | |
| #map9 = affine_map<(d0, d1) -> (d1)> | |
| #map10 = affine_map<(d0, d1, d2, d3) -> (d1)> | |
| #map11 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)> | |
| #map12 = affine_map<(d0, d1, d2, d3) -> ()> | |
| #map13 = affine_map<(d0, d1, d2, d3) -> (d0, d1, 0, 0)> | |
| #map14 = affine_map<(d0, d1, d2, d3) -> (d1, 0, 0)> | |
| #map15 = affine_map<(d0, d1, d2, d3) -> (d0, d2, d3, d1)> | |
| #map16 = affine_map<(d0, d1, d2) -> (d0, d1, d2)> | |
| #map17 = affine_map<(d0, d1, d2) -> (d0, d1, 0)> | |
| #map18 = affine_map<(d0, d1, d2) -> (d2)> | |
| #map19 = affine_map<(d0, d1, d2, d3) -> (d0, d2, d1, d3)> | |
| #map20 = affine_map<(d0, d1, d2) -> (d0, d2, d1)> | |
| #map21 = affine_map<(d0, d1, d2) -> ()> | |
| #map22 = affine_map<(d0, d1, d2, d3) -> (d0, d3, d1, d2)> | |
| module attributes {torch.debug_module_name = "_lambda"} { | |
| func.func @forward(%arg0: tensor<2x4x64x64xf16>, %arg1: tensor<1xf16>, %arg2: tensor<2x77x768xf16>) -> tensor<2x4x64x64xf16> { | |
| %cst = arith.constant dense<0.079056941504209485> : tensor<f64> | |
| %cst_0 = arith.constant dense<0.11180339887498948> : tensor<f64> | |
| %cst_1 = arith.constant dense<0.15811388300841897> : tensor<f64> | |
| %cst_2 = arith.constant dense<9.9999999999999995E-7> : tensor<f64> | |
| %cst_3 = arith.constant dense<1.000000e+00> : tensor<f64> | |
| %cst_4 = arith.constant dense<1.000000e-05> : tensor<f64> | |
| %cst_5 = arith.constant dense<1> : tensor<i64> | |
| %cst_6 = arith.constant dense<160> : tensor<i64> | |
| %cst_7 = arith.constant dense<-9.2103403719761836> : tensor<f64> | |
| %cst_8 = arith.constant dense_resource<__elided__> : tensor<1280x320xf16> | |
| %cst_9 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_10 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_11 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_12 = arith.constant dense_resource<__elided__> : tensor<320x4x3x3xf16> | |
| %cst_13 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_14 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_15 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_16 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xf16> | |
| %cst_17 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_18 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16> | |
| %cst_19 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_20 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_21 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_22 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xf16> | |
| %cst_23 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_24 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_25 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_26 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xf16> | |
| %cst_27 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_28 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_29 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_30 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_31 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_32 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_33 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_34 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_35 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_36 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_37 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_38 = arith.constant dense_resource<__elided__> : tensor<320x768xf16> | |
| %cst_39 = arith.constant dense_resource<__elided__> : tensor<320x768xf16> | |
| %cst_40 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_41 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_42 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_43 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_44 = arith.constant dense_resource<__elided__> : tensor<2560x320xf16> | |
| %cst_45 = arith.constant dense_resource<__elided__> : tensor<2560xf16> | |
| %cst_46 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16> | |
| %cst_47 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_48 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xf16> | |
| %cst_49 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_50 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_51 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_52 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xf16> | |
| %cst_53 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_54 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16> | |
| %cst_55 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_56 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_57 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_58 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xf16> | |
| %cst_59 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_60 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_61 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_62 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xf16> | |
| %cst_63 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_64 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_65 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_66 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_67 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_68 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_69 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_70 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_71 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_72 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_73 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_74 = arith.constant dense_resource<__elided__> : tensor<320x768xf16> | |
| %cst_75 = arith.constant dense_resource<__elided__> : tensor<320x768xf16> | |
| %cst_76 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_77 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_78 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_79 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_80 = arith.constant dense_resource<__elided__> : tensor<2560x320xf16> | |
| %cst_81 = arith.constant dense_resource<__elided__> : tensor<2560xf16> | |
| %cst_82 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16> | |
| %cst_83 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_84 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xf16> | |
| %cst_85 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_86 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xf16> | |
| %cst_87 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_88 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_89 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_90 = arith.constant dense_resource<__elided__> : tensor<640x320x3x3xf16> | |
| %cst_91 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_92 = arith.constant dense_resource<__elided__> : tensor<640x1280xf16> | |
| %cst_93 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_94 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_95 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_96 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xf16> | |
| %cst_97 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_98 = arith.constant dense_resource<__elided__> : tensor<640x320x1x1xf16> | |
| %cst_99 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_100 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_101 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_102 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xf16> | |
| %cst_103 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_104 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_105 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_106 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_107 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_108 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_109 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_110 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_111 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_112 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_113 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_114 = arith.constant dense_resource<__elided__> : tensor<640x768xf16> | |
| %cst_115 = arith.constant dense_resource<__elided__> : tensor<640x768xf16> | |
| %cst_116 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_117 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_118 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_119 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_120 = arith.constant dense_resource<__elided__> : tensor<5120x640xf16> | |
| %cst_121 = arith.constant dense_resource<__elided__> : tensor<5120xf16> | |
| %cst_122 = arith.constant dense_resource<__elided__> : tensor<640x2560xf16> | |
| %cst_123 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_124 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xf16> | |
| %cst_125 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_126 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_127 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_128 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xf16> | |
| %cst_129 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_130 = arith.constant dense_resource<__elided__> : tensor<640x1280xf16> | |
| %cst_131 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_132 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_133 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_134 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xf16> | |
| %cst_135 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_136 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_137 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_138 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xf16> | |
| %cst_139 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_140 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_141 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_142 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_143 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_144 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_145 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_146 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_147 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_148 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_149 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_150 = arith.constant dense_resource<__elided__> : tensor<640x768xf16> | |
| %cst_151 = arith.constant dense_resource<__elided__> : tensor<640x768xf16> | |
| %cst_152 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_153 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_154 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_155 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_156 = arith.constant dense_resource<__elided__> : tensor<5120x640xf16> | |
| %cst_157 = arith.constant dense_resource<__elided__> : tensor<5120xf16> | |
| %cst_158 = arith.constant dense_resource<__elided__> : tensor<640x2560xf16> | |
| %cst_159 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_160 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xf16> | |
| %cst_161 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_162 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xf16> | |
| %cst_163 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_164 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_165 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_166 = arith.constant dense_resource<__elided__> : tensor<1280x640x3x3xf16> | |
| %cst_167 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_168 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_169 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_170 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_171 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_172 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_173 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_174 = arith.constant dense_resource<__elided__> : tensor<1280x640x1x1xf16> | |
| %cst_175 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_176 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_177 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_178 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xf16> | |
| %cst_179 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_180 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_181 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_182 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_183 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_184 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_185 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_186 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_187 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_188 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_189 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_190 = arith.constant dense_resource<__elided__> : tensor<1280x768xf16> | |
| %cst_191 = arith.constant dense_resource<__elided__> : tensor<1280x768xf16> | |
| %cst_192 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_193 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_194 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_195 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_196 = arith.constant dense_resource<__elided__> : tensor<10240x1280xf16> | |
| %cst_197 = arith.constant dense_resource<__elided__> : tensor<10240xf16> | |
| %cst_198 = arith.constant dense_resource<__elided__> : tensor<1280x5120xf16> | |
| %cst_199 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_200 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xf16> | |
| %cst_201 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_202 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_203 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_204 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_205 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_206 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_207 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_208 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_209 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_210 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_211 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_212 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_213 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_214 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xf16> | |
| %cst_215 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_216 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_217 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_218 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_219 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_220 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_221 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_222 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_223 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_224 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_225 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_226 = arith.constant dense_resource<__elided__> : tensor<1280x768xf16> | |
| %cst_227 = arith.constant dense_resource<__elided__> : tensor<1280x768xf16> | |
| %cst_228 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_229 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_230 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_231 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_232 = arith.constant dense_resource<__elided__> : tensor<10240x1280xf16> | |
| %cst_233 = arith.constant dense_resource<__elided__> : tensor<10240xf16> | |
| %cst_234 = arith.constant dense_resource<__elided__> : tensor<1280x5120xf16> | |
| %cst_235 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_236 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xf16> | |
| %cst_237 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_238 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_239 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_240 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_241 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_242 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_243 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_244 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_245 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_246 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_247 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_248 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_249 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_250 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_251 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_252 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_253 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_254 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_255 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_256 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_257 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_258 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_259 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_260 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_261 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_262 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_263 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_264 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_265 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_266 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_267 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_268 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_269 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_270 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_271 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_272 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xf16> | |
| %cst_273 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_274 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_275 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_276 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_277 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_278 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_279 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_280 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_281 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_282 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_283 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_284 = arith.constant dense_resource<__elided__> : tensor<1280x768xf16> | |
| %cst_285 = arith.constant dense_resource<__elided__> : tensor<1280x768xf16> | |
| %cst_286 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_287 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_288 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_289 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_290 = arith.constant dense_resource<__elided__> : tensor<10240x1280xf16> | |
| %cst_291 = arith.constant dense_resource<__elided__> : tensor<10240xf16> | |
| %cst_292 = arith.constant dense_resource<__elided__> : tensor<1280x5120xf16> | |
| %cst_293 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_294 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xf16> | |
| %cst_295 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_296 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_297 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_298 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_299 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_300 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_301 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_302 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_303 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_304 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_305 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_306 = arith.constant dense_resource<__elided__> : tensor<2560xf16> | |
| %cst_307 = arith.constant dense_resource<__elided__> : tensor<2560xf16> | |
| %cst_308 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xf16> | |
| %cst_309 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_310 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_311 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_312 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_313 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_314 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_315 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_316 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xf16> | |
| %cst_317 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_318 = arith.constant dense_resource<__elided__> : tensor<2560xf16> | |
| %cst_319 = arith.constant dense_resource<__elided__> : tensor<2560xf16> | |
| %cst_320 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xf16> | |
| %cst_321 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_322 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_323 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_324 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_325 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_326 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_327 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_328 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xf16> | |
| %cst_329 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_330 = arith.constant dense_resource<__elided__> : tensor<2560xf16> | |
| %cst_331 = arith.constant dense_resource<__elided__> : tensor<2560xf16> | |
| %cst_332 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xf16> | |
| %cst_333 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_334 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_335 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_336 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_337 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_338 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_339 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_340 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xf16> | |
| %cst_341 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_342 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_343 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_344 = arith.constant dense_resource<__elided__> : tensor<2560xf16> | |
| %cst_345 = arith.constant dense_resource<__elided__> : tensor<2560xf16> | |
| %cst_346 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xf16> | |
| %cst_347 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_348 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_349 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_350 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_351 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_352 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_353 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_354 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xf16> | |
| %cst_355 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_356 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_357 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_358 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xf16> | |
| %cst_359 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_360 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_361 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_362 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_363 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_364 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_365 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_366 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_367 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_368 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_369 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_370 = arith.constant dense_resource<__elided__> : tensor<1280x768xf16> | |
| %cst_371 = arith.constant dense_resource<__elided__> : tensor<1280x768xf16> | |
| %cst_372 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_373 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_374 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_375 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_376 = arith.constant dense_resource<__elided__> : tensor<10240x1280xf16> | |
| %cst_377 = arith.constant dense_resource<__elided__> : tensor<10240xf16> | |
| %cst_378 = arith.constant dense_resource<__elided__> : tensor<1280x5120xf16> | |
| %cst_379 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_380 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xf16> | |
| %cst_381 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_382 = arith.constant dense_resource<__elided__> : tensor<2560xf16> | |
| %cst_383 = arith.constant dense_resource<__elided__> : tensor<2560xf16> | |
| %cst_384 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xf16> | |
| %cst_385 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_386 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_387 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_388 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_389 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_390 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_391 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_392 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xf16> | |
| %cst_393 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_394 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_395 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_396 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xf16> | |
| %cst_397 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_398 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_399 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_400 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_401 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_402 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_403 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_404 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_405 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_406 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_407 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_408 = arith.constant dense_resource<__elided__> : tensor<1280x768xf16> | |
| %cst_409 = arith.constant dense_resource<__elided__> : tensor<1280x768xf16> | |
| %cst_410 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_411 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_412 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_413 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_414 = arith.constant dense_resource<__elided__> : tensor<10240x1280xf16> | |
| %cst_415 = arith.constant dense_resource<__elided__> : tensor<10240xf16> | |
| %cst_416 = arith.constant dense_resource<__elided__> : tensor<1280x5120xf16> | |
| %cst_417 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_418 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xf16> | |
| %cst_419 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_420 = arith.constant dense_resource<__elided__> : tensor<1920xf16> | |
| %cst_421 = arith.constant dense_resource<__elided__> : tensor<1920xf16> | |
| %cst_422 = arith.constant dense_resource<__elided__> : tensor<1280x1920x3x3xf16> | |
| %cst_423 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_424 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_425 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_426 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_427 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_428 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_429 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_430 = arith.constant dense_resource<__elided__> : tensor<1280x1920x1x1xf16> | |
| %cst_431 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_432 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_433 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_434 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xf16> | |
| %cst_435 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_436 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_437 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_438 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_439 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_440 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_441 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_442 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_443 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_444 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_445 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_446 = arith.constant dense_resource<__elided__> : tensor<1280x768xf16> | |
| %cst_447 = arith.constant dense_resource<__elided__> : tensor<1280x768xf16> | |
| %cst_448 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16> | |
| %cst_449 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_450 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_451 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_452 = arith.constant dense_resource<__elided__> : tensor<10240x1280xf16> | |
| %cst_453 = arith.constant dense_resource<__elided__> : tensor<10240xf16> | |
| %cst_454 = arith.constant dense_resource<__elided__> : tensor<1280x5120xf16> | |
| %cst_455 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_456 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xf16> | |
| %cst_457 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_458 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf16> | |
| %cst_459 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_460 = arith.constant dense_resource<__elided__> : tensor<1920xf16> | |
| %cst_461 = arith.constant dense_resource<__elided__> : tensor<1920xf16> | |
| %cst_462 = arith.constant dense_resource<__elided__> : tensor<640x1920x3x3xf16> | |
| %cst_463 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_464 = arith.constant dense_resource<__elided__> : tensor<640x1280xf16> | |
| %cst_465 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_466 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_467 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_468 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xf16> | |
| %cst_469 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_470 = arith.constant dense_resource<__elided__> : tensor<640x1920x1x1xf16> | |
| %cst_471 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_472 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_473 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_474 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xf16> | |
| %cst_475 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_476 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_477 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_478 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_479 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_480 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_481 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_482 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_483 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_484 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_485 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_486 = arith.constant dense_resource<__elided__> : tensor<640x768xf16> | |
| %cst_487 = arith.constant dense_resource<__elided__> : tensor<640x768xf16> | |
| %cst_488 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_489 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_490 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_491 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_492 = arith.constant dense_resource<__elided__> : tensor<5120x640xf16> | |
| %cst_493 = arith.constant dense_resource<__elided__> : tensor<5120xf16> | |
| %cst_494 = arith.constant dense_resource<__elided__> : tensor<640x2560xf16> | |
| %cst_495 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_496 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xf16> | |
| %cst_497 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_498 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_499 = arith.constant dense_resource<__elided__> : tensor<1280xf16> | |
| %cst_500 = arith.constant dense_resource<__elided__> : tensor<640x1280x3x3xf16> | |
| %cst_501 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_502 = arith.constant dense_resource<__elided__> : tensor<640x1280xf16> | |
| %cst_503 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_504 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_505 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_506 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xf16> | |
| %cst_507 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_508 = arith.constant dense_resource<__elided__> : tensor<640x1280x1x1xf16> | |
| %cst_509 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_510 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_511 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_512 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xf16> | |
| %cst_513 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_514 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_515 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_516 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_517 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_518 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_519 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_520 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_521 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_522 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_523 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_524 = arith.constant dense_resource<__elided__> : tensor<640x768xf16> | |
| %cst_525 = arith.constant dense_resource<__elided__> : tensor<640x768xf16> | |
| %cst_526 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_527 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_528 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_529 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_530 = arith.constant dense_resource<__elided__> : tensor<5120x640xf16> | |
| %cst_531 = arith.constant dense_resource<__elided__> : tensor<5120xf16> | |
| %cst_532 = arith.constant dense_resource<__elided__> : tensor<640x2560xf16> | |
| %cst_533 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_534 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xf16> | |
| %cst_535 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_536 = arith.constant dense_resource<__elided__> : tensor<960xf16> | |
| %cst_537 = arith.constant dense_resource<__elided__> : tensor<960xf16> | |
| %cst_538 = arith.constant dense_resource<__elided__> : tensor<640x960x3x3xf16> | |
| %cst_539 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_540 = arith.constant dense_resource<__elided__> : tensor<640x1280xf16> | |
| %cst_541 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_542 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_543 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_544 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xf16> | |
| %cst_545 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_546 = arith.constant dense_resource<__elided__> : tensor<640x960x1x1xf16> | |
| %cst_547 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_548 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_549 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_550 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xf16> | |
| %cst_551 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_552 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_553 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_554 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_555 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_556 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_557 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_558 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_559 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_560 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_561 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_562 = arith.constant dense_resource<__elided__> : tensor<640x768xf16> | |
| %cst_563 = arith.constant dense_resource<__elided__> : tensor<640x768xf16> | |
| %cst_564 = arith.constant dense_resource<__elided__> : tensor<640x640xf16> | |
| %cst_565 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_566 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_567 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_568 = arith.constant dense_resource<__elided__> : tensor<5120x640xf16> | |
| %cst_569 = arith.constant dense_resource<__elided__> : tensor<5120xf16> | |
| %cst_570 = arith.constant dense_resource<__elided__> : tensor<640x2560xf16> | |
| %cst_571 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_572 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xf16> | |
| %cst_573 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_574 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xf16> | |
| %cst_575 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_576 = arith.constant dense_resource<__elided__> : tensor<960xf16> | |
| %cst_577 = arith.constant dense_resource<__elided__> : tensor<960xf16> | |
| %cst_578 = arith.constant dense_resource<__elided__> : tensor<320x960x3x3xf16> | |
| %cst_579 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_580 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16> | |
| %cst_581 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_582 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_583 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_584 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xf16> | |
| %cst_585 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_586 = arith.constant dense_resource<__elided__> : tensor<320x960x1x1xf16> | |
| %cst_587 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_588 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_589 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_590 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xf16> | |
| %cst_591 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_592 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_593 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_594 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_595 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_596 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_597 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_598 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_599 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_600 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_601 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_602 = arith.constant dense_resource<__elided__> : tensor<320x768xf16> | |
| %cst_603 = arith.constant dense_resource<__elided__> : tensor<320x768xf16> | |
| %cst_604 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_605 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_606 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_607 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_608 = arith.constant dense_resource<__elided__> : tensor<2560x320xf16> | |
| %cst_609 = arith.constant dense_resource<__elided__> : tensor<2560xf16> | |
| %cst_610 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16> | |
| %cst_611 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_612 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xf16> | |
| %cst_613 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_614 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_615 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_616 = arith.constant dense_resource<__elided__> : tensor<320x640x3x3xf16> | |
| %cst_617 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_618 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16> | |
| %cst_619 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_620 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_621 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_622 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xf16> | |
| %cst_623 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_624 = arith.constant dense_resource<__elided__> : tensor<320x640x1x1xf16> | |
| %cst_625 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_626 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_627 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_628 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xf16> | |
| %cst_629 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_630 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_631 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_632 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_633 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_634 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_635 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_636 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_637 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_638 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_639 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_640 = arith.constant dense_resource<__elided__> : tensor<320x768xf16> | |
| %cst_641 = arith.constant dense_resource<__elided__> : tensor<320x768xf16> | |
| %cst_642 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_643 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_644 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_645 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_646 = arith.constant dense_resource<__elided__> : tensor<2560x320xf16> | |
| %cst_647 = arith.constant dense_resource<__elided__> : tensor<2560xf16> | |
| %cst_648 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16> | |
| %cst_649 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_650 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xf16> | |
| %cst_651 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_652 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_653 = arith.constant dense_resource<__elided__> : tensor<640xf16> | |
| %cst_654 = arith.constant dense_resource<__elided__> : tensor<320x640x3x3xf16> | |
| %cst_655 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_656 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16> | |
| %cst_657 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_658 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_659 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_660 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xf16> | |
| %cst_661 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_662 = arith.constant dense_resource<__elided__> : tensor<320x640x1x1xf16> | |
| %cst_663 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_664 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_665 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_666 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xf16> | |
| %cst_667 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_668 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_669 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_670 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_671 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_672 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_673 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_674 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_675 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_676 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_677 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_678 = arith.constant dense_resource<__elided__> : tensor<320x768xf16> | |
| %cst_679 = arith.constant dense_resource<__elided__> : tensor<320x768xf16> | |
| %cst_680 = arith.constant dense_resource<__elided__> : tensor<320x320xf16> | |
| %cst_681 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_682 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_683 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_684 = arith.constant dense_resource<__elided__> : tensor<2560x320xf16> | |
| %cst_685 = arith.constant dense_resource<__elided__> : tensor<2560xf16> | |
| %cst_686 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16> | |
| %cst_687 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_688 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xf16> | |
| %cst_689 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_690 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_691 = arith.constant dense_resource<__elided__> : tensor<320xf16> | |
| %cst_692 = arith.constant dense_resource<__elided__> : tensor<4x320x3x3xf16> | |
| %cst_693 = arith.constant dense<[-1.393320e-03, -1.588820e-03, -2.624990e-04, -2.531050e-03]> : tensor<4xf16> | |
| %cst_694 = arith.constant 0.000000e+00 : f16 | |
| %cst_695 = arith.constant 1.000000e+00 : f16 | |
| %cst_696 = arith.constant -6.550400e+04 : f16 | |
| %cst_697 = arith.constant 2.000000e+00 : f16 | |
| %cst_698 = arith.constant 5.000000e-01 : f16 | |
| %cst_699 = arith.constant 0.000000e+00 : f64 | |
| %cst_700 = arith.constant 1.000000e-05 : f64 | |
| %cst_701 = arith.constant 0.000000e+00 : f32 | |
| %cst_702 = arith.constant 4.096000e+04 : f64 | |
| %cst_703 = arith.constant 4.096000e+04 : f32 | |
| %cst_704 = arith.constant 3.200000e+02 : f16 | |
| %cst_705 = arith.constant 1.024000e+04 : f64 | |
| %cst_706 = arith.constant 1.024000e+04 : f32 | |
| %cst_707 = arith.constant 2.048000e+04 : f64 | |
| %cst_708 = arith.constant 2.048000e+04 : f32 | |
| %cst_709 = arith.constant 6.400000e+02 : f16 | |
| %cst_710 = arith.constant 5.120000e+03 : f64 | |
| %cst_711 = arith.constant 5.120000e+03 : f32 | |
| %cst_712 = arith.constant 1.280000e+03 : f16 | |
| %cst_713 = arith.constant 2.560000e+03 : f64 | |
| %cst_714 = arith.constant 2.560000e+03 : f32 | |
| %cst_715 = arith.constant 1.536000e+04 : f64 | |
| %cst_716 = arith.constant 1.536000e+04 : f32 | |
| %cst_717 = arith.constant 6.144000e+04 : f64 | |
| %cst_718 = arith.constant 6.144000e+04 : f32 | |
| %cst_719 = arith.constant 3.072000e+04 : f64 | |
| %cst_720 = arith.constant 3.072000e+04 : f32 | |
| %cst_721 = arith.constant 1.228800e+05 : f64 | |
| %cst_722 = arith.constant 1.228800e+05 : f32 | |
| %cst_723 = arith.constant 8.192000e+04 : f64 | |
| %cst_724 = arith.constant 8.192000e+04 : f32 | |
| %c0_i64 = arith.constant 0 : i64 | |
| %c2 = arith.constant 2 : index | |
| %0 = tensor.empty() : tensor<2xf16> | |
| %1 = linalg.generic {indexing_maps = [#map0, #map1], iterator_types = ["parallel"]} ins(%arg1 : tensor<1xf16>) outs(%0 : tensor<2xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2xf16> | |
| %2 = tensor.empty() : tensor<160xf32> | |
| %3 = linalg.generic {indexing_maps = [#map1], iterator_types = ["parallel"]} outs(%2 : tensor<160xf32>) { | |
| ^bb0(%out: f32): | |
| %4280 = linalg.index 0 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.sitofp %4281 : i64 to f32 | |
| %4283 = arith.addf %4282, %cst_701 : f32 | |
| linalg.yield %4283 : f32 | |
| } -> tensor<160xf32> | |
| %4 = linalg.generic {indexing_maps = [#map1, #map2, #map1], iterator_types = ["parallel"]} ins(%3, %cst_7 : tensor<160xf32>, tensor<f64>) outs(%2 : tensor<160xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<160xf32> | |
| %5 = linalg.generic {indexing_maps = [#map1, #map2, #map1], iterator_types = ["parallel"]} ins(%4, %cst_6 : tensor<160xf32>, tensor<i64>) outs(%2 : tensor<160xf32>) { | |
| ^bb0(%in: f32, %in_1640: i64, %out: f32): | |
| %4280 = arith.sitofp %in_1640 : i64 to f32 | |
| %4281 = arith.divf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<160xf32> | |
| %6 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel"]} ins(%5 : tensor<160xf32>) outs(%2 : tensor<160xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.exp %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<160xf32> | |
| %expanded = tensor.expand_shape %1 [[0, 1]] : tensor<2xf16> into tensor<2x1xf16> | |
| %7 = tensor.empty() : tensor<f64> | |
| %8 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %9 = tensor.empty() : tensor<f32> | |
| %10 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%8 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %11 = tensor.empty() : tensor<2x1xf32> | |
| %12 = linalg.generic {indexing_maps = [#map4, #map5], iterator_types = ["parallel", "parallel"]} ins(%10 : tensor<f32>) outs(%11 : tensor<2x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x1xf32> | |
| %13 = tensor.empty() : tensor<2x1xf16> | |
| %14 = linalg.generic {indexing_maps = [#map6, #map5], iterator_types = ["parallel", "parallel"]} ins(%expanded : tensor<2x1xf16>) outs(%13 : tensor<2x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1xf16> | |
| %15 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%14 : tensor<2x1xf16>) outs(%12 : tensor<2x1xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x1xf32> | |
| %expanded_725 = tensor.expand_shape %6 [[0, 1]] : tensor<160xf32> into tensor<1x160xf32> | |
| %16 = tensor.empty() : tensor<2x160xf32> | |
| %17 = linalg.generic {indexing_maps = [#map6, #map7, #map5], iterator_types = ["parallel", "parallel"]} ins(%15, %expanded_725 : tensor<2x1xf32>, tensor<1x160xf32>) outs(%16 : tensor<2x160xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x160xf32> | |
| %18 = linalg.generic {indexing_maps = [#map5, #map4, #map5], iterator_types = ["parallel", "parallel"]} ins(%17, %cst_5 : tensor<2x160xf32>, tensor<i64>) outs(%16 : tensor<2x160xf32>) { | |
| ^bb0(%in: f32, %in_1640: i64, %out: f32): | |
| %4280 = arith.sitofp %in_1640 : i64 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x160xf32> | |
| %19 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%18 : tensor<2x160xf32>) outs(%16 : tensor<2x160xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.sin %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x160xf32> | |
| %20 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%18 : tensor<2x160xf32>) outs(%16 : tensor<2x160xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.cos %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x160xf32> | |
| %21 = tensor.empty() : tensor<2x320xf32> | |
| %inserted_slice = tensor.insert_slice %19 into %21[0, 0] [2, 160] [1, 1] : tensor<2x160xf32> into tensor<2x320xf32> | |
| %inserted_slice_726 = tensor.insert_slice %20 into %inserted_slice[0, 160] [2, 160] [1, 1] : tensor<2x160xf32> into tensor<2x320xf32> | |
| %extracted_slice = tensor.extract_slice %inserted_slice_726[0, 0] [2, 160] [1, 1] : tensor<2x320xf32> to tensor<2x160xf32> | |
| %inserted_slice_727 = tensor.insert_slice %20 into %21[0, 0] [2, 160] [1, 1] : tensor<2x160xf32> into tensor<2x320xf32> | |
| %inserted_slice_728 = tensor.insert_slice %extracted_slice into %inserted_slice_727[0, 160] [2, 160] [1, 1] : tensor<2x160xf32> into tensor<2x320xf32> | |
| %22 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %23 = tensor.empty() : tensor<f16> | |
| %24 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%22 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %25 = tensor.empty() : tensor<2x320xf16> | |
| %26 = linalg.generic {indexing_maps = [#map4, #map5], iterator_types = ["parallel", "parallel"]} ins(%24 : tensor<f16>) outs(%25 : tensor<2x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320xf16> | |
| %27 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%inserted_slice_728 : tensor<2x320xf32>) outs(%26 : tensor<2x320xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320xf16> | |
| %28 = tensor.empty() : tensor<320x1280xf16> | |
| %29 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_8 : tensor<1280x320xf16>) outs(%28 : tensor<320x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<320x1280xf16> | |
| %30 = tensor.empty() : tensor<2x1280xf16> | |
| %31 = linalg.fill ins(%cst_694 : f16) outs(%30 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %32 = linalg.matmul ins(%27, %29 : tensor<2x320xf16>, tensor<320x1280xf16>) outs(%31 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %33 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_9, %32 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %34 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%33 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280xf16> | |
| %35 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%34, %33 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %36 = tensor.empty() : tensor<1280x1280xf16> | |
| %37 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_10 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %38 = linalg.fill ins(%cst_694 : f16) outs(%30 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %39 = linalg.matmul ins(%35, %37 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%38 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %40 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_11, %39 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %padded = tensor.pad %arg0 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x4x64x64xf16> to tensor<2x4x66x66xf16> | |
| %41 = tensor.empty() : tensor<2x320x64x64xf16> | |
| %42 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_13 : tensor<320xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %43 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded, %cst_12 : tensor<2x4x66x66xf16>, tensor<320x4x3x3xf16>) outs(%42 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16> | |
| %collapsed = tensor.collapse_shape %43 [[0], [1], [2, 3]] : tensor<2x320x64x64xf16> into tensor<2x320x4096xf16> | |
| %expanded_729 = tensor.expand_shape %collapsed [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16> | |
| %44 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %45 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%44 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %46 = tensor.empty() : tensor<2x32x10x4096xf32> | |
| %47 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%45 : tensor<f32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %48 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_729 : tensor<2x32x10x4096xf16>) outs(%47 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %49 = tensor.empty() : tensor<2x32x10x4096xf64> | |
| %50 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%48 : tensor<2x32x10x4096xf32>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %51 = tensor.empty() : tensor<2x32x1x1xf64> | |
| %52 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %53 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%50 : tensor<2x32x10x4096xf64>) outs(%52 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %54 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%53 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_702 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %55 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%50, %54 : tensor<2x32x10x4096xf64>, tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %56 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%55 : tensor<2x32x10x4096xf64>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %57 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %58 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%56 : tensor<2x32x10x4096xf64>) outs(%57 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %59 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%58 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_702 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %60 = tensor.empty() : tensor<2x32x1x1xf32> | |
| %61 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%59 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %62 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %63 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%48 : tensor<2x32x10x4096xf32>) outs(%62 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %64 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%63 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_703 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %65 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%61, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %66 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%65 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %67 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_729, %64 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %68 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%67, %66 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %collapsed_730 = tensor.collapse_shape %68 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32> | |
| %expanded_731 = tensor.expand_shape %collapsed_730 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32> | |
| %expanded_732 = tensor.expand_shape %cst_14 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16> | |
| %69 = tensor.empty() : tensor<2x320x64x64xf32> | |
| %70 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_731, %expanded_732 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%69 : tensor<2x320x64x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x320x64x64xf32> | |
| %expanded_733 = tensor.expand_shape %cst_15 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16> | |
| %71 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%70, %expanded_733 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%69 : tensor<2x320x64x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x320x64x64xf32> | |
| %72 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %73 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%72 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %74 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%73 : tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %75 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%71 : tensor<2x320x64x64xf32>) outs(%74 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %76 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%75 : tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %77 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%76, %75 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %padded_734 = tensor.pad %77 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x320x64x64xf16> to tensor<2x320x66x66xf16> | |
| %78 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_17 : tensor<320xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %79 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_734, %cst_16 : tensor<2x320x66x66xf16>, tensor<320x320x3x3xf16>) outs(%78 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16> | |
| %80 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280xf16> | |
| %81 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%80, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %82 = tensor.empty() : tensor<1280x320xf16> | |
| %83 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_18 : tensor<320x1280xf16>) outs(%82 : tensor<1280x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x320xf16> | |
| %84 = linalg.fill ins(%cst_694 : f16) outs(%25 : tensor<2x320xf16>) -> tensor<2x320xf16> | |
| %85 = linalg.matmul ins(%81, %83 : tensor<2x1280xf16>, tensor<1280x320xf16>) outs(%84 : tensor<2x320xf16>) -> tensor<2x320xf16> | |
| %86 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_19, %85 : tensor<320xf16>, tensor<2x320xf16>) outs(%25 : tensor<2x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320xf16> | |
| %expanded_735 = tensor.expand_shape %86 [[0], [1, 2, 3]] : tensor<2x320xf16> into tensor<2x320x1x1xf16> | |
| %87 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%79, %expanded_735 : tensor<2x320x64x64xf16>, tensor<2x320x1x1xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %collapsed_736 = tensor.collapse_shape %87 [[0], [1], [2, 3]] : tensor<2x320x64x64xf16> into tensor<2x320x4096xf16> | |
| %expanded_737 = tensor.expand_shape %collapsed_736 [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16> | |
| %88 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %89 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%88 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %90 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%89 : tensor<f32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %91 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_737 : tensor<2x32x10x4096xf16>) outs(%90 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %92 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%91 : tensor<2x32x10x4096xf32>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %93 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %94 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%92 : tensor<2x32x10x4096xf64>) outs(%93 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %95 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%94 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_702 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %96 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%92, %95 : tensor<2x32x10x4096xf64>, tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %97 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%96 : tensor<2x32x10x4096xf64>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %98 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %99 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%97 : tensor<2x32x10x4096xf64>) outs(%98 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %100 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%99 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_702 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %101 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%100 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %102 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %103 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%91 : tensor<2x32x10x4096xf32>) outs(%102 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %104 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%103 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_703 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %105 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%101, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %106 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%105 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %107 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_737, %104 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %108 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%107, %106 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %collapsed_738 = tensor.collapse_shape %108 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32> | |
| %expanded_739 = tensor.expand_shape %collapsed_738 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32> | |
| %expanded_740 = tensor.expand_shape %cst_20 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16> | |
| %109 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_739, %expanded_740 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%69 : tensor<2x320x64x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x320x64x64xf32> | |
| %expanded_741 = tensor.expand_shape %cst_21 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16> | |
| %110 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%109, %expanded_741 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%69 : tensor<2x320x64x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x320x64x64xf32> | |
| %111 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %112 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%111 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %113 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%112 : tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %114 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%110 : tensor<2x320x64x64xf32>) outs(%113 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %115 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%114 : tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %116 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%115, %114 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %padded_742 = tensor.pad %116 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x320x64x64xf16> to tensor<2x320x66x66xf16> | |
| %117 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_23 : tensor<320xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %118 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_742, %cst_22 : tensor<2x320x66x66xf16>, tensor<320x320x3x3xf16>) outs(%117 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16> | |
| %119 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%43, %118 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %120 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%119, %cst_3 : tensor<2x320x64x64xf16>, tensor<f64>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %collapsed_743 = tensor.collapse_shape %120 [[0], [1], [2, 3]] : tensor<2x320x64x64xf16> into tensor<2x320x4096xf16> | |
| %expanded_744 = tensor.expand_shape %collapsed_743 [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16> | |
| %121 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %122 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%121 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %123 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%122 : tensor<f32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %124 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_744 : tensor<2x32x10x4096xf16>) outs(%123 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %125 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%124 : tensor<2x32x10x4096xf32>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %126 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %127 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%125 : tensor<2x32x10x4096xf64>) outs(%126 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %128 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%127 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_702 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %129 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%125, %128 : tensor<2x32x10x4096xf64>, tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %130 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%129 : tensor<2x32x10x4096xf64>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %131 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %132 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%130 : tensor<2x32x10x4096xf64>) outs(%131 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %133 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%132 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_702 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %134 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%133 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %135 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %136 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%124 : tensor<2x32x10x4096xf32>) outs(%135 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %137 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%136 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_703 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %138 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%134, %cst_2 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %139 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%138 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %140 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_744, %137 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %141 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%140, %139 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %collapsed_745 = tensor.collapse_shape %141 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32> | |
| %expanded_746 = tensor.expand_shape %collapsed_745 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32> | |
| %expanded_747 = tensor.expand_shape %cst_24 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16> | |
| %142 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_746, %expanded_747 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%69 : tensor<2x320x64x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x320x64x64xf32> | |
| %expanded_748 = tensor.expand_shape %cst_25 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16> | |
| %143 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%142, %expanded_748 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%69 : tensor<2x320x64x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x320x64x64xf32> | |
| %144 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %145 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%144 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %146 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%145 : tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %147 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%143 : tensor<2x320x64x64xf32>) outs(%146 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %148 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_27 : tensor<320xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %149 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%147, %cst_26 : tensor<2x320x64x64xf16>, tensor<320x320x1x1xf16>) outs(%148 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16> | |
| %150 = tensor.empty() : tensor<2x64x64x320xf16> | |
| %151 = linalg.generic {indexing_maps = [#map11, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%149 : tensor<2x320x64x64xf16>) outs(%150 : tensor<2x64x64x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x64x64x320xf16> | |
| %collapsed_749 = tensor.collapse_shape %151 [[0], [1, 2], [3]] : tensor<2x64x64x320xf16> into tensor<2x4096x320xf16> | |
| %152 = tensor.empty() : tensor<2x4096x1xf16> | |
| %153 = linalg.fill ins(%cst_694 : f16) outs(%152 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16> | |
| %154 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_749 : tensor<2x4096x320xf16>) outs(%153 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %155 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%154 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_704 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %156 = tensor.empty() : tensor<2x4096x320xf16> | |
| %157 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%155 : tensor<2x4096x1xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %158 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_749, %157 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %159 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%158 : tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %160 = linalg.fill ins(%cst_694 : f16) outs(%152 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16> | |
| %161 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%159 : tensor<2x4096x320xf16>) outs(%160 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %162 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%161 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_704 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %163 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%162 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %164 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%163 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %165 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%164 : tensor<2x4096x1xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %166 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%158, %165 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %167 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%166, %cst_28 : tensor<2x4096x320xf16>, tensor<320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %168 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%167, %cst_29 : tensor<2x4096x320xf16>, tensor<320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %169 = tensor.empty() : tensor<320x320xf16> | |
| %170 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_30 : tensor<320x320xf16>) outs(%169 : tensor<320x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<320x320xf16> | |
| %collapsed_750 = tensor.collapse_shape %168 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16> | |
| %171 = tensor.empty() : tensor<8192x320xf16> | |
| %172 = linalg.fill ins(%cst_694 : f16) outs(%171 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %173 = linalg.matmul ins(%collapsed_750, %170 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%172 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %174 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_31 : tensor<320x320xf16>) outs(%169 : tensor<320x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<320x320xf16> | |
| %175 = linalg.fill ins(%cst_694 : f16) outs(%171 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %176 = linalg.matmul ins(%collapsed_750, %174 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%175 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %177 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_32 : tensor<320x320xf16>) outs(%169 : tensor<320x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<320x320xf16> | |
| %178 = linalg.fill ins(%cst_694 : f16) outs(%171 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %179 = linalg.matmul ins(%collapsed_750, %177 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%178 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %expanded_751 = tensor.expand_shape %173 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x8x40xf16> | |
| %180 = tensor.empty() : tensor<2x8x4096x40xf16> | |
| %181 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_751 : tensor<2x4096x8x40xf16>) outs(%180 : tensor<2x8x4096x40xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x4096x40xf16> | |
| %collapsed_752 = tensor.collapse_shape %181 [[0, 1], [2], [3]] : tensor<2x8x4096x40xf16> into tensor<16x4096x40xf16> | |
| %expanded_753 = tensor.expand_shape %176 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x8x40xf16> | |
| %182 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_753 : tensor<2x4096x8x40xf16>) outs(%180 : tensor<2x8x4096x40xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x4096x40xf16> | |
| %collapsed_754 = tensor.collapse_shape %182 [[0, 1], [2], [3]] : tensor<2x8x4096x40xf16> into tensor<16x4096x40xf16> | |
| %expanded_755 = tensor.expand_shape %179 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x8x40xf16> | |
| %183 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_755 : tensor<2x4096x8x40xf16>) outs(%180 : tensor<2x8x4096x40xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x4096x40xf16> | |
| %collapsed_756 = tensor.collapse_shape %183 [[0, 1], [2], [3]] : tensor<2x8x4096x40xf16> into tensor<16x4096x40xf16> | |
| %184 = tensor.empty() : tensor<16x40x4096xf16> | |
| %185 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_754 : tensor<16x4096x40xf16>) outs(%184 : tensor<16x40x4096xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x40x4096xf16> | |
| %186 = tensor.empty() : tensor<16x4096x4096xf16> | |
| %187 = linalg.fill ins(%cst_694 : f16) outs(%186 : tensor<16x4096x4096xf16>) -> tensor<16x4096x4096xf16> | |
| %188 = linalg.batch_matmul ins(%collapsed_752, %185 : tensor<16x4096x40xf16>, tensor<16x40x4096xf16>) outs(%187 : tensor<16x4096x4096xf16>) -> tensor<16x4096x4096xf16> | |
| %189 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%188, %cst_1 : tensor<16x4096x4096xf16>, tensor<f64>) outs(%186 : tensor<16x4096x4096xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x4096x4096xf16> | |
| %190 = tensor.empty() : tensor<16x4096x1xi64> | |
| %191 = linalg.fill ins(%c0_i64 : i64) outs(%190 : tensor<16x4096x1xi64>) -> tensor<16x4096x1xi64> | |
| %192 = tensor.empty() : tensor<16x4096x1xf16> | |
| %193 = linalg.fill ins(%cst_696 : f16) outs(%192 : tensor<16x4096x1xf16>) -> tensor<16x4096x1xf16> | |
| %194:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%189 : tensor<16x4096x4096xf16>) outs(%193, %191 : tensor<16x4096x1xf16>, tensor<16x4096x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x4096x1xf16>, tensor<16x4096x1xi64>) | |
| %195 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%189, %194#0 : tensor<16x4096x4096xf16>, tensor<16x4096x1xf16>) outs(%186 : tensor<16x4096x4096xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x4096x4096xf16> | |
| %196 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%195 : tensor<16x4096x4096xf16>) outs(%186 : tensor<16x4096x4096xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x4096x4096xf16> | |
| %197 = linalg.fill ins(%cst_694 : f16) outs(%192 : tensor<16x4096x1xf16>) -> tensor<16x4096x1xf16> | |
| %198 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%196 : tensor<16x4096x4096xf16>) outs(%197 : tensor<16x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x4096x1xf16> | |
| %199 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%196, %198 : tensor<16x4096x4096xf16>, tensor<16x4096x1xf16>) outs(%186 : tensor<16x4096x4096xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x4096x4096xf16> | |
| %200 = tensor.empty() : tensor<16x4096x40xf16> | |
| %201 = linalg.fill ins(%cst_694 : f16) outs(%200 : tensor<16x4096x40xf16>) -> tensor<16x4096x40xf16> | |
| %202 = linalg.batch_matmul ins(%199, %collapsed_756 : tensor<16x4096x4096xf16>, tensor<16x4096x40xf16>) outs(%201 : tensor<16x4096x40xf16>) -> tensor<16x4096x40xf16> | |
| %expanded_757 = tensor.expand_shape %202 [[0, 1], [2], [3]] : tensor<16x4096x40xf16> into tensor<2x8x4096x40xf16> | |
| %203 = tensor.empty() : tensor<2x4096x8x40xf16> | |
| %204 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_757 : tensor<2x8x4096x40xf16>) outs(%203 : tensor<2x4096x8x40xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x4096x8x40xf16> | |
| %205 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_33 : tensor<320x320xf16>) outs(%169 : tensor<320x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<320x320xf16> | |
| %collapsed_758 = tensor.collapse_shape %204 [[0, 1], [2, 3]] : tensor<2x4096x8x40xf16> into tensor<8192x320xf16> | |
| %206 = linalg.fill ins(%cst_694 : f16) outs(%171 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %207 = linalg.matmul ins(%collapsed_758, %205 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%206 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %208 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_34, %207 : tensor<320xf16>, tensor<8192x320xf16>) outs(%171 : tensor<8192x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<8192x320xf16> | |
| %expanded_759 = tensor.expand_shape %208 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16> | |
| %209 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_759, %collapsed_749 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %210 = linalg.fill ins(%cst_694 : f16) outs(%152 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16> | |
| %211 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%209 : tensor<2x4096x320xf16>) outs(%210 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %212 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%211 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_704 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %213 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%212 : tensor<2x4096x1xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %214 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%209, %213 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %215 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%214 : tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %216 = linalg.fill ins(%cst_694 : f16) outs(%152 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16> | |
| %217 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%215 : tensor<2x4096x320xf16>) outs(%216 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %218 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%217 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_704 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %219 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%218 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %220 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%219 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %221 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%220 : tensor<2x4096x1xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %222 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%214, %221 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %223 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%222, %cst_35 : tensor<2x4096x320xf16>, tensor<320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %224 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%223, %cst_36 : tensor<2x4096x320xf16>, tensor<320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %225 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_37 : tensor<320x320xf16>) outs(%169 : tensor<320x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<320x320xf16> | |
| %collapsed_760 = tensor.collapse_shape %224 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16> | |
| %226 = linalg.fill ins(%cst_694 : f16) outs(%171 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %227 = linalg.matmul ins(%collapsed_760, %225 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%226 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %228 = tensor.empty() : tensor<768x320xf16> | |
| %229 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_38 : tensor<320x768xf16>) outs(%228 : tensor<768x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x320xf16> | |
| %collapsed_761 = tensor.collapse_shape %arg2 [[0, 1], [2]] : tensor<2x77x768xf16> into tensor<154x768xf16> | |
| %230 = tensor.empty() : tensor<154x320xf16> | |
| %231 = linalg.fill ins(%cst_694 : f16) outs(%230 : tensor<154x320xf16>) -> tensor<154x320xf16> | |
| %232 = linalg.matmul ins(%collapsed_761, %229 : tensor<154x768xf16>, tensor<768x320xf16>) outs(%231 : tensor<154x320xf16>) -> tensor<154x320xf16> | |
| %233 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_39 : tensor<320x768xf16>) outs(%228 : tensor<768x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x320xf16> | |
| %234 = linalg.fill ins(%cst_694 : f16) outs(%230 : tensor<154x320xf16>) -> tensor<154x320xf16> | |
| %235 = linalg.matmul ins(%collapsed_761, %233 : tensor<154x768xf16>, tensor<768x320xf16>) outs(%234 : tensor<154x320xf16>) -> tensor<154x320xf16> | |
| %expanded_762 = tensor.expand_shape %227 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x8x40xf16> | |
| %236 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_762 : tensor<2x4096x8x40xf16>) outs(%180 : tensor<2x8x4096x40xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x4096x40xf16> | |
| %collapsed_763 = tensor.collapse_shape %236 [[0, 1], [2], [3]] : tensor<2x8x4096x40xf16> into tensor<16x4096x40xf16> | |
| %expanded_764 = tensor.expand_shape %232 [[0, 1], [2, 3]] : tensor<154x320xf16> into tensor<2x77x8x40xf16> | |
| %237 = tensor.empty() : tensor<2x8x77x40xf16> | |
| %238 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_764 : tensor<2x77x8x40xf16>) outs(%237 : tensor<2x8x77x40xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x40xf16> | |
| %collapsed_765 = tensor.collapse_shape %238 [[0, 1], [2], [3]] : tensor<2x8x77x40xf16> into tensor<16x77x40xf16> | |
| %expanded_766 = tensor.expand_shape %235 [[0, 1], [2, 3]] : tensor<154x320xf16> into tensor<2x77x8x40xf16> | |
| %239 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_766 : tensor<2x77x8x40xf16>) outs(%237 : tensor<2x8x77x40xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x40xf16> | |
| %collapsed_767 = tensor.collapse_shape %239 [[0, 1], [2], [3]] : tensor<2x8x77x40xf16> into tensor<16x77x40xf16> | |
| %240 = tensor.empty() : tensor<16x40x77xf16> | |
| %241 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_765 : tensor<16x77x40xf16>) outs(%240 : tensor<16x40x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x40x77xf16> | |
| %242 = tensor.empty() : tensor<16x4096x77xf16> | |
| %243 = linalg.fill ins(%cst_694 : f16) outs(%242 : tensor<16x4096x77xf16>) -> tensor<16x4096x77xf16> | |
| %244 = linalg.batch_matmul ins(%collapsed_763, %241 : tensor<16x4096x40xf16>, tensor<16x40x77xf16>) outs(%243 : tensor<16x4096x77xf16>) -> tensor<16x4096x77xf16> | |
| %245 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%244, %cst_1 : tensor<16x4096x77xf16>, tensor<f64>) outs(%242 : tensor<16x4096x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x4096x77xf16> | |
| %246 = linalg.fill ins(%c0_i64 : i64) outs(%190 : tensor<16x4096x1xi64>) -> tensor<16x4096x1xi64> | |
| %247 = linalg.fill ins(%cst_696 : f16) outs(%192 : tensor<16x4096x1xf16>) -> tensor<16x4096x1xf16> | |
| %248:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%245 : tensor<16x4096x77xf16>) outs(%247, %246 : tensor<16x4096x1xf16>, tensor<16x4096x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x4096x1xf16>, tensor<16x4096x1xi64>) | |
| %249 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%245, %248#0 : tensor<16x4096x77xf16>, tensor<16x4096x1xf16>) outs(%242 : tensor<16x4096x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x4096x77xf16> | |
| %250 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%249 : tensor<16x4096x77xf16>) outs(%242 : tensor<16x4096x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x4096x77xf16> | |
| %251 = linalg.fill ins(%cst_694 : f16) outs(%192 : tensor<16x4096x1xf16>) -> tensor<16x4096x1xf16> | |
| %252 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%250 : tensor<16x4096x77xf16>) outs(%251 : tensor<16x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x4096x1xf16> | |
| %253 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%250, %252 : tensor<16x4096x77xf16>, tensor<16x4096x1xf16>) outs(%242 : tensor<16x4096x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x4096x77xf16> | |
| %254 = linalg.fill ins(%cst_694 : f16) outs(%200 : tensor<16x4096x40xf16>) -> tensor<16x4096x40xf16> | |
| %255 = linalg.batch_matmul ins(%253, %collapsed_767 : tensor<16x4096x77xf16>, tensor<16x77x40xf16>) outs(%254 : tensor<16x4096x40xf16>) -> tensor<16x4096x40xf16> | |
| %expanded_768 = tensor.expand_shape %255 [[0, 1], [2], [3]] : tensor<16x4096x40xf16> into tensor<2x8x4096x40xf16> | |
| %256 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_768 : tensor<2x8x4096x40xf16>) outs(%203 : tensor<2x4096x8x40xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x4096x8x40xf16> | |
| %257 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_40 : tensor<320x320xf16>) outs(%169 : tensor<320x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<320x320xf16> | |
| %collapsed_769 = tensor.collapse_shape %256 [[0, 1], [2, 3]] : tensor<2x4096x8x40xf16> into tensor<8192x320xf16> | |
| %258 = linalg.fill ins(%cst_694 : f16) outs(%171 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %259 = linalg.matmul ins(%collapsed_769, %257 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%258 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %260 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_41, %259 : tensor<320xf16>, tensor<8192x320xf16>) outs(%171 : tensor<8192x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<8192x320xf16> | |
| %expanded_770 = tensor.expand_shape %260 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16> | |
| %261 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_770, %209 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %262 = linalg.fill ins(%cst_694 : f16) outs(%152 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16> | |
| %263 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%261 : tensor<2x4096x320xf16>) outs(%262 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %264 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%263 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_704 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %265 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%264 : tensor<2x4096x1xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %266 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%261, %265 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %267 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%266 : tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %268 = linalg.fill ins(%cst_694 : f16) outs(%152 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16> | |
| %269 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%267 : tensor<2x4096x320xf16>) outs(%268 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %270 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%269 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_704 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %271 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%270 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %272 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%271 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %273 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%272 : tensor<2x4096x1xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %274 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%266, %273 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %275 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%274, %cst_42 : tensor<2x4096x320xf16>, tensor<320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %276 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%275, %cst_43 : tensor<2x4096x320xf16>, tensor<320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %277 = tensor.empty() : tensor<320x2560xf16> | |
| %278 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_44 : tensor<2560x320xf16>) outs(%277 : tensor<320x2560xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<320x2560xf16> | |
| %collapsed_771 = tensor.collapse_shape %276 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16> | |
| %279 = tensor.empty() : tensor<8192x2560xf16> | |
| %280 = linalg.fill ins(%cst_694 : f16) outs(%279 : tensor<8192x2560xf16>) -> tensor<8192x2560xf16> | |
| %281 = linalg.matmul ins(%collapsed_771, %278 : tensor<8192x320xf16>, tensor<320x2560xf16>) outs(%280 : tensor<8192x2560xf16>) -> tensor<8192x2560xf16> | |
| %282 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_45, %281 : tensor<2560xf16>, tensor<8192x2560xf16>) outs(%279 : tensor<8192x2560xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<8192x2560xf16> | |
| %expanded_772 = tensor.expand_shape %282 [[0, 1], [2]] : tensor<8192x2560xf16> into tensor<2x4096x2560xf16> | |
| %extracted_slice_773 = tensor.extract_slice %expanded_772[0, 0, 0] [2, 4096, 1280] [1, 1, 1] : tensor<2x4096x2560xf16> to tensor<2x4096x1280xf16> | |
| %extracted_slice_774 = tensor.extract_slice %expanded_772[0, 0, 1280] [2, 4096, 1280] [1, 1, 1] : tensor<2x4096x2560xf16> to tensor<2x4096x1280xf16> | |
| %283 = tensor.empty() : tensor<2x4096x1280xf16> | |
| %284 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_774 : tensor<2x4096x1280xf16>) outs(%283 : tensor<2x4096x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.sqrt %cst_697 : f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| %4282 = math.erf %4281 : f16 | |
| %4283 = arith.addf %4282, %cst_695 : f16 | |
| %4284 = arith.mulf %4283, %cst_698 : f16 | |
| %4285 = arith.mulf %in, %4284 : f16 | |
| linalg.yield %4285 : f16 | |
| } -> tensor<2x4096x1280xf16> | |
| %285 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_773, %284 : tensor<2x4096x1280xf16>, tensor<2x4096x1280xf16>) outs(%283 : tensor<2x4096x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1280xf16> | |
| %286 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_46 : tensor<320x1280xf16>) outs(%82 : tensor<1280x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x320xf16> | |
| %collapsed_775 = tensor.collapse_shape %285 [[0, 1], [2]] : tensor<2x4096x1280xf16> into tensor<8192x1280xf16> | |
| %287 = linalg.fill ins(%cst_694 : f16) outs(%171 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %288 = linalg.matmul ins(%collapsed_775, %286 : tensor<8192x1280xf16>, tensor<1280x320xf16>) outs(%287 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %289 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_47, %288 : tensor<320xf16>, tensor<8192x320xf16>) outs(%171 : tensor<8192x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<8192x320xf16> | |
| %expanded_776 = tensor.expand_shape %289 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16> | |
| %290 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_776, %261 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %expanded_777 = tensor.expand_shape %290 [[0], [1, 2], [3]] : tensor<2x4096x320xf16> into tensor<2x64x64x320xf16> | |
| %291 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_777 : tensor<2x64x64x320xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %292 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_49 : tensor<320xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %293 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%291, %cst_48 : tensor<2x320x64x64xf16>, tensor<320x320x1x1xf16>) outs(%292 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16> | |
| %294 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%293, %120 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %collapsed_778 = tensor.collapse_shape %294 [[0], [1], [2, 3]] : tensor<2x320x64x64xf16> into tensor<2x320x4096xf16> | |
| %expanded_779 = tensor.expand_shape %collapsed_778 [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16> | |
| %295 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %296 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%295 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %297 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%296 : tensor<f32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %298 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_779 : tensor<2x32x10x4096xf16>) outs(%297 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %299 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%298 : tensor<2x32x10x4096xf32>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %300 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %301 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%299 : tensor<2x32x10x4096xf64>) outs(%300 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %302 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%301 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_702 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %303 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%299, %302 : tensor<2x32x10x4096xf64>, tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %304 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%303 : tensor<2x32x10x4096xf64>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %305 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %306 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%304 : tensor<2x32x10x4096xf64>) outs(%305 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %307 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%306 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_702 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %308 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%307 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %309 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %310 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%298 : tensor<2x32x10x4096xf32>) outs(%309 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %311 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%310 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_703 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %312 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%308, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %313 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%312 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %314 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_779, %311 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %315 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%314, %313 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %collapsed_780 = tensor.collapse_shape %315 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32> | |
| %expanded_781 = tensor.expand_shape %collapsed_780 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32> | |
| %expanded_782 = tensor.expand_shape %cst_50 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16> | |
| %316 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_781, %expanded_782 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%69 : tensor<2x320x64x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x320x64x64xf32> | |
| %expanded_783 = tensor.expand_shape %cst_51 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16> | |
| %317 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%316, %expanded_783 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%69 : tensor<2x320x64x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x320x64x64xf32> | |
| %318 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %319 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%318 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %320 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%319 : tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %321 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%317 : tensor<2x320x64x64xf32>) outs(%320 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %322 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%321 : tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %323 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%322, %321 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %padded_784 = tensor.pad %323 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x320x64x64xf16> to tensor<2x320x66x66xf16> | |
| %324 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_53 : tensor<320xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %325 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_784, %cst_52 : tensor<2x320x66x66xf16>, tensor<320x320x3x3xf16>) outs(%324 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16> | |
| %326 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280xf16> | |
| %327 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%326, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %328 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_54 : tensor<320x1280xf16>) outs(%82 : tensor<1280x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x320xf16> | |
| %329 = linalg.fill ins(%cst_694 : f16) outs(%25 : tensor<2x320xf16>) -> tensor<2x320xf16> | |
| %330 = linalg.matmul ins(%327, %328 : tensor<2x1280xf16>, tensor<1280x320xf16>) outs(%329 : tensor<2x320xf16>) -> tensor<2x320xf16> | |
| %331 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_55, %330 : tensor<320xf16>, tensor<2x320xf16>) outs(%25 : tensor<2x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320xf16> | |
| %expanded_785 = tensor.expand_shape %331 [[0], [1, 2, 3]] : tensor<2x320xf16> into tensor<2x320x1x1xf16> | |
| %332 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%325, %expanded_785 : tensor<2x320x64x64xf16>, tensor<2x320x1x1xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %collapsed_786 = tensor.collapse_shape %332 [[0], [1], [2, 3]] : tensor<2x320x64x64xf16> into tensor<2x320x4096xf16> | |
| %expanded_787 = tensor.expand_shape %collapsed_786 [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16> | |
| %333 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %334 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%333 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %335 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%334 : tensor<f32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %336 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_787 : tensor<2x32x10x4096xf16>) outs(%335 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %337 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%336 : tensor<2x32x10x4096xf32>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %338 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %339 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%337 : tensor<2x32x10x4096xf64>) outs(%338 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %340 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%339 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_702 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %341 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%337, %340 : tensor<2x32x10x4096xf64>, tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %342 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%341 : tensor<2x32x10x4096xf64>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %343 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %344 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%342 : tensor<2x32x10x4096xf64>) outs(%343 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %345 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%344 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_702 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %346 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%345 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %347 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %348 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%336 : tensor<2x32x10x4096xf32>) outs(%347 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %349 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%348 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_703 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %350 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%346, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %351 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%350 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %352 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_787, %349 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %353 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%352, %351 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %collapsed_788 = tensor.collapse_shape %353 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32> | |
| %expanded_789 = tensor.expand_shape %collapsed_788 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32> | |
| %expanded_790 = tensor.expand_shape %cst_56 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16> | |
| %354 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_789, %expanded_790 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%69 : tensor<2x320x64x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x320x64x64xf32> | |
| %expanded_791 = tensor.expand_shape %cst_57 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16> | |
| %355 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%354, %expanded_791 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%69 : tensor<2x320x64x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x320x64x64xf32> | |
| %356 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %357 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%356 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %358 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%357 : tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %359 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%355 : tensor<2x320x64x64xf32>) outs(%358 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %360 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%359 : tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %361 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%360, %359 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %padded_792 = tensor.pad %361 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x320x64x64xf16> to tensor<2x320x66x66xf16> | |
| %362 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_59 : tensor<320xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %363 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_792, %cst_58 : tensor<2x320x66x66xf16>, tensor<320x320x3x3xf16>) outs(%362 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16> | |
| %364 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%294, %363 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %365 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%364, %cst_3 : tensor<2x320x64x64xf16>, tensor<f64>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %collapsed_793 = tensor.collapse_shape %365 [[0], [1], [2, 3]] : tensor<2x320x64x64xf16> into tensor<2x320x4096xf16> | |
| %expanded_794 = tensor.expand_shape %collapsed_793 [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16> | |
| %366 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %367 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%366 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %368 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%367 : tensor<f32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %369 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_794 : tensor<2x32x10x4096xf16>) outs(%368 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %370 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%369 : tensor<2x32x10x4096xf32>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %371 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %372 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%370 : tensor<2x32x10x4096xf64>) outs(%371 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %373 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%372 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_702 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %374 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%370, %373 : tensor<2x32x10x4096xf64>, tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %375 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%374 : tensor<2x32x10x4096xf64>) outs(%49 : tensor<2x32x10x4096xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x4096xf64> | |
| %376 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %377 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%375 : tensor<2x32x10x4096xf64>) outs(%376 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %378 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%377 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_702 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %379 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%378 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %380 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %381 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%369 : tensor<2x32x10x4096xf32>) outs(%380 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %382 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%381 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_703 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %383 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%379, %cst_2 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %384 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%383 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %385 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_794, %382 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %386 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%385, %384 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x10x4096xf32> | |
| %collapsed_795 = tensor.collapse_shape %386 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32> | |
| %expanded_796 = tensor.expand_shape %collapsed_795 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32> | |
| %expanded_797 = tensor.expand_shape %cst_60 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16> | |
| %387 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_796, %expanded_797 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%69 : tensor<2x320x64x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x320x64x64xf32> | |
| %expanded_798 = tensor.expand_shape %cst_61 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16> | |
| %388 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%387, %expanded_798 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%69 : tensor<2x320x64x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x320x64x64xf32> | |
| %389 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %390 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%389 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %391 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%390 : tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %392 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%388 : tensor<2x320x64x64xf32>) outs(%391 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %393 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_63 : tensor<320xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %394 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%392, %cst_62 : tensor<2x320x64x64xf16>, tensor<320x320x1x1xf16>) outs(%393 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16> | |
| %395 = linalg.generic {indexing_maps = [#map11, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%394 : tensor<2x320x64x64xf16>) outs(%150 : tensor<2x64x64x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x64x64x320xf16> | |
| %collapsed_799 = tensor.collapse_shape %395 [[0], [1, 2], [3]] : tensor<2x64x64x320xf16> into tensor<2x4096x320xf16> | |
| %396 = linalg.fill ins(%cst_694 : f16) outs(%152 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16> | |
| %397 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_799 : tensor<2x4096x320xf16>) outs(%396 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %398 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%397 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_704 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %399 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%398 : tensor<2x4096x1xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %400 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_799, %399 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %401 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%400 : tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %402 = linalg.fill ins(%cst_694 : f16) outs(%152 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16> | |
| %403 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%401 : tensor<2x4096x320xf16>) outs(%402 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %404 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%403 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_704 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %405 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%404 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %406 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%405 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %407 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%406 : tensor<2x4096x1xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %408 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%400, %407 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %409 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%408, %cst_64 : tensor<2x4096x320xf16>, tensor<320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %410 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%409, %cst_65 : tensor<2x4096x320xf16>, tensor<320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %411 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_66 : tensor<320x320xf16>) outs(%169 : tensor<320x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<320x320xf16> | |
| %collapsed_800 = tensor.collapse_shape %410 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16> | |
| %412 = linalg.fill ins(%cst_694 : f16) outs(%171 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %413 = linalg.matmul ins(%collapsed_800, %411 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%412 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %414 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_67 : tensor<320x320xf16>) outs(%169 : tensor<320x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<320x320xf16> | |
| %415 = linalg.fill ins(%cst_694 : f16) outs(%171 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %416 = linalg.matmul ins(%collapsed_800, %414 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%415 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %417 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_68 : tensor<320x320xf16>) outs(%169 : tensor<320x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<320x320xf16> | |
| %418 = linalg.fill ins(%cst_694 : f16) outs(%171 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %419 = linalg.matmul ins(%collapsed_800, %417 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%418 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %expanded_801 = tensor.expand_shape %413 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x8x40xf16> | |
| %420 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_801 : tensor<2x4096x8x40xf16>) outs(%180 : tensor<2x8x4096x40xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x4096x40xf16> | |
| %collapsed_802 = tensor.collapse_shape %420 [[0, 1], [2], [3]] : tensor<2x8x4096x40xf16> into tensor<16x4096x40xf16> | |
| %expanded_803 = tensor.expand_shape %416 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x8x40xf16> | |
| %421 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_803 : tensor<2x4096x8x40xf16>) outs(%180 : tensor<2x8x4096x40xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x4096x40xf16> | |
| %collapsed_804 = tensor.collapse_shape %421 [[0, 1], [2], [3]] : tensor<2x8x4096x40xf16> into tensor<16x4096x40xf16> | |
| %expanded_805 = tensor.expand_shape %419 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x8x40xf16> | |
| %422 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_805 : tensor<2x4096x8x40xf16>) outs(%180 : tensor<2x8x4096x40xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x4096x40xf16> | |
| %collapsed_806 = tensor.collapse_shape %422 [[0, 1], [2], [3]] : tensor<2x8x4096x40xf16> into tensor<16x4096x40xf16> | |
| %423 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_804 : tensor<16x4096x40xf16>) outs(%184 : tensor<16x40x4096xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x40x4096xf16> | |
| %424 = linalg.fill ins(%cst_694 : f16) outs(%186 : tensor<16x4096x4096xf16>) -> tensor<16x4096x4096xf16> | |
| %425 = linalg.batch_matmul ins(%collapsed_802, %423 : tensor<16x4096x40xf16>, tensor<16x40x4096xf16>) outs(%424 : tensor<16x4096x4096xf16>) -> tensor<16x4096x4096xf16> | |
| %426 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%425, %cst_1 : tensor<16x4096x4096xf16>, tensor<f64>) outs(%186 : tensor<16x4096x4096xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x4096x4096xf16> | |
| %427 = linalg.fill ins(%c0_i64 : i64) outs(%190 : tensor<16x4096x1xi64>) -> tensor<16x4096x1xi64> | |
| %428 = linalg.fill ins(%cst_696 : f16) outs(%192 : tensor<16x4096x1xf16>) -> tensor<16x4096x1xf16> | |
| %429:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%426 : tensor<16x4096x4096xf16>) outs(%428, %427 : tensor<16x4096x1xf16>, tensor<16x4096x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x4096x1xf16>, tensor<16x4096x1xi64>) | |
| %430 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%426, %429#0 : tensor<16x4096x4096xf16>, tensor<16x4096x1xf16>) outs(%186 : tensor<16x4096x4096xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x4096x4096xf16> | |
| %431 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%430 : tensor<16x4096x4096xf16>) outs(%186 : tensor<16x4096x4096xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x4096x4096xf16> | |
| %432 = linalg.fill ins(%cst_694 : f16) outs(%192 : tensor<16x4096x1xf16>) -> tensor<16x4096x1xf16> | |
| %433 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%431 : tensor<16x4096x4096xf16>) outs(%432 : tensor<16x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x4096x1xf16> | |
| %434 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%431, %433 : tensor<16x4096x4096xf16>, tensor<16x4096x1xf16>) outs(%186 : tensor<16x4096x4096xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x4096x4096xf16> | |
| %435 = linalg.fill ins(%cst_694 : f16) outs(%200 : tensor<16x4096x40xf16>) -> tensor<16x4096x40xf16> | |
| %436 = linalg.batch_matmul ins(%434, %collapsed_806 : tensor<16x4096x4096xf16>, tensor<16x4096x40xf16>) outs(%435 : tensor<16x4096x40xf16>) -> tensor<16x4096x40xf16> | |
| %expanded_807 = tensor.expand_shape %436 [[0, 1], [2], [3]] : tensor<16x4096x40xf16> into tensor<2x8x4096x40xf16> | |
| %437 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_807 : tensor<2x8x4096x40xf16>) outs(%203 : tensor<2x4096x8x40xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x4096x8x40xf16> | |
| %438 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_69 : tensor<320x320xf16>) outs(%169 : tensor<320x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<320x320xf16> | |
| %collapsed_808 = tensor.collapse_shape %437 [[0, 1], [2, 3]] : tensor<2x4096x8x40xf16> into tensor<8192x320xf16> | |
| %439 = linalg.fill ins(%cst_694 : f16) outs(%171 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %440 = linalg.matmul ins(%collapsed_808, %438 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%439 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %441 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_70, %440 : tensor<320xf16>, tensor<8192x320xf16>) outs(%171 : tensor<8192x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<8192x320xf16> | |
| %expanded_809 = tensor.expand_shape %441 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16> | |
| %442 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_809, %collapsed_799 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %443 = linalg.fill ins(%cst_694 : f16) outs(%152 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16> | |
| %444 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%442 : tensor<2x4096x320xf16>) outs(%443 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %445 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%444 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_704 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %446 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%445 : tensor<2x4096x1xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %447 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%442, %446 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %448 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%447 : tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %449 = linalg.fill ins(%cst_694 : f16) outs(%152 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16> | |
| %450 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%448 : tensor<2x4096x320xf16>) outs(%449 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %451 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%450 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_704 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %452 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%451 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %453 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%452 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %454 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%453 : tensor<2x4096x1xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %455 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%447, %454 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %456 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%455, %cst_71 : tensor<2x4096x320xf16>, tensor<320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %457 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%456, %cst_72 : tensor<2x4096x320xf16>, tensor<320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %458 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_73 : tensor<320x320xf16>) outs(%169 : tensor<320x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<320x320xf16> | |
| %collapsed_810 = tensor.collapse_shape %457 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16> | |
| %459 = linalg.fill ins(%cst_694 : f16) outs(%171 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %460 = linalg.matmul ins(%collapsed_810, %458 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%459 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %461 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_74 : tensor<320x768xf16>) outs(%228 : tensor<768x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x320xf16> | |
| %462 = linalg.fill ins(%cst_694 : f16) outs(%230 : tensor<154x320xf16>) -> tensor<154x320xf16> | |
| %463 = linalg.matmul ins(%collapsed_761, %461 : tensor<154x768xf16>, tensor<768x320xf16>) outs(%462 : tensor<154x320xf16>) -> tensor<154x320xf16> | |
| %464 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_75 : tensor<320x768xf16>) outs(%228 : tensor<768x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x320xf16> | |
| %465 = linalg.fill ins(%cst_694 : f16) outs(%230 : tensor<154x320xf16>) -> tensor<154x320xf16> | |
| %466 = linalg.matmul ins(%collapsed_761, %464 : tensor<154x768xf16>, tensor<768x320xf16>) outs(%465 : tensor<154x320xf16>) -> tensor<154x320xf16> | |
| %expanded_811 = tensor.expand_shape %460 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x8x40xf16> | |
| %467 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_811 : tensor<2x4096x8x40xf16>) outs(%180 : tensor<2x8x4096x40xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x4096x40xf16> | |
| %collapsed_812 = tensor.collapse_shape %467 [[0, 1], [2], [3]] : tensor<2x8x4096x40xf16> into tensor<16x4096x40xf16> | |
| %expanded_813 = tensor.expand_shape %463 [[0, 1], [2, 3]] : tensor<154x320xf16> into tensor<2x77x8x40xf16> | |
| %468 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_813 : tensor<2x77x8x40xf16>) outs(%237 : tensor<2x8x77x40xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x40xf16> | |
| %collapsed_814 = tensor.collapse_shape %468 [[0, 1], [2], [3]] : tensor<2x8x77x40xf16> into tensor<16x77x40xf16> | |
| %expanded_815 = tensor.expand_shape %466 [[0, 1], [2, 3]] : tensor<154x320xf16> into tensor<2x77x8x40xf16> | |
| %469 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_815 : tensor<2x77x8x40xf16>) outs(%237 : tensor<2x8x77x40xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x40xf16> | |
| %collapsed_816 = tensor.collapse_shape %469 [[0, 1], [2], [3]] : tensor<2x8x77x40xf16> into tensor<16x77x40xf16> | |
| %470 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_814 : tensor<16x77x40xf16>) outs(%240 : tensor<16x40x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x40x77xf16> | |
| %471 = linalg.fill ins(%cst_694 : f16) outs(%242 : tensor<16x4096x77xf16>) -> tensor<16x4096x77xf16> | |
| %472 = linalg.batch_matmul ins(%collapsed_812, %470 : tensor<16x4096x40xf16>, tensor<16x40x77xf16>) outs(%471 : tensor<16x4096x77xf16>) -> tensor<16x4096x77xf16> | |
| %473 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%472, %cst_1 : tensor<16x4096x77xf16>, tensor<f64>) outs(%242 : tensor<16x4096x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x4096x77xf16> | |
| %474 = linalg.fill ins(%c0_i64 : i64) outs(%190 : tensor<16x4096x1xi64>) -> tensor<16x4096x1xi64> | |
| %475 = linalg.fill ins(%cst_696 : f16) outs(%192 : tensor<16x4096x1xf16>) -> tensor<16x4096x1xf16> | |
| %476:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%473 : tensor<16x4096x77xf16>) outs(%475, %474 : tensor<16x4096x1xf16>, tensor<16x4096x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x4096x1xf16>, tensor<16x4096x1xi64>) | |
| %477 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%473, %476#0 : tensor<16x4096x77xf16>, tensor<16x4096x1xf16>) outs(%242 : tensor<16x4096x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x4096x77xf16> | |
| %478 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%477 : tensor<16x4096x77xf16>) outs(%242 : tensor<16x4096x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x4096x77xf16> | |
| %479 = linalg.fill ins(%cst_694 : f16) outs(%192 : tensor<16x4096x1xf16>) -> tensor<16x4096x1xf16> | |
| %480 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%478 : tensor<16x4096x77xf16>) outs(%479 : tensor<16x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x4096x1xf16> | |
| %481 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%478, %480 : tensor<16x4096x77xf16>, tensor<16x4096x1xf16>) outs(%242 : tensor<16x4096x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x4096x77xf16> | |
| %482 = linalg.fill ins(%cst_694 : f16) outs(%200 : tensor<16x4096x40xf16>) -> tensor<16x4096x40xf16> | |
| %483 = linalg.batch_matmul ins(%481, %collapsed_816 : tensor<16x4096x77xf16>, tensor<16x77x40xf16>) outs(%482 : tensor<16x4096x40xf16>) -> tensor<16x4096x40xf16> | |
| %expanded_817 = tensor.expand_shape %483 [[0, 1], [2], [3]] : tensor<16x4096x40xf16> into tensor<2x8x4096x40xf16> | |
| %484 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_817 : tensor<2x8x4096x40xf16>) outs(%203 : tensor<2x4096x8x40xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x4096x8x40xf16> | |
| %485 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_76 : tensor<320x320xf16>) outs(%169 : tensor<320x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<320x320xf16> | |
| %collapsed_818 = tensor.collapse_shape %484 [[0, 1], [2, 3]] : tensor<2x4096x8x40xf16> into tensor<8192x320xf16> | |
| %486 = linalg.fill ins(%cst_694 : f16) outs(%171 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %487 = linalg.matmul ins(%collapsed_818, %485 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%486 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %488 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_77, %487 : tensor<320xf16>, tensor<8192x320xf16>) outs(%171 : tensor<8192x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<8192x320xf16> | |
| %expanded_819 = tensor.expand_shape %488 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16> | |
| %489 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_819, %442 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %490 = linalg.fill ins(%cst_694 : f16) outs(%152 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16> | |
| %491 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%489 : tensor<2x4096x320xf16>) outs(%490 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %492 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%491 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_704 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %493 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%492 : tensor<2x4096x1xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %494 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%489, %493 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %495 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%494 : tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %496 = linalg.fill ins(%cst_694 : f16) outs(%152 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16> | |
| %497 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%495 : tensor<2x4096x320xf16>) outs(%496 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %498 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%497 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_704 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %499 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%498 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %500 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%499 : tensor<2x4096x1xf16>) outs(%152 : tensor<2x4096x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1xf16> | |
| %501 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%500 : tensor<2x4096x1xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %502 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%494, %501 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %503 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%502, %cst_78 : tensor<2x4096x320xf16>, tensor<320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %504 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%503, %cst_79 : tensor<2x4096x320xf16>, tensor<320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %505 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_80 : tensor<2560x320xf16>) outs(%277 : tensor<320x2560xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<320x2560xf16> | |
| %collapsed_820 = tensor.collapse_shape %504 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16> | |
| %506 = linalg.fill ins(%cst_694 : f16) outs(%279 : tensor<8192x2560xf16>) -> tensor<8192x2560xf16> | |
| %507 = linalg.matmul ins(%collapsed_820, %505 : tensor<8192x320xf16>, tensor<320x2560xf16>) outs(%506 : tensor<8192x2560xf16>) -> tensor<8192x2560xf16> | |
| %508 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_81, %507 : tensor<2560xf16>, tensor<8192x2560xf16>) outs(%279 : tensor<8192x2560xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<8192x2560xf16> | |
| %expanded_821 = tensor.expand_shape %508 [[0, 1], [2]] : tensor<8192x2560xf16> into tensor<2x4096x2560xf16> | |
| %extracted_slice_822 = tensor.extract_slice %expanded_821[0, 0, 0] [2, 4096, 1280] [1, 1, 1] : tensor<2x4096x2560xf16> to tensor<2x4096x1280xf16> | |
| %extracted_slice_823 = tensor.extract_slice %expanded_821[0, 0, 1280] [2, 4096, 1280] [1, 1, 1] : tensor<2x4096x2560xf16> to tensor<2x4096x1280xf16> | |
| %509 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_823 : tensor<2x4096x1280xf16>) outs(%283 : tensor<2x4096x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.sqrt %cst_697 : f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| %4282 = math.erf %4281 : f16 | |
| %4283 = arith.addf %4282, %cst_695 : f16 | |
| %4284 = arith.mulf %4283, %cst_698 : f16 | |
| %4285 = arith.mulf %in, %4284 : f16 | |
| linalg.yield %4285 : f16 | |
| } -> tensor<2x4096x1280xf16> | |
| %510 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_822, %509 : tensor<2x4096x1280xf16>, tensor<2x4096x1280xf16>) outs(%283 : tensor<2x4096x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x1280xf16> | |
| %511 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_82 : tensor<320x1280xf16>) outs(%82 : tensor<1280x320xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x320xf16> | |
| %collapsed_824 = tensor.collapse_shape %510 [[0, 1], [2]] : tensor<2x4096x1280xf16> into tensor<8192x1280xf16> | |
| %512 = linalg.fill ins(%cst_694 : f16) outs(%171 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %513 = linalg.matmul ins(%collapsed_824, %511 : tensor<8192x1280xf16>, tensor<1280x320xf16>) outs(%512 : tensor<8192x320xf16>) -> tensor<8192x320xf16> | |
| %514 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_83, %513 : tensor<320xf16>, tensor<8192x320xf16>) outs(%171 : tensor<8192x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<8192x320xf16> | |
| %expanded_825 = tensor.expand_shape %514 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16> | |
| %515 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_825, %489 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%156 : tensor<2x4096x320xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x4096x320xf16> | |
| %expanded_826 = tensor.expand_shape %515 [[0], [1, 2], [3]] : tensor<2x4096x320xf16> into tensor<2x64x64x320xf16> | |
| %516 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_826 : tensor<2x64x64x320xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %517 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_85 : tensor<320xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %518 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%516, %cst_84 : tensor<2x320x64x64xf16>, tensor<320x320x1x1xf16>) outs(%517 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16> | |
| %519 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%518, %365 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x64x64xf16> | |
| %padded_827 = tensor.pad %519 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x320x64x64xf16> to tensor<2x320x66x66xf16> | |
| %520 = tensor.empty() : tensor<2x320x32x32xf16> | |
| %521 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_87 : tensor<320xf16>) outs(%520 : tensor<2x320x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x32x32xf16> | |
| %522 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%padded_827, %cst_86 : tensor<2x320x66x66xf16>, tensor<320x320x3x3xf16>) outs(%521 : tensor<2x320x32x32xf16>) -> tensor<2x320x32x32xf16> | |
| %collapsed_828 = tensor.collapse_shape %522 [[0], [1], [2, 3]] : tensor<2x320x32x32xf16> into tensor<2x320x1024xf16> | |
| %expanded_829 = tensor.expand_shape %collapsed_828 [[0], [1, 2], [3]] : tensor<2x320x1024xf16> into tensor<2x32x10x1024xf16> | |
| %523 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %524 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%523 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %525 = tensor.empty() : tensor<2x32x10x1024xf32> | |
| %526 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%524 : tensor<f32>) outs(%525 : tensor<2x32x10x1024xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x10x1024xf32> | |
| %527 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_829 : tensor<2x32x10x1024xf16>) outs(%526 : tensor<2x32x10x1024xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x10x1024xf32> | |
| %528 = tensor.empty() : tensor<2x32x10x1024xf64> | |
| %529 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%527 : tensor<2x32x10x1024xf32>) outs(%528 : tensor<2x32x10x1024xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x1024xf64> | |
| %530 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %531 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%529 : tensor<2x32x10x1024xf64>) outs(%530 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %532 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%531 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %533 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%529, %532 : tensor<2x32x10x1024xf64>, tensor<2x32x1x1xf64>) outs(%528 : tensor<2x32x10x1024xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x1024xf64> | |
| %534 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%533 : tensor<2x32x10x1024xf64>) outs(%528 : tensor<2x32x10x1024xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x10x1024xf64> | |
| %535 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %536 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%534 : tensor<2x32x10x1024xf64>) outs(%535 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %537 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%536 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %538 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%537 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %539 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %540 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%527 : tensor<2x32x10x1024xf32>) outs(%539 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %541 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%540 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_706 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %542 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%538, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %543 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%542 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %544 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_829, %541 : tensor<2x32x10x1024xf16>, tensor<2x32x1x1xf32>) outs(%525 : tensor<2x32x10x1024xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x10x1024xf32> | |
| %545 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%544, %543 : tensor<2x32x10x1024xf32>, tensor<2x32x1x1xf32>) outs(%525 : tensor<2x32x10x1024xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x10x1024xf32> | |
| %collapsed_830 = tensor.collapse_shape %545 [[0], [1, 2], [3]] : tensor<2x32x10x1024xf32> into tensor<2x320x1024xf32> | |
| %expanded_831 = tensor.expand_shape %collapsed_830 [[0], [1], [2, 3]] : tensor<2x320x1024xf32> into tensor<2x320x32x32xf32> | |
| %expanded_832 = tensor.expand_shape %cst_88 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16> | |
| %546 = tensor.empty() : tensor<2x320x32x32xf32> | |
| %547 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_831, %expanded_832 : tensor<2x320x32x32xf32>, tensor<320x1x1xf16>) outs(%546 : tensor<2x320x32x32xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x320x32x32xf32> | |
| %expanded_833 = tensor.expand_shape %cst_89 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16> | |
| %548 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%547, %expanded_833 : tensor<2x320x32x32xf32>, tensor<320x1x1xf16>) outs(%546 : tensor<2x320x32x32xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x320x32x32xf32> | |
| %549 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %550 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%549 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %551 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%550 : tensor<f16>) outs(%520 : tensor<2x320x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x320x32x32xf16> | |
| %552 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%548 : tensor<2x320x32x32xf32>) outs(%551 : tensor<2x320x32x32xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x32x32xf16> | |
| %553 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%552 : tensor<2x320x32x32xf16>) outs(%520 : tensor<2x320x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x320x32x32xf16> | |
| %554 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%553, %552 : tensor<2x320x32x32xf16>, tensor<2x320x32x32xf16>) outs(%520 : tensor<2x320x32x32xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x320x32x32xf16> | |
| %padded_834 = tensor.pad %554 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x320x32x32xf16> to tensor<2x320x34x34xf16> | |
| %555 = tensor.empty() : tensor<2x640x32x32xf16> | |
| %556 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_91 : tensor<640xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %557 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_834, %cst_90 : tensor<2x320x34x34xf16>, tensor<640x320x3x3xf16>) outs(%556 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16> | |
| %558 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280xf16> | |
| %559 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%558, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %560 = tensor.empty() : tensor<1280x640xf16> | |
| %561 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_92 : tensor<640x1280xf16>) outs(%560 : tensor<1280x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x640xf16> | |
| %562 = tensor.empty() : tensor<2x640xf16> | |
| %563 = linalg.fill ins(%cst_694 : f16) outs(%562 : tensor<2x640xf16>) -> tensor<2x640xf16> | |
| %564 = linalg.matmul ins(%559, %561 : tensor<2x1280xf16>, tensor<1280x640xf16>) outs(%563 : tensor<2x640xf16>) -> tensor<2x640xf16> | |
| %565 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_93, %564 : tensor<640xf16>, tensor<2x640xf16>) outs(%562 : tensor<2x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640xf16> | |
| %expanded_835 = tensor.expand_shape %565 [[0], [1, 2, 3]] : tensor<2x640xf16> into tensor<2x640x1x1xf16> | |
| %566 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%557, %expanded_835 : tensor<2x640x32x32xf16>, tensor<2x640x1x1xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %collapsed_836 = tensor.collapse_shape %566 [[0], [1], [2, 3]] : tensor<2x640x32x32xf16> into tensor<2x640x1024xf16> | |
| %expanded_837 = tensor.expand_shape %collapsed_836 [[0], [1, 2], [3]] : tensor<2x640x1024xf16> into tensor<2x32x20x1024xf16> | |
| %567 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %568 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%567 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %569 = tensor.empty() : tensor<2x32x20x1024xf32> | |
| %570 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%568 : tensor<f32>) outs(%569 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %571 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_837 : tensor<2x32x20x1024xf16>) outs(%570 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %572 = tensor.empty() : tensor<2x32x20x1024xf64> | |
| %573 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%571 : tensor<2x32x20x1024xf32>) outs(%572 : tensor<2x32x20x1024xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x1024xf64> | |
| %574 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %575 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%573 : tensor<2x32x20x1024xf64>) outs(%574 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %576 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%575 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_707 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %577 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%573, %576 : tensor<2x32x20x1024xf64>, tensor<2x32x1x1xf64>) outs(%572 : tensor<2x32x20x1024xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x1024xf64> | |
| %578 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%577 : tensor<2x32x20x1024xf64>) outs(%572 : tensor<2x32x20x1024xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x1024xf64> | |
| %579 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %580 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%578 : tensor<2x32x20x1024xf64>) outs(%579 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %581 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%580 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_707 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %582 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%581 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %583 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %584 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%571 : tensor<2x32x20x1024xf32>) outs(%583 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %585 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%584 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_708 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %586 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%582, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %587 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%586 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %588 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_837, %585 : tensor<2x32x20x1024xf16>, tensor<2x32x1x1xf32>) outs(%569 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %589 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%588, %587 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%569 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %collapsed_838 = tensor.collapse_shape %589 [[0], [1, 2], [3]] : tensor<2x32x20x1024xf32> into tensor<2x640x1024xf32> | |
| %expanded_839 = tensor.expand_shape %collapsed_838 [[0], [1], [2, 3]] : tensor<2x640x1024xf32> into tensor<2x640x32x32xf32> | |
| %expanded_840 = tensor.expand_shape %cst_94 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16> | |
| %590 = tensor.empty() : tensor<2x640x32x32xf32> | |
| %591 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_839, %expanded_840 : tensor<2x640x32x32xf32>, tensor<640x1x1xf16>) outs(%590 : tensor<2x640x32x32xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x640x32x32xf32> | |
| %expanded_841 = tensor.expand_shape %cst_95 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16> | |
| %592 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%591, %expanded_841 : tensor<2x640x32x32xf32>, tensor<640x1x1xf16>) outs(%590 : tensor<2x640x32x32xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x640x32x32xf32> | |
| %593 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %594 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%593 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %595 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%594 : tensor<f16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %596 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%592 : tensor<2x640x32x32xf32>) outs(%595 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %597 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%596 : tensor<2x640x32x32xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %598 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%597, %596 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %padded_842 = tensor.pad %598 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x640x32x32xf16> to tensor<2x640x34x34xf16> | |
| %599 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_97 : tensor<640xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %600 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_842, %cst_96 : tensor<2x640x34x34xf16>, tensor<640x640x3x3xf16>) outs(%599 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16> | |
| %601 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_99 : tensor<640xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %602 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%522, %cst_98 : tensor<2x320x32x32xf16>, tensor<640x320x1x1xf16>) outs(%601 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16> | |
| %603 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%602, %600 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %604 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%603, %cst_3 : tensor<2x640x32x32xf16>, tensor<f64>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %collapsed_843 = tensor.collapse_shape %604 [[0], [1], [2, 3]] : tensor<2x640x32x32xf16> into tensor<2x640x1024xf16> | |
| %expanded_844 = tensor.expand_shape %collapsed_843 [[0], [1, 2], [3]] : tensor<2x640x1024xf16> into tensor<2x32x20x1024xf16> | |
| %605 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %606 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%605 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %607 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%606 : tensor<f32>) outs(%569 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %608 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_844 : tensor<2x32x20x1024xf16>) outs(%607 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %609 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%608 : tensor<2x32x20x1024xf32>) outs(%572 : tensor<2x32x20x1024xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x1024xf64> | |
| %610 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %611 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%609 : tensor<2x32x20x1024xf64>) outs(%610 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %612 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%611 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_707 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %613 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%609, %612 : tensor<2x32x20x1024xf64>, tensor<2x32x1x1xf64>) outs(%572 : tensor<2x32x20x1024xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x1024xf64> | |
| %614 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%613 : tensor<2x32x20x1024xf64>) outs(%572 : tensor<2x32x20x1024xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x1024xf64> | |
| %615 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %616 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%614 : tensor<2x32x20x1024xf64>) outs(%615 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %617 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%616 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_707 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %618 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%617 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %619 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %620 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%608 : tensor<2x32x20x1024xf32>) outs(%619 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %621 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%620 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_708 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %622 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%618, %cst_2 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %623 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%622 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %624 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_844, %621 : tensor<2x32x20x1024xf16>, tensor<2x32x1x1xf32>) outs(%569 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %625 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%624, %623 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%569 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %collapsed_845 = tensor.collapse_shape %625 [[0], [1, 2], [3]] : tensor<2x32x20x1024xf32> into tensor<2x640x1024xf32> | |
| %expanded_846 = tensor.expand_shape %collapsed_845 [[0], [1], [2, 3]] : tensor<2x640x1024xf32> into tensor<2x640x32x32xf32> | |
| %expanded_847 = tensor.expand_shape %cst_100 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16> | |
| %626 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_846, %expanded_847 : tensor<2x640x32x32xf32>, tensor<640x1x1xf16>) outs(%590 : tensor<2x640x32x32xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x640x32x32xf32> | |
| %expanded_848 = tensor.expand_shape %cst_101 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16> | |
| %627 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%626, %expanded_848 : tensor<2x640x32x32xf32>, tensor<640x1x1xf16>) outs(%590 : tensor<2x640x32x32xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x640x32x32xf32> | |
| %628 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %629 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%628 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %630 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%629 : tensor<f16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %631 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%627 : tensor<2x640x32x32xf32>) outs(%630 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %632 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_103 : tensor<640xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %633 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%631, %cst_102 : tensor<2x640x32x32xf16>, tensor<640x640x1x1xf16>) outs(%632 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16> | |
| %634 = tensor.empty() : tensor<2x32x32x640xf16> | |
| %635 = linalg.generic {indexing_maps = [#map11, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%633 : tensor<2x640x32x32xf16>) outs(%634 : tensor<2x32x32x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x32x32x640xf16> | |
| %collapsed_849 = tensor.collapse_shape %635 [[0], [1, 2], [3]] : tensor<2x32x32x640xf16> into tensor<2x1024x640xf16> | |
| %636 = tensor.empty() : tensor<2x1024x1xf16> | |
| %637 = linalg.fill ins(%cst_694 : f16) outs(%636 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16> | |
| %638 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_849 : tensor<2x1024x640xf16>) outs(%637 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %639 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%638 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_709 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %640 = tensor.empty() : tensor<2x1024x640xf16> | |
| %641 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%639 : tensor<2x1024x1xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %642 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_849, %641 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %643 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%642 : tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %644 = linalg.fill ins(%cst_694 : f16) outs(%636 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16> | |
| %645 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%643 : tensor<2x1024x640xf16>) outs(%644 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %646 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%645 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_709 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %647 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%646 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %648 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%647 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %649 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%648 : tensor<2x1024x1xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %650 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%642, %649 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %651 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%650, %cst_104 : tensor<2x1024x640xf16>, tensor<640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %652 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%651, %cst_105 : tensor<2x1024x640xf16>, tensor<640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %653 = tensor.empty() : tensor<640x640xf16> | |
| %654 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_106 : tensor<640x640xf16>) outs(%653 : tensor<640x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<640x640xf16> | |
| %collapsed_850 = tensor.collapse_shape %652 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16> | |
| %655 = tensor.empty() : tensor<2048x640xf16> | |
| %656 = linalg.fill ins(%cst_694 : f16) outs(%655 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %657 = linalg.matmul ins(%collapsed_850, %654 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%656 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %658 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_107 : tensor<640x640xf16>) outs(%653 : tensor<640x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<640x640xf16> | |
| %659 = linalg.fill ins(%cst_694 : f16) outs(%655 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %660 = linalg.matmul ins(%collapsed_850, %658 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%659 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %661 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_108 : tensor<640x640xf16>) outs(%653 : tensor<640x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<640x640xf16> | |
| %662 = linalg.fill ins(%cst_694 : f16) outs(%655 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %663 = linalg.matmul ins(%collapsed_850, %661 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%662 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %expanded_851 = tensor.expand_shape %657 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x8x80xf16> | |
| %664 = tensor.empty() : tensor<2x8x1024x80xf16> | |
| %665 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_851 : tensor<2x1024x8x80xf16>) outs(%664 : tensor<2x8x1024x80xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x1024x80xf16> | |
| %collapsed_852 = tensor.collapse_shape %665 [[0, 1], [2], [3]] : tensor<2x8x1024x80xf16> into tensor<16x1024x80xf16> | |
| %expanded_853 = tensor.expand_shape %660 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x8x80xf16> | |
| %666 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_853 : tensor<2x1024x8x80xf16>) outs(%664 : tensor<2x8x1024x80xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x1024x80xf16> | |
| %collapsed_854 = tensor.collapse_shape %666 [[0, 1], [2], [3]] : tensor<2x8x1024x80xf16> into tensor<16x1024x80xf16> | |
| %expanded_855 = tensor.expand_shape %663 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x8x80xf16> | |
| %667 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_855 : tensor<2x1024x8x80xf16>) outs(%664 : tensor<2x8x1024x80xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x1024x80xf16> | |
| %collapsed_856 = tensor.collapse_shape %667 [[0, 1], [2], [3]] : tensor<2x8x1024x80xf16> into tensor<16x1024x80xf16> | |
| %668 = tensor.empty() : tensor<16x80x1024xf16> | |
| %669 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_854 : tensor<16x1024x80xf16>) outs(%668 : tensor<16x80x1024xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x80x1024xf16> | |
| %670 = tensor.empty() : tensor<16x1024x1024xf16> | |
| %671 = linalg.fill ins(%cst_694 : f16) outs(%670 : tensor<16x1024x1024xf16>) -> tensor<16x1024x1024xf16> | |
| %672 = linalg.batch_matmul ins(%collapsed_852, %669 : tensor<16x1024x80xf16>, tensor<16x80x1024xf16>) outs(%671 : tensor<16x1024x1024xf16>) -> tensor<16x1024x1024xf16> | |
| %673 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%672, %cst_0 : tensor<16x1024x1024xf16>, tensor<f64>) outs(%670 : tensor<16x1024x1024xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x1024x1024xf16> | |
| %674 = tensor.empty() : tensor<16x1024x1xi64> | |
| %675 = linalg.fill ins(%c0_i64 : i64) outs(%674 : tensor<16x1024x1xi64>) -> tensor<16x1024x1xi64> | |
| %676 = tensor.empty() : tensor<16x1024x1xf16> | |
| %677 = linalg.fill ins(%cst_696 : f16) outs(%676 : tensor<16x1024x1xf16>) -> tensor<16x1024x1xf16> | |
| %678:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%673 : tensor<16x1024x1024xf16>) outs(%677, %675 : tensor<16x1024x1xf16>, tensor<16x1024x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x1024x1xf16>, tensor<16x1024x1xi64>) | |
| %679 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%673, %678#0 : tensor<16x1024x1024xf16>, tensor<16x1024x1xf16>) outs(%670 : tensor<16x1024x1024xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x1024x1024xf16> | |
| %680 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%679 : tensor<16x1024x1024xf16>) outs(%670 : tensor<16x1024x1024xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x1024x1024xf16> | |
| %681 = linalg.fill ins(%cst_694 : f16) outs(%676 : tensor<16x1024x1xf16>) -> tensor<16x1024x1xf16> | |
| %682 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%680 : tensor<16x1024x1024xf16>) outs(%681 : tensor<16x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x1024x1xf16> | |
| %683 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%680, %682 : tensor<16x1024x1024xf16>, tensor<16x1024x1xf16>) outs(%670 : tensor<16x1024x1024xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x1024x1024xf16> | |
| %684 = tensor.empty() : tensor<16x1024x80xf16> | |
| %685 = linalg.fill ins(%cst_694 : f16) outs(%684 : tensor<16x1024x80xf16>) -> tensor<16x1024x80xf16> | |
| %686 = linalg.batch_matmul ins(%683, %collapsed_856 : tensor<16x1024x1024xf16>, tensor<16x1024x80xf16>) outs(%685 : tensor<16x1024x80xf16>) -> tensor<16x1024x80xf16> | |
| %expanded_857 = tensor.expand_shape %686 [[0, 1], [2], [3]] : tensor<16x1024x80xf16> into tensor<2x8x1024x80xf16> | |
| %687 = tensor.empty() : tensor<2x1024x8x80xf16> | |
| %688 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_857 : tensor<2x8x1024x80xf16>) outs(%687 : tensor<2x1024x8x80xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1024x8x80xf16> | |
| %689 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_109 : tensor<640x640xf16>) outs(%653 : tensor<640x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<640x640xf16> | |
| %collapsed_858 = tensor.collapse_shape %688 [[0, 1], [2, 3]] : tensor<2x1024x8x80xf16> into tensor<2048x640xf16> | |
| %690 = linalg.fill ins(%cst_694 : f16) outs(%655 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %691 = linalg.matmul ins(%collapsed_858, %689 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%690 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %692 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_110, %691 : tensor<640xf16>, tensor<2048x640xf16>) outs(%655 : tensor<2048x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2048x640xf16> | |
| %expanded_859 = tensor.expand_shape %692 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16> | |
| %693 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_859, %collapsed_849 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %694 = linalg.fill ins(%cst_694 : f16) outs(%636 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16> | |
| %695 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%693 : tensor<2x1024x640xf16>) outs(%694 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %696 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%695 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_709 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %697 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%696 : tensor<2x1024x1xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %698 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%693, %697 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %699 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%698 : tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %700 = linalg.fill ins(%cst_694 : f16) outs(%636 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16> | |
| %701 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%699 : tensor<2x1024x640xf16>) outs(%700 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %702 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%701 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_709 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %703 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%702 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %704 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%703 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %705 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%704 : tensor<2x1024x1xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %706 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%698, %705 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %707 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%706, %cst_111 : tensor<2x1024x640xf16>, tensor<640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %708 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%707, %cst_112 : tensor<2x1024x640xf16>, tensor<640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %709 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_113 : tensor<640x640xf16>) outs(%653 : tensor<640x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<640x640xf16> | |
| %collapsed_860 = tensor.collapse_shape %708 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16> | |
| %710 = linalg.fill ins(%cst_694 : f16) outs(%655 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %711 = linalg.matmul ins(%collapsed_860, %709 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%710 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %712 = tensor.empty() : tensor<768x640xf16> | |
| %713 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_114 : tensor<640x768xf16>) outs(%712 : tensor<768x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x640xf16> | |
| %714 = tensor.empty() : tensor<154x640xf16> | |
| %715 = linalg.fill ins(%cst_694 : f16) outs(%714 : tensor<154x640xf16>) -> tensor<154x640xf16> | |
| %716 = linalg.matmul ins(%collapsed_761, %713 : tensor<154x768xf16>, tensor<768x640xf16>) outs(%715 : tensor<154x640xf16>) -> tensor<154x640xf16> | |
| %717 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_115 : tensor<640x768xf16>) outs(%712 : tensor<768x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x640xf16> | |
| %718 = linalg.fill ins(%cst_694 : f16) outs(%714 : tensor<154x640xf16>) -> tensor<154x640xf16> | |
| %719 = linalg.matmul ins(%collapsed_761, %717 : tensor<154x768xf16>, tensor<768x640xf16>) outs(%718 : tensor<154x640xf16>) -> tensor<154x640xf16> | |
| %expanded_861 = tensor.expand_shape %711 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x8x80xf16> | |
| %720 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_861 : tensor<2x1024x8x80xf16>) outs(%664 : tensor<2x8x1024x80xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x1024x80xf16> | |
| %collapsed_862 = tensor.collapse_shape %720 [[0, 1], [2], [3]] : tensor<2x8x1024x80xf16> into tensor<16x1024x80xf16> | |
| %expanded_863 = tensor.expand_shape %716 [[0, 1], [2, 3]] : tensor<154x640xf16> into tensor<2x77x8x80xf16> | |
| %721 = tensor.empty() : tensor<2x8x77x80xf16> | |
| %722 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_863 : tensor<2x77x8x80xf16>) outs(%721 : tensor<2x8x77x80xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x80xf16> | |
| %collapsed_864 = tensor.collapse_shape %722 [[0, 1], [2], [3]] : tensor<2x8x77x80xf16> into tensor<16x77x80xf16> | |
| %expanded_865 = tensor.expand_shape %719 [[0, 1], [2, 3]] : tensor<154x640xf16> into tensor<2x77x8x80xf16> | |
| %723 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_865 : tensor<2x77x8x80xf16>) outs(%721 : tensor<2x8x77x80xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x80xf16> | |
| %collapsed_866 = tensor.collapse_shape %723 [[0, 1], [2], [3]] : tensor<2x8x77x80xf16> into tensor<16x77x80xf16> | |
| %724 = tensor.empty() : tensor<16x80x77xf16> | |
| %725 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_864 : tensor<16x77x80xf16>) outs(%724 : tensor<16x80x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x80x77xf16> | |
| %726 = tensor.empty() : tensor<16x1024x77xf16> | |
| %727 = linalg.fill ins(%cst_694 : f16) outs(%726 : tensor<16x1024x77xf16>) -> tensor<16x1024x77xf16> | |
| %728 = linalg.batch_matmul ins(%collapsed_862, %725 : tensor<16x1024x80xf16>, tensor<16x80x77xf16>) outs(%727 : tensor<16x1024x77xf16>) -> tensor<16x1024x77xf16> | |
| %729 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%728, %cst_0 : tensor<16x1024x77xf16>, tensor<f64>) outs(%726 : tensor<16x1024x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x1024x77xf16> | |
| %730 = linalg.fill ins(%c0_i64 : i64) outs(%674 : tensor<16x1024x1xi64>) -> tensor<16x1024x1xi64> | |
| %731 = linalg.fill ins(%cst_696 : f16) outs(%676 : tensor<16x1024x1xf16>) -> tensor<16x1024x1xf16> | |
| %732:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%729 : tensor<16x1024x77xf16>) outs(%731, %730 : tensor<16x1024x1xf16>, tensor<16x1024x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x1024x1xf16>, tensor<16x1024x1xi64>) | |
| %733 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%729, %732#0 : tensor<16x1024x77xf16>, tensor<16x1024x1xf16>) outs(%726 : tensor<16x1024x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x1024x77xf16> | |
| %734 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%733 : tensor<16x1024x77xf16>) outs(%726 : tensor<16x1024x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x1024x77xf16> | |
| %735 = linalg.fill ins(%cst_694 : f16) outs(%676 : tensor<16x1024x1xf16>) -> tensor<16x1024x1xf16> | |
| %736 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%734 : tensor<16x1024x77xf16>) outs(%735 : tensor<16x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x1024x1xf16> | |
| %737 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%734, %736 : tensor<16x1024x77xf16>, tensor<16x1024x1xf16>) outs(%726 : tensor<16x1024x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x1024x77xf16> | |
| %738 = linalg.fill ins(%cst_694 : f16) outs(%684 : tensor<16x1024x80xf16>) -> tensor<16x1024x80xf16> | |
| %739 = linalg.batch_matmul ins(%737, %collapsed_866 : tensor<16x1024x77xf16>, tensor<16x77x80xf16>) outs(%738 : tensor<16x1024x80xf16>) -> tensor<16x1024x80xf16> | |
| %expanded_867 = tensor.expand_shape %739 [[0, 1], [2], [3]] : tensor<16x1024x80xf16> into tensor<2x8x1024x80xf16> | |
| %740 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_867 : tensor<2x8x1024x80xf16>) outs(%687 : tensor<2x1024x8x80xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1024x8x80xf16> | |
| %741 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_116 : tensor<640x640xf16>) outs(%653 : tensor<640x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<640x640xf16> | |
| %collapsed_868 = tensor.collapse_shape %740 [[0, 1], [2, 3]] : tensor<2x1024x8x80xf16> into tensor<2048x640xf16> | |
| %742 = linalg.fill ins(%cst_694 : f16) outs(%655 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %743 = linalg.matmul ins(%collapsed_868, %741 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%742 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %744 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_117, %743 : tensor<640xf16>, tensor<2048x640xf16>) outs(%655 : tensor<2048x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2048x640xf16> | |
| %expanded_869 = tensor.expand_shape %744 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16> | |
| %745 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_869, %693 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %746 = linalg.fill ins(%cst_694 : f16) outs(%636 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16> | |
| %747 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%745 : tensor<2x1024x640xf16>) outs(%746 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %748 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%747 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_709 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %749 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%748 : tensor<2x1024x1xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %750 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%745, %749 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %751 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%750 : tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %752 = linalg.fill ins(%cst_694 : f16) outs(%636 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16> | |
| %753 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%751 : tensor<2x1024x640xf16>) outs(%752 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %754 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%753 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_709 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %755 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%754 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %756 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%755 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %757 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%756 : tensor<2x1024x1xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %758 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%750, %757 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %759 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%758, %cst_118 : tensor<2x1024x640xf16>, tensor<640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %760 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%759, %cst_119 : tensor<2x1024x640xf16>, tensor<640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %761 = tensor.empty() : tensor<640x5120xf16> | |
| %762 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_120 : tensor<5120x640xf16>) outs(%761 : tensor<640x5120xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<640x5120xf16> | |
| %collapsed_870 = tensor.collapse_shape %760 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16> | |
| %763 = tensor.empty() : tensor<2048x5120xf16> | |
| %764 = linalg.fill ins(%cst_694 : f16) outs(%763 : tensor<2048x5120xf16>) -> tensor<2048x5120xf16> | |
| %765 = linalg.matmul ins(%collapsed_870, %762 : tensor<2048x640xf16>, tensor<640x5120xf16>) outs(%764 : tensor<2048x5120xf16>) -> tensor<2048x5120xf16> | |
| %766 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_121, %765 : tensor<5120xf16>, tensor<2048x5120xf16>) outs(%763 : tensor<2048x5120xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2048x5120xf16> | |
| %expanded_871 = tensor.expand_shape %766 [[0, 1], [2]] : tensor<2048x5120xf16> into tensor<2x1024x5120xf16> | |
| %extracted_slice_872 = tensor.extract_slice %expanded_871[0, 0, 0] [2, 1024, 2560] [1, 1, 1] : tensor<2x1024x5120xf16> to tensor<2x1024x2560xf16> | |
| %extracted_slice_873 = tensor.extract_slice %expanded_871[0, 0, 2560] [2, 1024, 2560] [1, 1, 1] : tensor<2x1024x5120xf16> to tensor<2x1024x2560xf16> | |
| %767 = tensor.empty() : tensor<2x1024x2560xf16> | |
| %768 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_873 : tensor<2x1024x2560xf16>) outs(%767 : tensor<2x1024x2560xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.sqrt %cst_697 : f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| %4282 = math.erf %4281 : f16 | |
| %4283 = arith.addf %4282, %cst_695 : f16 | |
| %4284 = arith.mulf %4283, %cst_698 : f16 | |
| %4285 = arith.mulf %in, %4284 : f16 | |
| linalg.yield %4285 : f16 | |
| } -> tensor<2x1024x2560xf16> | |
| %769 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_872, %768 : tensor<2x1024x2560xf16>, tensor<2x1024x2560xf16>) outs(%767 : tensor<2x1024x2560xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x2560xf16> | |
| %770 = tensor.empty() : tensor<2560x640xf16> | |
| %771 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_122 : tensor<640x2560xf16>) outs(%770 : tensor<2560x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2560x640xf16> | |
| %collapsed_874 = tensor.collapse_shape %769 [[0, 1], [2]] : tensor<2x1024x2560xf16> into tensor<2048x2560xf16> | |
| %772 = linalg.fill ins(%cst_694 : f16) outs(%655 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %773 = linalg.matmul ins(%collapsed_874, %771 : tensor<2048x2560xf16>, tensor<2560x640xf16>) outs(%772 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %774 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_123, %773 : tensor<640xf16>, tensor<2048x640xf16>) outs(%655 : tensor<2048x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2048x640xf16> | |
| %expanded_875 = tensor.expand_shape %774 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16> | |
| %775 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_875, %745 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %expanded_876 = tensor.expand_shape %775 [[0], [1, 2], [3]] : tensor<2x1024x640xf16> into tensor<2x32x32x640xf16> | |
| %776 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_876 : tensor<2x32x32x640xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %777 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_125 : tensor<640xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %778 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%776, %cst_124 : tensor<2x640x32x32xf16>, tensor<640x640x1x1xf16>) outs(%777 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16> | |
| %779 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%778, %604 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %collapsed_877 = tensor.collapse_shape %779 [[0], [1], [2, 3]] : tensor<2x640x32x32xf16> into tensor<2x640x1024xf16> | |
| %expanded_878 = tensor.expand_shape %collapsed_877 [[0], [1, 2], [3]] : tensor<2x640x1024xf16> into tensor<2x32x20x1024xf16> | |
| %780 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %781 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%780 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %782 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%781 : tensor<f32>) outs(%569 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %783 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_878 : tensor<2x32x20x1024xf16>) outs(%782 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %784 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%783 : tensor<2x32x20x1024xf32>) outs(%572 : tensor<2x32x20x1024xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x1024xf64> | |
| %785 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %786 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%784 : tensor<2x32x20x1024xf64>) outs(%785 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %787 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%786 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_707 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %788 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%784, %787 : tensor<2x32x20x1024xf64>, tensor<2x32x1x1xf64>) outs(%572 : tensor<2x32x20x1024xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x1024xf64> | |
| %789 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%788 : tensor<2x32x20x1024xf64>) outs(%572 : tensor<2x32x20x1024xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x1024xf64> | |
| %790 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %791 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%789 : tensor<2x32x20x1024xf64>) outs(%790 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %792 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%791 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_707 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %793 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%792 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %794 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %795 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%783 : tensor<2x32x20x1024xf32>) outs(%794 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %796 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%795 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_708 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %797 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%793, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %798 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%797 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %799 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_878, %796 : tensor<2x32x20x1024xf16>, tensor<2x32x1x1xf32>) outs(%569 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %800 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%799, %798 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%569 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %collapsed_879 = tensor.collapse_shape %800 [[0], [1, 2], [3]] : tensor<2x32x20x1024xf32> into tensor<2x640x1024xf32> | |
| %expanded_880 = tensor.expand_shape %collapsed_879 [[0], [1], [2, 3]] : tensor<2x640x1024xf32> into tensor<2x640x32x32xf32> | |
| %expanded_881 = tensor.expand_shape %cst_126 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16> | |
| %801 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_880, %expanded_881 : tensor<2x640x32x32xf32>, tensor<640x1x1xf16>) outs(%590 : tensor<2x640x32x32xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x640x32x32xf32> | |
| %expanded_882 = tensor.expand_shape %cst_127 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16> | |
| %802 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%801, %expanded_882 : tensor<2x640x32x32xf32>, tensor<640x1x1xf16>) outs(%590 : tensor<2x640x32x32xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x640x32x32xf32> | |
| %803 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %804 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%803 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %805 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%804 : tensor<f16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %806 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%802 : tensor<2x640x32x32xf32>) outs(%805 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %807 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%806 : tensor<2x640x32x32xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %808 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%807, %806 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %padded_883 = tensor.pad %808 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x640x32x32xf16> to tensor<2x640x34x34xf16> | |
| %809 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_129 : tensor<640xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %810 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_883, %cst_128 : tensor<2x640x34x34xf16>, tensor<640x640x3x3xf16>) outs(%809 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16> | |
| %811 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280xf16> | |
| %812 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%811, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %813 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_130 : tensor<640x1280xf16>) outs(%560 : tensor<1280x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x640xf16> | |
| %814 = linalg.fill ins(%cst_694 : f16) outs(%562 : tensor<2x640xf16>) -> tensor<2x640xf16> | |
| %815 = linalg.matmul ins(%812, %813 : tensor<2x1280xf16>, tensor<1280x640xf16>) outs(%814 : tensor<2x640xf16>) -> tensor<2x640xf16> | |
| %816 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_131, %815 : tensor<640xf16>, tensor<2x640xf16>) outs(%562 : tensor<2x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640xf16> | |
| %expanded_884 = tensor.expand_shape %816 [[0], [1, 2, 3]] : tensor<2x640xf16> into tensor<2x640x1x1xf16> | |
| %817 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%810, %expanded_884 : tensor<2x640x32x32xf16>, tensor<2x640x1x1xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %collapsed_885 = tensor.collapse_shape %817 [[0], [1], [2, 3]] : tensor<2x640x32x32xf16> into tensor<2x640x1024xf16> | |
| %expanded_886 = tensor.expand_shape %collapsed_885 [[0], [1, 2], [3]] : tensor<2x640x1024xf16> into tensor<2x32x20x1024xf16> | |
| %818 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %819 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%818 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %820 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%819 : tensor<f32>) outs(%569 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %821 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_886 : tensor<2x32x20x1024xf16>) outs(%820 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %822 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%821 : tensor<2x32x20x1024xf32>) outs(%572 : tensor<2x32x20x1024xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x1024xf64> | |
| %823 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %824 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%822 : tensor<2x32x20x1024xf64>) outs(%823 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %825 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%824 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_707 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %826 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%822, %825 : tensor<2x32x20x1024xf64>, tensor<2x32x1x1xf64>) outs(%572 : tensor<2x32x20x1024xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x1024xf64> | |
| %827 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%826 : tensor<2x32x20x1024xf64>) outs(%572 : tensor<2x32x20x1024xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x1024xf64> | |
| %828 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %829 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%827 : tensor<2x32x20x1024xf64>) outs(%828 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %830 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%829 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_707 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %831 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%830 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %832 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %833 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%821 : tensor<2x32x20x1024xf32>) outs(%832 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %834 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%833 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_708 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %835 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%831, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %836 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%835 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %837 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_886, %834 : tensor<2x32x20x1024xf16>, tensor<2x32x1x1xf32>) outs(%569 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %838 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%837, %836 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%569 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %collapsed_887 = tensor.collapse_shape %838 [[0], [1, 2], [3]] : tensor<2x32x20x1024xf32> into tensor<2x640x1024xf32> | |
| %expanded_888 = tensor.expand_shape %collapsed_887 [[0], [1], [2, 3]] : tensor<2x640x1024xf32> into tensor<2x640x32x32xf32> | |
| %expanded_889 = tensor.expand_shape %cst_132 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16> | |
| %839 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_888, %expanded_889 : tensor<2x640x32x32xf32>, tensor<640x1x1xf16>) outs(%590 : tensor<2x640x32x32xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x640x32x32xf32> | |
| %expanded_890 = tensor.expand_shape %cst_133 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16> | |
| %840 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%839, %expanded_890 : tensor<2x640x32x32xf32>, tensor<640x1x1xf16>) outs(%590 : tensor<2x640x32x32xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x640x32x32xf32> | |
| %841 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %842 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%841 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %843 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%842 : tensor<f16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %844 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%840 : tensor<2x640x32x32xf32>) outs(%843 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %845 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%844 : tensor<2x640x32x32xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %846 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%845, %844 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %padded_891 = tensor.pad %846 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x640x32x32xf16> to tensor<2x640x34x34xf16> | |
| %847 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_135 : tensor<640xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %848 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_891, %cst_134 : tensor<2x640x34x34xf16>, tensor<640x640x3x3xf16>) outs(%847 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16> | |
| %849 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%779, %848 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %850 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%849, %cst_3 : tensor<2x640x32x32xf16>, tensor<f64>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %collapsed_892 = tensor.collapse_shape %850 [[0], [1], [2, 3]] : tensor<2x640x32x32xf16> into tensor<2x640x1024xf16> | |
| %expanded_893 = tensor.expand_shape %collapsed_892 [[0], [1, 2], [3]] : tensor<2x640x1024xf16> into tensor<2x32x20x1024xf16> | |
| %851 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %852 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%851 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %853 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%852 : tensor<f32>) outs(%569 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %854 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_893 : tensor<2x32x20x1024xf16>) outs(%853 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %855 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%854 : tensor<2x32x20x1024xf32>) outs(%572 : tensor<2x32x20x1024xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x1024xf64> | |
| %856 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %857 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%855 : tensor<2x32x20x1024xf64>) outs(%856 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %858 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%857 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_707 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %859 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%855, %858 : tensor<2x32x20x1024xf64>, tensor<2x32x1x1xf64>) outs(%572 : tensor<2x32x20x1024xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x1024xf64> | |
| %860 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%859 : tensor<2x32x20x1024xf64>) outs(%572 : tensor<2x32x20x1024xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x1024xf64> | |
| %861 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %862 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%860 : tensor<2x32x20x1024xf64>) outs(%861 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %863 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%862 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_707 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %864 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%863 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %865 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %866 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%854 : tensor<2x32x20x1024xf32>) outs(%865 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %867 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%866 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_708 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %868 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%864, %cst_2 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %869 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%868 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %870 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_893, %867 : tensor<2x32x20x1024xf16>, tensor<2x32x1x1xf32>) outs(%569 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %871 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%870, %869 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%569 : tensor<2x32x20x1024xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x20x1024xf32> | |
| %collapsed_894 = tensor.collapse_shape %871 [[0], [1, 2], [3]] : tensor<2x32x20x1024xf32> into tensor<2x640x1024xf32> | |
| %expanded_895 = tensor.expand_shape %collapsed_894 [[0], [1], [2, 3]] : tensor<2x640x1024xf32> into tensor<2x640x32x32xf32> | |
| %expanded_896 = tensor.expand_shape %cst_136 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16> | |
| %872 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_895, %expanded_896 : tensor<2x640x32x32xf32>, tensor<640x1x1xf16>) outs(%590 : tensor<2x640x32x32xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x640x32x32xf32> | |
| %expanded_897 = tensor.expand_shape %cst_137 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16> | |
| %873 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%872, %expanded_897 : tensor<2x640x32x32xf32>, tensor<640x1x1xf16>) outs(%590 : tensor<2x640x32x32xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x640x32x32xf32> | |
| %874 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %875 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%874 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %876 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%875 : tensor<f16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %877 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%873 : tensor<2x640x32x32xf32>) outs(%876 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %878 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_139 : tensor<640xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %879 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%877, %cst_138 : tensor<2x640x32x32xf16>, tensor<640x640x1x1xf16>) outs(%878 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16> | |
| %880 = linalg.generic {indexing_maps = [#map11, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%879 : tensor<2x640x32x32xf16>) outs(%634 : tensor<2x32x32x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x32x32x640xf16> | |
| %collapsed_898 = tensor.collapse_shape %880 [[0], [1, 2], [3]] : tensor<2x32x32x640xf16> into tensor<2x1024x640xf16> | |
| %881 = linalg.fill ins(%cst_694 : f16) outs(%636 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16> | |
| %882 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_898 : tensor<2x1024x640xf16>) outs(%881 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %883 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%882 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_709 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %884 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%883 : tensor<2x1024x1xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %885 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_898, %884 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %886 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%885 : tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %887 = linalg.fill ins(%cst_694 : f16) outs(%636 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16> | |
| %888 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%886 : tensor<2x1024x640xf16>) outs(%887 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %889 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%888 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_709 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %890 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%889 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %891 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%890 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %892 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%891 : tensor<2x1024x1xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %893 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%885, %892 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %894 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%893, %cst_140 : tensor<2x1024x640xf16>, tensor<640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %895 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%894, %cst_141 : tensor<2x1024x640xf16>, tensor<640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %896 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_142 : tensor<640x640xf16>) outs(%653 : tensor<640x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<640x640xf16> | |
| %collapsed_899 = tensor.collapse_shape %895 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16> | |
| %897 = linalg.fill ins(%cst_694 : f16) outs(%655 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %898 = linalg.matmul ins(%collapsed_899, %896 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%897 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %899 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_143 : tensor<640x640xf16>) outs(%653 : tensor<640x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<640x640xf16> | |
| %900 = linalg.fill ins(%cst_694 : f16) outs(%655 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %901 = linalg.matmul ins(%collapsed_899, %899 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%900 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %902 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_144 : tensor<640x640xf16>) outs(%653 : tensor<640x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<640x640xf16> | |
| %903 = linalg.fill ins(%cst_694 : f16) outs(%655 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %904 = linalg.matmul ins(%collapsed_899, %902 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%903 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %expanded_900 = tensor.expand_shape %898 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x8x80xf16> | |
| %905 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_900 : tensor<2x1024x8x80xf16>) outs(%664 : tensor<2x8x1024x80xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x1024x80xf16> | |
| %collapsed_901 = tensor.collapse_shape %905 [[0, 1], [2], [3]] : tensor<2x8x1024x80xf16> into tensor<16x1024x80xf16> | |
| %expanded_902 = tensor.expand_shape %901 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x8x80xf16> | |
| %906 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_902 : tensor<2x1024x8x80xf16>) outs(%664 : tensor<2x8x1024x80xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x1024x80xf16> | |
| %collapsed_903 = tensor.collapse_shape %906 [[0, 1], [2], [3]] : tensor<2x8x1024x80xf16> into tensor<16x1024x80xf16> | |
| %expanded_904 = tensor.expand_shape %904 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x8x80xf16> | |
| %907 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_904 : tensor<2x1024x8x80xf16>) outs(%664 : tensor<2x8x1024x80xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x1024x80xf16> | |
| %collapsed_905 = tensor.collapse_shape %907 [[0, 1], [2], [3]] : tensor<2x8x1024x80xf16> into tensor<16x1024x80xf16> | |
| %908 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_903 : tensor<16x1024x80xf16>) outs(%668 : tensor<16x80x1024xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x80x1024xf16> | |
| %909 = linalg.fill ins(%cst_694 : f16) outs(%670 : tensor<16x1024x1024xf16>) -> tensor<16x1024x1024xf16> | |
| %910 = linalg.batch_matmul ins(%collapsed_901, %908 : tensor<16x1024x80xf16>, tensor<16x80x1024xf16>) outs(%909 : tensor<16x1024x1024xf16>) -> tensor<16x1024x1024xf16> | |
| %911 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%910, %cst_0 : tensor<16x1024x1024xf16>, tensor<f64>) outs(%670 : tensor<16x1024x1024xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x1024x1024xf16> | |
| %912 = linalg.fill ins(%c0_i64 : i64) outs(%674 : tensor<16x1024x1xi64>) -> tensor<16x1024x1xi64> | |
| %913 = linalg.fill ins(%cst_696 : f16) outs(%676 : tensor<16x1024x1xf16>) -> tensor<16x1024x1xf16> | |
| %914:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%911 : tensor<16x1024x1024xf16>) outs(%913, %912 : tensor<16x1024x1xf16>, tensor<16x1024x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x1024x1xf16>, tensor<16x1024x1xi64>) | |
| %915 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%911, %914#0 : tensor<16x1024x1024xf16>, tensor<16x1024x1xf16>) outs(%670 : tensor<16x1024x1024xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x1024x1024xf16> | |
| %916 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%915 : tensor<16x1024x1024xf16>) outs(%670 : tensor<16x1024x1024xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x1024x1024xf16> | |
| %917 = linalg.fill ins(%cst_694 : f16) outs(%676 : tensor<16x1024x1xf16>) -> tensor<16x1024x1xf16> | |
| %918 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%916 : tensor<16x1024x1024xf16>) outs(%917 : tensor<16x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x1024x1xf16> | |
| %919 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%916, %918 : tensor<16x1024x1024xf16>, tensor<16x1024x1xf16>) outs(%670 : tensor<16x1024x1024xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x1024x1024xf16> | |
| %920 = linalg.fill ins(%cst_694 : f16) outs(%684 : tensor<16x1024x80xf16>) -> tensor<16x1024x80xf16> | |
| %921 = linalg.batch_matmul ins(%919, %collapsed_905 : tensor<16x1024x1024xf16>, tensor<16x1024x80xf16>) outs(%920 : tensor<16x1024x80xf16>) -> tensor<16x1024x80xf16> | |
| %expanded_906 = tensor.expand_shape %921 [[0, 1], [2], [3]] : tensor<16x1024x80xf16> into tensor<2x8x1024x80xf16> | |
| %922 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_906 : tensor<2x8x1024x80xf16>) outs(%687 : tensor<2x1024x8x80xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1024x8x80xf16> | |
| %923 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_145 : tensor<640x640xf16>) outs(%653 : tensor<640x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<640x640xf16> | |
| %collapsed_907 = tensor.collapse_shape %922 [[0, 1], [2, 3]] : tensor<2x1024x8x80xf16> into tensor<2048x640xf16> | |
| %924 = linalg.fill ins(%cst_694 : f16) outs(%655 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %925 = linalg.matmul ins(%collapsed_907, %923 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%924 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %926 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_146, %925 : tensor<640xf16>, tensor<2048x640xf16>) outs(%655 : tensor<2048x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2048x640xf16> | |
| %expanded_908 = tensor.expand_shape %926 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16> | |
| %927 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_908, %collapsed_898 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %928 = linalg.fill ins(%cst_694 : f16) outs(%636 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16> | |
| %929 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%927 : tensor<2x1024x640xf16>) outs(%928 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %930 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%929 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_709 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %931 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%930 : tensor<2x1024x1xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %932 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%927, %931 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %933 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%932 : tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %934 = linalg.fill ins(%cst_694 : f16) outs(%636 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16> | |
| %935 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%933 : tensor<2x1024x640xf16>) outs(%934 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %936 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%935 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_709 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %937 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%936 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %938 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%937 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %939 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%938 : tensor<2x1024x1xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %940 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%932, %939 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %941 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%940, %cst_147 : tensor<2x1024x640xf16>, tensor<640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %942 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%941, %cst_148 : tensor<2x1024x640xf16>, tensor<640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %943 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_149 : tensor<640x640xf16>) outs(%653 : tensor<640x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<640x640xf16> | |
| %collapsed_909 = tensor.collapse_shape %942 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16> | |
| %944 = linalg.fill ins(%cst_694 : f16) outs(%655 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %945 = linalg.matmul ins(%collapsed_909, %943 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%944 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %946 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_150 : tensor<640x768xf16>) outs(%712 : tensor<768x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x640xf16> | |
| %947 = linalg.fill ins(%cst_694 : f16) outs(%714 : tensor<154x640xf16>) -> tensor<154x640xf16> | |
| %948 = linalg.matmul ins(%collapsed_761, %946 : tensor<154x768xf16>, tensor<768x640xf16>) outs(%947 : tensor<154x640xf16>) -> tensor<154x640xf16> | |
| %949 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_151 : tensor<640x768xf16>) outs(%712 : tensor<768x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x640xf16> | |
| %950 = linalg.fill ins(%cst_694 : f16) outs(%714 : tensor<154x640xf16>) -> tensor<154x640xf16> | |
| %951 = linalg.matmul ins(%collapsed_761, %949 : tensor<154x768xf16>, tensor<768x640xf16>) outs(%950 : tensor<154x640xf16>) -> tensor<154x640xf16> | |
| %expanded_910 = tensor.expand_shape %945 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x8x80xf16> | |
| %952 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_910 : tensor<2x1024x8x80xf16>) outs(%664 : tensor<2x8x1024x80xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x1024x80xf16> | |
| %collapsed_911 = tensor.collapse_shape %952 [[0, 1], [2], [3]] : tensor<2x8x1024x80xf16> into tensor<16x1024x80xf16> | |
| %expanded_912 = tensor.expand_shape %948 [[0, 1], [2, 3]] : tensor<154x640xf16> into tensor<2x77x8x80xf16> | |
| %953 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_912 : tensor<2x77x8x80xf16>) outs(%721 : tensor<2x8x77x80xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x80xf16> | |
| %collapsed_913 = tensor.collapse_shape %953 [[0, 1], [2], [3]] : tensor<2x8x77x80xf16> into tensor<16x77x80xf16> | |
| %expanded_914 = tensor.expand_shape %951 [[0, 1], [2, 3]] : tensor<154x640xf16> into tensor<2x77x8x80xf16> | |
| %954 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_914 : tensor<2x77x8x80xf16>) outs(%721 : tensor<2x8x77x80xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x80xf16> | |
| %collapsed_915 = tensor.collapse_shape %954 [[0, 1], [2], [3]] : tensor<2x8x77x80xf16> into tensor<16x77x80xf16> | |
| %955 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_913 : tensor<16x77x80xf16>) outs(%724 : tensor<16x80x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x80x77xf16> | |
| %956 = linalg.fill ins(%cst_694 : f16) outs(%726 : tensor<16x1024x77xf16>) -> tensor<16x1024x77xf16> | |
| %957 = linalg.batch_matmul ins(%collapsed_911, %955 : tensor<16x1024x80xf16>, tensor<16x80x77xf16>) outs(%956 : tensor<16x1024x77xf16>) -> tensor<16x1024x77xf16> | |
| %958 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%957, %cst_0 : tensor<16x1024x77xf16>, tensor<f64>) outs(%726 : tensor<16x1024x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x1024x77xf16> | |
| %959 = linalg.fill ins(%c0_i64 : i64) outs(%674 : tensor<16x1024x1xi64>) -> tensor<16x1024x1xi64> | |
| %960 = linalg.fill ins(%cst_696 : f16) outs(%676 : tensor<16x1024x1xf16>) -> tensor<16x1024x1xf16> | |
| %961:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%958 : tensor<16x1024x77xf16>) outs(%960, %959 : tensor<16x1024x1xf16>, tensor<16x1024x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x1024x1xf16>, tensor<16x1024x1xi64>) | |
| %962 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%958, %961#0 : tensor<16x1024x77xf16>, tensor<16x1024x1xf16>) outs(%726 : tensor<16x1024x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x1024x77xf16> | |
| %963 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%962 : tensor<16x1024x77xf16>) outs(%726 : tensor<16x1024x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x1024x77xf16> | |
| %964 = linalg.fill ins(%cst_694 : f16) outs(%676 : tensor<16x1024x1xf16>) -> tensor<16x1024x1xf16> | |
| %965 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%963 : tensor<16x1024x77xf16>) outs(%964 : tensor<16x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x1024x1xf16> | |
| %966 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%963, %965 : tensor<16x1024x77xf16>, tensor<16x1024x1xf16>) outs(%726 : tensor<16x1024x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x1024x77xf16> | |
| %967 = linalg.fill ins(%cst_694 : f16) outs(%684 : tensor<16x1024x80xf16>) -> tensor<16x1024x80xf16> | |
| %968 = linalg.batch_matmul ins(%966, %collapsed_915 : tensor<16x1024x77xf16>, tensor<16x77x80xf16>) outs(%967 : tensor<16x1024x80xf16>) -> tensor<16x1024x80xf16> | |
| %expanded_916 = tensor.expand_shape %968 [[0, 1], [2], [3]] : tensor<16x1024x80xf16> into tensor<2x8x1024x80xf16> | |
| %969 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_916 : tensor<2x8x1024x80xf16>) outs(%687 : tensor<2x1024x8x80xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1024x8x80xf16> | |
| %970 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_152 : tensor<640x640xf16>) outs(%653 : tensor<640x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<640x640xf16> | |
| %collapsed_917 = tensor.collapse_shape %969 [[0, 1], [2, 3]] : tensor<2x1024x8x80xf16> into tensor<2048x640xf16> | |
| %971 = linalg.fill ins(%cst_694 : f16) outs(%655 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %972 = linalg.matmul ins(%collapsed_917, %970 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%971 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %973 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_153, %972 : tensor<640xf16>, tensor<2048x640xf16>) outs(%655 : tensor<2048x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2048x640xf16> | |
| %expanded_918 = tensor.expand_shape %973 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16> | |
| %974 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_918, %927 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %975 = linalg.fill ins(%cst_694 : f16) outs(%636 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16> | |
| %976 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%974 : tensor<2x1024x640xf16>) outs(%975 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %977 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%976 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_709 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %978 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%977 : tensor<2x1024x1xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %979 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%974, %978 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %980 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%979 : tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %981 = linalg.fill ins(%cst_694 : f16) outs(%636 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16> | |
| %982 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%980 : tensor<2x1024x640xf16>) outs(%981 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %983 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%982 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_709 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %984 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%983 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %985 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%984 : tensor<2x1024x1xf16>) outs(%636 : tensor<2x1024x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x1xf16> | |
| %986 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%985 : tensor<2x1024x1xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %987 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%979, %986 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %988 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%987, %cst_154 : tensor<2x1024x640xf16>, tensor<640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %989 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%988, %cst_155 : tensor<2x1024x640xf16>, tensor<640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %990 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_156 : tensor<5120x640xf16>) outs(%761 : tensor<640x5120xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<640x5120xf16> | |
| %collapsed_919 = tensor.collapse_shape %989 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16> | |
| %991 = linalg.fill ins(%cst_694 : f16) outs(%763 : tensor<2048x5120xf16>) -> tensor<2048x5120xf16> | |
| %992 = linalg.matmul ins(%collapsed_919, %990 : tensor<2048x640xf16>, tensor<640x5120xf16>) outs(%991 : tensor<2048x5120xf16>) -> tensor<2048x5120xf16> | |
| %993 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_157, %992 : tensor<5120xf16>, tensor<2048x5120xf16>) outs(%763 : tensor<2048x5120xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2048x5120xf16> | |
| %expanded_920 = tensor.expand_shape %993 [[0, 1], [2]] : tensor<2048x5120xf16> into tensor<2x1024x5120xf16> | |
| %extracted_slice_921 = tensor.extract_slice %expanded_920[0, 0, 0] [2, 1024, 2560] [1, 1, 1] : tensor<2x1024x5120xf16> to tensor<2x1024x2560xf16> | |
| %extracted_slice_922 = tensor.extract_slice %expanded_920[0, 0, 2560] [2, 1024, 2560] [1, 1, 1] : tensor<2x1024x5120xf16> to tensor<2x1024x2560xf16> | |
| %994 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_922 : tensor<2x1024x2560xf16>) outs(%767 : tensor<2x1024x2560xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.sqrt %cst_697 : f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| %4282 = math.erf %4281 : f16 | |
| %4283 = arith.addf %4282, %cst_695 : f16 | |
| %4284 = arith.mulf %4283, %cst_698 : f16 | |
| %4285 = arith.mulf %in, %4284 : f16 | |
| linalg.yield %4285 : f16 | |
| } -> tensor<2x1024x2560xf16> | |
| %995 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_921, %994 : tensor<2x1024x2560xf16>, tensor<2x1024x2560xf16>) outs(%767 : tensor<2x1024x2560xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x2560xf16> | |
| %996 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_158 : tensor<640x2560xf16>) outs(%770 : tensor<2560x640xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2560x640xf16> | |
| %collapsed_923 = tensor.collapse_shape %995 [[0, 1], [2]] : tensor<2x1024x2560xf16> into tensor<2048x2560xf16> | |
| %997 = linalg.fill ins(%cst_694 : f16) outs(%655 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %998 = linalg.matmul ins(%collapsed_923, %996 : tensor<2048x2560xf16>, tensor<2560x640xf16>) outs(%997 : tensor<2048x640xf16>) -> tensor<2048x640xf16> | |
| %999 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_159, %998 : tensor<640xf16>, tensor<2048x640xf16>) outs(%655 : tensor<2048x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2048x640xf16> | |
| %expanded_924 = tensor.expand_shape %999 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16> | |
| %1000 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_924, %974 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%640 : tensor<2x1024x640xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1024x640xf16> | |
| %expanded_925 = tensor.expand_shape %1000 [[0], [1, 2], [3]] : tensor<2x1024x640xf16> into tensor<2x32x32x640xf16> | |
| %1001 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_925 : tensor<2x32x32x640xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %1002 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_161 : tensor<640xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %1003 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1001, %cst_160 : tensor<2x640x32x32xf16>, tensor<640x640x1x1xf16>) outs(%1002 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16> | |
| %1004 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1003, %850 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%555 : tensor<2x640x32x32xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640x32x32xf16> | |
| %padded_926 = tensor.pad %1004 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x640x32x32xf16> to tensor<2x640x34x34xf16> | |
| %1005 = tensor.empty() : tensor<2x640x16x16xf16> | |
| %1006 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_163 : tensor<640xf16>) outs(%1005 : tensor<2x640x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x16x16xf16> | |
| %1007 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%padded_926, %cst_162 : tensor<2x640x34x34xf16>, tensor<640x640x3x3xf16>) outs(%1006 : tensor<2x640x16x16xf16>) -> tensor<2x640x16x16xf16> | |
| %collapsed_927 = tensor.collapse_shape %1007 [[0], [1], [2, 3]] : tensor<2x640x16x16xf16> into tensor<2x640x256xf16> | |
| %expanded_928 = tensor.expand_shape %collapsed_927 [[0], [1, 2], [3]] : tensor<2x640x256xf16> into tensor<2x32x20x256xf16> | |
| %1008 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1009 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1008 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %1010 = tensor.empty() : tensor<2x32x20x256xf32> | |
| %1011 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1009 : tensor<f32>) outs(%1010 : tensor<2x32x20x256xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x20x256xf32> | |
| %1012 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_928 : tensor<2x32x20x256xf16>) outs(%1011 : tensor<2x32x20x256xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x20x256xf32> | |
| %1013 = tensor.empty() : tensor<2x32x20x256xf64> | |
| %1014 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1012 : tensor<2x32x20x256xf32>) outs(%1013 : tensor<2x32x20x256xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x256xf64> | |
| %1015 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1016 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1014 : tensor<2x32x20x256xf64>) outs(%1015 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1017 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1016 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_710 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1018 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1014, %1017 : tensor<2x32x20x256xf64>, tensor<2x32x1x1xf64>) outs(%1013 : tensor<2x32x20x256xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x256xf64> | |
| %1019 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1018 : tensor<2x32x20x256xf64>) outs(%1013 : tensor<2x32x20x256xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x20x256xf64> | |
| %1020 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1021 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1019 : tensor<2x32x20x256xf64>) outs(%1020 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1022 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1021 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_710 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1023 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1022 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1024 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %1025 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1012 : tensor<2x32x20x256xf32>) outs(%1024 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1026 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1025 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_711 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1027 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1023, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1028 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1027 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1029 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_928, %1026 : tensor<2x32x20x256xf16>, tensor<2x32x1x1xf32>) outs(%1010 : tensor<2x32x20x256xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x20x256xf32> | |
| %1030 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1029, %1028 : tensor<2x32x20x256xf32>, tensor<2x32x1x1xf32>) outs(%1010 : tensor<2x32x20x256xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x20x256xf32> | |
| %collapsed_929 = tensor.collapse_shape %1030 [[0], [1, 2], [3]] : tensor<2x32x20x256xf32> into tensor<2x640x256xf32> | |
| %expanded_930 = tensor.expand_shape %collapsed_929 [[0], [1], [2, 3]] : tensor<2x640x256xf32> into tensor<2x640x16x16xf32> | |
| %expanded_931 = tensor.expand_shape %cst_164 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16> | |
| %1031 = tensor.empty() : tensor<2x640x16x16xf32> | |
| %1032 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_930, %expanded_931 : tensor<2x640x16x16xf32>, tensor<640x1x1xf16>) outs(%1031 : tensor<2x640x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x640x16x16xf32> | |
| %expanded_932 = tensor.expand_shape %cst_165 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16> | |
| %1033 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1032, %expanded_932 : tensor<2x640x16x16xf32>, tensor<640x1x1xf16>) outs(%1031 : tensor<2x640x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x640x16x16xf32> | |
| %1034 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1035 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1034 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %1036 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1035 : tensor<f16>) outs(%1005 : tensor<2x640x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x640x16x16xf16> | |
| %1037 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1033 : tensor<2x640x16x16xf32>) outs(%1036 : tensor<2x640x16x16xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640x16x16xf16> | |
| %1038 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1037 : tensor<2x640x16x16xf16>) outs(%1005 : tensor<2x640x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x640x16x16xf16> | |
| %1039 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1038, %1037 : tensor<2x640x16x16xf16>, tensor<2x640x16x16xf16>) outs(%1005 : tensor<2x640x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x640x16x16xf16> | |
| %padded_933 = tensor.pad %1039 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x640x16x16xf16> to tensor<2x640x18x18xf16> | |
| %1040 = tensor.empty() : tensor<2x1280x16x16xf16> | |
| %1041 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_167 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1042 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_933, %cst_166 : tensor<2x640x18x18xf16>, tensor<1280x640x3x3xf16>) outs(%1041 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %1043 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280xf16> | |
| %1044 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%1043, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %1045 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_168 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %1046 = linalg.fill ins(%cst_694 : f16) outs(%30 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %1047 = linalg.matmul ins(%1044, %1045 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%1046 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %1048 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_169, %1047 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %expanded_934 = tensor.expand_shape %1048 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16> | |
| %1049 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1042, %expanded_934 : tensor<2x1280x16x16xf16>, tensor<2x1280x1x1xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %collapsed_935 = tensor.collapse_shape %1049 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16> | |
| %expanded_936 = tensor.expand_shape %collapsed_935 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16> | |
| %1050 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1051 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1050 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %1052 = tensor.empty() : tensor<2x32x40x256xf32> | |
| %1053 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1051 : tensor<f32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %1054 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_936 : tensor<2x32x40x256xf16>) outs(%1053 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %1055 = tensor.empty() : tensor<2x32x40x256xf64> | |
| %1056 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1054 : tensor<2x32x40x256xf32>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %1057 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1058 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1056 : tensor<2x32x40x256xf64>) outs(%1057 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1059 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1058 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1060 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1056, %1059 : tensor<2x32x40x256xf64>, tensor<2x32x1x1xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %1061 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1060 : tensor<2x32x40x256xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %1062 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1063 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1061 : tensor<2x32x40x256xf64>) outs(%1062 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1064 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1063 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1065 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1064 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1066 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %1067 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1054 : tensor<2x32x40x256xf32>) outs(%1066 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1068 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1067 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_706 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1069 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1065, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1070 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1069 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1071 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_936, %1068 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %1072 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1071, %1070 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %collapsed_937 = tensor.collapse_shape %1072 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32> | |
| %expanded_938 = tensor.expand_shape %collapsed_937 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32> | |
| %expanded_939 = tensor.expand_shape %cst_170 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1073 = tensor.empty() : tensor<2x1280x16x16xf32> | |
| %1074 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_938, %expanded_939 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %expanded_940 = tensor.expand_shape %cst_171 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1075 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1074, %expanded_940 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %1076 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1077 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1076 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %1078 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1077 : tensor<f16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1079 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1075 : tensor<2x1280x16x16xf32>) outs(%1078 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1080 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1079 : tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1081 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1080, %1079 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %padded_941 = tensor.pad %1081 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x16x16xf16> to tensor<2x1280x18x18xf16> | |
| %1082 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_173 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1083 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_941, %cst_172 : tensor<2x1280x18x18xf16>, tensor<1280x1280x3x3xf16>) outs(%1082 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %1084 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_175 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1085 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1007, %cst_174 : tensor<2x640x16x16xf16>, tensor<1280x640x1x1xf16>) outs(%1084 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %1086 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1085, %1083 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1087 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1086, %cst_3 : tensor<2x1280x16x16xf16>, tensor<f64>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %collapsed_942 = tensor.collapse_shape %1087 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16> | |
| %expanded_943 = tensor.expand_shape %collapsed_942 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16> | |
| %1088 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1089 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1088 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %1090 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1089 : tensor<f32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %1091 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_943 : tensor<2x32x40x256xf16>) outs(%1090 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %1092 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1091 : tensor<2x32x40x256xf32>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %1093 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1094 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1092 : tensor<2x32x40x256xf64>) outs(%1093 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1095 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1094 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1096 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1092, %1095 : tensor<2x32x40x256xf64>, tensor<2x32x1x1xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %1097 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1096 : tensor<2x32x40x256xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %1098 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1099 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1097 : tensor<2x32x40x256xf64>) outs(%1098 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1100 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1099 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1101 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1100 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1102 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %1103 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1091 : tensor<2x32x40x256xf32>) outs(%1102 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1104 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1103 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_706 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1105 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1101, %cst_2 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1106 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1105 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1107 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_943, %1104 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %1108 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1107, %1106 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %collapsed_944 = tensor.collapse_shape %1108 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32> | |
| %expanded_945 = tensor.expand_shape %collapsed_944 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32> | |
| %expanded_946 = tensor.expand_shape %cst_176 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1109 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_945, %expanded_946 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %expanded_947 = tensor.expand_shape %cst_177 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1110 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1109, %expanded_947 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %1111 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1112 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1111 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %1113 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1112 : tensor<f16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1114 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1110 : tensor<2x1280x16x16xf32>) outs(%1113 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1115 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_179 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1116 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1114, %cst_178 : tensor<2x1280x16x16xf16>, tensor<1280x1280x1x1xf16>) outs(%1115 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %1117 = tensor.empty() : tensor<2x16x16x1280xf16> | |
| %1118 = linalg.generic {indexing_maps = [#map11, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1116 : tensor<2x1280x16x16xf16>) outs(%1117 : tensor<2x16x16x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x16x16x1280xf16> | |
| %collapsed_948 = tensor.collapse_shape %1118 [[0], [1, 2], [3]] : tensor<2x16x16x1280xf16> into tensor<2x256x1280xf16> | |
| %1119 = tensor.empty() : tensor<2x256x1xf16> | |
| %1120 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %1121 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_948 : tensor<2x256x1280xf16>) outs(%1120 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1122 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1121 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1123 = tensor.empty() : tensor<2x256x1280xf16> | |
| %1124 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1122 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1125 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_948, %1124 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1126 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1125 : tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1127 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %1128 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1126 : tensor<2x256x1280xf16>) outs(%1127 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1129 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1128 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1130 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1129 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1131 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1130 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1132 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1131 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1133 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1125, %1132 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1134 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1133, %cst_180 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1135 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1134, %cst_181 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1136 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_182 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_949 = tensor.collapse_shape %1135 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16> | |
| %1137 = tensor.empty() : tensor<512x1280xf16> | |
| %1138 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1139 = linalg.matmul ins(%collapsed_949, %1136 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1138 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1140 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_183 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %1141 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1142 = linalg.matmul ins(%collapsed_949, %1140 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1141 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1143 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_184 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %1144 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1145 = linalg.matmul ins(%collapsed_949, %1143 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1144 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %expanded_950 = tensor.expand_shape %1139 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %1146 = tensor.empty() : tensor<2x8x256x160xf16> | |
| %1147 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_950 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_951 = tensor.collapse_shape %1147 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %expanded_952 = tensor.expand_shape %1142 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %1148 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_952 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_953 = tensor.collapse_shape %1148 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %expanded_954 = tensor.expand_shape %1145 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %1149 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_954 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_955 = tensor.collapse_shape %1149 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %1150 = tensor.empty() : tensor<16x160x256xf16> | |
| %1151 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_953 : tensor<16x256x160xf16>) outs(%1150 : tensor<16x160x256xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x160x256xf16> | |
| %1152 = tensor.empty() : tensor<16x256x256xf16> | |
| %1153 = linalg.fill ins(%cst_694 : f16) outs(%1152 : tensor<16x256x256xf16>) -> tensor<16x256x256xf16> | |
| %1154 = linalg.batch_matmul ins(%collapsed_951, %1151 : tensor<16x256x160xf16>, tensor<16x160x256xf16>) outs(%1153 : tensor<16x256x256xf16>) -> tensor<16x256x256xf16> | |
| %1155 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1154, %cst : tensor<16x256x256xf16>, tensor<f64>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %1156 = tensor.empty() : tensor<16x256x1xi64> | |
| %1157 = linalg.fill ins(%c0_i64 : i64) outs(%1156 : tensor<16x256x1xi64>) -> tensor<16x256x1xi64> | |
| %1158 = tensor.empty() : tensor<16x256x1xf16> | |
| %1159 = linalg.fill ins(%cst_696 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %1160:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1155 : tensor<16x256x256xf16>) outs(%1159, %1157 : tensor<16x256x1xf16>, tensor<16x256x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x256x1xf16>, tensor<16x256x1xi64>) | |
| %1161 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1155, %1160#0 : tensor<16x256x256xf16>, tensor<16x256x1xf16>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %1162 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1161 : tensor<16x256x256xf16>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %1163 = linalg.fill ins(%cst_694 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %1164 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1162 : tensor<16x256x256xf16>) outs(%1163 : tensor<16x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x1xf16> | |
| %1165 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1162, %1164 : tensor<16x256x256xf16>, tensor<16x256x1xf16>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %1166 = tensor.empty() : tensor<16x256x160xf16> | |
| %1167 = linalg.fill ins(%cst_694 : f16) outs(%1166 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %1168 = linalg.batch_matmul ins(%1165, %collapsed_955 : tensor<16x256x256xf16>, tensor<16x256x160xf16>) outs(%1167 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %expanded_956 = tensor.expand_shape %1168 [[0, 1], [2], [3]] : tensor<16x256x160xf16> into tensor<2x8x256x160xf16> | |
| %1169 = tensor.empty() : tensor<2x256x8x160xf16> | |
| %1170 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_956 : tensor<2x8x256x160xf16>) outs(%1169 : tensor<2x256x8x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x8x160xf16> | |
| %1171 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_185 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_957 = tensor.collapse_shape %1170 [[0, 1], [2, 3]] : tensor<2x256x8x160xf16> into tensor<512x1280xf16> | |
| %1172 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1173 = linalg.matmul ins(%collapsed_957, %1171 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1172 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1174 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_186, %1173 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1137 : tensor<512x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<512x1280xf16> | |
| %expanded_958 = tensor.expand_shape %1174 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16> | |
| %1175 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_958, %collapsed_948 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1176 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %1177 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1175 : tensor<2x256x1280xf16>) outs(%1176 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1178 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1177 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1179 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1178 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1180 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1175, %1179 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1181 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1180 : tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1182 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %1183 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1181 : tensor<2x256x1280xf16>) outs(%1182 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1184 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1183 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1185 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1184 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1186 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1185 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1187 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1186 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1188 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1180, %1187 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1189 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1188, %cst_187 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1190 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1189, %cst_188 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1191 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_189 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_959 = tensor.collapse_shape %1190 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16> | |
| %1192 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1193 = linalg.matmul ins(%collapsed_959, %1191 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1192 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1194 = tensor.empty() : tensor<768x1280xf16> | |
| %1195 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_190 : tensor<1280x768xf16>) outs(%1194 : tensor<768x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x1280xf16> | |
| %1196 = tensor.empty() : tensor<154x1280xf16> | |
| %1197 = linalg.fill ins(%cst_694 : f16) outs(%1196 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %1198 = linalg.matmul ins(%collapsed_761, %1195 : tensor<154x768xf16>, tensor<768x1280xf16>) outs(%1197 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %1199 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_191 : tensor<1280x768xf16>) outs(%1194 : tensor<768x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x1280xf16> | |
| %1200 = linalg.fill ins(%cst_694 : f16) outs(%1196 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %1201 = linalg.matmul ins(%collapsed_761, %1199 : tensor<154x768xf16>, tensor<768x1280xf16>) outs(%1200 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %expanded_960 = tensor.expand_shape %1193 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %1202 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_960 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_961 = tensor.collapse_shape %1202 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %expanded_962 = tensor.expand_shape %1198 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x8x160xf16> | |
| %1203 = tensor.empty() : tensor<2x8x77x160xf16> | |
| %1204 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_962 : tensor<2x77x8x160xf16>) outs(%1203 : tensor<2x8x77x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x160xf16> | |
| %collapsed_963 = tensor.collapse_shape %1204 [[0, 1], [2], [3]] : tensor<2x8x77x160xf16> into tensor<16x77x160xf16> | |
| %expanded_964 = tensor.expand_shape %1201 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x8x160xf16> | |
| %1205 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_964 : tensor<2x77x8x160xf16>) outs(%1203 : tensor<2x8x77x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x160xf16> | |
| %collapsed_965 = tensor.collapse_shape %1205 [[0, 1], [2], [3]] : tensor<2x8x77x160xf16> into tensor<16x77x160xf16> | |
| %1206 = tensor.empty() : tensor<16x160x77xf16> | |
| %1207 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_963 : tensor<16x77x160xf16>) outs(%1206 : tensor<16x160x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x160x77xf16> | |
| %1208 = tensor.empty() : tensor<16x256x77xf16> | |
| %1209 = linalg.fill ins(%cst_694 : f16) outs(%1208 : tensor<16x256x77xf16>) -> tensor<16x256x77xf16> | |
| %1210 = linalg.batch_matmul ins(%collapsed_961, %1207 : tensor<16x256x160xf16>, tensor<16x160x77xf16>) outs(%1209 : tensor<16x256x77xf16>) -> tensor<16x256x77xf16> | |
| %1211 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1210, %cst : tensor<16x256x77xf16>, tensor<f64>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %1212 = linalg.fill ins(%c0_i64 : i64) outs(%1156 : tensor<16x256x1xi64>) -> tensor<16x256x1xi64> | |
| %1213 = linalg.fill ins(%cst_696 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %1214:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1211 : tensor<16x256x77xf16>) outs(%1213, %1212 : tensor<16x256x1xf16>, tensor<16x256x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x256x1xf16>, tensor<16x256x1xi64>) | |
| %1215 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1211, %1214#0 : tensor<16x256x77xf16>, tensor<16x256x1xf16>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %1216 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1215 : tensor<16x256x77xf16>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %1217 = linalg.fill ins(%cst_694 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %1218 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1216 : tensor<16x256x77xf16>) outs(%1217 : tensor<16x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x1xf16> | |
| %1219 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1216, %1218 : tensor<16x256x77xf16>, tensor<16x256x1xf16>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %1220 = linalg.fill ins(%cst_694 : f16) outs(%1166 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %1221 = linalg.batch_matmul ins(%1219, %collapsed_965 : tensor<16x256x77xf16>, tensor<16x77x160xf16>) outs(%1220 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %expanded_966 = tensor.expand_shape %1221 [[0, 1], [2], [3]] : tensor<16x256x160xf16> into tensor<2x8x256x160xf16> | |
| %1222 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_966 : tensor<2x8x256x160xf16>) outs(%1169 : tensor<2x256x8x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x8x160xf16> | |
| %1223 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_192 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_967 = tensor.collapse_shape %1222 [[0, 1], [2, 3]] : tensor<2x256x8x160xf16> into tensor<512x1280xf16> | |
| %1224 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1225 = linalg.matmul ins(%collapsed_967, %1223 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1224 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1226 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_193, %1225 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1137 : tensor<512x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<512x1280xf16> | |
| %expanded_968 = tensor.expand_shape %1226 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16> | |
| %1227 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_968, %1175 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1228 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %1229 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1227 : tensor<2x256x1280xf16>) outs(%1228 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1230 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1229 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1231 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1230 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1232 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1227, %1231 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1233 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1232 : tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1234 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %1235 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1233 : tensor<2x256x1280xf16>) outs(%1234 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1236 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1235 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1237 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1236 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1238 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1237 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1239 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1238 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1240 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1232, %1239 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1241 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1240, %cst_194 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1242 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1241, %cst_195 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1243 = tensor.empty() : tensor<1280x10240xf16> | |
| %1244 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_196 : tensor<10240x1280xf16>) outs(%1243 : tensor<1280x10240xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x10240xf16> | |
| %collapsed_969 = tensor.collapse_shape %1242 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16> | |
| %1245 = tensor.empty() : tensor<512x10240xf16> | |
| %1246 = linalg.fill ins(%cst_694 : f16) outs(%1245 : tensor<512x10240xf16>) -> tensor<512x10240xf16> | |
| %1247 = linalg.matmul ins(%collapsed_969, %1244 : tensor<512x1280xf16>, tensor<1280x10240xf16>) outs(%1246 : tensor<512x10240xf16>) -> tensor<512x10240xf16> | |
| %1248 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_197, %1247 : tensor<10240xf16>, tensor<512x10240xf16>) outs(%1245 : tensor<512x10240xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<512x10240xf16> | |
| %expanded_970 = tensor.expand_shape %1248 [[0, 1], [2]] : tensor<512x10240xf16> into tensor<2x256x10240xf16> | |
| %extracted_slice_971 = tensor.extract_slice %expanded_970[0, 0, 0] [2, 256, 5120] [1, 1, 1] : tensor<2x256x10240xf16> to tensor<2x256x5120xf16> | |
| %extracted_slice_972 = tensor.extract_slice %expanded_970[0, 0, 5120] [2, 256, 5120] [1, 1, 1] : tensor<2x256x10240xf16> to tensor<2x256x5120xf16> | |
| %1249 = tensor.empty() : tensor<2x256x5120xf16> | |
| %1250 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_972 : tensor<2x256x5120xf16>) outs(%1249 : tensor<2x256x5120xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.sqrt %cst_697 : f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| %4282 = math.erf %4281 : f16 | |
| %4283 = arith.addf %4282, %cst_695 : f16 | |
| %4284 = arith.mulf %4283, %cst_698 : f16 | |
| %4285 = arith.mulf %in, %4284 : f16 | |
| linalg.yield %4285 : f16 | |
| } -> tensor<2x256x5120xf16> | |
| %1251 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_971, %1250 : tensor<2x256x5120xf16>, tensor<2x256x5120xf16>) outs(%1249 : tensor<2x256x5120xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x5120xf16> | |
| %1252 = tensor.empty() : tensor<5120x1280xf16> | |
| %1253 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_198 : tensor<1280x5120xf16>) outs(%1252 : tensor<5120x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<5120x1280xf16> | |
| %collapsed_973 = tensor.collapse_shape %1251 [[0, 1], [2]] : tensor<2x256x5120xf16> into tensor<512x5120xf16> | |
| %1254 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1255 = linalg.matmul ins(%collapsed_973, %1253 : tensor<512x5120xf16>, tensor<5120x1280xf16>) outs(%1254 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1256 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_199, %1255 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1137 : tensor<512x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<512x1280xf16> | |
| %expanded_974 = tensor.expand_shape %1256 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16> | |
| %1257 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_974, %1227 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %expanded_975 = tensor.expand_shape %1257 [[0], [1, 2], [3]] : tensor<2x256x1280xf16> into tensor<2x16x16x1280xf16> | |
| %1258 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_975 : tensor<2x16x16x1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1259 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_201 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1260 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1258, %cst_200 : tensor<2x1280x16x16xf16>, tensor<1280x1280x1x1xf16>) outs(%1259 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %1261 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1260, %1087 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %collapsed_976 = tensor.collapse_shape %1261 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16> | |
| %expanded_977 = tensor.expand_shape %collapsed_976 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16> | |
| %1262 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1263 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1262 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %1264 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1263 : tensor<f32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %1265 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_977 : tensor<2x32x40x256xf16>) outs(%1264 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %1266 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1265 : tensor<2x32x40x256xf32>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %1267 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1268 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1266 : tensor<2x32x40x256xf64>) outs(%1267 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1269 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1268 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1270 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1266, %1269 : tensor<2x32x40x256xf64>, tensor<2x32x1x1xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %1271 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1270 : tensor<2x32x40x256xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %1272 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1273 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1271 : tensor<2x32x40x256xf64>) outs(%1272 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1274 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1273 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1275 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1274 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1276 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %1277 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1265 : tensor<2x32x40x256xf32>) outs(%1276 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1278 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1277 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_706 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1279 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1275, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1280 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1279 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1281 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_977, %1278 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %1282 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1281, %1280 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %collapsed_978 = tensor.collapse_shape %1282 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32> | |
| %expanded_979 = tensor.expand_shape %collapsed_978 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32> | |
| %expanded_980 = tensor.expand_shape %cst_202 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1283 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_979, %expanded_980 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %expanded_981 = tensor.expand_shape %cst_203 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1284 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1283, %expanded_981 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %1285 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1286 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1285 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %1287 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1286 : tensor<f16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1288 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1284 : tensor<2x1280x16x16xf32>) outs(%1287 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1289 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1288 : tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1290 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1289, %1288 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %padded_982 = tensor.pad %1290 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x16x16xf16> to tensor<2x1280x18x18xf16> | |
| %1291 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_205 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1292 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_982, %cst_204 : tensor<2x1280x18x18xf16>, tensor<1280x1280x3x3xf16>) outs(%1291 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %1293 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280xf16> | |
| %1294 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%1293, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %1295 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_206 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %1296 = linalg.fill ins(%cst_694 : f16) outs(%30 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %1297 = linalg.matmul ins(%1294, %1295 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%1296 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %1298 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_207, %1297 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %expanded_983 = tensor.expand_shape %1298 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16> | |
| %1299 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1292, %expanded_983 : tensor<2x1280x16x16xf16>, tensor<2x1280x1x1xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %collapsed_984 = tensor.collapse_shape %1299 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16> | |
| %expanded_985 = tensor.expand_shape %collapsed_984 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16> | |
| %1300 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1301 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1300 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %1302 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1301 : tensor<f32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %1303 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_985 : tensor<2x32x40x256xf16>) outs(%1302 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %1304 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1303 : tensor<2x32x40x256xf32>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %1305 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1306 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1304 : tensor<2x32x40x256xf64>) outs(%1305 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1307 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1306 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1308 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1304, %1307 : tensor<2x32x40x256xf64>, tensor<2x32x1x1xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %1309 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1308 : tensor<2x32x40x256xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %1310 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1311 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1309 : tensor<2x32x40x256xf64>) outs(%1310 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1312 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1311 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1313 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1312 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1314 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %1315 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1303 : tensor<2x32x40x256xf32>) outs(%1314 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1316 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1315 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_706 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1317 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1313, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1318 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1317 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1319 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_985, %1316 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %1320 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1319, %1318 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %collapsed_986 = tensor.collapse_shape %1320 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32> | |
| %expanded_987 = tensor.expand_shape %collapsed_986 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32> | |
| %expanded_988 = tensor.expand_shape %cst_208 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1321 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_987, %expanded_988 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %expanded_989 = tensor.expand_shape %cst_209 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1322 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1321, %expanded_989 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %1323 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1324 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1323 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %1325 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1324 : tensor<f16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1326 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1322 : tensor<2x1280x16x16xf32>) outs(%1325 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1327 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1326 : tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1328 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1327, %1326 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %padded_990 = tensor.pad %1328 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x16x16xf16> to tensor<2x1280x18x18xf16> | |
| %1329 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_211 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1330 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_990, %cst_210 : tensor<2x1280x18x18xf16>, tensor<1280x1280x3x3xf16>) outs(%1329 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %1331 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1261, %1330 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1332 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1331, %cst_3 : tensor<2x1280x16x16xf16>, tensor<f64>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %collapsed_991 = tensor.collapse_shape %1332 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16> | |
| %expanded_992 = tensor.expand_shape %collapsed_991 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16> | |
| %1333 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1334 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1333 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %1335 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1334 : tensor<f32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %1336 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_992 : tensor<2x32x40x256xf16>) outs(%1335 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %1337 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1336 : tensor<2x32x40x256xf32>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %1338 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1339 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1337 : tensor<2x32x40x256xf64>) outs(%1338 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1340 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1339 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1341 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1337, %1340 : tensor<2x32x40x256xf64>, tensor<2x32x1x1xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %1342 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1341 : tensor<2x32x40x256xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %1343 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1344 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1342 : tensor<2x32x40x256xf64>) outs(%1343 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1345 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1344 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1346 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1345 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1347 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %1348 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1336 : tensor<2x32x40x256xf32>) outs(%1347 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1349 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1348 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_706 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1350 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1346, %cst_2 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1351 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1350 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1352 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_992, %1349 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %1353 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1352, %1351 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %collapsed_993 = tensor.collapse_shape %1353 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32> | |
| %expanded_994 = tensor.expand_shape %collapsed_993 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32> | |
| %expanded_995 = tensor.expand_shape %cst_212 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1354 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_994, %expanded_995 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %expanded_996 = tensor.expand_shape %cst_213 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1355 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1354, %expanded_996 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %1356 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1357 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1356 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %1358 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1357 : tensor<f16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1359 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1355 : tensor<2x1280x16x16xf32>) outs(%1358 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1360 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_215 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1361 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1359, %cst_214 : tensor<2x1280x16x16xf16>, tensor<1280x1280x1x1xf16>) outs(%1360 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %1362 = linalg.generic {indexing_maps = [#map11, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1361 : tensor<2x1280x16x16xf16>) outs(%1117 : tensor<2x16x16x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x16x16x1280xf16> | |
| %collapsed_997 = tensor.collapse_shape %1362 [[0], [1, 2], [3]] : tensor<2x16x16x1280xf16> into tensor<2x256x1280xf16> | |
| %1363 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %1364 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_997 : tensor<2x256x1280xf16>) outs(%1363 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1365 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1364 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1366 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1365 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1367 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_997, %1366 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1368 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1367 : tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1369 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %1370 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1368 : tensor<2x256x1280xf16>) outs(%1369 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1371 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1370 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1372 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1371 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1373 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1372 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1374 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1373 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1375 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1367, %1374 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1376 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1375, %cst_216 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1377 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1376, %cst_217 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1378 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_218 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_998 = tensor.collapse_shape %1377 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16> | |
| %1379 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1380 = linalg.matmul ins(%collapsed_998, %1378 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1379 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1381 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_219 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %1382 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1383 = linalg.matmul ins(%collapsed_998, %1381 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1382 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1384 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_220 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %1385 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1386 = linalg.matmul ins(%collapsed_998, %1384 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1385 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %expanded_999 = tensor.expand_shape %1380 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %1387 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_999 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_1000 = tensor.collapse_shape %1387 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %expanded_1001 = tensor.expand_shape %1383 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %1388 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1001 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_1002 = tensor.collapse_shape %1388 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %expanded_1003 = tensor.expand_shape %1386 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %1389 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1003 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_1004 = tensor.collapse_shape %1389 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %1390 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1002 : tensor<16x256x160xf16>) outs(%1150 : tensor<16x160x256xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x160x256xf16> | |
| %1391 = linalg.fill ins(%cst_694 : f16) outs(%1152 : tensor<16x256x256xf16>) -> tensor<16x256x256xf16> | |
| %1392 = linalg.batch_matmul ins(%collapsed_1000, %1390 : tensor<16x256x160xf16>, tensor<16x160x256xf16>) outs(%1391 : tensor<16x256x256xf16>) -> tensor<16x256x256xf16> | |
| %1393 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1392, %cst : tensor<16x256x256xf16>, tensor<f64>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %1394 = linalg.fill ins(%c0_i64 : i64) outs(%1156 : tensor<16x256x1xi64>) -> tensor<16x256x1xi64> | |
| %1395 = linalg.fill ins(%cst_696 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %1396:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1393 : tensor<16x256x256xf16>) outs(%1395, %1394 : tensor<16x256x1xf16>, tensor<16x256x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x256x1xf16>, tensor<16x256x1xi64>) | |
| %1397 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1393, %1396#0 : tensor<16x256x256xf16>, tensor<16x256x1xf16>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %1398 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1397 : tensor<16x256x256xf16>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %1399 = linalg.fill ins(%cst_694 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %1400 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1398 : tensor<16x256x256xf16>) outs(%1399 : tensor<16x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x1xf16> | |
| %1401 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1398, %1400 : tensor<16x256x256xf16>, tensor<16x256x1xf16>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %1402 = linalg.fill ins(%cst_694 : f16) outs(%1166 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %1403 = linalg.batch_matmul ins(%1401, %collapsed_1004 : tensor<16x256x256xf16>, tensor<16x256x160xf16>) outs(%1402 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %expanded_1005 = tensor.expand_shape %1403 [[0, 1], [2], [3]] : tensor<16x256x160xf16> into tensor<2x8x256x160xf16> | |
| %1404 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1005 : tensor<2x8x256x160xf16>) outs(%1169 : tensor<2x256x8x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x8x160xf16> | |
| %1405 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_221 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1006 = tensor.collapse_shape %1404 [[0, 1], [2, 3]] : tensor<2x256x8x160xf16> into tensor<512x1280xf16> | |
| %1406 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1407 = linalg.matmul ins(%collapsed_1006, %1405 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1406 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1408 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_222, %1407 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1137 : tensor<512x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<512x1280xf16> | |
| %expanded_1007 = tensor.expand_shape %1408 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16> | |
| %1409 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1007, %collapsed_997 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1410 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %1411 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1409 : tensor<2x256x1280xf16>) outs(%1410 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1412 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1411 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1413 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1412 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1414 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1409, %1413 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1415 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1414 : tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1416 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %1417 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1415 : tensor<2x256x1280xf16>) outs(%1416 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1418 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1417 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1419 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1418 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1420 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1419 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1421 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1420 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1422 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1414, %1421 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1423 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1422, %cst_223 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1424 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1423, %cst_224 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1425 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_225 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1008 = tensor.collapse_shape %1424 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16> | |
| %1426 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1427 = linalg.matmul ins(%collapsed_1008, %1425 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1426 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1428 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_226 : tensor<1280x768xf16>) outs(%1194 : tensor<768x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x1280xf16> | |
| %1429 = linalg.fill ins(%cst_694 : f16) outs(%1196 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %1430 = linalg.matmul ins(%collapsed_761, %1428 : tensor<154x768xf16>, tensor<768x1280xf16>) outs(%1429 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %1431 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_227 : tensor<1280x768xf16>) outs(%1194 : tensor<768x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x1280xf16> | |
| %1432 = linalg.fill ins(%cst_694 : f16) outs(%1196 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %1433 = linalg.matmul ins(%collapsed_761, %1431 : tensor<154x768xf16>, tensor<768x1280xf16>) outs(%1432 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %expanded_1009 = tensor.expand_shape %1427 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %1434 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1009 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_1010 = tensor.collapse_shape %1434 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %expanded_1011 = tensor.expand_shape %1430 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x8x160xf16> | |
| %1435 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1011 : tensor<2x77x8x160xf16>) outs(%1203 : tensor<2x8x77x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x160xf16> | |
| %collapsed_1012 = tensor.collapse_shape %1435 [[0, 1], [2], [3]] : tensor<2x8x77x160xf16> into tensor<16x77x160xf16> | |
| %expanded_1013 = tensor.expand_shape %1433 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x8x160xf16> | |
| %1436 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1013 : tensor<2x77x8x160xf16>) outs(%1203 : tensor<2x8x77x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x160xf16> | |
| %collapsed_1014 = tensor.collapse_shape %1436 [[0, 1], [2], [3]] : tensor<2x8x77x160xf16> into tensor<16x77x160xf16> | |
| %1437 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1012 : tensor<16x77x160xf16>) outs(%1206 : tensor<16x160x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x160x77xf16> | |
| %1438 = linalg.fill ins(%cst_694 : f16) outs(%1208 : tensor<16x256x77xf16>) -> tensor<16x256x77xf16> | |
| %1439 = linalg.batch_matmul ins(%collapsed_1010, %1437 : tensor<16x256x160xf16>, tensor<16x160x77xf16>) outs(%1438 : tensor<16x256x77xf16>) -> tensor<16x256x77xf16> | |
| %1440 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1439, %cst : tensor<16x256x77xf16>, tensor<f64>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %1441 = linalg.fill ins(%c0_i64 : i64) outs(%1156 : tensor<16x256x1xi64>) -> tensor<16x256x1xi64> | |
| %1442 = linalg.fill ins(%cst_696 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %1443:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1440 : tensor<16x256x77xf16>) outs(%1442, %1441 : tensor<16x256x1xf16>, tensor<16x256x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x256x1xf16>, tensor<16x256x1xi64>) | |
| %1444 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1440, %1443#0 : tensor<16x256x77xf16>, tensor<16x256x1xf16>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %1445 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1444 : tensor<16x256x77xf16>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %1446 = linalg.fill ins(%cst_694 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %1447 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1445 : tensor<16x256x77xf16>) outs(%1446 : tensor<16x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x1xf16> | |
| %1448 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1445, %1447 : tensor<16x256x77xf16>, tensor<16x256x1xf16>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %1449 = linalg.fill ins(%cst_694 : f16) outs(%1166 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %1450 = linalg.batch_matmul ins(%1448, %collapsed_1014 : tensor<16x256x77xf16>, tensor<16x77x160xf16>) outs(%1449 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %expanded_1015 = tensor.expand_shape %1450 [[0, 1], [2], [3]] : tensor<16x256x160xf16> into tensor<2x8x256x160xf16> | |
| %1451 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1015 : tensor<2x8x256x160xf16>) outs(%1169 : tensor<2x256x8x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x8x160xf16> | |
| %1452 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_228 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1016 = tensor.collapse_shape %1451 [[0, 1], [2, 3]] : tensor<2x256x8x160xf16> into tensor<512x1280xf16> | |
| %1453 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1454 = linalg.matmul ins(%collapsed_1016, %1452 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1453 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1455 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_229, %1454 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1137 : tensor<512x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<512x1280xf16> | |
| %expanded_1017 = tensor.expand_shape %1455 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16> | |
| %1456 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1017, %1409 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1457 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %1458 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1456 : tensor<2x256x1280xf16>) outs(%1457 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1459 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1458 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1460 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1459 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1461 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1456, %1460 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1462 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1461 : tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1463 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %1464 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1462 : tensor<2x256x1280xf16>) outs(%1463 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1465 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1464 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1466 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1465 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1467 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1466 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %1468 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1467 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1469 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1461, %1468 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1470 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1469, %cst_230 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1471 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1470, %cst_231 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %1472 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_232 : tensor<10240x1280xf16>) outs(%1243 : tensor<1280x10240xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x10240xf16> | |
| %collapsed_1018 = tensor.collapse_shape %1471 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16> | |
| %1473 = linalg.fill ins(%cst_694 : f16) outs(%1245 : tensor<512x10240xf16>) -> tensor<512x10240xf16> | |
| %1474 = linalg.matmul ins(%collapsed_1018, %1472 : tensor<512x1280xf16>, tensor<1280x10240xf16>) outs(%1473 : tensor<512x10240xf16>) -> tensor<512x10240xf16> | |
| %1475 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_233, %1474 : tensor<10240xf16>, tensor<512x10240xf16>) outs(%1245 : tensor<512x10240xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<512x10240xf16> | |
| %expanded_1019 = tensor.expand_shape %1475 [[0, 1], [2]] : tensor<512x10240xf16> into tensor<2x256x10240xf16> | |
| %extracted_slice_1020 = tensor.extract_slice %expanded_1019[0, 0, 0] [2, 256, 5120] [1, 1, 1] : tensor<2x256x10240xf16> to tensor<2x256x5120xf16> | |
| %extracted_slice_1021 = tensor.extract_slice %expanded_1019[0, 0, 5120] [2, 256, 5120] [1, 1, 1] : tensor<2x256x10240xf16> to tensor<2x256x5120xf16> | |
| %1476 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1021 : tensor<2x256x5120xf16>) outs(%1249 : tensor<2x256x5120xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.sqrt %cst_697 : f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| %4282 = math.erf %4281 : f16 | |
| %4283 = arith.addf %4282, %cst_695 : f16 | |
| %4284 = arith.mulf %4283, %cst_698 : f16 | |
| %4285 = arith.mulf %in, %4284 : f16 | |
| linalg.yield %4285 : f16 | |
| } -> tensor<2x256x5120xf16> | |
| %1477 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1020, %1476 : tensor<2x256x5120xf16>, tensor<2x256x5120xf16>) outs(%1249 : tensor<2x256x5120xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x5120xf16> | |
| %1478 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_234 : tensor<1280x5120xf16>) outs(%1252 : tensor<5120x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<5120x1280xf16> | |
| %collapsed_1022 = tensor.collapse_shape %1477 [[0, 1], [2]] : tensor<2x256x5120xf16> into tensor<512x5120xf16> | |
| %1479 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1480 = linalg.matmul ins(%collapsed_1022, %1478 : tensor<512x5120xf16>, tensor<5120x1280xf16>) outs(%1479 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %1481 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_235, %1480 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1137 : tensor<512x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<512x1280xf16> | |
| %expanded_1023 = tensor.expand_shape %1481 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16> | |
| %1482 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1023, %1456 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %expanded_1024 = tensor.expand_shape %1482 [[0], [1, 2], [3]] : tensor<2x256x1280xf16> into tensor<2x16x16x1280xf16> | |
| %1483 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1024 : tensor<2x16x16x1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1484 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_237 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %1485 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1483, %cst_236 : tensor<2x1280x16x16xf16>, tensor<1280x1280x1x1xf16>) outs(%1484 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %1486 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1485, %1332 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %padded_1025 = tensor.pad %1486 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x16x16xf16> to tensor<2x1280x18x18xf16> | |
| %1487 = tensor.empty() : tensor<2x1280x8x8xf16> | |
| %1488 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_239 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1489 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%padded_1025, %cst_238 : tensor<2x1280x18x18xf16>, tensor<1280x1280x3x3xf16>) outs(%1488 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %collapsed_1026 = tensor.collapse_shape %1489 [[0], [1], [2, 3]] : tensor<2x1280x8x8xf16> into tensor<2x1280x64xf16> | |
| %expanded_1027 = tensor.expand_shape %collapsed_1026 [[0], [1, 2], [3]] : tensor<2x1280x64xf16> into tensor<2x32x40x64xf16> | |
| %1490 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1491 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1490 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %1492 = tensor.empty() : tensor<2x32x40x64xf32> | |
| %1493 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1491 : tensor<f32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1494 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1027 : tensor<2x32x40x64xf16>) outs(%1493 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1495 = tensor.empty() : tensor<2x32x40x64xf64> | |
| %1496 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1494 : tensor<2x32x40x64xf32>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1497 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1498 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1496 : tensor<2x32x40x64xf64>) outs(%1497 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1499 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1498 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1500 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1496, %1499 : tensor<2x32x40x64xf64>, tensor<2x32x1x1xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1501 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1500 : tensor<2x32x40x64xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1502 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1503 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1501 : tensor<2x32x40x64xf64>) outs(%1502 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1504 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1503 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1505 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1504 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1506 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %1507 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1494 : tensor<2x32x40x64xf32>) outs(%1506 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1508 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1507 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_714 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1509 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1505, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1510 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1509 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1511 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1027, %1508 : tensor<2x32x40x64xf16>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1512 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1511, %1510 : tensor<2x32x40x64xf32>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %collapsed_1028 = tensor.collapse_shape %1512 [[0], [1, 2], [3]] : tensor<2x32x40x64xf32> into tensor<2x1280x64xf32> | |
| %expanded_1029 = tensor.expand_shape %collapsed_1028 [[0], [1], [2, 3]] : tensor<2x1280x64xf32> into tensor<2x1280x8x8xf32> | |
| %expanded_1030 = tensor.expand_shape %cst_240 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1513 = tensor.empty() : tensor<2x1280x8x8xf32> | |
| %1514 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1029, %expanded_1030 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %expanded_1031 = tensor.expand_shape %cst_241 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1515 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1514, %expanded_1031 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %1516 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1517 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1516 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %1518 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1517 : tensor<f16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1519 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1515 : tensor<2x1280x8x8xf32>) outs(%1518 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1520 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1519 : tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1521 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1520, %1519 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %padded_1032 = tensor.pad %1521 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x8x8xf16> to tensor<2x1280x10x10xf16> | |
| %1522 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_243 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1523 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1032, %cst_242 : tensor<2x1280x10x10xf16>, tensor<1280x1280x3x3xf16>) outs(%1522 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %1524 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280xf16> | |
| %1525 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%1524, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %1526 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_244 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %1527 = linalg.fill ins(%cst_694 : f16) outs(%30 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %1528 = linalg.matmul ins(%1525, %1526 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%1527 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %1529 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_245, %1528 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %expanded_1033 = tensor.expand_shape %1529 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16> | |
| %1530 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1523, %expanded_1033 : tensor<2x1280x8x8xf16>, tensor<2x1280x1x1xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %collapsed_1034 = tensor.collapse_shape %1530 [[0], [1], [2, 3]] : tensor<2x1280x8x8xf16> into tensor<2x1280x64xf16> | |
| %expanded_1035 = tensor.expand_shape %collapsed_1034 [[0], [1, 2], [3]] : tensor<2x1280x64xf16> into tensor<2x32x40x64xf16> | |
| %1531 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1532 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1531 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %1533 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1532 : tensor<f32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1534 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1035 : tensor<2x32x40x64xf16>) outs(%1533 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1535 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1534 : tensor<2x32x40x64xf32>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1536 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1537 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1535 : tensor<2x32x40x64xf64>) outs(%1536 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1538 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1537 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1539 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1535, %1538 : tensor<2x32x40x64xf64>, tensor<2x32x1x1xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1540 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1539 : tensor<2x32x40x64xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1541 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1542 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1540 : tensor<2x32x40x64xf64>) outs(%1541 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1543 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1542 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1544 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1543 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1545 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %1546 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1534 : tensor<2x32x40x64xf32>) outs(%1545 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1547 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1546 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_714 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1548 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1544, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1549 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1548 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1550 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1035, %1547 : tensor<2x32x40x64xf16>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1551 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1550, %1549 : tensor<2x32x40x64xf32>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %collapsed_1036 = tensor.collapse_shape %1551 [[0], [1, 2], [3]] : tensor<2x32x40x64xf32> into tensor<2x1280x64xf32> | |
| %expanded_1037 = tensor.expand_shape %collapsed_1036 [[0], [1], [2, 3]] : tensor<2x1280x64xf32> into tensor<2x1280x8x8xf32> | |
| %expanded_1038 = tensor.expand_shape %cst_246 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1552 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1037, %expanded_1038 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %expanded_1039 = tensor.expand_shape %cst_247 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1553 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1552, %expanded_1039 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %1554 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1555 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1554 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %1556 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1555 : tensor<f16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1557 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1553 : tensor<2x1280x8x8xf32>) outs(%1556 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1558 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1557 : tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1559 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1558, %1557 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %padded_1040 = tensor.pad %1559 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x8x8xf16> to tensor<2x1280x10x10xf16> | |
| %1560 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_249 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1561 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1040, %cst_248 : tensor<2x1280x10x10xf16>, tensor<1280x1280x3x3xf16>) outs(%1560 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %1562 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1489, %1561 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1563 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1562, %cst_3 : tensor<2x1280x8x8xf16>, tensor<f64>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %collapsed_1041 = tensor.collapse_shape %1563 [[0], [1], [2, 3]] : tensor<2x1280x8x8xf16> into tensor<2x1280x64xf16> | |
| %expanded_1042 = tensor.expand_shape %collapsed_1041 [[0], [1, 2], [3]] : tensor<2x1280x64xf16> into tensor<2x32x40x64xf16> | |
| %1564 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1565 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1564 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %1566 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1565 : tensor<f32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1567 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1042 : tensor<2x32x40x64xf16>) outs(%1566 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1568 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1567 : tensor<2x32x40x64xf32>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1569 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1570 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1568 : tensor<2x32x40x64xf64>) outs(%1569 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1571 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1570 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1572 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1568, %1571 : tensor<2x32x40x64xf64>, tensor<2x32x1x1xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1573 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1572 : tensor<2x32x40x64xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1574 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1575 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1573 : tensor<2x32x40x64xf64>) outs(%1574 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1576 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1575 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1577 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1576 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1578 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %1579 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1567 : tensor<2x32x40x64xf32>) outs(%1578 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1580 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1579 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_714 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1581 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1577, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1582 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1581 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1583 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1042, %1580 : tensor<2x32x40x64xf16>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1584 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1583, %1582 : tensor<2x32x40x64xf32>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %collapsed_1043 = tensor.collapse_shape %1584 [[0], [1, 2], [3]] : tensor<2x32x40x64xf32> into tensor<2x1280x64xf32> | |
| %expanded_1044 = tensor.expand_shape %collapsed_1043 [[0], [1], [2, 3]] : tensor<2x1280x64xf32> into tensor<2x1280x8x8xf32> | |
| %expanded_1045 = tensor.expand_shape %cst_250 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1585 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1044, %expanded_1045 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %expanded_1046 = tensor.expand_shape %cst_251 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1586 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1585, %expanded_1046 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %1587 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1588 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1587 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %1589 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1588 : tensor<f16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1590 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1586 : tensor<2x1280x8x8xf32>) outs(%1589 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1591 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1590 : tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1592 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1591, %1590 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %padded_1047 = tensor.pad %1592 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x8x8xf16> to tensor<2x1280x10x10xf16> | |
| %1593 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_253 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1594 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1047, %cst_252 : tensor<2x1280x10x10xf16>, tensor<1280x1280x3x3xf16>) outs(%1593 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %1595 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280xf16> | |
| %1596 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%1595, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %1597 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_254 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %1598 = linalg.fill ins(%cst_694 : f16) outs(%30 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %1599 = linalg.matmul ins(%1596, %1597 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%1598 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %1600 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_255, %1599 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %expanded_1048 = tensor.expand_shape %1600 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16> | |
| %1601 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1594, %expanded_1048 : tensor<2x1280x8x8xf16>, tensor<2x1280x1x1xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %collapsed_1049 = tensor.collapse_shape %1601 [[0], [1], [2, 3]] : tensor<2x1280x8x8xf16> into tensor<2x1280x64xf16> | |
| %expanded_1050 = tensor.expand_shape %collapsed_1049 [[0], [1, 2], [3]] : tensor<2x1280x64xf16> into tensor<2x32x40x64xf16> | |
| %1602 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1603 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1602 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %1604 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1603 : tensor<f32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1605 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1050 : tensor<2x32x40x64xf16>) outs(%1604 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1606 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1605 : tensor<2x32x40x64xf32>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1607 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1608 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1606 : tensor<2x32x40x64xf64>) outs(%1607 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1609 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1608 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1610 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1606, %1609 : tensor<2x32x40x64xf64>, tensor<2x32x1x1xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1611 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1610 : tensor<2x32x40x64xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1612 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1613 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1611 : tensor<2x32x40x64xf64>) outs(%1612 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1614 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1613 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1615 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1614 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1616 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %1617 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1605 : tensor<2x32x40x64xf32>) outs(%1616 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1618 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1617 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_714 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1619 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1615, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1620 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1619 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1621 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1050, %1618 : tensor<2x32x40x64xf16>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1622 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1621, %1620 : tensor<2x32x40x64xf32>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %collapsed_1051 = tensor.collapse_shape %1622 [[0], [1, 2], [3]] : tensor<2x32x40x64xf32> into tensor<2x1280x64xf32> | |
| %expanded_1052 = tensor.expand_shape %collapsed_1051 [[0], [1], [2, 3]] : tensor<2x1280x64xf32> into tensor<2x1280x8x8xf32> | |
| %expanded_1053 = tensor.expand_shape %cst_256 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1623 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1052, %expanded_1053 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %expanded_1054 = tensor.expand_shape %cst_257 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1624 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1623, %expanded_1054 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %1625 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1626 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1625 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %1627 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1626 : tensor<f16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1628 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1624 : tensor<2x1280x8x8xf32>) outs(%1627 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1629 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1628 : tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1630 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1629, %1628 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %padded_1055 = tensor.pad %1630 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x8x8xf16> to tensor<2x1280x10x10xf16> | |
| %1631 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_259 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1632 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1055, %cst_258 : tensor<2x1280x10x10xf16>, tensor<1280x1280x3x3xf16>) outs(%1631 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %1633 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1563, %1632 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1634 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1633, %cst_3 : tensor<2x1280x8x8xf16>, tensor<f64>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %collapsed_1056 = tensor.collapse_shape %1634 [[0], [1], [2, 3]] : tensor<2x1280x8x8xf16> into tensor<2x1280x64xf16> | |
| %expanded_1057 = tensor.expand_shape %collapsed_1056 [[0], [1, 2], [3]] : tensor<2x1280x64xf16> into tensor<2x32x40x64xf16> | |
| %1635 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1636 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1635 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %1637 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1636 : tensor<f32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1638 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1057 : tensor<2x32x40x64xf16>) outs(%1637 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1639 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1638 : tensor<2x32x40x64xf32>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1640 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1641 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1639 : tensor<2x32x40x64xf64>) outs(%1640 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1642 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1641 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1643 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1639, %1642 : tensor<2x32x40x64xf64>, tensor<2x32x1x1xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1644 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1643 : tensor<2x32x40x64xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1645 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1646 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1644 : tensor<2x32x40x64xf64>) outs(%1645 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1647 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1646 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1648 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1647 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1649 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %1650 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1638 : tensor<2x32x40x64xf32>) outs(%1649 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1651 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1650 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_714 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1652 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1648, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1653 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1652 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1654 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1057, %1651 : tensor<2x32x40x64xf16>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1655 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1654, %1653 : tensor<2x32x40x64xf32>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %collapsed_1058 = tensor.collapse_shape %1655 [[0], [1, 2], [3]] : tensor<2x32x40x64xf32> into tensor<2x1280x64xf32> | |
| %expanded_1059 = tensor.expand_shape %collapsed_1058 [[0], [1], [2, 3]] : tensor<2x1280x64xf32> into tensor<2x1280x8x8xf32> | |
| %expanded_1060 = tensor.expand_shape %cst_260 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1656 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1059, %expanded_1060 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %expanded_1061 = tensor.expand_shape %cst_261 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1657 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1656, %expanded_1061 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %1658 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1659 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1658 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %1660 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1659 : tensor<f16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1661 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1657 : tensor<2x1280x8x8xf32>) outs(%1660 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1662 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1661 : tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1663 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1662, %1661 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %padded_1062 = tensor.pad %1663 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x8x8xf16> to tensor<2x1280x10x10xf16> | |
| %1664 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_263 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1665 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1062, %cst_262 : tensor<2x1280x10x10xf16>, tensor<1280x1280x3x3xf16>) outs(%1664 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %1666 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280xf16> | |
| %1667 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%1666, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %1668 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_264 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %1669 = linalg.fill ins(%cst_694 : f16) outs(%30 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %1670 = linalg.matmul ins(%1667, %1668 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%1669 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %1671 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_265, %1670 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %expanded_1063 = tensor.expand_shape %1671 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16> | |
| %1672 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1665, %expanded_1063 : tensor<2x1280x8x8xf16>, tensor<2x1280x1x1xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %collapsed_1064 = tensor.collapse_shape %1672 [[0], [1], [2, 3]] : tensor<2x1280x8x8xf16> into tensor<2x1280x64xf16> | |
| %expanded_1065 = tensor.expand_shape %collapsed_1064 [[0], [1, 2], [3]] : tensor<2x1280x64xf16> into tensor<2x32x40x64xf16> | |
| %1673 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1674 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1673 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %1675 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1674 : tensor<f32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1676 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1065 : tensor<2x32x40x64xf16>) outs(%1675 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1677 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1676 : tensor<2x32x40x64xf32>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1678 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1679 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1677 : tensor<2x32x40x64xf64>) outs(%1678 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1680 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1679 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1681 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1677, %1680 : tensor<2x32x40x64xf64>, tensor<2x32x1x1xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1682 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1681 : tensor<2x32x40x64xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1683 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1684 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1682 : tensor<2x32x40x64xf64>) outs(%1683 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1685 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1684 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1686 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1685 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1687 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %1688 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1676 : tensor<2x32x40x64xf32>) outs(%1687 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1689 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1688 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_714 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1690 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1686, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1691 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1690 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1692 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1065, %1689 : tensor<2x32x40x64xf16>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1693 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1692, %1691 : tensor<2x32x40x64xf32>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %collapsed_1066 = tensor.collapse_shape %1693 [[0], [1, 2], [3]] : tensor<2x32x40x64xf32> into tensor<2x1280x64xf32> | |
| %expanded_1067 = tensor.expand_shape %collapsed_1066 [[0], [1], [2, 3]] : tensor<2x1280x64xf32> into tensor<2x1280x8x8xf32> | |
| %expanded_1068 = tensor.expand_shape %cst_266 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1694 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1067, %expanded_1068 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %expanded_1069 = tensor.expand_shape %cst_267 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1695 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1694, %expanded_1069 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %1696 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1697 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1696 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %1698 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1697 : tensor<f16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1699 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1695 : tensor<2x1280x8x8xf32>) outs(%1698 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1700 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1699 : tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1701 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1700, %1699 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %padded_1070 = tensor.pad %1701 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x8x8xf16> to tensor<2x1280x10x10xf16> | |
| %1702 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_269 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1703 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1070, %cst_268 : tensor<2x1280x10x10xf16>, tensor<1280x1280x3x3xf16>) outs(%1702 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %1704 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1634, %1703 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1705 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1704, %cst_5 : tensor<2x1280x8x8xf16>, tensor<i64>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: i64, %out: f16): | |
| %4280 = arith.sitofp %in_1640 : i64 to f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %collapsed_1071 = tensor.collapse_shape %1705 [[0], [1], [2, 3]] : tensor<2x1280x8x8xf16> into tensor<2x1280x64xf16> | |
| %expanded_1072 = tensor.expand_shape %collapsed_1071 [[0], [1, 2], [3]] : tensor<2x1280x64xf16> into tensor<2x32x40x64xf16> | |
| %1706 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1707 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1706 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %1708 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1707 : tensor<f32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1709 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1072 : tensor<2x32x40x64xf16>) outs(%1708 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1710 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1709 : tensor<2x32x40x64xf32>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1711 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1712 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1710 : tensor<2x32x40x64xf64>) outs(%1711 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1713 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1712 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1714 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1710, %1713 : tensor<2x32x40x64xf64>, tensor<2x32x1x1xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1715 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1714 : tensor<2x32x40x64xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1716 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1717 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1715 : tensor<2x32x40x64xf64>) outs(%1716 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1718 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1717 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1719 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1718 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1720 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %1721 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1709 : tensor<2x32x40x64xf32>) outs(%1720 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1722 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1721 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_714 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1723 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1719, %cst_2 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1724 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1723 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1725 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1072, %1722 : tensor<2x32x40x64xf16>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1726 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1725, %1724 : tensor<2x32x40x64xf32>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %collapsed_1073 = tensor.collapse_shape %1726 [[0], [1, 2], [3]] : tensor<2x32x40x64xf32> into tensor<2x1280x64xf32> | |
| %expanded_1074 = tensor.expand_shape %collapsed_1073 [[0], [1], [2, 3]] : tensor<2x1280x64xf32> into tensor<2x1280x8x8xf32> | |
| %expanded_1075 = tensor.expand_shape %cst_270 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1727 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1074, %expanded_1075 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %expanded_1076 = tensor.expand_shape %cst_271 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1728 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1727, %expanded_1076 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %1729 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1730 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1729 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %1731 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1730 : tensor<f16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1732 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1728 : tensor<2x1280x8x8xf32>) outs(%1731 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1733 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_273 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1734 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1732, %cst_272 : tensor<2x1280x8x8xf16>, tensor<1280x1280x1x1xf16>) outs(%1733 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %1735 = tensor.empty() : tensor<2x8x8x1280xf16> | |
| %1736 = linalg.generic {indexing_maps = [#map11, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1734 : tensor<2x1280x8x8xf16>) outs(%1735 : tensor<2x8x8x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x8x1280xf16> | |
| %collapsed_1077 = tensor.collapse_shape %1736 [[0], [1, 2], [3]] : tensor<2x8x8x1280xf16> into tensor<2x64x1280xf16> | |
| %1737 = tensor.empty() : tensor<2x64x1xf16> | |
| %1738 = linalg.fill ins(%cst_694 : f16) outs(%1737 : tensor<2x64x1xf16>) -> tensor<2x64x1xf16> | |
| %1739 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_1077 : tensor<2x64x1280xf16>) outs(%1738 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1740 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1739 : tensor<2x64x1xf16>) outs(%1737 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1741 = tensor.empty() : tensor<2x64x1280xf16> | |
| %1742 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1740 : tensor<2x64x1xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1743 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1077, %1742 : tensor<2x64x1280xf16>, tensor<2x64x1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1744 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1743 : tensor<2x64x1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1745 = linalg.fill ins(%cst_694 : f16) outs(%1737 : tensor<2x64x1xf16>) -> tensor<2x64x1xf16> | |
| %1746 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1744 : tensor<2x64x1280xf16>) outs(%1745 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1747 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1746 : tensor<2x64x1xf16>) outs(%1737 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1748 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1747 : tensor<2x64x1xf16>) outs(%1737 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1749 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1748 : tensor<2x64x1xf16>) outs(%1737 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1750 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1749 : tensor<2x64x1xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1751 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1743, %1750 : tensor<2x64x1280xf16>, tensor<2x64x1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1752 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1751, %cst_274 : tensor<2x64x1280xf16>, tensor<1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1753 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1752, %cst_275 : tensor<2x64x1280xf16>, tensor<1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1754 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_276 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1078 = tensor.collapse_shape %1753 [[0, 1], [2]] : tensor<2x64x1280xf16> into tensor<128x1280xf16> | |
| %1755 = tensor.empty() : tensor<128x1280xf16> | |
| %1756 = linalg.fill ins(%cst_694 : f16) outs(%1755 : tensor<128x1280xf16>) -> tensor<128x1280xf16> | |
| %1757 = linalg.matmul ins(%collapsed_1078, %1754 : tensor<128x1280xf16>, tensor<1280x1280xf16>) outs(%1756 : tensor<128x1280xf16>) -> tensor<128x1280xf16> | |
| %1758 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_277 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %1759 = linalg.fill ins(%cst_694 : f16) outs(%1755 : tensor<128x1280xf16>) -> tensor<128x1280xf16> | |
| %1760 = linalg.matmul ins(%collapsed_1078, %1758 : tensor<128x1280xf16>, tensor<1280x1280xf16>) outs(%1759 : tensor<128x1280xf16>) -> tensor<128x1280xf16> | |
| %1761 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_278 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %1762 = linalg.fill ins(%cst_694 : f16) outs(%1755 : tensor<128x1280xf16>) -> tensor<128x1280xf16> | |
| %1763 = linalg.matmul ins(%collapsed_1078, %1761 : tensor<128x1280xf16>, tensor<1280x1280xf16>) outs(%1762 : tensor<128x1280xf16>) -> tensor<128x1280xf16> | |
| %expanded_1079 = tensor.expand_shape %1757 [[0, 1], [2, 3]] : tensor<128x1280xf16> into tensor<2x64x8x160xf16> | |
| %1764 = tensor.empty() : tensor<2x8x64x160xf16> | |
| %1765 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1079 : tensor<2x64x8x160xf16>) outs(%1764 : tensor<2x8x64x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x64x160xf16> | |
| %collapsed_1080 = tensor.collapse_shape %1765 [[0, 1], [2], [3]] : tensor<2x8x64x160xf16> into tensor<16x64x160xf16> | |
| %expanded_1081 = tensor.expand_shape %1760 [[0, 1], [2, 3]] : tensor<128x1280xf16> into tensor<2x64x8x160xf16> | |
| %1766 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1081 : tensor<2x64x8x160xf16>) outs(%1764 : tensor<2x8x64x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x64x160xf16> | |
| %collapsed_1082 = tensor.collapse_shape %1766 [[0, 1], [2], [3]] : tensor<2x8x64x160xf16> into tensor<16x64x160xf16> | |
| %expanded_1083 = tensor.expand_shape %1763 [[0, 1], [2, 3]] : tensor<128x1280xf16> into tensor<2x64x8x160xf16> | |
| %1767 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1083 : tensor<2x64x8x160xf16>) outs(%1764 : tensor<2x8x64x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x64x160xf16> | |
| %collapsed_1084 = tensor.collapse_shape %1767 [[0, 1], [2], [3]] : tensor<2x8x64x160xf16> into tensor<16x64x160xf16> | |
| %1768 = tensor.empty() : tensor<16x160x64xf16> | |
| %1769 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1082 : tensor<16x64x160xf16>) outs(%1768 : tensor<16x160x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x160x64xf16> | |
| %1770 = tensor.empty() : tensor<16x64x64xf16> | |
| %1771 = linalg.fill ins(%cst_694 : f16) outs(%1770 : tensor<16x64x64xf16>) -> tensor<16x64x64xf16> | |
| %1772 = linalg.batch_matmul ins(%collapsed_1080, %1769 : tensor<16x64x160xf16>, tensor<16x160x64xf16>) outs(%1771 : tensor<16x64x64xf16>) -> tensor<16x64x64xf16> | |
| %1773 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1772, %cst : tensor<16x64x64xf16>, tensor<f64>) outs(%1770 : tensor<16x64x64xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x64x64xf16> | |
| %1774 = tensor.empty() : tensor<16x64x1xi64> | |
| %1775 = linalg.fill ins(%c0_i64 : i64) outs(%1774 : tensor<16x64x1xi64>) -> tensor<16x64x1xi64> | |
| %1776 = tensor.empty() : tensor<16x64x1xf16> | |
| %1777 = linalg.fill ins(%cst_696 : f16) outs(%1776 : tensor<16x64x1xf16>) -> tensor<16x64x1xf16> | |
| %1778:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1773 : tensor<16x64x64xf16>) outs(%1777, %1775 : tensor<16x64x1xf16>, tensor<16x64x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x64x1xf16>, tensor<16x64x1xi64>) | |
| %1779 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1773, %1778#0 : tensor<16x64x64xf16>, tensor<16x64x1xf16>) outs(%1770 : tensor<16x64x64xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x64x64xf16> | |
| %1780 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1779 : tensor<16x64x64xf16>) outs(%1770 : tensor<16x64x64xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x64x64xf16> | |
| %1781 = linalg.fill ins(%cst_694 : f16) outs(%1776 : tensor<16x64x1xf16>) -> tensor<16x64x1xf16> | |
| %1782 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1780 : tensor<16x64x64xf16>) outs(%1781 : tensor<16x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x64x1xf16> | |
| %1783 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1780, %1782 : tensor<16x64x64xf16>, tensor<16x64x1xf16>) outs(%1770 : tensor<16x64x64xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x64x64xf16> | |
| %1784 = tensor.empty() : tensor<16x64x160xf16> | |
| %1785 = linalg.fill ins(%cst_694 : f16) outs(%1784 : tensor<16x64x160xf16>) -> tensor<16x64x160xf16> | |
| %1786 = linalg.batch_matmul ins(%1783, %collapsed_1084 : tensor<16x64x64xf16>, tensor<16x64x160xf16>) outs(%1785 : tensor<16x64x160xf16>) -> tensor<16x64x160xf16> | |
| %expanded_1085 = tensor.expand_shape %1786 [[0, 1], [2], [3]] : tensor<16x64x160xf16> into tensor<2x8x64x160xf16> | |
| %1787 = tensor.empty() : tensor<2x64x8x160xf16> | |
| %1788 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1085 : tensor<2x8x64x160xf16>) outs(%1787 : tensor<2x64x8x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x64x8x160xf16> | |
| %1789 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_279 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1086 = tensor.collapse_shape %1788 [[0, 1], [2, 3]] : tensor<2x64x8x160xf16> into tensor<128x1280xf16> | |
| %1790 = linalg.fill ins(%cst_694 : f16) outs(%1755 : tensor<128x1280xf16>) -> tensor<128x1280xf16> | |
| %1791 = linalg.matmul ins(%collapsed_1086, %1789 : tensor<128x1280xf16>, tensor<1280x1280xf16>) outs(%1790 : tensor<128x1280xf16>) -> tensor<128x1280xf16> | |
| %1792 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_280, %1791 : tensor<1280xf16>, tensor<128x1280xf16>) outs(%1755 : tensor<128x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<128x1280xf16> | |
| %expanded_1087 = tensor.expand_shape %1792 [[0, 1], [2]] : tensor<128x1280xf16> into tensor<2x64x1280xf16> | |
| %1793 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1087, %collapsed_1077 : tensor<2x64x1280xf16>, tensor<2x64x1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1794 = linalg.fill ins(%cst_694 : f16) outs(%1737 : tensor<2x64x1xf16>) -> tensor<2x64x1xf16> | |
| %1795 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1793 : tensor<2x64x1280xf16>) outs(%1794 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1796 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1795 : tensor<2x64x1xf16>) outs(%1737 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1797 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1796 : tensor<2x64x1xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1798 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1793, %1797 : tensor<2x64x1280xf16>, tensor<2x64x1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1799 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1798 : tensor<2x64x1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1800 = linalg.fill ins(%cst_694 : f16) outs(%1737 : tensor<2x64x1xf16>) -> tensor<2x64x1xf16> | |
| %1801 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1799 : tensor<2x64x1280xf16>) outs(%1800 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1802 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1801 : tensor<2x64x1xf16>) outs(%1737 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1803 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1802 : tensor<2x64x1xf16>) outs(%1737 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1804 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1803 : tensor<2x64x1xf16>) outs(%1737 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1805 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1804 : tensor<2x64x1xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1806 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1798, %1805 : tensor<2x64x1280xf16>, tensor<2x64x1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1807 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1806, %cst_281 : tensor<2x64x1280xf16>, tensor<1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1808 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1807, %cst_282 : tensor<2x64x1280xf16>, tensor<1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1809 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_283 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1088 = tensor.collapse_shape %1808 [[0, 1], [2]] : tensor<2x64x1280xf16> into tensor<128x1280xf16> | |
| %1810 = linalg.fill ins(%cst_694 : f16) outs(%1755 : tensor<128x1280xf16>) -> tensor<128x1280xf16> | |
| %1811 = linalg.matmul ins(%collapsed_1088, %1809 : tensor<128x1280xf16>, tensor<1280x1280xf16>) outs(%1810 : tensor<128x1280xf16>) -> tensor<128x1280xf16> | |
| %1812 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_284 : tensor<1280x768xf16>) outs(%1194 : tensor<768x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x1280xf16> | |
| %1813 = linalg.fill ins(%cst_694 : f16) outs(%1196 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %1814 = linalg.matmul ins(%collapsed_761, %1812 : tensor<154x768xf16>, tensor<768x1280xf16>) outs(%1813 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %1815 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_285 : tensor<1280x768xf16>) outs(%1194 : tensor<768x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x1280xf16> | |
| %1816 = linalg.fill ins(%cst_694 : f16) outs(%1196 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %1817 = linalg.matmul ins(%collapsed_761, %1815 : tensor<154x768xf16>, tensor<768x1280xf16>) outs(%1816 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %expanded_1089 = tensor.expand_shape %1811 [[0, 1], [2, 3]] : tensor<128x1280xf16> into tensor<2x64x8x160xf16> | |
| %1818 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1089 : tensor<2x64x8x160xf16>) outs(%1764 : tensor<2x8x64x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x64x160xf16> | |
| %collapsed_1090 = tensor.collapse_shape %1818 [[0, 1], [2], [3]] : tensor<2x8x64x160xf16> into tensor<16x64x160xf16> | |
| %expanded_1091 = tensor.expand_shape %1814 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x8x160xf16> | |
| %1819 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1091 : tensor<2x77x8x160xf16>) outs(%1203 : tensor<2x8x77x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x160xf16> | |
| %collapsed_1092 = tensor.collapse_shape %1819 [[0, 1], [2], [3]] : tensor<2x8x77x160xf16> into tensor<16x77x160xf16> | |
| %expanded_1093 = tensor.expand_shape %1817 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x8x160xf16> | |
| %1820 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1093 : tensor<2x77x8x160xf16>) outs(%1203 : tensor<2x8x77x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x160xf16> | |
| %collapsed_1094 = tensor.collapse_shape %1820 [[0, 1], [2], [3]] : tensor<2x8x77x160xf16> into tensor<16x77x160xf16> | |
| %1821 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1092 : tensor<16x77x160xf16>) outs(%1206 : tensor<16x160x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x160x77xf16> | |
| %1822 = tensor.empty() : tensor<16x64x77xf16> | |
| %1823 = linalg.fill ins(%cst_694 : f16) outs(%1822 : tensor<16x64x77xf16>) -> tensor<16x64x77xf16> | |
| %1824 = linalg.batch_matmul ins(%collapsed_1090, %1821 : tensor<16x64x160xf16>, tensor<16x160x77xf16>) outs(%1823 : tensor<16x64x77xf16>) -> tensor<16x64x77xf16> | |
| %1825 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1824, %cst : tensor<16x64x77xf16>, tensor<f64>) outs(%1822 : tensor<16x64x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x64x77xf16> | |
| %1826 = linalg.fill ins(%c0_i64 : i64) outs(%1774 : tensor<16x64x1xi64>) -> tensor<16x64x1xi64> | |
| %1827 = linalg.fill ins(%cst_696 : f16) outs(%1776 : tensor<16x64x1xf16>) -> tensor<16x64x1xf16> | |
| %1828:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1825 : tensor<16x64x77xf16>) outs(%1827, %1826 : tensor<16x64x1xf16>, tensor<16x64x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x64x1xf16>, tensor<16x64x1xi64>) | |
| %1829 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1825, %1828#0 : tensor<16x64x77xf16>, tensor<16x64x1xf16>) outs(%1822 : tensor<16x64x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x64x77xf16> | |
| %1830 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1829 : tensor<16x64x77xf16>) outs(%1822 : tensor<16x64x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x64x77xf16> | |
| %1831 = linalg.fill ins(%cst_694 : f16) outs(%1776 : tensor<16x64x1xf16>) -> tensor<16x64x1xf16> | |
| %1832 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1830 : tensor<16x64x77xf16>) outs(%1831 : tensor<16x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x64x1xf16> | |
| %1833 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1830, %1832 : tensor<16x64x77xf16>, tensor<16x64x1xf16>) outs(%1822 : tensor<16x64x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x64x77xf16> | |
| %1834 = linalg.fill ins(%cst_694 : f16) outs(%1784 : tensor<16x64x160xf16>) -> tensor<16x64x160xf16> | |
| %1835 = linalg.batch_matmul ins(%1833, %collapsed_1094 : tensor<16x64x77xf16>, tensor<16x77x160xf16>) outs(%1834 : tensor<16x64x160xf16>) -> tensor<16x64x160xf16> | |
| %expanded_1095 = tensor.expand_shape %1835 [[0, 1], [2], [3]] : tensor<16x64x160xf16> into tensor<2x8x64x160xf16> | |
| %1836 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1095 : tensor<2x8x64x160xf16>) outs(%1787 : tensor<2x64x8x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x64x8x160xf16> | |
| %1837 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_286 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1096 = tensor.collapse_shape %1836 [[0, 1], [2, 3]] : tensor<2x64x8x160xf16> into tensor<128x1280xf16> | |
| %1838 = linalg.fill ins(%cst_694 : f16) outs(%1755 : tensor<128x1280xf16>) -> tensor<128x1280xf16> | |
| %1839 = linalg.matmul ins(%collapsed_1096, %1837 : tensor<128x1280xf16>, tensor<1280x1280xf16>) outs(%1838 : tensor<128x1280xf16>) -> tensor<128x1280xf16> | |
| %1840 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_287, %1839 : tensor<1280xf16>, tensor<128x1280xf16>) outs(%1755 : tensor<128x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<128x1280xf16> | |
| %expanded_1097 = tensor.expand_shape %1840 [[0, 1], [2]] : tensor<128x1280xf16> into tensor<2x64x1280xf16> | |
| %1841 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1097, %1793 : tensor<2x64x1280xf16>, tensor<2x64x1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1842 = linalg.fill ins(%cst_694 : f16) outs(%1737 : tensor<2x64x1xf16>) -> tensor<2x64x1xf16> | |
| %1843 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1841 : tensor<2x64x1280xf16>) outs(%1842 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1844 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1843 : tensor<2x64x1xf16>) outs(%1737 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1845 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1844 : tensor<2x64x1xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1846 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1841, %1845 : tensor<2x64x1280xf16>, tensor<2x64x1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1847 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1846 : tensor<2x64x1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1848 = linalg.fill ins(%cst_694 : f16) outs(%1737 : tensor<2x64x1xf16>) -> tensor<2x64x1xf16> | |
| %1849 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1847 : tensor<2x64x1280xf16>) outs(%1848 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1850 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1849 : tensor<2x64x1xf16>) outs(%1737 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1851 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1850 : tensor<2x64x1xf16>) outs(%1737 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1852 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1851 : tensor<2x64x1xf16>) outs(%1737 : tensor<2x64x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1xf16> | |
| %1853 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1852 : tensor<2x64x1xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1854 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1846, %1853 : tensor<2x64x1280xf16>, tensor<2x64x1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1855 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1854, %cst_288 : tensor<2x64x1280xf16>, tensor<1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1856 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1855, %cst_289 : tensor<2x64x1280xf16>, tensor<1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %1857 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_290 : tensor<10240x1280xf16>) outs(%1243 : tensor<1280x10240xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x10240xf16> | |
| %collapsed_1098 = tensor.collapse_shape %1856 [[0, 1], [2]] : tensor<2x64x1280xf16> into tensor<128x1280xf16> | |
| %1858 = tensor.empty() : tensor<128x10240xf16> | |
| %1859 = linalg.fill ins(%cst_694 : f16) outs(%1858 : tensor<128x10240xf16>) -> tensor<128x10240xf16> | |
| %1860 = linalg.matmul ins(%collapsed_1098, %1857 : tensor<128x1280xf16>, tensor<1280x10240xf16>) outs(%1859 : tensor<128x10240xf16>) -> tensor<128x10240xf16> | |
| %1861 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_291, %1860 : tensor<10240xf16>, tensor<128x10240xf16>) outs(%1858 : tensor<128x10240xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<128x10240xf16> | |
| %expanded_1099 = tensor.expand_shape %1861 [[0, 1], [2]] : tensor<128x10240xf16> into tensor<2x64x10240xf16> | |
| %extracted_slice_1100 = tensor.extract_slice %expanded_1099[0, 0, 0] [2, 64, 5120] [1, 1, 1] : tensor<2x64x10240xf16> to tensor<2x64x5120xf16> | |
| %extracted_slice_1101 = tensor.extract_slice %expanded_1099[0, 0, 5120] [2, 64, 5120] [1, 1, 1] : tensor<2x64x10240xf16> to tensor<2x64x5120xf16> | |
| %1862 = tensor.empty() : tensor<2x64x5120xf16> | |
| %1863 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1101 : tensor<2x64x5120xf16>) outs(%1862 : tensor<2x64x5120xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.sqrt %cst_697 : f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| %4282 = math.erf %4281 : f16 | |
| %4283 = arith.addf %4282, %cst_695 : f16 | |
| %4284 = arith.mulf %4283, %cst_698 : f16 | |
| %4285 = arith.mulf %in, %4284 : f16 | |
| linalg.yield %4285 : f16 | |
| } -> tensor<2x64x5120xf16> | |
| %1864 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1100, %1863 : tensor<2x64x5120xf16>, tensor<2x64x5120xf16>) outs(%1862 : tensor<2x64x5120xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x5120xf16> | |
| %1865 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_292 : tensor<1280x5120xf16>) outs(%1252 : tensor<5120x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<5120x1280xf16> | |
| %collapsed_1102 = tensor.collapse_shape %1864 [[0, 1], [2]] : tensor<2x64x5120xf16> into tensor<128x5120xf16> | |
| %1866 = linalg.fill ins(%cst_694 : f16) outs(%1755 : tensor<128x1280xf16>) -> tensor<128x1280xf16> | |
| %1867 = linalg.matmul ins(%collapsed_1102, %1865 : tensor<128x5120xf16>, tensor<5120x1280xf16>) outs(%1866 : tensor<128x1280xf16>) -> tensor<128x1280xf16> | |
| %1868 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_293, %1867 : tensor<1280xf16>, tensor<128x1280xf16>) outs(%1755 : tensor<128x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<128x1280xf16> | |
| %expanded_1103 = tensor.expand_shape %1868 [[0, 1], [2]] : tensor<128x1280xf16> into tensor<2x64x1280xf16> | |
| %1869 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1103, %1841 : tensor<2x64x1280xf16>, tensor<2x64x1280xf16>) outs(%1741 : tensor<2x64x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x64x1280xf16> | |
| %expanded_1104 = tensor.expand_shape %1869 [[0], [1, 2], [3]] : tensor<2x64x1280xf16> into tensor<2x8x8x1280xf16> | |
| %1870 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1104 : tensor<2x8x8x1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1871 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_295 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1872 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1870, %cst_294 : tensor<2x1280x8x8xf16>, tensor<1280x1280x1x1xf16>) outs(%1871 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %1873 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1872, %1705 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %collapsed_1105 = tensor.collapse_shape %1873 [[0], [1], [2, 3]] : tensor<2x1280x8x8xf16> into tensor<2x1280x64xf16> | |
| %expanded_1106 = tensor.expand_shape %collapsed_1105 [[0], [1, 2], [3]] : tensor<2x1280x64xf16> into tensor<2x32x40x64xf16> | |
| %1874 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1875 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1874 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %1876 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1875 : tensor<f32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1877 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1106 : tensor<2x32x40x64xf16>) outs(%1876 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1878 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1877 : tensor<2x32x40x64xf32>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1879 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1880 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1878 : tensor<2x32x40x64xf64>) outs(%1879 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1881 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1880 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1882 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1878, %1881 : tensor<2x32x40x64xf64>, tensor<2x32x1x1xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1883 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1882 : tensor<2x32x40x64xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1884 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1885 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1883 : tensor<2x32x40x64xf64>) outs(%1884 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1886 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1885 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1887 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1886 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1888 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %1889 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1877 : tensor<2x32x40x64xf32>) outs(%1888 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1890 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1889 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_714 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1891 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1887, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1892 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1891 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1893 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1106, %1890 : tensor<2x32x40x64xf16>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1894 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1893, %1892 : tensor<2x32x40x64xf32>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %collapsed_1107 = tensor.collapse_shape %1894 [[0], [1, 2], [3]] : tensor<2x32x40x64xf32> into tensor<2x1280x64xf32> | |
| %expanded_1108 = tensor.expand_shape %collapsed_1107 [[0], [1], [2, 3]] : tensor<2x1280x64xf32> into tensor<2x1280x8x8xf32> | |
| %expanded_1109 = tensor.expand_shape %cst_296 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1895 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1108, %expanded_1109 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %expanded_1110 = tensor.expand_shape %cst_297 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1896 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1895, %expanded_1110 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %1897 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1898 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1897 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %1899 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1898 : tensor<f16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1900 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1896 : tensor<2x1280x8x8xf32>) outs(%1899 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1901 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1900 : tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1902 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1901, %1900 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %padded_1111 = tensor.pad %1902 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x8x8xf16> to tensor<2x1280x10x10xf16> | |
| %1903 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_299 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1904 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1111, %cst_298 : tensor<2x1280x10x10xf16>, tensor<1280x1280x3x3xf16>) outs(%1903 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %1905 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280xf16> | |
| %1906 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%1905, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %1907 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_300 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %1908 = linalg.fill ins(%cst_694 : f16) outs(%30 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %1909 = linalg.matmul ins(%1906, %1907 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%1908 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %1910 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_301, %1909 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %expanded_1112 = tensor.expand_shape %1910 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16> | |
| %1911 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1904, %expanded_1112 : tensor<2x1280x8x8xf16>, tensor<2x1280x1x1xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %collapsed_1113 = tensor.collapse_shape %1911 [[0], [1], [2, 3]] : tensor<2x1280x8x8xf16> into tensor<2x1280x64xf16> | |
| %expanded_1114 = tensor.expand_shape %collapsed_1113 [[0], [1, 2], [3]] : tensor<2x1280x64xf16> into tensor<2x32x40x64xf16> | |
| %1912 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1913 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1912 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %1914 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1913 : tensor<f32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1915 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1114 : tensor<2x32x40x64xf16>) outs(%1914 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1916 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1915 : tensor<2x32x40x64xf32>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1917 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1918 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1916 : tensor<2x32x40x64xf64>) outs(%1917 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1919 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1918 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1920 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1916, %1919 : tensor<2x32x40x64xf64>, tensor<2x32x1x1xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1921 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1920 : tensor<2x32x40x64xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1922 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1923 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1921 : tensor<2x32x40x64xf64>) outs(%1922 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1924 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1923 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1925 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1924 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1926 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %1927 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1915 : tensor<2x32x40x64xf32>) outs(%1926 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1928 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1927 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_714 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1929 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1925, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1930 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1929 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1931 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1114, %1928 : tensor<2x32x40x64xf16>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1932 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1931, %1930 : tensor<2x32x40x64xf32>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %collapsed_1115 = tensor.collapse_shape %1932 [[0], [1, 2], [3]] : tensor<2x32x40x64xf32> into tensor<2x1280x64xf32> | |
| %expanded_1116 = tensor.expand_shape %collapsed_1115 [[0], [1], [2, 3]] : tensor<2x1280x64xf32> into tensor<2x1280x8x8xf32> | |
| %expanded_1117 = tensor.expand_shape %cst_302 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1933 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1116, %expanded_1117 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %expanded_1118 = tensor.expand_shape %cst_303 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %1934 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1933, %expanded_1118 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %1935 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1936 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1935 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %1937 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1936 : tensor<f16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1938 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1934 : tensor<2x1280x8x8xf32>) outs(%1937 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1939 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1938 : tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1940 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1939, %1938 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %padded_1119 = tensor.pad %1940 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x8x8xf16> to tensor<2x1280x10x10xf16> | |
| %1941 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_305 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1942 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1119, %cst_304 : tensor<2x1280x10x10xf16>, tensor<1280x1280x3x3xf16>) outs(%1941 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %1943 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1873, %1942 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1944 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1943, %cst_5 : tensor<2x1280x8x8xf16>, tensor<i64>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: i64, %out: f16): | |
| %4280 = arith.sitofp %in_1640 : i64 to f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1945 = tensor.empty() : tensor<2x2560x8x8xf16> | |
| %inserted_slice_1120 = tensor.insert_slice %1944 into %1945[0, 0, 0, 0] [2, 1280, 8, 8] [1, 1, 1, 1] : tensor<2x1280x8x8xf16> into tensor<2x2560x8x8xf16> | |
| %inserted_slice_1121 = tensor.insert_slice %1634 into %inserted_slice_1120[0, 1280, 0, 0] [2, 1280, 8, 8] [1, 1, 1, 1] : tensor<2x1280x8x8xf16> into tensor<2x2560x8x8xf16> | |
| %collapsed_1122 = tensor.collapse_shape %inserted_slice_1121 [[0], [1], [2, 3]] : tensor<2x2560x8x8xf16> into tensor<2x2560x64xf16> | |
| %expanded_1123 = tensor.expand_shape %collapsed_1122 [[0], [1, 2], [3]] : tensor<2x2560x64xf16> into tensor<2x32x80x64xf16> | |
| %1946 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1947 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1946 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %1948 = tensor.empty() : tensor<2x32x80x64xf32> | |
| %1949 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1947 : tensor<f32>) outs(%1948 : tensor<2x32x80x64xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x80x64xf32> | |
| %1950 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1123 : tensor<2x32x80x64xf16>) outs(%1949 : tensor<2x32x80x64xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x80x64xf32> | |
| %1951 = tensor.empty() : tensor<2x32x80x64xf64> | |
| %1952 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1950 : tensor<2x32x80x64xf32>) outs(%1951 : tensor<2x32x80x64xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x80x64xf64> | |
| %1953 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1954 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1952 : tensor<2x32x80x64xf64>) outs(%1953 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1955 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1954 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_710 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1956 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1952, %1955 : tensor<2x32x80x64xf64>, tensor<2x32x1x1xf64>) outs(%1951 : tensor<2x32x80x64xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x80x64xf64> | |
| %1957 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1956 : tensor<2x32x80x64xf64>) outs(%1951 : tensor<2x32x80x64xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x80x64xf64> | |
| %1958 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1959 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1957 : tensor<2x32x80x64xf64>) outs(%1958 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1960 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1959 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_710 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1961 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1960 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1962 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %1963 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1950 : tensor<2x32x80x64xf32>) outs(%1962 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1964 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1963 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_711 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1965 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1961, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1966 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1965 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %1967 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1123, %1964 : tensor<2x32x80x64xf16>, tensor<2x32x1x1xf32>) outs(%1948 : tensor<2x32x80x64xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x80x64xf32> | |
| %1968 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1967, %1966 : tensor<2x32x80x64xf32>, tensor<2x32x1x1xf32>) outs(%1948 : tensor<2x32x80x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x80x64xf32> | |
| %collapsed_1124 = tensor.collapse_shape %1968 [[0], [1, 2], [3]] : tensor<2x32x80x64xf32> into tensor<2x2560x64xf32> | |
| %expanded_1125 = tensor.expand_shape %collapsed_1124 [[0], [1], [2, 3]] : tensor<2x2560x64xf32> into tensor<2x2560x8x8xf32> | |
| %expanded_1126 = tensor.expand_shape %cst_306 [[0, 1, 2]] : tensor<2560xf16> into tensor<2560x1x1xf16> | |
| %1969 = tensor.empty() : tensor<2x2560x8x8xf32> | |
| %1970 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1125, %expanded_1126 : tensor<2x2560x8x8xf32>, tensor<2560x1x1xf16>) outs(%1969 : tensor<2x2560x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x2560x8x8xf32> | |
| %expanded_1127 = tensor.expand_shape %cst_307 [[0, 1, 2]] : tensor<2560xf16> into tensor<2560x1x1xf16> | |
| %1971 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1970, %expanded_1127 : tensor<2x2560x8x8xf32>, tensor<2560x1x1xf16>) outs(%1969 : tensor<2x2560x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x2560x8x8xf32> | |
| %1972 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1973 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1972 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %1974 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1973 : tensor<f16>) outs(%1945 : tensor<2x2560x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x2560x8x8xf16> | |
| %1975 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1971 : tensor<2x2560x8x8xf32>) outs(%1974 : tensor<2x2560x8x8xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x2560x8x8xf16> | |
| %1976 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1975 : tensor<2x2560x8x8xf16>) outs(%1945 : tensor<2x2560x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x2560x8x8xf16> | |
| %1977 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1976, %1975 : tensor<2x2560x8x8xf16>, tensor<2x2560x8x8xf16>) outs(%1945 : tensor<2x2560x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x2560x8x8xf16> | |
| %padded_1128 = tensor.pad %1977 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x2560x8x8xf16> to tensor<2x2560x10x10xf16> | |
| %1978 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_309 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %1979 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1128, %cst_308 : tensor<2x2560x10x10xf16>, tensor<1280x2560x3x3xf16>) outs(%1978 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %1980 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280xf16> | |
| %1981 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%1980, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %1982 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_310 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %1983 = linalg.fill ins(%cst_694 : f16) outs(%30 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %1984 = linalg.matmul ins(%1981, %1982 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%1983 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %1985 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_311, %1984 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %expanded_1129 = tensor.expand_shape %1985 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16> | |
| %1986 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1979, %expanded_1129 : tensor<2x1280x8x8xf16>, tensor<2x1280x1x1xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %collapsed_1130 = tensor.collapse_shape %1986 [[0], [1], [2, 3]] : tensor<2x1280x8x8xf16> into tensor<2x1280x64xf16> | |
| %expanded_1131 = tensor.expand_shape %collapsed_1130 [[0], [1, 2], [3]] : tensor<2x1280x64xf16> into tensor<2x32x40x64xf16> | |
| %1987 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %1988 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1987 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %1989 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1988 : tensor<f32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1990 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1131 : tensor<2x32x40x64xf16>) outs(%1989 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %1991 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1990 : tensor<2x32x40x64xf32>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1992 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1993 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1991 : tensor<2x32x40x64xf64>) outs(%1992 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1994 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1993 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1995 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1991, %1994 : tensor<2x32x40x64xf64>, tensor<2x32x1x1xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1996 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1995 : tensor<2x32x40x64xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %1997 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %1998 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1996 : tensor<2x32x40x64xf64>) outs(%1997 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %1999 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1998 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2000 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1999 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2001 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %2002 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1990 : tensor<2x32x40x64xf32>) outs(%2001 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2003 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2002 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_714 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2004 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2000, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2005 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2004 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2006 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1131, %2003 : tensor<2x32x40x64xf16>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %2007 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2006, %2005 : tensor<2x32x40x64xf32>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %collapsed_1132 = tensor.collapse_shape %2007 [[0], [1, 2], [3]] : tensor<2x32x40x64xf32> into tensor<2x1280x64xf32> | |
| %expanded_1133 = tensor.expand_shape %collapsed_1132 [[0], [1], [2, 3]] : tensor<2x1280x64xf32> into tensor<2x1280x8x8xf32> | |
| %expanded_1134 = tensor.expand_shape %cst_312 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2008 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1133, %expanded_1134 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %expanded_1135 = tensor.expand_shape %cst_313 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2009 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2008, %expanded_1135 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %2010 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2011 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2010 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %2012 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2011 : tensor<f16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2013 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2009 : tensor<2x1280x8x8xf32>) outs(%2012 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2014 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2013 : tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2015 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2014, %2013 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %padded_1136 = tensor.pad %2015 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x8x8xf16> to tensor<2x1280x10x10xf16> | |
| %2016 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_315 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2017 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1136, %cst_314 : tensor<2x1280x10x10xf16>, tensor<1280x1280x3x3xf16>) outs(%2016 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %2018 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_317 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2019 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%inserted_slice_1121, %cst_316 : tensor<2x2560x8x8xf16>, tensor<1280x2560x1x1xf16>) outs(%2018 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %2020 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2019, %2017 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2021 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2020, %cst_3 : tensor<2x1280x8x8xf16>, tensor<f64>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %inserted_slice_1137 = tensor.insert_slice %2021 into %1945[0, 0, 0, 0] [2, 1280, 8, 8] [1, 1, 1, 1] : tensor<2x1280x8x8xf16> into tensor<2x2560x8x8xf16> | |
| %inserted_slice_1138 = tensor.insert_slice %1563 into %inserted_slice_1137[0, 1280, 0, 0] [2, 1280, 8, 8] [1, 1, 1, 1] : tensor<2x1280x8x8xf16> into tensor<2x2560x8x8xf16> | |
| %collapsed_1139 = tensor.collapse_shape %inserted_slice_1138 [[0], [1], [2, 3]] : tensor<2x2560x8x8xf16> into tensor<2x2560x64xf16> | |
| %expanded_1140 = tensor.expand_shape %collapsed_1139 [[0], [1, 2], [3]] : tensor<2x2560x64xf16> into tensor<2x32x80x64xf16> | |
| %2022 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2023 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2022 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %2024 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2023 : tensor<f32>) outs(%1948 : tensor<2x32x80x64xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x80x64xf32> | |
| %2025 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1140 : tensor<2x32x80x64xf16>) outs(%2024 : tensor<2x32x80x64xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x80x64xf32> | |
| %2026 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2025 : tensor<2x32x80x64xf32>) outs(%1951 : tensor<2x32x80x64xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x80x64xf64> | |
| %2027 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2028 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2026 : tensor<2x32x80x64xf64>) outs(%2027 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2029 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2028 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_710 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2030 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2026, %2029 : tensor<2x32x80x64xf64>, tensor<2x32x1x1xf64>) outs(%1951 : tensor<2x32x80x64xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x80x64xf64> | |
| %2031 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2030 : tensor<2x32x80x64xf64>) outs(%1951 : tensor<2x32x80x64xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x80x64xf64> | |
| %2032 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2033 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2031 : tensor<2x32x80x64xf64>) outs(%2032 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2034 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2033 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_710 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2035 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2034 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2036 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %2037 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2025 : tensor<2x32x80x64xf32>) outs(%2036 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2038 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2037 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_711 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2039 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2035, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2040 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2039 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2041 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1140, %2038 : tensor<2x32x80x64xf16>, tensor<2x32x1x1xf32>) outs(%1948 : tensor<2x32x80x64xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x80x64xf32> | |
| %2042 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2041, %2040 : tensor<2x32x80x64xf32>, tensor<2x32x1x1xf32>) outs(%1948 : tensor<2x32x80x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x80x64xf32> | |
| %collapsed_1141 = tensor.collapse_shape %2042 [[0], [1, 2], [3]] : tensor<2x32x80x64xf32> into tensor<2x2560x64xf32> | |
| %expanded_1142 = tensor.expand_shape %collapsed_1141 [[0], [1], [2, 3]] : tensor<2x2560x64xf32> into tensor<2x2560x8x8xf32> | |
| %expanded_1143 = tensor.expand_shape %cst_318 [[0, 1, 2]] : tensor<2560xf16> into tensor<2560x1x1xf16> | |
| %2043 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1142, %expanded_1143 : tensor<2x2560x8x8xf32>, tensor<2560x1x1xf16>) outs(%1969 : tensor<2x2560x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x2560x8x8xf32> | |
| %expanded_1144 = tensor.expand_shape %cst_319 [[0, 1, 2]] : tensor<2560xf16> into tensor<2560x1x1xf16> | |
| %2044 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2043, %expanded_1144 : tensor<2x2560x8x8xf32>, tensor<2560x1x1xf16>) outs(%1969 : tensor<2x2560x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x2560x8x8xf32> | |
| %2045 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2046 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2045 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %2047 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2046 : tensor<f16>) outs(%1945 : tensor<2x2560x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x2560x8x8xf16> | |
| %2048 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2044 : tensor<2x2560x8x8xf32>) outs(%2047 : tensor<2x2560x8x8xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x2560x8x8xf16> | |
| %2049 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2048 : tensor<2x2560x8x8xf16>) outs(%1945 : tensor<2x2560x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x2560x8x8xf16> | |
| %2050 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2049, %2048 : tensor<2x2560x8x8xf16>, tensor<2x2560x8x8xf16>) outs(%1945 : tensor<2x2560x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x2560x8x8xf16> | |
| %padded_1145 = tensor.pad %2050 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x2560x8x8xf16> to tensor<2x2560x10x10xf16> | |
| %2051 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_321 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2052 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1145, %cst_320 : tensor<2x2560x10x10xf16>, tensor<1280x2560x3x3xf16>) outs(%2051 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %2053 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280xf16> | |
| %2054 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%2053, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %2055 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_322 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %2056 = linalg.fill ins(%cst_694 : f16) outs(%30 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %2057 = linalg.matmul ins(%2054, %2055 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%2056 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %2058 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_323, %2057 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %expanded_1146 = tensor.expand_shape %2058 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16> | |
| %2059 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2052, %expanded_1146 : tensor<2x1280x8x8xf16>, tensor<2x1280x1x1xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %collapsed_1147 = tensor.collapse_shape %2059 [[0], [1], [2, 3]] : tensor<2x1280x8x8xf16> into tensor<2x1280x64xf16> | |
| %expanded_1148 = tensor.expand_shape %collapsed_1147 [[0], [1, 2], [3]] : tensor<2x1280x64xf16> into tensor<2x32x40x64xf16> | |
| %2060 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2061 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2060 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %2062 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2061 : tensor<f32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %2063 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1148 : tensor<2x32x40x64xf16>) outs(%2062 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %2064 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2063 : tensor<2x32x40x64xf32>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %2065 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2066 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2064 : tensor<2x32x40x64xf64>) outs(%2065 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2067 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2066 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2068 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2064, %2067 : tensor<2x32x40x64xf64>, tensor<2x32x1x1xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %2069 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2068 : tensor<2x32x40x64xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %2070 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2071 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2069 : tensor<2x32x40x64xf64>) outs(%2070 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2072 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2071 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2073 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2072 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2074 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %2075 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2063 : tensor<2x32x40x64xf32>) outs(%2074 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2076 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2075 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_714 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2077 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2073, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2078 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2077 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2079 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1148, %2076 : tensor<2x32x40x64xf16>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %2080 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2079, %2078 : tensor<2x32x40x64xf32>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %collapsed_1149 = tensor.collapse_shape %2080 [[0], [1, 2], [3]] : tensor<2x32x40x64xf32> into tensor<2x1280x64xf32> | |
| %expanded_1150 = tensor.expand_shape %collapsed_1149 [[0], [1], [2, 3]] : tensor<2x1280x64xf32> into tensor<2x1280x8x8xf32> | |
| %expanded_1151 = tensor.expand_shape %cst_324 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2081 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1150, %expanded_1151 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %expanded_1152 = tensor.expand_shape %cst_325 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2082 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2081, %expanded_1152 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %2083 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2084 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2083 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %2085 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2084 : tensor<f16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2086 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2082 : tensor<2x1280x8x8xf32>) outs(%2085 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2087 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2086 : tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2088 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2087, %2086 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %padded_1153 = tensor.pad %2088 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x8x8xf16> to tensor<2x1280x10x10xf16> | |
| %2089 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_327 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2090 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1153, %cst_326 : tensor<2x1280x10x10xf16>, tensor<1280x1280x3x3xf16>) outs(%2089 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %2091 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_329 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2092 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%inserted_slice_1138, %cst_328 : tensor<2x2560x8x8xf16>, tensor<1280x2560x1x1xf16>) outs(%2091 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %2093 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2092, %2090 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2094 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2093, %cst_3 : tensor<2x1280x8x8xf16>, tensor<f64>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %inserted_slice_1154 = tensor.insert_slice %2094 into %1945[0, 0, 0, 0] [2, 1280, 8, 8] [1, 1, 1, 1] : tensor<2x1280x8x8xf16> into tensor<2x2560x8x8xf16> | |
| %inserted_slice_1155 = tensor.insert_slice %1489 into %inserted_slice_1154[0, 1280, 0, 0] [2, 1280, 8, 8] [1, 1, 1, 1] : tensor<2x1280x8x8xf16> into tensor<2x2560x8x8xf16> | |
| %collapsed_1156 = tensor.collapse_shape %inserted_slice_1155 [[0], [1], [2, 3]] : tensor<2x2560x8x8xf16> into tensor<2x2560x64xf16> | |
| %expanded_1157 = tensor.expand_shape %collapsed_1156 [[0], [1, 2], [3]] : tensor<2x2560x64xf16> into tensor<2x32x80x64xf16> | |
| %2095 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2096 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2095 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %2097 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2096 : tensor<f32>) outs(%1948 : tensor<2x32x80x64xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x80x64xf32> | |
| %2098 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1157 : tensor<2x32x80x64xf16>) outs(%2097 : tensor<2x32x80x64xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x80x64xf32> | |
| %2099 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2098 : tensor<2x32x80x64xf32>) outs(%1951 : tensor<2x32x80x64xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x80x64xf64> | |
| %2100 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2101 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2099 : tensor<2x32x80x64xf64>) outs(%2100 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2102 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2101 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_710 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2103 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2099, %2102 : tensor<2x32x80x64xf64>, tensor<2x32x1x1xf64>) outs(%1951 : tensor<2x32x80x64xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x80x64xf64> | |
| %2104 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2103 : tensor<2x32x80x64xf64>) outs(%1951 : tensor<2x32x80x64xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x80x64xf64> | |
| %2105 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2106 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2104 : tensor<2x32x80x64xf64>) outs(%2105 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2107 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2106 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_710 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2108 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2107 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2109 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %2110 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2098 : tensor<2x32x80x64xf32>) outs(%2109 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2111 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2110 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_711 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2112 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2108, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2113 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2112 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2114 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1157, %2111 : tensor<2x32x80x64xf16>, tensor<2x32x1x1xf32>) outs(%1948 : tensor<2x32x80x64xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x80x64xf32> | |
| %2115 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2114, %2113 : tensor<2x32x80x64xf32>, tensor<2x32x1x1xf32>) outs(%1948 : tensor<2x32x80x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x80x64xf32> | |
| %collapsed_1158 = tensor.collapse_shape %2115 [[0], [1, 2], [3]] : tensor<2x32x80x64xf32> into tensor<2x2560x64xf32> | |
| %expanded_1159 = tensor.expand_shape %collapsed_1158 [[0], [1], [2, 3]] : tensor<2x2560x64xf32> into tensor<2x2560x8x8xf32> | |
| %expanded_1160 = tensor.expand_shape %cst_330 [[0, 1, 2]] : tensor<2560xf16> into tensor<2560x1x1xf16> | |
| %2116 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1159, %expanded_1160 : tensor<2x2560x8x8xf32>, tensor<2560x1x1xf16>) outs(%1969 : tensor<2x2560x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x2560x8x8xf32> | |
| %expanded_1161 = tensor.expand_shape %cst_331 [[0, 1, 2]] : tensor<2560xf16> into tensor<2560x1x1xf16> | |
| %2117 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2116, %expanded_1161 : tensor<2x2560x8x8xf32>, tensor<2560x1x1xf16>) outs(%1969 : tensor<2x2560x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x2560x8x8xf32> | |
| %2118 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2119 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2118 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %2120 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2119 : tensor<f16>) outs(%1945 : tensor<2x2560x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x2560x8x8xf16> | |
| %2121 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2117 : tensor<2x2560x8x8xf32>) outs(%2120 : tensor<2x2560x8x8xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x2560x8x8xf16> | |
| %2122 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2121 : tensor<2x2560x8x8xf16>) outs(%1945 : tensor<2x2560x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x2560x8x8xf16> | |
| %2123 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2122, %2121 : tensor<2x2560x8x8xf16>, tensor<2x2560x8x8xf16>) outs(%1945 : tensor<2x2560x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x2560x8x8xf16> | |
| %padded_1162 = tensor.pad %2123 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x2560x8x8xf16> to tensor<2x2560x10x10xf16> | |
| %2124 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_333 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2125 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1162, %cst_332 : tensor<2x2560x10x10xf16>, tensor<1280x2560x3x3xf16>) outs(%2124 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %2126 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280xf16> | |
| %2127 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%2126, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %2128 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_334 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %2129 = linalg.fill ins(%cst_694 : f16) outs(%30 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %2130 = linalg.matmul ins(%2127, %2128 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%2129 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %2131 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_335, %2130 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %expanded_1163 = tensor.expand_shape %2131 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16> | |
| %2132 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2125, %expanded_1163 : tensor<2x1280x8x8xf16>, tensor<2x1280x1x1xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %collapsed_1164 = tensor.collapse_shape %2132 [[0], [1], [2, 3]] : tensor<2x1280x8x8xf16> into tensor<2x1280x64xf16> | |
| %expanded_1165 = tensor.expand_shape %collapsed_1164 [[0], [1, 2], [3]] : tensor<2x1280x64xf16> into tensor<2x32x40x64xf16> | |
| %2133 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2134 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2133 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %2135 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2134 : tensor<f32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %2136 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1165 : tensor<2x32x40x64xf16>) outs(%2135 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %2137 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2136 : tensor<2x32x40x64xf32>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %2138 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2139 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2137 : tensor<2x32x40x64xf64>) outs(%2138 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2140 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2139 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2141 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2137, %2140 : tensor<2x32x40x64xf64>, tensor<2x32x1x1xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %2142 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2141 : tensor<2x32x40x64xf64>) outs(%1495 : tensor<2x32x40x64xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x64xf64> | |
| %2143 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2144 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2142 : tensor<2x32x40x64xf64>) outs(%2143 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2145 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2144 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_713 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2146 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2145 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2147 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %2148 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2136 : tensor<2x32x40x64xf32>) outs(%2147 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2149 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2148 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_714 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2150 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2146, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2151 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2150 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2152 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1165, %2149 : tensor<2x32x40x64xf16>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %2153 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2152, %2151 : tensor<2x32x40x64xf32>, tensor<2x32x1x1xf32>) outs(%1492 : tensor<2x32x40x64xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x64xf32> | |
| %collapsed_1166 = tensor.collapse_shape %2153 [[0], [1, 2], [3]] : tensor<2x32x40x64xf32> into tensor<2x1280x64xf32> | |
| %expanded_1167 = tensor.expand_shape %collapsed_1166 [[0], [1], [2, 3]] : tensor<2x1280x64xf32> into tensor<2x1280x8x8xf32> | |
| %expanded_1168 = tensor.expand_shape %cst_336 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2154 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1167, %expanded_1168 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %expanded_1169 = tensor.expand_shape %cst_337 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2155 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2154, %expanded_1169 : tensor<2x1280x8x8xf32>, tensor<1280x1x1xf16>) outs(%1513 : tensor<2x1280x8x8xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x8x8xf32> | |
| %2156 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2157 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2156 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %2158 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2157 : tensor<f16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2159 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2155 : tensor<2x1280x8x8xf32>) outs(%2158 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2160 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2159 : tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2161 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2160, %2159 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %padded_1170 = tensor.pad %2161 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x8x8xf16> to tensor<2x1280x10x10xf16> | |
| %2162 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_339 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2163 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1170, %cst_338 : tensor<2x1280x10x10xf16>, tensor<1280x1280x3x3xf16>) outs(%2162 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %2164 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_341 : tensor<1280xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2165 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%inserted_slice_1155, %cst_340 : tensor<2x2560x8x8xf16>, tensor<1280x2560x1x1xf16>) outs(%2164 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16> | |
| %2166 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2165, %2163 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2167 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2166, %cst_3 : tensor<2x1280x8x8xf16>, tensor<f64>) outs(%1487 : tensor<2x1280x8x8xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1280x8x8xf16> | |
| %2168 = linalg.generic {indexing_maps = [#map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%out: f16): | |
| %4280 = linalg.index 0 : index | |
| %4281 = linalg.index 1 : index | |
| %4282 = linalg.index 2 : index | |
| %4283 = linalg.index 3 : index | |
| %4284 = arith.floordivsi %4282, %c2 : index | |
| %4285 = arith.floordivsi %4283, %c2 : index | |
| %extracted = tensor.extract %2167[%4280, %4281, %4284, %4285] : tensor<2x1280x8x8xf16> | |
| linalg.yield %extracted : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %padded_1171 = tensor.pad %2168 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x16x16xf16> to tensor<2x1280x18x18xf16> | |
| %2169 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_343 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2170 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1171, %cst_342 : tensor<2x1280x18x18xf16>, tensor<1280x1280x3x3xf16>) outs(%2169 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %2171 = tensor.empty() : tensor<2x2560x16x16xf16> | |
| %inserted_slice_1172 = tensor.insert_slice %2170 into %2171[0, 0, 0, 0] [2, 1280, 16, 16] [1, 1, 1, 1] : tensor<2x1280x16x16xf16> into tensor<2x2560x16x16xf16> | |
| %inserted_slice_1173 = tensor.insert_slice %1486 into %inserted_slice_1172[0, 1280, 0, 0] [2, 1280, 16, 16] [1, 1, 1, 1] : tensor<2x1280x16x16xf16> into tensor<2x2560x16x16xf16> | |
| %collapsed_1174 = tensor.collapse_shape %inserted_slice_1173 [[0], [1], [2, 3]] : tensor<2x2560x16x16xf16> into tensor<2x2560x256xf16> | |
| %expanded_1175 = tensor.expand_shape %collapsed_1174 [[0], [1, 2], [3]] : tensor<2x2560x256xf16> into tensor<2x32x80x256xf16> | |
| %2172 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2173 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2172 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %2174 = tensor.empty() : tensor<2x32x80x256xf32> | |
| %2175 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2173 : tensor<f32>) outs(%2174 : tensor<2x32x80x256xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x80x256xf32> | |
| %2176 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1175 : tensor<2x32x80x256xf16>) outs(%2175 : tensor<2x32x80x256xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x80x256xf32> | |
| %2177 = tensor.empty() : tensor<2x32x80x256xf64> | |
| %2178 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2176 : tensor<2x32x80x256xf32>) outs(%2177 : tensor<2x32x80x256xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x80x256xf64> | |
| %2179 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2180 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2178 : tensor<2x32x80x256xf64>) outs(%2179 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2181 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2180 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_707 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2182 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2178, %2181 : tensor<2x32x80x256xf64>, tensor<2x32x1x1xf64>) outs(%2177 : tensor<2x32x80x256xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x80x256xf64> | |
| %2183 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2182 : tensor<2x32x80x256xf64>) outs(%2177 : tensor<2x32x80x256xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x80x256xf64> | |
| %2184 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2185 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2183 : tensor<2x32x80x256xf64>) outs(%2184 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2186 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2185 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_707 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2187 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2186 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2188 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %2189 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2176 : tensor<2x32x80x256xf32>) outs(%2188 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2190 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2189 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_708 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2191 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2187, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2192 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2191 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2193 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1175, %2190 : tensor<2x32x80x256xf16>, tensor<2x32x1x1xf32>) outs(%2174 : tensor<2x32x80x256xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x80x256xf32> | |
| %2194 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2193, %2192 : tensor<2x32x80x256xf32>, tensor<2x32x1x1xf32>) outs(%2174 : tensor<2x32x80x256xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x80x256xf32> | |
| %collapsed_1176 = tensor.collapse_shape %2194 [[0], [1, 2], [3]] : tensor<2x32x80x256xf32> into tensor<2x2560x256xf32> | |
| %expanded_1177 = tensor.expand_shape %collapsed_1176 [[0], [1], [2, 3]] : tensor<2x2560x256xf32> into tensor<2x2560x16x16xf32> | |
| %expanded_1178 = tensor.expand_shape %cst_344 [[0, 1, 2]] : tensor<2560xf16> into tensor<2560x1x1xf16> | |
| %2195 = tensor.empty() : tensor<2x2560x16x16xf32> | |
| %2196 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1177, %expanded_1178 : tensor<2x2560x16x16xf32>, tensor<2560x1x1xf16>) outs(%2195 : tensor<2x2560x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x2560x16x16xf32> | |
| %expanded_1179 = tensor.expand_shape %cst_345 [[0, 1, 2]] : tensor<2560xf16> into tensor<2560x1x1xf16> | |
| %2197 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2196, %expanded_1179 : tensor<2x2560x16x16xf32>, tensor<2560x1x1xf16>) outs(%2195 : tensor<2x2560x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x2560x16x16xf32> | |
| %2198 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2199 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2198 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %2200 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2199 : tensor<f16>) outs(%2171 : tensor<2x2560x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x2560x16x16xf16> | |
| %2201 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2197 : tensor<2x2560x16x16xf32>) outs(%2200 : tensor<2x2560x16x16xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x2560x16x16xf16> | |
| %2202 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2201 : tensor<2x2560x16x16xf16>) outs(%2171 : tensor<2x2560x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x2560x16x16xf16> | |
| %2203 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2202, %2201 : tensor<2x2560x16x16xf16>, tensor<2x2560x16x16xf16>) outs(%2171 : tensor<2x2560x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x2560x16x16xf16> | |
| %padded_1180 = tensor.pad %2203 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x2560x16x16xf16> to tensor<2x2560x18x18xf16> | |
| %2204 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_347 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2205 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1180, %cst_346 : tensor<2x2560x18x18xf16>, tensor<1280x2560x3x3xf16>) outs(%2204 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %2206 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280xf16> | |
| %2207 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%2206, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %2208 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_348 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %2209 = linalg.fill ins(%cst_694 : f16) outs(%30 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %2210 = linalg.matmul ins(%2207, %2208 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%2209 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %2211 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_349, %2210 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %expanded_1181 = tensor.expand_shape %2211 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16> | |
| %2212 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2205, %expanded_1181 : tensor<2x1280x16x16xf16>, tensor<2x1280x1x1xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %collapsed_1182 = tensor.collapse_shape %2212 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16> | |
| %expanded_1183 = tensor.expand_shape %collapsed_1182 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16> | |
| %2213 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2214 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2213 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %2215 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2214 : tensor<f32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2216 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1183 : tensor<2x32x40x256xf16>) outs(%2215 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2217 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2216 : tensor<2x32x40x256xf32>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2218 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2219 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2217 : tensor<2x32x40x256xf64>) outs(%2218 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2220 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2219 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2221 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2217, %2220 : tensor<2x32x40x256xf64>, tensor<2x32x1x1xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2222 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2221 : tensor<2x32x40x256xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2223 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2224 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2222 : tensor<2x32x40x256xf64>) outs(%2223 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2225 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2224 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2226 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2225 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2227 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %2228 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2216 : tensor<2x32x40x256xf32>) outs(%2227 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2229 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2228 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_706 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2230 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2226, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2231 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2230 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2232 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1183, %2229 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2233 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2232, %2231 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %collapsed_1184 = tensor.collapse_shape %2233 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32> | |
| %expanded_1185 = tensor.expand_shape %collapsed_1184 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32> | |
| %expanded_1186 = tensor.expand_shape %cst_350 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2234 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1185, %expanded_1186 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %expanded_1187 = tensor.expand_shape %cst_351 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2235 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2234, %expanded_1187 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %2236 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2237 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2236 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %2238 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2237 : tensor<f16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2239 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2235 : tensor<2x1280x16x16xf32>) outs(%2238 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2240 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2239 : tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2241 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2240, %2239 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %padded_1188 = tensor.pad %2241 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x16x16xf16> to tensor<2x1280x18x18xf16> | |
| %2242 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_353 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2243 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1188, %cst_352 : tensor<2x1280x18x18xf16>, tensor<1280x1280x3x3xf16>) outs(%2242 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %2244 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_355 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2245 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%inserted_slice_1173, %cst_354 : tensor<2x2560x16x16xf16>, tensor<1280x2560x1x1xf16>) outs(%2244 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %2246 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2245, %2243 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2247 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2246, %cst_3 : tensor<2x1280x16x16xf16>, tensor<f64>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %collapsed_1189 = tensor.collapse_shape %2247 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16> | |
| %expanded_1190 = tensor.expand_shape %collapsed_1189 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16> | |
| %2248 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2249 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2248 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %2250 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2249 : tensor<f32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2251 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1190 : tensor<2x32x40x256xf16>) outs(%2250 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2252 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2251 : tensor<2x32x40x256xf32>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2253 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2254 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2252 : tensor<2x32x40x256xf64>) outs(%2253 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2255 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2254 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2256 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2252, %2255 : tensor<2x32x40x256xf64>, tensor<2x32x1x1xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2257 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2256 : tensor<2x32x40x256xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2258 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2259 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2257 : tensor<2x32x40x256xf64>) outs(%2258 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2260 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2259 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2261 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2260 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2262 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %2263 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2251 : tensor<2x32x40x256xf32>) outs(%2262 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2264 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2263 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_706 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2265 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2261, %cst_2 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2266 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2265 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2267 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1190, %2264 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2268 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2267, %2266 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %collapsed_1191 = tensor.collapse_shape %2268 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32> | |
| %expanded_1192 = tensor.expand_shape %collapsed_1191 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32> | |
| %expanded_1193 = tensor.expand_shape %cst_356 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2269 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1192, %expanded_1193 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %expanded_1194 = tensor.expand_shape %cst_357 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2270 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2269, %expanded_1194 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %2271 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2272 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2271 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %2273 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2272 : tensor<f16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2274 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2270 : tensor<2x1280x16x16xf32>) outs(%2273 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2275 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_359 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2276 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2274, %cst_358 : tensor<2x1280x16x16xf16>, tensor<1280x1280x1x1xf16>) outs(%2275 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %2277 = linalg.generic {indexing_maps = [#map11, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2276 : tensor<2x1280x16x16xf16>) outs(%1117 : tensor<2x16x16x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x16x16x1280xf16> | |
| %collapsed_1195 = tensor.collapse_shape %2277 [[0], [1, 2], [3]] : tensor<2x16x16x1280xf16> into tensor<2x256x1280xf16> | |
| %2278 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %2279 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_1195 : tensor<2x256x1280xf16>) outs(%2278 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2280 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2279 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2281 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2280 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2282 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1195, %2281 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2283 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2282 : tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2284 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %2285 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2283 : tensor<2x256x1280xf16>) outs(%2284 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2286 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2285 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2287 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2286 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2288 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2287 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2289 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2288 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2290 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2282, %2289 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2291 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2290, %cst_360 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2292 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2291, %cst_361 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2293 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_362 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1196 = tensor.collapse_shape %2292 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16> | |
| %2294 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2295 = linalg.matmul ins(%collapsed_1196, %2293 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2294 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2296 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_363 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %2297 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2298 = linalg.matmul ins(%collapsed_1196, %2296 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2297 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2299 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_364 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %2300 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2301 = linalg.matmul ins(%collapsed_1196, %2299 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2300 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %expanded_1197 = tensor.expand_shape %2295 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %2302 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1197 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_1198 = tensor.collapse_shape %2302 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %expanded_1199 = tensor.expand_shape %2298 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %2303 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1199 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_1200 = tensor.collapse_shape %2303 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %expanded_1201 = tensor.expand_shape %2301 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %2304 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1201 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_1202 = tensor.collapse_shape %2304 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %2305 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1200 : tensor<16x256x160xf16>) outs(%1150 : tensor<16x160x256xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x160x256xf16> | |
| %2306 = linalg.fill ins(%cst_694 : f16) outs(%1152 : tensor<16x256x256xf16>) -> tensor<16x256x256xf16> | |
| %2307 = linalg.batch_matmul ins(%collapsed_1198, %2305 : tensor<16x256x160xf16>, tensor<16x160x256xf16>) outs(%2306 : tensor<16x256x256xf16>) -> tensor<16x256x256xf16> | |
| %2308 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2307, %cst : tensor<16x256x256xf16>, tensor<f64>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %2309 = linalg.fill ins(%c0_i64 : i64) outs(%1156 : tensor<16x256x1xi64>) -> tensor<16x256x1xi64> | |
| %2310 = linalg.fill ins(%cst_696 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %2311:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2308 : tensor<16x256x256xf16>) outs(%2310, %2309 : tensor<16x256x1xf16>, tensor<16x256x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x256x1xf16>, tensor<16x256x1xi64>) | |
| %2312 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2308, %2311#0 : tensor<16x256x256xf16>, tensor<16x256x1xf16>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %2313 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2312 : tensor<16x256x256xf16>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %2314 = linalg.fill ins(%cst_694 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %2315 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2313 : tensor<16x256x256xf16>) outs(%2314 : tensor<16x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x1xf16> | |
| %2316 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2313, %2315 : tensor<16x256x256xf16>, tensor<16x256x1xf16>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %2317 = linalg.fill ins(%cst_694 : f16) outs(%1166 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %2318 = linalg.batch_matmul ins(%2316, %collapsed_1202 : tensor<16x256x256xf16>, tensor<16x256x160xf16>) outs(%2317 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %expanded_1203 = tensor.expand_shape %2318 [[0, 1], [2], [3]] : tensor<16x256x160xf16> into tensor<2x8x256x160xf16> | |
| %2319 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1203 : tensor<2x8x256x160xf16>) outs(%1169 : tensor<2x256x8x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x8x160xf16> | |
| %2320 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_365 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1204 = tensor.collapse_shape %2319 [[0, 1], [2, 3]] : tensor<2x256x8x160xf16> into tensor<512x1280xf16> | |
| %2321 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2322 = linalg.matmul ins(%collapsed_1204, %2320 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2321 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2323 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_366, %2322 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1137 : tensor<512x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<512x1280xf16> | |
| %expanded_1205 = tensor.expand_shape %2323 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16> | |
| %2324 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1205, %collapsed_1195 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2325 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %2326 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2324 : tensor<2x256x1280xf16>) outs(%2325 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2327 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2326 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2328 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2327 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2329 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2324, %2328 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2330 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2329 : tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2331 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %2332 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2330 : tensor<2x256x1280xf16>) outs(%2331 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2333 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2332 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2334 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2333 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2335 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2334 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2336 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2335 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2337 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2329, %2336 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2338 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2337, %cst_367 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2339 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2338, %cst_368 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2340 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_369 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1206 = tensor.collapse_shape %2339 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16> | |
| %2341 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2342 = linalg.matmul ins(%collapsed_1206, %2340 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2341 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2343 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_370 : tensor<1280x768xf16>) outs(%1194 : tensor<768x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x1280xf16> | |
| %2344 = linalg.fill ins(%cst_694 : f16) outs(%1196 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %2345 = linalg.matmul ins(%collapsed_761, %2343 : tensor<154x768xf16>, tensor<768x1280xf16>) outs(%2344 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %2346 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_371 : tensor<1280x768xf16>) outs(%1194 : tensor<768x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x1280xf16> | |
| %2347 = linalg.fill ins(%cst_694 : f16) outs(%1196 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %2348 = linalg.matmul ins(%collapsed_761, %2346 : tensor<154x768xf16>, tensor<768x1280xf16>) outs(%2347 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %expanded_1207 = tensor.expand_shape %2342 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %2349 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1207 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_1208 = tensor.collapse_shape %2349 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %expanded_1209 = tensor.expand_shape %2345 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x8x160xf16> | |
| %2350 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1209 : tensor<2x77x8x160xf16>) outs(%1203 : tensor<2x8x77x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x160xf16> | |
| %collapsed_1210 = tensor.collapse_shape %2350 [[0, 1], [2], [3]] : tensor<2x8x77x160xf16> into tensor<16x77x160xf16> | |
| %expanded_1211 = tensor.expand_shape %2348 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x8x160xf16> | |
| %2351 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1211 : tensor<2x77x8x160xf16>) outs(%1203 : tensor<2x8x77x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x160xf16> | |
| %collapsed_1212 = tensor.collapse_shape %2351 [[0, 1], [2], [3]] : tensor<2x8x77x160xf16> into tensor<16x77x160xf16> | |
| %2352 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1210 : tensor<16x77x160xf16>) outs(%1206 : tensor<16x160x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x160x77xf16> | |
| %2353 = linalg.fill ins(%cst_694 : f16) outs(%1208 : tensor<16x256x77xf16>) -> tensor<16x256x77xf16> | |
| %2354 = linalg.batch_matmul ins(%collapsed_1208, %2352 : tensor<16x256x160xf16>, tensor<16x160x77xf16>) outs(%2353 : tensor<16x256x77xf16>) -> tensor<16x256x77xf16> | |
| %2355 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2354, %cst : tensor<16x256x77xf16>, tensor<f64>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %2356 = linalg.fill ins(%c0_i64 : i64) outs(%1156 : tensor<16x256x1xi64>) -> tensor<16x256x1xi64> | |
| %2357 = linalg.fill ins(%cst_696 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %2358:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2355 : tensor<16x256x77xf16>) outs(%2357, %2356 : tensor<16x256x1xf16>, tensor<16x256x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x256x1xf16>, tensor<16x256x1xi64>) | |
| %2359 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2355, %2358#0 : tensor<16x256x77xf16>, tensor<16x256x1xf16>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %2360 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2359 : tensor<16x256x77xf16>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %2361 = linalg.fill ins(%cst_694 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %2362 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2360 : tensor<16x256x77xf16>) outs(%2361 : tensor<16x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x1xf16> | |
| %2363 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2360, %2362 : tensor<16x256x77xf16>, tensor<16x256x1xf16>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %2364 = linalg.fill ins(%cst_694 : f16) outs(%1166 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %2365 = linalg.batch_matmul ins(%2363, %collapsed_1212 : tensor<16x256x77xf16>, tensor<16x77x160xf16>) outs(%2364 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %expanded_1213 = tensor.expand_shape %2365 [[0, 1], [2], [3]] : tensor<16x256x160xf16> into tensor<2x8x256x160xf16> | |
| %2366 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1213 : tensor<2x8x256x160xf16>) outs(%1169 : tensor<2x256x8x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x8x160xf16> | |
| %2367 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_372 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1214 = tensor.collapse_shape %2366 [[0, 1], [2, 3]] : tensor<2x256x8x160xf16> into tensor<512x1280xf16> | |
| %2368 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2369 = linalg.matmul ins(%collapsed_1214, %2367 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2368 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2370 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_373, %2369 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1137 : tensor<512x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<512x1280xf16> | |
| %expanded_1215 = tensor.expand_shape %2370 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16> | |
| %2371 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1215, %2324 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2372 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %2373 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2371 : tensor<2x256x1280xf16>) outs(%2372 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2374 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2373 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2375 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2374 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2376 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2371, %2375 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2377 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2376 : tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2378 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %2379 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2377 : tensor<2x256x1280xf16>) outs(%2378 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2380 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2379 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2381 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2380 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2382 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2381 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2383 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2382 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2384 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2376, %2383 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2385 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2384, %cst_374 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2386 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2385, %cst_375 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2387 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_376 : tensor<10240x1280xf16>) outs(%1243 : tensor<1280x10240xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x10240xf16> | |
| %collapsed_1216 = tensor.collapse_shape %2386 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16> | |
| %2388 = linalg.fill ins(%cst_694 : f16) outs(%1245 : tensor<512x10240xf16>) -> tensor<512x10240xf16> | |
| %2389 = linalg.matmul ins(%collapsed_1216, %2387 : tensor<512x1280xf16>, tensor<1280x10240xf16>) outs(%2388 : tensor<512x10240xf16>) -> tensor<512x10240xf16> | |
| %2390 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_377, %2389 : tensor<10240xf16>, tensor<512x10240xf16>) outs(%1245 : tensor<512x10240xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<512x10240xf16> | |
| %expanded_1217 = tensor.expand_shape %2390 [[0, 1], [2]] : tensor<512x10240xf16> into tensor<2x256x10240xf16> | |
| %extracted_slice_1218 = tensor.extract_slice %expanded_1217[0, 0, 0] [2, 256, 5120] [1, 1, 1] : tensor<2x256x10240xf16> to tensor<2x256x5120xf16> | |
| %extracted_slice_1219 = tensor.extract_slice %expanded_1217[0, 0, 5120] [2, 256, 5120] [1, 1, 1] : tensor<2x256x10240xf16> to tensor<2x256x5120xf16> | |
| %2391 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1219 : tensor<2x256x5120xf16>) outs(%1249 : tensor<2x256x5120xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.sqrt %cst_697 : f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| %4282 = math.erf %4281 : f16 | |
| %4283 = arith.addf %4282, %cst_695 : f16 | |
| %4284 = arith.mulf %4283, %cst_698 : f16 | |
| %4285 = arith.mulf %in, %4284 : f16 | |
| linalg.yield %4285 : f16 | |
| } -> tensor<2x256x5120xf16> | |
| %2392 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1218, %2391 : tensor<2x256x5120xf16>, tensor<2x256x5120xf16>) outs(%1249 : tensor<2x256x5120xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x5120xf16> | |
| %2393 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_378 : tensor<1280x5120xf16>) outs(%1252 : tensor<5120x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<5120x1280xf16> | |
| %collapsed_1220 = tensor.collapse_shape %2392 [[0, 1], [2]] : tensor<2x256x5120xf16> into tensor<512x5120xf16> | |
| %2394 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2395 = linalg.matmul ins(%collapsed_1220, %2393 : tensor<512x5120xf16>, tensor<5120x1280xf16>) outs(%2394 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2396 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_379, %2395 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1137 : tensor<512x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<512x1280xf16> | |
| %expanded_1221 = tensor.expand_shape %2396 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16> | |
| %2397 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1221, %2371 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %expanded_1222 = tensor.expand_shape %2397 [[0], [1, 2], [3]] : tensor<2x256x1280xf16> into tensor<2x16x16x1280xf16> | |
| %2398 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1222 : tensor<2x16x16x1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2399 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_381 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2400 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2398, %cst_380 : tensor<2x1280x16x16xf16>, tensor<1280x1280x1x1xf16>) outs(%2399 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %2401 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2400, %2247 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %inserted_slice_1223 = tensor.insert_slice %2401 into %2171[0, 0, 0, 0] [2, 1280, 16, 16] [1, 1, 1, 1] : tensor<2x1280x16x16xf16> into tensor<2x2560x16x16xf16> | |
| %inserted_slice_1224 = tensor.insert_slice %1261 into %inserted_slice_1223[0, 1280, 0, 0] [2, 1280, 16, 16] [1, 1, 1, 1] : tensor<2x1280x16x16xf16> into tensor<2x2560x16x16xf16> | |
| %collapsed_1225 = tensor.collapse_shape %inserted_slice_1224 [[0], [1], [2, 3]] : tensor<2x2560x16x16xf16> into tensor<2x2560x256xf16> | |
| %expanded_1226 = tensor.expand_shape %collapsed_1225 [[0], [1, 2], [3]] : tensor<2x2560x256xf16> into tensor<2x32x80x256xf16> | |
| %2402 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2403 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2402 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %2404 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2403 : tensor<f32>) outs(%2174 : tensor<2x32x80x256xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x80x256xf32> | |
| %2405 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1226 : tensor<2x32x80x256xf16>) outs(%2404 : tensor<2x32x80x256xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x80x256xf32> | |
| %2406 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2405 : tensor<2x32x80x256xf32>) outs(%2177 : tensor<2x32x80x256xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x80x256xf64> | |
| %2407 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2408 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2406 : tensor<2x32x80x256xf64>) outs(%2407 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2409 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2408 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_707 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2410 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2406, %2409 : tensor<2x32x80x256xf64>, tensor<2x32x1x1xf64>) outs(%2177 : tensor<2x32x80x256xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x80x256xf64> | |
| %2411 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2410 : tensor<2x32x80x256xf64>) outs(%2177 : tensor<2x32x80x256xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x80x256xf64> | |
| %2412 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2413 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2411 : tensor<2x32x80x256xf64>) outs(%2412 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2414 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2413 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_707 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2415 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2414 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2416 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %2417 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2405 : tensor<2x32x80x256xf32>) outs(%2416 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2418 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2417 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_708 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2419 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2415, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2420 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2419 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2421 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1226, %2418 : tensor<2x32x80x256xf16>, tensor<2x32x1x1xf32>) outs(%2174 : tensor<2x32x80x256xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x80x256xf32> | |
| %2422 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2421, %2420 : tensor<2x32x80x256xf32>, tensor<2x32x1x1xf32>) outs(%2174 : tensor<2x32x80x256xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x80x256xf32> | |
| %collapsed_1227 = tensor.collapse_shape %2422 [[0], [1, 2], [3]] : tensor<2x32x80x256xf32> into tensor<2x2560x256xf32> | |
| %expanded_1228 = tensor.expand_shape %collapsed_1227 [[0], [1], [2, 3]] : tensor<2x2560x256xf32> into tensor<2x2560x16x16xf32> | |
| %expanded_1229 = tensor.expand_shape %cst_382 [[0, 1, 2]] : tensor<2560xf16> into tensor<2560x1x1xf16> | |
| %2423 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1228, %expanded_1229 : tensor<2x2560x16x16xf32>, tensor<2560x1x1xf16>) outs(%2195 : tensor<2x2560x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x2560x16x16xf32> | |
| %expanded_1230 = tensor.expand_shape %cst_383 [[0, 1, 2]] : tensor<2560xf16> into tensor<2560x1x1xf16> | |
| %2424 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2423, %expanded_1230 : tensor<2x2560x16x16xf32>, tensor<2560x1x1xf16>) outs(%2195 : tensor<2x2560x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x2560x16x16xf32> | |
| %2425 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2426 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2425 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %2427 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2426 : tensor<f16>) outs(%2171 : tensor<2x2560x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x2560x16x16xf16> | |
| %2428 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2424 : tensor<2x2560x16x16xf32>) outs(%2427 : tensor<2x2560x16x16xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x2560x16x16xf16> | |
| %2429 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2428 : tensor<2x2560x16x16xf16>) outs(%2171 : tensor<2x2560x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x2560x16x16xf16> | |
| %2430 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2429, %2428 : tensor<2x2560x16x16xf16>, tensor<2x2560x16x16xf16>) outs(%2171 : tensor<2x2560x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x2560x16x16xf16> | |
| %padded_1231 = tensor.pad %2430 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x2560x16x16xf16> to tensor<2x2560x18x18xf16> | |
| %2431 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_385 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2432 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1231, %cst_384 : tensor<2x2560x18x18xf16>, tensor<1280x2560x3x3xf16>) outs(%2431 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %2433 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280xf16> | |
| %2434 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%2433, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %2435 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_386 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %2436 = linalg.fill ins(%cst_694 : f16) outs(%30 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %2437 = linalg.matmul ins(%2434, %2435 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%2436 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %2438 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_387, %2437 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %expanded_1232 = tensor.expand_shape %2438 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16> | |
| %2439 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2432, %expanded_1232 : tensor<2x1280x16x16xf16>, tensor<2x1280x1x1xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %collapsed_1233 = tensor.collapse_shape %2439 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16> | |
| %expanded_1234 = tensor.expand_shape %collapsed_1233 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16> | |
| %2440 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2441 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2440 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %2442 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2441 : tensor<f32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2443 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1234 : tensor<2x32x40x256xf16>) outs(%2442 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2444 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2443 : tensor<2x32x40x256xf32>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2445 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2446 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2444 : tensor<2x32x40x256xf64>) outs(%2445 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2447 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2446 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2448 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2444, %2447 : tensor<2x32x40x256xf64>, tensor<2x32x1x1xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2449 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2448 : tensor<2x32x40x256xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2450 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2451 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2449 : tensor<2x32x40x256xf64>) outs(%2450 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2452 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2451 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2453 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2452 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2454 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %2455 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2443 : tensor<2x32x40x256xf32>) outs(%2454 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2456 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2455 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_706 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2457 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2453, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2458 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2457 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2459 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1234, %2456 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2460 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2459, %2458 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %collapsed_1235 = tensor.collapse_shape %2460 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32> | |
| %expanded_1236 = tensor.expand_shape %collapsed_1235 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32> | |
| %expanded_1237 = tensor.expand_shape %cst_388 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2461 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1236, %expanded_1237 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %expanded_1238 = tensor.expand_shape %cst_389 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2462 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2461, %expanded_1238 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %2463 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2464 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2463 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %2465 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2464 : tensor<f16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2466 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2462 : tensor<2x1280x16x16xf32>) outs(%2465 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2467 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2466 : tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2468 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2467, %2466 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %padded_1239 = tensor.pad %2468 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x16x16xf16> to tensor<2x1280x18x18xf16> | |
| %2469 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_391 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2470 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1239, %cst_390 : tensor<2x1280x18x18xf16>, tensor<1280x1280x3x3xf16>) outs(%2469 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %2471 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_393 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2472 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%inserted_slice_1224, %cst_392 : tensor<2x2560x16x16xf16>, tensor<1280x2560x1x1xf16>) outs(%2471 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %2473 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2472, %2470 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2474 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2473, %cst_3 : tensor<2x1280x16x16xf16>, tensor<f64>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %collapsed_1240 = tensor.collapse_shape %2474 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16> | |
| %expanded_1241 = tensor.expand_shape %collapsed_1240 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16> | |
| %2475 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2476 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2475 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %2477 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2476 : tensor<f32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2478 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1241 : tensor<2x32x40x256xf16>) outs(%2477 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2479 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2478 : tensor<2x32x40x256xf32>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2480 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2481 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2479 : tensor<2x32x40x256xf64>) outs(%2480 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2482 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2481 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2483 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2479, %2482 : tensor<2x32x40x256xf64>, tensor<2x32x1x1xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2484 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2483 : tensor<2x32x40x256xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2485 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2486 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2484 : tensor<2x32x40x256xf64>) outs(%2485 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2487 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2486 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2488 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2487 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2489 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %2490 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2478 : tensor<2x32x40x256xf32>) outs(%2489 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2491 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2490 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_706 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2492 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2488, %cst_2 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2493 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2492 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2494 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1241, %2491 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2495 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2494, %2493 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %collapsed_1242 = tensor.collapse_shape %2495 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32> | |
| %expanded_1243 = tensor.expand_shape %collapsed_1242 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32> | |
| %expanded_1244 = tensor.expand_shape %cst_394 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2496 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1243, %expanded_1244 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %expanded_1245 = tensor.expand_shape %cst_395 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2497 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2496, %expanded_1245 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %2498 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2499 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2498 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %2500 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2499 : tensor<f16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2501 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2497 : tensor<2x1280x16x16xf32>) outs(%2500 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2502 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_397 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2503 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2501, %cst_396 : tensor<2x1280x16x16xf16>, tensor<1280x1280x1x1xf16>) outs(%2502 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %2504 = linalg.generic {indexing_maps = [#map11, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2503 : tensor<2x1280x16x16xf16>) outs(%1117 : tensor<2x16x16x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x16x16x1280xf16> | |
| %collapsed_1246 = tensor.collapse_shape %2504 [[0], [1, 2], [3]] : tensor<2x16x16x1280xf16> into tensor<2x256x1280xf16> | |
| %2505 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %2506 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_1246 : tensor<2x256x1280xf16>) outs(%2505 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2507 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2506 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2508 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2507 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2509 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1246, %2508 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2510 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2509 : tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2511 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %2512 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2510 : tensor<2x256x1280xf16>) outs(%2511 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2513 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2512 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2514 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2513 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2515 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2514 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2516 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2515 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2517 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2509, %2516 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2518 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2517, %cst_398 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2519 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2518, %cst_399 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2520 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_400 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1247 = tensor.collapse_shape %2519 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16> | |
| %2521 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2522 = linalg.matmul ins(%collapsed_1247, %2520 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2521 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2523 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_401 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %2524 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2525 = linalg.matmul ins(%collapsed_1247, %2523 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2524 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2526 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_402 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %2527 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2528 = linalg.matmul ins(%collapsed_1247, %2526 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2527 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %expanded_1248 = tensor.expand_shape %2522 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %2529 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1248 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_1249 = tensor.collapse_shape %2529 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %expanded_1250 = tensor.expand_shape %2525 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %2530 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1250 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_1251 = tensor.collapse_shape %2530 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %expanded_1252 = tensor.expand_shape %2528 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %2531 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1252 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_1253 = tensor.collapse_shape %2531 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %2532 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1251 : tensor<16x256x160xf16>) outs(%1150 : tensor<16x160x256xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x160x256xf16> | |
| %2533 = linalg.fill ins(%cst_694 : f16) outs(%1152 : tensor<16x256x256xf16>) -> tensor<16x256x256xf16> | |
| %2534 = linalg.batch_matmul ins(%collapsed_1249, %2532 : tensor<16x256x160xf16>, tensor<16x160x256xf16>) outs(%2533 : tensor<16x256x256xf16>) -> tensor<16x256x256xf16> | |
| %2535 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2534, %cst : tensor<16x256x256xf16>, tensor<f64>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %2536 = linalg.fill ins(%c0_i64 : i64) outs(%1156 : tensor<16x256x1xi64>) -> tensor<16x256x1xi64> | |
| %2537 = linalg.fill ins(%cst_696 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %2538:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2535 : tensor<16x256x256xf16>) outs(%2537, %2536 : tensor<16x256x1xf16>, tensor<16x256x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x256x1xf16>, tensor<16x256x1xi64>) | |
| %2539 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2535, %2538#0 : tensor<16x256x256xf16>, tensor<16x256x1xf16>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %2540 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2539 : tensor<16x256x256xf16>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %2541 = linalg.fill ins(%cst_694 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %2542 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2540 : tensor<16x256x256xf16>) outs(%2541 : tensor<16x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x1xf16> | |
| %2543 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2540, %2542 : tensor<16x256x256xf16>, tensor<16x256x1xf16>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %2544 = linalg.fill ins(%cst_694 : f16) outs(%1166 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %2545 = linalg.batch_matmul ins(%2543, %collapsed_1253 : tensor<16x256x256xf16>, tensor<16x256x160xf16>) outs(%2544 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %expanded_1254 = tensor.expand_shape %2545 [[0, 1], [2], [3]] : tensor<16x256x160xf16> into tensor<2x8x256x160xf16> | |
| %2546 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1254 : tensor<2x8x256x160xf16>) outs(%1169 : tensor<2x256x8x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x8x160xf16> | |
| %2547 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_403 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1255 = tensor.collapse_shape %2546 [[0, 1], [2, 3]] : tensor<2x256x8x160xf16> into tensor<512x1280xf16> | |
| %2548 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2549 = linalg.matmul ins(%collapsed_1255, %2547 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2548 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2550 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_404, %2549 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1137 : tensor<512x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<512x1280xf16> | |
| %expanded_1256 = tensor.expand_shape %2550 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16> | |
| %2551 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1256, %collapsed_1246 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2552 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %2553 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2551 : tensor<2x256x1280xf16>) outs(%2552 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2554 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2553 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2555 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2554 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2556 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2551, %2555 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2557 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2556 : tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2558 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %2559 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2557 : tensor<2x256x1280xf16>) outs(%2558 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2560 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2559 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2561 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2560 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2562 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2561 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2563 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2562 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2564 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2556, %2563 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2565 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2564, %cst_405 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2566 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2565, %cst_406 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2567 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_407 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1257 = tensor.collapse_shape %2566 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16> | |
| %2568 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2569 = linalg.matmul ins(%collapsed_1257, %2567 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2568 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2570 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_408 : tensor<1280x768xf16>) outs(%1194 : tensor<768x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x1280xf16> | |
| %2571 = linalg.fill ins(%cst_694 : f16) outs(%1196 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %2572 = linalg.matmul ins(%collapsed_761, %2570 : tensor<154x768xf16>, tensor<768x1280xf16>) outs(%2571 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %2573 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_409 : tensor<1280x768xf16>) outs(%1194 : tensor<768x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x1280xf16> | |
| %2574 = linalg.fill ins(%cst_694 : f16) outs(%1196 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %2575 = linalg.matmul ins(%collapsed_761, %2573 : tensor<154x768xf16>, tensor<768x1280xf16>) outs(%2574 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %expanded_1258 = tensor.expand_shape %2569 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %2576 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1258 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_1259 = tensor.collapse_shape %2576 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %expanded_1260 = tensor.expand_shape %2572 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x8x160xf16> | |
| %2577 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1260 : tensor<2x77x8x160xf16>) outs(%1203 : tensor<2x8x77x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x160xf16> | |
| %collapsed_1261 = tensor.collapse_shape %2577 [[0, 1], [2], [3]] : tensor<2x8x77x160xf16> into tensor<16x77x160xf16> | |
| %expanded_1262 = tensor.expand_shape %2575 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x8x160xf16> | |
| %2578 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1262 : tensor<2x77x8x160xf16>) outs(%1203 : tensor<2x8x77x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x160xf16> | |
| %collapsed_1263 = tensor.collapse_shape %2578 [[0, 1], [2], [3]] : tensor<2x8x77x160xf16> into tensor<16x77x160xf16> | |
| %2579 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1261 : tensor<16x77x160xf16>) outs(%1206 : tensor<16x160x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x160x77xf16> | |
| %2580 = linalg.fill ins(%cst_694 : f16) outs(%1208 : tensor<16x256x77xf16>) -> tensor<16x256x77xf16> | |
| %2581 = linalg.batch_matmul ins(%collapsed_1259, %2579 : tensor<16x256x160xf16>, tensor<16x160x77xf16>) outs(%2580 : tensor<16x256x77xf16>) -> tensor<16x256x77xf16> | |
| %2582 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2581, %cst : tensor<16x256x77xf16>, tensor<f64>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %2583 = linalg.fill ins(%c0_i64 : i64) outs(%1156 : tensor<16x256x1xi64>) -> tensor<16x256x1xi64> | |
| %2584 = linalg.fill ins(%cst_696 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %2585:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2582 : tensor<16x256x77xf16>) outs(%2584, %2583 : tensor<16x256x1xf16>, tensor<16x256x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x256x1xf16>, tensor<16x256x1xi64>) | |
| %2586 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2582, %2585#0 : tensor<16x256x77xf16>, tensor<16x256x1xf16>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %2587 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2586 : tensor<16x256x77xf16>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %2588 = linalg.fill ins(%cst_694 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %2589 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2587 : tensor<16x256x77xf16>) outs(%2588 : tensor<16x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x1xf16> | |
| %2590 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2587, %2589 : tensor<16x256x77xf16>, tensor<16x256x1xf16>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %2591 = linalg.fill ins(%cst_694 : f16) outs(%1166 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %2592 = linalg.batch_matmul ins(%2590, %collapsed_1263 : tensor<16x256x77xf16>, tensor<16x77x160xf16>) outs(%2591 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %expanded_1264 = tensor.expand_shape %2592 [[0, 1], [2], [3]] : tensor<16x256x160xf16> into tensor<2x8x256x160xf16> | |
| %2593 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1264 : tensor<2x8x256x160xf16>) outs(%1169 : tensor<2x256x8x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x8x160xf16> | |
| %2594 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_410 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1265 = tensor.collapse_shape %2593 [[0, 1], [2, 3]] : tensor<2x256x8x160xf16> into tensor<512x1280xf16> | |
| %2595 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2596 = linalg.matmul ins(%collapsed_1265, %2594 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2595 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2597 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_411, %2596 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1137 : tensor<512x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<512x1280xf16> | |
| %expanded_1266 = tensor.expand_shape %2597 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16> | |
| %2598 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1266, %2551 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2599 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %2600 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2598 : tensor<2x256x1280xf16>) outs(%2599 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2601 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2600 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2602 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2601 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2603 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2598, %2602 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2604 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2603 : tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2605 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %2606 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2604 : tensor<2x256x1280xf16>) outs(%2605 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2607 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2606 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2608 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2607 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2609 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2608 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2610 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2609 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2611 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2603, %2610 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2612 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2611, %cst_412 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2613 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2612, %cst_413 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2614 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_414 : tensor<10240x1280xf16>) outs(%1243 : tensor<1280x10240xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x10240xf16> | |
| %collapsed_1267 = tensor.collapse_shape %2613 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16> | |
| %2615 = linalg.fill ins(%cst_694 : f16) outs(%1245 : tensor<512x10240xf16>) -> tensor<512x10240xf16> | |
| %2616 = linalg.matmul ins(%collapsed_1267, %2614 : tensor<512x1280xf16>, tensor<1280x10240xf16>) outs(%2615 : tensor<512x10240xf16>) -> tensor<512x10240xf16> | |
| %2617 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_415, %2616 : tensor<10240xf16>, tensor<512x10240xf16>) outs(%1245 : tensor<512x10240xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<512x10240xf16> | |
| %expanded_1268 = tensor.expand_shape %2617 [[0, 1], [2]] : tensor<512x10240xf16> into tensor<2x256x10240xf16> | |
| %extracted_slice_1269 = tensor.extract_slice %expanded_1268[0, 0, 0] [2, 256, 5120] [1, 1, 1] : tensor<2x256x10240xf16> to tensor<2x256x5120xf16> | |
| %extracted_slice_1270 = tensor.extract_slice %expanded_1268[0, 0, 5120] [2, 256, 5120] [1, 1, 1] : tensor<2x256x10240xf16> to tensor<2x256x5120xf16> | |
| %2618 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1270 : tensor<2x256x5120xf16>) outs(%1249 : tensor<2x256x5120xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.sqrt %cst_697 : f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| %4282 = math.erf %4281 : f16 | |
| %4283 = arith.addf %4282, %cst_695 : f16 | |
| %4284 = arith.mulf %4283, %cst_698 : f16 | |
| %4285 = arith.mulf %in, %4284 : f16 | |
| linalg.yield %4285 : f16 | |
| } -> tensor<2x256x5120xf16> | |
| %2619 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1269, %2618 : tensor<2x256x5120xf16>, tensor<2x256x5120xf16>) outs(%1249 : tensor<2x256x5120xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x5120xf16> | |
| %2620 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_416 : tensor<1280x5120xf16>) outs(%1252 : tensor<5120x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<5120x1280xf16> | |
| %collapsed_1271 = tensor.collapse_shape %2619 [[0, 1], [2]] : tensor<2x256x5120xf16> into tensor<512x5120xf16> | |
| %2621 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2622 = linalg.matmul ins(%collapsed_1271, %2620 : tensor<512x5120xf16>, tensor<5120x1280xf16>) outs(%2621 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2623 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_417, %2622 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1137 : tensor<512x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<512x1280xf16> | |
| %expanded_1272 = tensor.expand_shape %2623 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16> | |
| %2624 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1272, %2598 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %expanded_1273 = tensor.expand_shape %2624 [[0], [1, 2], [3]] : tensor<2x256x1280xf16> into tensor<2x16x16x1280xf16> | |
| %2625 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1273 : tensor<2x16x16x1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2626 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_419 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2627 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2625, %cst_418 : tensor<2x1280x16x16xf16>, tensor<1280x1280x1x1xf16>) outs(%2626 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %2628 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2627, %2474 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2629 = tensor.empty() : tensor<2x1920x16x16xf16> | |
| %inserted_slice_1274 = tensor.insert_slice %2628 into %2629[0, 0, 0, 0] [2, 1280, 16, 16] [1, 1, 1, 1] : tensor<2x1280x16x16xf16> into tensor<2x1920x16x16xf16> | |
| %inserted_slice_1275 = tensor.insert_slice %1007 into %inserted_slice_1274[0, 1280, 0, 0] [2, 640, 16, 16] [1, 1, 1, 1] : tensor<2x640x16x16xf16> into tensor<2x1920x16x16xf16> | |
| %collapsed_1276 = tensor.collapse_shape %inserted_slice_1275 [[0], [1], [2, 3]] : tensor<2x1920x16x16xf16> into tensor<2x1920x256xf16> | |
| %expanded_1277 = tensor.expand_shape %collapsed_1276 [[0], [1, 2], [3]] : tensor<2x1920x256xf16> into tensor<2x32x60x256xf16> | |
| %2630 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2631 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2630 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %2632 = tensor.empty() : tensor<2x32x60x256xf32> | |
| %2633 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2631 : tensor<f32>) outs(%2632 : tensor<2x32x60x256xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x60x256xf32> | |
| %2634 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1277 : tensor<2x32x60x256xf16>) outs(%2633 : tensor<2x32x60x256xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x60x256xf32> | |
| %2635 = tensor.empty() : tensor<2x32x60x256xf64> | |
| %2636 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2634 : tensor<2x32x60x256xf32>) outs(%2635 : tensor<2x32x60x256xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x60x256xf64> | |
| %2637 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2638 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2636 : tensor<2x32x60x256xf64>) outs(%2637 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2639 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2638 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_715 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2640 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2636, %2639 : tensor<2x32x60x256xf64>, tensor<2x32x1x1xf64>) outs(%2635 : tensor<2x32x60x256xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x60x256xf64> | |
| %2641 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2640 : tensor<2x32x60x256xf64>) outs(%2635 : tensor<2x32x60x256xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x60x256xf64> | |
| %2642 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2643 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2641 : tensor<2x32x60x256xf64>) outs(%2642 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2644 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2643 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_715 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2645 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2644 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2646 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %2647 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2634 : tensor<2x32x60x256xf32>) outs(%2646 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2648 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2647 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_716 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2649 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2645, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2650 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2649 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2651 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1277, %2648 : tensor<2x32x60x256xf16>, tensor<2x32x1x1xf32>) outs(%2632 : tensor<2x32x60x256xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x60x256xf32> | |
| %2652 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2651, %2650 : tensor<2x32x60x256xf32>, tensor<2x32x1x1xf32>) outs(%2632 : tensor<2x32x60x256xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x60x256xf32> | |
| %collapsed_1278 = tensor.collapse_shape %2652 [[0], [1, 2], [3]] : tensor<2x32x60x256xf32> into tensor<2x1920x256xf32> | |
| %expanded_1279 = tensor.expand_shape %collapsed_1278 [[0], [1], [2, 3]] : tensor<2x1920x256xf32> into tensor<2x1920x16x16xf32> | |
| %expanded_1280 = tensor.expand_shape %cst_420 [[0, 1, 2]] : tensor<1920xf16> into tensor<1920x1x1xf16> | |
| %2653 = tensor.empty() : tensor<2x1920x16x16xf32> | |
| %2654 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1279, %expanded_1280 : tensor<2x1920x16x16xf32>, tensor<1920x1x1xf16>) outs(%2653 : tensor<2x1920x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1920x16x16xf32> | |
| %expanded_1281 = tensor.expand_shape %cst_421 [[0, 1, 2]] : tensor<1920xf16> into tensor<1920x1x1xf16> | |
| %2655 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2654, %expanded_1281 : tensor<2x1920x16x16xf32>, tensor<1920x1x1xf16>) outs(%2653 : tensor<2x1920x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1920x16x16xf32> | |
| %2656 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2657 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2656 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %2658 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2657 : tensor<f16>) outs(%2629 : tensor<2x1920x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1920x16x16xf16> | |
| %2659 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2655 : tensor<2x1920x16x16xf32>) outs(%2658 : tensor<2x1920x16x16xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1920x16x16xf16> | |
| %2660 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2659 : tensor<2x1920x16x16xf16>) outs(%2629 : tensor<2x1920x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1920x16x16xf16> | |
| %2661 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2660, %2659 : tensor<2x1920x16x16xf16>, tensor<2x1920x16x16xf16>) outs(%2629 : tensor<2x1920x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1920x16x16xf16> | |
| %padded_1282 = tensor.pad %2661 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1920x16x16xf16> to tensor<2x1920x18x18xf16> | |
| %2662 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_423 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2663 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1282, %cst_422 : tensor<2x1920x18x18xf16>, tensor<1280x1920x3x3xf16>) outs(%2662 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %2664 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280xf16> | |
| %2665 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%2664, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %2666 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_424 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %2667 = linalg.fill ins(%cst_694 : f16) outs(%30 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %2668 = linalg.matmul ins(%2665, %2666 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%2667 : tensor<2x1280xf16>) -> tensor<2x1280xf16> | |
| %2669 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_425, %2668 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280xf16> | |
| %expanded_1283 = tensor.expand_shape %2669 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16> | |
| %2670 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2663, %expanded_1283 : tensor<2x1280x16x16xf16>, tensor<2x1280x1x1xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %collapsed_1284 = tensor.collapse_shape %2670 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16> | |
| %expanded_1285 = tensor.expand_shape %collapsed_1284 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16> | |
| %2671 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2672 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2671 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %2673 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2672 : tensor<f32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2674 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1285 : tensor<2x32x40x256xf16>) outs(%2673 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2675 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2674 : tensor<2x32x40x256xf32>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2676 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2677 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2675 : tensor<2x32x40x256xf64>) outs(%2676 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2678 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2677 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2679 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2675, %2678 : tensor<2x32x40x256xf64>, tensor<2x32x1x1xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2680 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2679 : tensor<2x32x40x256xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2681 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2682 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2680 : tensor<2x32x40x256xf64>) outs(%2681 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2683 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2682 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2684 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2683 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2685 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %2686 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2674 : tensor<2x32x40x256xf32>) outs(%2685 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2687 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2686 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_706 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2688 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2684, %cst_4 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2689 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2688 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2690 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1285, %2687 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2691 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2690, %2689 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %collapsed_1286 = tensor.collapse_shape %2691 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32> | |
| %expanded_1287 = tensor.expand_shape %collapsed_1286 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32> | |
| %expanded_1288 = tensor.expand_shape %cst_426 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2692 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1287, %expanded_1288 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %expanded_1289 = tensor.expand_shape %cst_427 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2693 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2692, %expanded_1289 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %2694 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2695 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2694 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %2696 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2695 : tensor<f16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2697 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2693 : tensor<2x1280x16x16xf32>) outs(%2696 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2698 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2697 : tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.negf %in : f16 | |
| %4281 = math.exp %4280 : f16 | |
| %4282 = arith.addf %4281, %cst_695 : f16 | |
| %4283 = arith.divf %cst_695, %4282 : f16 | |
| linalg.yield %4283 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2699 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2698, %2697 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %padded_1290 = tensor.pad %2699 low[0, 0, 1, 1] high[0, 0, 1, 1] { | |
| ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index): | |
| tensor.yield %cst_694 : f16 | |
| } : tensor<2x1280x16x16xf16> to tensor<2x1280x18x18xf16> | |
| %2700 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_429 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2701 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1290, %cst_428 : tensor<2x1280x18x18xf16>, tensor<1280x1280x3x3xf16>) outs(%2700 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %2702 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_431 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2703 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%inserted_slice_1275, %cst_430 : tensor<2x1920x16x16xf16>, tensor<1280x1920x1x1xf16>) outs(%2702 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %2704 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2703, %2701 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2705 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2704, %cst_3 : tensor<2x1280x16x16xf16>, tensor<f64>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.divf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %collapsed_1291 = tensor.collapse_shape %2705 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16> | |
| %expanded_1292 = tensor.expand_shape %collapsed_1291 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16> | |
| %2706 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2707 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2706 : tensor<f64>) outs(%9 : tensor<f32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<f32> | |
| %2708 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2707 : tensor<f32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| linalg.yield %in : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2709 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1292 : tensor<2x32x40x256xf16>) outs(%2708 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2710 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2709 : tensor<2x32x40x256xf32>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f32, %out: f64): | |
| %4280 = arith.extf %in : f32 to f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2711 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2712 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2710 : tensor<2x32x40x256xf64>) outs(%2711 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2713 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2712 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2714 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2710, %2713 : tensor<2x32x40x256xf64>, tensor<2x32x1x1xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %in_1640: f64, %out: f64): | |
| %4280 = arith.subf %in, %in_1640 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2715 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2714 : tensor<2x32x40x256xf64>) outs(%1055 : tensor<2x32x40x256xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.mulf %in, %in : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x40x256xf64> | |
| %2716 = linalg.fill ins(%cst_699 : f64) outs(%51 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64> | |
| %2717 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2715 : tensor<2x32x40x256xf64>) outs(%2716 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.addf %in, %out : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2718 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2717 : tensor<2x32x1x1xf64>) outs(%51 : tensor<2x32x1x1xf64>) { | |
| ^bb0(%in: f64, %out: f64): | |
| %4280 = arith.divf %in, %cst_705 : f64 | |
| linalg.yield %4280 : f64 | |
| } -> tensor<2x32x1x1xf64> | |
| %2719 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2718 : tensor<2x32x1x1xf64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f64, %out: f32): | |
| %4280 = arith.truncf %in : f64 to f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2720 = linalg.fill ins(%cst_701 : f32) outs(%60 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32> | |
| %2721 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2709 : tensor<2x32x40x256xf32>) outs(%2720 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.addf %in, %out : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2722 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2721 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = arith.divf %in, %cst_706 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2723 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2719, %cst_2 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %in_1640: f64, %out: f32): | |
| %4280 = arith.truncf %in_1640 : f64 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2724 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2723 : tensor<2x32x1x1xf32>) outs(%60 : tensor<2x32x1x1xf32>) { | |
| ^bb0(%in: f32, %out: f32): | |
| %4280 = math.rsqrt %in : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x1x1xf32> | |
| %2725 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1292, %2722 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f16, %in_1640: f32, %out: f32): | |
| %4280 = arith.extf %in : f16 to f32 | |
| %4281 = arith.subf %4280, %in_1640 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %2726 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2725, %2724 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%1052 : tensor<2x32x40x256xf32>) { | |
| ^bb0(%in: f32, %in_1640: f32, %out: f32): | |
| %4280 = arith.mulf %in, %in_1640 : f32 | |
| linalg.yield %4280 : f32 | |
| } -> tensor<2x32x40x256xf32> | |
| %collapsed_1293 = tensor.collapse_shape %2726 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32> | |
| %expanded_1294 = tensor.expand_shape %collapsed_1293 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32> | |
| %expanded_1295 = tensor.expand_shape %cst_432 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2727 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1294, %expanded_1295 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.mulf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %expanded_1296 = tensor.expand_shape %cst_433 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16> | |
| %2728 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2727, %expanded_1296 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%1073 : tensor<2x1280x16x16xf32>) { | |
| ^bb0(%in: f32, %in_1640: f16, %out: f32): | |
| %4280 = arith.extf %in_1640 : f16 to f32 | |
| %4281 = arith.addf %in, %4280 : f32 | |
| linalg.yield %4281 : f32 | |
| } -> tensor<2x1280x16x16xf32> | |
| %2729 = linalg.fill ins(%cst_699 : f64) outs(%7 : tensor<f64>) -> tensor<f64> | |
| %2730 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2729 : tensor<f64>) outs(%23 : tensor<f16>) { | |
| ^bb0(%in: f64, %out: f16): | |
| %4280 = arith.truncf %in : f64 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<f16> | |
| %2731 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2730 : tensor<f16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2732 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2728 : tensor<2x1280x16x16xf32>) outs(%2731 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f32, %out: f16): | |
| %4280 = arith.truncf %in : f32 to f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2733 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_435 : tensor<1280xf16>) outs(%1040 : tensor<2x1280x16x16xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x1280x16x16xf16> | |
| %2734 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2732, %cst_434 : tensor<2x1280x16x16xf16>, tensor<1280x1280x1x1xf16>) outs(%2733 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16> | |
| %2735 = linalg.generic {indexing_maps = [#map11, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2734 : tensor<2x1280x16x16xf16>) outs(%1117 : tensor<2x16x16x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x16x16x1280xf16> | |
| %collapsed_1297 = tensor.collapse_shape %2735 [[0], [1, 2], [3]] : tensor<2x16x16x1280xf16> into tensor<2x256x1280xf16> | |
| %2736 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %2737 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_1297 : tensor<2x256x1280xf16>) outs(%2736 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2738 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2737 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2739 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2738 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2740 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1297, %2739 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2741 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2740 : tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2742 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %2743 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2741 : tensor<2x256x1280xf16>) outs(%2742 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2744 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2743 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2745 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2744 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2746 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2745 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2747 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2746 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2748 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2740, %2747 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2749 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2748, %cst_436 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2750 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2749, %cst_437 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2751 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_438 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1298 = tensor.collapse_shape %2750 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16> | |
| %2752 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2753 = linalg.matmul ins(%collapsed_1298, %2751 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2752 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2754 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_439 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %2755 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2756 = linalg.matmul ins(%collapsed_1298, %2754 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2755 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2757 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_440 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %2758 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2759 = linalg.matmul ins(%collapsed_1298, %2757 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2758 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %expanded_1299 = tensor.expand_shape %2753 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %2760 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1299 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_1300 = tensor.collapse_shape %2760 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %expanded_1301 = tensor.expand_shape %2756 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %2761 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1301 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_1302 = tensor.collapse_shape %2761 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %expanded_1303 = tensor.expand_shape %2759 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %2762 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1303 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_1304 = tensor.collapse_shape %2762 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %2763 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1302 : tensor<16x256x160xf16>) outs(%1150 : tensor<16x160x256xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x160x256xf16> | |
| %2764 = linalg.fill ins(%cst_694 : f16) outs(%1152 : tensor<16x256x256xf16>) -> tensor<16x256x256xf16> | |
| %2765 = linalg.batch_matmul ins(%collapsed_1300, %2763 : tensor<16x256x160xf16>, tensor<16x160x256xf16>) outs(%2764 : tensor<16x256x256xf16>) -> tensor<16x256x256xf16> | |
| %2766 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2765, %cst : tensor<16x256x256xf16>, tensor<f64>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %2767 = linalg.fill ins(%c0_i64 : i64) outs(%1156 : tensor<16x256x1xi64>) -> tensor<16x256x1xi64> | |
| %2768 = linalg.fill ins(%cst_696 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %2769:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2766 : tensor<16x256x256xf16>) outs(%2768, %2767 : tensor<16x256x1xf16>, tensor<16x256x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x256x1xf16>, tensor<16x256x1xi64>) | |
| %2770 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2766, %2769#0 : tensor<16x256x256xf16>, tensor<16x256x1xf16>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %2771 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2770 : tensor<16x256x256xf16>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %2772 = linalg.fill ins(%cst_694 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %2773 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2771 : tensor<16x256x256xf16>) outs(%2772 : tensor<16x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x1xf16> | |
| %2774 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2771, %2773 : tensor<16x256x256xf16>, tensor<16x256x1xf16>) outs(%1152 : tensor<16x256x256xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x256xf16> | |
| %2775 = linalg.fill ins(%cst_694 : f16) outs(%1166 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %2776 = linalg.batch_matmul ins(%2774, %collapsed_1304 : tensor<16x256x256xf16>, tensor<16x256x160xf16>) outs(%2775 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %expanded_1305 = tensor.expand_shape %2776 [[0, 1], [2], [3]] : tensor<16x256x160xf16> into tensor<2x8x256x160xf16> | |
| %2777 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1305 : tensor<2x8x256x160xf16>) outs(%1169 : tensor<2x256x8x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x8x160xf16> | |
| %2778 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_441 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1306 = tensor.collapse_shape %2777 [[0, 1], [2, 3]] : tensor<2x256x8x160xf16> into tensor<512x1280xf16> | |
| %2779 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2780 = linalg.matmul ins(%collapsed_1306, %2778 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2779 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2781 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_442, %2780 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1137 : tensor<512x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<512x1280xf16> | |
| %expanded_1307 = tensor.expand_shape %2781 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16> | |
| %2782 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1307, %collapsed_1297 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2783 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %2784 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2782 : tensor<2x256x1280xf16>) outs(%2783 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2785 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2784 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2786 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2785 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2787 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2782, %2786 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2788 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2787 : tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2789 = linalg.fill ins(%cst_694 : f16) outs(%1119 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16> | |
| %2790 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2788 : tensor<2x256x1280xf16>) outs(%2789 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2791 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2790 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.divf %in, %cst_712 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2792 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2791 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.truncf %cst_700 : f64 to f16 | |
| %4281 = arith.addf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2793 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2792 : tensor<2x256x1xf16>) outs(%1119 : tensor<2x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.rsqrt %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1xf16> | |
| %2794 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2793 : tensor<2x256x1xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2795 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2787, %2794 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2796 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2795, %cst_443 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.mulf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2797 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2796, %cst_444 : tensor<2x256x1280xf16>, tensor<1280xf16>) outs(%1123 : tensor<2x256x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.addf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<2x256x1280xf16> | |
| %2798 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_445 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1308 = tensor.collapse_shape %2797 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16> | |
| %2799 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2800 = linalg.matmul ins(%collapsed_1308, %2798 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2799 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2801 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_446 : tensor<1280x768xf16>) outs(%1194 : tensor<768x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x1280xf16> | |
| %2802 = linalg.fill ins(%cst_694 : f16) outs(%1196 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %2803 = linalg.matmul ins(%collapsed_761, %2801 : tensor<154x768xf16>, tensor<768x1280xf16>) outs(%2802 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %2804 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_447 : tensor<1280x768xf16>) outs(%1194 : tensor<768x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<768x1280xf16> | |
| %2805 = linalg.fill ins(%cst_694 : f16) outs(%1196 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %2806 = linalg.matmul ins(%collapsed_761, %2804 : tensor<154x768xf16>, tensor<768x1280xf16>) outs(%2805 : tensor<154x1280xf16>) -> tensor<154x1280xf16> | |
| %expanded_1309 = tensor.expand_shape %2800 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x8x160xf16> | |
| %2807 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1309 : tensor<2x256x8x160xf16>) outs(%1146 : tensor<2x8x256x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x256x160xf16> | |
| %collapsed_1310 = tensor.collapse_shape %2807 [[0, 1], [2], [3]] : tensor<2x8x256x160xf16> into tensor<16x256x160xf16> | |
| %expanded_1311 = tensor.expand_shape %2803 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x8x160xf16> | |
| %2808 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1311 : tensor<2x77x8x160xf16>) outs(%1203 : tensor<2x8x77x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x160xf16> | |
| %collapsed_1312 = tensor.collapse_shape %2808 [[0, 1], [2], [3]] : tensor<2x8x77x160xf16> into tensor<16x77x160xf16> | |
| %expanded_1313 = tensor.expand_shape %2806 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x8x160xf16> | |
| %2809 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1313 : tensor<2x77x8x160xf16>) outs(%1203 : tensor<2x8x77x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x8x77x160xf16> | |
| %collapsed_1314 = tensor.collapse_shape %2809 [[0, 1], [2], [3]] : tensor<2x8x77x160xf16> into tensor<16x77x160xf16> | |
| %2810 = linalg.generic {indexing_maps = [#map16, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1312 : tensor<16x77x160xf16>) outs(%1206 : tensor<16x160x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<16x160x77xf16> | |
| %2811 = linalg.fill ins(%cst_694 : f16) outs(%1208 : tensor<16x256x77xf16>) -> tensor<16x256x77xf16> | |
| %2812 = linalg.batch_matmul ins(%collapsed_1310, %2810 : tensor<16x256x160xf16>, tensor<16x160x77xf16>) outs(%2811 : tensor<16x256x77xf16>) -> tensor<16x256x77xf16> | |
| %2813 = linalg.generic {indexing_maps = [#map16, #map21, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2812, %cst : tensor<16x256x77xf16>, tensor<f64>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f64, %out: f16): | |
| %4280 = arith.truncf %in_1640 : f64 to f16 | |
| %4281 = arith.mulf %in, %4280 : f16 | |
| linalg.yield %4281 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %2814 = linalg.fill ins(%c0_i64 : i64) outs(%1156 : tensor<16x256x1xi64>) -> tensor<16x256x1xi64> | |
| %2815 = linalg.fill ins(%cst_696 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %2816:2 = linalg.generic {indexing_maps = [#map16, #map17, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2813 : tensor<16x256x77xf16>) outs(%2815, %2814 : tensor<16x256x1xf16>, tensor<16x256x1xi64>) { | |
| ^bb0(%in: f16, %out: f16, %out_1640: i64): | |
| %4280 = linalg.index 2 : index | |
| %4281 = arith.index_cast %4280 : index to i64 | |
| %4282 = arith.maxf %in, %out : f16 | |
| %4283 = arith.cmpf ogt, %in, %out : f16 | |
| %4284 = arith.select %4283, %4281, %out_1640 : i64 | |
| linalg.yield %4282, %4284 : f16, i64 | |
| } -> (tensor<16x256x1xf16>, tensor<16x256x1xi64>) | |
| %2817 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2813, %2816#0 : tensor<16x256x77xf16>, tensor<16x256x1xf16>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.subf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %2818 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2817 : tensor<16x256x77xf16>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = math.exp %in : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %2819 = linalg.fill ins(%cst_694 : f16) outs(%1158 : tensor<16x256x1xf16>) -> tensor<16x256x1xf16> | |
| %2820 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2818 : tensor<16x256x77xf16>) outs(%2819 : tensor<16x256x1xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| %4280 = arith.addf %in, %out : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x1xf16> | |
| %2821 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2818, %2820 : tensor<16x256x77xf16>, tensor<16x256x1xf16>) outs(%1208 : tensor<16x256x77xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
| %4280 = arith.divf %in, %in_1640 : f16 | |
| linalg.yield %4280 : f16 | |
| } -> tensor<16x256x77xf16> | |
| %2822 = linalg.fill ins(%cst_694 : f16) outs(%1166 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %2823 = linalg.batch_matmul ins(%2821, %collapsed_1314 : tensor<16x256x77xf16>, tensor<16x77x160xf16>) outs(%2822 : tensor<16x256x160xf16>) -> tensor<16x256x160xf16> | |
| %expanded_1315 = tensor.expand_shape %2823 [[0, 1], [2], [3]] : tensor<16x256x160xf16> into tensor<2x8x256x160xf16> | |
| %2824 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1315 : tensor<2x8x256x160xf16>) outs(%1169 : tensor<2x256x8x160xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<2x256x8x160xf16> | |
| %2825 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_448 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) { | |
| ^bb0(%in: f16, %out: f16): | |
| linalg.yield %in : f16 | |
| } -> tensor<1280x1280xf16> | |
| %collapsed_1316 = tensor.collapse_shape %2824 [[0, 1], [2, 3]] : tensor<2x256x8x160xf16> into tensor<512x1280xf16> | |
| %2826 = linalg.fill ins(%cst_694 : f16) outs(%1137 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2827 = linalg.matmul ins(%collapsed_1316, %2825 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%2826 : tensor<512x1280xf16>) -> tensor<512x1280xf16> | |
| %2828 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_449, %2827 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1137 : tensor<512x1280xf16>) { | |
| ^bb0(%in: f16, %in_1640: f16, %out: f16): | |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment