Skip to content

Instantly share code, notes, and snippets.

@pashu123

pashu123/int8_elided.mlir

Created March 20, 2023 15:58
Show Gist options
  • Save pashu123/31227fa1ffa06910610ce9db50d7834f to your computer and use it in GitHub Desktop.
Save pashu123/31227fa1ffa06910610ce9db50d7834f to your computer and use it in GitHub Desktop.
Download ZIP
Raw
int8_elided.mlir
This file has been truncated, but you can view the full file.
#map = affine_map<(d0, d1, d2, d3) -> (0, d1, d2, d3)>
#map1 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
#map2 = affine_map<(d0) -> (0)>
#map3 = affine_map<(d0) -> (d0)>
#map4 = affine_map<(d0) -> ()>
#map5 = affine_map<(d0, d1) -> (d0, 0)>
#map6 = affine_map<(d0, d1) -> (0, d1)>
#map7 = affine_map<(d0, d1) -> (d0, 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, 0, 0)>
#map12 = affine_map<(d0, d1, d2, d3) -> ()>
#map13 = affine_map<(d0, d1, d2, d3) -> (0, d1, 0, 0)>
#map14 = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, d3, d4)>
#map15 = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, 0, d3, d4)>
#map16 = affine_map<(d0, d1, d2, d3) -> (d0, d2, d3, d1)>
#map17 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, 0)>
#map18 = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
#map19 = affine_map<(d0, d1, d2) -> ()>
#map20 = affine_map<(d0, d1, d2) -> (d0, d1, 0)>
#map21 = affine_map<(d0, d1, d2) -> (d2)>
#map22 = affine_map<(d0, d1, d2, d3) -> (d0, d2, d1, d3)>
#map23 = affine_map<(d0, d1, d2) -> (d0, d2, d1)>
#map24 = affine_map<(d0, d1, d2, d3) -> (d0, d3, d1, d2)>
#map25 = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, 0, 0, 0)>
#map26 = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, 0, 0)>
module attributes {torch.debug_module_name = "_lambda"} {
ml_program.global private mutable @global_seed(dense<0> : tensor<i64>) : tensor<i64>
func.func @forward(%arg0: tensor<1x4x64x64xf16>, %arg1: tensor<1xf32>, %arg2: tensor<2x77x1024xf16>, %arg3: tensor<f32>) -> tensor<1x4x64x64xf16> {
%cst = arith.constant dense<[-2.388000e-03, 3.170010e-03, 4.332070e-04, 1.795770e-03]> : tensor<4xf16>
%cst_0 = arith.constant dense_resource<__elided__> : tensor<4x320x3x3xf16>
%cst_1 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_2 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_3 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_4 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_5 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_6 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_7 = arith.constant dense_resource<__elided__> : tensor<320x1280xi8>
%cst_8 = arith.constant dense_resource<__elided__> : tensor<320x80x1xf16>
%cst_9 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_10 = arith.constant dense_resource<__elided__> : tensor<2560x320xi8>
%cst_11 = arith.constant dense_resource<__elided__> : tensor<2560x20x1xf16>
%cst_12 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%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<320x320xi8>
%cst_16 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_17 = arith.constant dense_resource<__elided__> : tensor<320x1024xi8>
%cst_18 = arith.constant dense_resource<__elided__> : tensor<320x64x1xf16>
%cst_19 = arith.constant dense_resource<__elided__> : tensor<320x1024xi8>
%cst_20 = arith.constant dense_resource<__elided__> : tensor<320x64x1xf16>
%cst_21 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_22 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_23 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_24 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_25 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_26 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_27 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_28 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%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<320x20x1xf16>
%cst_32 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_33 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_34 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_35 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_36 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_37 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_38 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_39 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_40 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_41 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_42 = arith.constant dense_resource<__elided__> : tensor<320x640x1x1xi8>
%cst_43 = arith.constant dense_resource<__elided__> : tensor<320x40x1x1x1xf16>
%cst_44 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_45 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xi8>
%cst_46 = arith.constant dense_resource<__elided__> : tensor<320x20x1x3x3xf16>
%cst_47 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_48 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_49 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_50 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16>
%cst_51 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_52 = arith.constant dense_resource<__elided__> : tensor<320x640x3x3xi8>
%cst_53 = arith.constant dense_resource<__elided__> : tensor<320x40x1x3x3xf16>
%cst_54 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_55 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_56 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_57 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_58 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_59 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_60 = arith.constant dense_resource<__elided__> : tensor<320x1280xi8>
%cst_61 = arith.constant dense_resource<__elided__> : tensor<320x80x1xf16>
%cst_62 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_63 = arith.constant dense_resource<__elided__> : tensor<2560x320xi8>
%cst_64 = arith.constant dense_resource<__elided__> : tensor<2560x20x1xf16>
%cst_65 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%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<320x320xi8>
%cst_69 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_70 = arith.constant dense_resource<__elided__> : tensor<320x1024xi8>
%cst_71 = arith.constant dense_resource<__elided__> : tensor<320x64x1xf16>
%cst_72 = arith.constant dense_resource<__elided__> : tensor<320x1024xi8>
%cst_73 = arith.constant dense_resource<__elided__> : tensor<320x64x1xf16>
%cst_74 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_75 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_76 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_77 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_78 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_79 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_80 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_81 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_82 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_83 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%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<320x20x1xf16>
%cst_87 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_88 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_89 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_90 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_91 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_92 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_93 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_94 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_95 = arith.constant dense_resource<__elided__> : tensor<320x640x1x1xi8>
%cst_96 = arith.constant dense_resource<__elided__> : tensor<320x40x1x1x1xf16>
%cst_97 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_98 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xi8>
%cst_99 = arith.constant dense_resource<__elided__> : tensor<320x20x1x3x3xf16>
%cst_100 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_101 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_102 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_103 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16>
%cst_104 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_105 = arith.constant dense_resource<__elided__> : tensor<320x640x3x3xi8>
%cst_106 = arith.constant dense_resource<__elided__> : tensor<320x40x1x3x3xf16>
%cst_107 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_108 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_109 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_110 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_111 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_112 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_113 = arith.constant dense_resource<__elided__> : tensor<320x1280xi8>
%cst_114 = arith.constant dense_resource<__elided__> : tensor<320x80x1xf16>
%cst_115 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_116 = arith.constant dense_resource<__elided__> : tensor<2560x320xi8>
%cst_117 = arith.constant dense_resource<__elided__> : tensor<2560x20x1xf16>
%cst_118 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_119 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_120 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_121 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_122 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_123 = arith.constant dense_resource<__elided__> : tensor<320x1024xi8>
%cst_124 = arith.constant dense_resource<__elided__> : tensor<320x64x1xf16>
%cst_125 = arith.constant dense_resource<__elided__> : tensor<320x1024xi8>
%cst_126 = arith.constant dense_resource<__elided__> : tensor<320x64x1xf16>
%cst_127 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_128 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_129 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_130 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_131 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_132 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_133 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_134 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_135 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_136 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_137 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_138 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_139 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_140 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_141 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_142 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_143 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_144 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_145 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_146 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_147 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_148 = arith.constant dense_resource<__elided__> : tensor<320x960x1x1xi8>
%cst_149 = arith.constant dense_resource<__elided__> : tensor<320x60x1x1x1xf16>
%cst_150 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_151 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xi8>
%cst_152 = arith.constant dense_resource<__elided__> : tensor<320x20x1x3x3xf16>
%cst_153 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_154 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_155 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_156 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16>
%cst_157 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_158 = arith.constant dense_resource<__elided__> : tensor<320x960x3x3xi8>
%cst_159 = arith.constant dense_resource<__elided__> : tensor<320x60x1x3x3xf16>
%cst_160 = arith.constant dense_resource<__elided__> : tensor<960xf16>
%cst_161 = arith.constant dense_resource<__elided__> : tensor<960xf16>
%cst_162 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_163 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xi8>
%cst_164 = arith.constant dense_resource<__elided__> : tensor<640x40x1x3x3xf16>
%cst_165 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_166 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_167 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_168 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_169 = arith.constant dense_resource<__elided__> : tensor<640x2560xi8>
%cst_170 = arith.constant dense_resource<__elided__> : tensor<640x160x1xf16>
%cst_171 = arith.constant dense_resource<__elided__> : tensor<5120xf16>
%cst_172 = arith.constant dense_resource<__elided__> : tensor<5120x640xi8>
%cst_173 = arith.constant dense_resource<__elided__> : tensor<5120x40x1xf16>
%cst_174 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_175 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_176 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_177 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_178 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_179 = arith.constant dense_resource<__elided__> : tensor<640x1024xi8>
%cst_180 = arith.constant dense_resource<__elided__> : tensor<640x64x1xf16>
%cst_181 = arith.constant dense_resource<__elided__> : tensor<640x1024xi8>
%cst_182 = arith.constant dense_resource<__elided__> : tensor<640x64x1xf16>
%cst_183 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_184 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_185 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_186 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_187 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_188 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_189 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_190 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_191 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_192 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_193 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_194 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_195 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_196 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_197 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_198 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_199 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_200 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_201 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_202 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_203 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_204 = arith.constant dense_resource<__elided__> : tensor<640x960x1x1xi8>
%cst_205 = arith.constant dense_resource<__elided__> : tensor<640x60x1x1x1xf16>
%cst_206 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_207 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xi8>
%cst_208 = arith.constant dense_resource<__elided__> : tensor<640x40x1x3x3xf16>
%cst_209 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_210 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_211 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_212 = arith.constant dense_resource<__elided__> : tensor<640x1280xf16>
%cst_213 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_214 = arith.constant dense_resource<__elided__> : tensor<640x960x3x3xi8>
%cst_215 = arith.constant dense_resource<__elided__> : tensor<640x60x1x3x3xf16>
%cst_216 = arith.constant dense_resource<__elided__> : tensor<960xf16>
%cst_217 = arith.constant dense_resource<__elided__> : tensor<960xf16>
%cst_218 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_219 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_220 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_221 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_222 = arith.constant dense_resource<__elided__> : tensor<640x2560xi8>
%cst_223 = arith.constant dense_resource<__elided__> : tensor<640x160x1xf16>
%cst_224 = arith.constant dense_resource<__elided__> : tensor<5120xf16>
%cst_225 = arith.constant dense_resource<__elided__> : tensor<5120x640xi8>
%cst_226 = arith.constant dense_resource<__elided__> : tensor<5120x40x1xf16>
%cst_227 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_228 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_229 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_230 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_231 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_232 = arith.constant dense_resource<__elided__> : tensor<640x1024xi8>
%cst_233 = arith.constant dense_resource<__elided__> : tensor<640x64x1xf16>
%cst_234 = arith.constant dense_resource<__elided__> : tensor<640x1024xi8>
%cst_235 = arith.constant dense_resource<__elided__> : tensor<640x64x1xf16>
%cst_236 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_237 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_238 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_239 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_240 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_241 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_242 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_243 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_244 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_245 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_246 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_247 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_248 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_249 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_250 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_251 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_252 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_253 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_254 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_255 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_256 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_257 = arith.constant dense_resource<__elided__> : tensor<640x1280x1x1xi8>
%cst_258 = arith.constant dense_resource<__elided__> : tensor<640x80x1x1x1xf16>
%cst_259 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_260 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xi8>
%cst_261 = arith.constant dense_resource<__elided__> : tensor<640x40x1x3x3xf16>
%cst_262 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_263 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_264 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_265 = arith.constant dense_resource<__elided__> : tensor<640x1280xf16>
%cst_266 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_267 = arith.constant dense_resource<__elided__> : tensor<640x1280x3x3xi8>
%cst_268 = arith.constant dense_resource<__elided__> : tensor<640x80x1x3x3xf16>
%cst_269 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_270 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_271 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_272 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_273 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_274 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_275 = arith.constant dense_resource<__elided__> : tensor<640x2560xi8>
%cst_276 = arith.constant dense_resource<__elided__> : tensor<640x160x1xf16>
%cst_277 = arith.constant dense_resource<__elided__> : tensor<5120xf16>
%cst_278 = arith.constant dense_resource<__elided__> : tensor<5120x640xi8>
%cst_279 = arith.constant dense_resource<__elided__> : tensor<5120x40x1xf16>
%cst_280 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_281 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_282 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_283 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_284 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_285 = arith.constant dense_resource<__elided__> : tensor<640x1024xi8>
%cst_286 = arith.constant dense_resource<__elided__> : tensor<640x64x1xf16>
%cst_287 = arith.constant dense_resource<__elided__> : tensor<640x1024xi8>
%cst_288 = arith.constant dense_resource<__elided__> : tensor<640x64x1xf16>
%cst_289 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_290 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_291 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_292 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_293 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_294 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_295 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_296 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_297 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_298 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_299 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_300 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_301 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_302 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_303 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_304 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_305 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_306 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_307 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_308 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_309 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_310 = arith.constant dense_resource<__elided__> : tensor<640x1920x1x1xi8>
%cst_311 = arith.constant dense_resource<__elided__> : tensor<640x120x1x1x1xf16>
%cst_312 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_313 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xi8>
%cst_314 = arith.constant dense_resource<__elided__> : tensor<640x40x1x3x3xf16>
%cst_315 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_316 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_317 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_318 = arith.constant dense_resource<__elided__> : tensor<640x1280xf16>
%cst_319 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_320 = arith.constant dense_resource<__elided__> : tensor<640x1920x3x3xi8>
%cst_321 = arith.constant dense_resource<__elided__> : tensor<640x120x1x3x3xf16>
%cst_322 = arith.constant dense_resource<__elided__> : tensor<1920xf16>
%cst_323 = arith.constant dense_resource<__elided__> : tensor<1920xf16>
%cst_324 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_325 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_326 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_327 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_328 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_329 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_330 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_331 = arith.constant dense_resource<__elided__> : tensor<1280x5120xi8>
%cst_332 = arith.constant dense_resource<__elided__> : tensor<1280x320x1xf16>
%cst_333 = arith.constant dense_resource<__elided__> : tensor<10240xf16>
%cst_334 = arith.constant dense_resource<__elided__> : tensor<10240x1280xi8>
%cst_335 = arith.constant dense_resource<__elided__> : tensor<10240x80x1xf16>
%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<1280xf16>
%cst_339 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_340 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_341 = arith.constant dense_resource<__elided__> : tensor<1280x1024xi8>
%cst_342 = arith.constant dense_resource<__elided__> : tensor<1280x64x1xf16>
%cst_343 = arith.constant dense_resource<__elided__> : tensor<1280x1024xi8>
%cst_344 = arith.constant dense_resource<__elided__> : tensor<1280x64x1xf16>
%cst_345 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_346 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_347 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_348 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_349 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_350 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_351 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_352 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%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<1280x80x1xf16>
%cst_356 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_357 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_358 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_359 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_360 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_361 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_362 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_363 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_364 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_365 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_366 = arith.constant dense_resource<__elided__> : tensor<1280x1920x1x1xi8>
%cst_367 = arith.constant dense_resource<__elided__> : tensor<1280x120x1x1x1xf16>
%cst_368 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_369 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_370 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_371 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_372 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_373 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_374 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_375 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_376 = arith.constant dense_resource<__elided__> : tensor<1280x1920x3x3xi8>
%cst_377 = arith.constant dense_resource<__elided__> : tensor<1280x120x1x3x3xf16>
%cst_378 = arith.constant dense_resource<__elided__> : tensor<1920xf16>
%cst_379 = arith.constant dense_resource<__elided__> : tensor<1920xf16>
%cst_380 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_381 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_382 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_383 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_384 = arith.constant dense_resource<__elided__> : tensor<1280x5120xi8>
%cst_385 = arith.constant dense_resource<__elided__> : tensor<1280x320x1xf16>
%cst_386 = arith.constant dense_resource<__elided__> : tensor<10240xf16>
%cst_387 = arith.constant dense_resource<__elided__> : tensor<10240x1280xi8>
%cst_388 = arith.constant dense_resource<__elided__> : tensor<10240x80x1xf16>
%cst_389 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_390 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_391 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_392 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_393 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_394 = arith.constant dense_resource<__elided__> : tensor<1280x1024xi8>
%cst_395 = arith.constant dense_resource<__elided__> : tensor<1280x64x1xf16>
%cst_396 = arith.constant dense_resource<__elided__> : tensor<1280x1024xi8>
%cst_397 = arith.constant dense_resource<__elided__> : tensor<1280x64x1xf16>
%cst_398 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_399 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_400 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_401 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_402 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_403 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_404 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_405 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_406 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_407 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_408 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_409 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_410 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_411 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_412 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_413 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_414 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_415 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_416 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_417 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_418 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_419 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xi8>
%cst_420 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x1x1xf16>
%cst_421 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_422 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_423 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%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<1280xf16>
%cst_427 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_428 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_429 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xi8>
%cst_430 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x3x3xf16>
%cst_431 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_432 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_433 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_434 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_435 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_436 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_437 = arith.constant dense_resource<__elided__> : tensor<1280x5120xi8>
%cst_438 = arith.constant dense_resource<__elided__> : tensor<1280x320x1xf16>
%cst_439 = arith.constant dense_resource<__elided__> : tensor<10240xf16>
%cst_440 = arith.constant dense_resource<__elided__> : tensor<10240x1280xi8>
%cst_441 = arith.constant dense_resource<__elided__> : tensor<10240x80x1xf16>
%cst_442 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_443 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_444 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_445 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_446 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_447 = arith.constant dense_resource<__elided__> : tensor<1280x1024xi8>
%cst_448 = arith.constant dense_resource<__elided__> : tensor<1280x64x1xf16>
%cst_449 = arith.constant dense_resource<__elided__> : tensor<1280x1024xi8>
%cst_450 = arith.constant dense_resource<__elided__> : tensor<1280x64x1xf16>
%cst_451 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_452 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_453 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_454 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_455 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_456 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_457 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_458 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%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<1280x80x1xf16>
%cst_462 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_463 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_464 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_465 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_466 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_467 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_468 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_469 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_470 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_471 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_472 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xi8>
%cst_473 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x1x1xf16>
%cst_474 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_475 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_476 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_477 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_478 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_479 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_480 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_481 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_482 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xi8>
%cst_483 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x3x3xf16>
%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<1280xf16>
%cst_487 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_488 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_489 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_490 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xi8>
%cst_491 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x1x1xf16>
%cst_492 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_493 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_494 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_495 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_496 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_497 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_498 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_499 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_500 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xi8>
%cst_501 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x3x3xf16>
%cst_502 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_503 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_504 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_505 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xi8>
%cst_506 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x1x1xf16>
%cst_507 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_508 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_509 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_510 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_511 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_512 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_513 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_514 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_515 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xi8>
%cst_516 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x3x3xf16>
%cst_517 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_518 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_519 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_520 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xi8>
%cst_521 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x1x1xf16>
%cst_522 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_523 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_524 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_525 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_526 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_527 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_528 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_529 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_530 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xi8>
%cst_531 = arith.constant dense_resource<__elided__> : tensor<1280x160x1x3x3xf16>
%cst_532 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_533 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_534 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_535 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_536 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_537 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_538 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_539 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_540 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_541 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_542 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_543 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_544 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_545 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_546 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_547 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_548 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_549 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_550 = arith.constant dense_resource<__elided__> : tensor<1280x5120xi8>
%cst_551 = arith.constant dense_resource<__elided__> : tensor<1280x320x1xf16>
%cst_552 = arith.constant dense_resource<__elided__> : tensor<10240xf16>
%cst_553 = arith.constant dense_resource<__elided__> : tensor<10240x1280xi8>
%cst_554 = arith.constant dense_resource<__elided__> : tensor<10240x80x1xf16>
%cst_555 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_556 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_557 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_558 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_559 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_560 = arith.constant dense_resource<__elided__> : tensor<1280x1024xi8>
%cst_561 = arith.constant dense_resource<__elided__> : tensor<1280x64x1xf16>
%cst_562 = arith.constant dense_resource<__elided__> : tensor<1280x1024xi8>
%cst_563 = arith.constant dense_resource<__elided__> : tensor<1280x64x1xf16>
%cst_564 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_565 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_566 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_567 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_568 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_569 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%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<1280x80x1xf16>
%cst_573 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_574 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_575 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_576 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_577 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_578 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_579 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_580 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_581 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_582 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_583 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_584 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_585 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_586 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_587 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_588 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_589 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_590 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_591 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_592 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_593 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_594 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_595 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_596 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_597 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_598 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_599 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_600 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_601 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_602 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_603 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_604 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_605 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_606 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_607 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_608 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_609 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_610 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_611 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_612 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_613 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_614 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_615 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_616 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_617 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_618 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_619 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_620 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_621 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_622 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_623 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_624 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_625 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_626 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_627 = arith.constant dense_resource<__elided__> : tensor<1280x5120xi8>
%cst_628 = arith.constant dense_resource<__elided__> : tensor<1280x320x1xf16>
%cst_629 = arith.constant dense_resource<__elided__> : tensor<10240xf16>
%cst_630 = arith.constant dense_resource<__elided__> : tensor<10240x1280xi8>
%cst_631 = arith.constant dense_resource<__elided__> : tensor<10240x80x1xf16>
%cst_632 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_633 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_634 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_635 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_636 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_637 = arith.constant dense_resource<__elided__> : tensor<1280x1024xi8>
%cst_638 = arith.constant dense_resource<__elided__> : tensor<1280x64x1xf16>
%cst_639 = arith.constant dense_resource<__elided__> : tensor<1280x1024xi8>
%cst_640 = arith.constant dense_resource<__elided__> : tensor<1280x64x1xf16>
%cst_641 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_642 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_643 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_644 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_645 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_646 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_647 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_648 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_649 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_650 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_651 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_652 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_653 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_654 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_655 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_656 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_657 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_658 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_659 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_660 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_661 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_662 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_663 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_664 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_665 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_666 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_667 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_668 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_669 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_670 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_671 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_672 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_673 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_674 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_675 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_676 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_677 = arith.constant dense_resource<__elided__> : tensor<1280x5120xi8>
%cst_678 = arith.constant dense_resource<__elided__> : tensor<1280x320x1xf16>
%cst_679 = arith.constant dense_resource<__elided__> : tensor<10240xf16>
%cst_680 = arith.constant dense_resource<__elided__> : tensor<10240x1280xi8>
%cst_681 = arith.constant dense_resource<__elided__> : tensor<10240x80x1xf16>
%cst_682 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_683 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_684 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_685 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_686 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_687 = arith.constant dense_resource<__elided__> : tensor<1280x1024xi8>
%cst_688 = arith.constant dense_resource<__elided__> : tensor<1280x64x1xf16>
%cst_689 = arith.constant dense_resource<__elided__> : tensor<1280x1024xi8>
%cst_690 = arith.constant dense_resource<__elided__> : tensor<1280x64x1xf16>
%cst_691 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_692 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_693 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_694 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_695 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_696 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_697 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_698 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_699 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_700 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_701 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_702 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_703 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_704 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_705 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_706 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_707 = arith.constant dense_resource<__elided__> : tensor<1280x1280xi8>
%cst_708 = arith.constant dense_resource<__elided__> : tensor<1280x80x1xf16>
%cst_709 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_710 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_711 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_712 = arith.constant dense_resource<__elided__> : tensor<1280x640x1x1xi8>
%cst_713 = arith.constant dense_resource<__elided__> : tensor<1280x40x1x1x1xf16>
%cst_714 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_715 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xi8>
%cst_716 = arith.constant dense_resource<__elided__> : tensor<1280x80x1x3x3xf16>
%cst_717 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_718 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_719 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_720 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_721 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_722 = arith.constant dense_resource<__elided__> : tensor<1280x640x3x3xi8>
%cst_723 = arith.constant dense_resource<__elided__> : tensor<1280x40x1x3x3xf16>
%cst_724 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_725 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_726 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_727 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xi8>
%cst_728 = arith.constant dense_resource<__elided__> : tensor<640x40x1x3x3xf16>
%cst_729 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_730 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_731 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_732 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_733 = arith.constant dense_resource<__elided__> : tensor<640x2560xi8>
%cst_734 = arith.constant dense_resource<__elided__> : tensor<640x160x1xf16>
%cst_735 = arith.constant dense_resource<__elided__> : tensor<5120xf16>
%cst_736 = arith.constant dense_resource<__elided__> : tensor<5120x640xi8>
%cst_737 = arith.constant dense_resource<__elided__> : tensor<5120x40x1xf16>
%cst_738 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_739 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_740 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_741 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_742 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_743 = arith.constant dense_resource<__elided__> : tensor<640x1024xi8>
%cst_744 = arith.constant dense_resource<__elided__> : tensor<640x64x1xf16>
%cst_745 = arith.constant dense_resource<__elided__> : tensor<640x1024xi8>
%cst_746 = arith.constant dense_resource<__elided__> : tensor<640x64x1xf16>
%cst_747 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_748 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_749 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_750 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_751 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_752 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_753 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_754 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_755 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_756 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_757 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_758 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_759 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_760 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_761 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_762 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_763 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_764 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_765 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_766 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_767 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_768 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xi8>
%cst_769 = arith.constant dense_resource<__elided__> : tensor<640x40x1x3x3xf16>
%cst_770 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_771 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_772 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_773 = arith.constant dense_resource<__elided__> : tensor<640x1280xf16>
%cst_774 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_775 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xi8>
%cst_776 = arith.constant dense_resource<__elided__> : tensor<640x40x1x3x3xf16>
%cst_777 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_778 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_779 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_780 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_781 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_782 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_783 = arith.constant dense_resource<__elided__> : tensor<640x2560xi8>
%cst_784 = arith.constant dense_resource<__elided__> : tensor<640x160x1xf16>
%cst_785 = arith.constant dense_resource<__elided__> : tensor<5120xf16>
%cst_786 = arith.constant dense_resource<__elided__> : tensor<5120x640xi8>
%cst_787 = arith.constant dense_resource<__elided__> : tensor<5120x40x1xf16>
%cst_788 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_789 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_790 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_791 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_792 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_793 = arith.constant dense_resource<__elided__> : tensor<640x1024xi8>
%cst_794 = arith.constant dense_resource<__elided__> : tensor<640x64x1xf16>
%cst_795 = arith.constant dense_resource<__elided__> : tensor<640x1024xi8>
%cst_796 = arith.constant dense_resource<__elided__> : tensor<640x64x1xf16>
%cst_797 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_798 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_799 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_800 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_801 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_802 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_803 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_804 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_805 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_806 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_807 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_808 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_809 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_810 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_811 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_812 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_813 = arith.constant dense_resource<__elided__> : tensor<640x640xi8>
%cst_814 = arith.constant dense_resource<__elided__> : tensor<640x40x1xf16>
%cst_815 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_816 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_817 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_818 = arith.constant dense_resource<__elided__> : tensor<640x320x1x1xi8>
%cst_819 = arith.constant dense_resource<__elided__> : tensor<640x20x1x1x1xf16>
%cst_820 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_821 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xi8>
%cst_822 = arith.constant dense_resource<__elided__> : tensor<640x40x1x3x3xf16>
%cst_823 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_824 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_825 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_826 = arith.constant dense_resource<__elided__> : tensor<640x1280xf16>
%cst_827 = arith.constant dense_resource<__elided__> : tensor<640xf16>
%cst_828 = arith.constant dense_resource<__elided__> : tensor<640x320x3x3xi8>
%cst_829 = arith.constant dense_resource<__elided__> : tensor<640x20x1x3x3xf16>
%cst_830 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_831 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_832 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_833 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xi8>
%cst_834 = arith.constant dense_resource<__elided__> : tensor<320x20x1x3x3xf16>
%cst_835 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_836 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_837 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_838 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_839 = arith.constant dense_resource<__elided__> : tensor<320x1280xi8>
%cst_840 = arith.constant dense_resource<__elided__> : tensor<320x80x1xf16>
%cst_841 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_842 = arith.constant dense_resource<__elided__> : tensor<2560x320xi8>
%cst_843 = arith.constant dense_resource<__elided__> : tensor<2560x20x1xf16>
%cst_844 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_845 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_846 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_847 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_848 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_849 = arith.constant dense_resource<__elided__> : tensor<320x1024xi8>
%cst_850 = arith.constant dense_resource<__elided__> : tensor<320x64x1xf16>
%cst_851 = arith.constant dense_resource<__elided__> : tensor<320x1024xi8>
%cst_852 = arith.constant dense_resource<__elided__> : tensor<320x64x1xf16>
%cst_853 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_854 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_855 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_856 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_857 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_858 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_859 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_860 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_861 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_862 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%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<320x20x1xf16>
%cst_866 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_867 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_868 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_869 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_870 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_871 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_872 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_873 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_874 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xi8>
%cst_875 = arith.constant dense_resource<__elided__> : tensor<320x20x1x3x3xf16>
%cst_876 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_877 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_878 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_879 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16>
%cst_880 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_881 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xi8>
%cst_882 = arith.constant dense_resource<__elided__> : tensor<320x20x1x3x3xf16>
%cst_883 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%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<320x320xi8>
%cst_887 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_888 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_889 = arith.constant dense_resource<__elided__> : tensor<320x1280xi8>
%cst_890 = arith.constant dense_resource<__elided__> : tensor<320x80x1xf16>
%cst_891 = arith.constant dense_resource<__elided__> : tensor<2560xf16>
%cst_892 = arith.constant dense_resource<__elided__> : tensor<2560x320xi8>
%cst_893 = arith.constant dense_resource<__elided__> : tensor<2560x20x1xf16>
%cst_894 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_895 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_896 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_897 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_898 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_899 = arith.constant dense_resource<__elided__> : tensor<320x1024xi8>
%cst_900 = arith.constant dense_resource<__elided__> : tensor<320x64x1xf16>
%cst_901 = arith.constant dense_resource<__elided__> : tensor<320x1024xi8>
%cst_902 = arith.constant dense_resource<__elided__> : tensor<320x64x1xf16>
%cst_903 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_904 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_905 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_906 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_907 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_908 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_909 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_910 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_911 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_912 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_913 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_914 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_915 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_916 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_917 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_918 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_919 = arith.constant dense_resource<__elided__> : tensor<320x320xi8>
%cst_920 = arith.constant dense_resource<__elided__> : tensor<320x20x1xf16>
%cst_921 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_922 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_923 = arith.constant dense<1.013280e-06> : tensor<f16>
%cst_924 = arith.constant dense<1.000000e+00> : tensor<f16>
%cst_925 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_926 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xi8>
%cst_927 = arith.constant dense_resource<__elided__> : tensor<320x20x1x3x3xf16>
%cst_928 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_929 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_930 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_931 = arith.constant dense_resource<__elided__> : tensor<320x1280xf16>
%cst_932 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_933 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xi8>
%cst_934 = arith.constant dense_resource<__elided__> : tensor<320x20x1x3x3xf16>
%cst_935 = arith.constant dense<0.000000e+00> : tensor<f16>
%cst_936 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_937 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_938 = arith.constant dense<1.001360e-05> : tensor<f16>
%cst_939 = arith.constant dense_resource<__elided__> : tensor<320xf16>
%cst_940 = arith.constant dense_resource<__elided__> : tensor<320x4x3x3xf16>
%cst_941 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_942 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf16>
%cst_943 = arith.constant dense_resource<__elided__> : tensor<1280xf16>
%cst_944 = arith.constant dense_resource<__elided__> : tensor<1280x320xf16>
%cst_945 = arith.constant dense<-9.210930e+00> : tensor<f16>
%cst_946 = arith.constant 0.000000e+00 : f16
%cst_947 = arith.constant 1.000000e+00 : f16
%cst_948 = arith.constant -6.550400e+04 : f16
%cst_949 = arith.constant 2.000000e+00 : f16
%cst_950 = arith.constant 5.000000e-01 : f16
%cst_951 = arith.constant 0.000000e+00 : f64
%cst_952 = arith.constant 0.000000e+00 : f32
%cst_953 = arith.constant 1.600000e+02 : f32
%cst_954 = arith.constant 4.096000e+04 : f64
%cst_955 = arith.constant 4.096000e+04 : f32
%cst_956 = arith.constant 1.280000e+02 : f16
%cst_957 = arith.constant -1.280000e+02 : f16
%cst_958 = arith.constant 1.270000e+02 : f16
%cst_959 = arith.constant 3.200000e+02 : f64
%cst_960 = arith.constant 3.200000e+02 : f32
%cst_961 = arith.constant 1.250000e-01 : f16
%cst_962 = arith.constant 1.024000e+04 : f64
%cst_963 = arith.constant 1.024000e+04 : f32
%cst_964 = arith.constant 2.048000e+04 : f64
%cst_965 = arith.constant 2.048000e+04 : f32
%cst_966 = arith.constant 6.400000e+02 : f64
%cst_967 = arith.constant 6.400000e+02 : f32
%cst_968 = arith.constant 5.120000e+03 : f64
%cst_969 = arith.constant 5.120000e+03 : f32
%cst_970 = arith.constant 1.280000e+03 : f64
%cst_971 = arith.constant 1.280000e+03 : f32
%cst_972 = arith.constant 2.560000e+03 : f64
%cst_973 = arith.constant 2.560000e+03 : f32
%cst_974 = arith.constant 1.536000e+04 : f64
%cst_975 = arith.constant 1.536000e+04 : f32
%cst_976 = arith.constant 6.144000e+04 : f64
%cst_977 = arith.constant 6.144000e+04 : f32
%cst_978 = arith.constant 3.072000e+04 : f64
%cst_979 = arith.constant 3.072000e+04 : f32
%cst_980 = arith.constant 1.228800e+05 : f64
%cst_981 = arith.constant 1.228800e+05 : f32
%cst_982 = arith.constant 8.192000e+04 : f64
%cst_983 = arith.constant 8.192000e+04 : f32
%c0_i64 = arith.constant 0 : i64
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%c2 = arith.constant 2 : index
%0 = tensor.empty() : tensor<1x4x64x64xf16>
%1 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg0 : tensor<1x4x64x64xf16>) outs(%0 : tensor<1x4x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1x4x64x64xf16>
%2 = tensor.empty() : tensor<2x4x64x64xf16>
%inserted_slice = tensor.insert_slice %1 into %2[0, 0, 0, 0] [1, 4, 64, 64] [1, 1, 1, 1] : tensor<1x4x64x64xf16> into tensor<2x4x64x64xf16>
%inserted_slice_984 = tensor.insert_slice %1 into %inserted_slice[1, 0, 0, 0] [1, 4, 64, 64] [1, 1, 1, 1] : tensor<1x4x64x64xf16> into tensor<2x4x64x64xf16>
%3 = tensor.empty() : tensor<2xf32>
%4 = linalg.generic {indexing_maps = [#map2, #map3], iterator_types = ["parallel"]} ins(%arg1 : tensor<1xf32>) outs(%3 : tensor<2xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<2xf32>
%5 = tensor.empty() : tensor<160xf32>
%6 = linalg.generic {indexing_maps = [#map3], iterator_types = ["parallel"]} outs(%5 : tensor<160xf32>) {
^bb0(%out: f32):
%6234 = linalg.index 0 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.sitofp %6235 : i64 to f32
%6237 = arith.addf %6236, %cst_952 : f32
linalg.yield %6237 : f32
} -> tensor<160xf32>
%7 = linalg.generic {indexing_maps = [#map3, #map4, #map3], iterator_types = ["parallel"]} ins(%6, %cst_945 : tensor<160xf32>, tensor<f16>) outs(%5 : tensor<160xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<160xf32>
%8 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = ["parallel"]} ins(%7 : tensor<160xf32>) outs(%5 : tensor<160xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_953 : f32
linalg.yield %6234 : f32
} -> tensor<160xf32>
%9 = linalg.generic {indexing_maps = [#map3, #map3], iterator_types = ["parallel"]} ins(%8 : tensor<160xf32>) outs(%5 : tensor<160xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.exp %in : f32
linalg.yield %6234 : f32
} -> tensor<160xf32>
%expanded = tensor.expand_shape %4 [[0, 1]] : tensor<2xf32> into tensor<2x1xf32>
%expanded_985 = tensor.expand_shape %9 [[0, 1]] : tensor<160xf32> into tensor<1x160xf32>
%10 = tensor.empty() : tensor<2x160xf32>
%11 = linalg.generic {indexing_maps = [#map5, #map6, #map7], iterator_types = ["parallel", "parallel"]} ins(%expanded, %expanded_985 : tensor<2x1xf32>, tensor<1x160xf32>) outs(%10 : tensor<2x160xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x160xf32>
%12 = linalg.generic {indexing_maps = [#map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%11 : tensor<2x160xf32>) outs(%10 : tensor<2x160xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.sin %in : f32
linalg.yield %6234 : f32
} -> tensor<2x160xf32>
%13 = linalg.generic {indexing_maps = [#map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%11 : tensor<2x160xf32>) outs(%10 : tensor<2x160xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.cos %in : f32
linalg.yield %6234 : f32
} -> tensor<2x160xf32>
%14 = tensor.empty() : tensor<2x320xf32>
%inserted_slice_986 = tensor.insert_slice %12 into %14[0, 0] [2, 160] [1, 1] : tensor<2x160xf32> into tensor<2x320xf32>
%inserted_slice_987 = tensor.insert_slice %13 into %inserted_slice_986[0, 160] [2, 160] [1, 1] : tensor<2x160xf32> into tensor<2x320xf32>
%extracted_slice = tensor.extract_slice %inserted_slice_987[0, 0] [2, 160] [1, 1] : tensor<2x320xf32> to tensor<2x160xf32>
%inserted_slice_988 = tensor.insert_slice %13 into %14[0, 0] [2, 160] [1, 1] : tensor<2x160xf32> into tensor<2x320xf32>
%inserted_slice_989 = tensor.insert_slice %extracted_slice into %inserted_slice_988[0, 160] [2, 160] [1, 1] : tensor<2x160xf32> into tensor<2x320xf32>
%15 = tensor.empty() : tensor<2x320xf16>
%16 = linalg.generic {indexing_maps = [#map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%inserted_slice_989 : tensor<2x320xf32>) outs(%15 : tensor<2x320xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x320xf16>
%17 = tensor.empty() : tensor<320x1280xf16>
%18 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_944 : tensor<1280x320xf16>) outs(%17 : tensor<320x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x1280xf16>
%19 = tensor.empty() : tensor<2x1280xf16>
%20 = linalg.fill ins(%cst_946 : f16) outs(%19 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
%21 = linalg.matmul ins(%16, %18 : tensor<2x320xf16>, tensor<320x1280xf16>) outs(%20 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
%22 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_943, %21 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%19 : tensor<2x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280xf16>
%23 = linalg.generic {indexing_maps = [#map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%22 : tensor<2x1280xf16>) outs(%19 : tensor<2x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.negf %in : f16
%6235 = math.exp %6234 : f16
%6236 = arith.addf %6235, %cst_947 : f16
%6237 = arith.divf %cst_947, %6236 : f16
linalg.yield %6237 : f16
} -> tensor<2x1280xf16>
%24 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%23, %22 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%19 : tensor<2x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280xf16>
%25 = tensor.empty() : tensor<1280x1280xf16>
%26 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_942 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%27 = linalg.matmul ins(%24, %26 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%20 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
%28 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_941, %27 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%19 : tensor<2x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280xf16>
%padded = tensor.pad %inserted_slice_984 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x4x64x64xf16> to tensor<2x4x66x66xf16>
%29 = tensor.empty() : tensor<2x320x64x64xf16>
%30 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_939 : tensor<320xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%31 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded, %cst_940 : tensor<2x4x66x66xf16>, tensor<320x4x3x3xf16>) outs(%30 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16>
%collapsed = tensor.collapse_shape %31 [[0], [1], [2, 3]] : tensor<2x320x64x64xf16> into tensor<2x320x4096xf16>
%expanded_990 = tensor.expand_shape %collapsed [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16>
%32 = tensor.empty() : tensor<2x32x10x4096xf32>
%33 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_990 : tensor<2x32x10x4096xf16>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x10x4096xf32>
%34 = tensor.empty() : tensor<2x32x10x4096xf64>
%35 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%33 : tensor<2x32x10x4096xf32>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%36 = tensor.empty() : tensor<2x32x1x1xf64>
%37 = linalg.fill ins(%cst_951 : f64) outs(%36 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
%38 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%35 : tensor<2x32x10x4096xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%39 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%38 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_954 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%40 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%35, %39 : tensor<2x32x10x4096xf64>, tensor<2x32x1x1xf64>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%41 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%40, %40 : tensor<2x32x10x4096xf64>, tensor<2x32x10x4096xf64>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%42 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%41 : tensor<2x32x10x4096xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%43 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%42 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_954 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%44 = tensor.empty() : tensor<2x32x1x1xf32>
%45 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%43 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%46 = linalg.fill ins(%cst_952 : f32) outs(%44 : tensor<2x32x1x1xf32>) -> tensor<2x32x1x1xf32>
%47 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%33 : tensor<2x32x10x4096xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%48 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%47 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_955 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%49 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%45, %cst_938 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%50 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%49 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%51 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_990, %48 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x10x4096xf32>
%52 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%51, %50 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x10x4096xf32>
%collapsed_991 = tensor.collapse_shape %52 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32>
%expanded_992 = tensor.expand_shape %collapsed_991 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32>
%expanded_993 = tensor.expand_shape %cst_937 [[0, 1, 2, 3]] : tensor<320xf16> into tensor<1x320x1x1xf16>
%expanded_994 = tensor.expand_shape %cst_936 [[0, 1, 2, 3]] : tensor<320xf16> into tensor<1x320x1x1xf16>
%53 = tensor.empty() : tensor<2x320x64x64xf32>
%54 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_992, %expanded_994 : tensor<2x320x64x64xf32>, tensor<1x320x1x1xf16>) outs(%53 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x320x64x64xf32>
%55 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%54, %expanded_993 : tensor<2x320x64x64xf32>, tensor<1x320x1x1xf16>) outs(%53 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x320x64x64xf32>
%56 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%55 : tensor<2x320x64x64xf32>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%57 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%56 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.negf %in : f16
%6235 = math.exp %6234 : f16
%6236 = arith.addf %6235, %cst_947 : f16
%6237 = arith.divf %cst_947, %6236 : f16
linalg.yield %6237 : f16
} -> tensor<2x320x64x64xf16>
%58 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%57, %56 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%expanded_995 = tensor.expand_shape %58 [[0], [1, 2], [3], [4]] : tensor<2x320x64x64xf16> into tensor<2x20x16x64x64xf16>
%59 = tensor.empty() : tensor<2x20x16x64x64xf16>
%60 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_995 : tensor<2x20x16x64x64xf16>) outs(%59 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x20x16x64x64xf16>
%61 = tensor.empty() : tensor<2x20x1x64x64xi64>
%62 = linalg.fill ins(%c0_i64 : i64) outs(%61 : tensor<2x20x1x64x64xi64>) -> tensor<2x20x1x64x64xi64>
%63 = tensor.empty() : tensor<2x20x1x64x64xf16>
%64 = linalg.fill ins(%cst_948 : f16) outs(%63 : tensor<2x20x1x64x64xf16>) -> tensor<2x20x1x64x64xf16>
%65:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%60 : tensor<2x20x16x64x64xf16>) outs(%64, %62 : tensor<2x20x1x64x64xf16>, tensor<2x20x1x64x64xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x20x1x64x64xf16>, tensor<2x20x1x64x64xi64>)
%66 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%65#0 : tensor<2x20x1x64x64xf16>) outs(%59 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x64x64xf16>
%collapsed_996 = tensor.collapse_shape %66 [[0], [1, 2], [3], [4]] : tensor<2x20x16x64x64xf16> into tensor<2x320x64x64xf16>
%67 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_996 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%68 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%58, %67 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%69 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%68, %cst_935 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%70 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%69 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%71 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%70 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x320x64x64xf16>
%72 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%71, %cst_935 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%73 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%72, %67 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%74 = tensor.empty() : tensor<320x20x16x3x3xf16>
%75 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_934 : tensor<320x20x1x3x3xf16>) outs(%74 : tensor<320x20x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x3x3xf16>
%collapsed_997 = tensor.collapse_shape %75 [[0], [1, 2], [3], [4]] : tensor<320x20x16x3x3xf16> into tensor<320x320x3x3xf16>
%76 = tensor.empty() : tensor<320x320x3x3xf16>
%77 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_933, %collapsed_997 : tensor<320x320x3x3xi8>, tensor<320x320x3x3xf16>) outs(%76 : tensor<320x320x3x3xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320x3x3xf16>
%padded_998 = tensor.pad %73 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x320x64x64xf16> to tensor<2x320x66x66xf16>
%78 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_932 : tensor<320xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%79 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_998, %77 : tensor<2x320x66x66xf16>, tensor<320x320x3x3xf16>) outs(%78 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16>
%80 = linalg.generic {indexing_maps = [#map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%28 : tensor<2x1280xf16>) outs(%19 : tensor<2x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.negf %in : f16
%6235 = math.exp %6234 : f16
%6236 = arith.addf %6235, %cst_947 : f16
%6237 = arith.divf %cst_947, %6236 : f16
linalg.yield %6237 : f16
} -> tensor<2x1280xf16>
%81 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%80, %28 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%19 : tensor<2x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280xf16>
%82 = tensor.empty() : tensor<1280x320xf16>
%83 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_931 : tensor<320x1280xf16>) outs(%82 : tensor<1280x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x320xf16>
%84 = linalg.fill ins(%cst_946 : f16) outs(%15 : tensor<2x320xf16>) -> tensor<2x320xf16>
%85 = linalg.matmul ins(%81, %83 : tensor<2x1280xf16>, tensor<1280x320xf16>) outs(%84 : tensor<2x320xf16>) -> tensor<2x320xf16>
%86 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_930, %85 : tensor<320xf16>, tensor<2x320xf16>) outs(%15 : tensor<2x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320xf16>
%expanded_999 = tensor.expand_shape %86 [[0], [1, 2, 3]] : tensor<2x320xf16> into tensor<2x320x1x1xf16>
%87 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%79, %expanded_999 : tensor<2x320x64x64xf16>, tensor<2x320x1x1xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%collapsed_1000 = tensor.collapse_shape %87 [[0], [1], [2, 3]] : tensor<2x320x64x64xf16> into tensor<2x320x4096xf16>
%expanded_1001 = tensor.expand_shape %collapsed_1000 [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16>
%88 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1001 : tensor<2x32x10x4096xf16>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x10x4096xf32>
%89 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%88 : tensor<2x32x10x4096xf32>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%90 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%89 : tensor<2x32x10x4096xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%91 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%90 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_954 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%92 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%89, %91 : tensor<2x32x10x4096xf64>, tensor<2x32x1x1xf64>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%93 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%92, %92 : tensor<2x32x10x4096xf64>, tensor<2x32x10x4096xf64>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%94 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%93 : tensor<2x32x10x4096xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%95 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%94 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_954 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%96 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%95 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%97 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%88 : tensor<2x32x10x4096xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%98 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%97 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_955 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%99 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%96, %cst_938 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%100 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%99 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%101 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1001, %98 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x10x4096xf32>
%102 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%101, %100 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x10x4096xf32>
%collapsed_1002 = tensor.collapse_shape %102 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32>
%expanded_1003 = tensor.expand_shape %collapsed_1002 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32>
%expanded_1004 = tensor.expand_shape %cst_929 [[0, 1, 2, 3]] : tensor<320xf16> into tensor<1x320x1x1xf16>
%expanded_1005 = tensor.expand_shape %cst_928 [[0, 1, 2, 3]] : tensor<320xf16> into tensor<1x320x1x1xf16>
%103 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1003, %expanded_1005 : tensor<2x320x64x64xf32>, tensor<1x320x1x1xf16>) outs(%53 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x320x64x64xf32>
%104 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%103, %expanded_1004 : tensor<2x320x64x64xf32>, tensor<1x320x1x1xf16>) outs(%53 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x320x64x64xf32>
%105 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%104 : tensor<2x320x64x64xf32>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%106 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%105 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.negf %in : f16
%6235 = math.exp %6234 : f16
%6236 = arith.addf %6235, %cst_947 : f16
%6237 = arith.divf %cst_947, %6236 : f16
linalg.yield %6237 : f16
} -> tensor<2x320x64x64xf16>
%107 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%106, %105 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%expanded_1006 = tensor.expand_shape %107 [[0], [1, 2], [3], [4]] : tensor<2x320x64x64xf16> into tensor<2x20x16x64x64xf16>
%108 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1006 : tensor<2x20x16x64x64xf16>) outs(%59 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x20x16x64x64xf16>
%109:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%108 : tensor<2x20x16x64x64xf16>) outs(%64, %62 : tensor<2x20x1x64x64xf16>, tensor<2x20x1x64x64xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x20x1x64x64xf16>, tensor<2x20x1x64x64xi64>)
%110 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%109#0 : tensor<2x20x1x64x64xf16>) outs(%59 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x64x64xf16>
%collapsed_1007 = tensor.collapse_shape %110 [[0], [1, 2], [3], [4]] : tensor<2x20x16x64x64xf16> into tensor<2x320x64x64xf16>
%111 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1007 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%112 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%107, %111 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%113 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%112, %cst_935 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%114 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%113 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%115 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%114 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x320x64x64xf16>
%116 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%115, %cst_935 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%117 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%116, %111 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%118 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_927 : tensor<320x20x1x3x3xf16>) outs(%74 : tensor<320x20x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x3x3xf16>
%collapsed_1008 = tensor.collapse_shape %118 [[0], [1, 2], [3], [4]] : tensor<320x20x16x3x3xf16> into tensor<320x320x3x3xf16>
%119 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_926, %collapsed_1008 : tensor<320x320x3x3xi8>, tensor<320x320x3x3xf16>) outs(%76 : tensor<320x320x3x3xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320x3x3xf16>
%padded_1009 = tensor.pad %117 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x320x64x64xf16> to tensor<2x320x66x66xf16>
%120 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_925 : tensor<320xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%121 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1009, %119 : tensor<2x320x66x66xf16>, tensor<320x320x3x3xf16>) outs(%120 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16>
%122 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%31, %121 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%123 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%122, %cst_924 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%collapsed_1010 = tensor.collapse_shape %123 [[0], [1], [2, 3]] : tensor<2x320x64x64xf16> into tensor<2x320x4096xf16>
%expanded_1011 = tensor.expand_shape %collapsed_1010 [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16>
%124 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1011 : tensor<2x32x10x4096xf16>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x10x4096xf32>
%125 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%124 : tensor<2x32x10x4096xf32>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%126 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%125 : tensor<2x32x10x4096xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%127 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%126 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_954 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%128 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%125, %127 : tensor<2x32x10x4096xf64>, tensor<2x32x1x1xf64>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%129 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%128, %128 : tensor<2x32x10x4096xf64>, tensor<2x32x10x4096xf64>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%130 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%129 : tensor<2x32x10x4096xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%131 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%130 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_954 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%132 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%131 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%133 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%124 : tensor<2x32x10x4096xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%134 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%133 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_955 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%135 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%132, %cst_923 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%136 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%135 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%137 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1011, %134 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x10x4096xf32>
%138 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%137, %136 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x10x4096xf32>
%collapsed_1012 = tensor.collapse_shape %138 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32>
%expanded_1013 = tensor.expand_shape %collapsed_1012 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32>
%expanded_1014 = tensor.expand_shape %cst_922 [[0, 1, 2, 3]] : tensor<320xf16> into tensor<1x320x1x1xf16>
%expanded_1015 = tensor.expand_shape %cst_921 [[0, 1, 2, 3]] : tensor<320xf16> into tensor<1x320x1x1xf16>
%139 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1013, %expanded_1015 : tensor<2x320x64x64xf32>, tensor<1x320x1x1xf16>) outs(%53 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x320x64x64xf32>
%140 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%139, %expanded_1014 : tensor<2x320x64x64xf32>, tensor<1x320x1x1xf16>) outs(%53 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x320x64x64xf32>
%141 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%140 : tensor<2x320x64x64xf32>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%142 = tensor.empty() : tensor<2x64x64x320xf16>
%143 = linalg.generic {indexing_maps = [#map1, #map16], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%141 : tensor<2x320x64x64xf16>) outs(%142 : tensor<2x64x64x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x64x64x320xf16>
%collapsed_1016 = tensor.collapse_shape %143 [[0], [1, 2], [3]] : tensor<2x64x64x320xf16> into tensor<2x4096x320xf16>
%expanded_1017 = tensor.expand_shape %collapsed_1016 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%144 = tensor.empty() : tensor<2x4096x20x16xf16>
%145 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1017 : tensor<2x4096x20x16xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x20x16xf16>
%146 = tensor.empty() : tensor<2x4096x20x1xi64>
%147 = linalg.fill ins(%c0_i64 : i64) outs(%146 : tensor<2x4096x20x1xi64>) -> tensor<2x4096x20x1xi64>
%148 = tensor.empty() : tensor<2x4096x20x1xf16>
%149 = linalg.fill ins(%cst_948 : f16) outs(%148 : tensor<2x4096x20x1xf16>) -> tensor<2x4096x20x1xf16>
%150:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%145 : tensor<2x4096x20x16xf16>) outs(%149, %147 : tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>)
%151 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%150#0 : tensor<2x4096x20x1xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1018 = tensor.collapse_shape %151 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%152 = tensor.empty() : tensor<2x4096x320xf16>
%153 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1018 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%154 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1016, %153 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%155 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%154, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%156 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%155 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%157 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%156 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x4096x320xf16>
%158 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%157, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%159 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%158, %153 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%160 = tensor.empty() : tensor<320x20x16xf16>
%161 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_920 : tensor<320x20x1xf16>) outs(%160 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1019 = tensor.collapse_shape %161 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%162 = tensor.empty() : tensor<320x320xf16>
%163 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_919, %collapsed_1019 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320xf16>
%164 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%163 : tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1020 = tensor.collapse_shape %159 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%165 = tensor.empty() : tensor<8192x320xf16>
%166 = linalg.fill ins(%cst_946 : f16) outs(%165 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%167 = linalg.matmul ins(%collapsed_1020, %164 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%168 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_918, %167 : tensor<320xf16>, tensor<8192x320xf16>) outs(%165 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<8192x320xf16>
%expanded_1021 = tensor.expand_shape %168 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16>
%169 = tensor.empty() : tensor<2x4096x320xf32>
%170 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1021 : tensor<2x4096x320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x4096x320xf32>
%171 = tensor.empty() : tensor<2x4096x320xf64>
%172 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%170 : tensor<2x4096x320xf32>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%173 = tensor.empty() : tensor<2x4096x1xf64>
%174 = linalg.fill ins(%cst_951 : f64) outs(%173 : tensor<2x4096x1xf64>) -> tensor<2x4096x1xf64>
%175 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%172 : tensor<2x4096x320xf64>) outs(%174 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%176 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%175 : tensor<2x4096x1xf64>) outs(%173 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_959 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%177 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%172, %176 : tensor<2x4096x320xf64>, tensor<2x4096x1xf64>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%178 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%177, %177 : tensor<2x4096x320xf64>, tensor<2x4096x320xf64>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%179 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%178 : tensor<2x4096x320xf64>) outs(%174 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%180 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%179 : tensor<2x4096x1xf64>) outs(%173 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_959 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%181 = tensor.empty() : tensor<2x4096x1xf32>
%182 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%180 : tensor<2x4096x1xf64>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%183 = linalg.fill ins(%cst_952 : f32) outs(%181 : tensor<2x4096x1xf32>) -> tensor<2x4096x1xf32>
%184 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%170 : tensor<2x4096x320xf32>) outs(%183 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%185 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%184 : tensor<2x4096x1xf32>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_960 : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%186 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%182, %cst_938 : tensor<2x4096x1xf32>, tensor<f16>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x1xf32>
%187 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%186 : tensor<2x4096x1xf32>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%188 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1021, %185 : tensor<2x4096x320xf16>, tensor<2x4096x1xf32>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%189 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%188, %187 : tensor<2x4096x320xf32>, tensor<2x4096x1xf32>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x320xf32>
%190 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%189, %cst_917 : tensor<2x4096x320xf32>, tensor<320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%191 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%190, %cst_916 : tensor<2x4096x320xf32>, tensor<320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%192 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%191 : tensor<2x4096x320xf32>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%expanded_1022 = tensor.expand_shape %192 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%193 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1022 : tensor<2x4096x20x16xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x20x16xf16>
%194:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%193 : tensor<2x4096x20x16xf16>) outs(%149, %147 : tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>)
%195 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%194#0 : tensor<2x4096x20x1xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1023 = tensor.collapse_shape %195 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%196 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1023 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%197 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%192, %196 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%198 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%197, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%199 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%198 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%200 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%199 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x4096x320xf16>
%201 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%200, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%202 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%201, %196 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%203 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_915 : tensor<320x20x1xf16>) outs(%160 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1024 = tensor.collapse_shape %203 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%204 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_914, %collapsed_1024 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320xf16>
%205 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%204 : tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1025 = tensor.collapse_shape %202 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%206 = linalg.matmul ins(%collapsed_1025, %205 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%expanded_1026 = tensor.expand_shape %206 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x5x64xf16>
%207 = tensor.empty() : tensor<2x5x4096x64xf16>
%208 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1026 : tensor<2x4096x5x64xf16>) outs(%207 : tensor<2x5x4096x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x5x4096x64xf16>
%collapsed_1027 = tensor.collapse_shape %208 [[0, 1], [2], [3]] : tensor<2x5x4096x64xf16> into tensor<10x4096x64xf16>
%209 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_913 : tensor<320x20x1xf16>) outs(%160 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1028 = tensor.collapse_shape %209 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%210 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_912, %collapsed_1028 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320xf16>
%211 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%210 : tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%212 = linalg.matmul ins(%collapsed_1025, %211 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%213 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_911 : tensor<320x20x1xf16>) outs(%160 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1029 = tensor.collapse_shape %213 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%214 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_910, %collapsed_1029 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320xf16>
%215 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%214 : tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%216 = linalg.matmul ins(%collapsed_1025, %215 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%expanded_1030 = tensor.expand_shape %212 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x5x64xf16>
%217 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1030 : tensor<2x4096x5x64xf16>) outs(%207 : tensor<2x5x4096x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x5x4096x64xf16>
%collapsed_1031 = tensor.collapse_shape %217 [[0, 1], [2], [3]] : tensor<2x5x4096x64xf16> into tensor<10x4096x64xf16>
%expanded_1032 = tensor.expand_shape %216 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x5x64xf16>
%218 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1032 : tensor<2x4096x5x64xf16>) outs(%207 : tensor<2x5x4096x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x5x4096x64xf16>
%collapsed_1033 = tensor.collapse_shape %218 [[0, 1], [2], [3]] : tensor<2x5x4096x64xf16> into tensor<10x4096x64xf16>
%219 = tensor.empty() : tensor<10x4096x4096xf16>
%220 = linalg.fill ins(%cst_946 : f16) outs(%219 : tensor<10x4096x4096xf16>) -> tensor<10x4096x4096xf16>
%221 = tensor.empty() : tensor<10x64x4096xf16>
%222 = linalg.generic {indexing_maps = [#map18, #map23], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1031 : tensor<10x4096x64xf16>) outs(%221 : tensor<10x64x4096xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<10x64x4096xf16>
%223 = linalg.batch_matmul ins(%collapsed_1027, %222 : tensor<10x4096x64xf16>, tensor<10x64x4096xf16>) outs(%220 : tensor<10x4096x4096xf16>) -> tensor<10x4096x4096xf16>
%224 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%223 : tensor<10x4096x4096xf16>) outs(%219 : tensor<10x4096x4096xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.mulf %in, %cst_961 : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x4096xf16>
%225 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%224, %220 : tensor<10x4096x4096xf16>, tensor<10x4096x4096xf16>) outs(%219 : tensor<10x4096x4096xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in_2625, %cst_946 : f16
%6235 = arith.addf %in, %6234 : f16
linalg.yield %6235 : f16
} -> tensor<10x4096x4096xf16>
%226 = tensor.empty() : tensor<10x4096x1xi64>
%227 = linalg.fill ins(%c0_i64 : i64) outs(%226 : tensor<10x4096x1xi64>) -> tensor<10x4096x1xi64>
%228 = tensor.empty() : tensor<10x4096x1xf16>
%229 = linalg.fill ins(%cst_948 : f16) outs(%228 : tensor<10x4096x1xf16>) -> tensor<10x4096x1xf16>
%230:2 = linalg.generic {indexing_maps = [#map18, #map20, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%225 : tensor<10x4096x4096xf16>) outs(%229, %227 : tensor<10x4096x1xf16>, tensor<10x4096x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<10x4096x1xf16>, tensor<10x4096x1xi64>)
%231 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%225, %230#0 : tensor<10x4096x4096xf16>, tensor<10x4096x1xf16>) outs(%219 : tensor<10x4096x4096xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x4096xf16>
%232 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%231 : tensor<10x4096x4096xf16>) outs(%219 : tensor<10x4096x4096xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.exp %in : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x4096xf16>
%233 = linalg.fill ins(%cst_946 : f16) outs(%228 : tensor<10x4096x1xf16>) -> tensor<10x4096x1xf16>
%234 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%232 : tensor<10x4096x4096xf16>) outs(%233 : tensor<10x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.addf %in, %out : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x1xf16>
%235 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%232, %234 : tensor<10x4096x4096xf16>, tensor<10x4096x1xf16>) outs(%219 : tensor<10x4096x4096xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x4096xf16>
%236 = tensor.empty() : tensor<10x4096x64xf16>
%237 = linalg.fill ins(%cst_946 : f16) outs(%236 : tensor<10x4096x64xf16>) -> tensor<10x4096x64xf16>
%238 = linalg.batch_matmul ins(%235, %collapsed_1033 : tensor<10x4096x4096xf16>, tensor<10x4096x64xf16>) outs(%237 : tensor<10x4096x64xf16>) -> tensor<10x4096x64xf16>
%expanded_1034 = tensor.expand_shape %238 [[0, 1], [2], [3]] : tensor<10x4096x64xf16> into tensor<2x5x4096x64xf16>
%239 = tensor.empty() : tensor<2x4096x5x64xf16>
%240 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1034 : tensor<2x5x4096x64xf16>) outs(%239 : tensor<2x4096x5x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x5x64xf16>
%collapsed_1035 = tensor.collapse_shape %240 [[0], [1], [2, 3]] : tensor<2x4096x5x64xf16> into tensor<2x4096x320xf16>
%expanded_1036 = tensor.expand_shape %collapsed_1035 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%241 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1036 : tensor<2x4096x20x16xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x20x16xf16>
%242:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%241 : tensor<2x4096x20x16xf16>) outs(%149, %147 : tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>)
%243 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%242#0 : tensor<2x4096x20x1xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1037 = tensor.collapse_shape %243 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%244 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1037 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%245 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1035, %244 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%246 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%245, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%247 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%246 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%248 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%247 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x4096x320xf16>
%249 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%248, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%250 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%249, %244 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%251 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_909 : tensor<320x20x1xf16>) outs(%160 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1038 = tensor.collapse_shape %251 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%252 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_908, %collapsed_1038 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320xf16>
%253 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%252 : tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1039 = tensor.collapse_shape %250 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%254 = linalg.matmul ins(%collapsed_1039, %253 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%255 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_907, %254 : tensor<320xf16>, tensor<8192x320xf16>) outs(%165 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<8192x320xf16>
%expanded_1040 = tensor.expand_shape %255 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16>
%256 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1040, %expanded_1021 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%257 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%256 : tensor<2x4096x320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x4096x320xf32>
%258 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%257 : tensor<2x4096x320xf32>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%259 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%258 : tensor<2x4096x320xf64>) outs(%174 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%260 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%259 : tensor<2x4096x1xf64>) outs(%173 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_959 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%261 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%258, %260 : tensor<2x4096x320xf64>, tensor<2x4096x1xf64>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%262 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%261, %261 : tensor<2x4096x320xf64>, tensor<2x4096x320xf64>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%263 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%262 : tensor<2x4096x320xf64>) outs(%174 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%264 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%263 : tensor<2x4096x1xf64>) outs(%173 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_959 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%265 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%264 : tensor<2x4096x1xf64>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%266 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%257 : tensor<2x4096x320xf32>) outs(%183 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%267 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%266 : tensor<2x4096x1xf32>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_960 : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%268 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%265, %cst_938 : tensor<2x4096x1xf32>, tensor<f16>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x1xf32>
%269 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%268 : tensor<2x4096x1xf32>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%270 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%256, %267 : tensor<2x4096x320xf16>, tensor<2x4096x1xf32>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%271 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%270, %269 : tensor<2x4096x320xf32>, tensor<2x4096x1xf32>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x320xf32>
%272 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%271, %cst_906 : tensor<2x4096x320xf32>, tensor<320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%273 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%272, %cst_905 : tensor<2x4096x320xf32>, tensor<320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%274 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%273 : tensor<2x4096x320xf32>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%expanded_1041 = tensor.expand_shape %274 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%275 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1041 : tensor<2x4096x20x16xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x20x16xf16>
%276:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%275 : tensor<2x4096x20x16xf16>) outs(%149, %147 : tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>)
%277 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%276#0 : tensor<2x4096x20x1xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1042 = tensor.collapse_shape %277 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%278 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1042 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%279 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%274, %278 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%280 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%279, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%281 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%280 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%282 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%281 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x4096x320xf16>
%283 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%282, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%284 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%283, %278 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%285 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_904 : tensor<320x20x1xf16>) outs(%160 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1043 = tensor.collapse_shape %285 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%286 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_903, %collapsed_1043 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320xf16>
%287 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%286 : tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1044 = tensor.collapse_shape %284 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%288 = linalg.matmul ins(%collapsed_1044, %287 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%expanded_1045 = tensor.expand_shape %288 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x5x64xf16>
%289 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1045 : tensor<2x4096x5x64xf16>) outs(%207 : tensor<2x5x4096x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x5x4096x64xf16>
%collapsed_1046 = tensor.collapse_shape %289 [[0, 1], [2], [3]] : tensor<2x5x4096x64xf16> into tensor<10x4096x64xf16>
%expanded_1047 = tensor.expand_shape %arg2 [[0], [1], [2, 3]] : tensor<2x77x1024xf16> into tensor<2x77x64x16xf16>
%290 = tensor.empty() : tensor<2x77x64x16xf16>
%291 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1047 : tensor<2x77x64x16xf16>) outs(%290 : tensor<2x77x64x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x77x64x16xf16>
%292 = tensor.empty() : tensor<2x77x64x1xi64>
%293 = linalg.fill ins(%c0_i64 : i64) outs(%292 : tensor<2x77x64x1xi64>) -> tensor<2x77x64x1xi64>
%294 = tensor.empty() : tensor<2x77x64x1xf16>
%295 = linalg.fill ins(%cst_948 : f16) outs(%294 : tensor<2x77x64x1xf16>) -> tensor<2x77x64x1xf16>
%296:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%291 : tensor<2x77x64x16xf16>) outs(%295, %293 : tensor<2x77x64x1xf16>, tensor<2x77x64x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x77x64x1xf16>, tensor<2x77x64x1xi64>)
%297 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%296#0 : tensor<2x77x64x1xf16>) outs(%290 : tensor<2x77x64x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x77x64x16xf16>
%collapsed_1048 = tensor.collapse_shape %297 [[0], [1], [2, 3]] : tensor<2x77x64x16xf16> into tensor<2x77x1024xf16>
%298 = tensor.empty() : tensor<2x77x1024xf16>
%299 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1048 : tensor<2x77x1024xf16>) outs(%298 : tensor<2x77x1024xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x77x1024xf16>
%300 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2, %299 : tensor<2x77x1024xf16>, tensor<2x77x1024xf16>) outs(%298 : tensor<2x77x1024xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x77x1024xf16>
%301 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%300, %cst_935 : tensor<2x77x1024xf16>, tensor<f16>) outs(%298 : tensor<2x77x1024xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x77x1024xf16>
%302 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%301 : tensor<2x77x1024xf16>) outs(%298 : tensor<2x77x1024xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x77x1024xf16>
%303 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%302 : tensor<2x77x1024xf16>) outs(%298 : tensor<2x77x1024xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x77x1024xf16>
%304 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%303, %cst_935 : tensor<2x77x1024xf16>, tensor<f16>) outs(%298 : tensor<2x77x1024xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x77x1024xf16>
%305 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%304, %299 : tensor<2x77x1024xf16>, tensor<2x77x1024xf16>) outs(%298 : tensor<2x77x1024xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x77x1024xf16>
%306 = tensor.empty() : tensor<320x64x16xf16>
%307 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_902 : tensor<320x64x1xf16>) outs(%306 : tensor<320x64x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x64x16xf16>
%collapsed_1049 = tensor.collapse_shape %307 [[0], [1, 2]] : tensor<320x64x16xf16> into tensor<320x1024xf16>
%308 = tensor.empty() : tensor<320x1024xf16>
%309 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_901, %collapsed_1049 : tensor<320x1024xi8>, tensor<320x1024xf16>) outs(%308 : tensor<320x1024xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x1024xf16>
%310 = tensor.empty() : tensor<1024x320xf16>
%311 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%309 : tensor<320x1024xf16>) outs(%310 : tensor<1024x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1024x320xf16>
%collapsed_1050 = tensor.collapse_shape %305 [[0, 1], [2]] : tensor<2x77x1024xf16> into tensor<154x1024xf16>
%312 = tensor.empty() : tensor<154x320xf16>
%313 = linalg.fill ins(%cst_946 : f16) outs(%312 : tensor<154x320xf16>) -> tensor<154x320xf16>
%314 = linalg.matmul ins(%collapsed_1050, %311 : tensor<154x1024xf16>, tensor<1024x320xf16>) outs(%313 : tensor<154x320xf16>) -> tensor<154x320xf16>
%315 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_900 : tensor<320x64x1xf16>) outs(%306 : tensor<320x64x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x64x16xf16>
%collapsed_1051 = tensor.collapse_shape %315 [[0], [1, 2]] : tensor<320x64x16xf16> into tensor<320x1024xf16>
%316 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_899, %collapsed_1051 : tensor<320x1024xi8>, tensor<320x1024xf16>) outs(%308 : tensor<320x1024xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x1024xf16>
%317 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%316 : tensor<320x1024xf16>) outs(%310 : tensor<1024x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1024x320xf16>
%318 = linalg.matmul ins(%collapsed_1050, %317 : tensor<154x1024xf16>, tensor<1024x320xf16>) outs(%313 : tensor<154x320xf16>) -> tensor<154x320xf16>
%expanded_1052 = tensor.expand_shape %314 [[0, 1], [2, 3]] : tensor<154x320xf16> into tensor<2x77x5x64xf16>
%319 = tensor.empty() : tensor<2x5x77x64xf16>
%320 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1052 : tensor<2x77x5x64xf16>) outs(%319 : tensor<2x5x77x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x5x77x64xf16>
%collapsed_1053 = tensor.collapse_shape %320 [[0, 1], [2], [3]] : tensor<2x5x77x64xf16> into tensor<10x77x64xf16>
%expanded_1054 = tensor.expand_shape %318 [[0, 1], [2, 3]] : tensor<154x320xf16> into tensor<2x77x5x64xf16>
%321 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1054 : tensor<2x77x5x64xf16>) outs(%319 : tensor<2x5x77x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x5x77x64xf16>
%collapsed_1055 = tensor.collapse_shape %321 [[0, 1], [2], [3]] : tensor<2x5x77x64xf16> into tensor<10x77x64xf16>
%322 = tensor.empty() : tensor<10x4096x77xf16>
%323 = linalg.fill ins(%cst_946 : f16) outs(%322 : tensor<10x4096x77xf16>) -> tensor<10x4096x77xf16>
%324 = tensor.empty() : tensor<10x64x77xf16>
%325 = linalg.generic {indexing_maps = [#map18, #map23], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1053 : tensor<10x77x64xf16>) outs(%324 : tensor<10x64x77xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<10x64x77xf16>
%326 = linalg.batch_matmul ins(%collapsed_1046, %325 : tensor<10x4096x64xf16>, tensor<10x64x77xf16>) outs(%323 : tensor<10x4096x77xf16>) -> tensor<10x4096x77xf16>
%327 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%326 : tensor<10x4096x77xf16>) outs(%322 : tensor<10x4096x77xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.mulf %in, %cst_961 : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x77xf16>
%328 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%327, %323 : tensor<10x4096x77xf16>, tensor<10x4096x77xf16>) outs(%322 : tensor<10x4096x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in_2625, %cst_946 : f16
%6235 = arith.addf %in, %6234 : f16
linalg.yield %6235 : f16
} -> tensor<10x4096x77xf16>
%329:2 = linalg.generic {indexing_maps = [#map18, #map20, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%328 : tensor<10x4096x77xf16>) outs(%229, %227 : tensor<10x4096x1xf16>, tensor<10x4096x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<10x4096x1xf16>, tensor<10x4096x1xi64>)
%330 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%328, %329#0 : tensor<10x4096x77xf16>, tensor<10x4096x1xf16>) outs(%322 : tensor<10x4096x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x77xf16>
%331 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%330 : tensor<10x4096x77xf16>) outs(%322 : tensor<10x4096x77xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.exp %in : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x77xf16>
%332 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%331 : tensor<10x4096x77xf16>) outs(%233 : tensor<10x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.addf %in, %out : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x1xf16>
%333 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%331, %332 : tensor<10x4096x77xf16>, tensor<10x4096x1xf16>) outs(%322 : tensor<10x4096x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x77xf16>
%334 = linalg.batch_matmul ins(%333, %collapsed_1055 : tensor<10x4096x77xf16>, tensor<10x77x64xf16>) outs(%237 : tensor<10x4096x64xf16>) -> tensor<10x4096x64xf16>
%expanded_1056 = tensor.expand_shape %334 [[0, 1], [2], [3]] : tensor<10x4096x64xf16> into tensor<2x5x4096x64xf16>
%335 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1056 : tensor<2x5x4096x64xf16>) outs(%239 : tensor<2x4096x5x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x5x64xf16>
%collapsed_1057 = tensor.collapse_shape %335 [[0], [1], [2, 3]] : tensor<2x4096x5x64xf16> into tensor<2x4096x320xf16>
%expanded_1058 = tensor.expand_shape %collapsed_1057 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%336 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1058 : tensor<2x4096x20x16xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x20x16xf16>
%337:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%336 : tensor<2x4096x20x16xf16>) outs(%149, %147 : tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>)
%338 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%337#0 : tensor<2x4096x20x1xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1059 = tensor.collapse_shape %338 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%339 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1059 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%340 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1057, %339 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%341 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%340, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%342 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%341 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%343 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%342 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x4096x320xf16>
%344 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%343, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%345 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%344, %339 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%346 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_898 : tensor<320x20x1xf16>) outs(%160 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1060 = tensor.collapse_shape %346 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%347 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_897, %collapsed_1060 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320xf16>
%348 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%347 : tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1061 = tensor.collapse_shape %345 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%349 = linalg.matmul ins(%collapsed_1061, %348 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%350 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_896, %349 : tensor<320xf16>, tensor<8192x320xf16>) outs(%165 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<8192x320xf16>
%expanded_1062 = tensor.expand_shape %350 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16>
%351 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1062, %256 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%352 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%351 : tensor<2x4096x320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x4096x320xf32>
%353 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%352 : tensor<2x4096x320xf32>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%354 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%353 : tensor<2x4096x320xf64>) outs(%174 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%355 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%354 : tensor<2x4096x1xf64>) outs(%173 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_959 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%356 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%353, %355 : tensor<2x4096x320xf64>, tensor<2x4096x1xf64>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%357 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%356, %356 : tensor<2x4096x320xf64>, tensor<2x4096x320xf64>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%358 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%357 : tensor<2x4096x320xf64>) outs(%174 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%359 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%358 : tensor<2x4096x1xf64>) outs(%173 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_959 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%360 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%359 : tensor<2x4096x1xf64>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%361 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%352 : tensor<2x4096x320xf32>) outs(%183 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%362 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%361 : tensor<2x4096x1xf32>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_960 : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%363 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%360, %cst_938 : tensor<2x4096x1xf32>, tensor<f16>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x1xf32>
%364 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%363 : tensor<2x4096x1xf32>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%365 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%351, %362 : tensor<2x4096x320xf16>, tensor<2x4096x1xf32>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%366 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%365, %364 : tensor<2x4096x320xf32>, tensor<2x4096x1xf32>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x320xf32>
%367 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%366, %cst_895 : tensor<2x4096x320xf32>, tensor<320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%368 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%367, %cst_894 : tensor<2x4096x320xf32>, tensor<320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%369 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%368 : tensor<2x4096x320xf32>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%expanded_1063 = tensor.expand_shape %369 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%370 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1063 : tensor<2x4096x20x16xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x20x16xf16>
%371:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%370 : tensor<2x4096x20x16xf16>) outs(%149, %147 : tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>)
%372 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%371#0 : tensor<2x4096x20x1xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1064 = tensor.collapse_shape %372 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%373 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1064 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%374 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%369, %373 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%375 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%374, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%376 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%375 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%377 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%376 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x4096x320xf16>
%378 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%377, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%379 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%378, %373 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%380 = tensor.empty() : tensor<2560x20x16xf16>
%381 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_893 : tensor<2560x20x1xf16>) outs(%380 : tensor<2560x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2560x20x16xf16>
%collapsed_1065 = tensor.collapse_shape %381 [[0], [1, 2]] : tensor<2560x20x16xf16> into tensor<2560x320xf16>
%382 = tensor.empty() : tensor<2560x320xf16>
%383 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_892, %collapsed_1065 : tensor<2560x320xi8>, tensor<2560x320xf16>) outs(%382 : tensor<2560x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<2560x320xf16>
%384 = tensor.empty() : tensor<320x2560xf16>
%385 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%383 : tensor<2560x320xf16>) outs(%384 : tensor<320x2560xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x2560xf16>
%collapsed_1066 = tensor.collapse_shape %379 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%386 = tensor.empty() : tensor<8192x2560xf16>
%387 = linalg.fill ins(%cst_946 : f16) outs(%386 : tensor<8192x2560xf16>) -> tensor<8192x2560xf16>
%388 = linalg.matmul ins(%collapsed_1066, %385 : tensor<8192x320xf16>, tensor<320x2560xf16>) outs(%387 : tensor<8192x2560xf16>) -> tensor<8192x2560xf16>
%389 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_891, %388 : tensor<2560xf16>, tensor<8192x2560xf16>) outs(%386 : tensor<8192x2560xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<8192x2560xf16>
%expanded_1067 = tensor.expand_shape %389 [[0, 1], [2]] : tensor<8192x2560xf16> into tensor<2x4096x2560xf16>
%extracted_slice_1068 = tensor.extract_slice %expanded_1067[0, 0, 0] [2, 4096, 1280] [1, 1, 1] : tensor<2x4096x2560xf16> to tensor<2x4096x1280xf16>
%extracted_slice_1069 = tensor.extract_slice %expanded_1067[0, 0, 1280] [2, 4096, 1280] [1, 1, 1] : tensor<2x4096x2560xf16> to tensor<2x4096x1280xf16>
%390 = tensor.empty() : tensor<2x4096x1280xf16>
%391 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1069 : tensor<2x4096x1280xf16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.sqrt %cst_949 : f16
%6235 = arith.divf %in, %6234 : f16
%6236 = math.erf %6235 : f16
%6237 = arith.addf %6236, %cst_947 : f16
%6238 = arith.mulf %6237, %cst_950 : f16
%6239 = arith.mulf %in, %6238 : f16
linalg.yield %6239 : f16
} -> tensor<2x4096x1280xf16>
%392 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1068, %391 : tensor<2x4096x1280xf16>, tensor<2x4096x1280xf16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x1280xf16>
%expanded_1070 = tensor.expand_shape %392 [[0], [1], [2, 3]] : tensor<2x4096x1280xf16> into tensor<2x4096x80x16xf16>
%393 = tensor.empty() : tensor<2x4096x80x16xf16>
%394 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1070 : tensor<2x4096x80x16xf16>) outs(%393 : tensor<2x4096x80x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x80x16xf16>
%395 = tensor.empty() : tensor<2x4096x80x1xi64>
%396 = linalg.fill ins(%c0_i64 : i64) outs(%395 : tensor<2x4096x80x1xi64>) -> tensor<2x4096x80x1xi64>
%397 = tensor.empty() : tensor<2x4096x80x1xf16>
%398 = linalg.fill ins(%cst_948 : f16) outs(%397 : tensor<2x4096x80x1xf16>) -> tensor<2x4096x80x1xf16>
%399:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%394 : tensor<2x4096x80x16xf16>) outs(%398, %396 : tensor<2x4096x80x1xf16>, tensor<2x4096x80x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x4096x80x1xf16>, tensor<2x4096x80x1xi64>)
%400 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%399#0 : tensor<2x4096x80x1xf16>) outs(%393 : tensor<2x4096x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x80x16xf16>
%collapsed_1071 = tensor.collapse_shape %400 [[0], [1], [2, 3]] : tensor<2x4096x80x16xf16> into tensor<2x4096x1280xf16>
%401 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1071 : tensor<2x4096x1280xf16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x1280xf16>
%402 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%392, %401 : tensor<2x4096x1280xf16>, tensor<2x4096x1280xf16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x1280xf16>
%403 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%402, %cst_935 : tensor<2x4096x1280xf16>, tensor<f16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x1280xf16>
%404 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%403 : tensor<2x4096x1280xf16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x1280xf16>
%405 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%404 : tensor<2x4096x1280xf16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x4096x1280xf16>
%406 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%405, %cst_935 : tensor<2x4096x1280xf16>, tensor<f16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x1280xf16>
%407 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%406, %401 : tensor<2x4096x1280xf16>, tensor<2x4096x1280xf16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x1280xf16>
%408 = tensor.empty() : tensor<320x80x16xf16>
%409 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_890 : tensor<320x80x1xf16>) outs(%408 : tensor<320x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x80x16xf16>
%collapsed_1072 = tensor.collapse_shape %409 [[0], [1, 2]] : tensor<320x80x16xf16> into tensor<320x1280xf16>
%410 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_889, %collapsed_1072 : tensor<320x1280xi8>, tensor<320x1280xf16>) outs(%17 : tensor<320x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x1280xf16>
%411 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%410 : tensor<320x1280xf16>) outs(%82 : tensor<1280x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x320xf16>
%collapsed_1073 = tensor.collapse_shape %407 [[0, 1], [2]] : tensor<2x4096x1280xf16> into tensor<8192x1280xf16>
%412 = linalg.matmul ins(%collapsed_1073, %411 : tensor<8192x1280xf16>, tensor<1280x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%413 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_888, %412 : tensor<320xf16>, tensor<8192x320xf16>) outs(%165 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<8192x320xf16>
%expanded_1074 = tensor.expand_shape %413 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16>
%414 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1074, %351 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%expanded_1075 = tensor.expand_shape %414 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%415 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1075 : tensor<2x4096x20x16xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x20x16xf16>
%416:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%415 : tensor<2x4096x20x16xf16>) outs(%149, %147 : tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>)
%417 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%416#0 : tensor<2x4096x20x1xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1076 = tensor.collapse_shape %417 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%418 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1076 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%419 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%414, %418 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%420 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%419, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%421 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%420 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%422 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%421 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x4096x320xf16>
%423 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%422, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%424 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%423, %418 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%425 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_887 : tensor<320x20x1xf16>) outs(%160 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1077 = tensor.collapse_shape %425 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%426 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_886, %collapsed_1077 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320xf16>
%427 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%426 : tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1078 = tensor.collapse_shape %424 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%428 = linalg.matmul ins(%collapsed_1078, %427 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%429 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_885, %428 : tensor<320xf16>, tensor<8192x320xf16>) outs(%165 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<8192x320xf16>
%expanded_1079 = tensor.expand_shape %429 [[0, 1, 2], [3]] : tensor<8192x320xf16> into tensor<2x64x64x320xf16>
%430 = linalg.generic {indexing_maps = [#map1, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1079 : tensor<2x64x64x320xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%431 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%430, %123 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%collapsed_1080 = tensor.collapse_shape %431 [[0], [1], [2, 3]] : tensor<2x320x64x64xf16> into tensor<2x320x4096xf16>
%expanded_1081 = tensor.expand_shape %collapsed_1080 [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16>
%432 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1081 : tensor<2x32x10x4096xf16>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x10x4096xf32>
%433 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%432 : tensor<2x32x10x4096xf32>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%434 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%433 : tensor<2x32x10x4096xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%435 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%434 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_954 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%436 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%433, %435 : tensor<2x32x10x4096xf64>, tensor<2x32x1x1xf64>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%437 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%436, %436 : tensor<2x32x10x4096xf64>, tensor<2x32x10x4096xf64>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%438 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%437 : tensor<2x32x10x4096xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%439 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%438 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_954 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%440 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%439 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%441 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%432 : tensor<2x32x10x4096xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%442 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%441 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_955 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%443 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%440, %cst_938 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%444 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%443 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%445 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1081, %442 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x10x4096xf32>
%446 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%445, %444 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x10x4096xf32>
%collapsed_1082 = tensor.collapse_shape %446 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32>
%expanded_1083 = tensor.expand_shape %collapsed_1082 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32>
%expanded_1084 = tensor.expand_shape %cst_884 [[0, 1, 2, 3]] : tensor<320xf16> into tensor<1x320x1x1xf16>
%expanded_1085 = tensor.expand_shape %cst_883 [[0, 1, 2, 3]] : tensor<320xf16> into tensor<1x320x1x1xf16>
%447 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1083, %expanded_1085 : tensor<2x320x64x64xf32>, tensor<1x320x1x1xf16>) outs(%53 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x320x64x64xf32>
%448 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%447, %expanded_1084 : tensor<2x320x64x64xf32>, tensor<1x320x1x1xf16>) outs(%53 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x320x64x64xf32>
%449 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%448 : tensor<2x320x64x64xf32>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%450 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%449 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.negf %in : f16
%6235 = math.exp %6234 : f16
%6236 = arith.addf %6235, %cst_947 : f16
%6237 = arith.divf %cst_947, %6236 : f16
linalg.yield %6237 : f16
} -> tensor<2x320x64x64xf16>
%451 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%450, %449 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%expanded_1086 = tensor.expand_shape %451 [[0], [1, 2], [3], [4]] : tensor<2x320x64x64xf16> into tensor<2x20x16x64x64xf16>
%452 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1086 : tensor<2x20x16x64x64xf16>) outs(%59 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x20x16x64x64xf16>
%453:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%452 : tensor<2x20x16x64x64xf16>) outs(%64, %62 : tensor<2x20x1x64x64xf16>, tensor<2x20x1x64x64xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x20x1x64x64xf16>, tensor<2x20x1x64x64xi64>)
%454 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%453#0 : tensor<2x20x1x64x64xf16>) outs(%59 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x64x64xf16>
%collapsed_1087 = tensor.collapse_shape %454 [[0], [1, 2], [3], [4]] : tensor<2x20x16x64x64xf16> into tensor<2x320x64x64xf16>
%455 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1087 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%456 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%451, %455 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%457 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%456, %cst_935 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%458 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%457 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%459 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%458 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x320x64x64xf16>
%460 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%459, %cst_935 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%461 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%460, %455 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%462 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_882 : tensor<320x20x1x3x3xf16>) outs(%74 : tensor<320x20x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x3x3xf16>
%collapsed_1088 = tensor.collapse_shape %462 [[0], [1, 2], [3], [4]] : tensor<320x20x16x3x3xf16> into tensor<320x320x3x3xf16>
%463 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_881, %collapsed_1088 : tensor<320x320x3x3xi8>, tensor<320x320x3x3xf16>) outs(%76 : tensor<320x320x3x3xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320x3x3xf16>
%padded_1089 = tensor.pad %461 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x320x64x64xf16> to tensor<2x320x66x66xf16>
%464 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_880 : tensor<320xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%465 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1089, %463 : tensor<2x320x66x66xf16>, tensor<320x320x3x3xf16>) outs(%464 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16>
%466 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_879 : tensor<320x1280xf16>) outs(%82 : tensor<1280x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x320xf16>
%467 = linalg.matmul ins(%81, %466 : tensor<2x1280xf16>, tensor<1280x320xf16>) outs(%84 : tensor<2x320xf16>) -> tensor<2x320xf16>
%468 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_878, %467 : tensor<320xf16>, tensor<2x320xf16>) outs(%15 : tensor<2x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320xf16>
%expanded_1090 = tensor.expand_shape %468 [[0], [1, 2, 3]] : tensor<2x320xf16> into tensor<2x320x1x1xf16>
%469 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%465, %expanded_1090 : tensor<2x320x64x64xf16>, tensor<2x320x1x1xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%collapsed_1091 = tensor.collapse_shape %469 [[0], [1], [2, 3]] : tensor<2x320x64x64xf16> into tensor<2x320x4096xf16>
%expanded_1092 = tensor.expand_shape %collapsed_1091 [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16>
%470 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1092 : tensor<2x32x10x4096xf16>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x10x4096xf32>
%471 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%470 : tensor<2x32x10x4096xf32>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%472 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%471 : tensor<2x32x10x4096xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%473 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%472 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_954 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%474 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%471, %473 : tensor<2x32x10x4096xf64>, tensor<2x32x1x1xf64>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%475 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%474, %474 : tensor<2x32x10x4096xf64>, tensor<2x32x10x4096xf64>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%476 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%475 : tensor<2x32x10x4096xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%477 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%476 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_954 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%478 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%477 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%479 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%470 : tensor<2x32x10x4096xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%480 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%479 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_955 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%481 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%478, %cst_938 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%482 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%481 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%483 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1092, %480 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x10x4096xf32>
%484 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%483, %482 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x10x4096xf32>
%collapsed_1093 = tensor.collapse_shape %484 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32>
%expanded_1094 = tensor.expand_shape %collapsed_1093 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32>
%expanded_1095 = tensor.expand_shape %cst_877 [[0, 1, 2, 3]] : tensor<320xf16> into tensor<1x320x1x1xf16>
%expanded_1096 = tensor.expand_shape %cst_876 [[0, 1, 2, 3]] : tensor<320xf16> into tensor<1x320x1x1xf16>
%485 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1094, %expanded_1096 : tensor<2x320x64x64xf32>, tensor<1x320x1x1xf16>) outs(%53 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x320x64x64xf32>
%486 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%485, %expanded_1095 : tensor<2x320x64x64xf32>, tensor<1x320x1x1xf16>) outs(%53 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x320x64x64xf32>
%487 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%486 : tensor<2x320x64x64xf32>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%488 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%487 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.negf %in : f16
%6235 = math.exp %6234 : f16
%6236 = arith.addf %6235, %cst_947 : f16
%6237 = arith.divf %cst_947, %6236 : f16
linalg.yield %6237 : f16
} -> tensor<2x320x64x64xf16>
%489 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%488, %487 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%expanded_1097 = tensor.expand_shape %489 [[0], [1, 2], [3], [4]] : tensor<2x320x64x64xf16> into tensor<2x20x16x64x64xf16>
%490 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1097 : tensor<2x20x16x64x64xf16>) outs(%59 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x20x16x64x64xf16>
%491:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%490 : tensor<2x20x16x64x64xf16>) outs(%64, %62 : tensor<2x20x1x64x64xf16>, tensor<2x20x1x64x64xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x20x1x64x64xf16>, tensor<2x20x1x64x64xi64>)
%492 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%491#0 : tensor<2x20x1x64x64xf16>) outs(%59 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x64x64xf16>
%collapsed_1098 = tensor.collapse_shape %492 [[0], [1, 2], [3], [4]] : tensor<2x20x16x64x64xf16> into tensor<2x320x64x64xf16>
%493 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1098 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%494 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%489, %493 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%495 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%494, %cst_935 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%496 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%495 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%497 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%496 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x320x64x64xf16>
%498 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%497, %cst_935 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%499 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%498, %493 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%500 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_875 : tensor<320x20x1x3x3xf16>) outs(%74 : tensor<320x20x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x3x3xf16>
%collapsed_1099 = tensor.collapse_shape %500 [[0], [1, 2], [3], [4]] : tensor<320x20x16x3x3xf16> into tensor<320x320x3x3xf16>
%501 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_874, %collapsed_1099 : tensor<320x320x3x3xi8>, tensor<320x320x3x3xf16>) outs(%76 : tensor<320x320x3x3xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320x3x3xf16>
%padded_1100 = tensor.pad %499 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x320x64x64xf16> to tensor<2x320x66x66xf16>
%502 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_873 : tensor<320xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%503 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1100, %501 : tensor<2x320x66x66xf16>, tensor<320x320x3x3xf16>) outs(%502 : tensor<2x320x64x64xf16>) -> tensor<2x320x64x64xf16>
%504 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%431, %503 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%505 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%504, %cst_924 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%collapsed_1101 = tensor.collapse_shape %505 [[0], [1], [2, 3]] : tensor<2x320x64x64xf16> into tensor<2x320x4096xf16>
%expanded_1102 = tensor.expand_shape %collapsed_1101 [[0], [1, 2], [3]] : tensor<2x320x4096xf16> into tensor<2x32x10x4096xf16>
%506 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1102 : tensor<2x32x10x4096xf16>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x10x4096xf32>
%507 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%506 : tensor<2x32x10x4096xf32>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%508 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%507 : tensor<2x32x10x4096xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%509 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%508 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_954 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%510 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%507, %509 : tensor<2x32x10x4096xf64>, tensor<2x32x1x1xf64>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%511 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%510, %510 : tensor<2x32x10x4096xf64>, tensor<2x32x10x4096xf64>) outs(%34 : tensor<2x32x10x4096xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x4096xf64>
%512 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%511 : tensor<2x32x10x4096xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%513 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%512 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_954 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%514 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%513 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%515 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%506 : tensor<2x32x10x4096xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%516 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%515 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_955 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%517 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%514, %cst_923 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%518 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%517 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%519 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1102, %516 : tensor<2x32x10x4096xf16>, tensor<2x32x1x1xf32>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x10x4096xf32>
%520 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%519, %518 : tensor<2x32x10x4096xf32>, tensor<2x32x1x1xf32>) outs(%32 : tensor<2x32x10x4096xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x10x4096xf32>
%collapsed_1103 = tensor.collapse_shape %520 [[0], [1, 2], [3]] : tensor<2x32x10x4096xf32> into tensor<2x320x4096xf32>
%expanded_1104 = tensor.expand_shape %collapsed_1103 [[0], [1], [2, 3]] : tensor<2x320x4096xf32> into tensor<2x320x64x64xf32>
%expanded_1105 = tensor.expand_shape %cst_872 [[0, 1, 2, 3]] : tensor<320xf16> into tensor<1x320x1x1xf16>
%expanded_1106 = tensor.expand_shape %cst_871 [[0, 1, 2, 3]] : tensor<320xf16> into tensor<1x320x1x1xf16>
%521 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1104, %expanded_1106 : tensor<2x320x64x64xf32>, tensor<1x320x1x1xf16>) outs(%53 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x320x64x64xf32>
%522 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%521, %expanded_1105 : tensor<2x320x64x64xf32>, tensor<1x320x1x1xf16>) outs(%53 : tensor<2x320x64x64xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x320x64x64xf32>
%523 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%522 : tensor<2x320x64x64xf32>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%524 = linalg.generic {indexing_maps = [#map1, #map16], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%523 : tensor<2x320x64x64xf16>) outs(%142 : tensor<2x64x64x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x64x64x320xf16>
%collapsed_1107 = tensor.collapse_shape %524 [[0], [1, 2], [3]] : tensor<2x64x64x320xf16> into tensor<2x4096x320xf16>
%expanded_1108 = tensor.expand_shape %collapsed_1107 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%525 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1108 : tensor<2x4096x20x16xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x20x16xf16>
%526:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%525 : tensor<2x4096x20x16xf16>) outs(%149, %147 : tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>)
%527 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%526#0 : tensor<2x4096x20x1xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1109 = tensor.collapse_shape %527 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%528 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1109 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%529 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1107, %528 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%530 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%529, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%531 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%530 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%532 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%531 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x4096x320xf16>
%533 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%532, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%534 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%533, %528 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%535 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_870 : tensor<320x20x1xf16>) outs(%160 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1110 = tensor.collapse_shape %535 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%536 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_869, %collapsed_1110 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320xf16>
%537 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%536 : tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1111 = tensor.collapse_shape %534 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%538 = linalg.matmul ins(%collapsed_1111, %537 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%539 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_868, %538 : tensor<320xf16>, tensor<8192x320xf16>) outs(%165 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<8192x320xf16>
%expanded_1112 = tensor.expand_shape %539 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16>
%540 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1112 : tensor<2x4096x320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x4096x320xf32>
%541 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%540 : tensor<2x4096x320xf32>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%542 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%541 : tensor<2x4096x320xf64>) outs(%174 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%543 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%542 : tensor<2x4096x1xf64>) outs(%173 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_959 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%544 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%541, %543 : tensor<2x4096x320xf64>, tensor<2x4096x1xf64>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%545 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%544, %544 : tensor<2x4096x320xf64>, tensor<2x4096x320xf64>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%546 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%545 : tensor<2x4096x320xf64>) outs(%174 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%547 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%546 : tensor<2x4096x1xf64>) outs(%173 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_959 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%548 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%547 : tensor<2x4096x1xf64>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%549 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%540 : tensor<2x4096x320xf32>) outs(%183 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%550 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%549 : tensor<2x4096x1xf32>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_960 : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%551 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%548, %cst_938 : tensor<2x4096x1xf32>, tensor<f16>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x1xf32>
%552 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%551 : tensor<2x4096x1xf32>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%553 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1112, %550 : tensor<2x4096x320xf16>, tensor<2x4096x1xf32>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%554 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%553, %552 : tensor<2x4096x320xf32>, tensor<2x4096x1xf32>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x320xf32>
%555 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%554, %cst_867 : tensor<2x4096x320xf32>, tensor<320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%556 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%555, %cst_866 : tensor<2x4096x320xf32>, tensor<320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%557 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%556 : tensor<2x4096x320xf32>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%expanded_1113 = tensor.expand_shape %557 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%558 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1113 : tensor<2x4096x20x16xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x20x16xf16>
%559:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%558 : tensor<2x4096x20x16xf16>) outs(%149, %147 : tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>)
%560 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%559#0 : tensor<2x4096x20x1xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1114 = tensor.collapse_shape %560 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%561 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1114 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%562 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%557, %561 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%563 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%562, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%564 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%563 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%565 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%564 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x4096x320xf16>
%566 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%565, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%567 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%566, %561 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%568 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_865 : tensor<320x20x1xf16>) outs(%160 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1115 = tensor.collapse_shape %568 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%569 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_864, %collapsed_1115 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320xf16>
%570 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%569 : tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1116 = tensor.collapse_shape %567 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%571 = linalg.matmul ins(%collapsed_1116, %570 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%expanded_1117 = tensor.expand_shape %571 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x5x64xf16>
%572 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1117 : tensor<2x4096x5x64xf16>) outs(%207 : tensor<2x5x4096x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x5x4096x64xf16>
%collapsed_1118 = tensor.collapse_shape %572 [[0, 1], [2], [3]] : tensor<2x5x4096x64xf16> into tensor<10x4096x64xf16>
%573 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_863 : tensor<320x20x1xf16>) outs(%160 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1119 = tensor.collapse_shape %573 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%574 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_862, %collapsed_1119 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320xf16>
%575 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%574 : tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%576 = linalg.matmul ins(%collapsed_1116, %575 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%577 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_861 : tensor<320x20x1xf16>) outs(%160 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1120 = tensor.collapse_shape %577 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%578 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_860, %collapsed_1120 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320xf16>
%579 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%578 : tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%580 = linalg.matmul ins(%collapsed_1116, %579 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%expanded_1121 = tensor.expand_shape %576 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x5x64xf16>
%581 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1121 : tensor<2x4096x5x64xf16>) outs(%207 : tensor<2x5x4096x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x5x4096x64xf16>
%collapsed_1122 = tensor.collapse_shape %581 [[0, 1], [2], [3]] : tensor<2x5x4096x64xf16> into tensor<10x4096x64xf16>
%expanded_1123 = tensor.expand_shape %580 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x5x64xf16>
%582 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1123 : tensor<2x4096x5x64xf16>) outs(%207 : tensor<2x5x4096x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x5x4096x64xf16>
%collapsed_1124 = tensor.collapse_shape %582 [[0, 1], [2], [3]] : tensor<2x5x4096x64xf16> into tensor<10x4096x64xf16>
%583 = linalg.generic {indexing_maps = [#map18, #map23], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1122 : tensor<10x4096x64xf16>) outs(%221 : tensor<10x64x4096xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<10x64x4096xf16>
%584 = linalg.batch_matmul ins(%collapsed_1118, %583 : tensor<10x4096x64xf16>, tensor<10x64x4096xf16>) outs(%220 : tensor<10x4096x4096xf16>) -> tensor<10x4096x4096xf16>
%585 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%584 : tensor<10x4096x4096xf16>) outs(%219 : tensor<10x4096x4096xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.mulf %in, %cst_961 : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x4096xf16>
%586 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%585, %220 : tensor<10x4096x4096xf16>, tensor<10x4096x4096xf16>) outs(%219 : tensor<10x4096x4096xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in_2625, %cst_946 : f16
%6235 = arith.addf %in, %6234 : f16
linalg.yield %6235 : f16
} -> tensor<10x4096x4096xf16>
%587:2 = linalg.generic {indexing_maps = [#map18, #map20, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%586 : tensor<10x4096x4096xf16>) outs(%229, %227 : tensor<10x4096x1xf16>, tensor<10x4096x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<10x4096x1xf16>, tensor<10x4096x1xi64>)
%588 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%586, %587#0 : tensor<10x4096x4096xf16>, tensor<10x4096x1xf16>) outs(%219 : tensor<10x4096x4096xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x4096xf16>
%589 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%588 : tensor<10x4096x4096xf16>) outs(%219 : tensor<10x4096x4096xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.exp %in : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x4096xf16>
%590 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%589 : tensor<10x4096x4096xf16>) outs(%233 : tensor<10x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.addf %in, %out : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x1xf16>
%591 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%589, %590 : tensor<10x4096x4096xf16>, tensor<10x4096x1xf16>) outs(%219 : tensor<10x4096x4096xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x4096xf16>
%592 = linalg.batch_matmul ins(%591, %collapsed_1124 : tensor<10x4096x4096xf16>, tensor<10x4096x64xf16>) outs(%237 : tensor<10x4096x64xf16>) -> tensor<10x4096x64xf16>
%expanded_1125 = tensor.expand_shape %592 [[0, 1], [2], [3]] : tensor<10x4096x64xf16> into tensor<2x5x4096x64xf16>
%593 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1125 : tensor<2x5x4096x64xf16>) outs(%239 : tensor<2x4096x5x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x5x64xf16>
%collapsed_1126 = tensor.collapse_shape %593 [[0], [1], [2, 3]] : tensor<2x4096x5x64xf16> into tensor<2x4096x320xf16>
%expanded_1127 = tensor.expand_shape %collapsed_1126 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%594 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1127 : tensor<2x4096x20x16xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x20x16xf16>
%595:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%594 : tensor<2x4096x20x16xf16>) outs(%149, %147 : tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>)
%596 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%595#0 : tensor<2x4096x20x1xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1128 = tensor.collapse_shape %596 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%597 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1128 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%598 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1126, %597 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%599 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%598, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%600 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%599 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%601 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%600 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x4096x320xf16>
%602 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%601, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%603 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%602, %597 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%604 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_859 : tensor<320x20x1xf16>) outs(%160 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1129 = tensor.collapse_shape %604 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%605 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_858, %collapsed_1129 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320xf16>
%606 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%605 : tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1130 = tensor.collapse_shape %603 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%607 = linalg.matmul ins(%collapsed_1130, %606 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%608 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_857, %607 : tensor<320xf16>, tensor<8192x320xf16>) outs(%165 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<8192x320xf16>
%expanded_1131 = tensor.expand_shape %608 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16>
%609 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1131, %expanded_1112 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%610 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%609 : tensor<2x4096x320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x4096x320xf32>
%611 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%610 : tensor<2x4096x320xf32>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%612 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%611 : tensor<2x4096x320xf64>) outs(%174 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%613 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%612 : tensor<2x4096x1xf64>) outs(%173 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_959 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%614 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%611, %613 : tensor<2x4096x320xf64>, tensor<2x4096x1xf64>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%615 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%614, %614 : tensor<2x4096x320xf64>, tensor<2x4096x320xf64>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%616 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%615 : tensor<2x4096x320xf64>) outs(%174 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%617 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%616 : tensor<2x4096x1xf64>) outs(%173 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_959 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%618 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%617 : tensor<2x4096x1xf64>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%619 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%610 : tensor<2x4096x320xf32>) outs(%183 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%620 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%619 : tensor<2x4096x1xf32>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_960 : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%621 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%618, %cst_938 : tensor<2x4096x1xf32>, tensor<f16>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x1xf32>
%622 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%621 : tensor<2x4096x1xf32>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%623 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%609, %620 : tensor<2x4096x320xf16>, tensor<2x4096x1xf32>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%624 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%623, %622 : tensor<2x4096x320xf32>, tensor<2x4096x1xf32>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x320xf32>
%625 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%624, %cst_856 : tensor<2x4096x320xf32>, tensor<320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%626 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%625, %cst_855 : tensor<2x4096x320xf32>, tensor<320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%627 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%626 : tensor<2x4096x320xf32>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%expanded_1132 = tensor.expand_shape %627 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%628 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1132 : tensor<2x4096x20x16xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x20x16xf16>
%629:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%628 : tensor<2x4096x20x16xf16>) outs(%149, %147 : tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>)
%630 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%629#0 : tensor<2x4096x20x1xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1133 = tensor.collapse_shape %630 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%631 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1133 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%632 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%627, %631 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%633 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%632, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%634 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%633 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%635 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%634 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x4096x320xf16>
%636 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%635, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%637 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%636, %631 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%638 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_854 : tensor<320x20x1xf16>) outs(%160 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1134 = tensor.collapse_shape %638 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%639 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_853, %collapsed_1134 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320xf16>
%640 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%639 : tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1135 = tensor.collapse_shape %637 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%641 = linalg.matmul ins(%collapsed_1135, %640 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%expanded_1136 = tensor.expand_shape %641 [[0, 1], [2, 3]] : tensor<8192x320xf16> into tensor<2x4096x5x64xf16>
%642 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1136 : tensor<2x4096x5x64xf16>) outs(%207 : tensor<2x5x4096x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x5x4096x64xf16>
%collapsed_1137 = tensor.collapse_shape %642 [[0, 1], [2], [3]] : tensor<2x5x4096x64xf16> into tensor<10x4096x64xf16>
%643 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_852 : tensor<320x64x1xf16>) outs(%306 : tensor<320x64x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x64x16xf16>
%collapsed_1138 = tensor.collapse_shape %643 [[0], [1, 2]] : tensor<320x64x16xf16> into tensor<320x1024xf16>
%644 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_851, %collapsed_1138 : tensor<320x1024xi8>, tensor<320x1024xf16>) outs(%308 : tensor<320x1024xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x1024xf16>
%645 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%644 : tensor<320x1024xf16>) outs(%310 : tensor<1024x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1024x320xf16>
%646 = linalg.matmul ins(%collapsed_1050, %645 : tensor<154x1024xf16>, tensor<1024x320xf16>) outs(%313 : tensor<154x320xf16>) -> tensor<154x320xf16>
%647 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_850 : tensor<320x64x1xf16>) outs(%306 : tensor<320x64x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x64x16xf16>
%collapsed_1139 = tensor.collapse_shape %647 [[0], [1, 2]] : tensor<320x64x16xf16> into tensor<320x1024xf16>
%648 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_849, %collapsed_1139 : tensor<320x1024xi8>, tensor<320x1024xf16>) outs(%308 : tensor<320x1024xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x1024xf16>
%649 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%648 : tensor<320x1024xf16>) outs(%310 : tensor<1024x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1024x320xf16>
%650 = linalg.matmul ins(%collapsed_1050, %649 : tensor<154x1024xf16>, tensor<1024x320xf16>) outs(%313 : tensor<154x320xf16>) -> tensor<154x320xf16>
%expanded_1140 = tensor.expand_shape %646 [[0, 1], [2, 3]] : tensor<154x320xf16> into tensor<2x77x5x64xf16>
%651 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1140 : tensor<2x77x5x64xf16>) outs(%319 : tensor<2x5x77x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x5x77x64xf16>
%collapsed_1141 = tensor.collapse_shape %651 [[0, 1], [2], [3]] : tensor<2x5x77x64xf16> into tensor<10x77x64xf16>
%expanded_1142 = tensor.expand_shape %650 [[0, 1], [2, 3]] : tensor<154x320xf16> into tensor<2x77x5x64xf16>
%652 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1142 : tensor<2x77x5x64xf16>) outs(%319 : tensor<2x5x77x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x5x77x64xf16>
%collapsed_1143 = tensor.collapse_shape %652 [[0, 1], [2], [3]] : tensor<2x5x77x64xf16> into tensor<10x77x64xf16>
%653 = linalg.generic {indexing_maps = [#map18, #map23], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1141 : tensor<10x77x64xf16>) outs(%324 : tensor<10x64x77xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<10x64x77xf16>
%654 = linalg.batch_matmul ins(%collapsed_1137, %653 : tensor<10x4096x64xf16>, tensor<10x64x77xf16>) outs(%323 : tensor<10x4096x77xf16>) -> tensor<10x4096x77xf16>
%655 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%654 : tensor<10x4096x77xf16>) outs(%322 : tensor<10x4096x77xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.mulf %in, %cst_961 : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x77xf16>
%656 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%655, %323 : tensor<10x4096x77xf16>, tensor<10x4096x77xf16>) outs(%322 : tensor<10x4096x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in_2625, %cst_946 : f16
%6235 = arith.addf %in, %6234 : f16
linalg.yield %6235 : f16
} -> tensor<10x4096x77xf16>
%657:2 = linalg.generic {indexing_maps = [#map18, #map20, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%656 : tensor<10x4096x77xf16>) outs(%229, %227 : tensor<10x4096x1xf16>, tensor<10x4096x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<10x4096x1xf16>, tensor<10x4096x1xi64>)
%658 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%656, %657#0 : tensor<10x4096x77xf16>, tensor<10x4096x1xf16>) outs(%322 : tensor<10x4096x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x77xf16>
%659 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%658 : tensor<10x4096x77xf16>) outs(%322 : tensor<10x4096x77xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.exp %in : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x77xf16>
%660 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%659 : tensor<10x4096x77xf16>) outs(%233 : tensor<10x4096x1xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.addf %in, %out : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x1xf16>
%661 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%659, %660 : tensor<10x4096x77xf16>, tensor<10x4096x1xf16>) outs(%322 : tensor<10x4096x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<10x4096x77xf16>
%662 = linalg.batch_matmul ins(%661, %collapsed_1143 : tensor<10x4096x77xf16>, tensor<10x77x64xf16>) outs(%237 : tensor<10x4096x64xf16>) -> tensor<10x4096x64xf16>
%expanded_1144 = tensor.expand_shape %662 [[0, 1], [2], [3]] : tensor<10x4096x64xf16> into tensor<2x5x4096x64xf16>
%663 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1144 : tensor<2x5x4096x64xf16>) outs(%239 : tensor<2x4096x5x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x5x64xf16>
%collapsed_1145 = tensor.collapse_shape %663 [[0], [1], [2, 3]] : tensor<2x4096x5x64xf16> into tensor<2x4096x320xf16>
%expanded_1146 = tensor.expand_shape %collapsed_1145 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%664 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1146 : tensor<2x4096x20x16xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x20x16xf16>
%665:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%664 : tensor<2x4096x20x16xf16>) outs(%149, %147 : tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>)
%666 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%665#0 : tensor<2x4096x20x1xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1147 = tensor.collapse_shape %666 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%667 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1147 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%668 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1145, %667 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%669 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%668, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%670 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%669 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%671 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%670 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x4096x320xf16>
%672 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%671, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%673 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%672, %667 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%674 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_848 : tensor<320x20x1xf16>) outs(%160 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1148 = tensor.collapse_shape %674 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%675 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_847, %collapsed_1148 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320xf16>
%676 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%675 : tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1149 = tensor.collapse_shape %673 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%677 = linalg.matmul ins(%collapsed_1149, %676 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%678 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_846, %677 : tensor<320xf16>, tensor<8192x320xf16>) outs(%165 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<8192x320xf16>
%expanded_1150 = tensor.expand_shape %678 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16>
%679 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1150, %609 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%680 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%679 : tensor<2x4096x320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x4096x320xf32>
%681 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%680 : tensor<2x4096x320xf32>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%682 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%681 : tensor<2x4096x320xf64>) outs(%174 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%683 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%682 : tensor<2x4096x1xf64>) outs(%173 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_959 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%684 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%681, %683 : tensor<2x4096x320xf64>, tensor<2x4096x1xf64>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%685 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%684, %684 : tensor<2x4096x320xf64>, tensor<2x4096x320xf64>) outs(%171 : tensor<2x4096x320xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x320xf64>
%686 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%685 : tensor<2x4096x320xf64>) outs(%174 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%687 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%686 : tensor<2x4096x1xf64>) outs(%173 : tensor<2x4096x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_959 : f64
linalg.yield %6234 : f64
} -> tensor<2x4096x1xf64>
%688 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%687 : tensor<2x4096x1xf64>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%689 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%680 : tensor<2x4096x320xf32>) outs(%183 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%690 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%689 : tensor<2x4096x1xf32>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_960 : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%691 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%688, %cst_938 : tensor<2x4096x1xf32>, tensor<f16>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x1xf32>
%692 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%691 : tensor<2x4096x1xf32>) outs(%181 : tensor<2x4096x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x1xf32>
%693 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%679, %690 : tensor<2x4096x320xf16>, tensor<2x4096x1xf32>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%694 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%693, %692 : tensor<2x4096x320xf32>, tensor<2x4096x1xf32>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x4096x320xf32>
%695 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%694, %cst_845 : tensor<2x4096x320xf32>, tensor<320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%696 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%695, %cst_844 : tensor<2x4096x320xf32>, tensor<320xf16>) outs(%169 : tensor<2x4096x320xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x4096x320xf32>
%697 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%696 : tensor<2x4096x320xf32>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%expanded_1151 = tensor.expand_shape %697 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%698 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1151 : tensor<2x4096x20x16xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x20x16xf16>
%699:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%698 : tensor<2x4096x20x16xf16>) outs(%149, %147 : tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>)
%700 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%699#0 : tensor<2x4096x20x1xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1152 = tensor.collapse_shape %700 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%701 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1152 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%702 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%697, %701 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%703 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%702, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%704 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%703 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%705 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%704 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x4096x320xf16>
%706 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%705, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%707 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%706, %701 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%708 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_843 : tensor<2560x20x1xf16>) outs(%380 : tensor<2560x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2560x20x16xf16>
%collapsed_1153 = tensor.collapse_shape %708 [[0], [1, 2]] : tensor<2560x20x16xf16> into tensor<2560x320xf16>
%709 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_842, %collapsed_1153 : tensor<2560x320xi8>, tensor<2560x320xf16>) outs(%382 : tensor<2560x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<2560x320xf16>
%710 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%709 : tensor<2560x320xf16>) outs(%384 : tensor<320x2560xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x2560xf16>
%collapsed_1154 = tensor.collapse_shape %707 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%711 = linalg.matmul ins(%collapsed_1154, %710 : tensor<8192x320xf16>, tensor<320x2560xf16>) outs(%387 : tensor<8192x2560xf16>) -> tensor<8192x2560xf16>
%712 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_841, %711 : tensor<2560xf16>, tensor<8192x2560xf16>) outs(%386 : tensor<8192x2560xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<8192x2560xf16>
%expanded_1155 = tensor.expand_shape %712 [[0, 1], [2]] : tensor<8192x2560xf16> into tensor<2x4096x2560xf16>
%extracted_slice_1156 = tensor.extract_slice %expanded_1155[0, 0, 0] [2, 4096, 1280] [1, 1, 1] : tensor<2x4096x2560xf16> to tensor<2x4096x1280xf16>
%extracted_slice_1157 = tensor.extract_slice %expanded_1155[0, 0, 1280] [2, 4096, 1280] [1, 1, 1] : tensor<2x4096x2560xf16> to tensor<2x4096x1280xf16>
%713 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1157 : tensor<2x4096x1280xf16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.sqrt %cst_949 : f16
%6235 = arith.divf %in, %6234 : f16
%6236 = math.erf %6235 : f16
%6237 = arith.addf %6236, %cst_947 : f16
%6238 = arith.mulf %6237, %cst_950 : f16
%6239 = arith.mulf %in, %6238 : f16
linalg.yield %6239 : f16
} -> tensor<2x4096x1280xf16>
%714 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1156, %713 : tensor<2x4096x1280xf16>, tensor<2x4096x1280xf16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x1280xf16>
%expanded_1158 = tensor.expand_shape %714 [[0], [1], [2, 3]] : tensor<2x4096x1280xf16> into tensor<2x4096x80x16xf16>
%715 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1158 : tensor<2x4096x80x16xf16>) outs(%393 : tensor<2x4096x80x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x80x16xf16>
%716:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%715 : tensor<2x4096x80x16xf16>) outs(%398, %396 : tensor<2x4096x80x1xf16>, tensor<2x4096x80x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x4096x80x1xf16>, tensor<2x4096x80x1xi64>)
%717 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%716#0 : tensor<2x4096x80x1xf16>) outs(%393 : tensor<2x4096x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x80x16xf16>
%collapsed_1159 = tensor.collapse_shape %717 [[0], [1], [2, 3]] : tensor<2x4096x80x16xf16> into tensor<2x4096x1280xf16>
%718 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1159 : tensor<2x4096x1280xf16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x1280xf16>
%719 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%714, %718 : tensor<2x4096x1280xf16>, tensor<2x4096x1280xf16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x1280xf16>
%720 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%719, %cst_935 : tensor<2x4096x1280xf16>, tensor<f16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x1280xf16>
%721 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%720 : tensor<2x4096x1280xf16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x1280xf16>
%722 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%721 : tensor<2x4096x1280xf16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x4096x1280xf16>
%723 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%722, %cst_935 : tensor<2x4096x1280xf16>, tensor<f16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x1280xf16>
%724 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%723, %718 : tensor<2x4096x1280xf16>, tensor<2x4096x1280xf16>) outs(%390 : tensor<2x4096x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x1280xf16>
%725 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_840 : tensor<320x80x1xf16>) outs(%408 : tensor<320x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x80x16xf16>
%collapsed_1160 = tensor.collapse_shape %725 [[0], [1, 2]] : tensor<320x80x16xf16> into tensor<320x1280xf16>
%726 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_839, %collapsed_1160 : tensor<320x1280xi8>, tensor<320x1280xf16>) outs(%17 : tensor<320x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x1280xf16>
%727 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%726 : tensor<320x1280xf16>) outs(%82 : tensor<1280x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x320xf16>
%collapsed_1161 = tensor.collapse_shape %724 [[0, 1], [2]] : tensor<2x4096x1280xf16> into tensor<8192x1280xf16>
%728 = linalg.matmul ins(%collapsed_1161, %727 : tensor<8192x1280xf16>, tensor<1280x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%729 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_838, %728 : tensor<320xf16>, tensor<8192x320xf16>) outs(%165 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<8192x320xf16>
%expanded_1162 = tensor.expand_shape %729 [[0, 1], [2]] : tensor<8192x320xf16> into tensor<2x4096x320xf16>
%730 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1162, %679 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%expanded_1163 = tensor.expand_shape %730 [[0], [1], [2, 3]] : tensor<2x4096x320xf16> into tensor<2x4096x20x16xf16>
%731 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1163 : tensor<2x4096x20x16xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x20x16xf16>
%732:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%731 : tensor<2x4096x20x16xf16>) outs(%149, %147 : tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x4096x20x1xf16>, tensor<2x4096x20x1xi64>)
%733 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%732#0 : tensor<2x4096x20x1xf16>) outs(%144 : tensor<2x4096x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x4096x20x16xf16>
%collapsed_1164 = tensor.collapse_shape %733 [[0], [1], [2, 3]] : tensor<2x4096x20x16xf16> into tensor<2x4096x320xf16>
%734 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1164 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%735 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%730, %734 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%736 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%735, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%737 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%736 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%738 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%737 : tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x4096x320xf16>
%739 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%738, %cst_935 : tensor<2x4096x320xf16>, tensor<f16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%740 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%739, %734 : tensor<2x4096x320xf16>, tensor<2x4096x320xf16>) outs(%152 : tensor<2x4096x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x4096x320xf16>
%741 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_837 : tensor<320x20x1xf16>) outs(%160 : tensor<320x20x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16xf16>
%collapsed_1165 = tensor.collapse_shape %741 [[0], [1, 2]] : tensor<320x20x16xf16> into tensor<320x320xf16>
%742 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_836, %collapsed_1165 : tensor<320x320xi8>, tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320xf16>
%743 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%742 : tensor<320x320xf16>) outs(%162 : tensor<320x320xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x320xf16>
%collapsed_1166 = tensor.collapse_shape %740 [[0, 1], [2]] : tensor<2x4096x320xf16> into tensor<8192x320xf16>
%744 = linalg.matmul ins(%collapsed_1166, %743 : tensor<8192x320xf16>, tensor<320x320xf16>) outs(%166 : tensor<8192x320xf16>) -> tensor<8192x320xf16>
%745 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_835, %744 : tensor<320xf16>, tensor<8192x320xf16>) outs(%165 : tensor<8192x320xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<8192x320xf16>
%expanded_1167 = tensor.expand_shape %745 [[0, 1, 2], [3]] : tensor<8192x320xf16> into tensor<2x64x64x320xf16>
%746 = linalg.generic {indexing_maps = [#map1, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1167 : tensor<2x64x64x320xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x64x64xf16>
%747 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%746, %505 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%expanded_1168 = tensor.expand_shape %747 [[0], [1, 2], [3], [4]] : tensor<2x320x64x64xf16> into tensor<2x20x16x64x64xf16>
%748 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1168 : tensor<2x20x16x64x64xf16>) outs(%59 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x20x16x64x64xf16>
%749:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%748 : tensor<2x20x16x64x64xf16>) outs(%64, %62 : tensor<2x20x1x64x64xf16>, tensor<2x20x1x64x64xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x20x1x64x64xf16>, tensor<2x20x1x64x64xi64>)
%750 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%749#0 : tensor<2x20x1x64x64xf16>) outs(%59 : tensor<2x20x16x64x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x64x64xf16>
%collapsed_1169 = tensor.collapse_shape %750 [[0], [1, 2], [3], [4]] : tensor<2x20x16x64x64xf16> into tensor<2x320x64x64xf16>
%751 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1169 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%752 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%747, %751 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%753 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%752, %cst_935 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%754 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%753 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%755 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%754 : tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x320x64x64xf16>
%756 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%755, %cst_935 : tensor<2x320x64x64xf16>, tensor<f16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%757 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%756, %751 : tensor<2x320x64x64xf16>, tensor<2x320x64x64xf16>) outs(%29 : tensor<2x320x64x64xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x64x64xf16>
%758 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_834 : tensor<320x20x1x3x3xf16>) outs(%74 : tensor<320x20x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<320x20x16x3x3xf16>
%collapsed_1170 = tensor.collapse_shape %758 [[0], [1, 2], [3], [4]] : tensor<320x20x16x3x3xf16> into tensor<320x320x3x3xf16>
%759 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_833, %collapsed_1170 : tensor<320x320x3x3xi8>, tensor<320x320x3x3xf16>) outs(%76 : tensor<320x320x3x3xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<320x320x3x3xf16>
%padded_1171 = tensor.pad %757 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x320x64x64xf16> to tensor<2x320x66x66xf16>
%760 = tensor.empty() : tensor<2x320x32x32xf16>
%761 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_832 : tensor<320xf16>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x320x32x32xf16>
%762 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%padded_1171, %759 : tensor<2x320x66x66xf16>, tensor<320x320x3x3xf16>) outs(%761 : tensor<2x320x32x32xf16>) -> tensor<2x320x32x32xf16>
%collapsed_1172 = tensor.collapse_shape %762 [[0], [1], [2, 3]] : tensor<2x320x32x32xf16> into tensor<2x320x1024xf16>
%expanded_1173 = tensor.expand_shape %collapsed_1172 [[0], [1, 2], [3]] : tensor<2x320x1024xf16> into tensor<2x32x10x1024xf16>
%763 = tensor.empty() : tensor<2x32x10x1024xf32>
%764 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1173 : tensor<2x32x10x1024xf16>) outs(%763 : tensor<2x32x10x1024xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x10x1024xf32>
%765 = tensor.empty() : tensor<2x32x10x1024xf64>
%766 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%764 : tensor<2x32x10x1024xf32>) outs(%765 : tensor<2x32x10x1024xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x1024xf64>
%767 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%766 : tensor<2x32x10x1024xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%768 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%767 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_962 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%769 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%766, %768 : tensor<2x32x10x1024xf64>, tensor<2x32x1x1xf64>) outs(%765 : tensor<2x32x10x1024xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x1024xf64>
%770 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%769, %769 : tensor<2x32x10x1024xf64>, tensor<2x32x10x1024xf64>) outs(%765 : tensor<2x32x10x1024xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x10x1024xf64>
%771 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%770 : tensor<2x32x10x1024xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%772 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%771 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_962 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%773 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%772 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%774 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%764 : tensor<2x32x10x1024xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%775 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%774 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_963 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%776 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%773, %cst_938 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%777 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%776 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%778 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1173, %775 : tensor<2x32x10x1024xf16>, tensor<2x32x1x1xf32>) outs(%763 : tensor<2x32x10x1024xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x10x1024xf32>
%779 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%778, %777 : tensor<2x32x10x1024xf32>, tensor<2x32x1x1xf32>) outs(%763 : tensor<2x32x10x1024xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x10x1024xf32>
%collapsed_1174 = tensor.collapse_shape %779 [[0], [1, 2], [3]] : tensor<2x32x10x1024xf32> into tensor<2x320x1024xf32>
%expanded_1175 = tensor.expand_shape %collapsed_1174 [[0], [1], [2, 3]] : tensor<2x320x1024xf32> into tensor<2x320x32x32xf32>
%expanded_1176 = tensor.expand_shape %cst_831 [[0, 1, 2, 3]] : tensor<320xf16> into tensor<1x320x1x1xf16>
%expanded_1177 = tensor.expand_shape %cst_830 [[0, 1, 2, 3]] : tensor<320xf16> into tensor<1x320x1x1xf16>
%780 = tensor.empty() : tensor<2x320x32x32xf32>
%781 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1175, %expanded_1177 : tensor<2x320x32x32xf32>, tensor<1x320x1x1xf16>) outs(%780 : tensor<2x320x32x32xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x320x32x32xf32>
%782 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%781, %expanded_1176 : tensor<2x320x32x32xf32>, tensor<1x320x1x1xf16>) outs(%780 : tensor<2x320x32x32xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x320x32x32xf32>
%783 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%782 : tensor<2x320x32x32xf32>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x320x32x32xf16>
%784 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%783 : tensor<2x320x32x32xf16>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.negf %in : f16
%6235 = math.exp %6234 : f16
%6236 = arith.addf %6235, %cst_947 : f16
%6237 = arith.divf %cst_947, %6236 : f16
linalg.yield %6237 : f16
} -> tensor<2x320x32x32xf16>
%785 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%784, %783 : tensor<2x320x32x32xf16>, tensor<2x320x32x32xf16>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x32x32xf16>
%expanded_1178 = tensor.expand_shape %785 [[0], [1, 2], [3], [4]] : tensor<2x320x32x32xf16> into tensor<2x20x16x32x32xf16>
%786 = tensor.empty() : tensor<2x20x16x32x32xf16>
%787 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1178 : tensor<2x20x16x32x32xf16>) outs(%786 : tensor<2x20x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x20x16x32x32xf16>
%788 = tensor.empty() : tensor<2x20x1x32x32xi64>
%789 = linalg.fill ins(%c0_i64 : i64) outs(%788 : tensor<2x20x1x32x32xi64>) -> tensor<2x20x1x32x32xi64>
%790 = tensor.empty() : tensor<2x20x1x32x32xf16>
%791 = linalg.fill ins(%cst_948 : f16) outs(%790 : tensor<2x20x1x32x32xf16>) -> tensor<2x20x1x32x32xf16>
%792:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%787 : tensor<2x20x16x32x32xf16>) outs(%791, %789 : tensor<2x20x1x32x32xf16>, tensor<2x20x1x32x32xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x20x1x32x32xf16>, tensor<2x20x1x32x32xi64>)
%793 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%792#0 : tensor<2x20x1x32x32xf16>) outs(%786 : tensor<2x20x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x32x32xf16>
%collapsed_1179 = tensor.collapse_shape %793 [[0], [1, 2], [3], [4]] : tensor<2x20x16x32x32xf16> into tensor<2x320x32x32xf16>
%794 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1179 : tensor<2x320x32x32xf16>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x32x32xf16>
%795 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%785, %794 : tensor<2x320x32x32xf16>, tensor<2x320x32x32xf16>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x32x32xf16>
%796 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%795, %cst_935 : tensor<2x320x32x32xf16>, tensor<f16>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x32x32xf16>
%797 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%796 : tensor<2x320x32x32xf16>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x320x32x32xf16>
%798 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%797 : tensor<2x320x32x32xf16>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x320x32x32xf16>
%799 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%798, %cst_935 : tensor<2x320x32x32xf16>, tensor<f16>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x32x32xf16>
%800 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%799, %794 : tensor<2x320x32x32xf16>, tensor<2x320x32x32xf16>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x32x32xf16>
%801 = tensor.empty() : tensor<640x20x16x3x3xf16>
%802 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_829 : tensor<640x20x1x3x3xf16>) outs(%801 : tensor<640x20x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x20x16x3x3xf16>
%collapsed_1180 = tensor.collapse_shape %802 [[0], [1, 2], [3], [4]] : tensor<640x20x16x3x3xf16> into tensor<640x320x3x3xf16>
%803 = tensor.empty() : tensor<640x320x3x3xf16>
%804 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_828, %collapsed_1180 : tensor<640x320x3x3xi8>, tensor<640x320x3x3xf16>) outs(%803 : tensor<640x320x3x3xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x320x3x3xf16>
%padded_1181 = tensor.pad %800 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x320x32x32xf16> to tensor<2x320x34x34xf16>
%805 = tensor.empty() : tensor<2x640x32x32xf16>
%806 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_827 : tensor<640xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%807 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1181, %804 : tensor<2x320x34x34xf16>, tensor<640x320x3x3xf16>) outs(%806 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16>
%808 = tensor.empty() : tensor<1280x640xf16>
%809 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_826 : tensor<640x1280xf16>) outs(%808 : tensor<1280x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x640xf16>
%810 = tensor.empty() : tensor<2x640xf16>
%811 = linalg.fill ins(%cst_946 : f16) outs(%810 : tensor<2x640xf16>) -> tensor<2x640xf16>
%812 = linalg.matmul ins(%81, %809 : tensor<2x1280xf16>, tensor<1280x640xf16>) outs(%811 : tensor<2x640xf16>) -> tensor<2x640xf16>
%813 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_825, %812 : tensor<640xf16>, tensor<2x640xf16>) outs(%810 : tensor<2x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640xf16>
%expanded_1182 = tensor.expand_shape %813 [[0], [1, 2, 3]] : tensor<2x640xf16> into tensor<2x640x1x1xf16>
%814 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%807, %expanded_1182 : tensor<2x640x32x32xf16>, tensor<2x640x1x1xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%collapsed_1183 = tensor.collapse_shape %814 [[0], [1], [2, 3]] : tensor<2x640x32x32xf16> into tensor<2x640x1024xf16>
%expanded_1184 = tensor.expand_shape %collapsed_1183 [[0], [1, 2], [3]] : tensor<2x640x1024xf16> into tensor<2x32x20x1024xf16>
%815 = tensor.empty() : tensor<2x32x20x1024xf32>
%816 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1184 : tensor<2x32x20x1024xf16>) outs(%815 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x20x1024xf32>
%817 = tensor.empty() : tensor<2x32x20x1024xf64>
%818 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%816 : tensor<2x32x20x1024xf32>) outs(%817 : tensor<2x32x20x1024xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x1024xf64>
%819 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%818 : tensor<2x32x20x1024xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%820 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%819 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_964 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%821 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%818, %820 : tensor<2x32x20x1024xf64>, tensor<2x32x1x1xf64>) outs(%817 : tensor<2x32x20x1024xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x1024xf64>
%822 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%821, %821 : tensor<2x32x20x1024xf64>, tensor<2x32x20x1024xf64>) outs(%817 : tensor<2x32x20x1024xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x1024xf64>
%823 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%822 : tensor<2x32x20x1024xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%824 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%823 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_964 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%825 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%824 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%826 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%816 : tensor<2x32x20x1024xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%827 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%826 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_965 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%828 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%825, %cst_938 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%829 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%828 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%830 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1184, %827 : tensor<2x32x20x1024xf16>, tensor<2x32x1x1xf32>) outs(%815 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x20x1024xf32>
%831 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%830, %829 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%815 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x20x1024xf32>
%collapsed_1185 = tensor.collapse_shape %831 [[0], [1, 2], [3]] : tensor<2x32x20x1024xf32> into tensor<2x640x1024xf32>
%expanded_1186 = tensor.expand_shape %collapsed_1185 [[0], [1], [2, 3]] : tensor<2x640x1024xf32> into tensor<2x640x32x32xf32>
%expanded_1187 = tensor.expand_shape %cst_824 [[0, 1, 2, 3]] : tensor<640xf16> into tensor<1x640x1x1xf16>
%expanded_1188 = tensor.expand_shape %cst_823 [[0, 1, 2, 3]] : tensor<640xf16> into tensor<1x640x1x1xf16>
%832 = tensor.empty() : tensor<2x640x32x32xf32>
%833 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1186, %expanded_1188 : tensor<2x640x32x32xf32>, tensor<1x640x1x1xf16>) outs(%832 : tensor<2x640x32x32xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x640x32x32xf32>
%834 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%833, %expanded_1187 : tensor<2x640x32x32xf32>, tensor<1x640x1x1xf16>) outs(%832 : tensor<2x640x32x32xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x640x32x32xf32>
%835 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%834 : tensor<2x640x32x32xf32>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%836 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%835 : tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.negf %in : f16
%6235 = math.exp %6234 : f16
%6236 = arith.addf %6235, %cst_947 : f16
%6237 = arith.divf %cst_947, %6236 : f16
linalg.yield %6237 : f16
} -> tensor<2x640x32x32xf16>
%837 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%836, %835 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%expanded_1189 = tensor.expand_shape %837 [[0], [1, 2], [3], [4]] : tensor<2x640x32x32xf16> into tensor<2x40x16x32x32xf16>
%838 = tensor.empty() : tensor<2x40x16x32x32xf16>
%839 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1189 : tensor<2x40x16x32x32xf16>) outs(%838 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x40x16x32x32xf16>
%840 = tensor.empty() : tensor<2x40x1x32x32xi64>
%841 = linalg.fill ins(%c0_i64 : i64) outs(%840 : tensor<2x40x1x32x32xi64>) -> tensor<2x40x1x32x32xi64>
%842 = tensor.empty() : tensor<2x40x1x32x32xf16>
%843 = linalg.fill ins(%cst_948 : f16) outs(%842 : tensor<2x40x1x32x32xf16>) -> tensor<2x40x1x32x32xf16>
%844:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%839 : tensor<2x40x16x32x32xf16>) outs(%843, %841 : tensor<2x40x1x32x32xf16>, tensor<2x40x1x32x32xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x40x1x32x32xf16>, tensor<2x40x1x32x32xi64>)
%845 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%844#0 : tensor<2x40x1x32x32xf16>) outs(%838 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x40x16x32x32xf16>
%collapsed_1190 = tensor.collapse_shape %845 [[0], [1, 2], [3], [4]] : tensor<2x40x16x32x32xf16> into tensor<2x640x32x32xf16>
%846 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1190 : tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%847 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%837, %846 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%848 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%847, %cst_935 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%849 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%848 : tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%850 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%849 : tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x640x32x32xf16>
%851 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%850, %cst_935 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%852 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%851, %846 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%853 = tensor.empty() : tensor<640x40x16x3x3xf16>
%854 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_822 : tensor<640x40x1x3x3xf16>) outs(%853 : tensor<640x40x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16x3x3xf16>
%collapsed_1191 = tensor.collapse_shape %854 [[0], [1, 2], [3], [4]] : tensor<640x40x16x3x3xf16> into tensor<640x640x3x3xf16>
%855 = tensor.empty() : tensor<640x640x3x3xf16>
%856 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_821, %collapsed_1191 : tensor<640x640x3x3xi8>, tensor<640x640x3x3xf16>) outs(%855 : tensor<640x640x3x3xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640x3x3xf16>
%padded_1192 = tensor.pad %852 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x640x32x32xf16> to tensor<2x640x34x34xf16>
%857 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_820 : tensor<640xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%858 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1192, %856 : tensor<2x640x34x34xf16>, tensor<640x640x3x3xf16>) outs(%857 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16>
%expanded_1193 = tensor.expand_shape %762 [[0], [1, 2], [3], [4]] : tensor<2x320x32x32xf16> into tensor<2x20x16x32x32xf16>
%859 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1193 : tensor<2x20x16x32x32xf16>) outs(%786 : tensor<2x20x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x20x16x32x32xf16>
%860:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%859 : tensor<2x20x16x32x32xf16>) outs(%791, %789 : tensor<2x20x1x32x32xf16>, tensor<2x20x1x32x32xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x20x1x32x32xf16>, tensor<2x20x1x32x32xi64>)
%861 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%860#0 : tensor<2x20x1x32x32xf16>) outs(%786 : tensor<2x20x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x16x32x32xf16>
%collapsed_1194 = tensor.collapse_shape %861 [[0], [1, 2], [3], [4]] : tensor<2x20x16x32x32xf16> into tensor<2x320x32x32xf16>
%862 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1194 : tensor<2x320x32x32xf16>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x32x32xf16>
%863 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%762, %862 : tensor<2x320x32x32xf16>, tensor<2x320x32x32xf16>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x32x32xf16>
%864 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%863, %cst_935 : tensor<2x320x32x32xf16>, tensor<f16>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x32x32xf16>
%865 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%864 : tensor<2x320x32x32xf16>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x320x32x32xf16>
%866 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%865 : tensor<2x320x32x32xf16>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x320x32x32xf16>
%867 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%866, %cst_935 : tensor<2x320x32x32xf16>, tensor<f16>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x32x32xf16>
%868 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%867, %862 : tensor<2x320x32x32xf16>, tensor<2x320x32x32xf16>) outs(%760 : tensor<2x320x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x320x32x32xf16>
%869 = tensor.empty() : tensor<640x20x16x1x1xf16>
%870 = linalg.generic {indexing_maps = [#map25, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_819 : tensor<640x20x1x1x1xf16>) outs(%869 : tensor<640x20x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x20x16x1x1xf16>
%871 = linalg.generic {indexing_maps = [#map26, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%870 : tensor<640x20x16x1x1xf16>) outs(%869 : tensor<640x20x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x20x16x1x1xf16>
%collapsed_1195 = tensor.collapse_shape %871 [[0], [1, 2], [3], [4]] : tensor<640x20x16x1x1xf16> into tensor<640x320x1x1xf16>
%872 = tensor.empty() : tensor<640x320x1x1xf16>
%873 = linalg.generic {indexing_maps = [#map11, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_818, %collapsed_1195 : tensor<640x320x1x1xi8>, tensor<640x320x1x1xf16>) outs(%872 : tensor<640x320x1x1xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x320x1x1xf16>
%874 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_817 : tensor<640xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%875 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%868, %873 : tensor<2x320x32x32xf16>, tensor<640x320x1x1xf16>) outs(%874 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16>
%876 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%875, %858 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%877 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%876, %cst_924 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%collapsed_1196 = tensor.collapse_shape %877 [[0], [1], [2, 3]] : tensor<2x640x32x32xf16> into tensor<2x640x1024xf16>
%expanded_1197 = tensor.expand_shape %collapsed_1196 [[0], [1, 2], [3]] : tensor<2x640x1024xf16> into tensor<2x32x20x1024xf16>
%878 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1197 : tensor<2x32x20x1024xf16>) outs(%815 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x20x1024xf32>
%879 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%878 : tensor<2x32x20x1024xf32>) outs(%817 : tensor<2x32x20x1024xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x1024xf64>
%880 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%879 : tensor<2x32x20x1024xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%881 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%880 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_964 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%882 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%879, %881 : tensor<2x32x20x1024xf64>, tensor<2x32x1x1xf64>) outs(%817 : tensor<2x32x20x1024xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x1024xf64>
%883 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%882, %882 : tensor<2x32x20x1024xf64>, tensor<2x32x20x1024xf64>) outs(%817 : tensor<2x32x20x1024xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x1024xf64>
%884 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%883 : tensor<2x32x20x1024xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%885 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%884 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_964 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%886 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%885 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%887 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%878 : tensor<2x32x20x1024xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%888 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%887 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_965 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%889 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%886, %cst_923 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%890 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%889 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%891 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1197, %888 : tensor<2x32x20x1024xf16>, tensor<2x32x1x1xf32>) outs(%815 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x20x1024xf32>
%892 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%891, %890 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%815 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x20x1024xf32>
%collapsed_1198 = tensor.collapse_shape %892 [[0], [1, 2], [3]] : tensor<2x32x20x1024xf32> into tensor<2x640x1024xf32>
%expanded_1199 = tensor.expand_shape %collapsed_1198 [[0], [1], [2, 3]] : tensor<2x640x1024xf32> into tensor<2x640x32x32xf32>
%expanded_1200 = tensor.expand_shape %cst_816 [[0, 1, 2, 3]] : tensor<640xf16> into tensor<1x640x1x1xf16>
%expanded_1201 = tensor.expand_shape %cst_815 [[0, 1, 2, 3]] : tensor<640xf16> into tensor<1x640x1x1xf16>
%893 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1199, %expanded_1201 : tensor<2x640x32x32xf32>, tensor<1x640x1x1xf16>) outs(%832 : tensor<2x640x32x32xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x640x32x32xf32>
%894 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%893, %expanded_1200 : tensor<2x640x32x32xf32>, tensor<1x640x1x1xf16>) outs(%832 : tensor<2x640x32x32xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x640x32x32xf32>
%895 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%894 : tensor<2x640x32x32xf32>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%896 = tensor.empty() : tensor<2x32x32x640xf16>
%897 = linalg.generic {indexing_maps = [#map1, #map16], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%895 : tensor<2x640x32x32xf16>) outs(%896 : tensor<2x32x32x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x32x32x640xf16>
%collapsed_1202 = tensor.collapse_shape %897 [[0], [1, 2], [3]] : tensor<2x32x32x640xf16> into tensor<2x1024x640xf16>
%expanded_1203 = tensor.expand_shape %collapsed_1202 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%898 = tensor.empty() : tensor<2x1024x40x16xf16>
%899 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1203 : tensor<2x1024x40x16xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x40x16xf16>
%900 = tensor.empty() : tensor<2x1024x40x1xi64>
%901 = linalg.fill ins(%c0_i64 : i64) outs(%900 : tensor<2x1024x40x1xi64>) -> tensor<2x1024x40x1xi64>
%902 = tensor.empty() : tensor<2x1024x40x1xf16>
%903 = linalg.fill ins(%cst_948 : f16) outs(%902 : tensor<2x1024x40x1xf16>) -> tensor<2x1024x40x1xf16>
%904:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%899 : tensor<2x1024x40x16xf16>) outs(%903, %901 : tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>)
%905 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%904#0 : tensor<2x1024x40x1xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1204 = tensor.collapse_shape %905 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%906 = tensor.empty() : tensor<2x1024x640xf16>
%907 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1204 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%908 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1202, %907 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%909 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%908, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%910 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%909 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%911 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%910 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1024x640xf16>
%912 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%911, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%913 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%912, %907 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%914 = tensor.empty() : tensor<640x40x16xf16>
%915 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_814 : tensor<640x40x1xf16>) outs(%914 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1205 = tensor.collapse_shape %915 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%916 = tensor.empty() : tensor<640x640xf16>
%917 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_813, %collapsed_1205 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640xf16>
%918 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%917 : tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1206 = tensor.collapse_shape %913 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%919 = tensor.empty() : tensor<2048x640xf16>
%920 = linalg.fill ins(%cst_946 : f16) outs(%919 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%921 = linalg.matmul ins(%collapsed_1206, %918 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%922 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_812, %921 : tensor<640xf16>, tensor<2048x640xf16>) outs(%919 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2048x640xf16>
%expanded_1207 = tensor.expand_shape %922 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16>
%923 = tensor.empty() : tensor<2x1024x640xf32>
%924 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1207 : tensor<2x1024x640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x1024x640xf32>
%925 = tensor.empty() : tensor<2x1024x640xf64>
%926 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%924 : tensor<2x1024x640xf32>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%927 = tensor.empty() : tensor<2x1024x1xf64>
%928 = linalg.fill ins(%cst_951 : f64) outs(%927 : tensor<2x1024x1xf64>) -> tensor<2x1024x1xf64>
%929 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%926 : tensor<2x1024x640xf64>) outs(%928 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%930 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%929 : tensor<2x1024x1xf64>) outs(%927 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_966 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%931 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%926, %930 : tensor<2x1024x640xf64>, tensor<2x1024x1xf64>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%932 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%931, %931 : tensor<2x1024x640xf64>, tensor<2x1024x640xf64>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%933 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%932 : tensor<2x1024x640xf64>) outs(%928 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%934 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%933 : tensor<2x1024x1xf64>) outs(%927 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_966 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%935 = tensor.empty() : tensor<2x1024x1xf32>
%936 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%934 : tensor<2x1024x1xf64>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%937 = linalg.fill ins(%cst_952 : f32) outs(%935 : tensor<2x1024x1xf32>) -> tensor<2x1024x1xf32>
%938 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%924 : tensor<2x1024x640xf32>) outs(%937 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%939 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%938 : tensor<2x1024x1xf32>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_967 : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%940 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%936, %cst_938 : tensor<2x1024x1xf32>, tensor<f16>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x1xf32>
%941 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%940 : tensor<2x1024x1xf32>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%942 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1207, %939 : tensor<2x1024x640xf16>, tensor<2x1024x1xf32>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%943 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%942, %941 : tensor<2x1024x640xf32>, tensor<2x1024x1xf32>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x640xf32>
%944 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%943, %cst_811 : tensor<2x1024x640xf32>, tensor<640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%945 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%944, %cst_810 : tensor<2x1024x640xf32>, tensor<640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%946 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%945 : tensor<2x1024x640xf32>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%expanded_1208 = tensor.expand_shape %946 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%947 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1208 : tensor<2x1024x40x16xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x40x16xf16>
%948:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%947 : tensor<2x1024x40x16xf16>) outs(%903, %901 : tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>)
%949 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%948#0 : tensor<2x1024x40x1xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1209 = tensor.collapse_shape %949 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%950 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1209 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%951 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%946, %950 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%952 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%951, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%953 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%952 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%954 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%953 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1024x640xf16>
%955 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%954, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%956 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%955, %950 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%957 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_809 : tensor<640x40x1xf16>) outs(%914 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1210 = tensor.collapse_shape %957 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%958 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_808, %collapsed_1210 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640xf16>
%959 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%958 : tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1211 = tensor.collapse_shape %956 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%960 = linalg.matmul ins(%collapsed_1211, %959 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%expanded_1212 = tensor.expand_shape %960 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x10x64xf16>
%961 = tensor.empty() : tensor<2x10x1024x64xf16>
%962 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1212 : tensor<2x1024x10x64xf16>) outs(%961 : tensor<2x10x1024x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x10x1024x64xf16>
%collapsed_1213 = tensor.collapse_shape %962 [[0, 1], [2], [3]] : tensor<2x10x1024x64xf16> into tensor<20x1024x64xf16>
%963 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_807 : tensor<640x40x1xf16>) outs(%914 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1214 = tensor.collapse_shape %963 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%964 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_806, %collapsed_1214 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640xf16>
%965 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%964 : tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%966 = linalg.matmul ins(%collapsed_1211, %965 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%967 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_805 : tensor<640x40x1xf16>) outs(%914 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1215 = tensor.collapse_shape %967 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%968 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_804, %collapsed_1215 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640xf16>
%969 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%968 : tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%970 = linalg.matmul ins(%collapsed_1211, %969 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%expanded_1216 = tensor.expand_shape %966 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x10x64xf16>
%971 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1216 : tensor<2x1024x10x64xf16>) outs(%961 : tensor<2x10x1024x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x10x1024x64xf16>
%collapsed_1217 = tensor.collapse_shape %971 [[0, 1], [2], [3]] : tensor<2x10x1024x64xf16> into tensor<20x1024x64xf16>
%expanded_1218 = tensor.expand_shape %970 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x10x64xf16>
%972 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1218 : tensor<2x1024x10x64xf16>) outs(%961 : tensor<2x10x1024x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x10x1024x64xf16>
%collapsed_1219 = tensor.collapse_shape %972 [[0, 1], [2], [3]] : tensor<2x10x1024x64xf16> into tensor<20x1024x64xf16>
%973 = tensor.empty() : tensor<20x1024x1024xf16>
%974 = linalg.fill ins(%cst_946 : f16) outs(%973 : tensor<20x1024x1024xf16>) -> tensor<20x1024x1024xf16>
%975 = tensor.empty() : tensor<20x64x1024xf16>
%976 = linalg.generic {indexing_maps = [#map18, #map23], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1217 : tensor<20x1024x64xf16>) outs(%975 : tensor<20x64x1024xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<20x64x1024xf16>
%977 = linalg.batch_matmul ins(%collapsed_1213, %976 : tensor<20x1024x64xf16>, tensor<20x64x1024xf16>) outs(%974 : tensor<20x1024x1024xf16>) -> tensor<20x1024x1024xf16>
%978 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%977 : tensor<20x1024x1024xf16>) outs(%973 : tensor<20x1024x1024xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.mulf %in, %cst_961 : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x1024xf16>
%979 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%978, %974 : tensor<20x1024x1024xf16>, tensor<20x1024x1024xf16>) outs(%973 : tensor<20x1024x1024xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in_2625, %cst_946 : f16
%6235 = arith.addf %in, %6234 : f16
linalg.yield %6235 : f16
} -> tensor<20x1024x1024xf16>
%980 = tensor.empty() : tensor<20x1024x1xi64>
%981 = linalg.fill ins(%c0_i64 : i64) outs(%980 : tensor<20x1024x1xi64>) -> tensor<20x1024x1xi64>
%982 = tensor.empty() : tensor<20x1024x1xf16>
%983 = linalg.fill ins(%cst_948 : f16) outs(%982 : tensor<20x1024x1xf16>) -> tensor<20x1024x1xf16>
%984:2 = linalg.generic {indexing_maps = [#map18, #map20, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%979 : tensor<20x1024x1024xf16>) outs(%983, %981 : tensor<20x1024x1xf16>, tensor<20x1024x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<20x1024x1xf16>, tensor<20x1024x1xi64>)
%985 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%979, %984#0 : tensor<20x1024x1024xf16>, tensor<20x1024x1xf16>) outs(%973 : tensor<20x1024x1024xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x1024xf16>
%986 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%985 : tensor<20x1024x1024xf16>) outs(%973 : tensor<20x1024x1024xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.exp %in : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x1024xf16>
%987 = linalg.fill ins(%cst_946 : f16) outs(%982 : tensor<20x1024x1xf16>) -> tensor<20x1024x1xf16>
%988 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%986 : tensor<20x1024x1024xf16>) outs(%987 : tensor<20x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.addf %in, %out : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x1xf16>
%989 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%986, %988 : tensor<20x1024x1024xf16>, tensor<20x1024x1xf16>) outs(%973 : tensor<20x1024x1024xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x1024xf16>
%990 = tensor.empty() : tensor<20x1024x64xf16>
%991 = linalg.fill ins(%cst_946 : f16) outs(%990 : tensor<20x1024x64xf16>) -> tensor<20x1024x64xf16>
%992 = linalg.batch_matmul ins(%989, %collapsed_1219 : tensor<20x1024x1024xf16>, tensor<20x1024x64xf16>) outs(%991 : tensor<20x1024x64xf16>) -> tensor<20x1024x64xf16>
%expanded_1220 = tensor.expand_shape %992 [[0, 1], [2], [3]] : tensor<20x1024x64xf16> into tensor<2x10x1024x64xf16>
%993 = tensor.empty() : tensor<2x1024x10x64xf16>
%994 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1220 : tensor<2x10x1024x64xf16>) outs(%993 : tensor<2x1024x10x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x10x64xf16>
%collapsed_1221 = tensor.collapse_shape %994 [[0], [1], [2, 3]] : tensor<2x1024x10x64xf16> into tensor<2x1024x640xf16>
%expanded_1222 = tensor.expand_shape %collapsed_1221 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%995 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1222 : tensor<2x1024x40x16xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x40x16xf16>
%996:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%995 : tensor<2x1024x40x16xf16>) outs(%903, %901 : tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>)
%997 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%996#0 : tensor<2x1024x40x1xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1223 = tensor.collapse_shape %997 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%998 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1223 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%999 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1221, %998 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1000 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%999, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1001 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1000 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1002 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1001 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1024x640xf16>
%1003 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1002, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1004 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1003, %998 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1005 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_803 : tensor<640x40x1xf16>) outs(%914 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1224 = tensor.collapse_shape %1005 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1006 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_802, %collapsed_1224 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640xf16>
%1007 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1006 : tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1225 = tensor.collapse_shape %1004 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1008 = linalg.matmul ins(%collapsed_1225, %1007 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1009 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_801, %1008 : tensor<640xf16>, tensor<2048x640xf16>) outs(%919 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2048x640xf16>
%expanded_1226 = tensor.expand_shape %1009 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16>
%1010 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1226, %expanded_1207 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1011 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1010 : tensor<2x1024x640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x1024x640xf32>
%1012 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1011 : tensor<2x1024x640xf32>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%1013 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1012 : tensor<2x1024x640xf64>) outs(%928 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1014 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1013 : tensor<2x1024x1xf64>) outs(%927 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_966 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1015 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1012, %1014 : tensor<2x1024x640xf64>, tensor<2x1024x1xf64>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%1016 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1015, %1015 : tensor<2x1024x640xf64>, tensor<2x1024x640xf64>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%1017 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1016 : tensor<2x1024x640xf64>) outs(%928 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1018 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1017 : tensor<2x1024x1xf64>) outs(%927 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_966 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1019 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1018 : tensor<2x1024x1xf64>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1020 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1011 : tensor<2x1024x640xf32>) outs(%937 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1021 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1020 : tensor<2x1024x1xf32>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_967 : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1022 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1019, %cst_938 : tensor<2x1024x1xf32>, tensor<f16>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x1xf32>
%1023 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1022 : tensor<2x1024x1xf32>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1024 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1010, %1021 : tensor<2x1024x640xf16>, tensor<2x1024x1xf32>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%1025 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1024, %1023 : tensor<2x1024x640xf32>, tensor<2x1024x1xf32>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x640xf32>
%1026 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1025, %cst_800 : tensor<2x1024x640xf32>, tensor<640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%1027 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1026, %cst_799 : tensor<2x1024x640xf32>, tensor<640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%1028 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1027 : tensor<2x1024x640xf32>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%expanded_1227 = tensor.expand_shape %1028 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%1029 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1227 : tensor<2x1024x40x16xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x40x16xf16>
%1030:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1029 : tensor<2x1024x40x16xf16>) outs(%903, %901 : tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>)
%1031 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1030#0 : tensor<2x1024x40x1xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1228 = tensor.collapse_shape %1031 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1032 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1228 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1033 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1028, %1032 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1034 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1033, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1035 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1034 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1036 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1035 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1024x640xf16>
%1037 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1036, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1038 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1037, %1032 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1039 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_798 : tensor<640x40x1xf16>) outs(%914 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1229 = tensor.collapse_shape %1039 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1040 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_797, %collapsed_1229 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640xf16>
%1041 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1040 : tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1230 = tensor.collapse_shape %1038 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1042 = linalg.matmul ins(%collapsed_1230, %1041 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%expanded_1231 = tensor.expand_shape %1042 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x10x64xf16>
%1043 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1231 : tensor<2x1024x10x64xf16>) outs(%961 : tensor<2x10x1024x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x10x1024x64xf16>
%collapsed_1232 = tensor.collapse_shape %1043 [[0, 1], [2], [3]] : tensor<2x10x1024x64xf16> into tensor<20x1024x64xf16>
%1044 = tensor.empty() : tensor<640x64x16xf16>
%1045 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_796 : tensor<640x64x1xf16>) outs(%1044 : tensor<640x64x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x64x16xf16>
%collapsed_1233 = tensor.collapse_shape %1045 [[0], [1, 2]] : tensor<640x64x16xf16> into tensor<640x1024xf16>
%1046 = tensor.empty() : tensor<640x1024xf16>
%1047 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_795, %collapsed_1233 : tensor<640x1024xi8>, tensor<640x1024xf16>) outs(%1046 : tensor<640x1024xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x1024xf16>
%1048 = tensor.empty() : tensor<1024x640xf16>
%1049 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1047 : tensor<640x1024xf16>) outs(%1048 : tensor<1024x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1024x640xf16>
%1050 = tensor.empty() : tensor<154x640xf16>
%1051 = linalg.fill ins(%cst_946 : f16) outs(%1050 : tensor<154x640xf16>) -> tensor<154x640xf16>
%1052 = linalg.matmul ins(%collapsed_1050, %1049 : tensor<154x1024xf16>, tensor<1024x640xf16>) outs(%1051 : tensor<154x640xf16>) -> tensor<154x640xf16>
%1053 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_794 : tensor<640x64x1xf16>) outs(%1044 : tensor<640x64x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x64x16xf16>
%collapsed_1234 = tensor.collapse_shape %1053 [[0], [1, 2]] : tensor<640x64x16xf16> into tensor<640x1024xf16>
%1054 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_793, %collapsed_1234 : tensor<640x1024xi8>, tensor<640x1024xf16>) outs(%1046 : tensor<640x1024xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x1024xf16>
%1055 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1054 : tensor<640x1024xf16>) outs(%1048 : tensor<1024x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1024x640xf16>
%1056 = linalg.matmul ins(%collapsed_1050, %1055 : tensor<154x1024xf16>, tensor<1024x640xf16>) outs(%1051 : tensor<154x640xf16>) -> tensor<154x640xf16>
%expanded_1235 = tensor.expand_shape %1052 [[0, 1], [2, 3]] : tensor<154x640xf16> into tensor<2x77x10x64xf16>
%1057 = tensor.empty() : tensor<2x10x77x64xf16>
%1058 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1235 : tensor<2x77x10x64xf16>) outs(%1057 : tensor<2x10x77x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x10x77x64xf16>
%collapsed_1236 = tensor.collapse_shape %1058 [[0, 1], [2], [3]] : tensor<2x10x77x64xf16> into tensor<20x77x64xf16>
%expanded_1237 = tensor.expand_shape %1056 [[0, 1], [2, 3]] : tensor<154x640xf16> into tensor<2x77x10x64xf16>
%1059 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1237 : tensor<2x77x10x64xf16>) outs(%1057 : tensor<2x10x77x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x10x77x64xf16>
%collapsed_1238 = tensor.collapse_shape %1059 [[0, 1], [2], [3]] : tensor<2x10x77x64xf16> into tensor<20x77x64xf16>
%1060 = tensor.empty() : tensor<20x1024x77xf16>
%1061 = linalg.fill ins(%cst_946 : f16) outs(%1060 : tensor<20x1024x77xf16>) -> tensor<20x1024x77xf16>
%1062 = tensor.empty() : tensor<20x64x77xf16>
%1063 = linalg.generic {indexing_maps = [#map18, #map23], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1236 : tensor<20x77x64xf16>) outs(%1062 : tensor<20x64x77xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<20x64x77xf16>
%1064 = linalg.batch_matmul ins(%collapsed_1232, %1063 : tensor<20x1024x64xf16>, tensor<20x64x77xf16>) outs(%1061 : tensor<20x1024x77xf16>) -> tensor<20x1024x77xf16>
%1065 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1064 : tensor<20x1024x77xf16>) outs(%1060 : tensor<20x1024x77xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.mulf %in, %cst_961 : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x77xf16>
%1066 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1065, %1061 : tensor<20x1024x77xf16>, tensor<20x1024x77xf16>) outs(%1060 : tensor<20x1024x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in_2625, %cst_946 : f16
%6235 = arith.addf %in, %6234 : f16
linalg.yield %6235 : f16
} -> tensor<20x1024x77xf16>
%1067:2 = linalg.generic {indexing_maps = [#map18, #map20, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1066 : tensor<20x1024x77xf16>) outs(%983, %981 : tensor<20x1024x1xf16>, tensor<20x1024x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<20x1024x1xf16>, tensor<20x1024x1xi64>)
%1068 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1066, %1067#0 : tensor<20x1024x77xf16>, tensor<20x1024x1xf16>) outs(%1060 : tensor<20x1024x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x77xf16>
%1069 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1068 : tensor<20x1024x77xf16>) outs(%1060 : tensor<20x1024x77xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.exp %in : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x77xf16>
%1070 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1069 : tensor<20x1024x77xf16>) outs(%987 : tensor<20x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.addf %in, %out : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x1xf16>
%1071 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1069, %1070 : tensor<20x1024x77xf16>, tensor<20x1024x1xf16>) outs(%1060 : tensor<20x1024x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x77xf16>
%1072 = linalg.batch_matmul ins(%1071, %collapsed_1238 : tensor<20x1024x77xf16>, tensor<20x77x64xf16>) outs(%991 : tensor<20x1024x64xf16>) -> tensor<20x1024x64xf16>
%expanded_1239 = tensor.expand_shape %1072 [[0, 1], [2], [3]] : tensor<20x1024x64xf16> into tensor<2x10x1024x64xf16>
%1073 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1239 : tensor<2x10x1024x64xf16>) outs(%993 : tensor<2x1024x10x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x10x64xf16>
%collapsed_1240 = tensor.collapse_shape %1073 [[0], [1], [2, 3]] : tensor<2x1024x10x64xf16> into tensor<2x1024x640xf16>
%expanded_1241 = tensor.expand_shape %collapsed_1240 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%1074 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1241 : tensor<2x1024x40x16xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x40x16xf16>
%1075:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1074 : tensor<2x1024x40x16xf16>) outs(%903, %901 : tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>)
%1076 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1075#0 : tensor<2x1024x40x1xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1242 = tensor.collapse_shape %1076 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1077 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1242 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1078 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1240, %1077 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1079 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1078, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1080 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1079 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1081 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1080 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1024x640xf16>
%1082 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1081, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1083 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1082, %1077 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1084 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_792 : tensor<640x40x1xf16>) outs(%914 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1243 = tensor.collapse_shape %1084 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1085 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_791, %collapsed_1243 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640xf16>
%1086 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1085 : tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1244 = tensor.collapse_shape %1083 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1087 = linalg.matmul ins(%collapsed_1244, %1086 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1088 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_790, %1087 : tensor<640xf16>, tensor<2048x640xf16>) outs(%919 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2048x640xf16>
%expanded_1245 = tensor.expand_shape %1088 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16>
%1089 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1245, %1010 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1090 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1089 : tensor<2x1024x640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x1024x640xf32>
%1091 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1090 : tensor<2x1024x640xf32>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%1092 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1091 : tensor<2x1024x640xf64>) outs(%928 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1093 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1092 : tensor<2x1024x1xf64>) outs(%927 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_966 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1094 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1091, %1093 : tensor<2x1024x640xf64>, tensor<2x1024x1xf64>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%1095 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1094, %1094 : tensor<2x1024x640xf64>, tensor<2x1024x640xf64>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%1096 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1095 : tensor<2x1024x640xf64>) outs(%928 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1097 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1096 : tensor<2x1024x1xf64>) outs(%927 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_966 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1098 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1097 : tensor<2x1024x1xf64>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1099 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1090 : tensor<2x1024x640xf32>) outs(%937 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1100 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1099 : tensor<2x1024x1xf32>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_967 : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1101 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1098, %cst_938 : tensor<2x1024x1xf32>, tensor<f16>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x1xf32>
%1102 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1101 : tensor<2x1024x1xf32>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1103 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1089, %1100 : tensor<2x1024x640xf16>, tensor<2x1024x1xf32>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%1104 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1103, %1102 : tensor<2x1024x640xf32>, tensor<2x1024x1xf32>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x640xf32>
%1105 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1104, %cst_789 : tensor<2x1024x640xf32>, tensor<640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%1106 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1105, %cst_788 : tensor<2x1024x640xf32>, tensor<640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%1107 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1106 : tensor<2x1024x640xf32>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%expanded_1246 = tensor.expand_shape %1107 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%1108 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1246 : tensor<2x1024x40x16xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x40x16xf16>
%1109:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1108 : tensor<2x1024x40x16xf16>) outs(%903, %901 : tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>)
%1110 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1109#0 : tensor<2x1024x40x1xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1247 = tensor.collapse_shape %1110 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1111 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1247 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1112 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1107, %1111 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1113 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1112, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1114 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1113 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1115 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1114 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1024x640xf16>
%1116 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1115, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1117 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1116, %1111 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1118 = tensor.empty() : tensor<5120x40x16xf16>
%1119 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_787 : tensor<5120x40x1xf16>) outs(%1118 : tensor<5120x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<5120x40x16xf16>
%collapsed_1248 = tensor.collapse_shape %1119 [[0], [1, 2]] : tensor<5120x40x16xf16> into tensor<5120x640xf16>
%1120 = tensor.empty() : tensor<5120x640xf16>
%1121 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_786, %collapsed_1248 : tensor<5120x640xi8>, tensor<5120x640xf16>) outs(%1120 : tensor<5120x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<5120x640xf16>
%1122 = tensor.empty() : tensor<640x5120xf16>
%1123 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1121 : tensor<5120x640xf16>) outs(%1122 : tensor<640x5120xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x5120xf16>
%collapsed_1249 = tensor.collapse_shape %1117 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1124 = tensor.empty() : tensor<2048x5120xf16>
%1125 = linalg.fill ins(%cst_946 : f16) outs(%1124 : tensor<2048x5120xf16>) -> tensor<2048x5120xf16>
%1126 = linalg.matmul ins(%collapsed_1249, %1123 : tensor<2048x640xf16>, tensor<640x5120xf16>) outs(%1125 : tensor<2048x5120xf16>) -> tensor<2048x5120xf16>
%1127 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_785, %1126 : tensor<5120xf16>, tensor<2048x5120xf16>) outs(%1124 : tensor<2048x5120xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2048x5120xf16>
%expanded_1250 = tensor.expand_shape %1127 [[0, 1], [2]] : tensor<2048x5120xf16> into tensor<2x1024x5120xf16>
%extracted_slice_1251 = tensor.extract_slice %expanded_1250[0, 0, 0] [2, 1024, 2560] [1, 1, 1] : tensor<2x1024x5120xf16> to tensor<2x1024x2560xf16>
%extracted_slice_1252 = tensor.extract_slice %expanded_1250[0, 0, 2560] [2, 1024, 2560] [1, 1, 1] : tensor<2x1024x5120xf16> to tensor<2x1024x2560xf16>
%1128 = tensor.empty() : tensor<2x1024x2560xf16>
%1129 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1252 : tensor<2x1024x2560xf16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.sqrt %cst_949 : f16
%6235 = arith.divf %in, %6234 : f16
%6236 = math.erf %6235 : f16
%6237 = arith.addf %6236, %cst_947 : f16
%6238 = arith.mulf %6237, %cst_950 : f16
%6239 = arith.mulf %in, %6238 : f16
linalg.yield %6239 : f16
} -> tensor<2x1024x2560xf16>
%1130 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1251, %1129 : tensor<2x1024x2560xf16>, tensor<2x1024x2560xf16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x2560xf16>
%expanded_1253 = tensor.expand_shape %1130 [[0], [1], [2, 3]] : tensor<2x1024x2560xf16> into tensor<2x1024x160x16xf16>
%1131 = tensor.empty() : tensor<2x1024x160x16xf16>
%1132 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1253 : tensor<2x1024x160x16xf16>) outs(%1131 : tensor<2x1024x160x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x160x16xf16>
%1133 = tensor.empty() : tensor<2x1024x160x1xi64>
%1134 = linalg.fill ins(%c0_i64 : i64) outs(%1133 : tensor<2x1024x160x1xi64>) -> tensor<2x1024x160x1xi64>
%1135 = tensor.empty() : tensor<2x1024x160x1xf16>
%1136 = linalg.fill ins(%cst_948 : f16) outs(%1135 : tensor<2x1024x160x1xf16>) -> tensor<2x1024x160x1xf16>
%1137:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1132 : tensor<2x1024x160x16xf16>) outs(%1136, %1134 : tensor<2x1024x160x1xf16>, tensor<2x1024x160x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x1024x160x1xf16>, tensor<2x1024x160x1xi64>)
%1138 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1137#0 : tensor<2x1024x160x1xf16>) outs(%1131 : tensor<2x1024x160x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x160x16xf16>
%collapsed_1254 = tensor.collapse_shape %1138 [[0], [1], [2, 3]] : tensor<2x1024x160x16xf16> into tensor<2x1024x2560xf16>
%1139 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1254 : tensor<2x1024x2560xf16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x2560xf16>
%1140 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1130, %1139 : tensor<2x1024x2560xf16>, tensor<2x1024x2560xf16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x2560xf16>
%1141 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1140, %cst_935 : tensor<2x1024x2560xf16>, tensor<f16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x2560xf16>
%1142 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1141 : tensor<2x1024x2560xf16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x2560xf16>
%1143 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1142 : tensor<2x1024x2560xf16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1024x2560xf16>
%1144 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1143, %cst_935 : tensor<2x1024x2560xf16>, tensor<f16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x2560xf16>
%1145 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1144, %1139 : tensor<2x1024x2560xf16>, tensor<2x1024x2560xf16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x2560xf16>
%1146 = tensor.empty() : tensor<640x160x16xf16>
%1147 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_784 : tensor<640x160x1xf16>) outs(%1146 : tensor<640x160x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x160x16xf16>
%collapsed_1255 = tensor.collapse_shape %1147 [[0], [1, 2]] : tensor<640x160x16xf16> into tensor<640x2560xf16>
%1148 = tensor.empty() : tensor<640x2560xf16>
%1149 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_783, %collapsed_1255 : tensor<640x2560xi8>, tensor<640x2560xf16>) outs(%1148 : tensor<640x2560xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x2560xf16>
%1150 = tensor.empty() : tensor<2560x640xf16>
%1151 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1149 : tensor<640x2560xf16>) outs(%1150 : tensor<2560x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2560x640xf16>
%collapsed_1256 = tensor.collapse_shape %1145 [[0, 1], [2]] : tensor<2x1024x2560xf16> into tensor<2048x2560xf16>
%1152 = linalg.matmul ins(%collapsed_1256, %1151 : tensor<2048x2560xf16>, tensor<2560x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1153 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_782, %1152 : tensor<640xf16>, tensor<2048x640xf16>) outs(%919 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2048x640xf16>
%expanded_1257 = tensor.expand_shape %1153 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16>
%1154 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1257, %1089 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%expanded_1258 = tensor.expand_shape %1154 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%1155 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1258 : tensor<2x1024x40x16xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x40x16xf16>
%1156:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1155 : tensor<2x1024x40x16xf16>) outs(%903, %901 : tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>)
%1157 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1156#0 : tensor<2x1024x40x1xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1259 = tensor.collapse_shape %1157 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1158 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1259 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1159 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1154, %1158 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1160 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1159, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1161 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1160 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1162 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1161 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1024x640xf16>
%1163 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1162, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1164 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1163, %1158 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1165 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_781 : tensor<640x40x1xf16>) outs(%914 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1260 = tensor.collapse_shape %1165 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1166 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_780, %collapsed_1260 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640xf16>
%1167 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1166 : tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1261 = tensor.collapse_shape %1164 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1168 = linalg.matmul ins(%collapsed_1261, %1167 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1169 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_779, %1168 : tensor<640xf16>, tensor<2048x640xf16>) outs(%919 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2048x640xf16>
%expanded_1262 = tensor.expand_shape %1169 [[0, 1, 2], [3]] : tensor<2048x640xf16> into tensor<2x32x32x640xf16>
%1170 = linalg.generic {indexing_maps = [#map1, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1262 : tensor<2x32x32x640xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%1171 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1170, %877 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%collapsed_1263 = tensor.collapse_shape %1171 [[0], [1], [2, 3]] : tensor<2x640x32x32xf16> into tensor<2x640x1024xf16>
%expanded_1264 = tensor.expand_shape %collapsed_1263 [[0], [1, 2], [3]] : tensor<2x640x1024xf16> into tensor<2x32x20x1024xf16>
%1172 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1264 : tensor<2x32x20x1024xf16>) outs(%815 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x20x1024xf32>
%1173 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1172 : tensor<2x32x20x1024xf32>) outs(%817 : tensor<2x32x20x1024xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x1024xf64>
%1174 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1173 : tensor<2x32x20x1024xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1175 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1174 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_964 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1176 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1173, %1175 : tensor<2x32x20x1024xf64>, tensor<2x32x1x1xf64>) outs(%817 : tensor<2x32x20x1024xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x1024xf64>
%1177 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1176, %1176 : tensor<2x32x20x1024xf64>, tensor<2x32x20x1024xf64>) outs(%817 : tensor<2x32x20x1024xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x1024xf64>
%1178 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1177 : tensor<2x32x20x1024xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1179 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1178 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_964 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1180 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1179 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1181 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1172 : tensor<2x32x20x1024xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1182 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1181 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_965 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1183 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1180, %cst_938 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%1184 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1183 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1185 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1264, %1182 : tensor<2x32x20x1024xf16>, tensor<2x32x1x1xf32>) outs(%815 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x20x1024xf32>
%1186 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1185, %1184 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%815 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x20x1024xf32>
%collapsed_1265 = tensor.collapse_shape %1186 [[0], [1, 2], [3]] : tensor<2x32x20x1024xf32> into tensor<2x640x1024xf32>
%expanded_1266 = tensor.expand_shape %collapsed_1265 [[0], [1], [2, 3]] : tensor<2x640x1024xf32> into tensor<2x640x32x32xf32>
%expanded_1267 = tensor.expand_shape %cst_778 [[0, 1, 2, 3]] : tensor<640xf16> into tensor<1x640x1x1xf16>
%expanded_1268 = tensor.expand_shape %cst_777 [[0, 1, 2, 3]] : tensor<640xf16> into tensor<1x640x1x1xf16>
%1187 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1266, %expanded_1268 : tensor<2x640x32x32xf32>, tensor<1x640x1x1xf16>) outs(%832 : tensor<2x640x32x32xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x640x32x32xf32>
%1188 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1187, %expanded_1267 : tensor<2x640x32x32xf32>, tensor<1x640x1x1xf16>) outs(%832 : tensor<2x640x32x32xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x640x32x32xf32>
%1189 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1188 : tensor<2x640x32x32xf32>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1190 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1189 : tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.negf %in : f16
%6235 = math.exp %6234 : f16
%6236 = arith.addf %6235, %cst_947 : f16
%6237 = arith.divf %cst_947, %6236 : f16
linalg.yield %6237 : f16
} -> tensor<2x640x32x32xf16>
%1191 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1190, %1189 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%expanded_1269 = tensor.expand_shape %1191 [[0], [1, 2], [3], [4]] : tensor<2x640x32x32xf16> into tensor<2x40x16x32x32xf16>
%1192 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1269 : tensor<2x40x16x32x32xf16>) outs(%838 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x40x16x32x32xf16>
%1193:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1192 : tensor<2x40x16x32x32xf16>) outs(%843, %841 : tensor<2x40x1x32x32xf16>, tensor<2x40x1x32x32xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x40x1x32x32xf16>, tensor<2x40x1x32x32xi64>)
%1194 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1193#0 : tensor<2x40x1x32x32xf16>) outs(%838 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x40x16x32x32xf16>
%collapsed_1270 = tensor.collapse_shape %1194 [[0], [1, 2], [3], [4]] : tensor<2x40x16x32x32xf16> into tensor<2x640x32x32xf16>
%1195 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1270 : tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1196 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1191, %1195 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1197 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1196, %cst_935 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1198 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1197 : tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1199 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1198 : tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x640x32x32xf16>
%1200 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1199, %cst_935 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1201 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1200, %1195 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1202 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_776 : tensor<640x40x1x3x3xf16>) outs(%853 : tensor<640x40x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16x3x3xf16>
%collapsed_1271 = tensor.collapse_shape %1202 [[0], [1, 2], [3], [4]] : tensor<640x40x16x3x3xf16> into tensor<640x640x3x3xf16>
%1203 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_775, %collapsed_1271 : tensor<640x640x3x3xi8>, tensor<640x640x3x3xf16>) outs(%855 : tensor<640x640x3x3xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640x3x3xf16>
%padded_1272 = tensor.pad %1201 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x640x32x32xf16> to tensor<2x640x34x34xf16>
%1204 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_774 : tensor<640xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%1205 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1272, %1203 : tensor<2x640x34x34xf16>, tensor<640x640x3x3xf16>) outs(%1204 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16>
%1206 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_773 : tensor<640x1280xf16>) outs(%808 : tensor<1280x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x640xf16>
%1207 = linalg.matmul ins(%81, %1206 : tensor<2x1280xf16>, tensor<1280x640xf16>) outs(%811 : tensor<2x640xf16>) -> tensor<2x640xf16>
%1208 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_772, %1207 : tensor<640xf16>, tensor<2x640xf16>) outs(%810 : tensor<2x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640xf16>
%expanded_1273 = tensor.expand_shape %1208 [[0], [1, 2, 3]] : tensor<2x640xf16> into tensor<2x640x1x1xf16>
%1209 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1205, %expanded_1273 : tensor<2x640x32x32xf16>, tensor<2x640x1x1xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%collapsed_1274 = tensor.collapse_shape %1209 [[0], [1], [2, 3]] : tensor<2x640x32x32xf16> into tensor<2x640x1024xf16>
%expanded_1275 = tensor.expand_shape %collapsed_1274 [[0], [1, 2], [3]] : tensor<2x640x1024xf16> into tensor<2x32x20x1024xf16>
%1210 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1275 : tensor<2x32x20x1024xf16>) outs(%815 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x20x1024xf32>
%1211 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1210 : tensor<2x32x20x1024xf32>) outs(%817 : tensor<2x32x20x1024xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x1024xf64>
%1212 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1211 : tensor<2x32x20x1024xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1213 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1212 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_964 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1214 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1211, %1213 : tensor<2x32x20x1024xf64>, tensor<2x32x1x1xf64>) outs(%817 : tensor<2x32x20x1024xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x1024xf64>
%1215 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1214, %1214 : tensor<2x32x20x1024xf64>, tensor<2x32x20x1024xf64>) outs(%817 : tensor<2x32x20x1024xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x1024xf64>
%1216 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1215 : tensor<2x32x20x1024xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1217 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1216 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_964 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1218 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1217 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1219 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1210 : tensor<2x32x20x1024xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1220 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1219 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_965 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1221 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1218, %cst_938 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%1222 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1221 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1223 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1275, %1220 : tensor<2x32x20x1024xf16>, tensor<2x32x1x1xf32>) outs(%815 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x20x1024xf32>
%1224 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1223, %1222 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%815 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x20x1024xf32>
%collapsed_1276 = tensor.collapse_shape %1224 [[0], [1, 2], [3]] : tensor<2x32x20x1024xf32> into tensor<2x640x1024xf32>
%expanded_1277 = tensor.expand_shape %collapsed_1276 [[0], [1], [2, 3]] : tensor<2x640x1024xf32> into tensor<2x640x32x32xf32>
%expanded_1278 = tensor.expand_shape %cst_771 [[0, 1, 2, 3]] : tensor<640xf16> into tensor<1x640x1x1xf16>
%expanded_1279 = tensor.expand_shape %cst_770 [[0, 1, 2, 3]] : tensor<640xf16> into tensor<1x640x1x1xf16>
%1225 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1277, %expanded_1279 : tensor<2x640x32x32xf32>, tensor<1x640x1x1xf16>) outs(%832 : tensor<2x640x32x32xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x640x32x32xf32>
%1226 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1225, %expanded_1278 : tensor<2x640x32x32xf32>, tensor<1x640x1x1xf16>) outs(%832 : tensor<2x640x32x32xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x640x32x32xf32>
%1227 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1226 : tensor<2x640x32x32xf32>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1228 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1227 : tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.negf %in : f16
%6235 = math.exp %6234 : f16
%6236 = arith.addf %6235, %cst_947 : f16
%6237 = arith.divf %cst_947, %6236 : f16
linalg.yield %6237 : f16
} -> tensor<2x640x32x32xf16>
%1229 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1228, %1227 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%expanded_1280 = tensor.expand_shape %1229 [[0], [1, 2], [3], [4]] : tensor<2x640x32x32xf16> into tensor<2x40x16x32x32xf16>
%1230 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1280 : tensor<2x40x16x32x32xf16>) outs(%838 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x40x16x32x32xf16>
%1231:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1230 : tensor<2x40x16x32x32xf16>) outs(%843, %841 : tensor<2x40x1x32x32xf16>, tensor<2x40x1x32x32xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x40x1x32x32xf16>, tensor<2x40x1x32x32xi64>)
%1232 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1231#0 : tensor<2x40x1x32x32xf16>) outs(%838 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x40x16x32x32xf16>
%collapsed_1281 = tensor.collapse_shape %1232 [[0], [1, 2], [3], [4]] : tensor<2x40x16x32x32xf16> into tensor<2x640x32x32xf16>
%1233 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1281 : tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1234 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1229, %1233 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1235 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1234, %cst_935 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1236 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1235 : tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1237 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1236 : tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x640x32x32xf16>
%1238 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1237, %cst_935 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1239 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1238, %1233 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1240 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_769 : tensor<640x40x1x3x3xf16>) outs(%853 : tensor<640x40x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16x3x3xf16>
%collapsed_1282 = tensor.collapse_shape %1240 [[0], [1, 2], [3], [4]] : tensor<640x40x16x3x3xf16> into tensor<640x640x3x3xf16>
%1241 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_768, %collapsed_1282 : tensor<640x640x3x3xi8>, tensor<640x640x3x3xf16>) outs(%855 : tensor<640x640x3x3xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640x3x3xf16>
%padded_1283 = tensor.pad %1239 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x640x32x32xf16> to tensor<2x640x34x34xf16>
%1242 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_767 : tensor<640xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%1243 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1283, %1241 : tensor<2x640x34x34xf16>, tensor<640x640x3x3xf16>) outs(%1242 : tensor<2x640x32x32xf16>) -> tensor<2x640x32x32xf16>
%1244 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1171, %1243 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1245 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1244, %cst_924 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%collapsed_1284 = tensor.collapse_shape %1245 [[0], [1], [2, 3]] : tensor<2x640x32x32xf16> into tensor<2x640x1024xf16>
%expanded_1285 = tensor.expand_shape %collapsed_1284 [[0], [1, 2], [3]] : tensor<2x640x1024xf16> into tensor<2x32x20x1024xf16>
%1246 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1285 : tensor<2x32x20x1024xf16>) outs(%815 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x20x1024xf32>
%1247 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1246 : tensor<2x32x20x1024xf32>) outs(%817 : tensor<2x32x20x1024xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x1024xf64>
%1248 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1247 : tensor<2x32x20x1024xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1249 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1248 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_964 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1250 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1247, %1249 : tensor<2x32x20x1024xf64>, tensor<2x32x1x1xf64>) outs(%817 : tensor<2x32x20x1024xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x1024xf64>
%1251 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1250, %1250 : tensor<2x32x20x1024xf64>, tensor<2x32x20x1024xf64>) outs(%817 : tensor<2x32x20x1024xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x1024xf64>
%1252 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1251 : tensor<2x32x20x1024xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1253 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1252 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_964 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1254 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1253 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1255 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1246 : tensor<2x32x20x1024xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1256 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1255 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_965 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1257 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1254, %cst_923 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%1258 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1257 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1259 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1285, %1256 : tensor<2x32x20x1024xf16>, tensor<2x32x1x1xf32>) outs(%815 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x20x1024xf32>
%1260 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1259, %1258 : tensor<2x32x20x1024xf32>, tensor<2x32x1x1xf32>) outs(%815 : tensor<2x32x20x1024xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x20x1024xf32>
%collapsed_1286 = tensor.collapse_shape %1260 [[0], [1, 2], [3]] : tensor<2x32x20x1024xf32> into tensor<2x640x1024xf32>
%expanded_1287 = tensor.expand_shape %collapsed_1286 [[0], [1], [2, 3]] : tensor<2x640x1024xf32> into tensor<2x640x32x32xf32>
%expanded_1288 = tensor.expand_shape %cst_766 [[0, 1, 2, 3]] : tensor<640xf16> into tensor<1x640x1x1xf16>
%expanded_1289 = tensor.expand_shape %cst_765 [[0, 1, 2, 3]] : tensor<640xf16> into tensor<1x640x1x1xf16>
%1261 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1287, %expanded_1289 : tensor<2x640x32x32xf32>, tensor<1x640x1x1xf16>) outs(%832 : tensor<2x640x32x32xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x640x32x32xf32>
%1262 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1261, %expanded_1288 : tensor<2x640x32x32xf32>, tensor<1x640x1x1xf16>) outs(%832 : tensor<2x640x32x32xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x640x32x32xf32>
%1263 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1262 : tensor<2x640x32x32xf32>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1264 = linalg.generic {indexing_maps = [#map1, #map16], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1263 : tensor<2x640x32x32xf16>) outs(%896 : tensor<2x32x32x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x32x32x640xf16>
%collapsed_1290 = tensor.collapse_shape %1264 [[0], [1, 2], [3]] : tensor<2x32x32x640xf16> into tensor<2x1024x640xf16>
%expanded_1291 = tensor.expand_shape %collapsed_1290 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%1265 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1291 : tensor<2x1024x40x16xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x40x16xf16>
%1266:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1265 : tensor<2x1024x40x16xf16>) outs(%903, %901 : tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>)
%1267 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1266#0 : tensor<2x1024x40x1xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1292 = tensor.collapse_shape %1267 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1268 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1292 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1269 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1290, %1268 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1270 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1269, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1271 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1270 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1272 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1271 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1024x640xf16>
%1273 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1272, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1274 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1273, %1268 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1275 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_764 : tensor<640x40x1xf16>) outs(%914 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1293 = tensor.collapse_shape %1275 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1276 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_763, %collapsed_1293 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640xf16>
%1277 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1276 : tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1294 = tensor.collapse_shape %1274 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1278 = linalg.matmul ins(%collapsed_1294, %1277 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1279 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_762, %1278 : tensor<640xf16>, tensor<2048x640xf16>) outs(%919 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2048x640xf16>
%expanded_1295 = tensor.expand_shape %1279 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16>
%1280 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1295 : tensor<2x1024x640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x1024x640xf32>
%1281 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1280 : tensor<2x1024x640xf32>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%1282 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1281 : tensor<2x1024x640xf64>) outs(%928 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1283 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1282 : tensor<2x1024x1xf64>) outs(%927 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_966 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1284 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1281, %1283 : tensor<2x1024x640xf64>, tensor<2x1024x1xf64>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%1285 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1284, %1284 : tensor<2x1024x640xf64>, tensor<2x1024x640xf64>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%1286 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1285 : tensor<2x1024x640xf64>) outs(%928 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1287 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1286 : tensor<2x1024x1xf64>) outs(%927 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_966 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1288 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1287 : tensor<2x1024x1xf64>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1289 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1280 : tensor<2x1024x640xf32>) outs(%937 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1290 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1289 : tensor<2x1024x1xf32>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_967 : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1291 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1288, %cst_938 : tensor<2x1024x1xf32>, tensor<f16>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x1xf32>
%1292 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1291 : tensor<2x1024x1xf32>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1293 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1295, %1290 : tensor<2x1024x640xf16>, tensor<2x1024x1xf32>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%1294 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1293, %1292 : tensor<2x1024x640xf32>, tensor<2x1024x1xf32>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x640xf32>
%1295 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1294, %cst_761 : tensor<2x1024x640xf32>, tensor<640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%1296 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1295, %cst_760 : tensor<2x1024x640xf32>, tensor<640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%1297 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1296 : tensor<2x1024x640xf32>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%expanded_1296 = tensor.expand_shape %1297 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%1298 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1296 : tensor<2x1024x40x16xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x40x16xf16>
%1299:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1298 : tensor<2x1024x40x16xf16>) outs(%903, %901 : tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>)
%1300 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1299#0 : tensor<2x1024x40x1xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1297 = tensor.collapse_shape %1300 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1301 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1297 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1302 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1297, %1301 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1303 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1302, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1304 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1303 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1305 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1304 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1024x640xf16>
%1306 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1305, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1307 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1306, %1301 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1308 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_759 : tensor<640x40x1xf16>) outs(%914 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1298 = tensor.collapse_shape %1308 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1309 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_758, %collapsed_1298 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640xf16>
%1310 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1309 : tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1299 = tensor.collapse_shape %1307 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1311 = linalg.matmul ins(%collapsed_1299, %1310 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%expanded_1300 = tensor.expand_shape %1311 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x10x64xf16>
%1312 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1300 : tensor<2x1024x10x64xf16>) outs(%961 : tensor<2x10x1024x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x10x1024x64xf16>
%collapsed_1301 = tensor.collapse_shape %1312 [[0, 1], [2], [3]] : tensor<2x10x1024x64xf16> into tensor<20x1024x64xf16>
%1313 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_757 : tensor<640x40x1xf16>) outs(%914 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1302 = tensor.collapse_shape %1313 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1314 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_756, %collapsed_1302 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640xf16>
%1315 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1314 : tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%1316 = linalg.matmul ins(%collapsed_1299, %1315 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1317 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_755 : tensor<640x40x1xf16>) outs(%914 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1303 = tensor.collapse_shape %1317 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1318 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_754, %collapsed_1303 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640xf16>
%1319 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1318 : tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%1320 = linalg.matmul ins(%collapsed_1299, %1319 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%expanded_1304 = tensor.expand_shape %1316 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x10x64xf16>
%1321 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1304 : tensor<2x1024x10x64xf16>) outs(%961 : tensor<2x10x1024x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x10x1024x64xf16>
%collapsed_1305 = tensor.collapse_shape %1321 [[0, 1], [2], [3]] : tensor<2x10x1024x64xf16> into tensor<20x1024x64xf16>
%expanded_1306 = tensor.expand_shape %1320 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x10x64xf16>
%1322 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1306 : tensor<2x1024x10x64xf16>) outs(%961 : tensor<2x10x1024x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x10x1024x64xf16>
%collapsed_1307 = tensor.collapse_shape %1322 [[0, 1], [2], [3]] : tensor<2x10x1024x64xf16> into tensor<20x1024x64xf16>
%1323 = linalg.generic {indexing_maps = [#map18, #map23], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1305 : tensor<20x1024x64xf16>) outs(%975 : tensor<20x64x1024xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<20x64x1024xf16>
%1324 = linalg.batch_matmul ins(%collapsed_1301, %1323 : tensor<20x1024x64xf16>, tensor<20x64x1024xf16>) outs(%974 : tensor<20x1024x1024xf16>) -> tensor<20x1024x1024xf16>
%1325 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1324 : tensor<20x1024x1024xf16>) outs(%973 : tensor<20x1024x1024xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.mulf %in, %cst_961 : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x1024xf16>
%1326 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1325, %974 : tensor<20x1024x1024xf16>, tensor<20x1024x1024xf16>) outs(%973 : tensor<20x1024x1024xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in_2625, %cst_946 : f16
%6235 = arith.addf %in, %6234 : f16
linalg.yield %6235 : f16
} -> tensor<20x1024x1024xf16>
%1327:2 = linalg.generic {indexing_maps = [#map18, #map20, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1326 : tensor<20x1024x1024xf16>) outs(%983, %981 : tensor<20x1024x1xf16>, tensor<20x1024x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<20x1024x1xf16>, tensor<20x1024x1xi64>)
%1328 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1326, %1327#0 : tensor<20x1024x1024xf16>, tensor<20x1024x1xf16>) outs(%973 : tensor<20x1024x1024xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x1024xf16>
%1329 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1328 : tensor<20x1024x1024xf16>) outs(%973 : tensor<20x1024x1024xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.exp %in : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x1024xf16>
%1330 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1329 : tensor<20x1024x1024xf16>) outs(%987 : tensor<20x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.addf %in, %out : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x1xf16>
%1331 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1329, %1330 : tensor<20x1024x1024xf16>, tensor<20x1024x1xf16>) outs(%973 : tensor<20x1024x1024xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x1024xf16>
%1332 = linalg.batch_matmul ins(%1331, %collapsed_1307 : tensor<20x1024x1024xf16>, tensor<20x1024x64xf16>) outs(%991 : tensor<20x1024x64xf16>) -> tensor<20x1024x64xf16>
%expanded_1308 = tensor.expand_shape %1332 [[0, 1], [2], [3]] : tensor<20x1024x64xf16> into tensor<2x10x1024x64xf16>
%1333 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1308 : tensor<2x10x1024x64xf16>) outs(%993 : tensor<2x1024x10x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x10x64xf16>
%collapsed_1309 = tensor.collapse_shape %1333 [[0], [1], [2, 3]] : tensor<2x1024x10x64xf16> into tensor<2x1024x640xf16>
%expanded_1310 = tensor.expand_shape %collapsed_1309 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%1334 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1310 : tensor<2x1024x40x16xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x40x16xf16>
%1335:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1334 : tensor<2x1024x40x16xf16>) outs(%903, %901 : tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>)
%1336 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1335#0 : tensor<2x1024x40x1xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1311 = tensor.collapse_shape %1336 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1337 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1311 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1338 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1309, %1337 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1339 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1338, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1340 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1339 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1341 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1340 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1024x640xf16>
%1342 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1341, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1343 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1342, %1337 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1344 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_753 : tensor<640x40x1xf16>) outs(%914 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1312 = tensor.collapse_shape %1344 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1345 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_752, %collapsed_1312 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640xf16>
%1346 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1345 : tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1313 = tensor.collapse_shape %1343 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1347 = linalg.matmul ins(%collapsed_1313, %1346 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1348 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_751, %1347 : tensor<640xf16>, tensor<2048x640xf16>) outs(%919 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2048x640xf16>
%expanded_1314 = tensor.expand_shape %1348 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16>
%1349 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1314, %expanded_1295 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1350 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1349 : tensor<2x1024x640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x1024x640xf32>
%1351 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1350 : tensor<2x1024x640xf32>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%1352 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1351 : tensor<2x1024x640xf64>) outs(%928 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1353 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1352 : tensor<2x1024x1xf64>) outs(%927 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_966 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1354 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1351, %1353 : tensor<2x1024x640xf64>, tensor<2x1024x1xf64>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%1355 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1354, %1354 : tensor<2x1024x640xf64>, tensor<2x1024x640xf64>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%1356 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1355 : tensor<2x1024x640xf64>) outs(%928 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1357 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1356 : tensor<2x1024x1xf64>) outs(%927 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_966 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1358 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1357 : tensor<2x1024x1xf64>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1359 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1350 : tensor<2x1024x640xf32>) outs(%937 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1360 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1359 : tensor<2x1024x1xf32>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_967 : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1361 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1358, %cst_938 : tensor<2x1024x1xf32>, tensor<f16>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x1xf32>
%1362 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1361 : tensor<2x1024x1xf32>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1363 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1349, %1360 : tensor<2x1024x640xf16>, tensor<2x1024x1xf32>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%1364 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1363, %1362 : tensor<2x1024x640xf32>, tensor<2x1024x1xf32>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x640xf32>
%1365 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1364, %cst_750 : tensor<2x1024x640xf32>, tensor<640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%1366 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1365, %cst_749 : tensor<2x1024x640xf32>, tensor<640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%1367 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1366 : tensor<2x1024x640xf32>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%expanded_1315 = tensor.expand_shape %1367 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%1368 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1315 : tensor<2x1024x40x16xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x40x16xf16>
%1369:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1368 : tensor<2x1024x40x16xf16>) outs(%903, %901 : tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>)
%1370 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1369#0 : tensor<2x1024x40x1xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1316 = tensor.collapse_shape %1370 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1371 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1316 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1372 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1367, %1371 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1373 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1372, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1374 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1373 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1375 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1374 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1024x640xf16>
%1376 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1375, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1377 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1376, %1371 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1378 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_748 : tensor<640x40x1xf16>) outs(%914 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1317 = tensor.collapse_shape %1378 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1379 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_747, %collapsed_1317 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640xf16>
%1380 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1379 : tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1318 = tensor.collapse_shape %1377 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1381 = linalg.matmul ins(%collapsed_1318, %1380 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%expanded_1319 = tensor.expand_shape %1381 [[0, 1], [2, 3]] : tensor<2048x640xf16> into tensor<2x1024x10x64xf16>
%1382 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1319 : tensor<2x1024x10x64xf16>) outs(%961 : tensor<2x10x1024x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x10x1024x64xf16>
%collapsed_1320 = tensor.collapse_shape %1382 [[0, 1], [2], [3]] : tensor<2x10x1024x64xf16> into tensor<20x1024x64xf16>
%1383 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_746 : tensor<640x64x1xf16>) outs(%1044 : tensor<640x64x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x64x16xf16>
%collapsed_1321 = tensor.collapse_shape %1383 [[0], [1, 2]] : tensor<640x64x16xf16> into tensor<640x1024xf16>
%1384 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_745, %collapsed_1321 : tensor<640x1024xi8>, tensor<640x1024xf16>) outs(%1046 : tensor<640x1024xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x1024xf16>
%1385 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1384 : tensor<640x1024xf16>) outs(%1048 : tensor<1024x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1024x640xf16>
%1386 = linalg.matmul ins(%collapsed_1050, %1385 : tensor<154x1024xf16>, tensor<1024x640xf16>) outs(%1051 : tensor<154x640xf16>) -> tensor<154x640xf16>
%1387 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_744 : tensor<640x64x1xf16>) outs(%1044 : tensor<640x64x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x64x16xf16>
%collapsed_1322 = tensor.collapse_shape %1387 [[0], [1, 2]] : tensor<640x64x16xf16> into tensor<640x1024xf16>
%1388 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_743, %collapsed_1322 : tensor<640x1024xi8>, tensor<640x1024xf16>) outs(%1046 : tensor<640x1024xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x1024xf16>
%1389 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1388 : tensor<640x1024xf16>) outs(%1048 : tensor<1024x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1024x640xf16>
%1390 = linalg.matmul ins(%collapsed_1050, %1389 : tensor<154x1024xf16>, tensor<1024x640xf16>) outs(%1051 : tensor<154x640xf16>) -> tensor<154x640xf16>
%expanded_1323 = tensor.expand_shape %1386 [[0, 1], [2, 3]] : tensor<154x640xf16> into tensor<2x77x10x64xf16>
%1391 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1323 : tensor<2x77x10x64xf16>) outs(%1057 : tensor<2x10x77x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x10x77x64xf16>
%collapsed_1324 = tensor.collapse_shape %1391 [[0, 1], [2], [3]] : tensor<2x10x77x64xf16> into tensor<20x77x64xf16>
%expanded_1325 = tensor.expand_shape %1390 [[0, 1], [2, 3]] : tensor<154x640xf16> into tensor<2x77x10x64xf16>
%1392 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1325 : tensor<2x77x10x64xf16>) outs(%1057 : tensor<2x10x77x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x10x77x64xf16>
%collapsed_1326 = tensor.collapse_shape %1392 [[0, 1], [2], [3]] : tensor<2x10x77x64xf16> into tensor<20x77x64xf16>
%1393 = linalg.generic {indexing_maps = [#map18, #map23], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1324 : tensor<20x77x64xf16>) outs(%1062 : tensor<20x64x77xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<20x64x77xf16>
%1394 = linalg.batch_matmul ins(%collapsed_1320, %1393 : tensor<20x1024x64xf16>, tensor<20x64x77xf16>) outs(%1061 : tensor<20x1024x77xf16>) -> tensor<20x1024x77xf16>
%1395 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1394 : tensor<20x1024x77xf16>) outs(%1060 : tensor<20x1024x77xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.mulf %in, %cst_961 : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x77xf16>
%1396 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1395, %1061 : tensor<20x1024x77xf16>, tensor<20x1024x77xf16>) outs(%1060 : tensor<20x1024x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in_2625, %cst_946 : f16
%6235 = arith.addf %in, %6234 : f16
linalg.yield %6235 : f16
} -> tensor<20x1024x77xf16>
%1397:2 = linalg.generic {indexing_maps = [#map18, #map20, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1396 : tensor<20x1024x77xf16>) outs(%983, %981 : tensor<20x1024x1xf16>, tensor<20x1024x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<20x1024x1xf16>, tensor<20x1024x1xi64>)
%1398 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1396, %1397#0 : tensor<20x1024x77xf16>, tensor<20x1024x1xf16>) outs(%1060 : tensor<20x1024x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x77xf16>
%1399 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1398 : tensor<20x1024x77xf16>) outs(%1060 : tensor<20x1024x77xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.exp %in : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x77xf16>
%1400 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1399 : tensor<20x1024x77xf16>) outs(%987 : tensor<20x1024x1xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.addf %in, %out : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x1xf16>
%1401 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1399, %1400 : tensor<20x1024x77xf16>, tensor<20x1024x1xf16>) outs(%1060 : tensor<20x1024x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<20x1024x77xf16>
%1402 = linalg.batch_matmul ins(%1401, %collapsed_1326 : tensor<20x1024x77xf16>, tensor<20x77x64xf16>) outs(%991 : tensor<20x1024x64xf16>) -> tensor<20x1024x64xf16>
%expanded_1327 = tensor.expand_shape %1402 [[0, 1], [2], [3]] : tensor<20x1024x64xf16> into tensor<2x10x1024x64xf16>
%1403 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1327 : tensor<2x10x1024x64xf16>) outs(%993 : tensor<2x1024x10x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x10x64xf16>
%collapsed_1328 = tensor.collapse_shape %1403 [[0], [1], [2, 3]] : tensor<2x1024x10x64xf16> into tensor<2x1024x640xf16>
%expanded_1329 = tensor.expand_shape %collapsed_1328 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%1404 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1329 : tensor<2x1024x40x16xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x40x16xf16>
%1405:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1404 : tensor<2x1024x40x16xf16>) outs(%903, %901 : tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>)
%1406 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1405#0 : tensor<2x1024x40x1xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1330 = tensor.collapse_shape %1406 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1407 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1330 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1408 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1328, %1407 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1409 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1408, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1410 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1409 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1411 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1410 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1024x640xf16>
%1412 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1411, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1413 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1412, %1407 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1414 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_742 : tensor<640x40x1xf16>) outs(%914 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1331 = tensor.collapse_shape %1414 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1415 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_741, %collapsed_1331 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640xf16>
%1416 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1415 : tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1332 = tensor.collapse_shape %1413 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1417 = linalg.matmul ins(%collapsed_1332, %1416 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1418 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_740, %1417 : tensor<640xf16>, tensor<2048x640xf16>) outs(%919 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2048x640xf16>
%expanded_1333 = tensor.expand_shape %1418 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16>
%1419 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1333, %1349 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1420 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1419 : tensor<2x1024x640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x1024x640xf32>
%1421 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1420 : tensor<2x1024x640xf32>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%1422 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1421 : tensor<2x1024x640xf64>) outs(%928 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1423 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1422 : tensor<2x1024x1xf64>) outs(%927 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_966 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1424 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1421, %1423 : tensor<2x1024x640xf64>, tensor<2x1024x1xf64>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%1425 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1424, %1424 : tensor<2x1024x640xf64>, tensor<2x1024x640xf64>) outs(%925 : tensor<2x1024x640xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x640xf64>
%1426 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1425 : tensor<2x1024x640xf64>) outs(%928 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1427 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1426 : tensor<2x1024x1xf64>) outs(%927 : tensor<2x1024x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_966 : f64
linalg.yield %6234 : f64
} -> tensor<2x1024x1xf64>
%1428 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1427 : tensor<2x1024x1xf64>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1429 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1420 : tensor<2x1024x640xf32>) outs(%937 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1430 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1429 : tensor<2x1024x1xf32>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_967 : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1431 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1428, %cst_938 : tensor<2x1024x1xf32>, tensor<f16>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x1xf32>
%1432 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1431 : tensor<2x1024x1xf32>) outs(%935 : tensor<2x1024x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x1xf32>
%1433 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1419, %1430 : tensor<2x1024x640xf16>, tensor<2x1024x1xf32>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%1434 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1433, %1432 : tensor<2x1024x640xf32>, tensor<2x1024x1xf32>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x1024x640xf32>
%1435 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1434, %cst_739 : tensor<2x1024x640xf32>, tensor<640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%1436 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1435, %cst_738 : tensor<2x1024x640xf32>, tensor<640xf16>) outs(%923 : tensor<2x1024x640xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1024x640xf32>
%1437 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1436 : tensor<2x1024x640xf32>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%expanded_1334 = tensor.expand_shape %1437 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%1438 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1334 : tensor<2x1024x40x16xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x40x16xf16>
%1439:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1438 : tensor<2x1024x40x16xf16>) outs(%903, %901 : tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>)
%1440 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1439#0 : tensor<2x1024x40x1xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1335 = tensor.collapse_shape %1440 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1441 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1335 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1442 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1437, %1441 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1443 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1442, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1444 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1443 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1445 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1444 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1024x640xf16>
%1446 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1445, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1447 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1446, %1441 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1448 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_737 : tensor<5120x40x1xf16>) outs(%1118 : tensor<5120x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<5120x40x16xf16>
%collapsed_1336 = tensor.collapse_shape %1448 [[0], [1, 2]] : tensor<5120x40x16xf16> into tensor<5120x640xf16>
%1449 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_736, %collapsed_1336 : tensor<5120x640xi8>, tensor<5120x640xf16>) outs(%1120 : tensor<5120x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<5120x640xf16>
%1450 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1449 : tensor<5120x640xf16>) outs(%1122 : tensor<640x5120xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x5120xf16>
%collapsed_1337 = tensor.collapse_shape %1447 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1451 = linalg.matmul ins(%collapsed_1337, %1450 : tensor<2048x640xf16>, tensor<640x5120xf16>) outs(%1125 : tensor<2048x5120xf16>) -> tensor<2048x5120xf16>
%1452 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_735, %1451 : tensor<5120xf16>, tensor<2048x5120xf16>) outs(%1124 : tensor<2048x5120xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2048x5120xf16>
%expanded_1338 = tensor.expand_shape %1452 [[0, 1], [2]] : tensor<2048x5120xf16> into tensor<2x1024x5120xf16>
%extracted_slice_1339 = tensor.extract_slice %expanded_1338[0, 0, 0] [2, 1024, 2560] [1, 1, 1] : tensor<2x1024x5120xf16> to tensor<2x1024x2560xf16>
%extracted_slice_1340 = tensor.extract_slice %expanded_1338[0, 0, 2560] [2, 1024, 2560] [1, 1, 1] : tensor<2x1024x5120xf16> to tensor<2x1024x2560xf16>
%1453 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1340 : tensor<2x1024x2560xf16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.sqrt %cst_949 : f16
%6235 = arith.divf %in, %6234 : f16
%6236 = math.erf %6235 : f16
%6237 = arith.addf %6236, %cst_947 : f16
%6238 = arith.mulf %6237, %cst_950 : f16
%6239 = arith.mulf %in, %6238 : f16
linalg.yield %6239 : f16
} -> tensor<2x1024x2560xf16>
%1454 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1339, %1453 : tensor<2x1024x2560xf16>, tensor<2x1024x2560xf16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x2560xf16>
%expanded_1341 = tensor.expand_shape %1454 [[0], [1], [2, 3]] : tensor<2x1024x2560xf16> into tensor<2x1024x160x16xf16>
%1455 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1341 : tensor<2x1024x160x16xf16>) outs(%1131 : tensor<2x1024x160x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x160x16xf16>
%1456:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1455 : tensor<2x1024x160x16xf16>) outs(%1136, %1134 : tensor<2x1024x160x1xf16>, tensor<2x1024x160x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x1024x160x1xf16>, tensor<2x1024x160x1xi64>)
%1457 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1456#0 : tensor<2x1024x160x1xf16>) outs(%1131 : tensor<2x1024x160x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x160x16xf16>
%collapsed_1342 = tensor.collapse_shape %1457 [[0], [1], [2, 3]] : tensor<2x1024x160x16xf16> into tensor<2x1024x2560xf16>
%1458 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1342 : tensor<2x1024x2560xf16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x2560xf16>
%1459 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1454, %1458 : tensor<2x1024x2560xf16>, tensor<2x1024x2560xf16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x2560xf16>
%1460 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1459, %cst_935 : tensor<2x1024x2560xf16>, tensor<f16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x2560xf16>
%1461 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1460 : tensor<2x1024x2560xf16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x2560xf16>
%1462 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1461 : tensor<2x1024x2560xf16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1024x2560xf16>
%1463 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1462, %cst_935 : tensor<2x1024x2560xf16>, tensor<f16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x2560xf16>
%1464 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1463, %1458 : tensor<2x1024x2560xf16>, tensor<2x1024x2560xf16>) outs(%1128 : tensor<2x1024x2560xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x2560xf16>
%1465 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_734 : tensor<640x160x1xf16>) outs(%1146 : tensor<640x160x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x160x16xf16>
%collapsed_1343 = tensor.collapse_shape %1465 [[0], [1, 2]] : tensor<640x160x16xf16> into tensor<640x2560xf16>
%1466 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_733, %collapsed_1343 : tensor<640x2560xi8>, tensor<640x2560xf16>) outs(%1148 : tensor<640x2560xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x2560xf16>
%1467 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1466 : tensor<640x2560xf16>) outs(%1150 : tensor<2560x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2560x640xf16>
%collapsed_1344 = tensor.collapse_shape %1464 [[0, 1], [2]] : tensor<2x1024x2560xf16> into tensor<2048x2560xf16>
%1468 = linalg.matmul ins(%collapsed_1344, %1467 : tensor<2048x2560xf16>, tensor<2560x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1469 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_732, %1468 : tensor<640xf16>, tensor<2048x640xf16>) outs(%919 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2048x640xf16>
%expanded_1345 = tensor.expand_shape %1469 [[0, 1], [2]] : tensor<2048x640xf16> into tensor<2x1024x640xf16>
%1470 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1345, %1419 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%expanded_1346 = tensor.expand_shape %1470 [[0], [1], [2, 3]] : tensor<2x1024x640xf16> into tensor<2x1024x40x16xf16>
%1471 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1346 : tensor<2x1024x40x16xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x40x16xf16>
%1472:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1471 : tensor<2x1024x40x16xf16>) outs(%903, %901 : tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x1024x40x1xf16>, tensor<2x1024x40x1xi64>)
%1473 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1472#0 : tensor<2x1024x40x1xf16>) outs(%898 : tensor<2x1024x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1024x40x16xf16>
%collapsed_1347 = tensor.collapse_shape %1473 [[0], [1], [2, 3]] : tensor<2x1024x40x16xf16> into tensor<2x1024x640xf16>
%1474 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1347 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1475 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1470, %1474 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1476 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1475, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1477 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1476 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1478 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1477 : tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1024x640xf16>
%1479 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1478, %cst_935 : tensor<2x1024x640xf16>, tensor<f16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1480 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1479, %1474 : tensor<2x1024x640xf16>, tensor<2x1024x640xf16>) outs(%906 : tensor<2x1024x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1024x640xf16>
%1481 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_731 : tensor<640x40x1xf16>) outs(%914 : tensor<640x40x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16xf16>
%collapsed_1348 = tensor.collapse_shape %1481 [[0], [1, 2]] : tensor<640x40x16xf16> into tensor<640x640xf16>
%1482 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_730, %collapsed_1348 : tensor<640x640xi8>, tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640xf16>
%1483 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1482 : tensor<640x640xf16>) outs(%916 : tensor<640x640xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x640xf16>
%collapsed_1349 = tensor.collapse_shape %1480 [[0, 1], [2]] : tensor<2x1024x640xf16> into tensor<2048x640xf16>
%1484 = linalg.matmul ins(%collapsed_1349, %1483 : tensor<2048x640xf16>, tensor<640x640xf16>) outs(%920 : tensor<2048x640xf16>) -> tensor<2048x640xf16>
%1485 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_729, %1484 : tensor<640xf16>, tensor<2048x640xf16>) outs(%919 : tensor<2048x640xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2048x640xf16>
%expanded_1350 = tensor.expand_shape %1485 [[0, 1, 2], [3]] : tensor<2048x640xf16> into tensor<2x32x32x640xf16>
%1486 = linalg.generic {indexing_maps = [#map1, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1350 : tensor<2x32x32x640xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x32x32xf16>
%1487 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1486, %1245 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%expanded_1351 = tensor.expand_shape %1487 [[0], [1, 2], [3], [4]] : tensor<2x640x32x32xf16> into tensor<2x40x16x32x32xf16>
%1488 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1351 : tensor<2x40x16x32x32xf16>) outs(%838 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x40x16x32x32xf16>
%1489:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1488 : tensor<2x40x16x32x32xf16>) outs(%843, %841 : tensor<2x40x1x32x32xf16>, tensor<2x40x1x32x32xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x40x1x32x32xf16>, tensor<2x40x1x32x32xi64>)
%1490 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1489#0 : tensor<2x40x1x32x32xf16>) outs(%838 : tensor<2x40x16x32x32xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x40x16x32x32xf16>
%collapsed_1352 = tensor.collapse_shape %1490 [[0], [1, 2], [3], [4]] : tensor<2x40x16x32x32xf16> into tensor<2x640x32x32xf16>
%1491 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1352 : tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1492 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1487, %1491 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1493 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1492, %cst_935 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1494 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1493 : tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1495 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1494 : tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x640x32x32xf16>
%1496 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1495, %cst_935 : tensor<2x640x32x32xf16>, tensor<f16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1497 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1496, %1491 : tensor<2x640x32x32xf16>, tensor<2x640x32x32xf16>) outs(%805 : tensor<2x640x32x32xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x32x32xf16>
%1498 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_728 : tensor<640x40x1x3x3xf16>) outs(%853 : tensor<640x40x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<640x40x16x3x3xf16>
%collapsed_1353 = tensor.collapse_shape %1498 [[0], [1, 2], [3], [4]] : tensor<640x40x16x3x3xf16> into tensor<640x640x3x3xf16>
%1499 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_727, %collapsed_1353 : tensor<640x640x3x3xi8>, tensor<640x640x3x3xf16>) outs(%855 : tensor<640x640x3x3xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<640x640x3x3xf16>
%padded_1354 = tensor.pad %1497 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x640x32x32xf16> to tensor<2x640x34x34xf16>
%1500 = tensor.empty() : tensor<2x640x16x16xf16>
%1501 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_726 : tensor<640xf16>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x640x16x16xf16>
%1502 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%padded_1354, %1499 : tensor<2x640x34x34xf16>, tensor<640x640x3x3xf16>) outs(%1501 : tensor<2x640x16x16xf16>) -> tensor<2x640x16x16xf16>
%collapsed_1355 = tensor.collapse_shape %1502 [[0], [1], [2, 3]] : tensor<2x640x16x16xf16> into tensor<2x640x256xf16>
%expanded_1356 = tensor.expand_shape %collapsed_1355 [[0], [1, 2], [3]] : tensor<2x640x256xf16> into tensor<2x32x20x256xf16>
%1503 = tensor.empty() : tensor<2x32x20x256xf32>
%1504 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1356 : tensor<2x32x20x256xf16>) outs(%1503 : tensor<2x32x20x256xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x20x256xf32>
%1505 = tensor.empty() : tensor<2x32x20x256xf64>
%1506 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1504 : tensor<2x32x20x256xf32>) outs(%1505 : tensor<2x32x20x256xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x256xf64>
%1507 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1506 : tensor<2x32x20x256xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1508 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1507 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_968 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1509 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1506, %1508 : tensor<2x32x20x256xf64>, tensor<2x32x1x1xf64>) outs(%1505 : tensor<2x32x20x256xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x256xf64>
%1510 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1509, %1509 : tensor<2x32x20x256xf64>, tensor<2x32x20x256xf64>) outs(%1505 : tensor<2x32x20x256xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x20x256xf64>
%1511 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1510 : tensor<2x32x20x256xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1512 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1511 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_968 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1513 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1512 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1514 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1504 : tensor<2x32x20x256xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1515 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1514 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_969 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1516 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1513, %cst_938 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%1517 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1516 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1518 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1356, %1515 : tensor<2x32x20x256xf16>, tensor<2x32x1x1xf32>) outs(%1503 : tensor<2x32x20x256xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x20x256xf32>
%1519 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1518, %1517 : tensor<2x32x20x256xf32>, tensor<2x32x1x1xf32>) outs(%1503 : tensor<2x32x20x256xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x20x256xf32>
%collapsed_1357 = tensor.collapse_shape %1519 [[0], [1, 2], [3]] : tensor<2x32x20x256xf32> into tensor<2x640x256xf32>
%expanded_1358 = tensor.expand_shape %collapsed_1357 [[0], [1], [2, 3]] : tensor<2x640x256xf32> into tensor<2x640x16x16xf32>
%expanded_1359 = tensor.expand_shape %cst_725 [[0, 1, 2, 3]] : tensor<640xf16> into tensor<1x640x1x1xf16>
%expanded_1360 = tensor.expand_shape %cst_724 [[0, 1, 2, 3]] : tensor<640xf16> into tensor<1x640x1x1xf16>
%1520 = tensor.empty() : tensor<2x640x16x16xf32>
%1521 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1358, %expanded_1360 : tensor<2x640x16x16xf32>, tensor<1x640x1x1xf16>) outs(%1520 : tensor<2x640x16x16xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x640x16x16xf32>
%1522 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1521, %expanded_1359 : tensor<2x640x16x16xf32>, tensor<1x640x1x1xf16>) outs(%1520 : tensor<2x640x16x16xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x640x16x16xf32>
%1523 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1522 : tensor<2x640x16x16xf32>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x640x16x16xf16>
%1524 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1523 : tensor<2x640x16x16xf16>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.negf %in : f16
%6235 = math.exp %6234 : f16
%6236 = arith.addf %6235, %cst_947 : f16
%6237 = arith.divf %cst_947, %6236 : f16
linalg.yield %6237 : f16
} -> tensor<2x640x16x16xf16>
%1525 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1524, %1523 : tensor<2x640x16x16xf16>, tensor<2x640x16x16xf16>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x16x16xf16>
%expanded_1361 = tensor.expand_shape %1525 [[0], [1, 2], [3], [4]] : tensor<2x640x16x16xf16> into tensor<2x40x16x16x16xf16>
%1526 = tensor.empty() : tensor<2x40x16x16x16xf16>
%1527 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1361 : tensor<2x40x16x16x16xf16>) outs(%1526 : tensor<2x40x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x40x16x16x16xf16>
%1528 = tensor.empty() : tensor<2x40x1x16x16xi64>
%1529 = linalg.fill ins(%c0_i64 : i64) outs(%1528 : tensor<2x40x1x16x16xi64>) -> tensor<2x40x1x16x16xi64>
%1530 = tensor.empty() : tensor<2x40x1x16x16xf16>
%1531 = linalg.fill ins(%cst_948 : f16) outs(%1530 : tensor<2x40x1x16x16xf16>) -> tensor<2x40x1x16x16xf16>
%1532:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1527 : tensor<2x40x16x16x16xf16>) outs(%1531, %1529 : tensor<2x40x1x16x16xf16>, tensor<2x40x1x16x16xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x40x1x16x16xf16>, tensor<2x40x1x16x16xi64>)
%1533 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1532#0 : tensor<2x40x1x16x16xf16>) outs(%1526 : tensor<2x40x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x40x16x16x16xf16>
%collapsed_1362 = tensor.collapse_shape %1533 [[0], [1, 2], [3], [4]] : tensor<2x40x16x16x16xf16> into tensor<2x640x16x16xf16>
%1534 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1362 : tensor<2x640x16x16xf16>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x16x16xf16>
%1535 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1525, %1534 : tensor<2x640x16x16xf16>, tensor<2x640x16x16xf16>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x16x16xf16>
%1536 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1535, %cst_935 : tensor<2x640x16x16xf16>, tensor<f16>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x16x16xf16>
%1537 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1536 : tensor<2x640x16x16xf16>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x640x16x16xf16>
%1538 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1537 : tensor<2x640x16x16xf16>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x640x16x16xf16>
%1539 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1538, %cst_935 : tensor<2x640x16x16xf16>, tensor<f16>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x16x16xf16>
%1540 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1539, %1534 : tensor<2x640x16x16xf16>, tensor<2x640x16x16xf16>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x16x16xf16>
%1541 = tensor.empty() : tensor<1280x40x16x3x3xf16>
%1542 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_723 : tensor<1280x40x1x3x3xf16>) outs(%1541 : tensor<1280x40x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x40x16x3x3xf16>
%collapsed_1363 = tensor.collapse_shape %1542 [[0], [1, 2], [3], [4]] : tensor<1280x40x16x3x3xf16> into tensor<1280x640x3x3xf16>
%1543 = tensor.empty() : tensor<1280x640x3x3xf16>
%1544 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_722, %collapsed_1363 : tensor<1280x640x3x3xi8>, tensor<1280x640x3x3xf16>) outs(%1543 : tensor<1280x640x3x3xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x640x3x3xf16>
%padded_1364 = tensor.pad %1540 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x640x16x16xf16> to tensor<2x640x18x18xf16>
%1545 = tensor.empty() : tensor<2x1280x16x16xf16>
%1546 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_721 : tensor<1280xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1280x16x16xf16>
%1547 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1364, %1544 : tensor<2x640x18x18xf16>, tensor<1280x640x3x3xf16>) outs(%1546 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16>
%1548 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_720 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%1549 = linalg.matmul ins(%81, %1548 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%20 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
%1550 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_719, %1549 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%19 : tensor<2x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280xf16>
%expanded_1365 = tensor.expand_shape %1550 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16>
%1551 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1547, %expanded_1365 : tensor<2x1280x16x16xf16>, tensor<2x1280x1x1xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%collapsed_1366 = tensor.collapse_shape %1551 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16>
%expanded_1367 = tensor.expand_shape %collapsed_1366 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16>
%1552 = tensor.empty() : tensor<2x32x40x256xf32>
%1553 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1367 : tensor<2x32x40x256xf16>) outs(%1552 : tensor<2x32x40x256xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x40x256xf32>
%1554 = tensor.empty() : tensor<2x32x40x256xf64>
%1555 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1553 : tensor<2x32x40x256xf32>) outs(%1554 : tensor<2x32x40x256xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x256xf64>
%1556 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1555 : tensor<2x32x40x256xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1557 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1556 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_962 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1558 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1555, %1557 : tensor<2x32x40x256xf64>, tensor<2x32x1x1xf64>) outs(%1554 : tensor<2x32x40x256xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x256xf64>
%1559 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1558, %1558 : tensor<2x32x40x256xf64>, tensor<2x32x40x256xf64>) outs(%1554 : tensor<2x32x40x256xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x256xf64>
%1560 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1559 : tensor<2x32x40x256xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1561 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1560 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_962 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1562 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1561 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1563 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1553 : tensor<2x32x40x256xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1564 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1563 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_963 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1565 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1562, %cst_938 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%1566 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1565 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1567 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1367, %1564 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%1552 : tensor<2x32x40x256xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x40x256xf32>
%1568 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1567, %1566 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%1552 : tensor<2x32x40x256xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x40x256xf32>
%collapsed_1368 = tensor.collapse_shape %1568 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32>
%expanded_1369 = tensor.expand_shape %collapsed_1368 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32>
%expanded_1370 = tensor.expand_shape %cst_718 [[0, 1, 2, 3]] : tensor<1280xf16> into tensor<1x1280x1x1xf16>
%expanded_1371 = tensor.expand_shape %cst_717 [[0, 1, 2, 3]] : tensor<1280xf16> into tensor<1x1280x1x1xf16>
%1569 = tensor.empty() : tensor<2x1280x16x16xf32>
%1570 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1369, %expanded_1371 : tensor<2x1280x16x16xf32>, tensor<1x1280x1x1xf16>) outs(%1569 : tensor<2x1280x16x16xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1280x16x16xf32>
%1571 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1570, %expanded_1370 : tensor<2x1280x16x16xf32>, tensor<1x1280x1x1xf16>) outs(%1569 : tensor<2x1280x16x16xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1280x16x16xf32>
%1572 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1571 : tensor<2x1280x16x16xf32>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1573 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1572 : tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.negf %in : f16
%6235 = math.exp %6234 : f16
%6236 = arith.addf %6235, %cst_947 : f16
%6237 = arith.divf %cst_947, %6236 : f16
linalg.yield %6237 : f16
} -> tensor<2x1280x16x16xf16>
%1574 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1573, %1572 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%expanded_1372 = tensor.expand_shape %1574 [[0], [1, 2], [3], [4]] : tensor<2x1280x16x16xf16> into tensor<2x80x16x16x16xf16>
%1575 = tensor.empty() : tensor<2x80x16x16x16xf16>
%1576 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1372 : tensor<2x80x16x16x16xf16>) outs(%1575 : tensor<2x80x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x80x16x16x16xf16>
%1577 = tensor.empty() : tensor<2x80x1x16x16xi64>
%1578 = linalg.fill ins(%c0_i64 : i64) outs(%1577 : tensor<2x80x1x16x16xi64>) -> tensor<2x80x1x16x16xi64>
%1579 = tensor.empty() : tensor<2x80x1x16x16xf16>
%1580 = linalg.fill ins(%cst_948 : f16) outs(%1579 : tensor<2x80x1x16x16xf16>) -> tensor<2x80x1x16x16xf16>
%1581:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1576 : tensor<2x80x16x16x16xf16>) outs(%1580, %1578 : tensor<2x80x1x16x16xf16>, tensor<2x80x1x16x16xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x80x1x16x16xf16>, tensor<2x80x1x16x16xi64>)
%1582 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1581#0 : tensor<2x80x1x16x16xf16>) outs(%1575 : tensor<2x80x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x80x16x16x16xf16>
%collapsed_1373 = tensor.collapse_shape %1582 [[0], [1, 2], [3], [4]] : tensor<2x80x16x16x16xf16> into tensor<2x1280x16x16xf16>
%1583 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1373 : tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1584 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1574, %1583 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1585 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1584, %cst_935 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1586 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1585 : tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1587 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1586 : tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1280x16x16xf16>
%1588 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1587, %cst_935 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1589 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1588, %1583 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1590 = tensor.empty() : tensor<1280x80x16x3x3xf16>
%1591 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_716 : tensor<1280x80x1x3x3xf16>) outs(%1590 : tensor<1280x80x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16x3x3xf16>
%collapsed_1374 = tensor.collapse_shape %1591 [[0], [1, 2], [3], [4]] : tensor<1280x80x16x3x3xf16> into tensor<1280x1280x3x3xf16>
%1592 = tensor.empty() : tensor<1280x1280x3x3xf16>
%1593 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_715, %collapsed_1374 : tensor<1280x1280x3x3xi8>, tensor<1280x1280x3x3xf16>) outs(%1592 : tensor<1280x1280x3x3xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280x3x3xf16>
%padded_1375 = tensor.pad %1589 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x1280x16x16xf16> to tensor<2x1280x18x18xf16>
%1594 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_714 : tensor<1280xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1280x16x16xf16>
%1595 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1375, %1593 : tensor<2x1280x18x18xf16>, tensor<1280x1280x3x3xf16>) outs(%1594 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16>
%expanded_1376 = tensor.expand_shape %1502 [[0], [1, 2], [3], [4]] : tensor<2x640x16x16xf16> into tensor<2x40x16x16x16xf16>
%1596 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1376 : tensor<2x40x16x16x16xf16>) outs(%1526 : tensor<2x40x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x40x16x16x16xf16>
%1597:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1596 : tensor<2x40x16x16x16xf16>) outs(%1531, %1529 : tensor<2x40x1x16x16xf16>, tensor<2x40x1x16x16xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x40x1x16x16xf16>, tensor<2x40x1x16x16xi64>)
%1598 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1597#0 : tensor<2x40x1x16x16xf16>) outs(%1526 : tensor<2x40x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x40x16x16x16xf16>
%collapsed_1377 = tensor.collapse_shape %1598 [[0], [1, 2], [3], [4]] : tensor<2x40x16x16x16xf16> into tensor<2x640x16x16xf16>
%1599 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1377 : tensor<2x640x16x16xf16>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x16x16xf16>
%1600 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1502, %1599 : tensor<2x640x16x16xf16>, tensor<2x640x16x16xf16>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x16x16xf16>
%1601 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1600, %cst_935 : tensor<2x640x16x16xf16>, tensor<f16>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x16x16xf16>
%1602 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1601 : tensor<2x640x16x16xf16>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x640x16x16xf16>
%1603 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1602 : tensor<2x640x16x16xf16>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x640x16x16xf16>
%1604 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1603, %cst_935 : tensor<2x640x16x16xf16>, tensor<f16>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x16x16xf16>
%1605 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1604, %1599 : tensor<2x640x16x16xf16>, tensor<2x640x16x16xf16>) outs(%1500 : tensor<2x640x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x640x16x16xf16>
%1606 = tensor.empty() : tensor<1280x40x16x1x1xf16>
%1607 = linalg.generic {indexing_maps = [#map25, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_713 : tensor<1280x40x1x1x1xf16>) outs(%1606 : tensor<1280x40x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x40x16x1x1xf16>
%1608 = linalg.generic {indexing_maps = [#map26, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1607 : tensor<1280x40x16x1x1xf16>) outs(%1606 : tensor<1280x40x16x1x1xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x40x16x1x1xf16>
%collapsed_1378 = tensor.collapse_shape %1608 [[0], [1, 2], [3], [4]] : tensor<1280x40x16x1x1xf16> into tensor<1280x640x1x1xf16>
%1609 = tensor.empty() : tensor<1280x640x1x1xf16>
%1610 = linalg.generic {indexing_maps = [#map11, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_712, %collapsed_1378 : tensor<1280x640x1x1xi8>, tensor<1280x640x1x1xf16>) outs(%1609 : tensor<1280x640x1x1xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x640x1x1xf16>
%1611 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_711 : tensor<1280xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1280x16x16xf16>
%1612 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1605, %1610 : tensor<2x640x16x16xf16>, tensor<1280x640x1x1xf16>) outs(%1611 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16>
%1613 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1612, %1595 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1614 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1613, %cst_924 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%collapsed_1379 = tensor.collapse_shape %1614 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16>
%expanded_1380 = tensor.expand_shape %collapsed_1379 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16>
%1615 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1380 : tensor<2x32x40x256xf16>) outs(%1552 : tensor<2x32x40x256xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x40x256xf32>
%1616 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1615 : tensor<2x32x40x256xf32>) outs(%1554 : tensor<2x32x40x256xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x256xf64>
%1617 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1616 : tensor<2x32x40x256xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1618 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1617 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_962 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1619 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1616, %1618 : tensor<2x32x40x256xf64>, tensor<2x32x1x1xf64>) outs(%1554 : tensor<2x32x40x256xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x256xf64>
%1620 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1619, %1619 : tensor<2x32x40x256xf64>, tensor<2x32x40x256xf64>) outs(%1554 : tensor<2x32x40x256xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x256xf64>
%1621 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1620 : tensor<2x32x40x256xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1622 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1621 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_962 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1623 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1622 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1624 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1615 : tensor<2x32x40x256xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1625 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1624 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_963 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1626 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1623, %cst_923 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%1627 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1626 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1628 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1380, %1625 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%1552 : tensor<2x32x40x256xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x40x256xf32>
%1629 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1628, %1627 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%1552 : tensor<2x32x40x256xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x40x256xf32>
%collapsed_1381 = tensor.collapse_shape %1629 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32>
%expanded_1382 = tensor.expand_shape %collapsed_1381 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32>
%expanded_1383 = tensor.expand_shape %cst_710 [[0, 1, 2, 3]] : tensor<1280xf16> into tensor<1x1280x1x1xf16>
%expanded_1384 = tensor.expand_shape %cst_709 [[0, 1, 2, 3]] : tensor<1280xf16> into tensor<1x1280x1x1xf16>
%1630 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1382, %expanded_1384 : tensor<2x1280x16x16xf32>, tensor<1x1280x1x1xf16>) outs(%1569 : tensor<2x1280x16x16xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1280x16x16xf32>
%1631 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1630, %expanded_1383 : tensor<2x1280x16x16xf32>, tensor<1x1280x1x1xf16>) outs(%1569 : tensor<2x1280x16x16xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1280x16x16xf32>
%1632 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1631 : tensor<2x1280x16x16xf32>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1633 = tensor.empty() : tensor<2x16x16x1280xf16>
%1634 = linalg.generic {indexing_maps = [#map1, #map16], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1632 : tensor<2x1280x16x16xf16>) outs(%1633 : tensor<2x16x16x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x16x16x1280xf16>
%collapsed_1385 = tensor.collapse_shape %1634 [[0], [1, 2], [3]] : tensor<2x16x16x1280xf16> into tensor<2x256x1280xf16>
%expanded_1386 = tensor.expand_shape %collapsed_1385 [[0], [1], [2, 3]] : tensor<2x256x1280xf16> into tensor<2x256x80x16xf16>
%1635 = tensor.empty() : tensor<2x256x80x16xf16>
%1636 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1386 : tensor<2x256x80x16xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x80x16xf16>
%1637 = tensor.empty() : tensor<2x256x80x1xi64>
%1638 = linalg.fill ins(%c0_i64 : i64) outs(%1637 : tensor<2x256x80x1xi64>) -> tensor<2x256x80x1xi64>
%1639 = tensor.empty() : tensor<2x256x80x1xf16>
%1640 = linalg.fill ins(%cst_948 : f16) outs(%1639 : tensor<2x256x80x1xf16>) -> tensor<2x256x80x1xf16>
%1641:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1636 : tensor<2x256x80x16xf16>) outs(%1640, %1638 : tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>)
%1642 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1641#0 : tensor<2x256x80x1xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x80x16xf16>
%collapsed_1387 = tensor.collapse_shape %1642 [[0], [1], [2, 3]] : tensor<2x256x80x16xf16> into tensor<2x256x1280xf16>
%1643 = tensor.empty() : tensor<2x256x1280xf16>
%1644 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1387 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1645 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1385, %1644 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1646 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1645, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1647 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1646 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1648 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1647 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x256x1280xf16>
%1649 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1648, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1650 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1649, %1644 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1651 = tensor.empty() : tensor<1280x80x16xf16>
%1652 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_708 : tensor<1280x80x1xf16>) outs(%1651 : tensor<1280x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16xf16>
%collapsed_1388 = tensor.collapse_shape %1652 [[0], [1, 2]] : tensor<1280x80x16xf16> into tensor<1280x1280xf16>
%1653 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_707, %collapsed_1388 : tensor<1280x1280xi8>, tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280xf16>
%1654 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1653 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%collapsed_1389 = tensor.collapse_shape %1650 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16>
%1655 = tensor.empty() : tensor<512x1280xf16>
%1656 = linalg.fill ins(%cst_946 : f16) outs(%1655 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%1657 = linalg.matmul ins(%collapsed_1389, %1654 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%1658 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_706, %1657 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1655 : tensor<512x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<512x1280xf16>
%expanded_1390 = tensor.expand_shape %1658 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16>
%1659 = tensor.empty() : tensor<2x256x1280xf32>
%1660 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1390 : tensor<2x256x1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x256x1280xf32>
%1661 = tensor.empty() : tensor<2x256x1280xf64>
%1662 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1660 : tensor<2x256x1280xf32>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%1663 = tensor.empty() : tensor<2x256x1xf64>
%1664 = linalg.fill ins(%cst_951 : f64) outs(%1663 : tensor<2x256x1xf64>) -> tensor<2x256x1xf64>
%1665 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1662 : tensor<2x256x1280xf64>) outs(%1664 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%1666 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1665 : tensor<2x256x1xf64>) outs(%1663 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_970 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%1667 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1662, %1666 : tensor<2x256x1280xf64>, tensor<2x256x1xf64>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%1668 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1667, %1667 : tensor<2x256x1280xf64>, tensor<2x256x1280xf64>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%1669 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1668 : tensor<2x256x1280xf64>) outs(%1664 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%1670 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1669 : tensor<2x256x1xf64>) outs(%1663 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_970 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%1671 = tensor.empty() : tensor<2x256x1xf32>
%1672 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1670 : tensor<2x256x1xf64>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%1673 = linalg.fill ins(%cst_952 : f32) outs(%1671 : tensor<2x256x1xf32>) -> tensor<2x256x1xf32>
%1674 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1660 : tensor<2x256x1280xf32>) outs(%1673 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%1675 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1674 : tensor<2x256x1xf32>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_971 : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%1676 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1672, %cst_938 : tensor<2x256x1xf32>, tensor<f16>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1xf32>
%1677 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1676 : tensor<2x256x1xf32>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%1678 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1390, %1675 : tensor<2x256x1280xf16>, tensor<2x256x1xf32>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%1679 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1678, %1677 : tensor<2x256x1280xf32>, tensor<2x256x1xf32>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1280xf32>
%1680 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1679, %cst_705 : tensor<2x256x1280xf32>, tensor<1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%1681 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1680, %cst_704 : tensor<2x256x1280xf32>, tensor<1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%1682 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1681 : tensor<2x256x1280xf32>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%expanded_1391 = tensor.expand_shape %1682 [[0], [1], [2, 3]] : tensor<2x256x1280xf16> into tensor<2x256x80x16xf16>
%1683 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1391 : tensor<2x256x80x16xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x80x16xf16>
%1684:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1683 : tensor<2x256x80x16xf16>) outs(%1640, %1638 : tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>)
%1685 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1684#0 : tensor<2x256x80x1xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x80x16xf16>
%collapsed_1392 = tensor.collapse_shape %1685 [[0], [1], [2, 3]] : tensor<2x256x80x16xf16> into tensor<2x256x1280xf16>
%1686 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1392 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1687 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1682, %1686 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1688 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1687, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1689 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1688 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1690 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1689 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x256x1280xf16>
%1691 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1690, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1692 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1691, %1686 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1693 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_703 : tensor<1280x80x1xf16>) outs(%1651 : tensor<1280x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16xf16>
%collapsed_1393 = tensor.collapse_shape %1693 [[0], [1, 2]] : tensor<1280x80x16xf16> into tensor<1280x1280xf16>
%1694 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_702, %collapsed_1393 : tensor<1280x1280xi8>, tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280xf16>
%1695 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1694 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%collapsed_1394 = tensor.collapse_shape %1692 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16>
%1696 = linalg.matmul ins(%collapsed_1394, %1695 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%expanded_1395 = tensor.expand_shape %1696 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x20x64xf16>
%1697 = tensor.empty() : tensor<2x20x256x64xf16>
%1698 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1395 : tensor<2x256x20x64xf16>) outs(%1697 : tensor<2x20x256x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x256x64xf16>
%collapsed_1396 = tensor.collapse_shape %1698 [[0, 1], [2], [3]] : tensor<2x20x256x64xf16> into tensor<40x256x64xf16>
%1699 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_701 : tensor<1280x80x1xf16>) outs(%1651 : tensor<1280x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16xf16>
%collapsed_1397 = tensor.collapse_shape %1699 [[0], [1, 2]] : tensor<1280x80x16xf16> into tensor<1280x1280xf16>
%1700 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_700, %collapsed_1397 : tensor<1280x1280xi8>, tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280xf16>
%1701 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1700 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%1702 = linalg.matmul ins(%collapsed_1394, %1701 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%1703 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_699 : tensor<1280x80x1xf16>) outs(%1651 : tensor<1280x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16xf16>
%collapsed_1398 = tensor.collapse_shape %1703 [[0], [1, 2]] : tensor<1280x80x16xf16> into tensor<1280x1280xf16>
%1704 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_698, %collapsed_1398 : tensor<1280x1280xi8>, tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280xf16>
%1705 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1704 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%1706 = linalg.matmul ins(%collapsed_1394, %1705 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%expanded_1399 = tensor.expand_shape %1702 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x20x64xf16>
%1707 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1399 : tensor<2x256x20x64xf16>) outs(%1697 : tensor<2x20x256x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x256x64xf16>
%collapsed_1400 = tensor.collapse_shape %1707 [[0, 1], [2], [3]] : tensor<2x20x256x64xf16> into tensor<40x256x64xf16>
%expanded_1401 = tensor.expand_shape %1706 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x20x64xf16>
%1708 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1401 : tensor<2x256x20x64xf16>) outs(%1697 : tensor<2x20x256x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x256x64xf16>
%collapsed_1402 = tensor.collapse_shape %1708 [[0, 1], [2], [3]] : tensor<2x20x256x64xf16> into tensor<40x256x64xf16>
%1709 = tensor.empty() : tensor<40x256x256xf16>
%1710 = linalg.fill ins(%cst_946 : f16) outs(%1709 : tensor<40x256x256xf16>) -> tensor<40x256x256xf16>
%1711 = tensor.empty() : tensor<40x64x256xf16>
%1712 = linalg.generic {indexing_maps = [#map18, #map23], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1400 : tensor<40x256x64xf16>) outs(%1711 : tensor<40x64x256xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<40x64x256xf16>
%1713 = linalg.batch_matmul ins(%collapsed_1396, %1712 : tensor<40x256x64xf16>, tensor<40x64x256xf16>) outs(%1710 : tensor<40x256x256xf16>) -> tensor<40x256x256xf16>
%1714 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1713 : tensor<40x256x256xf16>) outs(%1709 : tensor<40x256x256xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.mulf %in, %cst_961 : f16
linalg.yield %6234 : f16
} -> tensor<40x256x256xf16>
%1715 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1714, %1710 : tensor<40x256x256xf16>, tensor<40x256x256xf16>) outs(%1709 : tensor<40x256x256xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in_2625, %cst_946 : f16
%6235 = arith.addf %in, %6234 : f16
linalg.yield %6235 : f16
} -> tensor<40x256x256xf16>
%1716 = tensor.empty() : tensor<40x256x1xi64>
%1717 = linalg.fill ins(%c0_i64 : i64) outs(%1716 : tensor<40x256x1xi64>) -> tensor<40x256x1xi64>
%1718 = tensor.empty() : tensor<40x256x1xf16>
%1719 = linalg.fill ins(%cst_948 : f16) outs(%1718 : tensor<40x256x1xf16>) -> tensor<40x256x1xf16>
%1720:2 = linalg.generic {indexing_maps = [#map18, #map20, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1715 : tensor<40x256x256xf16>) outs(%1719, %1717 : tensor<40x256x1xf16>, tensor<40x256x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<40x256x1xf16>, tensor<40x256x1xi64>)
%1721 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1715, %1720#0 : tensor<40x256x256xf16>, tensor<40x256x1xf16>) outs(%1709 : tensor<40x256x256xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<40x256x256xf16>
%1722 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1721 : tensor<40x256x256xf16>) outs(%1709 : tensor<40x256x256xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.exp %in : f16
linalg.yield %6234 : f16
} -> tensor<40x256x256xf16>
%1723 = linalg.fill ins(%cst_946 : f16) outs(%1718 : tensor<40x256x1xf16>) -> tensor<40x256x1xf16>
%1724 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1722 : tensor<40x256x256xf16>) outs(%1723 : tensor<40x256x1xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.addf %in, %out : f16
linalg.yield %6234 : f16
} -> tensor<40x256x1xf16>
%1725 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1722, %1724 : tensor<40x256x256xf16>, tensor<40x256x1xf16>) outs(%1709 : tensor<40x256x256xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<40x256x256xf16>
%1726 = tensor.empty() : tensor<40x256x64xf16>
%1727 = linalg.fill ins(%cst_946 : f16) outs(%1726 : tensor<40x256x64xf16>) -> tensor<40x256x64xf16>
%1728 = linalg.batch_matmul ins(%1725, %collapsed_1402 : tensor<40x256x256xf16>, tensor<40x256x64xf16>) outs(%1727 : tensor<40x256x64xf16>) -> tensor<40x256x64xf16>
%expanded_1403 = tensor.expand_shape %1728 [[0, 1], [2], [3]] : tensor<40x256x64xf16> into tensor<2x20x256x64xf16>
%1729 = tensor.empty() : tensor<2x256x20x64xf16>
%1730 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1403 : tensor<2x20x256x64xf16>) outs(%1729 : tensor<2x256x20x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x20x64xf16>
%collapsed_1404 = tensor.collapse_shape %1730 [[0], [1], [2, 3]] : tensor<2x256x20x64xf16> into tensor<2x256x1280xf16>
%expanded_1405 = tensor.expand_shape %collapsed_1404 [[0], [1], [2, 3]] : tensor<2x256x1280xf16> into tensor<2x256x80x16xf16>
%1731 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1405 : tensor<2x256x80x16xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x80x16xf16>
%1732:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1731 : tensor<2x256x80x16xf16>) outs(%1640, %1638 : tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>)
%1733 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1732#0 : tensor<2x256x80x1xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x80x16xf16>
%collapsed_1406 = tensor.collapse_shape %1733 [[0], [1], [2, 3]] : tensor<2x256x80x16xf16> into tensor<2x256x1280xf16>
%1734 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1406 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1735 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1404, %1734 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1736 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1735, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1737 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1736 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1738 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1737 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x256x1280xf16>
%1739 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1738, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1740 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1739, %1734 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1741 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_697 : tensor<1280x80x1xf16>) outs(%1651 : tensor<1280x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16xf16>
%collapsed_1407 = tensor.collapse_shape %1741 [[0], [1, 2]] : tensor<1280x80x16xf16> into tensor<1280x1280xf16>
%1742 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_696, %collapsed_1407 : tensor<1280x1280xi8>, tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280xf16>
%1743 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1742 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%collapsed_1408 = tensor.collapse_shape %1740 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16>
%1744 = linalg.matmul ins(%collapsed_1408, %1743 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%1745 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_695, %1744 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1655 : tensor<512x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<512x1280xf16>
%expanded_1409 = tensor.expand_shape %1745 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16>
%1746 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1409, %expanded_1390 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1747 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1746 : tensor<2x256x1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x256x1280xf32>
%1748 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1747 : tensor<2x256x1280xf32>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%1749 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1748 : tensor<2x256x1280xf64>) outs(%1664 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%1750 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1749 : tensor<2x256x1xf64>) outs(%1663 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_970 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%1751 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1748, %1750 : tensor<2x256x1280xf64>, tensor<2x256x1xf64>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%1752 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1751, %1751 : tensor<2x256x1280xf64>, tensor<2x256x1280xf64>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%1753 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1752 : tensor<2x256x1280xf64>) outs(%1664 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%1754 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1753 : tensor<2x256x1xf64>) outs(%1663 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_970 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%1755 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1754 : tensor<2x256x1xf64>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%1756 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1747 : tensor<2x256x1280xf32>) outs(%1673 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%1757 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1756 : tensor<2x256x1xf32>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_971 : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%1758 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1755, %cst_938 : tensor<2x256x1xf32>, tensor<f16>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1xf32>
%1759 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1758 : tensor<2x256x1xf32>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%1760 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1746, %1757 : tensor<2x256x1280xf16>, tensor<2x256x1xf32>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%1761 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1760, %1759 : tensor<2x256x1280xf32>, tensor<2x256x1xf32>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1280xf32>
%1762 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1761, %cst_694 : tensor<2x256x1280xf32>, tensor<1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%1763 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1762, %cst_693 : tensor<2x256x1280xf32>, tensor<1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%1764 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1763 : tensor<2x256x1280xf32>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%expanded_1410 = tensor.expand_shape %1764 [[0], [1], [2, 3]] : tensor<2x256x1280xf16> into tensor<2x256x80x16xf16>
%1765 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1410 : tensor<2x256x80x16xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x80x16xf16>
%1766:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1765 : tensor<2x256x80x16xf16>) outs(%1640, %1638 : tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>)
%1767 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1766#0 : tensor<2x256x80x1xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x80x16xf16>
%collapsed_1411 = tensor.collapse_shape %1767 [[0], [1], [2, 3]] : tensor<2x256x80x16xf16> into tensor<2x256x1280xf16>
%1768 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1411 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1769 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1764, %1768 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1770 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1769, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1771 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1770 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1772 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1771 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x256x1280xf16>
%1773 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1772, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1774 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1773, %1768 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1775 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_692 : tensor<1280x80x1xf16>) outs(%1651 : tensor<1280x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16xf16>
%collapsed_1412 = tensor.collapse_shape %1775 [[0], [1, 2]] : tensor<1280x80x16xf16> into tensor<1280x1280xf16>
%1776 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_691, %collapsed_1412 : tensor<1280x1280xi8>, tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280xf16>
%1777 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1776 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%collapsed_1413 = tensor.collapse_shape %1774 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16>
%1778 = linalg.matmul ins(%collapsed_1413, %1777 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%expanded_1414 = tensor.expand_shape %1778 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x20x64xf16>
%1779 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1414 : tensor<2x256x20x64xf16>) outs(%1697 : tensor<2x20x256x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x256x64xf16>
%collapsed_1415 = tensor.collapse_shape %1779 [[0, 1], [2], [3]] : tensor<2x20x256x64xf16> into tensor<40x256x64xf16>
%1780 = tensor.empty() : tensor<1280x64x16xf16>
%1781 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_690 : tensor<1280x64x1xf16>) outs(%1780 : tensor<1280x64x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x64x16xf16>
%collapsed_1416 = tensor.collapse_shape %1781 [[0], [1, 2]] : tensor<1280x64x16xf16> into tensor<1280x1024xf16>
%1782 = tensor.empty() : tensor<1280x1024xf16>
%1783 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_689, %collapsed_1416 : tensor<1280x1024xi8>, tensor<1280x1024xf16>) outs(%1782 : tensor<1280x1024xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1024xf16>
%1784 = tensor.empty() : tensor<1024x1280xf16>
%1785 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1783 : tensor<1280x1024xf16>) outs(%1784 : tensor<1024x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1024x1280xf16>
%1786 = tensor.empty() : tensor<154x1280xf16>
%1787 = linalg.fill ins(%cst_946 : f16) outs(%1786 : tensor<154x1280xf16>) -> tensor<154x1280xf16>
%1788 = linalg.matmul ins(%collapsed_1050, %1785 : tensor<154x1024xf16>, tensor<1024x1280xf16>) outs(%1787 : tensor<154x1280xf16>) -> tensor<154x1280xf16>
%1789 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_688 : tensor<1280x64x1xf16>) outs(%1780 : tensor<1280x64x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x64x16xf16>
%collapsed_1417 = tensor.collapse_shape %1789 [[0], [1, 2]] : tensor<1280x64x16xf16> into tensor<1280x1024xf16>
%1790 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_687, %collapsed_1417 : tensor<1280x1024xi8>, tensor<1280x1024xf16>) outs(%1782 : tensor<1280x1024xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1024xf16>
%1791 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1790 : tensor<1280x1024xf16>) outs(%1784 : tensor<1024x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1024x1280xf16>
%1792 = linalg.matmul ins(%collapsed_1050, %1791 : tensor<154x1024xf16>, tensor<1024x1280xf16>) outs(%1787 : tensor<154x1280xf16>) -> tensor<154x1280xf16>
%expanded_1418 = tensor.expand_shape %1788 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x20x64xf16>
%1793 = tensor.empty() : tensor<2x20x77x64xf16>
%1794 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1418 : tensor<2x77x20x64xf16>) outs(%1793 : tensor<2x20x77x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x77x64xf16>
%collapsed_1419 = tensor.collapse_shape %1794 [[0, 1], [2], [3]] : tensor<2x20x77x64xf16> into tensor<40x77x64xf16>
%expanded_1420 = tensor.expand_shape %1792 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x20x64xf16>
%1795 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1420 : tensor<2x77x20x64xf16>) outs(%1793 : tensor<2x20x77x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x77x64xf16>
%collapsed_1421 = tensor.collapse_shape %1795 [[0, 1], [2], [3]] : tensor<2x20x77x64xf16> into tensor<40x77x64xf16>
%1796 = tensor.empty() : tensor<40x256x77xf16>
%1797 = linalg.fill ins(%cst_946 : f16) outs(%1796 : tensor<40x256x77xf16>) -> tensor<40x256x77xf16>
%1798 = tensor.empty() : tensor<40x64x77xf16>
%1799 = linalg.generic {indexing_maps = [#map18, #map23], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1419 : tensor<40x77x64xf16>) outs(%1798 : tensor<40x64x77xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<40x64x77xf16>
%1800 = linalg.batch_matmul ins(%collapsed_1415, %1799 : tensor<40x256x64xf16>, tensor<40x64x77xf16>) outs(%1797 : tensor<40x256x77xf16>) -> tensor<40x256x77xf16>
%1801 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1800 : tensor<40x256x77xf16>) outs(%1796 : tensor<40x256x77xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.mulf %in, %cst_961 : f16
linalg.yield %6234 : f16
} -> tensor<40x256x77xf16>
%1802 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1801, %1797 : tensor<40x256x77xf16>, tensor<40x256x77xf16>) outs(%1796 : tensor<40x256x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in_2625, %cst_946 : f16
%6235 = arith.addf %in, %6234 : f16
linalg.yield %6235 : f16
} -> tensor<40x256x77xf16>
%1803:2 = linalg.generic {indexing_maps = [#map18, #map20, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1802 : tensor<40x256x77xf16>) outs(%1719, %1717 : tensor<40x256x1xf16>, tensor<40x256x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<40x256x1xf16>, tensor<40x256x1xi64>)
%1804 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1802, %1803#0 : tensor<40x256x77xf16>, tensor<40x256x1xf16>) outs(%1796 : tensor<40x256x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<40x256x77xf16>
%1805 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1804 : tensor<40x256x77xf16>) outs(%1796 : tensor<40x256x77xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.exp %in : f16
linalg.yield %6234 : f16
} -> tensor<40x256x77xf16>
%1806 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1805 : tensor<40x256x77xf16>) outs(%1723 : tensor<40x256x1xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.addf %in, %out : f16
linalg.yield %6234 : f16
} -> tensor<40x256x1xf16>
%1807 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1805, %1806 : tensor<40x256x77xf16>, tensor<40x256x1xf16>) outs(%1796 : tensor<40x256x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<40x256x77xf16>
%1808 = linalg.batch_matmul ins(%1807, %collapsed_1421 : tensor<40x256x77xf16>, tensor<40x77x64xf16>) outs(%1727 : tensor<40x256x64xf16>) -> tensor<40x256x64xf16>
%expanded_1422 = tensor.expand_shape %1808 [[0, 1], [2], [3]] : tensor<40x256x64xf16> into tensor<2x20x256x64xf16>
%1809 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1422 : tensor<2x20x256x64xf16>) outs(%1729 : tensor<2x256x20x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x20x64xf16>
%collapsed_1423 = tensor.collapse_shape %1809 [[0], [1], [2, 3]] : tensor<2x256x20x64xf16> into tensor<2x256x1280xf16>
%expanded_1424 = tensor.expand_shape %collapsed_1423 [[0], [1], [2, 3]] : tensor<2x256x1280xf16> into tensor<2x256x80x16xf16>
%1810 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1424 : tensor<2x256x80x16xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x80x16xf16>
%1811:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1810 : tensor<2x256x80x16xf16>) outs(%1640, %1638 : tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>)
%1812 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1811#0 : tensor<2x256x80x1xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x80x16xf16>
%collapsed_1425 = tensor.collapse_shape %1812 [[0], [1], [2, 3]] : tensor<2x256x80x16xf16> into tensor<2x256x1280xf16>
%1813 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1425 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1814 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1423, %1813 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1815 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1814, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1816 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1815 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1817 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1816 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x256x1280xf16>
%1818 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1817, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1819 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1818, %1813 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1820 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_686 : tensor<1280x80x1xf16>) outs(%1651 : tensor<1280x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16xf16>
%collapsed_1426 = tensor.collapse_shape %1820 [[0], [1, 2]] : tensor<1280x80x16xf16> into tensor<1280x1280xf16>
%1821 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_685, %collapsed_1426 : tensor<1280x1280xi8>, tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280xf16>
%1822 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1821 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%collapsed_1427 = tensor.collapse_shape %1819 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16>
%1823 = linalg.matmul ins(%collapsed_1427, %1822 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%1824 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_684, %1823 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1655 : tensor<512x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<512x1280xf16>
%expanded_1428 = tensor.expand_shape %1824 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16>
%1825 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1428, %1746 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1826 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1825 : tensor<2x256x1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x256x1280xf32>
%1827 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1826 : tensor<2x256x1280xf32>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%1828 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1827 : tensor<2x256x1280xf64>) outs(%1664 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%1829 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1828 : tensor<2x256x1xf64>) outs(%1663 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_970 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%1830 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1827, %1829 : tensor<2x256x1280xf64>, tensor<2x256x1xf64>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%1831 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1830, %1830 : tensor<2x256x1280xf64>, tensor<2x256x1280xf64>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%1832 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1831 : tensor<2x256x1280xf64>) outs(%1664 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%1833 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1832 : tensor<2x256x1xf64>) outs(%1663 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_970 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%1834 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1833 : tensor<2x256x1xf64>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%1835 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1826 : tensor<2x256x1280xf32>) outs(%1673 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%1836 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1835 : tensor<2x256x1xf32>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_971 : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%1837 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1834, %cst_938 : tensor<2x256x1xf32>, tensor<f16>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1xf32>
%1838 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1837 : tensor<2x256x1xf32>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%1839 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1825, %1836 : tensor<2x256x1280xf16>, tensor<2x256x1xf32>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%1840 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1839, %1838 : tensor<2x256x1280xf32>, tensor<2x256x1xf32>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1280xf32>
%1841 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1840, %cst_683 : tensor<2x256x1280xf32>, tensor<1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%1842 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1841, %cst_682 : tensor<2x256x1280xf32>, tensor<1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%1843 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1842 : tensor<2x256x1280xf32>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%expanded_1429 = tensor.expand_shape %1843 [[0], [1], [2, 3]] : tensor<2x256x1280xf16> into tensor<2x256x80x16xf16>
%1844 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1429 : tensor<2x256x80x16xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x80x16xf16>
%1845:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1844 : tensor<2x256x80x16xf16>) outs(%1640, %1638 : tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>)
%1846 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1845#0 : tensor<2x256x80x1xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x80x16xf16>
%collapsed_1430 = tensor.collapse_shape %1846 [[0], [1], [2, 3]] : tensor<2x256x80x16xf16> into tensor<2x256x1280xf16>
%1847 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1430 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1848 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1843, %1847 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1849 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1848, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1850 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1849 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1851 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1850 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x256x1280xf16>
%1852 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1851, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1853 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1852, %1847 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1854 = tensor.empty() : tensor<10240x80x16xf16>
%1855 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_681 : tensor<10240x80x1xf16>) outs(%1854 : tensor<10240x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<10240x80x16xf16>
%collapsed_1431 = tensor.collapse_shape %1855 [[0], [1, 2]] : tensor<10240x80x16xf16> into tensor<10240x1280xf16>
%1856 = tensor.empty() : tensor<10240x1280xf16>
%1857 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_680, %collapsed_1431 : tensor<10240x1280xi8>, tensor<10240x1280xf16>) outs(%1856 : tensor<10240x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<10240x1280xf16>
%1858 = tensor.empty() : tensor<1280x10240xf16>
%1859 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1857 : tensor<10240x1280xf16>) outs(%1858 : tensor<1280x10240xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x10240xf16>
%collapsed_1432 = tensor.collapse_shape %1853 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16>
%1860 = tensor.empty() : tensor<512x10240xf16>
%1861 = linalg.fill ins(%cst_946 : f16) outs(%1860 : tensor<512x10240xf16>) -> tensor<512x10240xf16>
%1862 = linalg.matmul ins(%collapsed_1432, %1859 : tensor<512x1280xf16>, tensor<1280x10240xf16>) outs(%1861 : tensor<512x10240xf16>) -> tensor<512x10240xf16>
%1863 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_679, %1862 : tensor<10240xf16>, tensor<512x10240xf16>) outs(%1860 : tensor<512x10240xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<512x10240xf16>
%expanded_1433 = tensor.expand_shape %1863 [[0, 1], [2]] : tensor<512x10240xf16> into tensor<2x256x10240xf16>
%extracted_slice_1434 = tensor.extract_slice %expanded_1433[0, 0, 0] [2, 256, 5120] [1, 1, 1] : tensor<2x256x10240xf16> to tensor<2x256x5120xf16>
%extracted_slice_1435 = tensor.extract_slice %expanded_1433[0, 0, 5120] [2, 256, 5120] [1, 1, 1] : tensor<2x256x10240xf16> to tensor<2x256x5120xf16>
%1864 = tensor.empty() : tensor<2x256x5120xf16>
%1865 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1435 : tensor<2x256x5120xf16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.sqrt %cst_949 : f16
%6235 = arith.divf %in, %6234 : f16
%6236 = math.erf %6235 : f16
%6237 = arith.addf %6236, %cst_947 : f16
%6238 = arith.mulf %6237, %cst_950 : f16
%6239 = arith.mulf %in, %6238 : f16
linalg.yield %6239 : f16
} -> tensor<2x256x5120xf16>
%1866 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1434, %1865 : tensor<2x256x5120xf16>, tensor<2x256x5120xf16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x5120xf16>
%expanded_1436 = tensor.expand_shape %1866 [[0], [1], [2, 3]] : tensor<2x256x5120xf16> into tensor<2x256x320x16xf16>
%1867 = tensor.empty() : tensor<2x256x320x16xf16>
%1868 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1436 : tensor<2x256x320x16xf16>) outs(%1867 : tensor<2x256x320x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x320x16xf16>
%1869 = tensor.empty() : tensor<2x256x320x1xi64>
%1870 = linalg.fill ins(%c0_i64 : i64) outs(%1869 : tensor<2x256x320x1xi64>) -> tensor<2x256x320x1xi64>
%1871 = tensor.empty() : tensor<2x256x320x1xf16>
%1872 = linalg.fill ins(%cst_948 : f16) outs(%1871 : tensor<2x256x320x1xf16>) -> tensor<2x256x320x1xf16>
%1873:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1868 : tensor<2x256x320x16xf16>) outs(%1872, %1870 : tensor<2x256x320x1xf16>, tensor<2x256x320x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x256x320x1xf16>, tensor<2x256x320x1xi64>)
%1874 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1873#0 : tensor<2x256x320x1xf16>) outs(%1867 : tensor<2x256x320x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x320x16xf16>
%collapsed_1437 = tensor.collapse_shape %1874 [[0], [1], [2, 3]] : tensor<2x256x320x16xf16> into tensor<2x256x5120xf16>
%1875 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1437 : tensor<2x256x5120xf16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x5120xf16>
%1876 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1866, %1875 : tensor<2x256x5120xf16>, tensor<2x256x5120xf16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x5120xf16>
%1877 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1876, %cst_935 : tensor<2x256x5120xf16>, tensor<f16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x5120xf16>
%1878 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1877 : tensor<2x256x5120xf16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x5120xf16>
%1879 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1878 : tensor<2x256x5120xf16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x256x5120xf16>
%1880 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1879, %cst_935 : tensor<2x256x5120xf16>, tensor<f16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x5120xf16>
%1881 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1880, %1875 : tensor<2x256x5120xf16>, tensor<2x256x5120xf16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x5120xf16>
%1882 = tensor.empty() : tensor<1280x320x16xf16>
%1883 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_678 : tensor<1280x320x1xf16>) outs(%1882 : tensor<1280x320x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x320x16xf16>
%collapsed_1438 = tensor.collapse_shape %1883 [[0], [1, 2]] : tensor<1280x320x16xf16> into tensor<1280x5120xf16>
%1884 = tensor.empty() : tensor<1280x5120xf16>
%1885 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_677, %collapsed_1438 : tensor<1280x5120xi8>, tensor<1280x5120xf16>) outs(%1884 : tensor<1280x5120xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x5120xf16>
%1886 = tensor.empty() : tensor<5120x1280xf16>
%1887 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1885 : tensor<1280x5120xf16>) outs(%1886 : tensor<5120x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<5120x1280xf16>
%collapsed_1439 = tensor.collapse_shape %1881 [[0, 1], [2]] : tensor<2x256x5120xf16> into tensor<512x5120xf16>
%1888 = linalg.matmul ins(%collapsed_1439, %1887 : tensor<512x5120xf16>, tensor<5120x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%1889 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_676, %1888 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1655 : tensor<512x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<512x1280xf16>
%expanded_1440 = tensor.expand_shape %1889 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16>
%1890 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1440, %1825 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%expanded_1441 = tensor.expand_shape %1890 [[0], [1], [2, 3]] : tensor<2x256x1280xf16> into tensor<2x256x80x16xf16>
%1891 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1441 : tensor<2x256x80x16xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x80x16xf16>
%1892:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1891 : tensor<2x256x80x16xf16>) outs(%1640, %1638 : tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>)
%1893 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1892#0 : tensor<2x256x80x1xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x80x16xf16>
%collapsed_1442 = tensor.collapse_shape %1893 [[0], [1], [2, 3]] : tensor<2x256x80x16xf16> into tensor<2x256x1280xf16>
%1894 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1442 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1895 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1890, %1894 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1896 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1895, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1897 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1896 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1898 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1897 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x256x1280xf16>
%1899 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1898, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1900 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1899, %1894 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%1901 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_675 : tensor<1280x80x1xf16>) outs(%1651 : tensor<1280x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16xf16>
%collapsed_1443 = tensor.collapse_shape %1901 [[0], [1, 2]] : tensor<1280x80x16xf16> into tensor<1280x1280xf16>
%1902 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_674, %collapsed_1443 : tensor<1280x1280xi8>, tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280xf16>
%1903 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%1902 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%collapsed_1444 = tensor.collapse_shape %1900 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16>
%1904 = linalg.matmul ins(%collapsed_1444, %1903 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%1905 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_673, %1904 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1655 : tensor<512x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<512x1280xf16>
%expanded_1445 = tensor.expand_shape %1905 [[0, 1, 2], [3]] : tensor<512x1280xf16> into tensor<2x16x16x1280xf16>
%1906 = linalg.generic {indexing_maps = [#map1, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1445 : tensor<2x16x16x1280xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1280x16x16xf16>
%1907 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1906, %1614 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%collapsed_1446 = tensor.collapse_shape %1907 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16>
%expanded_1447 = tensor.expand_shape %collapsed_1446 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16>
%1908 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1447 : tensor<2x32x40x256xf16>) outs(%1552 : tensor<2x32x40x256xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x40x256xf32>
%1909 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1908 : tensor<2x32x40x256xf32>) outs(%1554 : tensor<2x32x40x256xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x256xf64>
%1910 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1909 : tensor<2x32x40x256xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1911 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1910 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_962 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1912 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1909, %1911 : tensor<2x32x40x256xf64>, tensor<2x32x1x1xf64>) outs(%1554 : tensor<2x32x40x256xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x256xf64>
%1913 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1912, %1912 : tensor<2x32x40x256xf64>, tensor<2x32x40x256xf64>) outs(%1554 : tensor<2x32x40x256xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x256xf64>
%1914 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1913 : tensor<2x32x40x256xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1915 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1914 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_962 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1916 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1915 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1917 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1908 : tensor<2x32x40x256xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1918 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1917 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_963 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1919 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1916, %cst_938 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%1920 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1919 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1921 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1447, %1918 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%1552 : tensor<2x32x40x256xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x40x256xf32>
%1922 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1921, %1920 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%1552 : tensor<2x32x40x256xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x40x256xf32>
%collapsed_1448 = tensor.collapse_shape %1922 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32>
%expanded_1449 = tensor.expand_shape %collapsed_1448 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32>
%expanded_1450 = tensor.expand_shape %cst_672 [[0, 1, 2, 3]] : tensor<1280xf16> into tensor<1x1280x1x1xf16>
%expanded_1451 = tensor.expand_shape %cst_671 [[0, 1, 2, 3]] : tensor<1280xf16> into tensor<1x1280x1x1xf16>
%1923 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1449, %expanded_1451 : tensor<2x1280x16x16xf32>, tensor<1x1280x1x1xf16>) outs(%1569 : tensor<2x1280x16x16xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1280x16x16xf32>
%1924 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1923, %expanded_1450 : tensor<2x1280x16x16xf32>, tensor<1x1280x1x1xf16>) outs(%1569 : tensor<2x1280x16x16xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1280x16x16xf32>
%1925 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1924 : tensor<2x1280x16x16xf32>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1926 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1925 : tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.negf %in : f16
%6235 = math.exp %6234 : f16
%6236 = arith.addf %6235, %cst_947 : f16
%6237 = arith.divf %cst_947, %6236 : f16
linalg.yield %6237 : f16
} -> tensor<2x1280x16x16xf16>
%1927 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1926, %1925 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%expanded_1452 = tensor.expand_shape %1927 [[0], [1, 2], [3], [4]] : tensor<2x1280x16x16xf16> into tensor<2x80x16x16x16xf16>
%1928 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1452 : tensor<2x80x16x16x16xf16>) outs(%1575 : tensor<2x80x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x80x16x16x16xf16>
%1929:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1928 : tensor<2x80x16x16x16xf16>) outs(%1580, %1578 : tensor<2x80x1x16x16xf16>, tensor<2x80x1x16x16xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x80x1x16x16xf16>, tensor<2x80x1x16x16xi64>)
%1930 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1929#0 : tensor<2x80x1x16x16xf16>) outs(%1575 : tensor<2x80x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x80x16x16x16xf16>
%collapsed_1453 = tensor.collapse_shape %1930 [[0], [1, 2], [3], [4]] : tensor<2x80x16x16x16xf16> into tensor<2x1280x16x16xf16>
%1931 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1453 : tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1932 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1927, %1931 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1933 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1932, %cst_935 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1934 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1933 : tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1935 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1934 : tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1280x16x16xf16>
%1936 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1935, %cst_935 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1937 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1936, %1931 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1938 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_670 : tensor<1280x80x1x3x3xf16>) outs(%1590 : tensor<1280x80x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16x3x3xf16>
%collapsed_1454 = tensor.collapse_shape %1938 [[0], [1, 2], [3], [4]] : tensor<1280x80x16x3x3xf16> into tensor<1280x1280x3x3xf16>
%1939 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_669, %collapsed_1454 : tensor<1280x1280x3x3xi8>, tensor<1280x1280x3x3xf16>) outs(%1592 : tensor<1280x1280x3x3xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280x3x3xf16>
%padded_1455 = tensor.pad %1937 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x1280x16x16xf16> to tensor<2x1280x18x18xf16>
%1940 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_668 : tensor<1280xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1280x16x16xf16>
%1941 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1455, %1939 : tensor<2x1280x18x18xf16>, tensor<1280x1280x3x3xf16>) outs(%1940 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16>
%1942 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_667 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%1943 = linalg.matmul ins(%81, %1942 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%20 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
%1944 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_666, %1943 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%19 : tensor<2x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280xf16>
%expanded_1456 = tensor.expand_shape %1944 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16>
%1945 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1941, %expanded_1456 : tensor<2x1280x16x16xf16>, tensor<2x1280x1x1xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%collapsed_1457 = tensor.collapse_shape %1945 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16>
%expanded_1458 = tensor.expand_shape %collapsed_1457 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16>
%1946 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1458 : tensor<2x32x40x256xf16>) outs(%1552 : tensor<2x32x40x256xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x40x256xf32>
%1947 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1946 : tensor<2x32x40x256xf32>) outs(%1554 : tensor<2x32x40x256xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x256xf64>
%1948 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1947 : tensor<2x32x40x256xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1949 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1948 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_962 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1950 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1947, %1949 : tensor<2x32x40x256xf64>, tensor<2x32x1x1xf64>) outs(%1554 : tensor<2x32x40x256xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x256xf64>
%1951 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1950, %1950 : tensor<2x32x40x256xf64>, tensor<2x32x40x256xf64>) outs(%1554 : tensor<2x32x40x256xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x256xf64>
%1952 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1951 : tensor<2x32x40x256xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1953 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1952 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_962 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1954 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1953 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1955 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1946 : tensor<2x32x40x256xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1956 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1955 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_963 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1957 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1954, %cst_938 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%1958 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1957 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1959 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1458, %1956 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%1552 : tensor<2x32x40x256xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x40x256xf32>
%1960 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1959, %1958 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%1552 : tensor<2x32x40x256xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x40x256xf32>
%collapsed_1459 = tensor.collapse_shape %1960 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32>
%expanded_1460 = tensor.expand_shape %collapsed_1459 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32>
%expanded_1461 = tensor.expand_shape %cst_665 [[0, 1, 2, 3]] : tensor<1280xf16> into tensor<1x1280x1x1xf16>
%expanded_1462 = tensor.expand_shape %cst_664 [[0, 1, 2, 3]] : tensor<1280xf16> into tensor<1x1280x1x1xf16>
%1961 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1460, %expanded_1462 : tensor<2x1280x16x16xf32>, tensor<1x1280x1x1xf16>) outs(%1569 : tensor<2x1280x16x16xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1280x16x16xf32>
%1962 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1961, %expanded_1461 : tensor<2x1280x16x16xf32>, tensor<1x1280x1x1xf16>) outs(%1569 : tensor<2x1280x16x16xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1280x16x16xf32>
%1963 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1962 : tensor<2x1280x16x16xf32>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1964 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1963 : tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.negf %in : f16
%6235 = math.exp %6234 : f16
%6236 = arith.addf %6235, %cst_947 : f16
%6237 = arith.divf %cst_947, %6236 : f16
linalg.yield %6237 : f16
} -> tensor<2x1280x16x16xf16>
%1965 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1964, %1963 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%expanded_1463 = tensor.expand_shape %1965 [[0], [1, 2], [3], [4]] : tensor<2x1280x16x16xf16> into tensor<2x80x16x16x16xf16>
%1966 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1463 : tensor<2x80x16x16x16xf16>) outs(%1575 : tensor<2x80x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x80x16x16x16xf16>
%1967:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%1966 : tensor<2x80x16x16x16xf16>) outs(%1580, %1578 : tensor<2x80x1x16x16xf16>, tensor<2x80x1x16x16xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x80x1x16x16xf16>, tensor<2x80x1x16x16xi64>)
%1968 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%1967#0 : tensor<2x80x1x16x16xf16>) outs(%1575 : tensor<2x80x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x80x16x16x16xf16>
%collapsed_1464 = tensor.collapse_shape %1968 [[0], [1, 2], [3], [4]] : tensor<2x80x16x16x16xf16> into tensor<2x1280x16x16xf16>
%1969 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1464 : tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1970 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1965, %1969 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1971 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1970, %cst_935 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1972 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1971 : tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1973 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1972 : tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1280x16x16xf16>
%1974 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1973, %cst_935 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1975 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1974, %1969 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1976 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_663 : tensor<1280x80x1x3x3xf16>) outs(%1590 : tensor<1280x80x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16x3x3xf16>
%collapsed_1465 = tensor.collapse_shape %1976 [[0], [1, 2], [3], [4]] : tensor<1280x80x16x3x3xf16> into tensor<1280x1280x3x3xf16>
%1977 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_662, %collapsed_1465 : tensor<1280x1280x3x3xi8>, tensor<1280x1280x3x3xf16>) outs(%1592 : tensor<1280x1280x3x3xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280x3x3xf16>
%padded_1466 = tensor.pad %1975 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x1280x16x16xf16> to tensor<2x1280x18x18xf16>
%1978 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_661 : tensor<1280xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1280x16x16xf16>
%1979 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1466, %1977 : tensor<2x1280x18x18xf16>, tensor<1280x1280x3x3xf16>) outs(%1978 : tensor<2x1280x16x16xf16>) -> tensor<2x1280x16x16xf16>
%1980 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1907, %1979 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%1981 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1980, %cst_924 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%collapsed_1467 = tensor.collapse_shape %1981 [[0], [1], [2, 3]] : tensor<2x1280x16x16xf16> into tensor<2x1280x256xf16>
%expanded_1468 = tensor.expand_shape %collapsed_1467 [[0], [1, 2], [3]] : tensor<2x1280x256xf16> into tensor<2x32x40x256xf16>
%1982 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1468 : tensor<2x32x40x256xf16>) outs(%1552 : tensor<2x32x40x256xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x40x256xf32>
%1983 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1982 : tensor<2x32x40x256xf32>) outs(%1554 : tensor<2x32x40x256xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x256xf64>
%1984 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1983 : tensor<2x32x40x256xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1985 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1984 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_962 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1986 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1983, %1985 : tensor<2x32x40x256xf64>, tensor<2x32x1x1xf64>) outs(%1554 : tensor<2x32x40x256xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x256xf64>
%1987 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1986, %1986 : tensor<2x32x40x256xf64>, tensor<2x32x40x256xf64>) outs(%1554 : tensor<2x32x40x256xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x256xf64>
%1988 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1987 : tensor<2x32x40x256xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1989 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1988 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_962 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%1990 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1989 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1991 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1982 : tensor<2x32x40x256xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1992 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1991 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_963 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1993 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1990, %cst_923 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%1994 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1993 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%1995 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1468, %1992 : tensor<2x32x40x256xf16>, tensor<2x32x1x1xf32>) outs(%1552 : tensor<2x32x40x256xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x40x256xf32>
%1996 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1995, %1994 : tensor<2x32x40x256xf32>, tensor<2x32x1x1xf32>) outs(%1552 : tensor<2x32x40x256xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x40x256xf32>
%collapsed_1469 = tensor.collapse_shape %1996 [[0], [1, 2], [3]] : tensor<2x32x40x256xf32> into tensor<2x1280x256xf32>
%expanded_1470 = tensor.expand_shape %collapsed_1469 [[0], [1], [2, 3]] : tensor<2x1280x256xf32> into tensor<2x1280x16x16xf32>
%expanded_1471 = tensor.expand_shape %cst_660 [[0, 1, 2, 3]] : tensor<1280xf16> into tensor<1x1280x1x1xf16>
%expanded_1472 = tensor.expand_shape %cst_659 [[0, 1, 2, 3]] : tensor<1280xf16> into tensor<1x1280x1x1xf16>
%1997 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1470, %expanded_1472 : tensor<2x1280x16x16xf32>, tensor<1x1280x1x1xf16>) outs(%1569 : tensor<2x1280x16x16xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1280x16x16xf32>
%1998 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1997, %expanded_1471 : tensor<2x1280x16x16xf32>, tensor<1x1280x1x1xf16>) outs(%1569 : tensor<2x1280x16x16xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1280x16x16xf32>
%1999 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1998 : tensor<2x1280x16x16xf32>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%2000 = linalg.generic {indexing_maps = [#map1, #map16], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1999 : tensor<2x1280x16x16xf16>) outs(%1633 : tensor<2x16x16x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x16x16x1280xf16>
%collapsed_1473 = tensor.collapse_shape %2000 [[0], [1, 2], [3]] : tensor<2x16x16x1280xf16> into tensor<2x256x1280xf16>
%expanded_1474 = tensor.expand_shape %collapsed_1473 [[0], [1], [2, 3]] : tensor<2x256x1280xf16> into tensor<2x256x80x16xf16>
%2001 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1474 : tensor<2x256x80x16xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x80x16xf16>
%2002:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2001 : tensor<2x256x80x16xf16>) outs(%1640, %1638 : tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>)
%2003 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2002#0 : tensor<2x256x80x1xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x80x16xf16>
%collapsed_1475 = tensor.collapse_shape %2003 [[0], [1], [2, 3]] : tensor<2x256x80x16xf16> into tensor<2x256x1280xf16>
%2004 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1475 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2005 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1473, %2004 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2006 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2005, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2007 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2006 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2008 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2007 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x256x1280xf16>
%2009 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2008, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2010 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2009, %2004 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2011 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_658 : tensor<1280x80x1xf16>) outs(%1651 : tensor<1280x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16xf16>
%collapsed_1476 = tensor.collapse_shape %2011 [[0], [1, 2]] : tensor<1280x80x16xf16> into tensor<1280x1280xf16>
%2012 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_657, %collapsed_1476 : tensor<1280x1280xi8>, tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280xf16>
%2013 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%2012 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%collapsed_1477 = tensor.collapse_shape %2010 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16>
%2014 = linalg.matmul ins(%collapsed_1477, %2013 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%2015 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_656, %2014 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1655 : tensor<512x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<512x1280xf16>
%expanded_1478 = tensor.expand_shape %2015 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16>
%2016 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1478 : tensor<2x256x1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x256x1280xf32>
%2017 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2016 : tensor<2x256x1280xf32>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%2018 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2017 : tensor<2x256x1280xf64>) outs(%1664 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%2019 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2018 : tensor<2x256x1xf64>) outs(%1663 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_970 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%2020 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2017, %2019 : tensor<2x256x1280xf64>, tensor<2x256x1xf64>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%2021 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2020, %2020 : tensor<2x256x1280xf64>, tensor<2x256x1280xf64>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%2022 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2021 : tensor<2x256x1280xf64>) outs(%1664 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%2023 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2022 : tensor<2x256x1xf64>) outs(%1663 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_970 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%2024 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2023 : tensor<2x256x1xf64>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%2025 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2016 : tensor<2x256x1280xf32>) outs(%1673 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%2026 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2025 : tensor<2x256x1xf32>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_971 : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%2027 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2024, %cst_938 : tensor<2x256x1xf32>, tensor<f16>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1xf32>
%2028 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2027 : tensor<2x256x1xf32>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%2029 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1478, %2026 : tensor<2x256x1280xf16>, tensor<2x256x1xf32>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%2030 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2029, %2028 : tensor<2x256x1280xf32>, tensor<2x256x1xf32>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1280xf32>
%2031 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2030, %cst_655 : tensor<2x256x1280xf32>, tensor<1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%2032 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2031, %cst_654 : tensor<2x256x1280xf32>, tensor<1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%2033 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2032 : tensor<2x256x1280xf32>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%expanded_1479 = tensor.expand_shape %2033 [[0], [1], [2, 3]] : tensor<2x256x1280xf16> into tensor<2x256x80x16xf16>
%2034 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1479 : tensor<2x256x80x16xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x80x16xf16>
%2035:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2034 : tensor<2x256x80x16xf16>) outs(%1640, %1638 : tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>)
%2036 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2035#0 : tensor<2x256x80x1xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x80x16xf16>
%collapsed_1480 = tensor.collapse_shape %2036 [[0], [1], [2, 3]] : tensor<2x256x80x16xf16> into tensor<2x256x1280xf16>
%2037 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1480 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2038 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2033, %2037 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2039 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2038, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2040 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2039 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2041 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2040 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x256x1280xf16>
%2042 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2041, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2043 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2042, %2037 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2044 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_653 : tensor<1280x80x1xf16>) outs(%1651 : tensor<1280x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16xf16>
%collapsed_1481 = tensor.collapse_shape %2044 [[0], [1, 2]] : tensor<1280x80x16xf16> into tensor<1280x1280xf16>
%2045 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_652, %collapsed_1481 : tensor<1280x1280xi8>, tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280xf16>
%2046 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%2045 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%collapsed_1482 = tensor.collapse_shape %2043 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16>
%2047 = linalg.matmul ins(%collapsed_1482, %2046 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%expanded_1483 = tensor.expand_shape %2047 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x20x64xf16>
%2048 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1483 : tensor<2x256x20x64xf16>) outs(%1697 : tensor<2x20x256x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x256x64xf16>
%collapsed_1484 = tensor.collapse_shape %2048 [[0, 1], [2], [3]] : tensor<2x20x256x64xf16> into tensor<40x256x64xf16>
%2049 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_651 : tensor<1280x80x1xf16>) outs(%1651 : tensor<1280x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16xf16>
%collapsed_1485 = tensor.collapse_shape %2049 [[0], [1, 2]] : tensor<1280x80x16xf16> into tensor<1280x1280xf16>
%2050 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_650, %collapsed_1485 : tensor<1280x1280xi8>, tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280xf16>
%2051 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%2050 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%2052 = linalg.matmul ins(%collapsed_1482, %2051 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%2053 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_649 : tensor<1280x80x1xf16>) outs(%1651 : tensor<1280x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16xf16>
%collapsed_1486 = tensor.collapse_shape %2053 [[0], [1, 2]] : tensor<1280x80x16xf16> into tensor<1280x1280xf16>
%2054 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_648, %collapsed_1486 : tensor<1280x1280xi8>, tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280xf16>
%2055 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%2054 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%2056 = linalg.matmul ins(%collapsed_1482, %2055 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%expanded_1487 = tensor.expand_shape %2052 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x20x64xf16>
%2057 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1487 : tensor<2x256x20x64xf16>) outs(%1697 : tensor<2x20x256x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x256x64xf16>
%collapsed_1488 = tensor.collapse_shape %2057 [[0, 1], [2], [3]] : tensor<2x20x256x64xf16> into tensor<40x256x64xf16>
%expanded_1489 = tensor.expand_shape %2056 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x20x64xf16>
%2058 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1489 : tensor<2x256x20x64xf16>) outs(%1697 : tensor<2x20x256x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x256x64xf16>
%collapsed_1490 = tensor.collapse_shape %2058 [[0, 1], [2], [3]] : tensor<2x20x256x64xf16> into tensor<40x256x64xf16>
%2059 = linalg.generic {indexing_maps = [#map18, #map23], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1488 : tensor<40x256x64xf16>) outs(%1711 : tensor<40x64x256xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<40x64x256xf16>
%2060 = linalg.batch_matmul ins(%collapsed_1484, %2059 : tensor<40x256x64xf16>, tensor<40x64x256xf16>) outs(%1710 : tensor<40x256x256xf16>) -> tensor<40x256x256xf16>
%2061 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2060 : tensor<40x256x256xf16>) outs(%1709 : tensor<40x256x256xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.mulf %in, %cst_961 : f16
linalg.yield %6234 : f16
} -> tensor<40x256x256xf16>
%2062 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2061, %1710 : tensor<40x256x256xf16>, tensor<40x256x256xf16>) outs(%1709 : tensor<40x256x256xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in_2625, %cst_946 : f16
%6235 = arith.addf %in, %6234 : f16
linalg.yield %6235 : f16
} -> tensor<40x256x256xf16>
%2063:2 = linalg.generic {indexing_maps = [#map18, #map20, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2062 : tensor<40x256x256xf16>) outs(%1719, %1717 : tensor<40x256x1xf16>, tensor<40x256x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<40x256x1xf16>, tensor<40x256x1xi64>)
%2064 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2062, %2063#0 : tensor<40x256x256xf16>, tensor<40x256x1xf16>) outs(%1709 : tensor<40x256x256xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<40x256x256xf16>
%2065 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2064 : tensor<40x256x256xf16>) outs(%1709 : tensor<40x256x256xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.exp %in : f16
linalg.yield %6234 : f16
} -> tensor<40x256x256xf16>
%2066 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2065 : tensor<40x256x256xf16>) outs(%1723 : tensor<40x256x1xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.addf %in, %out : f16
linalg.yield %6234 : f16
} -> tensor<40x256x1xf16>
%2067 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2065, %2066 : tensor<40x256x256xf16>, tensor<40x256x1xf16>) outs(%1709 : tensor<40x256x256xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<40x256x256xf16>
%2068 = linalg.batch_matmul ins(%2067, %collapsed_1490 : tensor<40x256x256xf16>, tensor<40x256x64xf16>) outs(%1727 : tensor<40x256x64xf16>) -> tensor<40x256x64xf16>
%expanded_1491 = tensor.expand_shape %2068 [[0, 1], [2], [3]] : tensor<40x256x64xf16> into tensor<2x20x256x64xf16>
%2069 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1491 : tensor<2x20x256x64xf16>) outs(%1729 : tensor<2x256x20x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x20x64xf16>
%collapsed_1492 = tensor.collapse_shape %2069 [[0], [1], [2, 3]] : tensor<2x256x20x64xf16> into tensor<2x256x1280xf16>
%expanded_1493 = tensor.expand_shape %collapsed_1492 [[0], [1], [2, 3]] : tensor<2x256x1280xf16> into tensor<2x256x80x16xf16>
%2070 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1493 : tensor<2x256x80x16xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x80x16xf16>
%2071:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2070 : tensor<2x256x80x16xf16>) outs(%1640, %1638 : tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>)
%2072 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2071#0 : tensor<2x256x80x1xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x80x16xf16>
%collapsed_1494 = tensor.collapse_shape %2072 [[0], [1], [2, 3]] : tensor<2x256x80x16xf16> into tensor<2x256x1280xf16>
%2073 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1494 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2074 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1492, %2073 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2075 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2074, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2076 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2075 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2077 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2076 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x256x1280xf16>
%2078 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2077, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2079 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2078, %2073 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2080 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_647 : tensor<1280x80x1xf16>) outs(%1651 : tensor<1280x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16xf16>
%collapsed_1495 = tensor.collapse_shape %2080 [[0], [1, 2]] : tensor<1280x80x16xf16> into tensor<1280x1280xf16>
%2081 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_646, %collapsed_1495 : tensor<1280x1280xi8>, tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280xf16>
%2082 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%2081 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%collapsed_1496 = tensor.collapse_shape %2079 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16>
%2083 = linalg.matmul ins(%collapsed_1496, %2082 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%2084 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_645, %2083 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1655 : tensor<512x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<512x1280xf16>
%expanded_1497 = tensor.expand_shape %2084 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16>
%2085 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1497, %expanded_1478 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2086 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2085 : tensor<2x256x1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x256x1280xf32>
%2087 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2086 : tensor<2x256x1280xf32>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%2088 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2087 : tensor<2x256x1280xf64>) outs(%1664 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%2089 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2088 : tensor<2x256x1xf64>) outs(%1663 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_970 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%2090 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2087, %2089 : tensor<2x256x1280xf64>, tensor<2x256x1xf64>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%2091 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2090, %2090 : tensor<2x256x1280xf64>, tensor<2x256x1280xf64>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%2092 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2091 : tensor<2x256x1280xf64>) outs(%1664 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%2093 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2092 : tensor<2x256x1xf64>) outs(%1663 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_970 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%2094 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2093 : tensor<2x256x1xf64>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%2095 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2086 : tensor<2x256x1280xf32>) outs(%1673 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%2096 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2095 : tensor<2x256x1xf32>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_971 : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%2097 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2094, %cst_938 : tensor<2x256x1xf32>, tensor<f16>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1xf32>
%2098 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2097 : tensor<2x256x1xf32>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%2099 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2085, %2096 : tensor<2x256x1280xf16>, tensor<2x256x1xf32>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%2100 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2099, %2098 : tensor<2x256x1280xf32>, tensor<2x256x1xf32>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1280xf32>
%2101 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2100, %cst_644 : tensor<2x256x1280xf32>, tensor<1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%2102 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2101, %cst_643 : tensor<2x256x1280xf32>, tensor<1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%2103 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2102 : tensor<2x256x1280xf32>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%expanded_1498 = tensor.expand_shape %2103 [[0], [1], [2, 3]] : tensor<2x256x1280xf16> into tensor<2x256x80x16xf16>
%2104 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1498 : tensor<2x256x80x16xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x80x16xf16>
%2105:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2104 : tensor<2x256x80x16xf16>) outs(%1640, %1638 : tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>)
%2106 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2105#0 : tensor<2x256x80x1xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x80x16xf16>
%collapsed_1499 = tensor.collapse_shape %2106 [[0], [1], [2, 3]] : tensor<2x256x80x16xf16> into tensor<2x256x1280xf16>
%2107 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1499 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2108 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2103, %2107 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2109 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2108, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2110 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2109 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2111 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2110 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x256x1280xf16>
%2112 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2111, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2113 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2112, %2107 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2114 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_642 : tensor<1280x80x1xf16>) outs(%1651 : tensor<1280x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16xf16>
%collapsed_1500 = tensor.collapse_shape %2114 [[0], [1, 2]] : tensor<1280x80x16xf16> into tensor<1280x1280xf16>
%2115 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_641, %collapsed_1500 : tensor<1280x1280xi8>, tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280xf16>
%2116 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%2115 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%collapsed_1501 = tensor.collapse_shape %2113 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16>
%2117 = linalg.matmul ins(%collapsed_1501, %2116 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%expanded_1502 = tensor.expand_shape %2117 [[0, 1], [2, 3]] : tensor<512x1280xf16> into tensor<2x256x20x64xf16>
%2118 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1502 : tensor<2x256x20x64xf16>) outs(%1697 : tensor<2x20x256x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x256x64xf16>
%collapsed_1503 = tensor.collapse_shape %2118 [[0, 1], [2], [3]] : tensor<2x20x256x64xf16> into tensor<40x256x64xf16>
%2119 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_640 : tensor<1280x64x1xf16>) outs(%1780 : tensor<1280x64x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x64x16xf16>
%collapsed_1504 = tensor.collapse_shape %2119 [[0], [1, 2]] : tensor<1280x64x16xf16> into tensor<1280x1024xf16>
%2120 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_639, %collapsed_1504 : tensor<1280x1024xi8>, tensor<1280x1024xf16>) outs(%1782 : tensor<1280x1024xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1024xf16>
%2121 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%2120 : tensor<1280x1024xf16>) outs(%1784 : tensor<1024x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1024x1280xf16>
%2122 = linalg.matmul ins(%collapsed_1050, %2121 : tensor<154x1024xf16>, tensor<1024x1280xf16>) outs(%1787 : tensor<154x1280xf16>) -> tensor<154x1280xf16>
%2123 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_638 : tensor<1280x64x1xf16>) outs(%1780 : tensor<1280x64x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x64x16xf16>
%collapsed_1505 = tensor.collapse_shape %2123 [[0], [1, 2]] : tensor<1280x64x16xf16> into tensor<1280x1024xf16>
%2124 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_637, %collapsed_1505 : tensor<1280x1024xi8>, tensor<1280x1024xf16>) outs(%1782 : tensor<1280x1024xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1024xf16>
%2125 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%2124 : tensor<1280x1024xf16>) outs(%1784 : tensor<1024x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1024x1280xf16>
%2126 = linalg.matmul ins(%collapsed_1050, %2125 : tensor<154x1024xf16>, tensor<1024x1280xf16>) outs(%1787 : tensor<154x1280xf16>) -> tensor<154x1280xf16>
%expanded_1506 = tensor.expand_shape %2122 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x20x64xf16>
%2127 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1506 : tensor<2x77x20x64xf16>) outs(%1793 : tensor<2x20x77x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x77x64xf16>
%collapsed_1507 = tensor.collapse_shape %2127 [[0, 1], [2], [3]] : tensor<2x20x77x64xf16> into tensor<40x77x64xf16>
%expanded_1508 = tensor.expand_shape %2126 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x20x64xf16>
%2128 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1508 : tensor<2x77x20x64xf16>) outs(%1793 : tensor<2x20x77x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x20x77x64xf16>
%collapsed_1509 = tensor.collapse_shape %2128 [[0, 1], [2], [3]] : tensor<2x20x77x64xf16> into tensor<40x77x64xf16>
%2129 = linalg.generic {indexing_maps = [#map18, #map23], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1507 : tensor<40x77x64xf16>) outs(%1798 : tensor<40x64x77xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<40x64x77xf16>
%2130 = linalg.batch_matmul ins(%collapsed_1503, %2129 : tensor<40x256x64xf16>, tensor<40x64x77xf16>) outs(%1797 : tensor<40x256x77xf16>) -> tensor<40x256x77xf16>
%2131 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2130 : tensor<40x256x77xf16>) outs(%1796 : tensor<40x256x77xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.mulf %in, %cst_961 : f16
linalg.yield %6234 : f16
} -> tensor<40x256x77xf16>
%2132 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2131, %1797 : tensor<40x256x77xf16>, tensor<40x256x77xf16>) outs(%1796 : tensor<40x256x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in_2625, %cst_946 : f16
%6235 = arith.addf %in, %6234 : f16
linalg.yield %6235 : f16
} -> tensor<40x256x77xf16>
%2133:2 = linalg.generic {indexing_maps = [#map18, #map20, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2132 : tensor<40x256x77xf16>) outs(%1719, %1717 : tensor<40x256x1xf16>, tensor<40x256x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<40x256x1xf16>, tensor<40x256x1xi64>)
%2134 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2132, %2133#0 : tensor<40x256x77xf16>, tensor<40x256x1xf16>) outs(%1796 : tensor<40x256x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<40x256x77xf16>
%2135 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2134 : tensor<40x256x77xf16>) outs(%1796 : tensor<40x256x77xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.exp %in : f16
linalg.yield %6234 : f16
} -> tensor<40x256x77xf16>
%2136 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2135 : tensor<40x256x77xf16>) outs(%1723 : tensor<40x256x1xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.addf %in, %out : f16
linalg.yield %6234 : f16
} -> tensor<40x256x1xf16>
%2137 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2135, %2136 : tensor<40x256x77xf16>, tensor<40x256x1xf16>) outs(%1796 : tensor<40x256x77xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<40x256x77xf16>
%2138 = linalg.batch_matmul ins(%2137, %collapsed_1509 : tensor<40x256x77xf16>, tensor<40x77x64xf16>) outs(%1727 : tensor<40x256x64xf16>) -> tensor<40x256x64xf16>
%expanded_1510 = tensor.expand_shape %2138 [[0, 1], [2], [3]] : tensor<40x256x64xf16> into tensor<2x20x256x64xf16>
%2139 = linalg.generic {indexing_maps = [#map1, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1510 : tensor<2x20x256x64xf16>) outs(%1729 : tensor<2x256x20x64xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x20x64xf16>
%collapsed_1511 = tensor.collapse_shape %2139 [[0], [1], [2, 3]] : tensor<2x256x20x64xf16> into tensor<2x256x1280xf16>
%expanded_1512 = tensor.expand_shape %collapsed_1511 [[0], [1], [2, 3]] : tensor<2x256x1280xf16> into tensor<2x256x80x16xf16>
%2140 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1512 : tensor<2x256x80x16xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x80x16xf16>
%2141:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2140 : tensor<2x256x80x16xf16>) outs(%1640, %1638 : tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>)
%2142 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2141#0 : tensor<2x256x80x1xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x80x16xf16>
%collapsed_1513 = tensor.collapse_shape %2142 [[0], [1], [2, 3]] : tensor<2x256x80x16xf16> into tensor<2x256x1280xf16>
%2143 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1513 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2144 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1511, %2143 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2145 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2144, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2146 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2145 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2147 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2146 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x256x1280xf16>
%2148 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2147, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2149 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2148, %2143 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2150 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_636 : tensor<1280x80x1xf16>) outs(%1651 : tensor<1280x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16xf16>
%collapsed_1514 = tensor.collapse_shape %2150 [[0], [1, 2]] : tensor<1280x80x16xf16> into tensor<1280x1280xf16>
%2151 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_635, %collapsed_1514 : tensor<1280x1280xi8>, tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280xf16>
%2152 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%2151 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%collapsed_1515 = tensor.collapse_shape %2149 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16>
%2153 = linalg.matmul ins(%collapsed_1515, %2152 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%2154 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_634, %2153 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1655 : tensor<512x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<512x1280xf16>
%expanded_1516 = tensor.expand_shape %2154 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16>
%2155 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1516, %2085 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2156 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2155 : tensor<2x256x1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x256x1280xf32>
%2157 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2156 : tensor<2x256x1280xf32>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%2158 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2157 : tensor<2x256x1280xf64>) outs(%1664 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%2159 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2158 : tensor<2x256x1xf64>) outs(%1663 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_970 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%2160 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2157, %2159 : tensor<2x256x1280xf64>, tensor<2x256x1xf64>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%2161 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2160, %2160 : tensor<2x256x1280xf64>, tensor<2x256x1280xf64>) outs(%1661 : tensor<2x256x1280xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1280xf64>
%2162 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2161 : tensor<2x256x1280xf64>) outs(%1664 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%2163 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2162 : tensor<2x256x1xf64>) outs(%1663 : tensor<2x256x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_970 : f64
linalg.yield %6234 : f64
} -> tensor<2x256x1xf64>
%2164 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2163 : tensor<2x256x1xf64>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%2165 = linalg.generic {indexing_maps = [#map18, #map20], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2156 : tensor<2x256x1280xf32>) outs(%1673 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%2166 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2165 : tensor<2x256x1xf32>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_971 : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%2167 = linalg.generic {indexing_maps = [#map20, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2164, %cst_938 : tensor<2x256x1xf32>, tensor<f16>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1xf32>
%2168 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2167 : tensor<2x256x1xf32>) outs(%1671 : tensor<2x256x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1xf32>
%2169 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2155, %2166 : tensor<2x256x1280xf16>, tensor<2x256x1xf32>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%2170 = linalg.generic {indexing_maps = [#map18, #map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2169, %2168 : tensor<2x256x1280xf32>, tensor<2x256x1xf32>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x256x1280xf32>
%2171 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2170, %cst_633 : tensor<2x256x1280xf32>, tensor<1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%2172 = linalg.generic {indexing_maps = [#map18, #map21, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2171, %cst_632 : tensor<2x256x1280xf32>, tensor<1280xf16>) outs(%1659 : tensor<2x256x1280xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x256x1280xf32>
%2173 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2172 : tensor<2x256x1280xf32>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%expanded_1517 = tensor.expand_shape %2173 [[0], [1], [2, 3]] : tensor<2x256x1280xf16> into tensor<2x256x80x16xf16>
%2174 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1517 : tensor<2x256x80x16xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x80x16xf16>
%2175:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2174 : tensor<2x256x80x16xf16>) outs(%1640, %1638 : tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>)
%2176 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2175#0 : tensor<2x256x80x1xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x80x16xf16>
%collapsed_1518 = tensor.collapse_shape %2176 [[0], [1], [2, 3]] : tensor<2x256x80x16xf16> into tensor<2x256x1280xf16>
%2177 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1518 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2178 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2173, %2177 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2179 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2178, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2180 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2179 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2181 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2180 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x256x1280xf16>
%2182 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2181, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2183 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2182, %2177 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2184 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_631 : tensor<10240x80x1xf16>) outs(%1854 : tensor<10240x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<10240x80x16xf16>
%collapsed_1519 = tensor.collapse_shape %2184 [[0], [1, 2]] : tensor<10240x80x16xf16> into tensor<10240x1280xf16>
%2185 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_630, %collapsed_1519 : tensor<10240x1280xi8>, tensor<10240x1280xf16>) outs(%1856 : tensor<10240x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<10240x1280xf16>
%2186 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%2185 : tensor<10240x1280xf16>) outs(%1858 : tensor<1280x10240xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x10240xf16>
%collapsed_1520 = tensor.collapse_shape %2183 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16>
%2187 = linalg.matmul ins(%collapsed_1520, %2186 : tensor<512x1280xf16>, tensor<1280x10240xf16>) outs(%1861 : tensor<512x10240xf16>) -> tensor<512x10240xf16>
%2188 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_629, %2187 : tensor<10240xf16>, tensor<512x10240xf16>) outs(%1860 : tensor<512x10240xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<512x10240xf16>
%expanded_1521 = tensor.expand_shape %2188 [[0, 1], [2]] : tensor<512x10240xf16> into tensor<2x256x10240xf16>
%extracted_slice_1522 = tensor.extract_slice %expanded_1521[0, 0, 0] [2, 256, 5120] [1, 1, 1] : tensor<2x256x10240xf16> to tensor<2x256x5120xf16>
%extracted_slice_1523 = tensor.extract_slice %expanded_1521[0, 0, 5120] [2, 256, 5120] [1, 1, 1] : tensor<2x256x10240xf16> to tensor<2x256x5120xf16>
%2189 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1523 : tensor<2x256x5120xf16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.sqrt %cst_949 : f16
%6235 = arith.divf %in, %6234 : f16
%6236 = math.erf %6235 : f16
%6237 = arith.addf %6236, %cst_947 : f16
%6238 = arith.mulf %6237, %cst_950 : f16
%6239 = arith.mulf %in, %6238 : f16
linalg.yield %6239 : f16
} -> tensor<2x256x5120xf16>
%2190 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1522, %2189 : tensor<2x256x5120xf16>, tensor<2x256x5120xf16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x5120xf16>
%expanded_1524 = tensor.expand_shape %2190 [[0], [1], [2, 3]] : tensor<2x256x5120xf16> into tensor<2x256x320x16xf16>
%2191 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1524 : tensor<2x256x320x16xf16>) outs(%1867 : tensor<2x256x320x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x320x16xf16>
%2192:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2191 : tensor<2x256x320x16xf16>) outs(%1872, %1870 : tensor<2x256x320x1xf16>, tensor<2x256x320x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x256x320x1xf16>, tensor<2x256x320x1xi64>)
%2193 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2192#0 : tensor<2x256x320x1xf16>) outs(%1867 : tensor<2x256x320x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x320x16xf16>
%collapsed_1525 = tensor.collapse_shape %2193 [[0], [1], [2, 3]] : tensor<2x256x320x16xf16> into tensor<2x256x5120xf16>
%2194 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1525 : tensor<2x256x5120xf16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x5120xf16>
%2195 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2190, %2194 : tensor<2x256x5120xf16>, tensor<2x256x5120xf16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x5120xf16>
%2196 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2195, %cst_935 : tensor<2x256x5120xf16>, tensor<f16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x5120xf16>
%2197 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2196 : tensor<2x256x5120xf16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x5120xf16>
%2198 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2197 : tensor<2x256x5120xf16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x256x5120xf16>
%2199 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2198, %cst_935 : tensor<2x256x5120xf16>, tensor<f16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x5120xf16>
%2200 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2199, %2194 : tensor<2x256x5120xf16>, tensor<2x256x5120xf16>) outs(%1864 : tensor<2x256x5120xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x5120xf16>
%2201 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_628 : tensor<1280x320x1xf16>) outs(%1882 : tensor<1280x320x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x320x16xf16>
%collapsed_1526 = tensor.collapse_shape %2201 [[0], [1, 2]] : tensor<1280x320x16xf16> into tensor<1280x5120xf16>
%2202 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_627, %collapsed_1526 : tensor<1280x5120xi8>, tensor<1280x5120xf16>) outs(%1884 : tensor<1280x5120xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x5120xf16>
%2203 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%2202 : tensor<1280x5120xf16>) outs(%1886 : tensor<5120x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<5120x1280xf16>
%collapsed_1527 = tensor.collapse_shape %2200 [[0, 1], [2]] : tensor<2x256x5120xf16> into tensor<512x5120xf16>
%2204 = linalg.matmul ins(%collapsed_1527, %2203 : tensor<512x5120xf16>, tensor<5120x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%2205 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_626, %2204 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1655 : tensor<512x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<512x1280xf16>
%expanded_1528 = tensor.expand_shape %2205 [[0, 1], [2]] : tensor<512x1280xf16> into tensor<2x256x1280xf16>
%2206 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1528, %2155 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%expanded_1529 = tensor.expand_shape %2206 [[0], [1], [2, 3]] : tensor<2x256x1280xf16> into tensor<2x256x80x16xf16>
%2207 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1529 : tensor<2x256x80x16xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x80x16xf16>
%2208:2 = linalg.generic {indexing_maps = [#map1, #map17, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2207 : tensor<2x256x80x16xf16>) outs(%1640, %1638 : tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 3 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x256x80x1xf16>, tensor<2x256x80x1xi64>)
%2209 = linalg.generic {indexing_maps = [#map17, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2208#0 : tensor<2x256x80x1xf16>) outs(%1635 : tensor<2x256x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x256x80x16xf16>
%collapsed_1530 = tensor.collapse_shape %2209 [[0], [1], [2, 3]] : tensor<2x256x80x16xf16> into tensor<2x256x1280xf16>
%2210 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1530 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2211 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2206, %2210 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2212 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2211, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2213 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2212 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2214 = linalg.generic {indexing_maps = [#map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2213 : tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x256x1280xf16>
%2215 = linalg.generic {indexing_maps = [#map18, #map19, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2214, %cst_935 : tensor<2x256x1280xf16>, tensor<f16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2216 = linalg.generic {indexing_maps = [#map18, #map18, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2215, %2210 : tensor<2x256x1280xf16>, tensor<2x256x1280xf16>) outs(%1643 : tensor<2x256x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x256x1280xf16>
%2217 = linalg.generic {indexing_maps = [#map20, #map18], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_625 : tensor<1280x80x1xf16>) outs(%1651 : tensor<1280x80x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16xf16>
%collapsed_1531 = tensor.collapse_shape %2217 [[0], [1, 2]] : tensor<1280x80x16xf16> into tensor<1280x1280xf16>
%2218 = linalg.generic {indexing_maps = [#map7, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_624, %collapsed_1531 : tensor<1280x1280xi8>, tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280xf16>
%2219 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%2218 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%collapsed_1532 = tensor.collapse_shape %2216 [[0, 1], [2]] : tensor<2x256x1280xf16> into tensor<512x1280xf16>
%2220 = linalg.matmul ins(%collapsed_1532, %2219 : tensor<512x1280xf16>, tensor<1280x1280xf16>) outs(%1656 : tensor<512x1280xf16>) -> tensor<512x1280xf16>
%2221 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_623, %2220 : tensor<1280xf16>, tensor<512x1280xf16>) outs(%1655 : tensor<512x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<512x1280xf16>
%expanded_1533 = tensor.expand_shape %2221 [[0, 1, 2], [3]] : tensor<512x1280xf16> into tensor<2x16x16x1280xf16>
%2222 = linalg.generic {indexing_maps = [#map1, #map24], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1533 : tensor<2x16x16x1280xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1280x16x16xf16>
%2223 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2222, %1981 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%expanded_1534 = tensor.expand_shape %2223 [[0], [1, 2], [3], [4]] : tensor<2x1280x16x16xf16> into tensor<2x80x16x16x16xf16>
%2224 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1534 : tensor<2x80x16x16x16xf16>) outs(%1575 : tensor<2x80x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x80x16x16x16xf16>
%2225:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%2224 : tensor<2x80x16x16x16xf16>) outs(%1580, %1578 : tensor<2x80x1x16x16xf16>, tensor<2x80x1x16x16xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x80x1x16x16xf16>, tensor<2x80x1x16x16xi64>)
%2226 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%2225#0 : tensor<2x80x1x16x16xf16>) outs(%1575 : tensor<2x80x16x16x16xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x80x16x16x16xf16>
%collapsed_1535 = tensor.collapse_shape %2226 [[0], [1, 2], [3], [4]] : tensor<2x80x16x16x16xf16> into tensor<2x1280x16x16xf16>
%2227 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1535 : tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%2228 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2223, %2227 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%2229 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2228, %cst_935 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%2230 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2229 : tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%2231 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2230 : tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1280x16x16xf16>
%2232 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2231, %cst_935 : tensor<2x1280x16x16xf16>, tensor<f16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%2233 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2232, %2227 : tensor<2x1280x16x16xf16>, tensor<2x1280x16x16xf16>) outs(%1545 : tensor<2x1280x16x16xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x16x16xf16>
%2234 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_622 : tensor<1280x80x1x3x3xf16>) outs(%1590 : tensor<1280x80x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16x3x3xf16>
%collapsed_1536 = tensor.collapse_shape %2234 [[0], [1, 2], [3], [4]] : tensor<1280x80x16x3x3xf16> into tensor<1280x1280x3x3xf16>
%2235 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_621, %collapsed_1536 : tensor<1280x1280x3x3xi8>, tensor<1280x1280x3x3xf16>) outs(%1592 : tensor<1280x1280x3x3xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280x3x3xf16>
%padded_1537 = tensor.pad %2233 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x1280x16x16xf16> to tensor<2x1280x18x18xf16>
%2236 = tensor.empty() : tensor<2x1280x8x8xf16>
%2237 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_620 : tensor<1280xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1280x8x8xf16>
%2238 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%padded_1537, %2235 : tensor<2x1280x18x18xf16>, tensor<1280x1280x3x3xf16>) outs(%2237 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16>
%collapsed_1538 = tensor.collapse_shape %2238 [[0], [1], [2, 3]] : tensor<2x1280x8x8xf16> into tensor<2x1280x64xf16>
%expanded_1539 = tensor.expand_shape %collapsed_1538 [[0], [1, 2], [3]] : tensor<2x1280x64xf16> into tensor<2x32x40x64xf16>
%2239 = tensor.empty() : tensor<2x32x40x64xf32>
%2240 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1539 : tensor<2x32x40x64xf16>) outs(%2239 : tensor<2x32x40x64xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x40x64xf32>
%2241 = tensor.empty() : tensor<2x32x40x64xf64>
%2242 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2240 : tensor<2x32x40x64xf32>) outs(%2241 : tensor<2x32x40x64xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x64xf64>
%2243 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2242 : tensor<2x32x40x64xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%2244 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2243 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_972 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%2245 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2242, %2244 : tensor<2x32x40x64xf64>, tensor<2x32x1x1xf64>) outs(%2241 : tensor<2x32x40x64xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x64xf64>
%2246 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2245, %2245 : tensor<2x32x40x64xf64>, tensor<2x32x40x64xf64>) outs(%2241 : tensor<2x32x40x64xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x64xf64>
%2247 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2246 : tensor<2x32x40x64xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%2248 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2247 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_972 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%2249 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2248 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%2250 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2240 : tensor<2x32x40x64xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%2251 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2250 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_973 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%2252 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2249, %cst_938 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%2253 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2252 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%2254 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1539, %2251 : tensor<2x32x40x64xf16>, tensor<2x32x1x1xf32>) outs(%2239 : tensor<2x32x40x64xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x40x64xf32>
%2255 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2254, %2253 : tensor<2x32x40x64xf32>, tensor<2x32x1x1xf32>) outs(%2239 : tensor<2x32x40x64xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x40x64xf32>
%collapsed_1540 = tensor.collapse_shape %2255 [[0], [1, 2], [3]] : tensor<2x32x40x64xf32> into tensor<2x1280x64xf32>
%expanded_1541 = tensor.expand_shape %collapsed_1540 [[0], [1], [2, 3]] : tensor<2x1280x64xf32> into tensor<2x1280x8x8xf32>
%expanded_1542 = tensor.expand_shape %cst_619 [[0, 1, 2, 3]] : tensor<1280xf16> into tensor<1x1280x1x1xf16>
%expanded_1543 = tensor.expand_shape %cst_618 [[0, 1, 2, 3]] : tensor<1280xf16> into tensor<1x1280x1x1xf16>
%2256 = tensor.empty() : tensor<2x1280x8x8xf32>
%2257 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1541, %expanded_1543 : tensor<2x1280x8x8xf32>, tensor<1x1280x1x1xf16>) outs(%2256 : tensor<2x1280x8x8xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1280x8x8xf32>
%2258 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2257, %expanded_1542 : tensor<2x1280x8x8xf32>, tensor<1x1280x1x1xf16>) outs(%2256 : tensor<2x1280x8x8xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1280x8x8xf32>
%2259 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2258 : tensor<2x1280x8x8xf32>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%2260 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2259 : tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.negf %in : f16
%6235 = math.exp %6234 : f16
%6236 = arith.addf %6235, %cst_947 : f16
%6237 = arith.divf %cst_947, %6236 : f16
linalg.yield %6237 : f16
} -> tensor<2x1280x8x8xf16>
%2261 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2260, %2259 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%expanded_1544 = tensor.expand_shape %2261 [[0], [1, 2], [3], [4]] : tensor<2x1280x8x8xf16> into tensor<2x80x16x8x8xf16>
%2262 = tensor.empty() : tensor<2x80x16x8x8xf16>
%2263 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1544 : tensor<2x80x16x8x8xf16>) outs(%2262 : tensor<2x80x16x8x8xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x80x16x8x8xf16>
%2264 = tensor.empty() : tensor<2x80x1x8x8xi64>
%2265 = linalg.fill ins(%c0_i64 : i64) outs(%2264 : tensor<2x80x1x8x8xi64>) -> tensor<2x80x1x8x8xi64>
%2266 = tensor.empty() : tensor<2x80x1x8x8xf16>
%2267 = linalg.fill ins(%cst_948 : f16) outs(%2266 : tensor<2x80x1x8x8xf16>) -> tensor<2x80x1x8x8xf16>
%2268:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%2263 : tensor<2x80x16x8x8xf16>) outs(%2267, %2265 : tensor<2x80x1x8x8xf16>, tensor<2x80x1x8x8xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x80x1x8x8xf16>, tensor<2x80x1x8x8xi64>)
%2269 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%2268#0 : tensor<2x80x1x8x8xf16>) outs(%2262 : tensor<2x80x16x8x8xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x80x16x8x8xf16>
%collapsed_1545 = tensor.collapse_shape %2269 [[0], [1, 2], [3], [4]] : tensor<2x80x16x8x8xf16> into tensor<2x1280x8x8xf16>
%2270 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1545 : tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%2271 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2261, %2270 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%2272 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2271, %cst_935 : tensor<2x1280x8x8xf16>, tensor<f16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%2273 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2272 : tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%2274 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2273 : tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1280x8x8xf16>
%2275 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2274, %cst_935 : tensor<2x1280x8x8xf16>, tensor<f16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%2276 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2275, %2270 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%2277 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_617 : tensor<1280x80x1x3x3xf16>) outs(%1590 : tensor<1280x80x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16x3x3xf16>
%collapsed_1546 = tensor.collapse_shape %2277 [[0], [1, 2], [3], [4]] : tensor<1280x80x16x3x3xf16> into tensor<1280x1280x3x3xf16>
%2278 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_616, %collapsed_1546 : tensor<1280x1280x3x3xi8>, tensor<1280x1280x3x3xf16>) outs(%1592 : tensor<1280x1280x3x3xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280x3x3xf16>
%padded_1547 = tensor.pad %2276 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x1280x8x8xf16> to tensor<2x1280x10x10xf16>
%2279 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_615 : tensor<1280xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1280x8x8xf16>
%2280 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1547, %2278 : tensor<2x1280x10x10xf16>, tensor<1280x1280x3x3xf16>) outs(%2279 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16>
%2281 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_614 : tensor<1280x1280xf16>) outs(%25 : tensor<1280x1280xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x1280xf16>
%2282 = linalg.matmul ins(%81, %2281 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%20 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
%2283 = linalg.generic {indexing_maps = [#map9, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_613, %2282 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%19 : tensor<2x1280xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280xf16>
%expanded_1548 = tensor.expand_shape %2283 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16>
%2284 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2280, %expanded_1548 : tensor<2x1280x8x8xf16>, tensor<2x1280x1x1xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%collapsed_1549 = tensor.collapse_shape %2284 [[0], [1], [2, 3]] : tensor<2x1280x8x8xf16> into tensor<2x1280x64xf16>
%expanded_1550 = tensor.expand_shape %collapsed_1549 [[0], [1, 2], [3]] : tensor<2x1280x64xf16> into tensor<2x32x40x64xf16>
%2285 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1550 : tensor<2x32x40x64xf16>) outs(%2239 : tensor<2x32x40x64xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x40x64xf32>
%2286 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2285 : tensor<2x32x40x64xf32>) outs(%2241 : tensor<2x32x40x64xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x64xf64>
%2287 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2286 : tensor<2x32x40x64xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%2288 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2287 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_972 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%2289 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2286, %2288 : tensor<2x32x40x64xf64>, tensor<2x32x1x1xf64>) outs(%2241 : tensor<2x32x40x64xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x64xf64>
%2290 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2289, %2289 : tensor<2x32x40x64xf64>, tensor<2x32x40x64xf64>) outs(%2241 : tensor<2x32x40x64xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x64xf64>
%2291 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2290 : tensor<2x32x40x64xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%2292 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2291 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_972 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%2293 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2292 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%2294 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2285 : tensor<2x32x40x64xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%2295 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2294 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_973 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%2296 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2293, %cst_938 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%2297 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2296 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%2298 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1550, %2295 : tensor<2x32x40x64xf16>, tensor<2x32x1x1xf32>) outs(%2239 : tensor<2x32x40x64xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x40x64xf32>
%2299 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2298, %2297 : tensor<2x32x40x64xf32>, tensor<2x32x1x1xf32>) outs(%2239 : tensor<2x32x40x64xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x40x64xf32>
%collapsed_1551 = tensor.collapse_shape %2299 [[0], [1, 2], [3]] : tensor<2x32x40x64xf32> into tensor<2x1280x64xf32>
%expanded_1552 = tensor.expand_shape %collapsed_1551 [[0], [1], [2, 3]] : tensor<2x1280x64xf32> into tensor<2x1280x8x8xf32>
%expanded_1553 = tensor.expand_shape %cst_612 [[0, 1, 2, 3]] : tensor<1280xf16> into tensor<1x1280x1x1xf16>
%expanded_1554 = tensor.expand_shape %cst_611 [[0, 1, 2, 3]] : tensor<1280xf16> into tensor<1x1280x1x1xf16>
%2300 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1552, %expanded_1554 : tensor<2x1280x8x8xf32>, tensor<1x1280x1x1xf16>) outs(%2256 : tensor<2x1280x8x8xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1280x8x8xf32>
%2301 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2300, %expanded_1553 : tensor<2x1280x8x8xf32>, tensor<1x1280x1x1xf16>) outs(%2256 : tensor<2x1280x8x8xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1280x8x8xf32>
%2302 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2301 : tensor<2x1280x8x8xf32>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%2303 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2302 : tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.negf %in : f16
%6235 = math.exp %6234 : f16
%6236 = arith.addf %6235, %cst_947 : f16
%6237 = arith.divf %cst_947, %6236 : f16
linalg.yield %6237 : f16
} -> tensor<2x1280x8x8xf16>
%2304 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2303, %2302 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%expanded_1555 = tensor.expand_shape %2304 [[0], [1, 2], [3], [4]] : tensor<2x1280x8x8xf16> into tensor<2x80x16x8x8xf16>
%2305 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1555 : tensor<2x80x16x8x8xf16>) outs(%2262 : tensor<2x80x16x8x8xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x80x16x8x8xf16>
%2306:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%2305 : tensor<2x80x16x8x8xf16>) outs(%2267, %2265 : tensor<2x80x1x8x8xf16>, tensor<2x80x1x8x8xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x80x1x8x8xf16>, tensor<2x80x1x8x8xi64>)
%2307 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%2306#0 : tensor<2x80x1x8x8xf16>) outs(%2262 : tensor<2x80x16x8x8xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x80x16x8x8xf16>
%collapsed_1556 = tensor.collapse_shape %2307 [[0], [1, 2], [3], [4]] : tensor<2x80x16x8x8xf16> into tensor<2x1280x8x8xf16>
%2308 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1556 : tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%2309 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2304, %2308 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%2310 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2309, %cst_935 : tensor<2x1280x8x8xf16>, tensor<f16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%2311 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2310 : tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.roundeven %in : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%2312 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2311 : tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.cmpf ult, %in, %cst_957 : f16
%6235 = arith.select %6234, %cst_957, %in : f16
%6236 = arith.cmpf ugt, %6235, %cst_958 : f16
%6237 = arith.select %6236, %cst_958, %6235 : f16
linalg.yield %6237 : f16
} -> tensor<2x1280x8x8xf16>
%2313 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2312, %cst_935 : tensor<2x1280x8x8xf16>, tensor<f16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.subf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%2314 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2313, %2308 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%2315 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%cst_610 : tensor<1280x80x1x3x3xf16>) outs(%1590 : tensor<1280x80x16x3x3xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<1280x80x16x3x3xf16>
%collapsed_1557 = tensor.collapse_shape %2315 [[0], [1, 2], [3], [4]] : tensor<1280x80x16x3x3xf16> into tensor<1280x1280x3x3xf16>
%2316 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_609, %collapsed_1557 : tensor<1280x1280x3x3xi8>, tensor<1280x1280x3x3xf16>) outs(%1592 : tensor<1280x1280x3x3xf16>) {
^bb0(%in: i8, %in_2625: f16, %out: f16):
%6234 = arith.sitofp %in : i8 to f16
%6235 = arith.mulf %6234, %in_2625 : f16
linalg.yield %6235 : f16
} -> tensor<1280x1280x3x3xf16>
%padded_1558 = tensor.pad %2314 low[0, 0, 1, 1] high[0, 0, 1, 1] {
^bb0(%arg4: index, %arg5: index, %arg6: index, %arg7: index):
tensor.yield %cst_946 : f16
} : tensor<2x1280x8x8xf16> to tensor<2x1280x10x10xf16>
%2317 = linalg.generic {indexing_maps = [#map10, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_608 : tensor<1280xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x1280x8x8xf16>
%2318 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1558, %2316 : tensor<2x1280x10x10xf16>, tensor<1280x1280x3x3xf16>) outs(%2317 : tensor<2x1280x8x8xf16>) -> tensor<2x1280x8x8xf16>
%2319 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2238, %2318 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.addf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%2320 = linalg.generic {indexing_maps = [#map1, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2319, %cst_924 : tensor<2x1280x8x8xf16>, tensor<f16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.divf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%collapsed_1559 = tensor.collapse_shape %2320 [[0], [1], [2, 3]] : tensor<2x1280x8x8xf16> into tensor<2x1280x64xf16>
%expanded_1560 = tensor.expand_shape %collapsed_1559 [[0], [1, 2], [3]] : tensor<2x1280x64xf16> into tensor<2x32x40x64xf16>
%2321 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1560 : tensor<2x32x40x64xf16>) outs(%2239 : tensor<2x32x40x64xf32>) {
^bb0(%in: f16, %out: f32):
%6234 = arith.extf %in : f16 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x40x64xf32>
%2322 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2321 : tensor<2x32x40x64xf32>) outs(%2241 : tensor<2x32x40x64xf64>) {
^bb0(%in: f32, %out: f64):
%6234 = arith.extf %in : f32 to f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x64xf64>
%2323 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2322 : tensor<2x32x40x64xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%2324 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2323 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_972 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%2325 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2322, %2324 : tensor<2x32x40x64xf64>, tensor<2x32x1x1xf64>) outs(%2241 : tensor<2x32x40x64xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.subf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x64xf64>
%2326 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2325, %2325 : tensor<2x32x40x64xf64>, tensor<2x32x40x64xf64>) outs(%2241 : tensor<2x32x40x64xf64>) {
^bb0(%in: f64, %in_2625: f64, %out: f64):
%6234 = arith.mulf %in, %in_2625 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x40x64xf64>
%2327 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2326 : tensor<2x32x40x64xf64>) outs(%37 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.addf %in, %out : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%2328 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2327 : tensor<2x32x1x1xf64>) outs(%36 : tensor<2x32x1x1xf64>) {
^bb0(%in: f64, %out: f64):
%6234 = arith.divf %in, %cst_972 : f64
linalg.yield %6234 : f64
} -> tensor<2x32x1x1xf64>
%2329 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2328 : tensor<2x32x1x1xf64>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f64, %out: f32):
%6234 = arith.truncf %in : f64 to f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%2330 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2321 : tensor<2x32x40x64xf32>) outs(%46 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.addf %in, %out : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%2331 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2330 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = arith.divf %in, %cst_973 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%2332 = linalg.generic {indexing_maps = [#map11, #map12, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2329, %cst_938 : tensor<2x32x1x1xf32>, tensor<f16>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x1x1xf32>
%2333 = linalg.generic {indexing_maps = [#map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2332 : tensor<2x32x1x1xf32>) outs(%44 : tensor<2x32x1x1xf32>) {
^bb0(%in: f32, %out: f32):
%6234 = math.rsqrt %in : f32
linalg.yield %6234 : f32
} -> tensor<2x32x1x1xf32>
%2334 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1560, %2331 : tensor<2x32x40x64xf16>, tensor<2x32x1x1xf32>) outs(%2239 : tensor<2x32x40x64xf32>) {
^bb0(%in: f16, %in_2625: f32, %out: f32):
%6234 = arith.extf %in : f16 to f32
%6235 = arith.subf %6234, %in_2625 : f32
linalg.yield %6235 : f32
} -> tensor<2x32x40x64xf32>
%2335 = linalg.generic {indexing_maps = [#map1, #map11, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2334, %2333 : tensor<2x32x40x64xf32>, tensor<2x32x1x1xf32>) outs(%2239 : tensor<2x32x40x64xf32>) {
^bb0(%in: f32, %in_2625: f32, %out: f32):
%6234 = arith.mulf %in, %in_2625 : f32
linalg.yield %6234 : f32
} -> tensor<2x32x40x64xf32>
%collapsed_1561 = tensor.collapse_shape %2335 [[0], [1, 2], [3]] : tensor<2x32x40x64xf32> into tensor<2x1280x64xf32>
%expanded_1562 = tensor.expand_shape %collapsed_1561 [[0], [1], [2, 3]] : tensor<2x1280x64xf32> into tensor<2x1280x8x8xf32>
%expanded_1563 = tensor.expand_shape %cst_607 [[0, 1, 2, 3]] : tensor<1280xf16> into tensor<1x1280x1x1xf16>
%expanded_1564 = tensor.expand_shape %cst_606 [[0, 1, 2, 3]] : tensor<1280xf16> into tensor<1x1280x1x1xf16>
%2336 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1562, %expanded_1564 : tensor<2x1280x8x8xf32>, tensor<1x1280x1x1xf16>) outs(%2256 : tensor<2x1280x8x8xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.mulf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1280x8x8xf32>
%2337 = linalg.generic {indexing_maps = [#map1, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2336, %expanded_1563 : tensor<2x1280x8x8xf32>, tensor<1x1280x1x1xf16>) outs(%2256 : tensor<2x1280x8x8xf32>) {
^bb0(%in: f32, %in_2625: f16, %out: f32):
%6234 = arith.extf %in_2625 : f16 to f32
%6235 = arith.addf %in, %6234 : f32
linalg.yield %6235 : f32
} -> tensor<2x1280x8x8xf32>
%2338 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2337 : tensor<2x1280x8x8xf32>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f32, %out: f16):
%6234 = arith.truncf %in : f32 to f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%2339 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2338 : tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.negf %in : f16
%6235 = math.exp %6234 : f16
%6236 = arith.addf %6235, %cst_947 : f16
%6237 = arith.divf %cst_947, %6236 : f16
linalg.yield %6237 : f16
} -> tensor<2x1280x8x8xf16>
%2340 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2339, %2338 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %in_2625: f16, %out: f16):
%6234 = arith.mulf %in, %in_2625 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%expanded_1565 = tensor.expand_shape %2340 [[0], [1, 2], [3], [4]] : tensor<2x1280x8x8xf16> into tensor<2x80x16x8x8xf16>
%2341 = linalg.generic {indexing_maps = [#map14, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1565 : tensor<2x80x16x8x8xf16>) outs(%2262 : tensor<2x80x16x8x8xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = math.absf %in : f16
linalg.yield %6234 : f16
} -> tensor<2x80x16x8x8xf16>
%2342:2 = linalg.generic {indexing_maps = [#map14, #map15, #map15], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel"]} ins(%2341 : tensor<2x80x16x8x8xf16>) outs(%2267, %2265 : tensor<2x80x1x8x8xf16>, tensor<2x80x1x8x8xi64>) {
^bb0(%in: f16, %out: f16, %out_2625: i64):
%6234 = linalg.index 2 : index
%6235 = arith.index_cast %6234 : index to i64
%6236 = arith.maxf %in, %out : f16
%6237 = arith.cmpf ogt, %in, %out : f16
%6238 = arith.select %6237, %6235, %out_2625 : i64
linalg.yield %6236, %6238 : f16, i64
} -> (tensor<2x80x1x8x8xf16>, tensor<2x80x1x8x8xi64>)
%2343 = linalg.generic {indexing_maps = [#map15, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%2342#0 : tensor<2x80x1x8x8xf16>) outs(%2262 : tensor<2x80x16x8x8xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<2x80x16x8x8xf16>
%collapsed_1566 = tensor.collapse_shape %2343 [[0], [1, 2], [3], [4]] : tensor<2x80x16x8x8xf16> into tensor<2x1280x8x8xf16>
%2344 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%collapsed_1566 : tensor<2x1280x8x8xf16>) outs(%2236 : tensor<2x1280x8x8xf16>) {
^bb0(%in: f16, %out: f16):
%6234 = arith.divf %in, %cst_956 : f16
linalg.yield %6234 : f16
} -> tensor<2x1280x8x8xf16>
%2345 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2340, %2344 : tensor<2x1280x8x8xf16>, tensor<2x1280x8x8xf16>) outs(%2
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment