Skip to content

Instantly share code, notes, and snippets.

@pashu123

pashu123/quant_elided.mlir

Created October 28, 2022 12:59
Show Gist options
  • Save pashu123/535af10a878cc7fd92c2e6f1f0859d76 to your computer and use it in GitHub Desktop.
Save pashu123/535af10a878cc7fd92c2e6f1f0859d76 to your computer and use it in GitHub Desktop.
Download ZIP
Raw
quant_elided.mlir
This file has been truncated, but you can view the full file.
#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, d4) -> (d0, d1, d2, d3, d4)>
#map16 = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, 0, d3, d4)>
#map17 = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, 0, 0, 0)>
#map18 = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, 0, 0)>
#map19 = affine_map<(d0, d1, d2, d3) -> (d0, d2, d3, d1)>
#map20 = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
#map21 = affine_map<(d0, d1, d2) -> (d0, d1, 0)>
#map22 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, 0)>
#map23 = affine_map<(d0, d1, d2) -> ()>
#map24 = affine_map<(d0, d1, d2, d3) -> (d0, d2, d1, d3)>
#map25 = affine_map<(d0, d1, d2) -> (d0, d2, d1)>
#map26 = 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<i64>, %arg2: tensor<2x77x768xf16>) -> tensor<2x4x64x64xf16> {
%cst = arith.constant dense<7.904050e-02> : tensor<f16>
%cst_0 = arith.constant dense<1.118160e-01> : tensor<f16>
%cst_1 = arith.constant dense<1.580810e-01> : tensor<f16>
%cst_2 = arith.constant dense<1.013280e-06> : tensor<f16>
%cst_3 = arith.constant dense<1.000000e+00> : tensor<f16>
%cst_4 = arith.constant dense<1.001360e-05> : tensor<f16>
%cst_5 = arith.constant dense<1> : tensor<i64>
%cst_6 = arith.constant dense<160> : tensor<i64>
%cst_7 = arith.constant dense<-9.210930e+00> : tensor<f16>
%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<320x20x1x3x3xf16>
%cst_17 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xi8>
%cst_18 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_19 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16>
%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<320xf16>
%cst_23 = arith.constant dense_resource<__elided__> : tensor<320x20x1x3x3xf16>
%cst_24 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xi8>
%cst_25 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_26 = arith.constant dense_resource<__elided__> : tensor<320x20x1x1x1xf16>
%cst_27 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xi8>
%cst_28 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_29 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_30 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_31 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_32 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_33 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_34 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_35 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_36 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_37 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_38 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_39 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_40 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_41 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_42 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_43 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_44 = arith.constant dense_resource<__elided__> : tensor<320x48x1xf16>
%cst_45 = arith.constant dense_resource<__elided__> : tensor<320x768xi8>
%cst_46 = arith.constant dense_resource<__elided__> : tensor<320x48x1xf16>
%cst_47 = arith.constant dense_resource<__elided__> : tensor<320x768xi8>
%cst_48 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_49 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_50 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_51 = arith.constant dense_resource<__elided__> : tensor<2560x20x1xf16>
%cst_52 = arith.constant dense_resource<__elided__> : tensor<2560x320xi8>
%cst_53 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_54 = arith.constant dense_resource<__elided__> : tensor<320x80x1xf16>
%cst_55 = arith.constant dense_resource<__elided__> : tensor<320x1280xi8>
%cst_56 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_57 = arith.constant dense_resource<__elided__> : tensor<320x20x1x1x1xf16>
%cst_58 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xi8>
%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<320x20x1x3x3xf16>
%cst_63 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xi8>
%cst_64 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_65 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16>
%cst_66 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_67 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_68 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_69 = arith.constant dense_resource<__elided__> : tensor<320x20x1x3x3xf16>
%cst_70 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xi8>
%cst_71 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_72 = arith.constant dense_resource<__elided__> : tensor<320x20x1x1x1xf16>
%cst_73 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xi8>
%cst_74 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_75 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_76 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_77 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_78 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_79 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_80 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_81 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_82 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_83 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_84 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_85 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_86 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_87 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_88 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_89 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_90 = arith.constant dense_resource<__elided__> : tensor<320x48x1xf16>
%cst_91 = arith.constant dense_resource<__elided__> : tensor<320x768xi8>
%cst_92 = arith.constant dense_resource<__elided__> : tensor<320x48x1xf16>
%cst_93 = arith.constant dense_resource<__elided__> : tensor<320x768xi8>
%cst_94 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_95 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_96 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_97 = arith.constant dense_resource<__elided__> : tensor<2560x20x1xf16>
%cst_98 = arith.constant dense_resource<__elided__> : tensor<2560x320xi8>
%cst_99 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_100 = arith.constant dense_resource<__elided__> : tensor<320x80x1xf16>
%cst_101 = arith.constant dense_resource<__elided__> : tensor<320x1280xi8>
%cst_102 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_103 = arith.constant dense_resource<__elided__> : tensor<320x20x1x1x1xf16>
%cst_104 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xi8>
%cst_105 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_106 = arith.constant dense_resource<__elided__> : tensor<320x20x1x3x3xf16>
%cst_107 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xi8>
%cst_108 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_109 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_110 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_111 = arith.constant dense_resource<__elided__> : tensor<640x20x1x3x3xf16>
%cst_112 = arith.constant dense_resource<__elided__> : tensor<640x320x3x3xi8>
%cst_113 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_114 = arith.constant dense_resource<__elided__> : tensor<640x1280xf16>
%cst_115 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_116 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_117 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_118 = arith.constant dense_resource<__elided__> : tensor<640x40x1x3x3xf16>
%cst_119 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xi8>
%cst_120 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_121 = arith.constant dense_resource<__elided__> : tensor<640x20x1x1x1xf16>
%cst_122 = arith.constant dense_resource<__elided__> : tensor<640x320x1x1xi8>
%cst_123 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_124 = arith.constant dense_resource<__elided__> : tensor<640x40x1x1x1xf16>
%cst_125 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xi8>
%cst_126 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_127 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_128 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_129 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_130 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_131 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_132 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_133 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_134 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_135 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_136 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_137 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_138 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_139 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_140 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_141 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_142 = arith.constant dense_resource<__elided__> : tensor<640x48x1xf16>
%cst_143 = arith.constant dense_resource<__elided__> : tensor<640x768xi8>
%cst_144 = arith.constant dense_resource<__elided__> : tensor<640x48x1xf16>
%cst_145 = arith.constant dense_resource<__elided__> : tensor<640x768xi8>
%cst_146 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_147 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_148 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_149 = arith.constant dense_resource<__elided__> : tensor<5120x40x1xf16>
%cst_150 = arith.constant dense_resource<__elided__> : tensor<5120x640xi8>
%cst_151 = arith.constant dense_resource<__elided__> : tensor<5120xf16>
%cst_152 = arith.constant dense_resource<__elided__> : tensor<640x160x1xf16>
%cst_153 = arith.constant dense_resource<__elided__> : tensor<640x2560xi8>
%cst_154 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_155 = arith.constant dense_resource<__elided__> : tensor<640x40x1x1x1xf16>
%cst_156 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xi8>
%cst_157 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_158 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_159 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_160 = arith.constant dense_resource<__elided__> : tensor<640x40x1x3x3xf16>
%cst_161 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xi8>
%cst_162 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_163 = arith.constant dense_resource<__elided__> : tensor<640x1280xf16>
%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<640xf16>
%cst_167 = arith.constant dense_resource<__elided__> : tensor<640x40x1x3x3xf16>
%cst_168 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xi8>
%cst_169 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_170 = arith.constant dense_resource<__elided__> : tensor<640x40x1x1x1xf16>
%cst_171 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xi8>
%cst_172 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_173 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_174 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_175 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_176 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_177 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_178 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_179 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_180 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_181 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_182 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_183 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_184 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_185 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_186 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_187 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_188 = arith.constant dense_resource<__elided__> : tensor<640x48x1xf16>
%cst_189 = arith.constant dense_resource<__elided__> : tensor<640x768xi8>
%cst_190 = arith.constant dense_resource<__elided__> : tensor<640x48x1xf16>
%cst_191 = arith.constant dense_resource<__elided__> : tensor<640x768xi8>
%cst_192 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_193 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_194 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_195 = arith.constant dense_resource<__elided__> : tensor<5120x40x1xf16>
%cst_196 = arith.constant dense_resource<__elided__> : tensor<5120x640xi8>
%cst_197 = arith.constant dense_resource<__elided__> : tensor<5120xf16>
%cst_198 = arith.constant dense_resource<__elided__> : tensor<640x160x1xf16>
%cst_199 = arith.constant dense_resource<__elided__> : tensor<640x2560xi8>
%cst_200 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_201 = arith.constant dense_resource<__elided__> : tensor<640x40x1x1x1xf16>
%cst_202 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xi8>
%cst_203 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_204 = arith.constant dense_resource<__elided__> : tensor<640x40x1x3x3xf16>
%cst_205 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xi8>
%cst_206 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_207 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_208 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_209 = arith.constant dense_resource<__elided__> : tensor<1280x40x1x3x3xf16>
%cst_210 = arith.constant dense_resource<__elided__> : tensor<1280x640x3x3xi8>
%cst_211 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_212 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_213 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_214 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_215 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_216 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_217 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_218 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_219 = arith.constant dense_resource<__elided__> : tensor<1280x40x1x1x1xf16>
%cst_220 = arith.constant dense_resource<__elided__> : tensor<1280x640x1x1xi8>
%cst_221 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_222 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x1x1xf16>
%cst_223 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xi8>
%cst_224 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_225 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_226 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_227 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_228 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_229 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_230 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_231 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_232 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_233 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_234 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_235 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_236 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_237 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_238 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_239 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_240 = arith.constant dense_resource<__elided__> : tensor<1280x48x1xf16>
%cst_241 = arith.constant dense_resource<__elided__> : tensor<1280x768xi8>
%cst_242 = arith.constant dense_resource<__elided__> : tensor<1280x48x1xf16>
%cst_243 = arith.constant dense_resource<__elided__> : tensor<1280x768xi8>
%cst_244 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_245 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_246 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_247 = arith.constant dense_resource<__elided__> : tensor<10240x80x1xf16>
%cst_248 = arith.constant dense_resource<__elided__> : tensor<10240x1280xi8>
%cst_249 = arith.constant dense_resource<__elided__> : tensor<10240xf16>
%cst_250 = arith.constant dense_resource<__elided__> : tensor<1280x320x1xf16>
%cst_251 = arith.constant dense_resource<__elided__> : tensor<1280x5120xi8>
%cst_252 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_253 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x1x1xf16>
%cst_254 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xi8>
%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<1280x80x1x3x3xf16>
%cst_259 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_260 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_261 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_262 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_263 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_264 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_265 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_266 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_267 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_268 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x1x1xf16>
%cst_269 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xi8>
%cst_270 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_271 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_272 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_273 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_274 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_275 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_276 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_277 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_278 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_279 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_280 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_281 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_282 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_283 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_284 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_285 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_286 = arith.constant dense_resource<__elided__> : tensor<1280x48x1xf16>
%cst_287 = arith.constant dense_resource<__elided__> : tensor<1280x768xi8>
%cst_288 = arith.constant dense_resource<__elided__> : tensor<1280x48x1xf16>
%cst_289 = arith.constant dense_resource<__elided__> : tensor<1280x768xi8>
%cst_290 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_291 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_292 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_293 = arith.constant dense_resource<__elided__> : tensor<10240x80x1xf16>
%cst_294 = arith.constant dense_resource<__elided__> : tensor<10240x1280xi8>
%cst_295 = arith.constant dense_resource<__elided__> : tensor<10240xf16>
%cst_296 = arith.constant dense_resource<__elided__> : tensor<1280x320x1xf16>
%cst_297 = arith.constant dense_resource<__elided__> : tensor<1280x5120xi8>
%cst_298 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_299 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x1x1xf16>
%cst_300 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xi8>
%cst_301 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_302 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_303 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_304 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_305 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_306 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_307 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_308 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%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<1280x80x1x3x3xf16>
%cst_315 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_316 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_317 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_318 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_319 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_320 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%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<1280x80x1x3x3xf16>
%cst_327 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_328 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_329 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_330 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_331 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_332 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%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<1280x80x1x3x3xf16>
%cst_339 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_340 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_341 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x1x1xf16>
%cst_342 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xi8>
%cst_343 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_344 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_345 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_346 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_347 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_348 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_349 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_350 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_351 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_352 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_353 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_354 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_355 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_356 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_357 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_358 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_359 = arith.constant dense_resource<__elided__> : tensor<1280x48x1xf16>
%cst_360 = arith.constant dense_resource<__elided__> : tensor<1280x768xi8>
%cst_361 = arith.constant dense_resource<__elided__> : tensor<1280x48x1xf16>
%cst_362 = arith.constant dense_resource<__elided__> : tensor<1280x768xi8>
%cst_363 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_364 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_365 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_366 = arith.constant dense_resource<__elided__> : tensor<10240x80x1xf16>
%cst_367 = arith.constant dense_resource<__elided__> : tensor<10240x1280xi8>
%cst_368 = arith.constant dense_resource<__elided__> : tensor<10240xf16>
%cst_369 = arith.constant dense_resource<__elided__> : tensor<1280x320x1xf16>
%cst_370 = arith.constant dense_resource<__elided__> : tensor<1280x5120xi8>
%cst_371 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_372 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x1x1xf16>
%cst_373 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xi8>
%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<1280xf16>
%cst_377 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_378 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_379 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_380 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_381 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_382 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_383 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_384 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_385 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_386 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_387 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_388 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_389 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x3x3xf16>
%cst_390 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xi8>
%cst_391 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_392 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%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<1280x80x1x3x3xf16>
%cst_397 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_398 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_399 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x1x1xf16>
%cst_400 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xi8>
%cst_401 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_402 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_403 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_404 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x3x3xf16>
%cst_405 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xi8>
%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<1280xf16>
%cst_409 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_410 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_411 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_412 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_413 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_414 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x1x1xf16>
%cst_415 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xi8>
%cst_416 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_417 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_418 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_419 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x3x3xf16>
%cst_420 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xi8>
%cst_421 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_422 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_423 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_424 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_425 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_426 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_427 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_428 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_429 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x1x1xf16>
%cst_430 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xi8>
%cst_431 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_432 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_433 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_434 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_435 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_436 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_437 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x3x3xf16>
%cst_438 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xi8>
%cst_439 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_440 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_441 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%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<1280x80x1x3x3xf16>
%cst_445 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_446 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_447 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x1x1xf16>
%cst_448 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xi8>
%cst_449 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_450 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x1x1xf16>
%cst_451 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xi8>
%cst_452 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_453 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_454 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_455 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_456 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_457 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_458 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_459 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_460 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_461 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_462 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_463 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_464 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_465 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_466 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_467 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_468 = arith.constant dense_resource<__elided__> : tensor<1280x48x1xf16>
%cst_469 = arith.constant dense_resource<__elided__> : tensor<1280x768xi8>
%cst_470 = arith.constant dense_resource<__elided__> : tensor<1280x48x1xf16>
%cst_471 = arith.constant dense_resource<__elided__> : tensor<1280x768xi8>
%cst_472 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_473 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_474 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_475 = arith.constant dense_resource<__elided__> : tensor<10240x80x1xf16>
%cst_476 = arith.constant dense_resource<__elided__> : tensor<10240x1280xi8>
%cst_477 = arith.constant dense_resource<__elided__> : tensor<10240xf16>
%cst_478 = arith.constant dense_resource<__elided__> : tensor<1280x320x1xf16>
%cst_479 = arith.constant dense_resource<__elided__> : tensor<1280x5120xi8>
%cst_480 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_481 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x1x1xf16>
%cst_482 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xi8>
%cst_483 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_484 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_485 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_486 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x3x3xf16>
%cst_487 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xi8>
%cst_488 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_489 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_490 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_491 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_492 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_493 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_494 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_495 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_496 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x1x1xf16>
%cst_497 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xi8>
%cst_498 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_499 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x1x1xf16>
%cst_500 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xi8>
%cst_501 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_502 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_503 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_504 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_505 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_506 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_507 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_508 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_509 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_510 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_511 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_512 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_513 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_514 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_515 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_516 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_517 = arith.constant dense_resource<__elided__> : tensor<1280x48x1xf16>
%cst_518 = arith.constant dense_resource<__elided__> : tensor<1280x768xi8>
%cst_519 = arith.constant dense_resource<__elided__> : tensor<1280x48x1xf16>
%cst_520 = arith.constant dense_resource<__elided__> : tensor<1280x768xi8>
%cst_521 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_522 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_523 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_524 = arith.constant dense_resource<__elided__> : tensor<10240x80x1xf16>
%cst_525 = arith.constant dense_resource<__elided__> : tensor<10240x1280xi8>
%cst_526 = arith.constant dense_resource<__elided__> : tensor<10240xf16>
%cst_527 = arith.constant dense_resource<__elided__> : tensor<1280x320x1xf16>
%cst_528 = arith.constant dense_resource<__elided__> : tensor<1280x5120xi8>
%cst_529 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_530 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x1x1xf16>
%cst_531 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xi8>
%cst_532 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_533 = arith.constant dense_resource<__elided__> : tensor<1920xf16>
%cst_534 = arith.constant dense_resource<__elided__> : tensor<1920xf16>
%cst_535 = arith.constant dense_resource<__elided__> : tensor<1280x120x1x3x3xf16>
%cst_536 = arith.constant dense_resource<__elided__> : tensor<1280x1920x3x3xi8>
%cst_537 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_538 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_539 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_540 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_541 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_542 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_543 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_544 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_545 = arith.constant dense_resource<__elided__> : tensor<1280x120x1x1x1xf16>
%cst_546 = arith.constant dense_resource<__elided__> : tensor<1280x1920x1x1xi8>
%cst_547 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_548 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x1x1xf16>
%cst_549 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xi8>
%cst_550 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_551 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_552 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_553 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_554 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_555 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_556 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_557 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_558 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_559 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_560 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_561 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_562 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_563 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_564 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_565 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_566 = arith.constant dense_resource<__elided__> : tensor<1280x48x1xf16>
%cst_567 = arith.constant dense_resource<__elided__> : tensor<1280x768xi8>
%cst_568 = arith.constant dense_resource<__elided__> : tensor<1280x48x1xf16>
%cst_569 = arith.constant dense_resource<__elided__> : tensor<1280x768xi8>
%cst_570 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_571 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_572 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_573 = arith.constant dense_resource<__elided__> : tensor<10240x80x1xf16>
%cst_574 = arith.constant dense_resource<__elided__> : tensor<10240x1280xi8>
%cst_575 = arith.constant dense_resource<__elided__> : tensor<10240xf16>
%cst_576 = arith.constant dense_resource<__elided__> : tensor<1280x320x1xf16>
%cst_577 = arith.constant dense_resource<__elided__> : tensor<1280x5120xi8>
%cst_578 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_579 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x1x1xf16>
%cst_580 = arith.constant dense_resource<__elided__> : tensor<1280x1280x1x1xi8>
%cst_581 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_582 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_583 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_584 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_585 = arith.constant dense_resource<__elided__> : tensor<1920xf16>
%cst_586 = arith.constant dense_resource<__elided__> : tensor<1920xf16>
%cst_587 = arith.constant dense_resource<__elided__> : tensor<640x120x1x3x3xf16>
%cst_588 = arith.constant dense_resource<__elided__> : tensor<640x1920x3x3xi8>
%cst_589 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_590 = arith.constant dense_resource<__elided__> : tensor<640x1280xf16>
%cst_591 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_592 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_593 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_594 = arith.constant dense_resource<__elided__> : tensor<640x40x1x3x3xf16>
%cst_595 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xi8>
%cst_596 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_597 = arith.constant dense_resource<__elided__> : tensor<640x120x1x1x1xf16>
%cst_598 = arith.constant dense_resource<__elided__> : tensor<640x1920x1x1xi8>
%cst_599 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_600 = arith.constant dense_resource<__elided__> : tensor<640x40x1x1x1xf16>
%cst_601 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xi8>
%cst_602 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_603 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_604 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_605 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_606 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_607 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_608 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_609 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_610 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_611 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_612 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_613 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_614 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_615 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_616 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_617 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_618 = arith.constant dense_resource<__elided__> : tensor<640x48x1xf16>
%cst_619 = arith.constant dense_resource<__elided__> : tensor<640x768xi8>
%cst_620 = arith.constant dense_resource<__elided__> : tensor<640x48x1xf16>
%cst_621 = arith.constant dense_resource<__elided__> : tensor<640x768xi8>
%cst_622 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_623 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_624 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_625 = arith.constant dense_resource<__elided__> : tensor<5120x40x1xf16>
%cst_626 = arith.constant dense_resource<__elided__> : tensor<5120x640xi8>
%cst_627 = arith.constant dense_resource<__elided__> : tensor<5120xf16>
%cst_628 = arith.constant dense_resource<__elided__> : tensor<640x160x1xf16>
%cst_629 = arith.constant dense_resource<__elided__> : tensor<640x2560xi8>
%cst_630 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_631 = arith.constant dense_resource<__elided__> : tensor<640x40x1x1x1xf16>
%cst_632 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xi8>
%cst_633 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_634 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_635 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_636 = arith.constant dense_resource<__elided__> : tensor<640x80x1x3x3xf16>
%cst_637 = arith.constant dense_resource<__elided__> : tensor<640x1280x3x3xi8>
%cst_638 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_639 = arith.constant dense_resource<__elided__> : tensor<640x1280xf16>
%cst_640 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_641 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_642 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_643 = arith.constant dense_resource<__elided__> : tensor<640x40x1x3x3xf16>
%cst_644 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xi8>
%cst_645 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_646 = arith.constant dense_resource<__elided__> : tensor<640x80x1x1x1xf16>
%cst_647 = arith.constant dense_resource<__elided__> : tensor<640x1280x1x1xi8>
%cst_648 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_649 = arith.constant dense_resource<__elided__> : tensor<640x40x1x1x1xf16>
%cst_650 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xi8>
%cst_651 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_652 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_653 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_654 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_655 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_656 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_657 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_658 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_659 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_660 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_661 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_662 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_663 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_664 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_665 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_666 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_667 = arith.constant dense_resource<__elided__> : tensor<640x48x1xf16>
%cst_668 = arith.constant dense_resource<__elided__> : tensor<640x768xi8>
%cst_669 = arith.constant dense_resource<__elided__> : tensor<640x48x1xf16>
%cst_670 = arith.constant dense_resource<__elided__> : tensor<640x768xi8>
%cst_671 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_672 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_673 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_674 = arith.constant dense_resource<__elided__> : tensor<5120x40x1xf16>
%cst_675 = arith.constant dense_resource<__elided__> : tensor<5120x640xi8>
%cst_676 = arith.constant dense_resource<__elided__> : tensor<5120xf16>
%cst_677 = arith.constant dense_resource<__elided__> : tensor<640x160x1xf16>
%cst_678 = arith.constant dense_resource<__elided__> : tensor<640x2560xi8>
%cst_679 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_680 = arith.constant dense_resource<__elided__> : tensor<640x40x1x1x1xf16>
%cst_681 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xi8>
%cst_682 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_683 = arith.constant dense_resource<__elided__> : tensor<960xf16>
%cst_684 = arith.constant dense_resource<__elided__> : tensor<960xf16>
%cst_685 = arith.constant dense_resource<__elided__> : tensor<640x60x1x3x3xf16>
%cst_686 = arith.constant dense_resource<__elided__> : tensor<640x960x3x3xi8>
%cst_687 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_688 = arith.constant dense_resource<__elided__> : tensor<640x1280xf16>
%cst_689 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_690 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_691 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_692 = arith.constant dense_resource<__elided__> : tensor<640x40x1x3x3xf16>
%cst_693 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xi8>
%cst_694 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_695 = arith.constant dense_resource<__elided__> : tensor<640x60x1x1x1xf16>
%cst_696 = arith.constant dense_resource<__elided__> : tensor<640x960x1x1xi8>
%cst_697 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_698 = arith.constant dense_resource<__elided__> : tensor<640x40x1x1x1xf16>
%cst_699 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xi8>
%cst_700 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_701 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_702 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_703 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_704 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_705 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_706 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_707 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_708 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_709 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_710 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_711 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_712 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_713 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_714 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_715 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_716 = arith.constant dense_resource<__elided__> : tensor<640x48x1xf16>
%cst_717 = arith.constant dense_resource<__elided__> : tensor<640x768xi8>
%cst_718 = arith.constant dense_resource<__elided__> : tensor<640x48x1xf16>
%cst_719 = arith.constant dense_resource<__elided__> : tensor<640x768xi8>
%cst_720 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_721 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_722 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_723 = arith.constant dense_resource<__elided__> : tensor<5120x40x1xf16>
%cst_724 = arith.constant dense_resource<__elided__> : tensor<5120x640xi8>
%cst_725 = arith.constant dense_resource<__elided__> : tensor<5120xf16>
%cst_726 = arith.constant dense_resource<__elided__> : tensor<640x160x1xf16>
%cst_727 = arith.constant dense_resource<__elided__> : tensor<640x2560xi8>
%cst_728 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_729 = arith.constant dense_resource<__elided__> : tensor<640x40x1x1x1xf16>
%cst_730 = arith.constant dense_resource<__elided__> : tensor<640x640x1x1xi8>
%cst_731 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_732 = arith.constant dense_resource<__elided__> : tensor<640x40x1x3x3xf16>
%cst_733 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xi8>
%cst_734 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_735 = arith.constant dense_resource<__elided__> : tensor<960xf16>
%cst_736 = arith.constant dense_resource<__elided__> : tensor<960xf16>
%cst_737 = arith.constant dense_resource<__elided__> : tensor<320x60x1x3x3xf16>
%cst_738 = arith.constant dense_resource<__elided__> : tensor<320x960x3x3xi8>
%cst_739 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_740 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16>
%cst_741 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_742 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_743 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_744 = arith.constant dense_resource<__elided__> : tensor<320x20x1x3x3xf16>
%cst_745 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xi8>
%cst_746 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_747 = arith.constant dense_resource<__elided__> : tensor<320x60x1x1x1xf16>
%cst_748 = arith.constant dense_resource<__elided__> : tensor<320x960x1x1xi8>
%cst_749 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_750 = arith.constant dense_resource<__elided__> : tensor<320x20x1x1x1xf16>
%cst_751 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xi8>
%cst_752 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_753 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_754 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_755 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_756 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_757 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_758 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_759 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_760 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_761 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_762 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_763 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_764 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_765 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_766 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_767 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_768 = arith.constant dense_resource<__elided__> : tensor<320x48x1xf16>
%cst_769 = arith.constant dense_resource<__elided__> : tensor<320x768xi8>
%cst_770 = arith.constant dense_resource<__elided__> : tensor<320x48x1xf16>
%cst_771 = arith.constant dense_resource<__elided__> : tensor<320x768xi8>
%cst_772 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_773 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_774 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_775 = arith.constant dense_resource<__elided__> : tensor<2560x20x1xf16>
%cst_776 = arith.constant dense_resource<__elided__> : tensor<2560x320xi8>
%cst_777 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_778 = arith.constant dense_resource<__elided__> : tensor<320x80x1xf16>
%cst_779 = arith.constant dense_resource<__elided__> : tensor<320x1280xi8>
%cst_780 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_781 = arith.constant dense_resource<__elided__> : tensor<320x20x1x1x1xf16>
%cst_782 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xi8>
%cst_783 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_784 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_785 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_786 = arith.constant dense_resource<__elided__> : tensor<320x40x1x3x3xf16>
%cst_787 = arith.constant dense_resource<__elided__> : tensor<320x640x3x3xi8>
%cst_788 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_789 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16>
%cst_790 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_791 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_792 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_793 = arith.constant dense_resource<__elided__> : tensor<320x20x1x3x3xf16>
%cst_794 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xi8>
%cst_795 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_796 = arith.constant dense_resource<__elided__> : tensor<320x40x1x1x1xf16>
%cst_797 = arith.constant dense_resource<__elided__> : tensor<320x640x1x1xi8>
%cst_798 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_799 = arith.constant dense_resource<__elided__> : tensor<320x20x1x1x1xf16>
%cst_800 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xi8>
%cst_801 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_802 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_803 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_804 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_805 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_806 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_807 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_808 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_809 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_810 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_811 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_812 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_813 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_814 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_815 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_816 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_817 = arith.constant dense_resource<__elided__> : tensor<320x48x1xf16>
%cst_818 = arith.constant dense_resource<__elided__> : tensor<320x768xi8>
%cst_819 = arith.constant dense_resource<__elided__> : tensor<320x48x1xf16>
%cst_820 = arith.constant dense_resource<__elided__> : tensor<320x768xi8>
%cst_821 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_822 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_823 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_824 = arith.constant dense_resource<__elided__> : tensor<2560x20x1xf16>
%cst_825 = arith.constant dense_resource<__elided__> : tensor<2560x320xi8>
%cst_826 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_827 = arith.constant dense_resource<__elided__> : tensor<320x80x1xf16>
%cst_828 = arith.constant dense_resource<__elided__> : tensor<320x1280xi8>
%cst_829 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_830 = arith.constant dense_resource<__elided__> : tensor<320x20x1x1x1xf16>
%cst_831 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xi8>
%cst_832 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_833 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_834 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_835 = arith.constant dense_resource<__elided__> : tensor<320x40x1x3x3xf16>
%cst_836 = arith.constant dense_resource<__elided__> : tensor<320x640x3x3xi8>
%cst_837 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_838 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16>
%cst_839 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_840 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_841 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_842 = arith.constant dense_resource<__elided__> : tensor<320x20x1x3x3xf16>
%cst_843 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xi8>
%cst_844 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_845 = arith.constant dense_resource<__elided__> : tensor<320x40x1x1x1xf16>
%cst_846 = arith.constant dense_resource<__elided__> : tensor<320x640x1x1xi8>
%cst_847 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_848 = arith.constant dense_resource<__elided__> : tensor<320x20x1x1x1xf16>
%cst_849 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xi8>
%cst_850 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_851 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_852 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_853 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_854 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_855 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_856 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_857 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_858 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_859 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_860 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_861 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_862 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_863 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_864 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_865 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_866 = arith.constant dense_resource<__elided__> : tensor<320x48x1xf16>
%cst_867 = arith.constant dense_resource<__elided__> : tensor<320x768xi8>
%cst_868 = arith.constant dense_resource<__elided__> : tensor<320x48x1xf16>
%cst_869 = arith.constant dense_resource<__elided__> : tensor<320x768xi8>
%cst_870 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_871 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_872 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_873 = arith.constant dense_resource<__elided__> : tensor<2560x20x1xf16>
%cst_874 = arith.constant dense_resource<__elided__> : tensor<2560x320xi8>
%cst_875 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_876 = arith.constant dense_resource<__elided__> : tensor<320x80x1xf16>
%cst_877 = arith.constant dense_resource<__elided__> : tensor<320x1280xi8>
%cst_878 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_879 = arith.constant dense<8.000000e+00> : tensor<f16>
%cst_880 = arith.constant dense<0.000000e+00> : tensor<f16>
%cst_881 = arith.constant dense<2.000000e+00> : tensor<f16>
%cst_882 = arith.constant dense_resource<__elided__> : tensor<320x20x1x1x1xf16>
%cst_883 = arith.constant dense_resource<__elided__> : tensor<320x320x1x1xi8>
%cst_884 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_885 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_886 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_887 = arith.constant dense_resource<__elided__> : tensor<4x320x3x3xf16>
%cst_888 = arith.constant dense<[-1.393320e-03, -1.588820e-03, -2.624990e-04, -2.531050e-03]> : tensor<4xf16>
%cst_889 = arith.constant 0.000000e+00 : f16
%cst_890 = arith.constant 1.000000e+00 : f16
%cst_891 = arith.constant -6.550400e+04 : f16
%cst_892 = arith.constant 2.000000e+00 : f16
%cst_893 = arith.constant 5.000000e-01 : f16
%cst_894 = arith.constant 0.000000e+00 : f64
%cst_895 = arith.constant 1.000000e-05 : f64
%cst_896 = arith.constant 0.000000e+00 : f32
%cst_897 = arith.constant 4.096000e+04 : f32
%cst_898 = arith.constant 4.096000e+04 : f64
%cst_899 = arith.constant 3.200000e+02 : f16
%cst_900 = arith.constant 1.024000e+04 : f32
%cst_901 = arith.constant 1.024000e+04 : f64
%cst_902 = arith.constant 2.048000e+04 : f32
%cst_903 = arith.constant 2.048000e+04 : f64
%cst_904 = arith.constant 6.400000e+02 : f16
%cst_905 = arith.constant 5.120000e+03 : f32
%cst_906 = arith.constant 5.120000e+03 : f64
%cst_907 = arith.constant 1.280000e+03 : f16
%cst_908 = arith.constant 2.560000e+03 : f32
%cst_909 = arith.constant 2.560000e+03 : f64
%cst_910 = arith.constant 1.536000e+04 : f32
%cst_911 = arith.constant 1.536000e+04 : f64
%cst_912 = arith.constant 6.144000e+04 : f32
%cst_913 = arith.constant 6.144000e+04 : f64
%cst_914 = arith.constant 3.072000e+04 : f32
%cst_915 = arith.constant 3.072000e+04 : f64
%cst_916 = arith.constant 1.228800e+05 : f32
%cst_917 = arith.constant 1.228800e+05 : f64
%cst_918 = arith.constant 8.192000e+04 : f32
%cst_919 = arith.constant 8.192000e+04 : f64
%c2 = arith.constant 2 : index
%expanded = tensor.expand_shape %arg1 [] : tensor<i64> into tensor<1xi64>
%0 = tensor.empty() : tensor<2xi64>
%1 = linalg.generic {indexing_maps = [#map0, #map1], iterator_types = ["parallel"]} ins(%expanded : tensor<1xi64>) outs(%0 : tensor<2xi64>) {
^bb0(%in: i64, %out: i64):
linalg.yield %in : i64
} -> tensor<2xi64>
%2 = tensor.empty() : tensor<160xf32>
%3 = linalg.generic {indexing_maps = [#map1], iterator_types = ["parallel"]} outs(%2 : tensor<160xf32>) {
^bb0(%out: f32):
%10165 = linalg.index 0 : index
%10166 = arith.index_cast %10165 : index to i64
%10167 = arith.sitofp %10166 : i64 to f32
%10168 = arith.addf %10167, %cst_896 : f32
linalg.yield %10168 : f32
} -> tensor<160xf32>
%4 = linalg.generic {indexing_maps = [#map1, #map2, #map1], iterator_types = ["parallel"]} ins(%3, %cst_7 : tensor<160xf32>, tensor<f16>) outs(%2 : tensor<160xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.mulf %in, %10165 : f32
linalg.yield %10166 : 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_2540: i64, %out: f32):
%10165 = arith.sitofp %in_2540 : i64 to f32
%10166 = arith.divf %in, %10165 : f32
linalg.yield %10166 : 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):
%10165 = math.exp %in : f32
linalg.yield %10165 : f32
} -> tensor<160xf32>
%expanded_920 = tensor.expand_shape %1 [[0, 1]] : tensor<2xi64> into tensor<2x1xi64>
%7 = tensor.empty() : tensor<f64>
%8 = linalg.fill ins(%cst_894 : 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):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : 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<2x1xi64>
%14 = linalg.generic {indexing_maps = [#map6, #map5], iterator_types = ["parallel", "parallel"]} ins(%expanded_920 : tensor<2x1xi64>) outs(%13 : tensor<2x1xi64>) {
^bb0(%in: i64, %out: i64):
linalg.yield %in : i64
} -> tensor<2x1xi64>
%15 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%14 : tensor<2x1xi64>) outs(%12 : tensor<2x1xf32>) {
^bb0(%in: i64, %out: f32):
%10165 = arith.sitofp %in : i64 to f32
linalg.yield %10165 : f32
} -> tensor<2x1xf32>
%expanded_921 = 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_921 : tensor<2x1xf32>, tensor<1x160xf32>) outs(%16 : tensor<2x160xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.mulf %in, %in_2540 : f32
linalg.yield %10165 : 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_2540: i64, %out: f32):
%10165 = arith.sitofp %in_2540 : i64 to f32
%10166 = arith.mulf %in, %10165 : f32
linalg.yield %10166 : 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):
%10165 = math.sin %in : f32
linalg.yield %10165 : 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):
%10165 = math.cos %in : f32
linalg.yield %10165 : 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_922 = 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_922[0, 0] [2, 160] [1, 1] : tensor<2x320xf32> to tensor<2x160xf32>
%inserted_slice_923 = tensor.insert_slice %20 into %21[0, 0] [2, 160] [1, 1] : tensor<2x160xf32> into tensor<2x320xf32>
%inserted_slice_924 = tensor.insert_slice %extracted_slice into %inserted_slice_923[0, 160] [2, 160] [1, 1] : tensor<2x160xf32> into tensor<2x320xf32>
%22 = linalg.fill ins(%cst_894 : 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):
%10165 = arith.truncf %in : f64 to f16
linalg.yield %10165 : 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_924 : tensor<2x320xf32>) outs(%26 : tensor<2x320xf16>) {
^bb0(%in: f32, %out: f16):
%10165 = arith.truncf %in : f32 to f16
linalg.yield %10165 : 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_889 : 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_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : 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):
%10165 = arith.negf %in : f16
%10166 = math.exp %10165 : f16
%10167 = arith.addf %10166, %cst_890 : f16
%10168 = arith.divf %cst_890, %10167 : f16
linalg.yield %10168 : 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_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : 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_889 : 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_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : 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_889 : 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_925 = tensor.expand_shape %collapsed [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16>
%44 = linalg.fill ins(%cst_894 : 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):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : 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_925 : tensor<2x32x10x4096xf16>) outs(%47 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %out: f32):
%10165 = arith.extf %in : f16 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%49 = tensor.empty() : tensor<2x32x1x1xf32>
%50 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%51 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%48 : tensor<2x32x10x4096xf32>) outs(%50 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%52 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%51 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_897 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%53 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%48, %52 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.subf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%54 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%53 : tensor<2x32x10x4096xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.mulf %in, %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%55 = tensor.empty() : tensor<2x32x1x1xf64>
%56 = linalg.fill ins(%cst_894 : f64) outs(%55 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
%57 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%54 : tensor<2x32x10x4096xf32>) outs(%56 : tensor<2x32x1x1xf64>) {
^bb0(%in: f32, %out: f64):
%10165 = arith.extf %in : f32 to f64
%10166 = arith.addf %10165, %out : f64
linalg.yield %10166 : f64
} -> tensor<2x32x1x1xf64>
%58 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%57 : tensor<2x32x1x1xf64>) outs(%55 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%10165 = arith.divf %in, %cst_898 : f64
linalg.yield %10165 : f64
} -> tensor<2x32x1x1xf64>
%59 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%58 : tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%60 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%61 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%48 : tensor<2x32x10x4096xf32>) outs(%60 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%62 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%61 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_897 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%63 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%59, %cst_4 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x1x1xf32>
%64 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%63 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = math.rsqrt %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%65 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_925, %62 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %in_2540: f32, %out: f32):
%10165 = arith.extf %in : f16 to f32
%10166 = arith.subf %10165, %in_2540 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x10x4096xf32>
%66 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%65, %64 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.mulf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%collapsed_926 = tensor.collapse_shape %66 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32>
%expanded_927 = tensor.expand_shape %collapsed_926 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32>
%expanded_928 = tensor.expand_shape %cst_14 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16>
%67 = tensor.empty() : tensor<2x320x64x64xf32>
%68 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_927, %expanded_928 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%67 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.mulf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x320x64x64xf32>
%expanded_929 = tensor.expand_shape %cst_15 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16>
%69 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%68, %expanded_929 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%67 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x320x64x64xf32>
%70 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%71 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%70 : tensor<f64>) outs(%23 : tensor<f16>) {
^bb0(%in: f64, %out: f16):
%10165 = arith.truncf %in : f64 to f16
linalg.yield %10165 : f16
} -> tensor<f16>
%72 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%71 : tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%73 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%69 : tensor<2x320x64x64xf32>) outs(%72 : tensor<2x320x64x64xf16>) {
^bb0(%in: f32, %out: f16):
%10165 = arith.truncf %in : f32 to f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%74 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%73 : tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
%10166 = math.exp %10165 : f16
%10167 = arith.addf %10166, %cst_890 : f16
%10168 = arith.divf %cst_890, %10167 : f16
linalg.yield %10168 : f16
} -> tensor<2x320x64x64xf16>
%75 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%74, %73 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%expanded_930 = tensor.expand_shape %75 [[0], [1, 2], [3], [4]] : tensor<2x320x64x64xf16> into tensor<2x20x16x64x64xf16>
%76 = tensor.empty() : tensor<2x20x16x64x64xf16>
%77 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_930 : tensor<2x20x16x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x20x16x64x64xf16>
%78 = tensor.empty() : tensor<2x20x1x64x64xf16>
%79 = linalg.fill ins(%cst_891 : f16) outs(%78 : tensor<2x20x1x64x64xf16>) -> tensor<2x20x1x64x64xf16>
%80 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%77 : tensor<2x20x16x64x64xf16>) outs(%79 : tensor<2x20x1x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x20x1x64x64xf16>
%81 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%80 : tensor<2x20x1x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x64x64xf16>
%collapsed_931 = tensor.collapse_shape %81 [[0], [1, 2], [3], [4]] : tensor<2x20x16x64x64xf16> into tensor<2x320x64x64xf16>
%82 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%83 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %82 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%84 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%83 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%85 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%84 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%86 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_931, %85 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%87 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%75, %86 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%88 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%87, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%89 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%90 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %89 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%91 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%90 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%92 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%93 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %92 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%94 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%93, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%95 = tensor.empty() : tensor<2x320x64x64xi1>
%96 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%88, %94 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%97 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%96, %94, %88 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%98 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%97, %91 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%99 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%98, %91, %97 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%100 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%99 : tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%101 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%100, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%102 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%101, %86 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%103 = tensor.empty() : tensor<320x20x16x3x3xf16>
%104 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_16 : tensor<320x20x1x3x3xf16>) outs(%103 : tensor<320x20x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x3x3xf16>
%collapsed_932 = tensor.collapse_shape %104 [[0], [1, 2], [3], [4]] : tensor<320x20x16x3x3xf16> into tensor<320x320x3x3xf16>
%105 = tensor.empty() : tensor<320x320x3x3xf16>
%106 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_17, %collapsed_932 : tensor<320x320x3x3xi8>, tensor<320x320x3x3xf16>) outs(%105 : tensor<320x320x3x3xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320x3x3xf16>
%padded_933 = tensor.pad %102 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
tensor.yield %cst_889 : f16
} : tensor<2x320x64x64xf16> to tensor<2x320x66x66xf16>
%107 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_18 : tensor<320xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%108 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_933, %106 : tensor<2x320x66x66xf16>, tensor<320x320x3x3xf16>) outs(%107 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16>
%109 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
%10166 = math.exp %10165 : f16
%10167 = arith.addf %10166, %cst_890 : f16
%10168 = arith.divf %cst_890, %10167 : f16
linalg.yield %10168 : f16
} -> tensor<2x1280xf16>
%110 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%109, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280xf16>
%111 = tensor.empty() : tensor<1280x320xf16>
%112 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_19 : tensor<320x1280xf16>) outs(%111 : tensor<1280x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x320xf16>
%113 = linalg.fill ins(%cst_889 : f16) outs(%25 : tensor<2x320xf16>) -> tensor<2x320xf16>
%114 = linalg.matmul ins(%110, %112 : tensor<2x1280xf16>, tensor<1280x320xf16>) outs(%113 : tensor<2x320xf16>) -> tensor<2x320xf16>
%115 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_20, %114 : tensor<320xf16>, tensor<2x320xf16>) outs(%25 : tensor<2x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320xf16>
%expanded_934 = tensor.expand_shape %115 [[0], [1, 2, 3]] : tensor<2x320xf16> into tensor<2x320x1x1xf16>
%116 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%108, %expanded_934 : tensor<2x320x64x64xf16>, tensor<2x320x1x1xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%collapsed_935 = tensor.collapse_shape %116 [[0], [1], [2, 3]] : tensor<2x320x64x64xf16> into tensor<2x320x4096xf16>
%expanded_936 = tensor.expand_shape %collapsed_935 [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16>
%117 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%118 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%117 : tensor<f64>) outs(%9 : tensor<f32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<f32>
%119 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%118 : tensor<f32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<2x32x10x4096xf32>
%120 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_936 : tensor<2x32x10x4096xf16>) outs(%119 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %out: f32):
%10165 = arith.extf %in : f16 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%121 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%122 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%120 : tensor<2x32x10x4096xf32>) outs(%121 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%123 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%122 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_897 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%124 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%120, %123 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.subf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%125 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%124 : tensor<2x32x10x4096xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.mulf %in, %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%126 = linalg.fill ins(%cst_894 : f64) outs(%55 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
%127 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%125 : tensor<2x32x10x4096xf32>) outs(%126 : tensor<2x32x1x1xf64>) {
^bb0(%in: f32, %out: f64):
%10165 = arith.extf %in : f32 to f64
%10166 = arith.addf %10165, %out : f64
linalg.yield %10166 : f64
} -> tensor<2x32x1x1xf64>
%128 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%127 : tensor<2x32x1x1xf64>) outs(%55 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%10165 = arith.divf %in, %cst_898 : f64
linalg.yield %10165 : f64
} -> tensor<2x32x1x1xf64>
%129 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%128 : tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%130 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%131 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%120 : tensor<2x32x10x4096xf32>) outs(%130 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%132 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%131 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_897 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%133 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%129, %cst_4 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x1x1xf32>
%134 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%133 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = math.rsqrt %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%135 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_936, %132 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %in_2540: f32, %out: f32):
%10165 = arith.extf %in : f16 to f32
%10166 = arith.subf %10165, %in_2540 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x10x4096xf32>
%136 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%135, %134 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.mulf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%collapsed_937 = tensor.collapse_shape %136 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32>
%expanded_938 = tensor.expand_shape %collapsed_937 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32>
%expanded_939 = tensor.expand_shape %cst_21 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16>
%137 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_938, %expanded_939 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%67 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.mulf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x320x64x64xf32>
%expanded_940 = tensor.expand_shape %cst_22 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16>
%138 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%137, %expanded_940 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%67 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x320x64x64xf32>
%139 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%140 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%139 : tensor<f64>) outs(%23 : tensor<f16>) {
^bb0(%in: f64, %out: f16):
%10165 = arith.truncf %in : f64 to f16
linalg.yield %10165 : f16
} -> tensor<f16>
%141 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%140 : tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%142 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%138 : tensor<2x320x64x64xf32>) outs(%141 : tensor<2x320x64x64xf16>) {
^bb0(%in: f32, %out: f16):
%10165 = arith.truncf %in : f32 to f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%143 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%142 : tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
%10166 = math.exp %10165 : f16
%10167 = arith.addf %10166, %cst_890 : f16
%10168 = arith.divf %cst_890, %10167 : f16
linalg.yield %10168 : f16
} -> tensor<2x320x64x64xf16>
%144 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%143, %142 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%expanded_941 = tensor.expand_shape %144 [[0], [1, 2], [3], [4]] : tensor<2x320x64x64xf16> into tensor<2x20x16x64x64xf16>
%145 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_941 : tensor<2x20x16x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x20x16x64x64xf16>
%146 = linalg.fill ins(%cst_891 : f16) outs(%78 : tensor<2x20x1x64x64xf16>) -> tensor<2x20x1x64x64xf16>
%147 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%145 : tensor<2x20x16x64x64xf16>) outs(%146 : tensor<2x20x1x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x20x1x64x64xf16>
%148 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%147 : tensor<2x20x1x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x64x64xf16>
%collapsed_942 = tensor.collapse_shape %148 [[0], [1, 2], [3], [4]] : tensor<2x20x16x64x64xf16> into tensor<2x320x64x64xf16>
%149 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%150 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %149 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%151 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%150 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%152 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%151 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%153 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_942, %152 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%154 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%144, %153 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%155 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%154, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%156 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%157 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %156 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%158 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%157 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%159 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%160 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %159 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%161 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%160, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%162 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%155, %161 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%163 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%162, %161, %155 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%164 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%163, %158 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%165 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%164, %158, %163 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%166 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%165 : tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%167 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%166, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%168 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%167, %153 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%169 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_23 : tensor<320x20x1x3x3xf16>) outs(%103 : tensor<320x20x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x3x3xf16>
%collapsed_943 = tensor.collapse_shape %169 [[0], [1, 2], [3], [4]] : tensor<320x20x16x3x3xf16> into tensor<320x320x3x3xf16>
%170 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_24, %collapsed_943 : tensor<320x320x3x3xi8>, tensor<320x320x3x3xf16>) outs(%105 : tensor<320x320x3x3xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320x3x3xf16>
%padded_944 = tensor.pad %168 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
tensor.yield %cst_889 : f16
} : tensor<2x320x64x64xf16> to tensor<2x320x66x66xf16>
%171 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_25 : tensor<320xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%172 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_944, %170 : tensor<2x320x66x66xf16>, tensor<320x320x3x3xf16>) outs(%171 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16>
%173 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%43, %172 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%174 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%173, %cst_3 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%collapsed_945 = tensor.collapse_shape %174 [[0], [1], [2, 3]] : tensor<2x320x64x64xf16> into tensor<2x320x4096xf16>
%expanded_946 = tensor.expand_shape %collapsed_945 [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16>
%175 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%176 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%175 : tensor<f64>) outs(%9 : tensor<f32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<f32>
%177 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%176 : tensor<f32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<2x32x10x4096xf32>
%178 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_946 : tensor<2x32x10x4096xf16>) outs(%177 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %out: f32):
%10165 = arith.extf %in : f16 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%179 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%180 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%178 : tensor<2x32x10x4096xf32>) outs(%179 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%181 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%180 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_897 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%182 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%178, %181 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.subf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%183 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%182 : tensor<2x32x10x4096xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.mulf %in, %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%184 = linalg.fill ins(%cst_894 : f64) outs(%55 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
%185 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%183 : tensor<2x32x10x4096xf32>) outs(%184 : tensor<2x32x1x1xf64>) {
^bb0(%in: f32, %out: f64):
%10165 = arith.extf %in : f32 to f64
%10166 = arith.addf %10165, %out : f64
linalg.yield %10166 : f64
} -> tensor<2x32x1x1xf64>
%186 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%185 : tensor<2x32x1x1xf64>) outs(%55 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%10165 = arith.divf %in, %cst_898 : f64
linalg.yield %10165 : f64
} -> tensor<2x32x1x1xf64>
%187 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%186 : tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%188 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%189 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%178 : tensor<2x32x10x4096xf32>) outs(%188 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%190 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%189 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_897 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%191 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%187, %cst_2 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x1x1xf32>
%192 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%191 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = math.rsqrt %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%193 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_946, %190 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %in_2540: f32, %out: f32):
%10165 = arith.extf %in : f16 to f32
%10166 = arith.subf %10165, %in_2540 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x10x4096xf32>
%194 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%193, %192 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.mulf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%collapsed_947 = tensor.collapse_shape %194 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32>
%expanded_948 = tensor.expand_shape %collapsed_947 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32>
%195 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%196 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%195 : tensor<f64>) outs(%23 : tensor<f16>) {
^bb0(%in: f64, %out: f16):
%10165 = arith.truncf %in : f64 to f16
linalg.yield %10165 : f16
} -> tensor<f16>
%197 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%196 : tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%198 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_948 : tensor<2x320x64x64xf32>) outs(%197 : tensor<2x320x64x64xf16>) {
^bb0(%in: f32, %out: f16):
%10165 = arith.truncf %in : f32 to f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%expanded_949 = tensor.expand_shape %198 [[0], [1, 2], [3], [4]] : tensor<2x320x64x64xf16> into tensor<2x20x16x64x64xf16>
%199 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_949 : tensor<2x20x16x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x20x16x64x64xf16>
%200 = linalg.fill ins(%cst_891 : f16) outs(%78 : tensor<2x20x1x64x64xf16>) -> tensor<2x20x1x64x64xf16>
%201 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%199 : tensor<2x20x16x64x64xf16>) outs(%200 : tensor<2x20x1x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x20x1x64x64xf16>
%202 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%201 : tensor<2x20x1x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x64x64xf16>
%collapsed_950 = tensor.collapse_shape %202 [[0], [1, 2], [3], [4]] : tensor<2x20x16x64x64xf16> into tensor<2x320x64x64xf16>
%203 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%204 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %203 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%205 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%204 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%206 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%205 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%207 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_950, %206 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%208 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%198, %207 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%209 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%208, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%210 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%211 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %210 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%212 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%211 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%213 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%214 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %213 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%215 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%214, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%216 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%209, %215 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%217 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%216, %215, %209 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%218 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%217, %212 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%219 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%218, %212, %217 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%220 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%219 : tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%221 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%220, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%222 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%221, %207 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%223 = tensor.empty() : tensor<320x20x16x1x1xf16>
%224 = linalg.generic {indexing_maps = [#map17, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_26 : tensor<320x20x1x1x1xf16>) outs(%223 : tensor<320x20x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x1x1xf16>
%225 = linalg.generic {indexing_maps = [#map18, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%224 : tensor<320x20x16x1x1xf16>) outs(%223 : tensor<320x20x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x1x1xf16>
%collapsed_951 = tensor.collapse_shape %225 [[0], [1, 2], [3], [4]] : tensor<320x20x16x1x1xf16> into tensor<320x320x1x1xf16>
%226 = tensor.empty() : tensor<320x320x1x1xf16>
%227 = linalg.generic {indexing_maps = [#map13, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_27, %collapsed_951 : tensor<320x320x1x1xi8>, tensor<320x320x1x1xf16>) outs(%226 : tensor<320x320x1x1xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320x1x1xf16>
%228 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_28 : tensor<320xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%229 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%222, %227 : tensor<2x320x64x64xf16>, tensor<320x320x1x1xf16>) outs(%228 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16>
%230 = tensor.empty() : tensor<2x64x64x320xf16>
%231 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%229 : tensor<2x320x64x64xf16>) outs(%230 : tensor<2x64x64x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x64x64x320xf16>
%collapsed_952 = tensor.collapse_shape %231 [[0], [1, 2], [3]] : tensor<2x64x64x320xf16> into tensor<2x4096x320xf16>
%232 = tensor.empty() : tensor<2x4096x1xf16>
%233 = linalg.fill ins(%cst_889 : f16) outs(%232 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16>
%234 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_952 : tensor<2x4096x320xf16>) outs(%233 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%235 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%234 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_899 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%236 = tensor.empty() : tensor<2x4096x320xf16>
%237 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%235 : tensor<2x4096x1xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x320xf16>
%238 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_952, %237 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%239 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%238 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.mulf %in, %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%240 = linalg.fill ins(%cst_889 : f16) outs(%232 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16>
%241 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%239 : tensor<2x4096x320xf16>) outs(%240 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%242 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%241 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_899 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%243 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%242 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.truncf %cst_895 : f64 to f16
%10166 = arith.addf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<2x4096x1xf16>
%244 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%243 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.rsqrt %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%245 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%244 : tensor<2x4096x1xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x320xf16>
%246 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%238, %245 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%expanded_953 = tensor.expand_shape %246 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%247 = tensor.empty() : tensor<2x4096x20x16xf16>
%248 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_953 : tensor<2x4096x20x16xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x16xf16>
%249 = tensor.empty() : tensor<2x4096x20x1xf16>
%250 = linalg.fill ins(%cst_891 : f16) outs(%249 : tensor<2x4096x20x1xf16>) -> tensor<2x4096x20x1xf16>
%251 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%248 : tensor<2x4096x20x16xf16>) outs(%250 : tensor<2x4096x20x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x1xf16>
%252 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%251 : tensor<2x4096x20x1xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_954 = tensor.collapse_shape %252 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%253 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%254 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %253 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%255 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%254 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%256 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%255 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%257 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_954, %256 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%258 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%246, %257 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%259 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%258, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%260 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%261 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %260 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%262 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%261 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%263 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%264 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %263 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%265 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%264, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%266 = tensor.empty() : tensor<2x4096x320xi1>
%267 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%259, %265 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%268 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%267, %265, %259 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%269 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%268, %262 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%270 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%269, %262, %268 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%271 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%270 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%272 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%271, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%273 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%272, %257 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%274 = tensor.empty() : tensor<320x20x16xf16>
%275 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_29 : tensor<320x20x1xf16>) outs(%274 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_955 = tensor.collapse_shape %275 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%276 = tensor.empty() : tensor<320x320xf16>
%277 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_30, %collapsed_955 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320xf16>
%278 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%277 : tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_956 = tensor.collapse_shape %273 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%279 = tensor.empty() : tensor<8192x320xf16>
%280 = linalg.fill ins(%cst_889 : f16) outs(%279 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%281 = linalg.matmul ins(%collapsed_956, %278 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%280 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%282 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_31, %281 : tensor<320xf16>, tensor<8192x320xf16>) outs(%279 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<8192x320xf16>
%283 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_953 : tensor<2x4096x20x16xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x16xf16>
%284 = linalg.fill ins(%cst_891 : f16) outs(%249 : tensor<2x4096x20x1xf16>) -> tensor<2x4096x20x1xf16>
%285 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%283 : tensor<2x4096x20x16xf16>) outs(%284 : tensor<2x4096x20x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x1xf16>
%286 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%285 : tensor<2x4096x20x1xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_957 = tensor.collapse_shape %286 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%287 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%288 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %287 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%289 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%288 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%290 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%289 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%291 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_957, %290 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%292 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%246, %291 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%293 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%292, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%294 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%295 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %294 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%296 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%295 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%297 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%298 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %297 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%299 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%298, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%300 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%293, %299 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%301 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%300, %299, %293 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%302 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%301, %296 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%303 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%302, %296, %301 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%304 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%303 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%305 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%304, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%306 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%305, %291 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%307 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_32 : tensor<320x20x1xf16>) outs(%274 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_958 = tensor.collapse_shape %307 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%308 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_33, %collapsed_958 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320xf16>
%309 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%308 : tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_959 = tensor.collapse_shape %306 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%310 = linalg.fill ins(%cst_889 : f16) outs(%279 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%311 = linalg.matmul ins(%collapsed_959, %309 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%310 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%312 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_34, %311 : tensor<320xf16>, tensor<8192x320xf16>) outs(%279 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<8192x320xf16>
%313 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_953 : tensor<2x4096x20x16xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x16xf16>
%314 = linalg.fill ins(%cst_891 : f16) outs(%249 : tensor<2x4096x20x1xf16>) -> tensor<2x4096x20x1xf16>
%315 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%313 : tensor<2x4096x20x16xf16>) outs(%314 : tensor<2x4096x20x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x1xf16>
%316 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%315 : tensor<2x4096x20x1xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_960 = tensor.collapse_shape %316 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%317 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%318 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %317 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%319 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%318 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%320 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%319 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%321 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_960, %320 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%322 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%246, %321 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%323 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%322, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%324 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%325 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %324 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%326 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%325 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%327 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%328 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %327 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%329 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%328, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%330 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%323, %329 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%331 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%330, %329, %323 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%332 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%331, %326 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%333 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%332, %326, %331 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%334 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%333 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%335 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%334, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%336 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%335, %321 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%337 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_35 : tensor<320x20x1xf16>) outs(%274 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_961 = tensor.collapse_shape %337 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%338 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_36, %collapsed_961 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320xf16>
%339 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%338 : tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_962 = tensor.collapse_shape %336 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%340 = linalg.fill ins(%cst_889 : f16) outs(%279 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%341 = linalg.matmul ins(%collapsed_962, %339 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%340 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%342 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_37, %341 : tensor<320xf16>, tensor<8192x320xf16>) outs(%279 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<8192x320xf16>
%expanded_963 = tensor.expand_shape %282 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x8x40xf16>
%343 = tensor.empty() : tensor<2x8x4096x40xf16>
%344 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_963 : tensor<2x4096x8x40xf16>) outs(%343 : tensor<2x8x4096x40xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x4096x40xf16>
%collapsed_964 = tensor.collapse_shape %344 [[0, 1], [2], [3]] : tensor<2x8x4096x40xf16> into tensor<16x4096x40xf16>
%expanded_965 = tensor.expand_shape %312 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x8x40xf16>
%345 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_965 : tensor<2x4096x8x40xf16>) outs(%343 : tensor<2x8x4096x40xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x4096x40xf16>
%collapsed_966 = tensor.collapse_shape %345 [[0, 1], [2], [3]] : tensor<2x8x4096x40xf16> into tensor<16x4096x40xf16>
%expanded_967 = tensor.expand_shape %342 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x8x40xf16>
%346 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_967 : tensor<2x4096x8x40xf16>) outs(%343 : tensor<2x8x4096x40xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x4096x40xf16>
%collapsed_968 = tensor.collapse_shape %346 [[0, 1], [2], [3]] : tensor<2x8x4096x40xf16> into tensor<16x4096x40xf16>
%347 = tensor.empty() : tensor<16x40x4096xf16>
%348 = linalg.generic {indexing_maps = [#map20, #map25], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_966 : tensor<16x4096x40xf16>) outs(%347 : tensor<16x40x4096xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<16x40x4096xf16>
%349 = tensor.empty() : tensor<16x4096x4096xf16>
%350 = linalg.fill ins(%cst_889 : f16) outs(%349 : tensor<16x4096x4096xf16>) -> tensor<16x4096x4096xf16>
%351 = linalg.batch_matmul ins(%collapsed_964, %348 : tensor<16x4096x40xf16>, tensor<16x40x4096xf16>) outs(%350 : tensor<16x4096x4096xf16>) -> tensor<16x4096x4096xf16>
%352 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%351, %cst_1 : tensor<16x4096x4096xf16>, tensor<f16>) outs(%349 : tensor<16x4096x4096xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x4096xf16>
%353 = tensor.empty() : tensor<16x4096x1xf16>
%354 = linalg.fill ins(%cst_891 : f16) outs(%353 : tensor<16x4096x1xf16>) -> tensor<16x4096x1xf16>
%355 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%352 : tensor<16x4096x4096xf16>) outs(%354 : tensor<16x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x1xf16>
%356 = linalg.generic {indexing_maps = [#map20, #map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%352, %355 : tensor<16x4096x4096xf16>, tensor<16x4096x1xf16>) outs(%349 : tensor<16x4096x4096xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x4096xf16>
%357 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%356 : tensor<16x4096x4096xf16>) outs(%349 : tensor<16x4096x4096xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.exp %in : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x4096xf16>
%358 = linalg.fill ins(%cst_889 : f16) outs(%353 : tensor<16x4096x1xf16>) -> tensor<16x4096x1xf16>
%359 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%357 : tensor<16x4096x4096xf16>) outs(%358 : tensor<16x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x1xf16>
%360 = linalg.generic {indexing_maps = [#map20, #map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%357, %359 : tensor<16x4096x4096xf16>, tensor<16x4096x1xf16>) outs(%349 : tensor<16x4096x4096xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x4096xf16>
%361 = tensor.empty() : tensor<16x4096x40xf16>
%362 = linalg.fill ins(%cst_889 : f16) outs(%361 : tensor<16x4096x40xf16>) -> tensor<16x4096x40xf16>
%363 = linalg.batch_matmul ins(%360, %collapsed_968 : tensor<16x4096x4096xf16>, tensor<16x4096x40xf16>) outs(%362 : tensor<16x4096x40xf16>) -> tensor<16x4096x40xf16>
%expanded_969 = tensor.expand_shape %363 [[0, 1], [2], [3]] : tensor<16x4096x40xf16> into tensor<2x8x4096x40xf16>
%364 = tensor.empty() : tensor<2x4096x8x40xf16>
%365 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_969 : tensor<2x8x4096x40xf16>) outs(%364 : tensor<2x4096x8x40xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x8x40xf16>
%collapsed_970 = tensor.collapse_shape %365 [[0], [1], [2, 3]] : tensor<2x4096x8x40xf16> into tensor<2x4096x320xf16>
%expanded_971 = tensor.expand_shape %collapsed_970 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%366 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_971 : tensor<2x4096x20x16xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x16xf16>
%367 = linalg.fill ins(%cst_891 : f16) outs(%249 : tensor<2x4096x20x1xf16>) -> tensor<2x4096x20x1xf16>
%368 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%366 : tensor<2x4096x20x16xf16>) outs(%367 : tensor<2x4096x20x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x1xf16>
%369 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%368 : tensor<2x4096x20x1xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_972 = tensor.collapse_shape %369 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%370 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%371 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %370 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%372 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%371 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%373 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%372 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%374 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_972, %373 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%375 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_970, %374 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%376 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%375, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%377 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%378 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %377 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%379 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%378 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%380 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%381 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %380 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%382 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%381, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%383 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%376, %382 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%384 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%383, %382, %376 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%385 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%384, %379 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%386 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%385, %379, %384 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%387 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%386 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%388 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%387, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%389 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%388, %374 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%390 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_38 : tensor<320x20x1xf16>) outs(%274 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_973 = tensor.collapse_shape %390 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%391 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_39, %collapsed_973 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320xf16>
%392 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%391 : tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_974 = tensor.collapse_shape %389 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%393 = linalg.fill ins(%cst_889 : f16) outs(%279 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%394 = linalg.matmul ins(%collapsed_974, %392 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%393 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%395 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_40, %394 : tensor<320xf16>, tensor<8192x320xf16>) outs(%279 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<8192x320xf16>
%expanded_975 = tensor.expand_shape %395 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16>
%396 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_975, %collapsed_952 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%397 = linalg.fill ins(%cst_889 : f16) outs(%232 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16>
%398 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%396 : tensor<2x4096x320xf16>) outs(%397 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%399 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%398 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_899 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%400 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%399 : tensor<2x4096x1xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x320xf16>
%401 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%396, %400 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%402 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%401 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.mulf %in, %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%403 = linalg.fill ins(%cst_889 : f16) outs(%232 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16>
%404 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%402 : tensor<2x4096x320xf16>) outs(%403 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%405 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%404 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_899 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%406 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%405 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.truncf %cst_895 : f64 to f16
%10166 = arith.addf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<2x4096x1xf16>
%407 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%406 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.rsqrt %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%408 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%407 : tensor<2x4096x1xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x320xf16>
%409 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%401, %408 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%expanded_976 = tensor.expand_shape %409 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%410 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_976 : tensor<2x4096x20x16xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x16xf16>
%411 = linalg.fill ins(%cst_891 : f16) outs(%249 : tensor<2x4096x20x1xf16>) -> tensor<2x4096x20x1xf16>
%412 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%410 : tensor<2x4096x20x16xf16>) outs(%411 : tensor<2x4096x20x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x1xf16>
%413 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%412 : tensor<2x4096x20x1xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_977 = tensor.collapse_shape %413 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%414 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%415 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %414 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%416 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%415 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%417 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%416 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%418 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_977, %417 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%419 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%409, %418 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%420 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%419, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%421 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%422 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %421 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%423 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%422 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%424 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%425 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %424 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%426 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%425, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%427 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%420, %426 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%428 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%427, %426, %420 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%429 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%428, %423 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%430 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%429, %423, %428 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%431 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%430 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%432 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%431, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%433 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%432, %418 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%434 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_41 : tensor<320x20x1xf16>) outs(%274 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_978 = tensor.collapse_shape %434 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%435 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_42, %collapsed_978 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320xf16>
%436 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%435 : tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_979 = tensor.collapse_shape %433 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%437 = linalg.fill ins(%cst_889 : f16) outs(%279 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%438 = linalg.matmul ins(%collapsed_979, %436 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%437 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%439 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_43, %438 : tensor<320xf16>, tensor<8192x320xf16>) outs(%279 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<8192x320xf16>
%expanded_980 = tensor.expand_shape %arg2 [[0], [1], [2, 3]] : tensor<2x77x768xf16> into tensor<2x77x48x16xf16>
%440 = tensor.empty() : tensor<2x77x48x16xf16>
%441 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_980 : tensor<2x77x48x16xf16>) outs(%440 : tensor<2x77x48x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x77x48x16xf16>
%442 = tensor.empty() : tensor<2x77x48x1xf16>
%443 = linalg.fill ins(%cst_891 : f16) outs(%442 : tensor<2x77x48x1xf16>) -> tensor<2x77x48x1xf16>
%444 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%441 : tensor<2x77x48x16xf16>) outs(%443 : tensor<2x77x48x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x77x48x1xf16>
%445 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%444 : tensor<2x77x48x1xf16>) outs(%440 : tensor<2x77x48x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x77x48x16xf16>
%collapsed_981 = tensor.collapse_shape %445 [[0], [1], [2, 3]] : tensor<2x77x48x16xf16> into tensor<2x77x768xf16>
%446 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%447 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %446 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%448 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%447 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%449 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%448 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%450 = tensor.empty() : tensor<2x77x768xf16>
%451 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_981, %449 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%452 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2, %451 : tensor<2x77x768xf16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%453 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%452, %cst_880 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%454 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%455 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %454 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%456 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%455 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%457 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%458 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %457 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%459 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%458, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%460 = tensor.empty() : tensor<2x77x768xi1>
%461 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%453, %459 : tensor<2x77x768xf16>, tensor<f16>) outs(%460 : tensor<2x77x768xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x77x768xi1>
%462 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%461, %459, %453 : tensor<2x77x768xi1>, tensor<f16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%463 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%462, %456 : tensor<2x77x768xf16>, tensor<f16>) outs(%460 : tensor<2x77x768xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x77x768xi1>
%464 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%463, %456, %462 : tensor<2x77x768xi1>, tensor<f16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%465 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%464 : tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%466 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%465, %cst_880 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%467 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%466, %451 : tensor<2x77x768xf16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%468 = tensor.empty() : tensor<320x48x16xf16>
%469 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_44 : tensor<320x48x1xf16>) outs(%468 : tensor<320x48x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x48x16xf16>
%collapsed_982 = tensor.collapse_shape %469 [[0], [1, 2]] : tensor<320x48x16xf16> into tensor<320x768xf16>
%470 = tensor.empty() : tensor<320x768xf16>
%471 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_45, %collapsed_982 : tensor<320x768xi8>, tensor<320x768xf16>) outs(%470 : tensor<320x768xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x768xf16>
%472 = tensor.empty() : tensor<768x320xf16>
%473 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%471 : tensor<320x768xf16>) outs(%472 : tensor<768x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<768x320xf16>
%collapsed_983 = tensor.collapse_shape %467 [[0, 1], [2]] : tensor<2x77x768xf16> into tensor<154x768xf16>
%474 = tensor.empty() : tensor<154x320xf16>
%475 = linalg.fill ins(%cst_889 : f16) outs(%474 : tensor<154x320xf16>) -> tensor<154x320xf16>
%476 = linalg.matmul ins(%collapsed_983, %473 : tensor<154x768xf16>, tensor<768x320xf16>) outs(%475 : tensor<154x320xf16>) -> tensor<154x320xf16>
%477 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_980 : tensor<2x77x48x16xf16>) outs(%440 : tensor<2x77x48x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x77x48x16xf16>
%478 = linalg.fill ins(%cst_891 : f16) outs(%442 : tensor<2x77x48x1xf16>) -> tensor<2x77x48x1xf16>
%479 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%477 : tensor<2x77x48x16xf16>) outs(%478 : tensor<2x77x48x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x77x48x1xf16>
%480 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%479 : tensor<2x77x48x1xf16>) outs(%440 : tensor<2x77x48x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x77x48x16xf16>
%collapsed_984 = tensor.collapse_shape %480 [[0], [1], [2, 3]] : tensor<2x77x48x16xf16> into tensor<2x77x768xf16>
%481 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%482 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %481 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%483 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%482 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%484 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%483 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%485 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_984, %484 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%486 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2, %485 : tensor<2x77x768xf16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%487 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%486, %cst_880 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%488 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%489 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %488 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%490 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%489 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%491 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%492 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %491 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%493 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%492, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%494 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%487, %493 : tensor<2x77x768xf16>, tensor<f16>) outs(%460 : tensor<2x77x768xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x77x768xi1>
%495 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%494, %493, %487 : tensor<2x77x768xi1>, tensor<f16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%496 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%495, %490 : tensor<2x77x768xf16>, tensor<f16>) outs(%460 : tensor<2x77x768xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x77x768xi1>
%497 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%496, %490, %495 : tensor<2x77x768xi1>, tensor<f16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%498 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%497 : tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%499 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%498, %cst_880 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%500 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%499, %485 : tensor<2x77x768xf16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%501 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_46 : tensor<320x48x1xf16>) outs(%468 : tensor<320x48x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x48x16xf16>
%collapsed_985 = tensor.collapse_shape %501 [[0], [1, 2]] : tensor<320x48x16xf16> into tensor<320x768xf16>
%502 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_47, %collapsed_985 : tensor<320x768xi8>, tensor<320x768xf16>) outs(%470 : tensor<320x768xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x768xf16>
%503 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%502 : tensor<320x768xf16>) outs(%472 : tensor<768x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<768x320xf16>
%collapsed_986 = tensor.collapse_shape %500 [[0, 1], [2]] : tensor<2x77x768xf16> into tensor<154x768xf16>
%504 = linalg.fill ins(%cst_889 : f16) outs(%474 : tensor<154x320xf16>) -> tensor<154x320xf16>
%505 = linalg.matmul ins(%collapsed_986, %503 : tensor<154x768xf16>, tensor<768x320xf16>) outs(%504 : tensor<154x320xf16>) -> tensor<154x320xf16>
%expanded_987 = tensor.expand_shape %439 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x8x40xf16>
%506 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_987 : tensor<2x4096x8x40xf16>) outs(%343 : tensor<2x8x4096x40xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x4096x40xf16>
%collapsed_988 = tensor.collapse_shape %506 [[0, 1], [2], [3]] : tensor<2x8x4096x40xf16> into tensor<16x4096x40xf16>
%expanded_989 = tensor.expand_shape %476 [[0, 1], [2, 3]] : tensor<154x320xf16> into tensor<2x77x8x40xf16>
%507 = tensor.empty() : tensor<2x8x77x40xf16>
%508 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_989 : tensor<2x77x8x40xf16>) outs(%507 : tensor<2x8x77x40xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x77x40xf16>
%collapsed_990 = tensor.collapse_shape %508 [[0, 1], [2], [3]] : tensor<2x8x77x40xf16> into tensor<16x77x40xf16>
%expanded_991 = tensor.expand_shape %505 [[0, 1], [2, 3]] : tensor<154x320xf16> into tensor<2x77x8x40xf16>
%509 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_991 : tensor<2x77x8x40xf16>) outs(%507 : tensor<2x8x77x40xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x77x40xf16>
%collapsed_992 = tensor.collapse_shape %509 [[0, 1], [2], [3]] : tensor<2x8x77x40xf16> into tensor<16x77x40xf16>
%510 = tensor.empty() : tensor<16x40x77xf16>
%511 = linalg.generic {indexing_maps = [#map20, #map25], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_990 : tensor<16x77x40xf16>) outs(%510 : tensor<16x40x77xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<16x40x77xf16>
%512 = tensor.empty() : tensor<16x4096x77xf16>
%513 = linalg.fill ins(%cst_889 : f16) outs(%512 : tensor<16x4096x77xf16>) -> tensor<16x4096x77xf16>
%514 = linalg.batch_matmul ins(%collapsed_988, %511 : tensor<16x4096x40xf16>, tensor<16x40x77xf16>) outs(%513 : tensor<16x4096x77xf16>) -> tensor<16x4096x77xf16>
%515 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%514, %cst_1 : tensor<16x4096x77xf16>, tensor<f16>) outs(%512 : tensor<16x4096x77xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x77xf16>
%516 = linalg.fill ins(%cst_891 : f16) outs(%353 : tensor<16x4096x1xf16>) -> tensor<16x4096x1xf16>
%517 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%515 : tensor<16x4096x77xf16>) outs(%516 : tensor<16x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x1xf16>
%518 = linalg.generic {indexing_maps = [#map20, #map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%515, %517 : tensor<16x4096x77xf16>, tensor<16x4096x1xf16>) outs(%512 : tensor<16x4096x77xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x77xf16>
%519 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%518 : tensor<16x4096x77xf16>) outs(%512 : tensor<16x4096x77xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.exp %in : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x77xf16>
%520 = linalg.fill ins(%cst_889 : f16) outs(%353 : tensor<16x4096x1xf16>) -> tensor<16x4096x1xf16>
%521 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%519 : tensor<16x4096x77xf16>) outs(%520 : tensor<16x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x1xf16>
%522 = linalg.generic {indexing_maps = [#map20, #map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%519, %521 : tensor<16x4096x77xf16>, tensor<16x4096x1xf16>) outs(%512 : tensor<16x4096x77xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x77xf16>
%523 = linalg.fill ins(%cst_889 : f16) outs(%361 : tensor<16x4096x40xf16>) -> tensor<16x4096x40xf16>
%524 = linalg.batch_matmul ins(%522, %collapsed_992 : tensor<16x4096x77xf16>, tensor<16x77x40xf16>) outs(%523 : tensor<16x4096x40xf16>) -> tensor<16x4096x40xf16>
%expanded_993 = tensor.expand_shape %524 [[0, 1], [2], [3]] : tensor<16x4096x40xf16> into tensor<2x8x4096x40xf16>
%525 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_993 : tensor<2x8x4096x40xf16>) outs(%364 : tensor<2x4096x8x40xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x8x40xf16>
%collapsed_994 = tensor.collapse_shape %525 [[0], [1], [2, 3]] : tensor<2x4096x8x40xf16> into tensor<2x4096x320xf16>
%expanded_995 = tensor.expand_shape %collapsed_994 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%526 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_995 : tensor<2x4096x20x16xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x16xf16>
%527 = linalg.fill ins(%cst_891 : f16) outs(%249 : tensor<2x4096x20x1xf16>) -> tensor<2x4096x20x1xf16>
%528 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%526 : tensor<2x4096x20x16xf16>) outs(%527 : tensor<2x4096x20x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x1xf16>
%529 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%528 : tensor<2x4096x20x1xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_996 = tensor.collapse_shape %529 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%530 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%531 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %530 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%532 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%531 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%533 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%532 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%534 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_996, %533 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%535 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_994, %534 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%536 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%535, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%537 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%538 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %537 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%539 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%538 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%540 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%541 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %540 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%542 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%541, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%543 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%536, %542 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%544 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%543, %542, %536 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%545 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%544, %539 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%546 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%545, %539, %544 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%547 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%546 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%548 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%547, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%549 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%548, %534 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%550 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_48 : tensor<320x20x1xf16>) outs(%274 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_997 = tensor.collapse_shape %550 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%551 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_49, %collapsed_997 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320xf16>
%552 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%551 : tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_998 = tensor.collapse_shape %549 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%553 = linalg.fill ins(%cst_889 : f16) outs(%279 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%554 = linalg.matmul ins(%collapsed_998, %552 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%553 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%555 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_50, %554 : tensor<320xf16>, tensor<8192x320xf16>) outs(%279 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<8192x320xf16>
%expanded_999 = tensor.expand_shape %555 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16>
%556 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_999, %396 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%557 = linalg.fill ins(%cst_889 : f16) outs(%232 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16>
%558 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%556 : tensor<2x4096x320xf16>) outs(%557 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%559 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%558 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_899 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%560 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%559 : tensor<2x4096x1xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x320xf16>
%561 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%556, %560 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%562 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%561 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.mulf %in, %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%563 = linalg.fill ins(%cst_889 : f16) outs(%232 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16>
%564 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%562 : tensor<2x4096x320xf16>) outs(%563 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%565 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%564 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_899 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%566 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%565 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.truncf %cst_895 : f64 to f16
%10166 = arith.addf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<2x4096x1xf16>
%567 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%566 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.rsqrt %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%568 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%567 : tensor<2x4096x1xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x320xf16>
%569 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%561, %568 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%570 = tensor.empty() : tensor<2560x20x16xf16>
%571 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_51 : tensor<2560x20x1xf16>) outs(%570 : tensor<2560x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2560x20x16xf16>
%collapsed_1000 = tensor.collapse_shape %571 [[0], [1, 2]] : tensor<2560x20x16xf16> into tensor<2560x320xf16>
%572 = tensor.empty() : tensor<2560x320xf16>
%573 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_52, %collapsed_1000 : tensor<2560x320xi8>, tensor<2560x320xf16>) outs(%572 : tensor<2560x320xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<2560x320xf16>
%574 = tensor.empty() : tensor<320x2560xf16>
%575 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%573 : tensor<2560x320xf16>) outs(%574 : tensor<320x2560xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x2560xf16>
%collapsed_1001 = tensor.collapse_shape %569 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%576 = tensor.empty() : tensor<8192x2560xf16>
%577 = linalg.fill ins(%cst_889 : f16) outs(%576 : tensor<8192x2560xf16>) -> tensor<8192x2560xf16>
%578 = linalg.matmul ins(%collapsed_1001, %575 : tensor<8192x320xf16>, tensor<320x2560xf16>) outs(%577 : tensor<8192x2560xf16>) -> tensor<8192x2560xf16>
%579 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_53, %578 : tensor<2560xf16>, tensor<8192x2560xf16>) outs(%576 : tensor<8192x2560xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<8192x2560xf16>
%expanded_1002 = tensor.expand_shape %579 [[0, 1], [2]] : tensor<8192x2560xf16> into tensor<2x4096x2560xf16>
%extracted_slice_1003 = tensor.extract_slice %expanded_1002[0, 0, 0] [2, 4096, 1280] [1, 1, 1] : tensor<2x4096x2560xf16> to tensor<2x4096x1280xf16>
%extracted_slice_1004 = tensor.extract_slice %expanded_1002[0, 0, 1280] [2, 4096, 1280] [1, 1, 1] : tensor<2x4096x2560xf16> to tensor<2x4096x1280xf16>
%580 = tensor.empty() : tensor<2x4096x1280xf16>
%581 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1004 : tensor<2x4096x1280xf16>) outs(%580 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.sqrt %cst_892 : f16
%10166 = arith.divf %in, %10165 : f16
%10167 = math.erf %10166 : f16
%10168 = arith.addf %10167, %cst_890 : f16
%10169 = arith.mulf %10168, %cst_893 : f16
%10170 = arith.mulf %in, %10169 : f16
linalg.yield %10170 : f16
} -> tensor<2x4096x1280xf16>
%582 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1003, %581 : tensor<2x4096x1280xf16>, tensor<2x4096x1280xf16>) outs(%580 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1280xf16>
%583 = tensor.empty() : tensor<320x80x16xf16>
%584 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_54 : tensor<320x80x1xf16>) outs(%583 : tensor<320x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x80x16xf16>
%collapsed_1005 = tensor.collapse_shape %584 [[0], [1, 2]] : tensor<320x80x16xf16> into tensor<320x1280xf16>
%585 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_55, %collapsed_1005 : tensor<320x1280xi8>, tensor<320x1280xf16>) outs(%28 : tensor<320x1280xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x1280xf16>
%586 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%585 : tensor<320x1280xf16>) outs(%111 : tensor<1280x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x320xf16>
%collapsed_1006 = tensor.collapse_shape %582 [[0, 1], [2]] : tensor<2x4096x1280xf16> into tensor<8192x1280xf16>
%587 = linalg.fill ins(%cst_889 : f16) outs(%279 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%588 = linalg.matmul ins(%collapsed_1006, %586 : tensor<8192x1280xf16>, tensor<1280x320xf16>) outs(%587 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%589 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_56, %588 : tensor<320xf16>, tensor<8192x320xf16>) outs(%279 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<8192x320xf16>
%expanded_1007 = tensor.expand_shape %589 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16>
%590 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1007, %556 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%expanded_1008 = tensor.expand_shape %590 [[0], [1, 2], [3]] : tensor<2x4096x320xf16> into tensor<2x64x64x320xf16>
%591 = linalg.generic {indexing_maps = [#map11, #map26], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1008 : tensor<2x64x64x320xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%expanded_1009 = tensor.expand_shape %591 [[0], [1, 2], [3], [4]] : tensor<2x320x64x64xf16> into tensor<2x20x16x64x64xf16>
%592 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1009 : tensor<2x20x16x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x20x16x64x64xf16>
%593 = linalg.fill ins(%cst_891 : f16) outs(%78 : tensor<2x20x1x64x64xf16>) -> tensor<2x20x1x64x64xf16>
%594 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%592 : tensor<2x20x16x64x64xf16>) outs(%593 : tensor<2x20x1x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x20x1x64x64xf16>
%595 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%594 : tensor<2x20x1x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x64x64xf16>
%collapsed_1010 = tensor.collapse_shape %595 [[0], [1, 2], [3], [4]] : tensor<2x20x16x64x64xf16> into tensor<2x320x64x64xf16>
%596 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%597 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %596 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%598 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%597 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%599 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%598 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%600 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1010, %599 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%601 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%591, %600 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%602 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%601, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%603 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%604 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %603 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%605 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%604 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%606 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%607 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %606 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%608 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%607, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%609 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%602, %608 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%610 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%609, %608, %602 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%611 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%610, %605 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%612 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%611, %605, %610 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%613 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%612 : tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%614 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%613, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%615 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%614, %600 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%616 = linalg.generic {indexing_maps = [#map17, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_57 : tensor<320x20x1x1x1xf16>) outs(%223 : tensor<320x20x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x1x1xf16>
%617 = linalg.generic {indexing_maps = [#map18, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%616 : tensor<320x20x16x1x1xf16>) outs(%223 : tensor<320x20x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x1x1xf16>
%collapsed_1011 = tensor.collapse_shape %617 [[0], [1, 2], [3], [4]] : tensor<320x20x16x1x1xf16> into tensor<320x320x1x1xf16>
%618 = linalg.generic {indexing_maps = [#map13, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_58, %collapsed_1011 : tensor<320x320x1x1xi8>, tensor<320x320x1x1xf16>) outs(%226 : tensor<320x320x1x1xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320x1x1xf16>
%619 = 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>
%620 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%615, %618 : tensor<2x320x64x64xf16>, tensor<320x320x1x1xf16>) outs(%619 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16>
%621 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%620, %174 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%collapsed_1012 = tensor.collapse_shape %621 [[0], [1], [2, 3]] : tensor<2x320x64x64xf16> into tensor<2x320x4096xf16>
%expanded_1013 = tensor.expand_shape %collapsed_1012 [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16>
%622 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%623 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%622 : tensor<f64>) outs(%9 : tensor<f32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<f32>
%624 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%623 : tensor<f32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<2x32x10x4096xf32>
%625 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1013 : tensor<2x32x10x4096xf16>) outs(%624 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %out: f32):
%10165 = arith.extf %in : f16 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%626 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%627 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%625 : tensor<2x32x10x4096xf32>) outs(%626 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%628 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%627 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_897 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%629 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%625, %628 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.subf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%630 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%629 : tensor<2x32x10x4096xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.mulf %in, %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%631 = linalg.fill ins(%cst_894 : f64) outs(%55 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
%632 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%630 : tensor<2x32x10x4096xf32>) outs(%631 : tensor<2x32x1x1xf64>) {
^bb0(%in: f32, %out: f64):
%10165 = arith.extf %in : f32 to f64
%10166 = arith.addf %10165, %out : f64
linalg.yield %10166 : f64
} -> tensor<2x32x1x1xf64>
%633 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%632 : tensor<2x32x1x1xf64>) outs(%55 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%10165 = arith.divf %in, %cst_898 : f64
linalg.yield %10165 : f64
} -> tensor<2x32x1x1xf64>
%634 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%633 : tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%635 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%636 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%625 : tensor<2x32x10x4096xf32>) outs(%635 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%637 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%636 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_897 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%638 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%634, %cst_4 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x1x1xf32>
%639 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%638 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = math.rsqrt %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%640 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1013, %637 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %in_2540: f32, %out: f32):
%10165 = arith.extf %in : f16 to f32
%10166 = arith.subf %10165, %in_2540 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x10x4096xf32>
%641 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%640, %639 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.mulf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%collapsed_1014 = tensor.collapse_shape %641 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32>
%expanded_1015 = tensor.expand_shape %collapsed_1014 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32>
%expanded_1016 = tensor.expand_shape %cst_60 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16>
%642 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1015, %expanded_1016 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%67 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.mulf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x320x64x64xf32>
%expanded_1017 = tensor.expand_shape %cst_61 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16>
%643 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%642, %expanded_1017 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%67 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x320x64x64xf32>
%644 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%645 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%644 : tensor<f64>) outs(%23 : tensor<f16>) {
^bb0(%in: f64, %out: f16):
%10165 = arith.truncf %in : f64 to f16
linalg.yield %10165 : f16
} -> tensor<f16>
%646 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%645 : tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%647 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%643 : tensor<2x320x64x64xf32>) outs(%646 : tensor<2x320x64x64xf16>) {
^bb0(%in: f32, %out: f16):
%10165 = arith.truncf %in : f32 to f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%648 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%647 : tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
%10166 = math.exp %10165 : f16
%10167 = arith.addf %10166, %cst_890 : f16
%10168 = arith.divf %cst_890, %10167 : f16
linalg.yield %10168 : f16
} -> tensor<2x320x64x64xf16>
%649 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%648, %647 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%expanded_1018 = tensor.expand_shape %649 [[0], [1, 2], [3], [4]] : tensor<2x320x64x64xf16> into tensor<2x20x16x64x64xf16>
%650 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1018 : tensor<2x20x16x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x20x16x64x64xf16>
%651 = linalg.fill ins(%cst_891 : f16) outs(%78 : tensor<2x20x1x64x64xf16>) -> tensor<2x20x1x64x64xf16>
%652 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%650 : tensor<2x20x16x64x64xf16>) outs(%651 : tensor<2x20x1x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x20x1x64x64xf16>
%653 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%652 : tensor<2x20x1x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x64x64xf16>
%collapsed_1019 = tensor.collapse_shape %653 [[0], [1, 2], [3], [4]] : tensor<2x20x16x64x64xf16> into tensor<2x320x64x64xf16>
%654 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%655 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %654 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%656 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%655 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%657 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%656 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%658 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1019, %657 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%659 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%649, %658 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%660 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%659, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%661 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%662 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %661 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%663 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%662 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%664 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%665 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %664 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%666 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%665, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%667 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%660, %666 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%668 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%667, %666, %660 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%669 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%668, %663 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%670 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%669, %663, %668 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%671 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%670 : tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%672 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%671, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%673 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%672, %658 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%674 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_62 : tensor<320x20x1x3x3xf16>) outs(%103 : tensor<320x20x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x3x3xf16>
%collapsed_1020 = tensor.collapse_shape %674 [[0], [1, 2], [3], [4]] : tensor<320x20x16x3x3xf16> into tensor<320x320x3x3xf16>
%675 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_63, %collapsed_1020 : tensor<320x320x3x3xi8>, tensor<320x320x3x3xf16>) outs(%105 : tensor<320x320x3x3xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320x3x3xf16>
%padded_1021 = tensor.pad %673 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
tensor.yield %cst_889 : f16
} : tensor<2x320x64x64xf16> to tensor<2x320x66x66xf16>
%676 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_64 : tensor<320xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%677 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1021, %675 : tensor<2x320x66x66xf16>, tensor<320x320x3x3xf16>) outs(%676 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16>
%678 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
%10166 = math.exp %10165 : f16
%10167 = arith.addf %10166, %cst_890 : f16
%10168 = arith.divf %cst_890, %10167 : f16
linalg.yield %10168 : f16
} -> tensor<2x1280xf16>
%679 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%678, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280xf16>
%680 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_65 : tensor<320x1280xf16>) outs(%111 : tensor<1280x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x320xf16>
%681 = linalg.fill ins(%cst_889 : f16) outs(%25 : tensor<2x320xf16>) -> tensor<2x320xf16>
%682 = linalg.matmul ins(%679, %680 : tensor<2x1280xf16>, tensor<1280x320xf16>) outs(%681 : tensor<2x320xf16>) -> tensor<2x320xf16>
%683 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_66, %682 : tensor<320xf16>, tensor<2x320xf16>) outs(%25 : tensor<2x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320xf16>
%expanded_1022 = tensor.expand_shape %683 [[0], [1, 2, 3]] : tensor<2x320xf16> into tensor<2x320x1x1xf16>
%684 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%677, %expanded_1022 : tensor<2x320x64x64xf16>, tensor<2x320x1x1xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%collapsed_1023 = tensor.collapse_shape %684 [[0], [1], [2, 3]] : tensor<2x320x64x64xf16> into tensor<2x320x4096xf16>
%expanded_1024 = tensor.expand_shape %collapsed_1023 [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16>
%685 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%686 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%685 : tensor<f64>) outs(%9 : tensor<f32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<f32>
%687 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%686 : tensor<f32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<2x32x10x4096xf32>
%688 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1024 : tensor<2x32x10x4096xf16>) outs(%687 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %out: f32):
%10165 = arith.extf %in : f16 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%689 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%690 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%688 : tensor<2x32x10x4096xf32>) outs(%689 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%691 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%690 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_897 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%692 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%688, %691 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.subf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%693 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%692 : tensor<2x32x10x4096xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.mulf %in, %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%694 = linalg.fill ins(%cst_894 : f64) outs(%55 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
%695 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%693 : tensor<2x32x10x4096xf32>) outs(%694 : tensor<2x32x1x1xf64>) {
^bb0(%in: f32, %out: f64):
%10165 = arith.extf %in : f32 to f64
%10166 = arith.addf %10165, %out : f64
linalg.yield %10166 : f64
} -> tensor<2x32x1x1xf64>
%696 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%695 : tensor<2x32x1x1xf64>) outs(%55 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%10165 = arith.divf %in, %cst_898 : f64
linalg.yield %10165 : f64
} -> tensor<2x32x1x1xf64>
%697 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%696 : tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%698 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%699 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%688 : tensor<2x32x10x4096xf32>) outs(%698 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%700 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%699 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_897 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%701 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%697, %cst_4 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x1x1xf32>
%702 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%701 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = math.rsqrt %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%703 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1024, %700 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %in_2540: f32, %out: f32):
%10165 = arith.extf %in : f16 to f32
%10166 = arith.subf %10165, %in_2540 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x10x4096xf32>
%704 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%703, %702 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.mulf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%collapsed_1025 = tensor.collapse_shape %704 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32>
%expanded_1026 = tensor.expand_shape %collapsed_1025 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32>
%expanded_1027 = tensor.expand_shape %cst_67 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16>
%705 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1026, %expanded_1027 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%67 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.mulf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x320x64x64xf32>
%expanded_1028 = tensor.expand_shape %cst_68 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16>
%706 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%705, %expanded_1028 : tensor<2x320x64x64xf32>, tensor<320x1x1xf16>) outs(%67 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x320x64x64xf32>
%707 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%708 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%707 : tensor<f64>) outs(%23 : tensor<f16>) {
^bb0(%in: f64, %out: f16):
%10165 = arith.truncf %in : f64 to f16
linalg.yield %10165 : f16
} -> tensor<f16>
%709 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%708 : tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%710 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%706 : tensor<2x320x64x64xf32>) outs(%709 : tensor<2x320x64x64xf16>) {
^bb0(%in: f32, %out: f16):
%10165 = arith.truncf %in : f32 to f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%711 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%710 : tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
%10166 = math.exp %10165 : f16
%10167 = arith.addf %10166, %cst_890 : f16
%10168 = arith.divf %cst_890, %10167 : f16
linalg.yield %10168 : f16
} -> tensor<2x320x64x64xf16>
%712 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%711, %710 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%expanded_1029 = tensor.expand_shape %712 [[0], [1, 2], [3], [4]] : tensor<2x320x64x64xf16> into tensor<2x20x16x64x64xf16>
%713 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1029 : tensor<2x20x16x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x20x16x64x64xf16>
%714 = linalg.fill ins(%cst_891 : f16) outs(%78 : tensor<2x20x1x64x64xf16>) -> tensor<2x20x1x64x64xf16>
%715 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%713 : tensor<2x20x16x64x64xf16>) outs(%714 : tensor<2x20x1x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x20x1x64x64xf16>
%716 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%715 : tensor<2x20x1x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x64x64xf16>
%collapsed_1030 = tensor.collapse_shape %716 [[0], [1, 2], [3], [4]] : tensor<2x20x16x64x64xf16> into tensor<2x320x64x64xf16>
%717 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%718 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %717 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%719 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%718 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%720 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%719 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%721 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1030, %720 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%722 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%712, %721 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%723 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%722, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%724 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%725 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %724 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%726 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%725 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%727 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%728 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %727 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%729 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%728, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%730 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%723, %729 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%731 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%730, %729, %723 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%732 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%731, %726 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%733 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%732, %726, %731 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%734 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%733 : tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%735 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%734, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%736 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%735, %721 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%737 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_69 : tensor<320x20x1x3x3xf16>) outs(%103 : tensor<320x20x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x3x3xf16>
%collapsed_1031 = tensor.collapse_shape %737 [[0], [1, 2], [3], [4]] : tensor<320x20x16x3x3xf16> into tensor<320x320x3x3xf16>
%738 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_70, %collapsed_1031 : tensor<320x320x3x3xi8>, tensor<320x320x3x3xf16>) outs(%105 : tensor<320x320x3x3xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320x3x3xf16>
%padded_1032 = tensor.pad %736 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
tensor.yield %cst_889 : f16
} : tensor<2x320x64x64xf16> to tensor<2x320x66x66xf16>
%739 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_71 : tensor<320xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%740 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1032, %738 : tensor<2x320x66x66xf16>, tensor<320x320x3x3xf16>) outs(%739 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16>
%741 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%621, %740 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%742 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%741, %cst_3 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%collapsed_1033 = tensor.collapse_shape %742 [[0], [1], [2, 3]] : tensor<2x320x64x64xf16> into tensor<2x320x4096xf16>
%expanded_1034 = tensor.expand_shape %collapsed_1033 [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16>
%743 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%744 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%743 : tensor<f64>) outs(%9 : tensor<f32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<f32>
%745 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%744 : tensor<f32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<2x32x10x4096xf32>
%746 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1034 : tensor<2x32x10x4096xf16>) outs(%745 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %out: f32):
%10165 = arith.extf %in : f16 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%747 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%748 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%746 : tensor<2x32x10x4096xf32>) outs(%747 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%749 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%748 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_897 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%750 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%746, %749 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.subf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%751 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%750 : tensor<2x32x10x4096xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.mulf %in, %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%752 = linalg.fill ins(%cst_894 : f64) outs(%55 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
%753 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%751 : tensor<2x32x10x4096xf32>) outs(%752 : tensor<2x32x1x1xf64>) {
^bb0(%in: f32, %out: f64):
%10165 = arith.extf %in : f32 to f64
%10166 = arith.addf %10165, %out : f64
linalg.yield %10166 : f64
} -> tensor<2x32x1x1xf64>
%754 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%753 : tensor<2x32x1x1xf64>) outs(%55 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%10165 = arith.divf %in, %cst_898 : f64
linalg.yield %10165 : f64
} -> tensor<2x32x1x1xf64>
%755 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%754 : tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%756 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%757 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%746 : tensor<2x32x10x4096xf32>) outs(%756 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%758 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%757 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_897 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%759 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%755, %cst_2 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x1x1xf32>
%760 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%759 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = math.rsqrt %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%761 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1034, %758 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %in_2540: f32, %out: f32):
%10165 = arith.extf %in : f16 to f32
%10166 = arith.subf %10165, %in_2540 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x10x4096xf32>
%762 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%761, %760 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%46 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.mulf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x4096xf32>
%collapsed_1035 = tensor.collapse_shape %762 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32>
%expanded_1036 = tensor.expand_shape %collapsed_1035 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32>
%763 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%764 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%763 : tensor<f64>) outs(%23 : tensor<f16>) {
^bb0(%in: f64, %out: f16):
%10165 = arith.truncf %in : f64 to f16
linalg.yield %10165 : f16
} -> tensor<f16>
%765 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%764 : tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%766 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1036 : tensor<2x320x64x64xf32>) outs(%765 : tensor<2x320x64x64xf16>) {
^bb0(%in: f32, %out: f16):
%10165 = arith.truncf %in : f32 to f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%expanded_1037 = tensor.expand_shape %766 [[0], [1, 2], [3], [4]] : tensor<2x320x64x64xf16> into tensor<2x20x16x64x64xf16>
%767 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1037 : tensor<2x20x16x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x20x16x64x64xf16>
%768 = linalg.fill ins(%cst_891 : f16) outs(%78 : tensor<2x20x1x64x64xf16>) -> tensor<2x20x1x64x64xf16>
%769 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%767 : tensor<2x20x16x64x64xf16>) outs(%768 : tensor<2x20x1x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x20x1x64x64xf16>
%770 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%769 : tensor<2x20x1x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x64x64xf16>
%collapsed_1038 = tensor.collapse_shape %770 [[0], [1, 2], [3], [4]] : tensor<2x20x16x64x64xf16> into tensor<2x320x64x64xf16>
%771 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%772 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %771 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%773 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%772 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%774 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%773 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%775 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1038, %774 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%776 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%766, %775 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%777 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%776, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%778 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%779 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %778 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%780 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%779 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%781 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%782 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %781 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%783 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%782, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%784 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%777, %783 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%785 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%784, %783, %777 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%786 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%785, %780 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%787 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%786, %780, %785 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%788 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%787 : tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%789 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%788, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%790 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%789, %775 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%791 = linalg.generic {indexing_maps = [#map17, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_72 : tensor<320x20x1x1x1xf16>) outs(%223 : tensor<320x20x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x1x1xf16>
%792 = linalg.generic {indexing_maps = [#map18, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%791 : tensor<320x20x16x1x1xf16>) outs(%223 : tensor<320x20x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x1x1xf16>
%collapsed_1039 = tensor.collapse_shape %792 [[0], [1, 2], [3], [4]] : tensor<320x20x16x1x1xf16> into tensor<320x320x1x1xf16>
%793 = linalg.generic {indexing_maps = [#map13, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_73, %collapsed_1039 : tensor<320x320x1x1xi8>, tensor<320x320x1x1xf16>) outs(%226 : tensor<320x320x1x1xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320x1x1xf16>
%794 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_74 : tensor<320xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%795 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%790, %793 : tensor<2x320x64x64xf16>, tensor<320x320x1x1xf16>) outs(%794 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16>
%796 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%795 : tensor<2x320x64x64xf16>) outs(%230 : tensor<2x64x64x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x64x64x320xf16>
%collapsed_1040 = tensor.collapse_shape %796 [[0], [1, 2], [3]] : tensor<2x64x64x320xf16> into tensor<2x4096x320xf16>
%797 = linalg.fill ins(%cst_889 : f16) outs(%232 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16>
%798 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_1040 : tensor<2x4096x320xf16>) outs(%797 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%799 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%798 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_899 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%800 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%799 : tensor<2x4096x1xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x320xf16>
%801 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1040, %800 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%802 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%801 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.mulf %in, %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%803 = linalg.fill ins(%cst_889 : f16) outs(%232 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16>
%804 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%802 : tensor<2x4096x320xf16>) outs(%803 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%805 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%804 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_899 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%806 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%805 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.truncf %cst_895 : f64 to f16
%10166 = arith.addf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<2x4096x1xf16>
%807 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%806 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.rsqrt %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%808 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%807 : tensor<2x4096x1xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x320xf16>
%809 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%801, %808 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%expanded_1041 = tensor.expand_shape %809 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%810 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1041 : tensor<2x4096x20x16xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x16xf16>
%811 = linalg.fill ins(%cst_891 : f16) outs(%249 : tensor<2x4096x20x1xf16>) -> tensor<2x4096x20x1xf16>
%812 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%810 : tensor<2x4096x20x16xf16>) outs(%811 : tensor<2x4096x20x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x1xf16>
%813 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%812 : tensor<2x4096x20x1xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1042 = tensor.collapse_shape %813 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%814 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%815 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %814 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%816 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%815 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%817 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%816 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%818 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1042, %817 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%819 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%809, %818 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%820 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%819, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%821 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%822 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %821 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%823 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%822 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%824 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%825 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %824 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%826 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%825, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%827 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%820, %826 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%828 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%827, %826, %820 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%829 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%828, %823 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%830 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%829, %823, %828 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%831 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%830 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%832 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%831, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%833 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%832, %818 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%834 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_75 : tensor<320x20x1xf16>) outs(%274 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1043 = tensor.collapse_shape %834 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%835 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_76, %collapsed_1043 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320xf16>
%836 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%835 : tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1044 = tensor.collapse_shape %833 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%837 = linalg.fill ins(%cst_889 : f16) outs(%279 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%838 = linalg.matmul ins(%collapsed_1044, %836 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%837 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%839 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_77, %838 : tensor<320xf16>, tensor<8192x320xf16>) outs(%279 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<8192x320xf16>
%840 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1041 : tensor<2x4096x20x16xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x16xf16>
%841 = linalg.fill ins(%cst_891 : f16) outs(%249 : tensor<2x4096x20x1xf16>) -> tensor<2x4096x20x1xf16>
%842 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%840 : tensor<2x4096x20x16xf16>) outs(%841 : tensor<2x4096x20x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x1xf16>
%843 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%842 : tensor<2x4096x20x1xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1045 = tensor.collapse_shape %843 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%844 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%845 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %844 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%846 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%845 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%847 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%846 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%848 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1045, %847 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%849 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%809, %848 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%850 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%849, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%851 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%852 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %851 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%853 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%852 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%854 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%855 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %854 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%856 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%855, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%857 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%850, %856 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%858 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%857, %856, %850 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%859 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%858, %853 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%860 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%859, %853, %858 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%861 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%860 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%862 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%861, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%863 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%862, %848 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%864 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_78 : tensor<320x20x1xf16>) outs(%274 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1046 = tensor.collapse_shape %864 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%865 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_79, %collapsed_1046 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320xf16>
%866 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%865 : tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1047 = tensor.collapse_shape %863 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%867 = linalg.fill ins(%cst_889 : f16) outs(%279 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%868 = linalg.matmul ins(%collapsed_1047, %866 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%867 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%869 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_80, %868 : tensor<320xf16>, tensor<8192x320xf16>) outs(%279 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<8192x320xf16>
%870 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1041 : tensor<2x4096x20x16xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x16xf16>
%871 = linalg.fill ins(%cst_891 : f16) outs(%249 : tensor<2x4096x20x1xf16>) -> tensor<2x4096x20x1xf16>
%872 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%870 : tensor<2x4096x20x16xf16>) outs(%871 : tensor<2x4096x20x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x1xf16>
%873 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%872 : tensor<2x4096x20x1xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1048 = tensor.collapse_shape %873 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%874 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%875 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %874 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%876 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%875 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%877 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%876 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%878 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1048, %877 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%879 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%809, %878 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%880 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%879, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%881 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%882 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %881 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%883 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%882 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%884 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%885 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %884 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%886 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%885, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%887 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%880, %886 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%888 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%887, %886, %880 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%889 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%888, %883 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%890 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%889, %883, %888 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%891 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%890 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%892 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%891, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%893 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%892, %878 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%894 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_81 : tensor<320x20x1xf16>) outs(%274 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1049 = tensor.collapse_shape %894 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%895 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_82, %collapsed_1049 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320xf16>
%896 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%895 : tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1050 = tensor.collapse_shape %893 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%897 = linalg.fill ins(%cst_889 : f16) outs(%279 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%898 = linalg.matmul ins(%collapsed_1050, %896 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%897 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%899 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_83, %898 : tensor<320xf16>, tensor<8192x320xf16>) outs(%279 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<8192x320xf16>
%expanded_1051 = tensor.expand_shape %839 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x8x40xf16>
%900 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1051 : tensor<2x4096x8x40xf16>) outs(%343 : tensor<2x8x4096x40xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x4096x40xf16>
%collapsed_1052 = tensor.collapse_shape %900 [[0, 1], [2], [3]] : tensor<2x8x4096x40xf16> into tensor<16x4096x40xf16>
%expanded_1053 = tensor.expand_shape %869 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x8x40xf16>
%901 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1053 : tensor<2x4096x8x40xf16>) outs(%343 : tensor<2x8x4096x40xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x4096x40xf16>
%collapsed_1054 = tensor.collapse_shape %901 [[0, 1], [2], [3]] : tensor<2x8x4096x40xf16> into tensor<16x4096x40xf16>
%expanded_1055 = tensor.expand_shape %899 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x8x40xf16>
%902 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1055 : tensor<2x4096x8x40xf16>) outs(%343 : tensor<2x8x4096x40xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x4096x40xf16>
%collapsed_1056 = tensor.collapse_shape %902 [[0, 1], [2], [3]] : tensor<2x8x4096x40xf16> into tensor<16x4096x40xf16>
%903 = linalg.generic {indexing_maps = [#map20, #map25], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1054 : tensor<16x4096x40xf16>) outs(%347 : tensor<16x40x4096xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<16x40x4096xf16>
%904 = linalg.fill ins(%cst_889 : f16) outs(%349 : tensor<16x4096x4096xf16>) -> tensor<16x4096x4096xf16>
%905 = linalg.batch_matmul ins(%collapsed_1052, %903 : tensor<16x4096x40xf16>, tensor<16x40x4096xf16>) outs(%904 : tensor<16x4096x4096xf16>) -> tensor<16x4096x4096xf16>
%906 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%905, %cst_1 : tensor<16x4096x4096xf16>, tensor<f16>) outs(%349 : tensor<16x4096x4096xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x4096xf16>
%907 = linalg.fill ins(%cst_891 : f16) outs(%353 : tensor<16x4096x1xf16>) -> tensor<16x4096x1xf16>
%908 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%906 : tensor<16x4096x4096xf16>) outs(%907 : tensor<16x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x1xf16>
%909 = linalg.generic {indexing_maps = [#map20, #map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%906, %908 : tensor<16x4096x4096xf16>, tensor<16x4096x1xf16>) outs(%349 : tensor<16x4096x4096xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x4096xf16>
%910 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%909 : tensor<16x4096x4096xf16>) outs(%349 : tensor<16x4096x4096xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.exp %in : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x4096xf16>
%911 = linalg.fill ins(%cst_889 : f16) outs(%353 : tensor<16x4096x1xf16>) -> tensor<16x4096x1xf16>
%912 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%910 : tensor<16x4096x4096xf16>) outs(%911 : tensor<16x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x1xf16>
%913 = linalg.generic {indexing_maps = [#map20, #map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%910, %912 : tensor<16x4096x4096xf16>, tensor<16x4096x1xf16>) outs(%349 : tensor<16x4096x4096xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x4096xf16>
%914 = linalg.fill ins(%cst_889 : f16) outs(%361 : tensor<16x4096x40xf16>) -> tensor<16x4096x40xf16>
%915 = linalg.batch_matmul ins(%913, %collapsed_1056 : tensor<16x4096x4096xf16>, tensor<16x4096x40xf16>) outs(%914 : tensor<16x4096x40xf16>) -> tensor<16x4096x40xf16>
%expanded_1057 = tensor.expand_shape %915 [[0, 1], [2], [3]] : tensor<16x4096x40xf16> into tensor<2x8x4096x40xf16>
%916 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1057 : tensor<2x8x4096x40xf16>) outs(%364 : tensor<2x4096x8x40xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x8x40xf16>
%collapsed_1058 = tensor.collapse_shape %916 [[0], [1], [2, 3]] : tensor<2x4096x8x40xf16> into tensor<2x4096x320xf16>
%expanded_1059 = tensor.expand_shape %collapsed_1058 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%917 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1059 : tensor<2x4096x20x16xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x16xf16>
%918 = linalg.fill ins(%cst_891 : f16) outs(%249 : tensor<2x4096x20x1xf16>) -> tensor<2x4096x20x1xf16>
%919 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%917 : tensor<2x4096x20x16xf16>) outs(%918 : tensor<2x4096x20x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x1xf16>
%920 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%919 : tensor<2x4096x20x1xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1060 = tensor.collapse_shape %920 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%921 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%922 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %921 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%923 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%922 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%924 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%923 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%925 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1060, %924 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%926 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1058, %925 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%927 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%926, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%928 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%929 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %928 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%930 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%929 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%931 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%932 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %931 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%933 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%932, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%934 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%927, %933 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%935 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%934, %933, %927 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%936 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%935, %930 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%937 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%936, %930, %935 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%938 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%937 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%939 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%938, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%940 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%939, %925 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%941 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_84 : tensor<320x20x1xf16>) outs(%274 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1061 = tensor.collapse_shape %941 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%942 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_85, %collapsed_1061 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320xf16>
%943 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%942 : tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1062 = tensor.collapse_shape %940 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%944 = linalg.fill ins(%cst_889 : f16) outs(%279 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%945 = linalg.matmul ins(%collapsed_1062, %943 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%944 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%946 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_86, %945 : tensor<320xf16>, tensor<8192x320xf16>) outs(%279 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<8192x320xf16>
%expanded_1063 = tensor.expand_shape %946 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16>
%947 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1063, %collapsed_1040 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%948 = linalg.fill ins(%cst_889 : f16) outs(%232 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16>
%949 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%947 : tensor<2x4096x320xf16>) outs(%948 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%950 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%949 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_899 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%951 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%950 : tensor<2x4096x1xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x320xf16>
%952 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%947, %951 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%953 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%952 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.mulf %in, %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%954 = linalg.fill ins(%cst_889 : f16) outs(%232 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16>
%955 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%953 : tensor<2x4096x320xf16>) outs(%954 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%956 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%955 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_899 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%957 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%956 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.truncf %cst_895 : f64 to f16
%10166 = arith.addf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<2x4096x1xf16>
%958 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%957 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.rsqrt %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%959 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%958 : tensor<2x4096x1xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x320xf16>
%960 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%952, %959 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%expanded_1064 = tensor.expand_shape %960 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%961 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1064 : tensor<2x4096x20x16xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x16xf16>
%962 = linalg.fill ins(%cst_891 : f16) outs(%249 : tensor<2x4096x20x1xf16>) -> tensor<2x4096x20x1xf16>
%963 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%961 : tensor<2x4096x20x16xf16>) outs(%962 : tensor<2x4096x20x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x1xf16>
%964 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%963 : tensor<2x4096x20x1xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1065 = tensor.collapse_shape %964 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%965 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%966 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %965 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%967 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%966 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%968 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%967 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%969 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1065, %968 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%970 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%960, %969 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%971 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%970, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%972 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%973 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %972 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%974 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%973 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%975 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%976 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %975 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%977 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%976, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%978 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%971, %977 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%979 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%978, %977, %971 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%980 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%979, %974 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%981 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%980, %974, %979 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%982 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%981 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%983 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%982, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%984 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%983, %969 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%985 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_87 : tensor<320x20x1xf16>) outs(%274 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1066 = tensor.collapse_shape %985 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%986 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_88, %collapsed_1066 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320xf16>
%987 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%986 : tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1067 = tensor.collapse_shape %984 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%988 = linalg.fill ins(%cst_889 : f16) outs(%279 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%989 = linalg.matmul ins(%collapsed_1067, %987 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%988 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%990 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_89, %989 : tensor<320xf16>, tensor<8192x320xf16>) outs(%279 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<8192x320xf16>
%991 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_980 : tensor<2x77x48x16xf16>) outs(%440 : tensor<2x77x48x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x77x48x16xf16>
%992 = linalg.fill ins(%cst_891 : f16) outs(%442 : tensor<2x77x48x1xf16>) -> tensor<2x77x48x1xf16>
%993 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%991 : tensor<2x77x48x16xf16>) outs(%992 : tensor<2x77x48x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x77x48x1xf16>
%994 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%993 : tensor<2x77x48x1xf16>) outs(%440 : tensor<2x77x48x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x77x48x16xf16>
%collapsed_1068 = tensor.collapse_shape %994 [[0], [1], [2, 3]] : tensor<2x77x48x16xf16> into tensor<2x77x768xf16>
%995 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%996 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %995 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%997 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%996 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%998 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%997 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%999 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1068, %998 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1000 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2, %999 : tensor<2x77x768xf16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1001 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1000, %cst_880 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1002 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1003 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1002 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1004 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1003 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1005 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1006 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1005 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1007 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1006, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1008 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1001, %1007 : tensor<2x77x768xf16>, tensor<f16>) outs(%460 : tensor<2x77x768xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x77x768xi1>
%1009 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1008, %1007, %1001 : tensor<2x77x768xi1>, tensor<f16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1010 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1009, %1004 : tensor<2x77x768xf16>, tensor<f16>) outs(%460 : tensor<2x77x768xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x77x768xi1>
%1011 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1010, %1004, %1009 : tensor<2x77x768xi1>, tensor<f16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1012 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1011 : tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1013 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1012, %cst_880 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1014 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1013, %999 : tensor<2x77x768xf16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1015 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_90 : tensor<320x48x1xf16>) outs(%468 : tensor<320x48x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x48x16xf16>
%collapsed_1069 = tensor.collapse_shape %1015 [[0], [1, 2]] : tensor<320x48x16xf16> into tensor<320x768xf16>
%1016 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_91, %collapsed_1069 : tensor<320x768xi8>, tensor<320x768xf16>) outs(%470 : tensor<320x768xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x768xf16>
%1017 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%1016 : tensor<320x768xf16>) outs(%472 : tensor<768x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<768x320xf16>
%collapsed_1070 = tensor.collapse_shape %1014 [[0, 1], [2]] : tensor<2x77x768xf16> into tensor<154x768xf16>
%1018 = linalg.fill ins(%cst_889 : f16) outs(%474 : tensor<154x320xf16>) -> tensor<154x320xf16>
%1019 = linalg.matmul ins(%collapsed_1070, %1017 : tensor<154x768xf16>, tensor<768x320xf16>) outs(%1018 : tensor<154x320xf16>) -> tensor<154x320xf16>
%1020 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_980 : tensor<2x77x48x16xf16>) outs(%440 : tensor<2x77x48x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x77x48x16xf16>
%1021 = linalg.fill ins(%cst_891 : f16) outs(%442 : tensor<2x77x48x1xf16>) -> tensor<2x77x48x1xf16>
%1022 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1020 : tensor<2x77x48x16xf16>) outs(%1021 : tensor<2x77x48x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x77x48x1xf16>
%1023 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1022 : tensor<2x77x48x1xf16>) outs(%440 : tensor<2x77x48x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x77x48x16xf16>
%collapsed_1071 = tensor.collapse_shape %1023 [[0], [1], [2, 3]] : tensor<2x77x48x16xf16> into tensor<2x77x768xf16>
%1024 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1025 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1024 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1026 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1025 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1027 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1026 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1028 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1071, %1027 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1029 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2, %1028 : tensor<2x77x768xf16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1030 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1029, %cst_880 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1031 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1032 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1031 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1033 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1032 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1034 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1035 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1034 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1036 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1035, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1037 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1030, %1036 : tensor<2x77x768xf16>, tensor<f16>) outs(%460 : tensor<2x77x768xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x77x768xi1>
%1038 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1037, %1036, %1030 : tensor<2x77x768xi1>, tensor<f16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1039 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1038, %1033 : tensor<2x77x768xf16>, tensor<f16>) outs(%460 : tensor<2x77x768xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x77x768xi1>
%1040 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1039, %1033, %1038 : tensor<2x77x768xi1>, tensor<f16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1041 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1040 : tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1042 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1041, %cst_880 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1043 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1042, %1028 : tensor<2x77x768xf16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1044 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_92 : tensor<320x48x1xf16>) outs(%468 : tensor<320x48x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x48x16xf16>
%collapsed_1072 = tensor.collapse_shape %1044 [[0], [1, 2]] : tensor<320x48x16xf16> into tensor<320x768xf16>
%1045 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_93, %collapsed_1072 : tensor<320x768xi8>, tensor<320x768xf16>) outs(%470 : tensor<320x768xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x768xf16>
%1046 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%1045 : tensor<320x768xf16>) outs(%472 : tensor<768x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<768x320xf16>
%collapsed_1073 = tensor.collapse_shape %1043 [[0, 1], [2]] : tensor<2x77x768xf16> into tensor<154x768xf16>
%1047 = linalg.fill ins(%cst_889 : f16) outs(%474 : tensor<154x320xf16>) -> tensor<154x320xf16>
%1048 = linalg.matmul ins(%collapsed_1073, %1046 : tensor<154x768xf16>, tensor<768x320xf16>) outs(%1047 : tensor<154x320xf16>) -> tensor<154x320xf16>
%expanded_1074 = tensor.expand_shape %990 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x8x40xf16>
%1049 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1074 : tensor<2x4096x8x40xf16>) outs(%343 : tensor<2x8x4096x40xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x4096x40xf16>
%collapsed_1075 = tensor.collapse_shape %1049 [[0, 1], [2], [3]] : tensor<2x8x4096x40xf16> into tensor<16x4096x40xf16>
%expanded_1076 = tensor.expand_shape %1019 [[0, 1], [2, 3]] : tensor<154x320xf16> into tensor<2x77x8x40xf16>
%1050 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1076 : tensor<2x77x8x40xf16>) outs(%507 : tensor<2x8x77x40xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x77x40xf16>
%collapsed_1077 = tensor.collapse_shape %1050 [[0, 1], [2], [3]] : tensor<2x8x77x40xf16> into tensor<16x77x40xf16>
%expanded_1078 = tensor.expand_shape %1048 [[0, 1], [2, 3]] : tensor<154x320xf16> into tensor<2x77x8x40xf16>
%1051 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1078 : tensor<2x77x8x40xf16>) outs(%507 : tensor<2x8x77x40xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x77x40xf16>
%collapsed_1079 = tensor.collapse_shape %1051 [[0, 1], [2], [3]] : tensor<2x8x77x40xf16> into tensor<16x77x40xf16>
%1052 = linalg.generic {indexing_maps = [#map20, #map25], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1077 : tensor<16x77x40xf16>) outs(%510 : tensor<16x40x77xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<16x40x77xf16>
%1053 = linalg.fill ins(%cst_889 : f16) outs(%512 : tensor<16x4096x77xf16>) -> tensor<16x4096x77xf16>
%1054 = linalg.batch_matmul ins(%collapsed_1075, %1052 : tensor<16x4096x40xf16>, tensor<16x40x77xf16>) outs(%1053 : tensor<16x4096x77xf16>) -> tensor<16x4096x77xf16>
%1055 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1054, %cst_1 : tensor<16x4096x77xf16>, tensor<f16>) outs(%512 : tensor<16x4096x77xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x77xf16>
%1056 = linalg.fill ins(%cst_891 : f16) outs(%353 : tensor<16x4096x1xf16>) -> tensor<16x4096x1xf16>
%1057 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1055 : tensor<16x4096x77xf16>) outs(%1056 : tensor<16x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x1xf16>
%1058 = linalg.generic {indexing_maps = [#map20, #map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1055, %1057 : tensor<16x4096x77xf16>, tensor<16x4096x1xf16>) outs(%512 : tensor<16x4096x77xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x77xf16>
%1059 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1058 : tensor<16x4096x77xf16>) outs(%512 : tensor<16x4096x77xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.exp %in : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x77xf16>
%1060 = linalg.fill ins(%cst_889 : f16) outs(%353 : tensor<16x4096x1xf16>) -> tensor<16x4096x1xf16>
%1061 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1059 : tensor<16x4096x77xf16>) outs(%1060 : tensor<16x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x1xf16>
%1062 = linalg.generic {indexing_maps = [#map20, #map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1059, %1061 : tensor<16x4096x77xf16>, tensor<16x4096x1xf16>) outs(%512 : tensor<16x4096x77xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x4096x77xf16>
%1063 = linalg.fill ins(%cst_889 : f16) outs(%361 : tensor<16x4096x40xf16>) -> tensor<16x4096x40xf16>
%1064 = linalg.batch_matmul ins(%1062, %collapsed_1079 : tensor<16x4096x77xf16>, tensor<16x77x40xf16>) outs(%1063 : tensor<16x4096x40xf16>) -> tensor<16x4096x40xf16>
%expanded_1080 = tensor.expand_shape %1064 [[0, 1], [2], [3]] : tensor<16x4096x40xf16> into tensor<2x8x4096x40xf16>
%1065 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1080 : tensor<2x8x4096x40xf16>) outs(%364 : tensor<2x4096x8x40xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x8x40xf16>
%collapsed_1081 = tensor.collapse_shape %1065 [[0], [1], [2, 3]] : tensor<2x4096x8x40xf16> into tensor<2x4096x320xf16>
%expanded_1082 = tensor.expand_shape %collapsed_1081 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%1066 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1082 : tensor<2x4096x20x16xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x16xf16>
%1067 = linalg.fill ins(%cst_891 : f16) outs(%249 : tensor<2x4096x20x1xf16>) -> tensor<2x4096x20x1xf16>
%1068 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1066 : tensor<2x4096x20x16xf16>) outs(%1067 : tensor<2x4096x20x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x20x1xf16>
%1069 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1068 : tensor<2x4096x20x1xf16>) outs(%247 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1083 = tensor.collapse_shape %1069 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%1070 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1071 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1070 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1072 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1071 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1073 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1072 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1074 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1083, %1073 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%1075 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1081, %1074 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%1076 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1075, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%1077 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1078 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1077 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1079 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1078 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1080 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1081 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1080 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1082 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1081, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1083 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1076, %1082 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%1084 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1083, %1082, %1076 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%1085 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1084, %1079 : tensor<2x4096x320xf16>, tensor<f16>) outs(%266 : tensor<2x4096x320xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x4096x320xi1>
%1086 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1085, %1079, %1084 : tensor<2x4096x320xi1>, tensor<f16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%1087 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1086 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%1088 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1087, %cst_880 : tensor<2x4096x320xf16>, tensor<f16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%1089 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1088, %1074 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%1090 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_94 : tensor<320x20x1xf16>) outs(%274 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1084 = tensor.collapse_shape %1090 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%1091 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_95, %collapsed_1084 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320xf16>
%1092 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%1091 : tensor<320x320xf16>) outs(%276 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1085 = tensor.collapse_shape %1089 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%1093 = linalg.fill ins(%cst_889 : f16) outs(%279 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%1094 = linalg.matmul ins(%collapsed_1085, %1092 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%1093 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%1095 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_96, %1094 : tensor<320xf16>, tensor<8192x320xf16>) outs(%279 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<8192x320xf16>
%expanded_1086 = tensor.expand_shape %1095 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16>
%1096 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1086, %947 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%1097 = linalg.fill ins(%cst_889 : f16) outs(%232 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16>
%1098 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1096 : tensor<2x4096x320xf16>) outs(%1097 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%1099 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1098 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_899 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%1100 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1099 : tensor<2x4096x1xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x320xf16>
%1101 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1096, %1100 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%1102 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1101 : tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.mulf %in, %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%1103 = linalg.fill ins(%cst_889 : f16) outs(%232 : tensor<2x4096x1xf16>) -> tensor<2x4096x1xf16>
%1104 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1102 : tensor<2x4096x320xf16>) outs(%1103 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%1105 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1104 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_899 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%1106 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1105 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.truncf %cst_895 : f64 to f16
%10166 = arith.addf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<2x4096x1xf16>
%1107 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1106 : tensor<2x4096x1xf16>) outs(%232 : tensor<2x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.rsqrt %in : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1xf16>
%1108 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1107 : tensor<2x4096x1xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x320xf16>
%1109 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1101, %1108 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%1110 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_97 : tensor<2560x20x1xf16>) outs(%570 : tensor<2560x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2560x20x16xf16>
%collapsed_1087 = tensor.collapse_shape %1110 [[0], [1, 2]] : tensor<2560x20x16xf16> into tensor<2560x320xf16>
%1111 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_98, %collapsed_1087 : tensor<2560x320xi8>, tensor<2560x320xf16>) outs(%572 : tensor<2560x320xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<2560x320xf16>
%1112 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%1111 : tensor<2560x320xf16>) outs(%574 : tensor<320x2560xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x2560xf16>
%collapsed_1088 = tensor.collapse_shape %1109 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%1113 = linalg.fill ins(%cst_889 : f16) outs(%576 : tensor<8192x2560xf16>) -> tensor<8192x2560xf16>
%1114 = linalg.matmul ins(%collapsed_1088, %1112 : tensor<8192x320xf16>, tensor<320x2560xf16>) outs(%1113 : tensor<8192x2560xf16>) -> tensor<8192x2560xf16>
%1115 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_99, %1114 : tensor<2560xf16>, tensor<8192x2560xf16>) outs(%576 : tensor<8192x2560xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<8192x2560xf16>
%expanded_1089 = tensor.expand_shape %1115 [[0, 1], [2]] : tensor<8192x2560xf16> into tensor<2x4096x2560xf16>
%extracted_slice_1090 = tensor.extract_slice %expanded_1089[0, 0, 0] [2, 4096, 1280] [1, 1, 1] : tensor<2x4096x2560xf16> to tensor<2x4096x1280xf16>
%extracted_slice_1091 = tensor.extract_slice %expanded_1089[0, 0, 1280] [2, 4096, 1280] [1, 1, 1] : tensor<2x4096x2560xf16> to tensor<2x4096x1280xf16>
%1116 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1091 : tensor<2x4096x1280xf16>) outs(%580 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.sqrt %cst_892 : f16
%10166 = arith.divf %in, %10165 : f16
%10167 = math.erf %10166 : f16
%10168 = arith.addf %10167, %cst_890 : f16
%10169 = arith.mulf %10168, %cst_893 : f16
%10170 = arith.mulf %in, %10169 : f16
linalg.yield %10170 : f16
} -> tensor<2x4096x1280xf16>
%1117 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1090, %1116 : tensor<2x4096x1280xf16>, tensor<2x4096x1280xf16>) outs(%580 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x1280xf16>
%1118 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_100 : tensor<320x80x1xf16>) outs(%583 : tensor<320x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x80x16xf16>
%collapsed_1092 = tensor.collapse_shape %1118 [[0], [1, 2]] : tensor<320x80x16xf16> into tensor<320x1280xf16>
%1119 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_101, %collapsed_1092 : tensor<320x1280xi8>, tensor<320x1280xf16>) outs(%28 : tensor<320x1280xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x1280xf16>
%1120 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%1119 : tensor<320x1280xf16>) outs(%111 : tensor<1280x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x320xf16>
%collapsed_1093 = tensor.collapse_shape %1117 [[0, 1], [2]] : tensor<2x4096x1280xf16> into tensor<8192x1280xf16>
%1121 = linalg.fill ins(%cst_889 : f16) outs(%279 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%1122 = linalg.matmul ins(%collapsed_1093, %1120 : tensor<8192x1280xf16>, tensor<1280x320xf16>) outs(%1121 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%1123 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_102, %1122 : tensor<320xf16>, tensor<8192x320xf16>) outs(%279 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<8192x320xf16>
%expanded_1094 = tensor.expand_shape %1123 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16>
%1124 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1094, %1096 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%236 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x4096x320xf16>
%expanded_1095 = tensor.expand_shape %1124 [[0], [1, 2], [3]] : tensor<2x4096x320xf16> into tensor<2x64x64x320xf16>
%1125 = linalg.generic {indexing_maps = [#map11, #map26], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1095 : tensor<2x64x64x320xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%expanded_1096 = tensor.expand_shape %1125 [[0], [1, 2], [3], [4]] : tensor<2x320x64x64xf16> into tensor<2x20x16x64x64xf16>
%1126 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1096 : tensor<2x20x16x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x20x16x64x64xf16>
%1127 = linalg.fill ins(%cst_891 : f16) outs(%78 : tensor<2x20x1x64x64xf16>) -> tensor<2x20x1x64x64xf16>
%1128 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1126 : tensor<2x20x16x64x64xf16>) outs(%1127 : tensor<2x20x1x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x20x1x64x64xf16>
%1129 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1128 : tensor<2x20x1x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x64x64xf16>
%collapsed_1097 = tensor.collapse_shape %1129 [[0], [1, 2], [3], [4]] : tensor<2x20x16x64x64xf16> into tensor<2x320x64x64xf16>
%1130 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1131 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1130 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1132 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1131 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1133 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1132 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1134 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1097, %1133 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%1135 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1125, %1134 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%1136 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1135, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%1137 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1138 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1137 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1139 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1138 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1140 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1141 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1140 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1142 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1141, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1143 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1136, %1142 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%1144 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1143, %1142, %1136 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%1145 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1144, %1139 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%1146 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1145, %1139, %1144 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%1147 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1146 : tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%1148 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1147, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%1149 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1148, %1134 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%1150 = linalg.generic {indexing_maps = [#map17, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_103 : tensor<320x20x1x1x1xf16>) outs(%223 : tensor<320x20x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x1x1xf16>
%1151 = linalg.generic {indexing_maps = [#map18, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1150 : tensor<320x20x16x1x1xf16>) outs(%223 : tensor<320x20x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x1x1xf16>
%collapsed_1098 = tensor.collapse_shape %1151 [[0], [1, 2], [3], [4]] : tensor<320x20x16x1x1xf16> into tensor<320x320x1x1xf16>
%1152 = linalg.generic {indexing_maps = [#map13, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_104, %collapsed_1098 : tensor<320x320x1x1xi8>, tensor<320x320x1x1xf16>) outs(%226 : tensor<320x320x1x1xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320x1x1xf16>
%1153 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_105 : tensor<320xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%1154 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1149, %1152 : tensor<2x320x64x64xf16>, tensor<320x320x1x1xf16>) outs(%1153 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16>
%1155 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1154, %742 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%expanded_1099 = tensor.expand_shape %1155 [[0], [1, 2], [3], [4]] : tensor<2x320x64x64xf16> into tensor<2x20x16x64x64xf16>
%1156 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1099 : tensor<2x20x16x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x20x16x64x64xf16>
%1157 = linalg.fill ins(%cst_891 : f16) outs(%78 : tensor<2x20x1x64x64xf16>) -> tensor<2x20x1x64x64xf16>
%1158 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1156 : tensor<2x20x16x64x64xf16>) outs(%1157 : tensor<2x20x1x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x20x1x64x64xf16>
%1159 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1158 : tensor<2x20x1x64x64xf16>) outs(%76 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x64x64xf16>
%collapsed_1100 = tensor.collapse_shape %1159 [[0], [1, 2], [3], [4]] : tensor<2x20x16x64x64xf16> into tensor<2x320x64x64xf16>
%1160 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1161 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1160 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1162 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1161 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1163 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1162 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1164 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1100, %1163 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%1165 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1155, %1164 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%1166 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1165, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%1167 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1168 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1167 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1169 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1168 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1170 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1171 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1170 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1172 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1171, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1173 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1166, %1172 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%1174 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1173, %1172, %1166 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%1175 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1174, %1169 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%95 : tensor<2x320x64x64xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x64x64xi1>
%1176 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1175, %1169, %1174 : tensor<2x320x64x64xi1>, tensor<f16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%1177 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1176 : tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%1178 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1177, %cst_880 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%1179 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1178, %1164 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%41 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x64x64xf16>
%1180 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_106 : tensor<320x20x1x3x3xf16>) outs(%103 : tensor<320x20x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x3x3xf16>
%collapsed_1101 = tensor.collapse_shape %1180 [[0], [1, 2], [3], [4]] : tensor<320x20x16x3x3xf16> into tensor<320x320x3x3xf16>
%1181 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_107, %collapsed_1101 : tensor<320x320x3x3xi8>, tensor<320x320x3x3xf16>) outs(%105 : tensor<320x320x3x3xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<320x320x3x3xf16>
%padded_1102 = tensor.pad %1179 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
tensor.yield %cst_889 : f16
} : tensor<2x320x64x64xf16> to tensor<2x320x66x66xf16>
%1182 = tensor.empty() : tensor<2x320x32x32xf16>
%1183 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_108 : tensor<320xf16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x32x32xf16>
%1184 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%padded_1102, %1181 : tensor<2x320x66x66xf16>, tensor<320x320x3x3xf16>) outs(%1183 : tensor<2x320x32x32xf16>) -> tensor<2x320x32x32xf16>
%collapsed_1103 = tensor.collapse_shape %1184 [[0], [1], [2, 3]] : tensor<2x320x32x32xf16> into tensor<2x320x1024xf16>
%expanded_1104 = tensor.expand_shape %collapsed_1103 [[0], [1, 2], [3]] : tensor<2x320x1024xf16> into tensor<2x32x10x1024xf16>
%1185 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%1186 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1185 : tensor<f64>) outs(%9 : tensor<f32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<f32>
%1187 = tensor.empty() : tensor<2x32x10x1024xf32>
%1188 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1186 : tensor<f32>) outs(%1187 : tensor<2x32x10x1024xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<2x32x10x1024xf32>
%1189 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1104 : tensor<2x32x10x1024xf16>) outs(%1188 : tensor<2x32x10x1024xf32>) {
^bb0(%in: f16, %out: f32):
%10165 = arith.extf %in : f16 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x1024xf32>
%1190 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%1191 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1189 : tensor<2x32x10x1024xf32>) outs(%1190 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1192 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1191 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_900 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1193 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1189, %1192 : tensor<2x32x10x1024xf32>, tensor<2x32x1x1xf32>) outs(%1187 : tensor<2x32x10x1024xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.subf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x1024xf32>
%1194 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1193 : tensor<2x32x10x1024xf32>) outs(%1187 : tensor<2x32x10x1024xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.mulf %in, %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x1024xf32>
%1195 = linalg.fill ins(%cst_894 : f64) outs(%55 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
%1196 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1194 : tensor<2x32x10x1024xf32>) outs(%1195 : tensor<2x32x1x1xf64>) {
^bb0(%in: f32, %out: f64):
%10165 = arith.extf %in : f32 to f64
%10166 = arith.addf %10165, %out : f64
linalg.yield %10166 : f64
} -> tensor<2x32x1x1xf64>
%1197 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1196 : tensor<2x32x1x1xf64>) outs(%55 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%10165 = arith.divf %in, %cst_901 : f64
linalg.yield %10165 : f64
} -> tensor<2x32x1x1xf64>
%1198 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1197 : tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1199 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%1200 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1189 : tensor<2x32x10x1024xf32>) outs(%1199 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1201 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1200 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_900 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1202 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1198, %cst_4 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x1x1xf32>
%1203 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1202 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = math.rsqrt %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1204 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1104, %1201 : tensor<2x32x10x1024xf16>, tensor<2x32x1x1xf32>) outs(%1187 : tensor<2x32x10x1024xf32>) {
^bb0(%in: f16, %in_2540: f32, %out: f32):
%10165 = arith.extf %in : f16 to f32
%10166 = arith.subf %10165, %in_2540 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x10x1024xf32>
%1205 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1204, %1203 : tensor<2x32x10x1024xf32>, tensor<2x32x1x1xf32>) outs(%1187 : tensor<2x32x10x1024xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.mulf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x10x1024xf32>
%collapsed_1105 = tensor.collapse_shape %1205 [[0], [1, 2], [3]] : tensor<2x32x10x1024xf32> into tensor<2x320x1024xf32>
%expanded_1106 = tensor.expand_shape %collapsed_1105 [[0], [1], [2, 3]] : tensor<2x320x1024xf32> into tensor<2x320x32x32xf32>
%expanded_1107 = tensor.expand_shape %cst_109 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16>
%1206 = tensor.empty() : tensor<2x320x32x32xf32>
%1207 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1106, %expanded_1107 : tensor<2x320x32x32xf32>, tensor<320x1x1xf16>) outs(%1206 : tensor<2x320x32x32xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.mulf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x320x32x32xf32>
%expanded_1108 = tensor.expand_shape %cst_110 [[0, 1, 2]] : tensor<320xf16> into tensor<320x1x1xf16>
%1208 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1207, %expanded_1108 : tensor<2x320x32x32xf32>, tensor<320x1x1xf16>) outs(%1206 : tensor<2x320x32x32xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x320x32x32xf32>
%1209 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%1210 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1209 : tensor<f64>) outs(%23 : tensor<f16>) {
^bb0(%in: f64, %out: f16):
%10165 = arith.truncf %in : f64 to f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1211 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1210 : tensor<f16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x32x32xf16>
%1212 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1208 : tensor<2x320x32x32xf32>) outs(%1211 : tensor<2x320x32x32xf16>) {
^bb0(%in: f32, %out: f16):
%10165 = arith.truncf %in : f32 to f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%1213 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1212 : tensor<2x320x32x32xf16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
%10166 = math.exp %10165 : f16
%10167 = arith.addf %10166, %cst_890 : f16
%10168 = arith.divf %cst_890, %10167 : f16
linalg.yield %10168 : f16
} -> tensor<2x320x32x32xf16>
%1214 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1213, %1212 : tensor<2x320x32x32xf16>, tensor<2x320x32x32xf16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%expanded_1109 = tensor.expand_shape %1214 [[0], [1, 2], [3], [4]] : tensor<2x320x32x32xf16> into tensor<2x20x16x32x32xf16>
%1215 = tensor.empty() : tensor<2x20x16x32x32xf16>
%1216 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1109 : tensor<2x20x16x32x32xf16>) outs(%1215 : tensor<2x20x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x20x16x32x32xf16>
%1217 = tensor.empty() : tensor<2x20x1x32x32xf16>
%1218 = linalg.fill ins(%cst_891 : f16) outs(%1217 : tensor<2x20x1x32x32xf16>) -> tensor<2x20x1x32x32xf16>
%1219 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1216 : tensor<2x20x16x32x32xf16>) outs(%1218 : tensor<2x20x1x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x20x1x32x32xf16>
%1220 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1219 : tensor<2x20x1x32x32xf16>) outs(%1215 : tensor<2x20x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x32x32xf16>
%collapsed_1110 = tensor.collapse_shape %1220 [[0], [1, 2], [3], [4]] : tensor<2x20x16x32x32xf16> into tensor<2x320x32x32xf16>
%1221 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1222 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1221 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1223 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1222 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1224 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1223 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1225 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1110, %1224 : tensor<2x320x32x32xf16>, tensor<f16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%1226 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1214, %1225 : tensor<2x320x32x32xf16>, tensor<2x320x32x32xf16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%1227 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1226, %cst_880 : tensor<2x320x32x32xf16>, tensor<f16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%1228 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1229 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1228 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1230 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1229 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1231 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1232 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1231 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1233 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1232, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1234 = tensor.empty() : tensor<2x320x32x32xi1>
%1235 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1227, %1233 : tensor<2x320x32x32xf16>, tensor<f16>) outs(%1234 : tensor<2x320x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x32x32xi1>
%1236 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1235, %1233, %1227 : tensor<2x320x32x32xi1>, tensor<f16>, tensor<2x320x32x32xf16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%1237 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1236, %1230 : tensor<2x320x32x32xf16>, tensor<f16>) outs(%1234 : tensor<2x320x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x32x32xi1>
%1238 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1237, %1230, %1236 : tensor<2x320x32x32xi1>, tensor<f16>, tensor<2x320x32x32xf16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%1239 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1238 : tensor<2x320x32x32xf16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%1240 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1239, %cst_880 : tensor<2x320x32x32xf16>, tensor<f16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%1241 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1240, %1225 : tensor<2x320x32x32xf16>, tensor<2x320x32x32xf16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%1242 = tensor.empty() : tensor<640x20x16x3x3xf16>
%1243 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_111 : tensor<640x20x1x3x3xf16>) outs(%1242 : tensor<640x20x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x20x16x3x3xf16>
%collapsed_1111 = tensor.collapse_shape %1243 [[0], [1, 2], [3], [4]] : tensor<640x20x16x3x3xf16> into tensor<640x320x3x3xf16>
%1244 = tensor.empty() : tensor<640x320x3x3xf16>
%1245 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_112, %collapsed_1111 : tensor<640x320x3x3xi8>, tensor<640x320x3x3xf16>) outs(%1244 : tensor<640x320x3x3xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x320x3x3xf16>
%padded_1112 = tensor.pad %1241 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
tensor.yield %cst_889 : f16
} : tensor<2x320x32x32xf16> to tensor<2x320x34x34xf16>
%1246 = tensor.empty() : tensor<2x640x32x32xf16>
%1247 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_113 : tensor<640xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%1248 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1112, %1245 : tensor<2x320x34x34xf16>, tensor<640x320x3x3xf16>) outs(%1247 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16>
%1249 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
%10166 = math.exp %10165 : f16
%10167 = arith.addf %10166, %cst_890 : f16
%10168 = arith.divf %cst_890, %10167 : f16
linalg.yield %10168 : f16
} -> tensor<2x1280xf16>
%1250 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%1249, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280xf16>
%1251 = tensor.empty() : tensor<1280x640xf16>
%1252 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_114 : tensor<640x1280xf16>) outs(%1251 : tensor<1280x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x640xf16>
%1253 = tensor.empty() : tensor<2x640xf16>
%1254 = linalg.fill ins(%cst_889 : f16) outs(%1253 : tensor<2x640xf16>) -> tensor<2x640xf16>
%1255 = linalg.matmul ins(%1250, %1252 : tensor<2x1280xf16>, tensor<1280x640xf16>) outs(%1254 : tensor<2x640xf16>) -> tensor<2x640xf16>
%1256 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_115, %1255 : tensor<640xf16>, tensor<2x640xf16>) outs(%1253 : tensor<2x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640xf16>
%expanded_1113 = tensor.expand_shape %1256 [[0], [1, 2, 3]] : tensor<2x640xf16> into tensor<2x640x1x1xf16>
%1257 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1248, %expanded_1113 : tensor<2x640x32x32xf16>, tensor<2x640x1x1xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%collapsed_1114 = tensor.collapse_shape %1257 [[0], [1], [2, 3]] : tensor<2x640x32x32xf16> into tensor<2x640x1024xf16>
%expanded_1115 = tensor.expand_shape %collapsed_1114 [[0], [1, 2], [3]] : tensor<2x640x1024xf16> into tensor<2x32x20x1024xf16>
%1258 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%1259 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1258 : tensor<f64>) outs(%9 : tensor<f32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<f32>
%1260 = tensor.empty() : tensor<2x32x20x1024xf32>
%1261 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1259 : tensor<f32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<2x32x20x1024xf32>
%1262 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1115 : tensor<2x32x20x1024xf16>) outs(%1261 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %out: f32):
%10165 = arith.extf %in : f16 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%1263 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%1264 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1262 : tensor<2x32x20x1024xf32>) outs(%1263 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1265 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1264 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_902 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1266 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1262, %1265 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.subf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%1267 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1266 : tensor<2x32x20x1024xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.mulf %in, %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%1268 = linalg.fill ins(%cst_894 : f64) outs(%55 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
%1269 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1267 : tensor<2x32x20x1024xf32>) outs(%1268 : tensor<2x32x1x1xf64>) {
^bb0(%in: f32, %out: f64):
%10165 = arith.extf %in : f32 to f64
%10166 = arith.addf %10165, %out : f64
linalg.yield %10166 : f64
} -> tensor<2x32x1x1xf64>
%1270 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1269 : tensor<2x32x1x1xf64>) outs(%55 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%10165 = arith.divf %in, %cst_903 : f64
linalg.yield %10165 : f64
} -> tensor<2x32x1x1xf64>
%1271 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1270 : tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1272 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%1273 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1262 : tensor<2x32x20x1024xf32>) outs(%1272 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1274 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1273 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_902 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1275 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1271, %cst_4 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x1x1xf32>
%1276 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1275 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = math.rsqrt %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1277 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1115, %1274 : tensor<2x32x20x1024xf16>, tensor<2x32x1x1xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %in_2540: f32, %out: f32):
%10165 = arith.extf %in : f16 to f32
%10166 = arith.subf %10165, %in_2540 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x20x1024xf32>
%1278 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1277, %1276 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.mulf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%collapsed_1116 = tensor.collapse_shape %1278 [[0], [1, 2], [3]] : tensor<2x32x20x1024xf32> into tensor<2x640x1024xf32>
%expanded_1117 = tensor.expand_shape %collapsed_1116 [[0], [1], [2, 3]] : tensor<2x640x1024xf32> into tensor<2x640x32x32xf32>
%expanded_1118 = tensor.expand_shape %cst_116 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16>
%1279 = tensor.empty() : tensor<2x640x32x32xf32>
%1280 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1117, %expanded_1118 : tensor<2x640x32x32xf32>, tensor<640x1x1xf16>) outs(%1279 : tensor<2x640x32x32xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.mulf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x640x32x32xf32>
%expanded_1119 = tensor.expand_shape %cst_117 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16>
%1281 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1280, %expanded_1119 : tensor<2x640x32x32xf32>, tensor<640x1x1xf16>) outs(%1279 : tensor<2x640x32x32xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x640x32x32xf32>
%1282 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%1283 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1282 : tensor<f64>) outs(%23 : tensor<f16>) {
^bb0(%in: f64, %out: f16):
%10165 = arith.truncf %in : f64 to f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1284 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1283 : tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%1285 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1281 : tensor<2x640x32x32xf32>) outs(%1284 : tensor<2x640x32x32xf16>) {
^bb0(%in: f32, %out: f16):
%10165 = arith.truncf %in : f32 to f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1286 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1285 : tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
%10166 = math.exp %10165 : f16
%10167 = arith.addf %10166, %cst_890 : f16
%10168 = arith.divf %cst_890, %10167 : f16
linalg.yield %10168 : f16
} -> tensor<2x640x32x32xf16>
%1287 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1286, %1285 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%expanded_1120 = tensor.expand_shape %1287 [[0], [1, 2], [3], [4]] : tensor<2x640x32x32xf16> into tensor<2x40x16x32x32xf16>
%1288 = tensor.empty() : tensor<2x40x16x32x32xf16>
%1289 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1120 : tensor<2x40x16x32x32xf16>) outs(%1288 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x40x16x32x32xf16>
%1290 = tensor.empty() : tensor<2x40x1x32x32xf16>
%1291 = linalg.fill ins(%cst_891 : f16) outs(%1290 : tensor<2x40x1x32x32xf16>) -> tensor<2x40x1x32x32xf16>
%1292 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1289 : tensor<2x40x16x32x32xf16>) outs(%1291 : tensor<2x40x1x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x40x1x32x32xf16>
%1293 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1292 : tensor<2x40x1x32x32xf16>) outs(%1288 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x40x16x32x32xf16>
%collapsed_1121 = tensor.collapse_shape %1293 [[0], [1, 2], [3], [4]] : tensor<2x40x16x32x32xf16> into tensor<2x640x32x32xf16>
%1294 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1295 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1294 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1296 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1295 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1297 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1296 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1298 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1121, %1297 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1299 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1287, %1298 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1300 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1299, %cst_880 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1301 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1302 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1301 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1303 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1302 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1304 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1305 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1304 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1306 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1305, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1307 = tensor.empty() : tensor<2x640x32x32xi1>
%1308 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1300, %1306 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1307 : tensor<2x640x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x32x32xi1>
%1309 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1308, %1306, %1300 : tensor<2x640x32x32xi1>, tensor<f16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1310 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1309, %1303 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1307 : tensor<2x640x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x32x32xi1>
%1311 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1310, %1303, %1309 : tensor<2x640x32x32xi1>, tensor<f16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1312 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1311 : tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1313 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1312, %cst_880 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1314 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1313, %1298 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1315 = tensor.empty() : tensor<640x40x16x3x3xf16>
%1316 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_118 : tensor<640x40x1x3x3xf16>) outs(%1315 : tensor<640x40x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16x3x3xf16>
%collapsed_1122 = tensor.collapse_shape %1316 [[0], [1, 2], [3], [4]] : tensor<640x40x16x3x3xf16> into tensor<640x640x3x3xf16>
%1317 = tensor.empty() : tensor<640x640x3x3xf16>
%1318 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_119, %collapsed_1122 : tensor<640x640x3x3xi8>, tensor<640x640x3x3xf16>) outs(%1317 : tensor<640x640x3x3xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640x3x3xf16>
%padded_1123 = tensor.pad %1314 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
tensor.yield %cst_889 : f16
} : tensor<2x640x32x32xf16> to tensor<2x640x34x34xf16>
%1319 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_120 : tensor<640xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%1320 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1123, %1318 : tensor<2x640x34x34xf16>, tensor<640x640x3x3xf16>) outs(%1319 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16>
%expanded_1124 = tensor.expand_shape %1184 [[0], [1, 2], [3], [4]] : tensor<2x320x32x32xf16> into tensor<2x20x16x32x32xf16>
%1321 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1124 : tensor<2x20x16x32x32xf16>) outs(%1215 : tensor<2x20x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x20x16x32x32xf16>
%1322 = linalg.fill ins(%cst_891 : f16) outs(%1217 : tensor<2x20x1x32x32xf16>) -> tensor<2x20x1x32x32xf16>
%1323 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1321 : tensor<2x20x16x32x32xf16>) outs(%1322 : tensor<2x20x1x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x20x1x32x32xf16>
%1324 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1323 : tensor<2x20x1x32x32xf16>) outs(%1215 : tensor<2x20x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x32x32xf16>
%collapsed_1125 = tensor.collapse_shape %1324 [[0], [1, 2], [3], [4]] : tensor<2x20x16x32x32xf16> into tensor<2x320x32x32xf16>
%1325 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1326 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1325 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1327 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1326 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1328 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1327 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1329 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1125, %1328 : tensor<2x320x32x32xf16>, tensor<f16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%1330 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1184, %1329 : tensor<2x320x32x32xf16>, tensor<2x320x32x32xf16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%1331 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1330, %cst_880 : tensor<2x320x32x32xf16>, tensor<f16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%1332 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1333 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1332 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1334 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1333 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1335 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1336 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1335 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1337 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1336, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1338 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1331, %1337 : tensor<2x320x32x32xf16>, tensor<f16>) outs(%1234 : tensor<2x320x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x32x32xi1>
%1339 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1338, %1337, %1331 : tensor<2x320x32x32xi1>, tensor<f16>, tensor<2x320x32x32xf16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%1340 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1339, %1334 : tensor<2x320x32x32xf16>, tensor<f16>) outs(%1234 : tensor<2x320x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x320x32x32xi1>
%1341 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1340, %1334, %1339 : tensor<2x320x32x32xi1>, tensor<f16>, tensor<2x320x32x32xf16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%1342 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1341 : tensor<2x320x32x32xf16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%1343 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1342, %cst_880 : tensor<2x320x32x32xf16>, tensor<f16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%1344 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1343, %1329 : tensor<2x320x32x32xf16>, tensor<2x320x32x32xf16>) outs(%1182 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x320x32x32xf16>
%1345 = tensor.empty() : tensor<640x20x16x1x1xf16>
%1346 = linalg.generic {indexing_maps = [#map17, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_121 : tensor<640x20x1x1x1xf16>) outs(%1345 : tensor<640x20x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x20x16x1x1xf16>
%1347 = linalg.generic {indexing_maps = [#map18, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1346 : tensor<640x20x16x1x1xf16>) outs(%1345 : tensor<640x20x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x20x16x1x1xf16>
%collapsed_1126 = tensor.collapse_shape %1347 [[0], [1, 2], [3], [4]] : tensor<640x20x16x1x1xf16> into tensor<640x320x1x1xf16>
%1348 = tensor.empty() : tensor<640x320x1x1xf16>
%1349 = linalg.generic {indexing_maps = [#map13, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_122, %collapsed_1126 : tensor<640x320x1x1xi8>, tensor<640x320x1x1xf16>) outs(%1348 : tensor<640x320x1x1xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x320x1x1xf16>
%1350 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_123 : tensor<640xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%1351 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1344, %1349 : tensor<2x320x32x32xf16>, tensor<640x320x1x1xf16>) outs(%1350 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16>
%1352 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1351, %1320 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1353 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1352, %cst_3 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%collapsed_1127 = tensor.collapse_shape %1353 [[0], [1], [2, 3]] : tensor<2x640x32x32xf16> into tensor<2x640x1024xf16>
%expanded_1128 = tensor.expand_shape %collapsed_1127 [[0], [1, 2], [3]] : tensor<2x640x1024xf16> into tensor<2x32x20x1024xf16>
%1354 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%1355 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1354 : tensor<f64>) outs(%9 : tensor<f32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<f32>
%1356 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1355 : tensor<f32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<2x32x20x1024xf32>
%1357 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1128 : tensor<2x32x20x1024xf16>) outs(%1356 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %out: f32):
%10165 = arith.extf %in : f16 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%1358 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%1359 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1357 : tensor<2x32x20x1024xf32>) outs(%1358 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1360 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1359 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_902 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1361 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1357, %1360 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.subf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%1362 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1361 : tensor<2x32x20x1024xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.mulf %in, %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%1363 = linalg.fill ins(%cst_894 : f64) outs(%55 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
%1364 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1362 : tensor<2x32x20x1024xf32>) outs(%1363 : tensor<2x32x1x1xf64>) {
^bb0(%in: f32, %out: f64):
%10165 = arith.extf %in : f32 to f64
%10166 = arith.addf %10165, %out : f64
linalg.yield %10166 : f64
} -> tensor<2x32x1x1xf64>
%1365 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1364 : tensor<2x32x1x1xf64>) outs(%55 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%10165 = arith.divf %in, %cst_903 : f64
linalg.yield %10165 : f64
} -> tensor<2x32x1x1xf64>
%1366 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1365 : tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1367 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%1368 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1357 : tensor<2x32x20x1024xf32>) outs(%1367 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1369 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1368 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_902 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1370 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1366, %cst_2 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x1x1xf32>
%1371 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1370 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = math.rsqrt %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1372 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1128, %1369 : tensor<2x32x20x1024xf16>, tensor<2x32x1x1xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %in_2540: f32, %out: f32):
%10165 = arith.extf %in : f16 to f32
%10166 = arith.subf %10165, %in_2540 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x20x1024xf32>
%1373 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1372, %1371 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.mulf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%collapsed_1129 = tensor.collapse_shape %1373 [[0], [1, 2], [3]] : tensor<2x32x20x1024xf32> into tensor<2x640x1024xf32>
%expanded_1130 = tensor.expand_shape %collapsed_1129 [[0], [1], [2, 3]] : tensor<2x640x1024xf32> into tensor<2x640x32x32xf32>
%1374 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%1375 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1374 : tensor<f64>) outs(%23 : tensor<f16>) {
^bb0(%in: f64, %out: f16):
%10165 = arith.truncf %in : f64 to f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1376 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1375 : tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%1377 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1130 : tensor<2x640x32x32xf32>) outs(%1376 : tensor<2x640x32x32xf16>) {
^bb0(%in: f32, %out: f16):
%10165 = arith.truncf %in : f32 to f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%expanded_1131 = tensor.expand_shape %1377 [[0], [1, 2], [3], [4]] : tensor<2x640x32x32xf16> into tensor<2x40x16x32x32xf16>
%1378 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1131 : tensor<2x40x16x32x32xf16>) outs(%1288 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x40x16x32x32xf16>
%1379 = linalg.fill ins(%cst_891 : f16) outs(%1290 : tensor<2x40x1x32x32xf16>) -> tensor<2x40x1x32x32xf16>
%1380 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1378 : tensor<2x40x16x32x32xf16>) outs(%1379 : tensor<2x40x1x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x40x1x32x32xf16>
%1381 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1380 : tensor<2x40x1x32x32xf16>) outs(%1288 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x40x16x32x32xf16>
%collapsed_1132 = tensor.collapse_shape %1381 [[0], [1, 2], [3], [4]] : tensor<2x40x16x32x32xf16> into tensor<2x640x32x32xf16>
%1382 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1383 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1382 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1384 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1383 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1385 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1384 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1386 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1132, %1385 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1387 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1377, %1386 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1388 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1387, %cst_880 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1389 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1390 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1389 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1391 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1390 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1392 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1393 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1392 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1394 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1393, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1395 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1388, %1394 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1307 : tensor<2x640x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x32x32xi1>
%1396 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1395, %1394, %1388 : tensor<2x640x32x32xi1>, tensor<f16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1397 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1396, %1391 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1307 : tensor<2x640x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x32x32xi1>
%1398 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1397, %1391, %1396 : tensor<2x640x32x32xi1>, tensor<f16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1399 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1398 : tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1400 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1399, %cst_880 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1401 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1400, %1386 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1402 = tensor.empty() : tensor<640x40x16x1x1xf16>
%1403 = linalg.generic {indexing_maps = [#map17, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_124 : tensor<640x40x1x1x1xf16>) outs(%1402 : tensor<640x40x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16x1x1xf16>
%1404 = linalg.generic {indexing_maps = [#map18, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1403 : tensor<640x40x16x1x1xf16>) outs(%1402 : tensor<640x40x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16x1x1xf16>
%collapsed_1133 = tensor.collapse_shape %1404 [[0], [1, 2], [3], [4]] : tensor<640x40x16x1x1xf16> into tensor<640x640x1x1xf16>
%1405 = tensor.empty() : tensor<640x640x1x1xf16>
%1406 = linalg.generic {indexing_maps = [#map13, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_125, %collapsed_1133 : tensor<640x640x1x1xi8>, tensor<640x640x1x1xf16>) outs(%1405 : tensor<640x640x1x1xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640x1x1xf16>
%1407 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_126 : tensor<640xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%1408 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1401, %1406 : tensor<2x640x32x32xf16>, tensor<640x640x1x1xf16>) outs(%1407 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16>
%1409 = tensor.empty() : tensor<2x32x32x640xf16>
%1410 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1408 : tensor<2x640x32x32xf16>) outs(%1409 : tensor<2x32x32x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x32x32x640xf16>
%collapsed_1134 = tensor.collapse_shape %1410 [[0], [1, 2], [3]] : tensor<2x32x32x640xf16> into tensor<2x1024x640xf16>
%1411 = tensor.empty() : tensor<2x1024x1xf16>
%1412 = linalg.fill ins(%cst_889 : f16) outs(%1411 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16>
%1413 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_1134 : tensor<2x1024x640xf16>) outs(%1412 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1414 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1413 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_904 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1415 = tensor.empty() : tensor<2x1024x640xf16>
%1416 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1414 : tensor<2x1024x1xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x640xf16>
%1417 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1134, %1416 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1418 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1417 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.mulf %in, %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1419 = linalg.fill ins(%cst_889 : f16) outs(%1411 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16>
%1420 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1418 : tensor<2x1024x640xf16>) outs(%1419 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1421 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1420 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_904 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1422 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1421 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.truncf %cst_895 : f64 to f16
%10166 = arith.addf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<2x1024x1xf16>
%1423 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1422 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.rsqrt %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1424 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1423 : tensor<2x1024x1xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x640xf16>
%1425 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1417, %1424 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%expanded_1135 = tensor.expand_shape %1425 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%1426 = tensor.empty() : tensor<2x1024x40x16xf16>
%1427 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1135 : tensor<2x1024x40x16xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x16xf16>
%1428 = tensor.empty() : tensor<2x1024x40x1xf16>
%1429 = linalg.fill ins(%cst_891 : f16) outs(%1428 : tensor<2x1024x40x1xf16>) -> tensor<2x1024x40x1xf16>
%1430 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1427 : tensor<2x1024x40x16xf16>) outs(%1429 : tensor<2x1024x40x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x1xf16>
%1431 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1430 : tensor<2x1024x40x1xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1136 = tensor.collapse_shape %1431 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1432 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1433 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1432 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1434 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1433 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1435 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1434 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1436 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1136, %1435 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1437 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1425, %1436 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1438 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1437, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1439 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1440 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1439 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1441 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1440 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1442 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1443 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1442 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1444 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1443, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1445 = tensor.empty() : tensor<2x1024x640xi1>
%1446 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1438, %1444 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%1447 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1446, %1444, %1438 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1448 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1447, %1441 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%1449 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1448, %1441, %1447 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1450 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1449 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1451 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1450, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1452 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1451, %1436 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1453 = tensor.empty() : tensor<640x40x16xf16>
%1454 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_127 : tensor<640x40x1xf16>) outs(%1453 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1137 = tensor.collapse_shape %1454 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1455 = tensor.empty() : tensor<640x640xf16>
%1456 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_128, %collapsed_1137 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640xf16>
%1457 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%1456 : tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1138 = tensor.collapse_shape %1452 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1458 = tensor.empty() : tensor<2048x640xf16>
%1459 = linalg.fill ins(%cst_889 : f16) outs(%1458 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1460 = linalg.matmul ins(%collapsed_1138, %1457 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%1459 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1461 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_129, %1460 : tensor<640xf16>, tensor<2048x640xf16>) outs(%1458 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2048x640xf16>
%1462 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1135 : tensor<2x1024x40x16xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x16xf16>
%1463 = linalg.fill ins(%cst_891 : f16) outs(%1428 : tensor<2x1024x40x1xf16>) -> tensor<2x1024x40x1xf16>
%1464 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1462 : tensor<2x1024x40x16xf16>) outs(%1463 : tensor<2x1024x40x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x1xf16>
%1465 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1464 : tensor<2x1024x40x1xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1139 = tensor.collapse_shape %1465 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1466 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1467 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1466 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1468 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1467 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1469 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1468 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1470 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1139, %1469 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1471 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1425, %1470 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1472 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1471, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1473 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1474 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1473 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1475 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1474 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1476 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1477 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1476 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1478 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1477, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1479 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1472, %1478 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%1480 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1479, %1478, %1472 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1481 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1480, %1475 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%1482 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1481, %1475, %1480 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1483 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1482 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1484 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1483, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1485 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1484, %1470 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1486 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_130 : tensor<640x40x1xf16>) outs(%1453 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1140 = tensor.collapse_shape %1486 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1487 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_131, %collapsed_1140 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640xf16>
%1488 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%1487 : tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1141 = tensor.collapse_shape %1485 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1489 = linalg.fill ins(%cst_889 : f16) outs(%1458 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1490 = linalg.matmul ins(%collapsed_1141, %1488 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%1489 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1491 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_132, %1490 : tensor<640xf16>, tensor<2048x640xf16>) outs(%1458 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2048x640xf16>
%1492 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1135 : tensor<2x1024x40x16xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x16xf16>
%1493 = linalg.fill ins(%cst_891 : f16) outs(%1428 : tensor<2x1024x40x1xf16>) -> tensor<2x1024x40x1xf16>
%1494 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1492 : tensor<2x1024x40x16xf16>) outs(%1493 : tensor<2x1024x40x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x1xf16>
%1495 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1494 : tensor<2x1024x40x1xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1142 = tensor.collapse_shape %1495 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1496 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1497 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1496 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1498 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1497 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1499 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1498 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1500 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1142, %1499 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1501 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1425, %1500 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1502 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1501, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1503 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1504 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1503 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1505 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1504 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1506 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1507 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1506 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1508 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1507, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1509 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1502, %1508 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%1510 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1509, %1508, %1502 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1511 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1510, %1505 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%1512 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1511, %1505, %1510 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1513 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1512 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1514 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1513, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1515 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1514, %1500 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1516 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_133 : tensor<640x40x1xf16>) outs(%1453 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1143 = tensor.collapse_shape %1516 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1517 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_134, %collapsed_1143 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640xf16>
%1518 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%1517 : tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1144 = tensor.collapse_shape %1515 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1519 = linalg.fill ins(%cst_889 : f16) outs(%1458 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1520 = linalg.matmul ins(%collapsed_1144, %1518 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%1519 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1521 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_135, %1520 : tensor<640xf16>, tensor<2048x640xf16>) outs(%1458 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2048x640xf16>
%expanded_1145 = tensor.expand_shape %1461 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x8x80xf16>
%1522 = tensor.empty() : tensor<2x8x1024x80xf16>
%1523 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1145 : tensor<2x1024x8x80xf16>) outs(%1522 : tensor<2x8x1024x80xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x1024x80xf16>
%collapsed_1146 = tensor.collapse_shape %1523 [[0, 1], [2], [3]] : tensor<2x8x1024x80xf16> into tensor<16x1024x80xf16>
%expanded_1147 = tensor.expand_shape %1491 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x8x80xf16>
%1524 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1147 : tensor<2x1024x8x80xf16>) outs(%1522 : tensor<2x8x1024x80xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x1024x80xf16>
%collapsed_1148 = tensor.collapse_shape %1524 [[0, 1], [2], [3]] : tensor<2x8x1024x80xf16> into tensor<16x1024x80xf16>
%expanded_1149 = tensor.expand_shape %1521 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x8x80xf16>
%1525 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1149 : tensor<2x1024x8x80xf16>) outs(%1522 : tensor<2x8x1024x80xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x1024x80xf16>
%collapsed_1150 = tensor.collapse_shape %1525 [[0, 1], [2], [3]] : tensor<2x8x1024x80xf16> into tensor<16x1024x80xf16>
%1526 = tensor.empty() : tensor<16x80x1024xf16>
%1527 = linalg.generic {indexing_maps = [#map20, #map25], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1148 : tensor<16x1024x80xf16>) outs(%1526 : tensor<16x80x1024xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<16x80x1024xf16>
%1528 = tensor.empty() : tensor<16x1024x1024xf16>
%1529 = linalg.fill ins(%cst_889 : f16) outs(%1528 : tensor<16x1024x1024xf16>) -> tensor<16x1024x1024xf16>
%1530 = linalg.batch_matmul ins(%collapsed_1146, %1527 : tensor<16x1024x80xf16>, tensor<16x80x1024xf16>) outs(%1529 : tensor<16x1024x1024xf16>) -> tensor<16x1024x1024xf16>
%1531 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1530, %cst_0 : tensor<16x1024x1024xf16>, tensor<f16>) outs(%1528 : tensor<16x1024x1024xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x1024xf16>
%1532 = tensor.empty() : tensor<16x1024x1xf16>
%1533 = linalg.fill ins(%cst_891 : f16) outs(%1532 : tensor<16x1024x1xf16>) -> tensor<16x1024x1xf16>
%1534 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1531 : tensor<16x1024x1024xf16>) outs(%1533 : tensor<16x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x1xf16>
%1535 = linalg.generic {indexing_maps = [#map20, #map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1531, %1534 : tensor<16x1024x1024xf16>, tensor<16x1024x1xf16>) outs(%1528 : tensor<16x1024x1024xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x1024xf16>
%1536 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1535 : tensor<16x1024x1024xf16>) outs(%1528 : tensor<16x1024x1024xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.exp %in : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x1024xf16>
%1537 = linalg.fill ins(%cst_889 : f16) outs(%1532 : tensor<16x1024x1xf16>) -> tensor<16x1024x1xf16>
%1538 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1536 : tensor<16x1024x1024xf16>) outs(%1537 : tensor<16x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x1xf16>
%1539 = linalg.generic {indexing_maps = [#map20, #map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1536, %1538 : tensor<16x1024x1024xf16>, tensor<16x1024x1xf16>) outs(%1528 : tensor<16x1024x1024xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x1024xf16>
%1540 = tensor.empty() : tensor<16x1024x80xf16>
%1541 = linalg.fill ins(%cst_889 : f16) outs(%1540 : tensor<16x1024x80xf16>) -> tensor<16x1024x80xf16>
%1542 = linalg.batch_matmul ins(%1539, %collapsed_1150 : tensor<16x1024x1024xf16>, tensor<16x1024x80xf16>) outs(%1541 : tensor<16x1024x80xf16>) -> tensor<16x1024x80xf16>
%expanded_1151 = tensor.expand_shape %1542 [[0, 1], [2], [3]] : tensor<16x1024x80xf16> into tensor<2x8x1024x80xf16>
%1543 = tensor.empty() : tensor<2x1024x8x80xf16>
%1544 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1151 : tensor<2x8x1024x80xf16>) outs(%1543 : tensor<2x1024x8x80xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x8x80xf16>
%collapsed_1152 = tensor.collapse_shape %1544 [[0], [1], [2, 3]] : tensor<2x1024x8x80xf16> into tensor<2x1024x640xf16>
%expanded_1153 = tensor.expand_shape %collapsed_1152 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%1545 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1153 : tensor<2x1024x40x16xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x16xf16>
%1546 = linalg.fill ins(%cst_891 : f16) outs(%1428 : tensor<2x1024x40x1xf16>) -> tensor<2x1024x40x1xf16>
%1547 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1545 : tensor<2x1024x40x16xf16>) outs(%1546 : tensor<2x1024x40x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x1xf16>
%1548 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1547 : tensor<2x1024x40x1xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1154 = tensor.collapse_shape %1548 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1549 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1550 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1549 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1551 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1550 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1552 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1551 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1553 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1154, %1552 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1554 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1152, %1553 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1555 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1554, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1556 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1557 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1556 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1558 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1557 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1559 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1560 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1559 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1561 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1560, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1562 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1555, %1561 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%1563 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1562, %1561, %1555 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1564 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1563, %1558 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%1565 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1564, %1558, %1563 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1566 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1565 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1567 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1566, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1568 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1567, %1553 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1569 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_136 : tensor<640x40x1xf16>) outs(%1453 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1155 = tensor.collapse_shape %1569 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1570 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_137, %collapsed_1155 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640xf16>
%1571 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%1570 : tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1156 = tensor.collapse_shape %1568 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1572 = linalg.fill ins(%cst_889 : f16) outs(%1458 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1573 = linalg.matmul ins(%collapsed_1156, %1571 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%1572 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1574 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_138, %1573 : tensor<640xf16>, tensor<2048x640xf16>) outs(%1458 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2048x640xf16>
%expanded_1157 = tensor.expand_shape %1574 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16>
%1575 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1157, %collapsed_1134 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1576 = linalg.fill ins(%cst_889 : f16) outs(%1411 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16>
%1577 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1575 : tensor<2x1024x640xf16>) outs(%1576 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1578 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1577 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_904 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1579 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1578 : tensor<2x1024x1xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x640xf16>
%1580 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1575, %1579 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1581 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1580 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.mulf %in, %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1582 = linalg.fill ins(%cst_889 : f16) outs(%1411 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16>
%1583 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1581 : tensor<2x1024x640xf16>) outs(%1582 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1584 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1583 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_904 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1585 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1584 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.truncf %cst_895 : f64 to f16
%10166 = arith.addf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<2x1024x1xf16>
%1586 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1585 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.rsqrt %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1587 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1586 : tensor<2x1024x1xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x640xf16>
%1588 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1580, %1587 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%expanded_1158 = tensor.expand_shape %1588 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%1589 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1158 : tensor<2x1024x40x16xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x16xf16>
%1590 = linalg.fill ins(%cst_891 : f16) outs(%1428 : tensor<2x1024x40x1xf16>) -> tensor<2x1024x40x1xf16>
%1591 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1589 : tensor<2x1024x40x16xf16>) outs(%1590 : tensor<2x1024x40x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x1xf16>
%1592 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1591 : tensor<2x1024x40x1xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1159 = tensor.collapse_shape %1592 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1593 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1594 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1593 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1595 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1594 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1596 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1595 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1597 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1159, %1596 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1598 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1588, %1597 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1599 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1598, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1600 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1601 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1600 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1602 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1601 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1603 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1604 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1603 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1605 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1604, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1606 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1599, %1605 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%1607 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1606, %1605, %1599 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1608 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1607, %1602 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%1609 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1608, %1602, %1607 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1610 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1609 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1611 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1610, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1612 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1611, %1597 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1613 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_139 : tensor<640x40x1xf16>) outs(%1453 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1160 = tensor.collapse_shape %1613 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1614 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_140, %collapsed_1160 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640xf16>
%1615 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%1614 : tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1161 = tensor.collapse_shape %1612 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1616 = linalg.fill ins(%cst_889 : f16) outs(%1458 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1617 = linalg.matmul ins(%collapsed_1161, %1615 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%1616 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1618 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_141, %1617 : tensor<640xf16>, tensor<2048x640xf16>) outs(%1458 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2048x640xf16>
%1619 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_980 : tensor<2x77x48x16xf16>) outs(%440 : tensor<2x77x48x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x77x48x16xf16>
%1620 = linalg.fill ins(%cst_891 : f16) outs(%442 : tensor<2x77x48x1xf16>) -> tensor<2x77x48x1xf16>
%1621 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1619 : tensor<2x77x48x16xf16>) outs(%1620 : tensor<2x77x48x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x77x48x1xf16>
%1622 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1621 : tensor<2x77x48x1xf16>) outs(%440 : tensor<2x77x48x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x77x48x16xf16>
%collapsed_1162 = tensor.collapse_shape %1622 [[0], [1], [2, 3]] : tensor<2x77x48x16xf16> into tensor<2x77x768xf16>
%1623 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1624 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1623 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1625 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1624 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1626 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1625 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1627 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1162, %1626 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1628 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2, %1627 : tensor<2x77x768xf16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1629 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1628, %cst_880 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1630 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1631 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1630 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1632 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1631 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1633 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1634 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1633 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1635 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1634, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1636 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1629, %1635 : tensor<2x77x768xf16>, tensor<f16>) outs(%460 : tensor<2x77x768xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x77x768xi1>
%1637 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1636, %1635, %1629 : tensor<2x77x768xi1>, tensor<f16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1638 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1637, %1632 : tensor<2x77x768xf16>, tensor<f16>) outs(%460 : tensor<2x77x768xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x77x768xi1>
%1639 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1638, %1632, %1637 : tensor<2x77x768xi1>, tensor<f16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1640 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1639 : tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1641 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1640, %cst_880 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1642 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1641, %1627 : tensor<2x77x768xf16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1643 = tensor.empty() : tensor<640x48x16xf16>
%1644 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_142 : tensor<640x48x1xf16>) outs(%1643 : tensor<640x48x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x48x16xf16>
%collapsed_1163 = tensor.collapse_shape %1644 [[0], [1, 2]] : tensor<640x48x16xf16> into tensor<640x768xf16>
%1645 = tensor.empty() : tensor<640x768xf16>
%1646 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_143, %collapsed_1163 : tensor<640x768xi8>, tensor<640x768xf16>) outs(%1645 : tensor<640x768xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x768xf16>
%1647 = tensor.empty() : tensor<768x640xf16>
%1648 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%1646 : tensor<640x768xf16>) outs(%1647 : tensor<768x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<768x640xf16>
%collapsed_1164 = tensor.collapse_shape %1642 [[0, 1], [2]] : tensor<2x77x768xf16> into tensor<154x768xf16>
%1649 = tensor.empty() : tensor<154x640xf16>
%1650 = linalg.fill ins(%cst_889 : f16) outs(%1649 : tensor<154x640xf16>) -> tensor<154x640xf16>
%1651 = linalg.matmul ins(%collapsed_1164, %1648 : tensor<154x768xf16>, tensor<768x640xf16>) outs(%1650 : tensor<154x640xf16>) -> tensor<154x640xf16>
%1652 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_980 : tensor<2x77x48x16xf16>) outs(%440 : tensor<2x77x48x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x77x48x16xf16>
%1653 = linalg.fill ins(%cst_891 : f16) outs(%442 : tensor<2x77x48x1xf16>) -> tensor<2x77x48x1xf16>
%1654 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1652 : tensor<2x77x48x16xf16>) outs(%1653 : tensor<2x77x48x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x77x48x1xf16>
%1655 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1654 : tensor<2x77x48x1xf16>) outs(%440 : tensor<2x77x48x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x77x48x16xf16>
%collapsed_1165 = tensor.collapse_shape %1655 [[0], [1], [2, 3]] : tensor<2x77x48x16xf16> into tensor<2x77x768xf16>
%1656 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1657 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1656 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1658 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1657 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1659 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1658 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1660 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1165, %1659 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1661 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2, %1660 : tensor<2x77x768xf16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1662 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1661, %cst_880 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1663 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1664 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1663 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1665 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1664 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1666 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1667 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1666 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1668 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1667, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1669 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1662, %1668 : tensor<2x77x768xf16>, tensor<f16>) outs(%460 : tensor<2x77x768xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x77x768xi1>
%1670 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1669, %1668, %1662 : tensor<2x77x768xi1>, tensor<f16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1671 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1670, %1665 : tensor<2x77x768xf16>, tensor<f16>) outs(%460 : tensor<2x77x768xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x77x768xi1>
%1672 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1671, %1665, %1670 : tensor<2x77x768xi1>, tensor<f16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1673 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1672 : tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1674 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1673, %cst_880 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1675 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1674, %1660 : tensor<2x77x768xf16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%1676 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_144 : tensor<640x48x1xf16>) outs(%1643 : tensor<640x48x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x48x16xf16>
%collapsed_1166 = tensor.collapse_shape %1676 [[0], [1, 2]] : tensor<640x48x16xf16> into tensor<640x768xf16>
%1677 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_145, %collapsed_1166 : tensor<640x768xi8>, tensor<640x768xf16>) outs(%1645 : tensor<640x768xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x768xf16>
%1678 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%1677 : tensor<640x768xf16>) outs(%1647 : tensor<768x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<768x640xf16>
%collapsed_1167 = tensor.collapse_shape %1675 [[0, 1], [2]] : tensor<2x77x768xf16> into tensor<154x768xf16>
%1679 = linalg.fill ins(%cst_889 : f16) outs(%1649 : tensor<154x640xf16>) -> tensor<154x640xf16>
%1680 = linalg.matmul ins(%collapsed_1167, %1678 : tensor<154x768xf16>, tensor<768x640xf16>) outs(%1679 : tensor<154x640xf16>) -> tensor<154x640xf16>
%expanded_1168 = tensor.expand_shape %1618 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x8x80xf16>
%1681 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1168 : tensor<2x1024x8x80xf16>) outs(%1522 : tensor<2x8x1024x80xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x1024x80xf16>
%collapsed_1169 = tensor.collapse_shape %1681 [[0, 1], [2], [3]] : tensor<2x8x1024x80xf16> into tensor<16x1024x80xf16>
%expanded_1170 = tensor.expand_shape %1651 [[0, 1], [2, 3]] : tensor<154x640xf16> into tensor<2x77x8x80xf16>
%1682 = tensor.empty() : tensor<2x8x77x80xf16>
%1683 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1170 : tensor<2x77x8x80xf16>) outs(%1682 : tensor<2x8x77x80xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x77x80xf16>
%collapsed_1171 = tensor.collapse_shape %1683 [[0, 1], [2], [3]] : tensor<2x8x77x80xf16> into tensor<16x77x80xf16>
%expanded_1172 = tensor.expand_shape %1680 [[0, 1], [2, 3]] : tensor<154x640xf16> into tensor<2x77x8x80xf16>
%1684 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1172 : tensor<2x77x8x80xf16>) outs(%1682 : tensor<2x8x77x80xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x77x80xf16>
%collapsed_1173 = tensor.collapse_shape %1684 [[0, 1], [2], [3]] : tensor<2x8x77x80xf16> into tensor<16x77x80xf16>
%1685 = tensor.empty() : tensor<16x80x77xf16>
%1686 = linalg.generic {indexing_maps = [#map20, #map25], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1171 : tensor<16x77x80xf16>) outs(%1685 : tensor<16x80x77xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<16x80x77xf16>
%1687 = tensor.empty() : tensor<16x1024x77xf16>
%1688 = linalg.fill ins(%cst_889 : f16) outs(%1687 : tensor<16x1024x77xf16>) -> tensor<16x1024x77xf16>
%1689 = linalg.batch_matmul ins(%collapsed_1169, %1686 : tensor<16x1024x80xf16>, tensor<16x80x77xf16>) outs(%1688 : tensor<16x1024x77xf16>) -> tensor<16x1024x77xf16>
%1690 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1689, %cst_0 : tensor<16x1024x77xf16>, tensor<f16>) outs(%1687 : tensor<16x1024x77xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x77xf16>
%1691 = linalg.fill ins(%cst_891 : f16) outs(%1532 : tensor<16x1024x1xf16>) -> tensor<16x1024x1xf16>
%1692 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1690 : tensor<16x1024x77xf16>) outs(%1691 : tensor<16x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x1xf16>
%1693 = linalg.generic {indexing_maps = [#map20, #map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1690, %1692 : tensor<16x1024x77xf16>, tensor<16x1024x1xf16>) outs(%1687 : tensor<16x1024x77xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x77xf16>
%1694 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1693 : tensor<16x1024x77xf16>) outs(%1687 : tensor<16x1024x77xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.exp %in : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x77xf16>
%1695 = linalg.fill ins(%cst_889 : f16) outs(%1532 : tensor<16x1024x1xf16>) -> tensor<16x1024x1xf16>
%1696 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1694 : tensor<16x1024x77xf16>) outs(%1695 : tensor<16x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x1xf16>
%1697 = linalg.generic {indexing_maps = [#map20, #map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1694, %1696 : tensor<16x1024x77xf16>, tensor<16x1024x1xf16>) outs(%1687 : tensor<16x1024x77xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x77xf16>
%1698 = linalg.fill ins(%cst_889 : f16) outs(%1540 : tensor<16x1024x80xf16>) -> tensor<16x1024x80xf16>
%1699 = linalg.batch_matmul ins(%1697, %collapsed_1173 : tensor<16x1024x77xf16>, tensor<16x77x80xf16>) outs(%1698 : tensor<16x1024x80xf16>) -> tensor<16x1024x80xf16>
%expanded_1174 = tensor.expand_shape %1699 [[0, 1], [2], [3]] : tensor<16x1024x80xf16> into tensor<2x8x1024x80xf16>
%1700 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1174 : tensor<2x8x1024x80xf16>) outs(%1543 : tensor<2x1024x8x80xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x8x80xf16>
%collapsed_1175 = tensor.collapse_shape %1700 [[0], [1], [2, 3]] : tensor<2x1024x8x80xf16> into tensor<2x1024x640xf16>
%expanded_1176 = tensor.expand_shape %collapsed_1175 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%1701 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1176 : tensor<2x1024x40x16xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x16xf16>
%1702 = linalg.fill ins(%cst_891 : f16) outs(%1428 : tensor<2x1024x40x1xf16>) -> tensor<2x1024x40x1xf16>
%1703 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1701 : tensor<2x1024x40x16xf16>) outs(%1702 : tensor<2x1024x40x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x1xf16>
%1704 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1703 : tensor<2x1024x40x1xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1177 = tensor.collapse_shape %1704 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1705 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1706 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1705 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1707 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1706 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1708 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1707 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1709 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1177, %1708 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1710 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1175, %1709 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1711 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1710, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1712 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1713 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1712 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1714 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1713 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1715 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1716 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1715 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1717 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1716, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1718 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1711, %1717 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%1719 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1718, %1717, %1711 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1720 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1719, %1714 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%1721 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1720, %1714, %1719 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1722 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1721 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1723 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1722, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1724 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1723, %1709 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1725 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_146 : tensor<640x40x1xf16>) outs(%1453 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1178 = tensor.collapse_shape %1725 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1726 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_147, %collapsed_1178 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640xf16>
%1727 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%1726 : tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1179 = tensor.collapse_shape %1724 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1728 = linalg.fill ins(%cst_889 : f16) outs(%1458 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1729 = linalg.matmul ins(%collapsed_1179, %1727 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%1728 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1730 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_148, %1729 : tensor<640xf16>, tensor<2048x640xf16>) outs(%1458 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2048x640xf16>
%expanded_1180 = tensor.expand_shape %1730 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16>
%1731 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1180, %1575 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1732 = linalg.fill ins(%cst_889 : f16) outs(%1411 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16>
%1733 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1731 : tensor<2x1024x640xf16>) outs(%1732 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1734 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1733 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_904 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1735 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1734 : tensor<2x1024x1xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x640xf16>
%1736 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1731, %1735 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1737 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1736 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.mulf %in, %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1738 = linalg.fill ins(%cst_889 : f16) outs(%1411 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16>
%1739 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1737 : tensor<2x1024x640xf16>) outs(%1738 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1740 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1739 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_904 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1741 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1740 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.truncf %cst_895 : f64 to f16
%10166 = arith.addf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<2x1024x1xf16>
%1742 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1741 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.rsqrt %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1743 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1742 : tensor<2x1024x1xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x640xf16>
%1744 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1736, %1743 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1745 = tensor.empty() : tensor<5120x40x16xf16>
%1746 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_149 : tensor<5120x40x1xf16>) outs(%1745 : tensor<5120x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<5120x40x16xf16>
%collapsed_1181 = tensor.collapse_shape %1746 [[0], [1, 2]] : tensor<5120x40x16xf16> into tensor<5120x640xf16>
%1747 = tensor.empty() : tensor<5120x640xf16>
%1748 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_150, %collapsed_1181 : tensor<5120x640xi8>, tensor<5120x640xf16>) outs(%1747 : tensor<5120x640xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<5120x640xf16>
%1749 = tensor.empty() : tensor<640x5120xf16>
%1750 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%1748 : tensor<5120x640xf16>) outs(%1749 : tensor<640x5120xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x5120xf16>
%collapsed_1182 = tensor.collapse_shape %1744 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1751 = tensor.empty() : tensor<2048x5120xf16>
%1752 = linalg.fill ins(%cst_889 : f16) outs(%1751 : tensor<2048x5120xf16>) -> tensor<2048x5120xf16>
%1753 = linalg.matmul ins(%collapsed_1182, %1750 : tensor<2048x640xf16>, tensor<640x5120xf16>) outs(%1752 : tensor<2048x5120xf16>) -> tensor<2048x5120xf16>
%1754 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_151, %1753 : tensor<5120xf16>, tensor<2048x5120xf16>) outs(%1751 : tensor<2048x5120xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2048x5120xf16>
%expanded_1183 = tensor.expand_shape %1754 [[0, 1], [2]] : tensor<2048x5120xf16> into tensor<2x1024x5120xf16>
%extracted_slice_1184 = tensor.extract_slice %expanded_1183[0, 0, 0] [2, 1024, 2560] [1, 1, 1] : tensor<2x1024x5120xf16> to tensor<2x1024x2560xf16>
%extracted_slice_1185 = tensor.extract_slice %expanded_1183[0, 0, 2560] [2, 1024, 2560] [1, 1, 1] : tensor<2x1024x5120xf16> to tensor<2x1024x2560xf16>
%1755 = tensor.empty() : tensor<2x1024x2560xf16>
%1756 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1185 : tensor<2x1024x2560xf16>) outs(%1755 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.sqrt %cst_892 : f16
%10166 = arith.divf %in, %10165 : f16
%10167 = math.erf %10166 : f16
%10168 = arith.addf %10167, %cst_890 : f16
%10169 = arith.mulf %10168, %cst_893 : f16
%10170 = arith.mulf %in, %10169 : f16
linalg.yield %10170 : f16
} -> tensor<2x1024x2560xf16>
%1757 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1184, %1756 : tensor<2x1024x2560xf16>, tensor<2x1024x2560xf16>) outs(%1755 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x2560xf16>
%1758 = tensor.empty() : tensor<640x160x16xf16>
%1759 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_152 : tensor<640x160x1xf16>) outs(%1758 : tensor<640x160x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x160x16xf16>
%collapsed_1186 = tensor.collapse_shape %1759 [[0], [1, 2]] : tensor<640x160x16xf16> into tensor<640x2560xf16>
%1760 = tensor.empty() : tensor<640x2560xf16>
%1761 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_153, %collapsed_1186 : tensor<640x2560xi8>, tensor<640x2560xf16>) outs(%1760 : tensor<640x2560xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x2560xf16>
%1762 = tensor.empty() : tensor<2560x640xf16>
%1763 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%1761 : tensor<640x2560xf16>) outs(%1762 : tensor<2560x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2560x640xf16>
%collapsed_1187 = tensor.collapse_shape %1757 [[0, 1], [2]] : tensor<2x1024x2560xf16> into tensor<2048x2560xf16>
%1764 = linalg.fill ins(%cst_889 : f16) outs(%1458 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1765 = linalg.matmul ins(%collapsed_1187, %1763 : tensor<2048x2560xf16>, tensor<2560x640xf16>) outs(%1764 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1766 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_154, %1765 : tensor<640xf16>, tensor<2048x640xf16>) outs(%1458 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2048x640xf16>
%expanded_1188 = tensor.expand_shape %1766 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16>
%1767 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1188, %1731 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%expanded_1189 = tensor.expand_shape %1767 [[0], [1, 2], [3]] : tensor<2x1024x640xf16> into tensor<2x32x32x640xf16>
%1768 = linalg.generic {indexing_maps = [#map11, #map26], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1189 : tensor<2x32x32x640xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%expanded_1190 = tensor.expand_shape %1768 [[0], [1, 2], [3], [4]] : tensor<2x640x32x32xf16> into tensor<2x40x16x32x32xf16>
%1769 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1190 : tensor<2x40x16x32x32xf16>) outs(%1288 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x40x16x32x32xf16>
%1770 = linalg.fill ins(%cst_891 : f16) outs(%1290 : tensor<2x40x1x32x32xf16>) -> tensor<2x40x1x32x32xf16>
%1771 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1769 : tensor<2x40x16x32x32xf16>) outs(%1770 : tensor<2x40x1x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x40x1x32x32xf16>
%1772 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1771 : tensor<2x40x1x32x32xf16>) outs(%1288 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x40x16x32x32xf16>
%collapsed_1191 = tensor.collapse_shape %1772 [[0], [1, 2], [3], [4]] : tensor<2x40x16x32x32xf16> into tensor<2x640x32x32xf16>
%1773 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1774 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1773 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1775 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1774 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1776 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1775 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1777 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1191, %1776 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1778 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1768, %1777 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1779 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1778, %cst_880 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1780 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1781 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1780 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1782 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1781 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1783 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1784 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1783 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1785 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1784, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1786 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1779, %1785 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1307 : tensor<2x640x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x32x32xi1>
%1787 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1786, %1785, %1779 : tensor<2x640x32x32xi1>, tensor<f16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1788 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1787, %1782 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1307 : tensor<2x640x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x32x32xi1>
%1789 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1788, %1782, %1787 : tensor<2x640x32x32xi1>, tensor<f16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1790 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1789 : tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1791 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1790, %cst_880 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1792 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1791, %1777 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1793 = linalg.generic {indexing_maps = [#map17, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_155 : tensor<640x40x1x1x1xf16>) outs(%1402 : tensor<640x40x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16x1x1xf16>
%1794 = linalg.generic {indexing_maps = [#map18, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1793 : tensor<640x40x16x1x1xf16>) outs(%1402 : tensor<640x40x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16x1x1xf16>
%collapsed_1192 = tensor.collapse_shape %1794 [[0], [1, 2], [3], [4]] : tensor<640x40x16x1x1xf16> into tensor<640x640x1x1xf16>
%1795 = linalg.generic {indexing_maps = [#map13, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_156, %collapsed_1192 : tensor<640x640x1x1xi8>, tensor<640x640x1x1xf16>) outs(%1405 : tensor<640x640x1x1xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640x1x1xf16>
%1796 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_157 : tensor<640xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%1797 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1792, %1795 : tensor<2x640x32x32xf16>, tensor<640x640x1x1xf16>) outs(%1796 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16>
%1798 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1797, %1353 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%collapsed_1193 = tensor.collapse_shape %1798 [[0], [1], [2, 3]] : tensor<2x640x32x32xf16> into tensor<2x640x1024xf16>
%expanded_1194 = tensor.expand_shape %collapsed_1193 [[0], [1, 2], [3]] : tensor<2x640x1024xf16> into tensor<2x32x20x1024xf16>
%1799 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%1800 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1799 : tensor<f64>) outs(%9 : tensor<f32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<f32>
%1801 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1800 : tensor<f32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<2x32x20x1024xf32>
%1802 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1194 : tensor<2x32x20x1024xf16>) outs(%1801 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %out: f32):
%10165 = arith.extf %in : f16 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%1803 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%1804 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1802 : tensor<2x32x20x1024xf32>) outs(%1803 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1805 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1804 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_902 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1806 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1802, %1805 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.subf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%1807 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1806 : tensor<2x32x20x1024xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.mulf %in, %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%1808 = linalg.fill ins(%cst_894 : f64) outs(%55 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
%1809 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1807 : tensor<2x32x20x1024xf32>) outs(%1808 : tensor<2x32x1x1xf64>) {
^bb0(%in: f32, %out: f64):
%10165 = arith.extf %in : f32 to f64
%10166 = arith.addf %10165, %out : f64
linalg.yield %10166 : f64
} -> tensor<2x32x1x1xf64>
%1810 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1809 : tensor<2x32x1x1xf64>) outs(%55 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%10165 = arith.divf %in, %cst_903 : f64
linalg.yield %10165 : f64
} -> tensor<2x32x1x1xf64>
%1811 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1810 : tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1812 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%1813 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1802 : tensor<2x32x20x1024xf32>) outs(%1812 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1814 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1813 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_902 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1815 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1811, %cst_4 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x1x1xf32>
%1816 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1815 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = math.rsqrt %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1817 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1194, %1814 : tensor<2x32x20x1024xf16>, tensor<2x32x1x1xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %in_2540: f32, %out: f32):
%10165 = arith.extf %in : f16 to f32
%10166 = arith.subf %10165, %in_2540 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x20x1024xf32>
%1818 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1817, %1816 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.mulf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%collapsed_1195 = tensor.collapse_shape %1818 [[0], [1, 2], [3]] : tensor<2x32x20x1024xf32> into tensor<2x640x1024xf32>
%expanded_1196 = tensor.expand_shape %collapsed_1195 [[0], [1], [2, 3]] : tensor<2x640x1024xf32> into tensor<2x640x32x32xf32>
%expanded_1197 = tensor.expand_shape %cst_158 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16>
%1819 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1196, %expanded_1197 : tensor<2x640x32x32xf32>, tensor<640x1x1xf16>) outs(%1279 : tensor<2x640x32x32xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.mulf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x640x32x32xf32>
%expanded_1198 = tensor.expand_shape %cst_159 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16>
%1820 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1819, %expanded_1198 : tensor<2x640x32x32xf32>, tensor<640x1x1xf16>) outs(%1279 : tensor<2x640x32x32xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x640x32x32xf32>
%1821 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%1822 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1821 : tensor<f64>) outs(%23 : tensor<f16>) {
^bb0(%in: f64, %out: f16):
%10165 = arith.truncf %in : f64 to f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1823 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1822 : tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%1824 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1820 : tensor<2x640x32x32xf32>) outs(%1823 : tensor<2x640x32x32xf16>) {
^bb0(%in: f32, %out: f16):
%10165 = arith.truncf %in : f32 to f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1825 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1824 : tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
%10166 = math.exp %10165 : f16
%10167 = arith.addf %10166, %cst_890 : f16
%10168 = arith.divf %cst_890, %10167 : f16
linalg.yield %10168 : f16
} -> tensor<2x640x32x32xf16>
%1826 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1825, %1824 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%expanded_1199 = tensor.expand_shape %1826 [[0], [1, 2], [3], [4]] : tensor<2x640x32x32xf16> into tensor<2x40x16x32x32xf16>
%1827 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1199 : tensor<2x40x16x32x32xf16>) outs(%1288 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x40x16x32x32xf16>
%1828 = linalg.fill ins(%cst_891 : f16) outs(%1290 : tensor<2x40x1x32x32xf16>) -> tensor<2x40x1x32x32xf16>
%1829 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1827 : tensor<2x40x16x32x32xf16>) outs(%1828 : tensor<2x40x1x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x40x1x32x32xf16>
%1830 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1829 : tensor<2x40x1x32x32xf16>) outs(%1288 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x40x16x32x32xf16>
%collapsed_1200 = tensor.collapse_shape %1830 [[0], [1, 2], [3], [4]] : tensor<2x40x16x32x32xf16> into tensor<2x640x32x32xf16>
%1831 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1832 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1831 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1833 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1832 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1834 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1833 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1835 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1200, %1834 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1836 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1826, %1835 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1837 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1836, %cst_880 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1838 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1839 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1838 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1840 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1839 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1841 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1842 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1841 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1843 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1842, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1844 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1837, %1843 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1307 : tensor<2x640x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x32x32xi1>
%1845 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1844, %1843, %1837 : tensor<2x640x32x32xi1>, tensor<f16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1846 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1845, %1840 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1307 : tensor<2x640x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x32x32xi1>
%1847 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1846, %1840, %1845 : tensor<2x640x32x32xi1>, tensor<f16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1848 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1847 : tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1849 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1848, %cst_880 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1850 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1849, %1835 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1851 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_160 : tensor<640x40x1x3x3xf16>) outs(%1315 : tensor<640x40x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16x3x3xf16>
%collapsed_1201 = tensor.collapse_shape %1851 [[0], [1, 2], [3], [4]] : tensor<640x40x16x3x3xf16> into tensor<640x640x3x3xf16>
%1852 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_161, %collapsed_1201 : tensor<640x640x3x3xi8>, tensor<640x640x3x3xf16>) outs(%1317 : tensor<640x640x3x3xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640x3x3xf16>
%padded_1202 = tensor.pad %1850 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
tensor.yield %cst_889 : f16
} : tensor<2x640x32x32xf16> to tensor<2x640x34x34xf16>
%1853 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_162 : tensor<640xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%1854 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1202, %1852 : tensor<2x640x34x34xf16>, tensor<640x640x3x3xf16>) outs(%1853 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16>
%1855 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
%10166 = math.exp %10165 : f16
%10167 = arith.addf %10166, %cst_890 : f16
%10168 = arith.divf %cst_890, %10167 : f16
linalg.yield %10168 : f16
} -> tensor<2x1280xf16>
%1856 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%1855, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280xf16>
%1857 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_163 : tensor<640x1280xf16>) outs(%1251 : tensor<1280x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x640xf16>
%1858 = linalg.fill ins(%cst_889 : f16) outs(%1253 : tensor<2x640xf16>) -> tensor<2x640xf16>
%1859 = linalg.matmul ins(%1856, %1857 : tensor<2x1280xf16>, tensor<1280x640xf16>) outs(%1858 : tensor<2x640xf16>) -> tensor<2x640xf16>
%1860 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_164, %1859 : tensor<640xf16>, tensor<2x640xf16>) outs(%1253 : tensor<2x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640xf16>
%expanded_1203 = tensor.expand_shape %1860 [[0], [1, 2, 3]] : tensor<2x640xf16> into tensor<2x640x1x1xf16>
%1861 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1854, %expanded_1203 : tensor<2x640x32x32xf16>, tensor<2x640x1x1xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%collapsed_1204 = tensor.collapse_shape %1861 [[0], [1], [2, 3]] : tensor<2x640x32x32xf16> into tensor<2x640x1024xf16>
%expanded_1205 = tensor.expand_shape %collapsed_1204 [[0], [1, 2], [3]] : tensor<2x640x1024xf16> into tensor<2x32x20x1024xf16>
%1862 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%1863 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1862 : tensor<f64>) outs(%9 : tensor<f32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<f32>
%1864 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1863 : tensor<f32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<2x32x20x1024xf32>
%1865 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1205 : tensor<2x32x20x1024xf16>) outs(%1864 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %out: f32):
%10165 = arith.extf %in : f16 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%1866 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%1867 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1865 : tensor<2x32x20x1024xf32>) outs(%1866 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1868 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1867 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_902 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1869 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1865, %1868 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.subf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%1870 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1869 : tensor<2x32x20x1024xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.mulf %in, %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%1871 = linalg.fill ins(%cst_894 : f64) outs(%55 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
%1872 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1870 : tensor<2x32x20x1024xf32>) outs(%1871 : tensor<2x32x1x1xf64>) {
^bb0(%in: f32, %out: f64):
%10165 = arith.extf %in : f32 to f64
%10166 = arith.addf %10165, %out : f64
linalg.yield %10166 : f64
} -> tensor<2x32x1x1xf64>
%1873 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1872 : tensor<2x32x1x1xf64>) outs(%55 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%10165 = arith.divf %in, %cst_903 : f64
linalg.yield %10165 : f64
} -> tensor<2x32x1x1xf64>
%1874 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1873 : tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1875 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%1876 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1865 : tensor<2x32x20x1024xf32>) outs(%1875 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1877 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1876 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_902 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1878 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1874, %cst_4 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x1x1xf32>
%1879 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1878 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = math.rsqrt %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1880 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1205, %1877 : tensor<2x32x20x1024xf16>, tensor<2x32x1x1xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %in_2540: f32, %out: f32):
%10165 = arith.extf %in : f16 to f32
%10166 = arith.subf %10165, %in_2540 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x20x1024xf32>
%1881 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1880, %1879 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.mulf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%collapsed_1206 = tensor.collapse_shape %1881 [[0], [1, 2], [3]] : tensor<2x32x20x1024xf32> into tensor<2x640x1024xf32>
%expanded_1207 = tensor.expand_shape %collapsed_1206 [[0], [1], [2, 3]] : tensor<2x640x1024xf32> into tensor<2x640x32x32xf32>
%expanded_1208 = tensor.expand_shape %cst_165 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16>
%1882 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1207, %expanded_1208 : tensor<2x640x32x32xf32>, tensor<640x1x1xf16>) outs(%1279 : tensor<2x640x32x32xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.mulf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x640x32x32xf32>
%expanded_1209 = tensor.expand_shape %cst_166 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16>
%1883 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1882, %expanded_1209 : tensor<2x640x32x32xf32>, tensor<640x1x1xf16>) outs(%1279 : tensor<2x640x32x32xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x640x32x32xf32>
%1884 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%1885 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1884 : tensor<f64>) outs(%23 : tensor<f16>) {
^bb0(%in: f64, %out: f16):
%10165 = arith.truncf %in : f64 to f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1886 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1885 : tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%1887 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1883 : tensor<2x640x32x32xf32>) outs(%1886 : tensor<2x640x32x32xf16>) {
^bb0(%in: f32, %out: f16):
%10165 = arith.truncf %in : f32 to f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1888 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1887 : tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
%10166 = math.exp %10165 : f16
%10167 = arith.addf %10166, %cst_890 : f16
%10168 = arith.divf %cst_890, %10167 : f16
linalg.yield %10168 : f16
} -> tensor<2x640x32x32xf16>
%1889 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1888, %1887 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%expanded_1210 = tensor.expand_shape %1889 [[0], [1, 2], [3], [4]] : tensor<2x640x32x32xf16> into tensor<2x40x16x32x32xf16>
%1890 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1210 : tensor<2x40x16x32x32xf16>) outs(%1288 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x40x16x32x32xf16>
%1891 = linalg.fill ins(%cst_891 : f16) outs(%1290 : tensor<2x40x1x32x32xf16>) -> tensor<2x40x1x32x32xf16>
%1892 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1890 : tensor<2x40x16x32x32xf16>) outs(%1891 : tensor<2x40x1x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x40x1x32x32xf16>
%1893 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1892 : tensor<2x40x1x32x32xf16>) outs(%1288 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x40x16x32x32xf16>
%collapsed_1211 = tensor.collapse_shape %1893 [[0], [1, 2], [3], [4]] : tensor<2x40x16x32x32xf16> into tensor<2x640x32x32xf16>
%1894 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1895 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1894 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1896 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1895 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1897 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1896 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1898 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1211, %1897 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1899 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1889, %1898 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1900 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1899, %cst_880 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1901 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1902 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1901 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1903 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1902 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1904 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1905 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1904 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1906 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1905, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1907 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1900, %1906 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1307 : tensor<2x640x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x32x32xi1>
%1908 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1907, %1906, %1900 : tensor<2x640x32x32xi1>, tensor<f16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1909 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1908, %1903 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1307 : tensor<2x640x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x32x32xi1>
%1910 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1909, %1903, %1908 : tensor<2x640x32x32xi1>, tensor<f16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1911 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1910 : tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1912 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1911, %cst_880 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1913 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1912, %1898 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1914 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_167 : tensor<640x40x1x3x3xf16>) outs(%1315 : tensor<640x40x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16x3x3xf16>
%collapsed_1212 = tensor.collapse_shape %1914 [[0], [1, 2], [3], [4]] : tensor<640x40x16x3x3xf16> into tensor<640x640x3x3xf16>
%1915 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_168, %collapsed_1212 : tensor<640x640x3x3xi8>, tensor<640x640x3x3xf16>) outs(%1317 : tensor<640x640x3x3xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640x3x3xf16>
%padded_1213 = tensor.pad %1913 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
tensor.yield %cst_889 : f16
} : tensor<2x640x32x32xf16> to tensor<2x640x34x34xf16>
%1916 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_169 : tensor<640xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%1917 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1213, %1915 : tensor<2x640x34x34xf16>, tensor<640x640x3x3xf16>) outs(%1916 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16>
%1918 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1798, %1917 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1919 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1918, %cst_3 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%collapsed_1214 = tensor.collapse_shape %1919 [[0], [1], [2, 3]] : tensor<2x640x32x32xf16> into tensor<2x640x1024xf16>
%expanded_1215 = tensor.expand_shape %collapsed_1214 [[0], [1, 2], [3]] : tensor<2x640x1024xf16> into tensor<2x32x20x1024xf16>
%1920 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%1921 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1920 : tensor<f64>) outs(%9 : tensor<f32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<f32>
%1922 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1921 : tensor<f32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<2x32x20x1024xf32>
%1923 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1215 : tensor<2x32x20x1024xf16>) outs(%1922 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %out: f32):
%10165 = arith.extf %in : f16 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%1924 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%1925 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1923 : tensor<2x32x20x1024xf32>) outs(%1924 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1926 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1925 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_902 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1927 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1923, %1926 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.subf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%1928 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1927 : tensor<2x32x20x1024xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.mulf %in, %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%1929 = linalg.fill ins(%cst_894 : f64) outs(%55 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
%1930 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1928 : tensor<2x32x20x1024xf32>) outs(%1929 : tensor<2x32x1x1xf64>) {
^bb0(%in: f32, %out: f64):
%10165 = arith.extf %in : f32 to f64
%10166 = arith.addf %10165, %out : f64
linalg.yield %10166 : f64
} -> tensor<2x32x1x1xf64>
%1931 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1930 : tensor<2x32x1x1xf64>) outs(%55 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%10165 = arith.divf %in, %cst_903 : f64
linalg.yield %10165 : f64
} -> tensor<2x32x1x1xf64>
%1932 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1931 : tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1933 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%1934 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1923 : tensor<2x32x20x1024xf32>) outs(%1933 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1935 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1934 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_902 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1936 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1932, %cst_2 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x1x1xf32>
%1937 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1936 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = math.rsqrt %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%1938 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1215, %1935 : tensor<2x32x20x1024xf16>, tensor<2x32x1x1xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %in_2540: f32, %out: f32):
%10165 = arith.extf %in : f16 to f32
%10166 = arith.subf %10165, %in_2540 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x20x1024xf32>
%1939 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1938, %1937 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%1260 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.mulf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x1024xf32>
%collapsed_1216 = tensor.collapse_shape %1939 [[0], [1, 2], [3]] : tensor<2x32x20x1024xf32> into tensor<2x640x1024xf32>
%expanded_1217 = tensor.expand_shape %collapsed_1216 [[0], [1], [2, 3]] : tensor<2x640x1024xf32> into tensor<2x640x32x32xf32>
%1940 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%1941 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1940 : tensor<f64>) outs(%23 : tensor<f16>) {
^bb0(%in: f64, %out: f16):
%10165 = arith.truncf %in : f64 to f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1942 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1941 : tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%1943 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1217 : tensor<2x640x32x32xf32>) outs(%1942 : tensor<2x640x32x32xf16>) {
^bb0(%in: f32, %out: f16):
%10165 = arith.truncf %in : f32 to f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%expanded_1218 = tensor.expand_shape %1943 [[0], [1, 2], [3], [4]] : tensor<2x640x32x32xf16> into tensor<2x40x16x32x32xf16>
%1944 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1218 : tensor<2x40x16x32x32xf16>) outs(%1288 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x40x16x32x32xf16>
%1945 = linalg.fill ins(%cst_891 : f16) outs(%1290 : tensor<2x40x1x32x32xf16>) -> tensor<2x40x1x32x32xf16>
%1946 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1944 : tensor<2x40x16x32x32xf16>) outs(%1945 : tensor<2x40x1x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x40x1x32x32xf16>
%1947 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1946 : tensor<2x40x1x32x32xf16>) outs(%1288 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x40x16x32x32xf16>
%collapsed_1219 = tensor.collapse_shape %1947 [[0], [1, 2], [3], [4]] : tensor<2x40x16x32x32xf16> into tensor<2x640x32x32xf16>
%1948 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1949 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1948 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1950 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1949 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1951 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1950 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1952 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1219, %1951 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1953 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1943, %1952 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1954 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1953, %cst_880 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1955 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1956 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1955 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1957 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1956 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1958 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1959 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1958 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1960 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%1959, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1961 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1954, %1960 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1307 : tensor<2x640x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x32x32xi1>
%1962 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1961, %1960, %1954 : tensor<2x640x32x32xi1>, tensor<f16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1963 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1962, %1957 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1307 : tensor<2x640x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x32x32xi1>
%1964 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1963, %1957, %1962 : tensor<2x640x32x32xi1>, tensor<f16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1965 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1964 : tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1966 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1965, %cst_880 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1967 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1966, %1952 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%1968 = linalg.generic {indexing_maps = [#map17, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_170 : tensor<640x40x1x1x1xf16>) outs(%1402 : tensor<640x40x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16x1x1xf16>
%1969 = linalg.generic {indexing_maps = [#map18, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1968 : tensor<640x40x16x1x1xf16>) outs(%1402 : tensor<640x40x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16x1x1xf16>
%collapsed_1220 = tensor.collapse_shape %1969 [[0], [1, 2], [3], [4]] : tensor<640x40x16x1x1xf16> into tensor<640x640x1x1xf16>
%1970 = linalg.generic {indexing_maps = [#map13, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_171, %collapsed_1220 : tensor<640x640x1x1xi8>, tensor<640x640x1x1xf16>) outs(%1405 : tensor<640x640x1x1xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640x1x1xf16>
%1971 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_172 : tensor<640xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%1972 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1967, %1970 : tensor<2x640x32x32xf16>, tensor<640x640x1x1xf16>) outs(%1971 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16>
%1973 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1972 : tensor<2x640x32x32xf16>) outs(%1409 : tensor<2x32x32x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x32x32x640xf16>
%collapsed_1221 = tensor.collapse_shape %1973 [[0], [1, 2], [3]] : tensor<2x32x32x640xf16> into tensor<2x1024x640xf16>
%1974 = linalg.fill ins(%cst_889 : f16) outs(%1411 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16>
%1975 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_1221 : tensor<2x1024x640xf16>) outs(%1974 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1976 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1975 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_904 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1977 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1976 : tensor<2x1024x1xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x640xf16>
%1978 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1221, %1977 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1979 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1978 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.mulf %in, %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1980 = linalg.fill ins(%cst_889 : f16) outs(%1411 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16>
%1981 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1979 : tensor<2x1024x640xf16>) outs(%1980 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1982 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1981 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_904 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1983 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1982 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.truncf %cst_895 : f64 to f16
%10166 = arith.addf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<2x1024x1xf16>
%1984 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1983 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.rsqrt %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%1985 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1984 : tensor<2x1024x1xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x640xf16>
%1986 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1978, %1985 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%expanded_1222 = tensor.expand_shape %1986 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%1987 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1222 : tensor<2x1024x40x16xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x16xf16>
%1988 = linalg.fill ins(%cst_891 : f16) outs(%1428 : tensor<2x1024x40x1xf16>) -> tensor<2x1024x40x1xf16>
%1989 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1987 : tensor<2x1024x40x16xf16>) outs(%1988 : tensor<2x1024x40x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x1xf16>
%1990 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1989 : tensor<2x1024x40x1xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1223 = tensor.collapse_shape %1990 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1991 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1992 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1991 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1993 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1992 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1994 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1993 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%1995 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1223, %1994 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1996 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1986, %1995 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1997 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1996, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%1998 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%1999 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %1998 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2000 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%1999 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2001 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2002 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2001 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2003 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%2002, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2004 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1997, %2003 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%2005 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2004, %2003, %1997 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2006 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2005, %2000 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%2007 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2006, %2000, %2005 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2008 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2007 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2009 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2008, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2010 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2009, %1995 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2011 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_173 : tensor<640x40x1xf16>) outs(%1453 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1224 = tensor.collapse_shape %2011 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%2012 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_174, %collapsed_1224 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640xf16>
%2013 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%2012 : tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1225 = tensor.collapse_shape %2010 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%2014 = linalg.fill ins(%cst_889 : f16) outs(%1458 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%2015 = linalg.matmul ins(%collapsed_1225, %2013 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%2014 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%2016 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_175, %2015 : tensor<640xf16>, tensor<2048x640xf16>) outs(%1458 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2048x640xf16>
%2017 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1222 : tensor<2x1024x40x16xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x16xf16>
%2018 = linalg.fill ins(%cst_891 : f16) outs(%1428 : tensor<2x1024x40x1xf16>) -> tensor<2x1024x40x1xf16>
%2019 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2017 : tensor<2x1024x40x16xf16>) outs(%2018 : tensor<2x1024x40x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x1xf16>
%2020 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2019 : tensor<2x1024x40x1xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1226 = tensor.collapse_shape %2020 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%2021 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2022 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2021 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2023 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2022 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2024 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2023 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2025 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1226, %2024 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2026 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1986, %2025 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2027 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2026, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2028 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2029 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2028 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2030 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2029 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2031 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2032 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2031 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2033 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%2032, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2034 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2027, %2033 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%2035 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2034, %2033, %2027 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2036 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2035, %2030 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%2037 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2036, %2030, %2035 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2038 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2037 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2039 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2038, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2040 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2039, %2025 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2041 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_176 : tensor<640x40x1xf16>) outs(%1453 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1227 = tensor.collapse_shape %2041 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%2042 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_177, %collapsed_1227 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640xf16>
%2043 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%2042 : tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1228 = tensor.collapse_shape %2040 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%2044 = linalg.fill ins(%cst_889 : f16) outs(%1458 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%2045 = linalg.matmul ins(%collapsed_1228, %2043 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%2044 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%2046 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_178, %2045 : tensor<640xf16>, tensor<2048x640xf16>) outs(%1458 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2048x640xf16>
%2047 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1222 : tensor<2x1024x40x16xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x16xf16>
%2048 = linalg.fill ins(%cst_891 : f16) outs(%1428 : tensor<2x1024x40x1xf16>) -> tensor<2x1024x40x1xf16>
%2049 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2047 : tensor<2x1024x40x16xf16>) outs(%2048 : tensor<2x1024x40x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x1xf16>
%2050 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2049 : tensor<2x1024x40x1xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1229 = tensor.collapse_shape %2050 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%2051 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2052 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2051 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2053 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2052 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2054 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2053 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2055 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1229, %2054 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2056 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1986, %2055 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2057 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2056, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2058 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2059 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2058 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2060 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2059 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2061 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2062 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2061 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2063 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%2062, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2064 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2057, %2063 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%2065 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2064, %2063, %2057 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2066 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2065, %2060 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%2067 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2066, %2060, %2065 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2068 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2067 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2069 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2068, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2070 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2069, %2055 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2071 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_179 : tensor<640x40x1xf16>) outs(%1453 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1230 = tensor.collapse_shape %2071 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%2072 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_180, %collapsed_1230 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640xf16>
%2073 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%2072 : tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1231 = tensor.collapse_shape %2070 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%2074 = linalg.fill ins(%cst_889 : f16) outs(%1458 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%2075 = linalg.matmul ins(%collapsed_1231, %2073 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%2074 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%2076 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_181, %2075 : tensor<640xf16>, tensor<2048x640xf16>) outs(%1458 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2048x640xf16>
%expanded_1232 = tensor.expand_shape %2016 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x8x80xf16>
%2077 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1232 : tensor<2x1024x8x80xf16>) outs(%1522 : tensor<2x8x1024x80xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x1024x80xf16>
%collapsed_1233 = tensor.collapse_shape %2077 [[0, 1], [2], [3]] : tensor<2x8x1024x80xf16> into tensor<16x1024x80xf16>
%expanded_1234 = tensor.expand_shape %2046 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x8x80xf16>
%2078 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1234 : tensor<2x1024x8x80xf16>) outs(%1522 : tensor<2x8x1024x80xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x1024x80xf16>
%collapsed_1235 = tensor.collapse_shape %2078 [[0, 1], [2], [3]] : tensor<2x8x1024x80xf16> into tensor<16x1024x80xf16>
%expanded_1236 = tensor.expand_shape %2076 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x8x80xf16>
%2079 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1236 : tensor<2x1024x8x80xf16>) outs(%1522 : tensor<2x8x1024x80xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x1024x80xf16>
%collapsed_1237 = tensor.collapse_shape %2079 [[0, 1], [2], [3]] : tensor<2x8x1024x80xf16> into tensor<16x1024x80xf16>
%2080 = linalg.generic {indexing_maps = [#map20, #map25], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1235 : tensor<16x1024x80xf16>) outs(%1526 : tensor<16x80x1024xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<16x80x1024xf16>
%2081 = linalg.fill ins(%cst_889 : f16) outs(%1528 : tensor<16x1024x1024xf16>) -> tensor<16x1024x1024xf16>
%2082 = linalg.batch_matmul ins(%collapsed_1233, %2080 : tensor<16x1024x80xf16>, tensor<16x80x1024xf16>) outs(%2081 : tensor<16x1024x1024xf16>) -> tensor<16x1024x1024xf16>
%2083 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2082, %cst_0 : tensor<16x1024x1024xf16>, tensor<f16>) outs(%1528 : tensor<16x1024x1024xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x1024xf16>
%2084 = linalg.fill ins(%cst_891 : f16) outs(%1532 : tensor<16x1024x1xf16>) -> tensor<16x1024x1xf16>
%2085 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2083 : tensor<16x1024x1024xf16>) outs(%2084 : tensor<16x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x1xf16>
%2086 = linalg.generic {indexing_maps = [#map20, #map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2083, %2085 : tensor<16x1024x1024xf16>, tensor<16x1024x1xf16>) outs(%1528 : tensor<16x1024x1024xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x1024xf16>
%2087 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2086 : tensor<16x1024x1024xf16>) outs(%1528 : tensor<16x1024x1024xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.exp %in : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x1024xf16>
%2088 = linalg.fill ins(%cst_889 : f16) outs(%1532 : tensor<16x1024x1xf16>) -> tensor<16x1024x1xf16>
%2089 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2087 : tensor<16x1024x1024xf16>) outs(%2088 : tensor<16x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x1xf16>
%2090 = linalg.generic {indexing_maps = [#map20, #map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2087, %2089 : tensor<16x1024x1024xf16>, tensor<16x1024x1xf16>) outs(%1528 : tensor<16x1024x1024xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x1024xf16>
%2091 = linalg.fill ins(%cst_889 : f16) outs(%1540 : tensor<16x1024x80xf16>) -> tensor<16x1024x80xf16>
%2092 = linalg.batch_matmul ins(%2090, %collapsed_1237 : tensor<16x1024x1024xf16>, tensor<16x1024x80xf16>) outs(%2091 : tensor<16x1024x80xf16>) -> tensor<16x1024x80xf16>
%expanded_1238 = tensor.expand_shape %2092 [[0, 1], [2], [3]] : tensor<16x1024x80xf16> into tensor<2x8x1024x80xf16>
%2093 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1238 : tensor<2x8x1024x80xf16>) outs(%1543 : tensor<2x1024x8x80xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x8x80xf16>
%collapsed_1239 = tensor.collapse_shape %2093 [[0], [1], [2, 3]] : tensor<2x1024x8x80xf16> into tensor<2x1024x640xf16>
%expanded_1240 = tensor.expand_shape %collapsed_1239 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%2094 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1240 : tensor<2x1024x40x16xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x16xf16>
%2095 = linalg.fill ins(%cst_891 : f16) outs(%1428 : tensor<2x1024x40x1xf16>) -> tensor<2x1024x40x1xf16>
%2096 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2094 : tensor<2x1024x40x16xf16>) outs(%2095 : tensor<2x1024x40x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x1xf16>
%2097 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2096 : tensor<2x1024x40x1xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1241 = tensor.collapse_shape %2097 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%2098 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2099 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2098 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2100 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2099 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2101 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2100 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2102 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1241, %2101 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2103 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1239, %2102 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2104 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2103, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2105 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2106 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2105 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2107 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2106 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2108 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2109 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2108 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2110 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%2109, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2111 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2104, %2110 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%2112 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2111, %2110, %2104 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2113 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2112, %2107 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%2114 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2113, %2107, %2112 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2115 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2114 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2116 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2115, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2117 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2116, %2102 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2118 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_182 : tensor<640x40x1xf16>) outs(%1453 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1242 = tensor.collapse_shape %2118 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%2119 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_183, %collapsed_1242 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640xf16>
%2120 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%2119 : tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1243 = tensor.collapse_shape %2117 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%2121 = linalg.fill ins(%cst_889 : f16) outs(%1458 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%2122 = linalg.matmul ins(%collapsed_1243, %2120 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%2121 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%2123 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_184, %2122 : tensor<640xf16>, tensor<2048x640xf16>) outs(%1458 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2048x640xf16>
%expanded_1244 = tensor.expand_shape %2123 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16>
%2124 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1244, %collapsed_1221 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2125 = linalg.fill ins(%cst_889 : f16) outs(%1411 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16>
%2126 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2124 : tensor<2x1024x640xf16>) outs(%2125 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%2127 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2126 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_904 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%2128 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2127 : tensor<2x1024x1xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x640xf16>
%2129 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2124, %2128 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2130 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2129 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.mulf %in, %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2131 = linalg.fill ins(%cst_889 : f16) outs(%1411 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16>
%2132 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2130 : tensor<2x1024x640xf16>) outs(%2131 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%2133 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2132 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_904 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%2134 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2133 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.truncf %cst_895 : f64 to f16
%10166 = arith.addf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<2x1024x1xf16>
%2135 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2134 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.rsqrt %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%2136 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2135 : tensor<2x1024x1xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x640xf16>
%2137 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2129, %2136 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%expanded_1245 = tensor.expand_shape %2137 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%2138 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1245 : tensor<2x1024x40x16xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x16xf16>
%2139 = linalg.fill ins(%cst_891 : f16) outs(%1428 : tensor<2x1024x40x1xf16>) -> tensor<2x1024x40x1xf16>
%2140 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2138 : tensor<2x1024x40x16xf16>) outs(%2139 : tensor<2x1024x40x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x1xf16>
%2141 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2140 : tensor<2x1024x40x1xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1246 = tensor.collapse_shape %2141 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%2142 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2143 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2142 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2144 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2143 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2145 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2144 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2146 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1246, %2145 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2147 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2137, %2146 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2148 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2147, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2149 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2150 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2149 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2151 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2150 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2152 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2153 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2152 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2154 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%2153, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2155 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2148, %2154 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%2156 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2155, %2154, %2148 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2157 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2156, %2151 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%2158 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2157, %2151, %2156 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2159 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2158 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2160 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2159, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2161 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2160, %2146 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2162 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_185 : tensor<640x40x1xf16>) outs(%1453 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1247 = tensor.collapse_shape %2162 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%2163 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_186, %collapsed_1247 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640xf16>
%2164 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%2163 : tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1248 = tensor.collapse_shape %2161 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%2165 = linalg.fill ins(%cst_889 : f16) outs(%1458 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%2166 = linalg.matmul ins(%collapsed_1248, %2164 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%2165 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%2167 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_187, %2166 : tensor<640xf16>, tensor<2048x640xf16>) outs(%1458 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2048x640xf16>
%2168 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_980 : tensor<2x77x48x16xf16>) outs(%440 : tensor<2x77x48x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x77x48x16xf16>
%2169 = linalg.fill ins(%cst_891 : f16) outs(%442 : tensor<2x77x48x1xf16>) -> tensor<2x77x48x1xf16>
%2170 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2168 : tensor<2x77x48x16xf16>) outs(%2169 : tensor<2x77x48x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x77x48x1xf16>
%2171 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2170 : tensor<2x77x48x1xf16>) outs(%440 : tensor<2x77x48x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x77x48x16xf16>
%collapsed_1249 = tensor.collapse_shape %2171 [[0], [1], [2, 3]] : tensor<2x77x48x16xf16> into tensor<2x77x768xf16>
%2172 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2173 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2172 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2174 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2173 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2175 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2174 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2176 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1249, %2175 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%2177 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2, %2176 : tensor<2x77x768xf16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%2178 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2177, %cst_880 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%2179 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2180 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2179 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2181 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2180 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2182 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2183 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2182 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2184 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%2183, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2185 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2178, %2184 : tensor<2x77x768xf16>, tensor<f16>) outs(%460 : tensor<2x77x768xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x77x768xi1>
%2186 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2185, %2184, %2178 : tensor<2x77x768xi1>, tensor<f16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%2187 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2186, %2181 : tensor<2x77x768xf16>, tensor<f16>) outs(%460 : tensor<2x77x768xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x77x768xi1>
%2188 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2187, %2181, %2186 : tensor<2x77x768xi1>, tensor<f16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%2189 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2188 : tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%2190 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2189, %cst_880 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%2191 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2190, %2176 : tensor<2x77x768xf16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%2192 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_188 : tensor<640x48x1xf16>) outs(%1643 : tensor<640x48x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x48x16xf16>
%collapsed_1250 = tensor.collapse_shape %2192 [[0], [1, 2]] : tensor<640x48x16xf16> into tensor<640x768xf16>
%2193 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_189, %collapsed_1250 : tensor<640x768xi8>, tensor<640x768xf16>) outs(%1645 : tensor<640x768xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x768xf16>
%2194 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%2193 : tensor<640x768xf16>) outs(%1647 : tensor<768x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<768x640xf16>
%collapsed_1251 = tensor.collapse_shape %2191 [[0, 1], [2]] : tensor<2x77x768xf16> into tensor<154x768xf16>
%2195 = linalg.fill ins(%cst_889 : f16) outs(%1649 : tensor<154x640xf16>) -> tensor<154x640xf16>
%2196 = linalg.matmul ins(%collapsed_1251, %2194 : tensor<154x768xf16>, tensor<768x640xf16>) outs(%2195 : tensor<154x640xf16>) -> tensor<154x640xf16>
%2197 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_980 : tensor<2x77x48x16xf16>) outs(%440 : tensor<2x77x48x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x77x48x16xf16>
%2198 = linalg.fill ins(%cst_891 : f16) outs(%442 : tensor<2x77x48x1xf16>) -> tensor<2x77x48x1xf16>
%2199 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2197 : tensor<2x77x48x16xf16>) outs(%2198 : tensor<2x77x48x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x77x48x1xf16>
%2200 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2199 : tensor<2x77x48x1xf16>) outs(%440 : tensor<2x77x48x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x77x48x16xf16>
%collapsed_1252 = tensor.collapse_shape %2200 [[0], [1], [2, 3]] : tensor<2x77x48x16xf16> into tensor<2x77x768xf16>
%2201 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2202 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2201 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2203 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2202 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2204 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2203 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2205 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1252, %2204 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%2206 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2, %2205 : tensor<2x77x768xf16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%2207 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2206, %cst_880 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%2208 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2209 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2208 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2210 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2209 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2211 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2212 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2211 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2213 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%2212, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2214 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2207, %2213 : tensor<2x77x768xf16>, tensor<f16>) outs(%460 : tensor<2x77x768xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x77x768xi1>
%2215 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2214, %2213, %2207 : tensor<2x77x768xi1>, tensor<f16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%2216 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2215, %2210 : tensor<2x77x768xf16>, tensor<f16>) outs(%460 : tensor<2x77x768xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x77x768xi1>
%2217 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2216, %2210, %2215 : tensor<2x77x768xi1>, tensor<f16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%2218 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2217 : tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%2219 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2218, %cst_880 : tensor<2x77x768xf16>, tensor<f16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%2220 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2219, %2205 : tensor<2x77x768xf16>, tensor<2x77x768xf16>) outs(%450 : tensor<2x77x768xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x77x768xf16>
%2221 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_190 : tensor<640x48x1xf16>) outs(%1643 : tensor<640x48x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x48x16xf16>
%collapsed_1253 = tensor.collapse_shape %2221 [[0], [1, 2]] : tensor<640x48x16xf16> into tensor<640x768xf16>
%2222 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_191, %collapsed_1253 : tensor<640x768xi8>, tensor<640x768xf16>) outs(%1645 : tensor<640x768xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x768xf16>
%2223 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%2222 : tensor<640x768xf16>) outs(%1647 : tensor<768x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<768x640xf16>
%collapsed_1254 = tensor.collapse_shape %2220 [[0, 1], [2]] : tensor<2x77x768xf16> into tensor<154x768xf16>
%2224 = linalg.fill ins(%cst_889 : f16) outs(%1649 : tensor<154x640xf16>) -> tensor<154x640xf16>
%2225 = linalg.matmul ins(%collapsed_1254, %2223 : tensor<154x768xf16>, tensor<768x640xf16>) outs(%2224 : tensor<154x640xf16>) -> tensor<154x640xf16>
%expanded_1255 = tensor.expand_shape %2167 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x8x80xf16>
%2226 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1255 : tensor<2x1024x8x80xf16>) outs(%1522 : tensor<2x8x1024x80xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x1024x80xf16>
%collapsed_1256 = tensor.collapse_shape %2226 [[0, 1], [2], [3]] : tensor<2x8x1024x80xf16> into tensor<16x1024x80xf16>
%expanded_1257 = tensor.expand_shape %2196 [[0, 1], [2, 3]] : tensor<154x640xf16> into tensor<2x77x8x80xf16>
%2227 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1257 : tensor<2x77x8x80xf16>) outs(%1682 : tensor<2x8x77x80xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x77x80xf16>
%collapsed_1258 = tensor.collapse_shape %2227 [[0, 1], [2], [3]] : tensor<2x8x77x80xf16> into tensor<16x77x80xf16>
%expanded_1259 = tensor.expand_shape %2225 [[0, 1], [2, 3]] : tensor<154x640xf16> into tensor<2x77x8x80xf16>
%2228 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1259 : tensor<2x77x8x80xf16>) outs(%1682 : tensor<2x8x77x80xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x8x77x80xf16>
%collapsed_1260 = tensor.collapse_shape %2228 [[0, 1], [2], [3]] : tensor<2x8x77x80xf16> into tensor<16x77x80xf16>
%2229 = linalg.generic {indexing_maps = [#map20, #map25], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1258 : tensor<16x77x80xf16>) outs(%1685 : tensor<16x80x77xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<16x80x77xf16>
%2230 = linalg.fill ins(%cst_889 : f16) outs(%1687 : tensor<16x1024x77xf16>) -> tensor<16x1024x77xf16>
%2231 = linalg.batch_matmul ins(%collapsed_1256, %2229 : tensor<16x1024x80xf16>, tensor<16x80x77xf16>) outs(%2230 : tensor<16x1024x77xf16>) -> tensor<16x1024x77xf16>
%2232 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2231, %cst_0 : tensor<16x1024x77xf16>, tensor<f16>) outs(%1687 : tensor<16x1024x77xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x77xf16>
%2233 = linalg.fill ins(%cst_891 : f16) outs(%1532 : tensor<16x1024x1xf16>) -> tensor<16x1024x1xf16>
%2234 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2232 : tensor<16x1024x77xf16>) outs(%2233 : tensor<16x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x1xf16>
%2235 = linalg.generic {indexing_maps = [#map20, #map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2232, %2234 : tensor<16x1024x77xf16>, tensor<16x1024x1xf16>) outs(%1687 : tensor<16x1024x77xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x77xf16>
%2236 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2235 : tensor<16x1024x77xf16>) outs(%1687 : tensor<16x1024x77xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.exp %in : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x77xf16>
%2237 = linalg.fill ins(%cst_889 : f16) outs(%1532 : tensor<16x1024x1xf16>) -> tensor<16x1024x1xf16>
%2238 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2236 : tensor<16x1024x77xf16>) outs(%2237 : tensor<16x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x1xf16>
%2239 = linalg.generic {indexing_maps = [#map20, #map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2236, %2238 : tensor<16x1024x77xf16>, tensor<16x1024x1xf16>) outs(%1687 : tensor<16x1024x77xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<16x1024x77xf16>
%2240 = linalg.fill ins(%cst_889 : f16) outs(%1540 : tensor<16x1024x80xf16>) -> tensor<16x1024x80xf16>
%2241 = linalg.batch_matmul ins(%2239, %collapsed_1260 : tensor<16x1024x77xf16>, tensor<16x77x80xf16>) outs(%2240 : tensor<16x1024x80xf16>) -> tensor<16x1024x80xf16>
%expanded_1261 = tensor.expand_shape %2241 [[0, 1], [2], [3]] : tensor<16x1024x80xf16> into tensor<2x8x1024x80xf16>
%2242 = linalg.generic {indexing_maps = [#map11, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1261 : tensor<2x8x1024x80xf16>) outs(%1543 : tensor<2x1024x8x80xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x8x80xf16>
%collapsed_1262 = tensor.collapse_shape %2242 [[0], [1], [2, 3]] : tensor<2x1024x8x80xf16> into tensor<2x1024x640xf16>
%expanded_1263 = tensor.expand_shape %collapsed_1262 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%2243 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1263 : tensor<2x1024x40x16xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x16xf16>
%2244 = linalg.fill ins(%cst_891 : f16) outs(%1428 : tensor<2x1024x40x1xf16>) -> tensor<2x1024x40x1xf16>
%2245 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2243 : tensor<2x1024x40x16xf16>) outs(%2244 : tensor<2x1024x40x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x40x1xf16>
%2246 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2245 : tensor<2x1024x40x1xf16>) outs(%1426 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1264 = tensor.collapse_shape %2246 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%2247 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2248 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2247 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2249 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2248 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2250 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2249 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2251 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1264, %2250 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2252 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1262, %2251 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2253 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2252, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2254 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2255 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2254 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2256 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2255 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2257 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2258 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2257 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2259 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%2258, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2260 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2253, %2259 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%2261 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2260, %2259, %2253 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2262 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2261, %2256 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1445 : tensor<2x1024x640xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1024x640xi1>
%2263 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2262, %2256, %2261 : tensor<2x1024x640xi1>, tensor<f16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2264 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2263 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2265 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2264, %cst_880 : tensor<2x1024x640xf16>, tensor<f16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2266 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2265, %2251 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2267 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_192 : tensor<640x40x1xf16>) outs(%1453 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1265 = tensor.collapse_shape %2267 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%2268 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_193, %collapsed_1265 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640xf16>
%2269 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%2268 : tensor<640x640xf16>) outs(%1455 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1266 = tensor.collapse_shape %2266 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%2270 = linalg.fill ins(%cst_889 : f16) outs(%1458 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%2271 = linalg.matmul ins(%collapsed_1266, %2269 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%2270 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%2272 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_194, %2271 : tensor<640xf16>, tensor<2048x640xf16>) outs(%1458 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2048x640xf16>
%expanded_1267 = tensor.expand_shape %2272 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16>
%2273 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1267, %2124 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2274 = linalg.fill ins(%cst_889 : f16) outs(%1411 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16>
%2275 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2273 : tensor<2x1024x640xf16>) outs(%2274 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%2276 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2275 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_904 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%2277 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2276 : tensor<2x1024x1xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x640xf16>
%2278 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2273, %2277 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2279 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2278 : tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.mulf %in, %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2280 = linalg.fill ins(%cst_889 : f16) outs(%1411 : tensor<2x1024x1xf16>) -> tensor<2x1024x1xf16>
%2281 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2279 : tensor<2x1024x640xf16>) outs(%2280 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%2282 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2281 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_904 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%2283 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2282 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.truncf %cst_895 : f64 to f16
%10166 = arith.addf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<2x1024x1xf16>
%2284 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2283 : tensor<2x1024x1xf16>) outs(%1411 : tensor<2x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.rsqrt %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x1xf16>
%2285 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2284 : tensor<2x1024x1xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x640xf16>
%2286 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2278, %2285 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%2287 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_195 : tensor<5120x40x1xf16>) outs(%1745 : tensor<5120x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<5120x40x16xf16>
%collapsed_1268 = tensor.collapse_shape %2287 [[0], [1, 2]] : tensor<5120x40x16xf16> into tensor<5120x640xf16>
%2288 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_196, %collapsed_1268 : tensor<5120x640xi8>, tensor<5120x640xf16>) outs(%1747 : tensor<5120x640xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<5120x640xf16>
%2289 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%2288 : tensor<5120x640xf16>) outs(%1749 : tensor<640x5120xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x5120xf16>
%collapsed_1269 = tensor.collapse_shape %2286 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%2290 = linalg.fill ins(%cst_889 : f16) outs(%1751 : tensor<2048x5120xf16>) -> tensor<2048x5120xf16>
%2291 = linalg.matmul ins(%collapsed_1269, %2289 : tensor<2048x640xf16>, tensor<640x5120xf16>) outs(%2290 : tensor<2048x5120xf16>) -> tensor<2048x5120xf16>
%2292 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_197, %2291 : tensor<5120xf16>, tensor<2048x5120xf16>) outs(%1751 : tensor<2048x5120xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2048x5120xf16>
%expanded_1270 = tensor.expand_shape %2292 [[0, 1], [2]] : tensor<2048x5120xf16> into tensor<2x1024x5120xf16>
%extracted_slice_1271 = tensor.extract_slice %expanded_1270[0, 0, 0] [2, 1024, 2560] [1, 1, 1] : tensor<2x1024x5120xf16> to tensor<2x1024x2560xf16>
%extracted_slice_1272 = tensor.extract_slice %expanded_1270[0, 0, 2560] [2, 1024, 2560] [1, 1, 1] : tensor<2x1024x5120xf16> to tensor<2x1024x2560xf16>
%2293 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1272 : tensor<2x1024x2560xf16>) outs(%1755 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.sqrt %cst_892 : f16
%10166 = arith.divf %in, %10165 : f16
%10167 = math.erf %10166 : f16
%10168 = arith.addf %10167, %cst_890 : f16
%10169 = arith.mulf %10168, %cst_893 : f16
%10170 = arith.mulf %in, %10169 : f16
linalg.yield %10170 : f16
} -> tensor<2x1024x2560xf16>
%2294 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1271, %2293 : tensor<2x1024x2560xf16>, tensor<2x1024x2560xf16>) outs(%1755 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x2560xf16>
%2295 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_198 : tensor<640x160x1xf16>) outs(%1758 : tensor<640x160x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x160x16xf16>
%collapsed_1273 = tensor.collapse_shape %2295 [[0], [1, 2]] : tensor<640x160x16xf16> into tensor<640x2560xf16>
%2296 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_199, %collapsed_1273 : tensor<640x2560xi8>, tensor<640x2560xf16>) outs(%1760 : tensor<640x2560xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x2560xf16>
%2297 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%2296 : tensor<640x2560xf16>) outs(%1762 : tensor<2560x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2560x640xf16>
%collapsed_1274 = tensor.collapse_shape %2294 [[0, 1], [2]] : tensor<2x1024x2560xf16> into tensor<2048x2560xf16>
%2298 = linalg.fill ins(%cst_889 : f16) outs(%1458 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%2299 = linalg.matmul ins(%collapsed_1274, %2297 : tensor<2048x2560xf16>, tensor<2560x640xf16>) outs(%2298 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%2300 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_200, %2299 : tensor<640xf16>, tensor<2048x640xf16>) outs(%1458 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2048x640xf16>
%expanded_1275 = tensor.expand_shape %2300 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16>
%2301 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1275, %2273 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%1415 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1024x640xf16>
%expanded_1276 = tensor.expand_shape %2301 [[0], [1, 2], [3]] : tensor<2x1024x640xf16> into tensor<2x32x32x640xf16>
%2302 = linalg.generic {indexing_maps = [#map11, #map26], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1276 : tensor<2x32x32x640xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%expanded_1277 = tensor.expand_shape %2302 [[0], [1, 2], [3], [4]] : tensor<2x640x32x32xf16> into tensor<2x40x16x32x32xf16>
%2303 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1277 : tensor<2x40x16x32x32xf16>) outs(%1288 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x40x16x32x32xf16>
%2304 = linalg.fill ins(%cst_891 : f16) outs(%1290 : tensor<2x40x1x32x32xf16>) -> tensor<2x40x1x32x32xf16>
%2305 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%2303 : tensor<2x40x16x32x32xf16>) outs(%2304 : tensor<2x40x1x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x40x1x32x32xf16>
%2306 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%2305 : tensor<2x40x1x32x32xf16>) outs(%1288 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x40x16x32x32xf16>
%collapsed_1278 = tensor.collapse_shape %2306 [[0], [1, 2], [3], [4]] : tensor<2x40x16x32x32xf16> into tensor<2x640x32x32xf16>
%2307 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2308 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2307 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2309 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2308 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2310 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2309 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2311 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1278, %2310 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%2312 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2302, %2311 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%2313 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2312, %cst_880 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%2314 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2315 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2314 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2316 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2315 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2317 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2318 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2317 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2319 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%2318, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2320 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2313, %2319 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1307 : tensor<2x640x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x32x32xi1>
%2321 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2320, %2319, %2313 : tensor<2x640x32x32xi1>, tensor<f16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%2322 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2321, %2316 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1307 : tensor<2x640x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x32x32xi1>
%2323 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2322, %2316, %2321 : tensor<2x640x32x32xi1>, tensor<f16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%2324 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2323 : tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%2325 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2324, %cst_880 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%2326 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2325, %2311 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%2327 = linalg.generic {indexing_maps = [#map17, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_201 : tensor<640x40x1x1x1xf16>) outs(%1402 : tensor<640x40x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16x1x1xf16>
%2328 = linalg.generic {indexing_maps = [#map18, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%2327 : tensor<640x40x16x1x1xf16>) outs(%1402 : tensor<640x40x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16x1x1xf16>
%collapsed_1279 = tensor.collapse_shape %2328 [[0], [1, 2], [3], [4]] : tensor<640x40x16x1x1xf16> into tensor<640x640x1x1xf16>
%2329 = linalg.generic {indexing_maps = [#map13, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_202, %collapsed_1279 : tensor<640x640x1x1xi8>, tensor<640x640x1x1xf16>) outs(%1405 : tensor<640x640x1x1xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640x1x1xf16>
%2330 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_203 : tensor<640xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%2331 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2326, %2329 : tensor<2x640x32x32xf16>, tensor<640x640x1x1xf16>) outs(%2330 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16>
%2332 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2331, %1919 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%expanded_1280 = tensor.expand_shape %2332 [[0], [1, 2], [3], [4]] : tensor<2x640x32x32xf16> into tensor<2x40x16x32x32xf16>
%2333 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1280 : tensor<2x40x16x32x32xf16>) outs(%1288 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x40x16x32x32xf16>
%2334 = linalg.fill ins(%cst_891 : f16) outs(%1290 : tensor<2x40x1x32x32xf16>) -> tensor<2x40x1x32x32xf16>
%2335 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%2333 : tensor<2x40x16x32x32xf16>) outs(%2334 : tensor<2x40x1x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x40x1x32x32xf16>
%2336 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%2335 : tensor<2x40x1x32x32xf16>) outs(%1288 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x40x16x32x32xf16>
%collapsed_1281 = tensor.collapse_shape %2336 [[0], [1, 2], [3], [4]] : tensor<2x40x16x32x32xf16> into tensor<2x640x32x32xf16>
%2337 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2338 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2337 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2339 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2338 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2340 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2339 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2341 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1281, %2340 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%2342 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2332, %2341 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%2343 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2342, %cst_880 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%2344 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2345 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2344 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2346 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2345 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2347 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2348 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2347 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2349 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%2348, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2350 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2343, %2349 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1307 : tensor<2x640x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x32x32xi1>
%2351 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2350, %2349, %2343 : tensor<2x640x32x32xi1>, tensor<f16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%2352 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2351, %2346 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1307 : tensor<2x640x32x32xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x32x32xi1>
%2353 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2352, %2346, %2351 : tensor<2x640x32x32xi1>, tensor<f16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%2354 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2353 : tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%2355 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2354, %cst_880 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%2356 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2355, %2341 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%1246 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x32x32xf16>
%2357 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_204 : tensor<640x40x1x3x3xf16>) outs(%1315 : tensor<640x40x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16x3x3xf16>
%collapsed_1282 = tensor.collapse_shape %2357 [[0], [1, 2], [3], [4]] : tensor<640x40x16x3x3xf16> into tensor<640x640x3x3xf16>
%2358 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_205, %collapsed_1282 : tensor<640x640x3x3xi8>, tensor<640x640x3x3xf16>) outs(%1317 : tensor<640x640x3x3xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<640x640x3x3xf16>
%padded_1283 = tensor.pad %2356 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
tensor.yield %cst_889 : f16
} : tensor<2x640x32x32xf16> to tensor<2x640x34x34xf16>
%2359 = tensor.empty() : tensor<2x640x16x16xf16>
%2360 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_206 : tensor<640xf16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x16x16xf16>
%2361 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%padded_1283, %2358 : tensor<2x640x34x34xf16>, tensor<640x640x3x3xf16>) outs(%2360 : tensor<2x640x16x16xf16>) -> tensor<2x640x16x16xf16>
%collapsed_1284 = tensor.collapse_shape %2361 [[0], [1], [2, 3]] : tensor<2x640x16x16xf16> into tensor<2x640x256xf16>
%expanded_1285 = tensor.expand_shape %collapsed_1284 [[0], [1, 2], [3]] : tensor<2x640x256xf16> into tensor<2x32x20x256xf16>
%2362 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%2363 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2362 : tensor<f64>) outs(%9 : tensor<f32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<f32>
%2364 = tensor.empty() : tensor<2x32x20x256xf32>
%2365 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2363 : tensor<f32>) outs(%2364 : tensor<2x32x20x256xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<2x32x20x256xf32>
%2366 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1285 : tensor<2x32x20x256xf16>) outs(%2365 : tensor<2x32x20x256xf32>) {
^bb0(%in: f16, %out: f32):
%10165 = arith.extf %in : f16 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x256xf32>
%2367 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%2368 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2366 : tensor<2x32x20x256xf32>) outs(%2367 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2369 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2368 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_905 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2370 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2366, %2369 : tensor<2x32x20x256xf32>, tensor<2x32x1x1xf32>) outs(%2364 : tensor<2x32x20x256xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.subf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x256xf32>
%2371 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2370 : tensor<2x32x20x256xf32>) outs(%2364 : tensor<2x32x20x256xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.mulf %in, %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x256xf32>
%2372 = linalg.fill ins(%cst_894 : f64) outs(%55 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
%2373 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2371 : tensor<2x32x20x256xf32>) outs(%2372 : tensor<2x32x1x1xf64>) {
^bb0(%in: f32, %out: f64):
%10165 = arith.extf %in : f32 to f64
%10166 = arith.addf %10165, %out : f64
linalg.yield %10166 : f64
} -> tensor<2x32x1x1xf64>
%2374 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2373 : tensor<2x32x1x1xf64>) outs(%55 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%10165 = arith.divf %in, %cst_906 : f64
linalg.yield %10165 : f64
} -> tensor<2x32x1x1xf64>
%2375 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2374 : tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2376 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%2377 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2366 : tensor<2x32x20x256xf32>) outs(%2376 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2378 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2377 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_905 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2379 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2375, %cst_4 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x1x1xf32>
%2380 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2379 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = math.rsqrt %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2381 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1285, %2378 : tensor<2x32x20x256xf16>, tensor<2x32x1x1xf32>) outs(%2364 : tensor<2x32x20x256xf32>) {
^bb0(%in: f16, %in_2540: f32, %out: f32):
%10165 = arith.extf %in : f16 to f32
%10166 = arith.subf %10165, %in_2540 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x20x256xf32>
%2382 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2381, %2380 : tensor<2x32x20x256xf32>, tensor<2x32x1x1xf32>) outs(%2364 : tensor<2x32x20x256xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.mulf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x20x256xf32>
%collapsed_1286 = tensor.collapse_shape %2382 [[0], [1, 2], [3]] : tensor<2x32x20x256xf32> into tensor<2x640x256xf32>
%expanded_1287 = tensor.expand_shape %collapsed_1286 [[0], [1], [2, 3]] : tensor<2x640x256xf32> into tensor<2x640x16x16xf32>
%expanded_1288 = tensor.expand_shape %cst_207 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16>
%2383 = tensor.empty() : tensor<2x640x16x16xf32>
%2384 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1287, %expanded_1288 : tensor<2x640x16x16xf32>, tensor<640x1x1xf16>) outs(%2383 : tensor<2x640x16x16xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.mulf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x640x16x16xf32>
%expanded_1289 = tensor.expand_shape %cst_208 [[0, 1, 2]] : tensor<640xf16> into tensor<640x1x1xf16>
%2385 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2384, %expanded_1289 : tensor<2x640x16x16xf32>, tensor<640x1x1xf16>) outs(%2383 : tensor<2x640x16x16xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x640x16x16xf32>
%2386 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%2387 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2386 : tensor<f64>) outs(%23 : tensor<f16>) {
^bb0(%in: f64, %out: f16):
%10165 = arith.truncf %in : f64 to f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2388 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2387 : tensor<f16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x16x16xf16>
%2389 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2385 : tensor<2x640x16x16xf32>) outs(%2388 : tensor<2x640x16x16xf16>) {
^bb0(%in: f32, %out: f16):
%10165 = arith.truncf %in : f32 to f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%2390 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2389 : tensor<2x640x16x16xf16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
%10166 = math.exp %10165 : f16
%10167 = arith.addf %10166, %cst_890 : f16
%10168 = arith.divf %cst_890, %10167 : f16
linalg.yield %10168 : f16
} -> tensor<2x640x16x16xf16>
%2391 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2390, %2389 : tensor<2x640x16x16xf16>, tensor<2x640x16x16xf16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%expanded_1290 = tensor.expand_shape %2391 [[0], [1, 2], [3], [4]] : tensor<2x640x16x16xf16> into tensor<2x40x16x16x16xf16>
%2392 = tensor.empty() : tensor<2x40x16x16x16xf16>
%2393 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1290 : tensor<2x40x16x16x16xf16>) outs(%2392 : tensor<2x40x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x40x16x16x16xf16>
%2394 = tensor.empty() : tensor<2x40x1x16x16xf16>
%2395 = linalg.fill ins(%cst_891 : f16) outs(%2394 : tensor<2x40x1x16x16xf16>) -> tensor<2x40x1x16x16xf16>
%2396 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%2393 : tensor<2x40x16x16x16xf16>) outs(%2395 : tensor<2x40x1x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x40x1x16x16xf16>
%2397 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%2396 : tensor<2x40x1x16x16xf16>) outs(%2392 : tensor<2x40x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x40x16x16x16xf16>
%collapsed_1291 = tensor.collapse_shape %2397 [[0], [1, 2], [3], [4]] : tensor<2x40x16x16x16xf16> into tensor<2x640x16x16xf16>
%2398 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2399 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2398 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2400 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2399 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2401 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2400 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2402 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1291, %2401 : tensor<2x640x16x16xf16>, tensor<f16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%2403 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2391, %2402 : tensor<2x640x16x16xf16>, tensor<2x640x16x16xf16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%2404 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2403, %cst_880 : tensor<2x640x16x16xf16>, tensor<f16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%2405 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2406 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2405 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2407 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2406 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2408 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2409 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2408 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2410 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%2409, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2411 = tensor.empty() : tensor<2x640x16x16xi1>
%2412 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2404, %2410 : tensor<2x640x16x16xf16>, tensor<f16>) outs(%2411 : tensor<2x640x16x16xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x16x16xi1>
%2413 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2412, %2410, %2404 : tensor<2x640x16x16xi1>, tensor<f16>, tensor<2x640x16x16xf16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%2414 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2413, %2407 : tensor<2x640x16x16xf16>, tensor<f16>) outs(%2411 : tensor<2x640x16x16xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x16x16xi1>
%2415 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2414, %2407, %2413 : tensor<2x640x16x16xi1>, tensor<f16>, tensor<2x640x16x16xf16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%2416 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2415 : tensor<2x640x16x16xf16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%2417 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2416, %cst_880 : tensor<2x640x16x16xf16>, tensor<f16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%2418 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2417, %2402 : tensor<2x640x16x16xf16>, tensor<2x640x16x16xf16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%2419 = tensor.empty() : tensor<1280x40x16x3x3xf16>
%2420 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_209 : tensor<1280x40x1x3x3xf16>) outs(%2419 : tensor<1280x40x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x40x16x3x3xf16>
%collapsed_1292 = tensor.collapse_shape %2420 [[0], [1, 2], [3], [4]] : tensor<1280x40x16x3x3xf16> into tensor<1280x640x3x3xf16>
%2421 = tensor.empty() : tensor<1280x640x3x3xf16>
%2422 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_210, %collapsed_1292 : tensor<1280x640x3x3xi8>, tensor<1280x640x3x3xf16>) outs(%2421 : tensor<1280x640x3x3xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<1280x640x3x3xf16>
%padded_1293 = tensor.pad %2418 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
tensor.yield %cst_889 : f16
} : tensor<2x640x16x16xf16> to tensor<2x640x18x18xf16>
%2423 = tensor.empty() : tensor<2x1280x16x16xf16>
%2424 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_211 : tensor<1280xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1280x16x16xf16>
%2425 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1293, %2422 : tensor<2x640x18x18xf16>, tensor<1280x640x3x3xf16>) outs(%2424 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16>
%2426 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
%10166 = math.exp %10165 : f16
%10167 = arith.addf %10166, %cst_890 : f16
%10168 = arith.divf %cst_890, %10167 : f16
linalg.yield %10168 : f16
} -> tensor<2x1280xf16>
%2427 = linalg.generic {indexing_maps = [#map5, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%2426, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280xf16>
%2428 = linalg.generic {indexing_maps = [#map5, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_212 : tensor<1280x1280xf16>) outs(%36 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%2429 = linalg.fill ins(%cst_889 : f16) outs(%30 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
%2430 = linalg.matmul ins(%2427, %2428 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%2429 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
%2431 = linalg.generic {indexing_maps = [#map9, #map5, #map5], iterator_types = ["parallel", "parallel"]} ins(%cst_213, %2430 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%30 : tensor<2x1280xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280xf16>
%expanded_1294 = tensor.expand_shape %2431 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16>
%2432 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2425, %expanded_1294 : tensor<2x1280x16x16xf16>, tensor<2x1280x1x1xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%collapsed_1295 = tensor.collapse_shape %2432 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16>
%expanded_1296 = tensor.expand_shape %collapsed_1295 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16>
%2433 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%2434 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2433 : tensor<f64>) outs(%9 : tensor<f32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<f32>
%2435 = tensor.empty() : tensor<2x32x40x256xf32>
%2436 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2434 : tensor<f32>) outs(%2435 : tensor<2x32x40x256xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<2x32x40x256xf32>
%2437 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1296 : tensor<2x32x40x256xf16>) outs(%2436 : tensor<2x32x40x256xf32>) {
^bb0(%in: f16, %out: f32):
%10165 = arith.extf %in : f16 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x40x256xf32>
%2438 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%2439 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2437 : tensor<2x32x40x256xf32>) outs(%2438 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2440 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2439 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_900 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2441 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2437, %2440 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%2435 : tensor<2x32x40x256xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.subf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x40x256xf32>
%2442 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2441 : tensor<2x32x40x256xf32>) outs(%2435 : tensor<2x32x40x256xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.mulf %in, %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x40x256xf32>
%2443 = linalg.fill ins(%cst_894 : f64) outs(%55 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
%2444 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2442 : tensor<2x32x40x256xf32>) outs(%2443 : tensor<2x32x1x1xf64>) {
^bb0(%in: f32, %out: f64):
%10165 = arith.extf %in : f32 to f64
%10166 = arith.addf %10165, %out : f64
linalg.yield %10166 : f64
} -> tensor<2x32x1x1xf64>
%2445 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2444 : tensor<2x32x1x1xf64>) outs(%55 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%10165 = arith.divf %in, %cst_901 : f64
linalg.yield %10165 : f64
} -> tensor<2x32x1x1xf64>
%2446 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2445 : tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2447 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%2448 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2437 : tensor<2x32x40x256xf32>) outs(%2447 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2449 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2448 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_900 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2450 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2446, %cst_4 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x1x1xf32>
%2451 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2450 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = math.rsqrt %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2452 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1296, %2449 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%2435 : tensor<2x32x40x256xf32>) {
^bb0(%in: f16, %in_2540: f32, %out: f32):
%10165 = arith.extf %in : f16 to f32
%10166 = arith.subf %10165, %in_2540 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x40x256xf32>
%2453 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2452, %2451 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%2435 : tensor<2x32x40x256xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.mulf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x40x256xf32>
%collapsed_1297 = tensor.collapse_shape %2453 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32>
%expanded_1298 = tensor.expand_shape %collapsed_1297 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32>
%expanded_1299 = tensor.expand_shape %cst_214 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16>
%2454 = tensor.empty() : tensor<2x1280x16x16xf32>
%2455 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1298, %expanded_1299 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%2454 : tensor<2x1280x16x16xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.mulf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x1280x16x16xf32>
%expanded_1300 = tensor.expand_shape %cst_215 [[0, 1, 2]] : tensor<1280xf16> into tensor<1280x1x1xf16>
%2456 = linalg.generic {indexing_maps = [#map11, #map14, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2455, %expanded_1300 : tensor<2x1280x16x16xf32>, tensor<1280x1x1xf16>) outs(%2454 : tensor<2x1280x16x16xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x1280x16x16xf32>
%2457 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%2458 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2457 : tensor<f64>) outs(%23 : tensor<f16>) {
^bb0(%in: f64, %out: f16):
%10165 = arith.truncf %in : f64 to f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2459 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2458 : tensor<f16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1280x16x16xf16>
%2460 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2456 : tensor<2x1280x16x16xf32>) outs(%2459 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f32, %out: f16):
%10165 = arith.truncf %in : f32 to f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2461 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2460 : tensor<2x1280x16x16xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
%10166 = math.exp %10165 : f16
%10167 = arith.addf %10166, %cst_890 : f16
%10168 = arith.divf %cst_890, %10167 : f16
linalg.yield %10168 : f16
} -> tensor<2x1280x16x16xf16>
%2462 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2461, %2460 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%expanded_1301 = tensor.expand_shape %2462 [[0], [1, 2], [3], [4]] : tensor<2x1280x16x16xf16> into tensor<2x80x16x16x16xf16>
%2463 = tensor.empty() : tensor<2x80x16x16x16xf16>
%2464 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1301 : tensor<2x80x16x16x16xf16>) outs(%2463 : tensor<2x80x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x80x16x16x16xf16>
%2465 = tensor.empty() : tensor<2x80x1x16x16xf16>
%2466 = linalg.fill ins(%cst_891 : f16) outs(%2465 : tensor<2x80x1x16x16xf16>) -> tensor<2x80x1x16x16xf16>
%2467 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%2464 : tensor<2x80x16x16x16xf16>) outs(%2466 : tensor<2x80x1x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x80x1x16x16xf16>
%2468 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%2467 : tensor<2x80x1x16x16xf16>) outs(%2463 : tensor<2x80x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x80x16x16x16xf16>
%collapsed_1302 = tensor.collapse_shape %2468 [[0], [1, 2], [3], [4]] : tensor<2x80x16x16x16xf16> into tensor<2x1280x16x16xf16>
%2469 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2470 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2469 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2471 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2470 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2472 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2471 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2473 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1302, %2472 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2474 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2462, %2473 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2475 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2474, %cst_880 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2476 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2477 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2476 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2478 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2477 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2479 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2480 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2479 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2481 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%2480, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2482 = tensor.empty() : tensor<2x1280x16x16xi1>
%2483 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2475, %2481 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%2482 : tensor<2x1280x16x16xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1280x16x16xi1>
%2484 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2483, %2481, %2475 : tensor<2x1280x16x16xi1>, tensor<f16>, tensor<2x1280x16x16xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2485 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2484, %2478 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%2482 : tensor<2x1280x16x16xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1280x16x16xi1>
%2486 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2485, %2478, %2484 : tensor<2x1280x16x16xi1>, tensor<f16>, tensor<2x1280x16x16xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2487 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2486 : tensor<2x1280x16x16xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2488 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2487, %cst_880 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2489 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2488, %2473 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2490 = tensor.empty() : tensor<1280x80x16x3x3xf16>
%2491 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_216 : tensor<1280x80x1x3x3xf16>) outs(%2490 : tensor<1280x80x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16x3x3xf16>
%collapsed_1303 = tensor.collapse_shape %2491 [[0], [1, 2], [3], [4]] : tensor<1280x80x16x3x3xf16> into tensor<1280x1280x3x3xf16>
%2492 = tensor.empty() : tensor<1280x1280x3x3xf16>
%2493 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_217, %collapsed_1303 : tensor<1280x1280x3x3xi8>, tensor<1280x1280x3x3xf16>) outs(%2492 : tensor<1280x1280x3x3xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<1280x1280x3x3xf16>
%padded_1304 = tensor.pad %2489 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
tensor.yield %cst_889 : f16
} : tensor<2x1280x16x16xf16> to tensor<2x1280x18x18xf16>
%2494 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_218 : tensor<1280xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1280x16x16xf16>
%2495 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1304, %2493 : tensor<2x1280x18x18xf16>, tensor<1280x1280x3x3xf16>) outs(%2494 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16>
%expanded_1305 = tensor.expand_shape %2361 [[0], [1, 2], [3], [4]] : tensor<2x640x16x16xf16> into tensor<2x40x16x16x16xf16>
%2496 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1305 : tensor<2x40x16x16x16xf16>) outs(%2392 : tensor<2x40x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x40x16x16x16xf16>
%2497 = linalg.fill ins(%cst_891 : f16) outs(%2394 : tensor<2x40x1x16x16xf16>) -> tensor<2x40x1x16x16xf16>
%2498 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%2496 : tensor<2x40x16x16x16xf16>) outs(%2497 : tensor<2x40x1x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x40x1x16x16xf16>
%2499 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%2498 : tensor<2x40x1x16x16xf16>) outs(%2392 : tensor<2x40x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x40x16x16x16xf16>
%collapsed_1306 = tensor.collapse_shape %2499 [[0], [1, 2], [3], [4]] : tensor<2x40x16x16x16xf16> into tensor<2x640x16x16xf16>
%2500 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2501 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2500 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2502 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2501 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2503 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2502 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2504 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1306, %2503 : tensor<2x640x16x16xf16>, tensor<f16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%2505 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2361, %2504 : tensor<2x640x16x16xf16>, tensor<2x640x16x16xf16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%2506 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2505, %cst_880 : tensor<2x640x16x16xf16>, tensor<f16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%2507 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2508 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2507 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2509 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2508 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2510 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2511 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2510 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2512 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%2511, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2513 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2506, %2512 : tensor<2x640x16x16xf16>, tensor<f16>) outs(%2411 : tensor<2x640x16x16xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x16x16xi1>
%2514 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2513, %2512, %2506 : tensor<2x640x16x16xi1>, tensor<f16>, tensor<2x640x16x16xf16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%2515 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2514, %2509 : tensor<2x640x16x16xf16>, tensor<f16>) outs(%2411 : tensor<2x640x16x16xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x640x16x16xi1>
%2516 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2515, %2509, %2514 : tensor<2x640x16x16xi1>, tensor<f16>, tensor<2x640x16x16xf16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%2517 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2516 : tensor<2x640x16x16xf16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%2518 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2517, %cst_880 : tensor<2x640x16x16xf16>, tensor<f16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%2519 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2518, %2504 : tensor<2x640x16x16xf16>, tensor<2x640x16x16xf16>) outs(%2359 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x640x16x16xf16>
%2520 = tensor.empty() : tensor<1280x40x16x1x1xf16>
%2521 = linalg.generic {indexing_maps = [#map17, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_219 : tensor<1280x40x1x1x1xf16>) outs(%2520 : tensor<1280x40x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x40x16x1x1xf16>
%2522 = linalg.generic {indexing_maps = [#map18, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%2521 : tensor<1280x40x16x1x1xf16>) outs(%2520 : tensor<1280x40x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x40x16x1x1xf16>
%collapsed_1307 = tensor.collapse_shape %2522 [[0], [1, 2], [3], [4]] : tensor<1280x40x16x1x1xf16> into tensor<1280x640x1x1xf16>
%2523 = tensor.empty() : tensor<1280x640x1x1xf16>
%2524 = linalg.generic {indexing_maps = [#map13, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_220, %collapsed_1307 : tensor<1280x640x1x1xi8>, tensor<1280x640x1x1xf16>) outs(%2523 : tensor<1280x640x1x1xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<1280x640x1x1xf16>
%2525 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_221 : tensor<1280xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1280x16x16xf16>
%2526 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2519, %2524 : tensor<2x640x16x16xf16>, tensor<1280x640x1x1xf16>) outs(%2525 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16>
%2527 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2526, %2495 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2528 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2527, %cst_3 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%collapsed_1308 = tensor.collapse_shape %2528 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16>
%expanded_1309 = tensor.expand_shape %collapsed_1308 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16>
%2529 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%2530 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2529 : tensor<f64>) outs(%9 : tensor<f32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<f32>
%2531 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2530 : tensor<f32>) outs(%2435 : tensor<2x32x40x256xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<2x32x40x256xf32>
%2532 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1309 : tensor<2x32x40x256xf16>) outs(%2531 : tensor<2x32x40x256xf32>) {
^bb0(%in: f16, %out: f32):
%10165 = arith.extf %in : f16 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x40x256xf32>
%2533 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%2534 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2532 : tensor<2x32x40x256xf32>) outs(%2533 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2535 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2534 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_900 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2536 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2532, %2535 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%2435 : tensor<2x32x40x256xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.subf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x40x256xf32>
%2537 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2536 : tensor<2x32x40x256xf32>) outs(%2435 : tensor<2x32x40x256xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.mulf %in, %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x40x256xf32>
%2538 = linalg.fill ins(%cst_894 : f64) outs(%55 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
%2539 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2537 : tensor<2x32x40x256xf32>) outs(%2538 : tensor<2x32x1x1xf64>) {
^bb0(%in: f32, %out: f64):
%10165 = arith.extf %in : f32 to f64
%10166 = arith.addf %10165, %out : f64
linalg.yield %10166 : f64
} -> tensor<2x32x1x1xf64>
%2540 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2539 : tensor<2x32x1x1xf64>) outs(%55 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%10165 = arith.divf %in, %cst_901 : f64
linalg.yield %10165 : f64
} -> tensor<2x32x1x1xf64>
%2541 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2540 : tensor<2x32x1x1xf64>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%10165 = arith.truncf %in : f64 to f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2542 = linalg.fill ins(%cst_896 : f32) outs(%49 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%2543 = linalg.generic {indexing_maps = [#map11, #map13], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2532 : tensor<2x32x40x256xf32>) outs(%2542 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.addf %in, %out : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2544 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2543 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = arith.divf %in, %cst_900 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2545 = linalg.generic {indexing_maps = [#map13, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2541, %cst_2 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2540: f16, %out: f32):
%10165 = arith.extf %in_2540 : f16 to f32
%10166 = arith.addf %in, %10165 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x1x1xf32>
%2546 = linalg.generic {indexing_maps = [#map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2545 : tensor<2x32x1x1xf32>) outs(%49 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%10165 = math.rsqrt %in : f32
linalg.yield %10165 : f32
} -> tensor<2x32x1x1xf32>
%2547 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1309, %2544 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%2435 : tensor<2x32x40x256xf32>) {
^bb0(%in: f16, %in_2540: f32, %out: f32):
%10165 = arith.extf %in : f16 to f32
%10166 = arith.subf %10165, %in_2540 : f32
linalg.yield %10166 : f32
} -> tensor<2x32x40x256xf32>
%2548 = linalg.generic {indexing_maps = [#map11, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2547, %2546 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%2435 : tensor<2x32x40x256xf32>) {
^bb0(%in: f32, %in_2540: f32, %out: f32):
%10165 = arith.mulf %in, %in_2540 : f32
linalg.yield %10165 : f32
} -> tensor<2x32x40x256xf32>
%collapsed_1310 = tensor.collapse_shape %2548 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32>
%expanded_1311 = tensor.expand_shape %collapsed_1310 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32>
%2549 = linalg.fill ins(%cst_894 : f64) outs(%7 : tensor<f64>) -> tensor<f64>
%2550 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2549 : tensor<f64>) outs(%23 : tensor<f16>) {
^bb0(%in: f64, %out: f16):
%10165 = arith.truncf %in : f64 to f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2551 = linalg.generic {indexing_maps = [#map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2550 : tensor<f16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1280x16x16xf16>
%2552 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1311 : tensor<2x1280x16x16xf32>) outs(%2551 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f32, %out: f16):
%10165 = arith.truncf %in : f32 to f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%expanded_1312 = tensor.expand_shape %2552 [[0], [1, 2], [3], [4]] : tensor<2x1280x16x16xf16> into tensor<2x80x16x16x16xf16>
%2553 = linalg.generic {indexing_maps = [#map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1312 : tensor<2x80x16x16x16xf16>) outs(%2463 : tensor<2x80x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x80x16x16x16xf16>
%2554 = linalg.fill ins(%cst_891 : f16) outs(%2465 : tensor<2x80x1x16x16xf16>) -> tensor<2x80x1x16x16xf16>
%2555 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%2553 : tensor<2x80x16x16x16xf16>) outs(%2554 : tensor<2x80x1x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x80x1x16x16xf16>
%2556 = linalg.generic {indexing_maps = [#map16, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%2555 : tensor<2x80x1x16x16xf16>) outs(%2463 : tensor<2x80x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x80x16x16x16xf16>
%collapsed_1313 = tensor.collapse_shape %2556 [[0], [1, 2], [3], [4]] : tensor<2x80x16x16x16xf16> into tensor<2x1280x16x16xf16>
%2557 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2558 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2557 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2559 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2558 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2560 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2559 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2561 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1313, %2560 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2562 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2552, %2561 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2563 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2562, %cst_880 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2564 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2565 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2564 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2566 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2565 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2567 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2568 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2567 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2569 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%2568, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2570 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2563, %2569 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%2482 : tensor<2x1280x16x16xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1280x16x16xi1>
%2571 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2570, %2569, %2563 : tensor<2x1280x16x16xi1>, tensor<f16>, tensor<2x1280x16x16xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2572 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2571, %2566 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%2482 : tensor<2x1280x16x16xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x1280x16x16xi1>
%2573 = linalg.generic {indexing_maps = [#map11, #map12, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2572, %2566, %2571 : tensor<2x1280x16x16xi1>, tensor<f16>, tensor<2x1280x16x16xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2574 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2573 : tensor<2x1280x16x16xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.round %in : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2575 = linalg.generic {indexing_maps = [#map11, #map12, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2574, %cst_880 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2576 = linalg.generic {indexing_maps = [#map11, #map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2575, %2561 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x1280x16x16xf16>
%2577 = tensor.empty() : tensor<1280x80x16x1x1xf16>
%2578 = linalg.generic {indexing_maps = [#map17, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_222 : tensor<1280x80x1x1x1xf16>) outs(%2577 : tensor<1280x80x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16x1x1xf16>
%2579 = linalg.generic {indexing_maps = [#map18, #map15], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%2578 : tensor<1280x80x16x1x1xf16>) outs(%2577 : tensor<1280x80x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16x1x1xf16>
%collapsed_1314 = tensor.collapse_shape %2579 [[0], [1, 2], [3], [4]] : tensor<1280x80x16x1x1xf16> into tensor<1280x1280x1x1xf16>
%2580 = tensor.empty() : tensor<1280x1280x1x1xf16>
%2581 = linalg.generic {indexing_maps = [#map13, #map13, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_223, %collapsed_1314 : tensor<1280x1280x1x1xi8>, tensor<1280x1280x1x1xf16>) outs(%2580 : tensor<1280x1280x1x1xf16>) {
^bb0(%in: i8, %in_2540: f16, %out: f16):
%10165 = arith.sitofp %in : i8 to f16
%10166 = arith.mulf %10165, %in_2540 : f16
linalg.yield %10166 : f16
} -> tensor<1280x1280x1x1xf16>
%2582 = linalg.generic {indexing_maps = [#map10, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_224 : tensor<1280xf16>) outs(%2423 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1280x16x16xf16>
%2583 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2576, %2581 : tensor<2x1280x16x16xf16>, tensor<1280x1280x1x1xf16>) outs(%2582 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16>
%2584 = tensor.empty() : tensor<2x16x16x1280xf16>
%2585 = linalg.generic {indexing_maps = [#map11, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2583 : tensor<2x1280x16x16xf16>) outs(%2584 : tensor<2x16x16x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x16x16x1280xf16>
%collapsed_1315 = tensor.collapse_shape %2585 [[0], [1, 2], [3]] : tensor<2x16x16x1280xf16> into tensor<2x256x1280xf16>
%2586 = tensor.empty() : tensor<2x256x1xf16>
%2587 = linalg.fill ins(%cst_889 : f16) outs(%2586 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16>
%2588 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_1315 : tensor<2x256x1280xf16>) outs(%2587 : tensor<2x256x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x256x1xf16>
%2589 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2588 : tensor<2x256x1xf16>) outs(%2586 : tensor<2x256x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_907 : f16
linalg.yield %10165 : f16
} -> tensor<2x256x1xf16>
%2590 = tensor.empty() : tensor<2x256x1280xf16>
%2591 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2589 : tensor<2x256x1xf16>) outs(%2590 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x1280xf16>
%2592 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1315, %2591 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%2590 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.subf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x256x1280xf16>
%2593 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2592 : tensor<2x256x1280xf16>) outs(%2590 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.mulf %in, %in : f16
linalg.yield %10165 : f16
} -> tensor<2x256x1280xf16>
%2594 = linalg.fill ins(%cst_889 : f16) outs(%2586 : tensor<2x256x1xf16>) -> tensor<2x256x1xf16>
%2595 = linalg.generic {indexing_maps = [#map20, #map21], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2593 : tensor<2x256x1280xf16>) outs(%2594 : tensor<2x256x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.addf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x256x1xf16>
%2596 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2595 : tensor<2x256x1xf16>) outs(%2586 : tensor<2x256x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.divf %in, %cst_907 : f16
linalg.yield %10165 : f16
} -> tensor<2x256x1xf16>
%2597 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2596 : tensor<2x256x1xf16>) outs(%2586 : tensor<2x256x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.truncf %cst_895 : f64 to f16
%10166 = arith.addf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<2x256x1xf16>
%2598 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2597 : tensor<2x256x1xf16>) outs(%2586 : tensor<2x256x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.rsqrt %in : f16
linalg.yield %10165 : f16
} -> tensor<2x256x1xf16>
%2599 = linalg.generic {indexing_maps = [#map21, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2598 : tensor<2x256x1xf16>) outs(%2590 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x1280xf16>
%2600 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2592, %2599 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%2590 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.mulf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x256x1280xf16>
%expanded_1316 = tensor.expand_shape %2600 [[0], [1], [2, 3]] : tensor<2x256x1280xf16> into tensor<2x256x80x16xf16>
%2601 = tensor.empty() : tensor<2x256x80x16xf16>
%2602 = linalg.generic {indexing_maps = [#map11, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1316 : tensor<2x256x80x16xf16>) outs(%2601 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = math.absf %in : f16
linalg.yield %10165 : f16
} -> tensor<2x256x80x16xf16>
%2603 = tensor.empty() : tensor<2x256x80x1xf16>
%2604 = linalg.fill ins(%cst_891 : f16) outs(%2603 : tensor<2x256x80x1xf16>) -> tensor<2x256x80x1xf16>
%2605 = linalg.generic {indexing_maps = [#map11, #map22], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2602 : tensor<2x256x80x16xf16>) outs(%2604 : tensor<2x256x80x1xf16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.maxf %in, %out : f16
linalg.yield %10165 : f16
} -> tensor<2x256x80x1xf16>
%2606 = linalg.generic {indexing_maps = [#map22, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2605 : tensor<2x256x80x1xf16>) outs(%2601 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x80x16xf16>
%collapsed_1317 = tensor.collapse_shape %2606 [[0], [1], [2, 3]] : tensor<2x256x80x16xf16> into tensor<2x256x1280xf16>
%2607 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2608 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2607 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2609 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2608 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2610 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2609 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2611 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1317, %2610 : tensor<2x256x1280xf16>, tensor<f16>) outs(%2590 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x256x1280xf16>
%2612 = linalg.generic {indexing_maps = [#map20, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2600, %2611 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%2590 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.divf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x256x1280xf16>
%2613 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2612, %cst_880 : tensor<2x256x1280xf16>, tensor<f16>) outs(%2590 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = arith.addf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<2x256x1280xf16>
%2614 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2615 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2614 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2616 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = []} ins(%2615 : tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %out: f16):
%10165 = arith.negf %in : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2617 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_879, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2618 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%cst_881, %2617 : tensor<f16>, tensor<f16>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: f16, %out: f16):
%10165 = math.powf %in, %in_2540 : f16
linalg.yield %10165 : f16
} -> tensor<f16>
%2619 = linalg.generic {indexing_maps = [#map3, #map3, #map3], iterator_types = []} ins(%2618, %cst_5 : tensor<f16>, tensor<i64>) outs(%23 : tensor<f16>) {
^bb0(%in: f16, %in_2540: i64, %out: f16):
%10165 = arith.sitofp %in_2540 : i64 to f16
%10166 = arith.subf %in, %10165 : f16
linalg.yield %10166 : f16
} -> tensor<f16>
%2620 = tensor.empty() : tensor<2x256x1280xi1>
%2621 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2613, %2619 : tensor<2x256x1280xf16>, tensor<f16>) outs(%2620 : tensor<2x256x1280xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ugt, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x256x1280xi1>
%2622 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2621, %2619, %2613 : tensor<2x256x1280xi1>, tensor<f16>, tensor<2x256x1280xf16>) outs(%2590 : tensor<2x256x1280xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x256x1280xf16>
%2623 = linalg.generic {indexing_maps = [#map20, #map23, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2622, %2616 : tensor<2x256x1280xf16>, tensor<f16>) outs(%2620 : tensor<2x256x1280xi1>) {
^bb0(%in: f16, %in_2540: f16, %out: i1):
%10165 = arith.cmpf ult, %in, %in_2540 : f16
linalg.yield %10165 : i1
} -> tensor<2x256x1280xi1>
%2624 = linalg.generic {indexing_maps = [#map20, #map23, #map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2623, %2616, %2622 : tensor<2x256x1280xi1>, tensor<f16>, tensor<2x256x1280xf16>) outs(%2590 : tensor<2x256x1280xf16>) {
^bb0(%in: i1, %in_2540: f16, %in_2541: f16, %out: f16):
%10165 = arith.select %in, %in_2540, %in_2541 : f16
linalg.yield %10165 : f16
} -> tensor<2x256x1280xf16>
%2625 = linalg.generic {indexing_maps = [#map20, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2624 : tensor<2x256x1280xf
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment