New Linalg IR - the one with the issue

linalg_elided_new.mlir

#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, d1, d2, d3) -> (d1)>
#map3 = affine_map<(d0, d1, d2, d3) -> (d0, d1, 0, 0)>
#map4 = affine_map<(d0, d1, d2, d3) -> (0, d1, 0, 0)>
#map5 = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
#map6 = affine_map<(d0, d1, d2) -> (d0, d2, d1)>
#map7 = affine_map<(d0, d1) -> (d0, d1)>
#map8 = affine_map<(d0, d1) -> (d1, d0)>
#map9 = affine_map<(d0, d1, d2) -> (0, d1, d2)>
#map10 = affine_map<(d0, d1, d2) -> (d2)>
#map11 = affine_map<(d0, d1, d2, d3) -> (d0, d2, d1, d3)>
#map12 = affine_map<(d0, d1, d2) -> (d0, d1, 0)>
#map13 = affine_map<(d0, d1, d2) -> (0, d1, 0)>
#map14 = affine_map<(d0, d1) -> (d1)>
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>) -> tensor<1x3x512x512xf16> {
    %cst = arith.constant dense<[1.574710e-02, -2.030940e-02, -4.635620e-02]> : tensor<3xf16>
    %cst_0 = arith.constant dense_resource<__elided__> : tensor<3x128x3x3xf16>
    %cst_1 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_2 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_3 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_4 = arith.constant dense_resource<__elided__> : tensor<128x128x3x3xf16>
    %cst_5 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_6 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_7 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_8 = arith.constant dense_resource<__elided__> : tensor<128x128x3x3xf16>
    %cst_9 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_10 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_11 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_12 = arith.constant dense_resource<__elided__> : tensor<128x128x3x3xf16>
    %cst_13 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_14 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_15 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_16 = arith.constant dense_resource<__elided__> : tensor<128x128x3x3xf16>
    %cst_17 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_18 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_19 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_20 = arith.constant dense_resource<__elided__> : tensor<128x256x1x1xf16>
    %cst_21 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_22 = arith.constant dense_resource<__elided__> : tensor<128x128x3x3xf16>
    %cst_23 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_24 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_25 = arith.constant dense_resource<__elided__> : tensor<128xf16>
    %cst_26 = arith.constant dense_resource<__elided__> : tensor<128x256x3x3xf16>
    %cst_27 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_28 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_29 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_30 = arith.constant dense_resource<__elided__> : tensor<256x256x3x3xf16>
    %cst_31 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_32 = arith.constant dense_resource<__elided__> : tensor<256x256x3x3xf16>
    %cst_33 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_34 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_35 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_36 = arith.constant dense_resource<__elided__> : tensor<256x256x3x3xf16>
    %cst_37 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_38 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_39 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_40 = arith.constant dense_resource<__elided__> : tensor<256x256x3x3xf16>
    %cst_41 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_42 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_43 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_44 = arith.constant dense_resource<__elided__> : tensor<256x256x3x3xf16>
    %cst_45 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_46 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_47 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_48 = arith.constant dense_resource<__elided__> : tensor<256x512x1x1xf16>
    %cst_49 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_50 = arith.constant dense_resource<__elided__> : tensor<256x256x3x3xf16>
    %cst_51 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_52 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_53 = arith.constant dense_resource<__elided__> : tensor<256xf16>
    %cst_54 = arith.constant dense_resource<__elided__> : tensor<256x512x3x3xf16>
    %cst_55 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_56 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_57 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_58 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_59 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_60 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_61 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_62 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_63 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_64 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_65 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_66 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_67 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_68 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_69 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_70 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_71 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_72 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_73 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_74 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_75 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_76 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_77 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_78 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_79 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_80 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_81 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_82 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_83 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_84 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_85 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_86 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_87 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_88 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_89 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_90 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_91 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_92 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_93 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_94 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_95 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_96 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_97 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_98 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_99 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_100 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_101 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_102 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_103 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_104 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_105 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_106 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_107 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_108 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_109 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_110 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_111 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_112 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_113 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_114 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_115 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_116 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_117 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_118 = arith.constant dense_resource<__elided__> : tensor<512x512xf16>
    %cst_119 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_120 = arith.constant dense_resource<__elided__> : tensor<512x512xf16>
    %cst_121 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_122 = arith.constant dense_resource<__elided__> : tensor<512x512xf16>
    %cst_123 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_124 = arith.constant dense_resource<__elided__> : tensor<512x512xf16>
    %cst_125 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_126 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_127 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_128 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_129 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_130 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_131 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_132 = arith.constant dense_resource<__elided__> : tensor<512x512x3x3xf16>
    %cst_133 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_134 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_135 = arith.constant dense_resource<__elided__> : tensor<512xf16>
    %cst_136 = arith.constant dense_resource<__elided__> : tensor<512x4x3x3xf16>
    %cst_137 = arith.constant dense<[3.207400e-02, -8.428960e-02, -2.432860e-01, 1.315920e-01]> : tensor<4xf16>
    %cst_138 = arith.constant dense_resource<__elided__> : tensor<4x4x1x1xf16>
    %cst_139 = arith.constant 0.000000e+00 : f16
    %cst_140 = arith.constant 0.000000e+00 : f32
    %cst_141 = arith.constant 1.000000e+00 : f16
    %cst_142 = arith.constant -6.550400e+04 : f16
    %cst_143 = arith.constant 0.000000e+00 : f64
    %cst_144 = arith.constant 5.4899807850672522 : f64
    %cst_145 = arith.constant 9.9999999999999995E-7 : f64
    %cst_146 = arith.constant 0.044194173824159216 : f64
    %cst_147 = arith.constant 6.553600e+04 : f64
    %cst_148 = arith.constant 6.553600e+04 : f32
    %cst_149 = arith.constant 2.621440e+05 : f64
    %cst_150 = arith.constant 2.621440e+05 : f32
    %cst_151 = arith.constant 0x4130000000000000 : f64
    %cst_152 = arith.constant 0x49800000 : f32
    %cst_153 = arith.constant 5.242880e+05 : f64
    %cst_154 = arith.constant 5.242880e+05 : f32
    %cst_155 = arith.constant 0x4140000000000000 : f64
    %cst_156 = arith.constant 0x4A000000 : f32
    %c0_i64 = arith.constant 0 : i64
    %c0 = arith.constant 0 : index
    %c1 = arith.constant 1 : index
    %cst_157 = arith.constant 2.000000e+00 : f16
    %cst_158 = arith.constant 5.000000e-01 : f16
    %cst_159 = arith.constant 2.550000e+02 : f16
    %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):
      %878 = arith.truncf %cst_144 : f64 to f16
      %879 = arith.mulf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x4x64x64xf16>
    %2 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_137 : tensor<4xf16>) outs(%0 : tensor<1x4x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x4x64x64xf16>
    %3 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1, %cst_138 : tensor<1x4x64x64xf16>, tensor<4x4x1x1xf16>) outs(%2 : tensor<1x4x64x64xf16>) -> tensor<1x4x64x64xf16>
    %padded = tensor.pad %3 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x4x64x64xf16> to tensor<1x4x66x66xf16>
    %4 = tensor.empty() : tensor<1x512x64x64xf16>
    %5 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_135 : tensor<512xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %6 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded, %cst_136 : tensor<1x4x66x66xf16>, tensor<512x4x3x3xf16>) outs(%5 : tensor<1x512x64x64xf16>) -> tensor<1x512x64x64xf16>
    %collapsed = tensor.collapse_shape %6 [[0], [1], [2, 3]] : tensor<1x512x64x64xf16> into tensor<1x512x4096xf16>
    %expanded = tensor.expand_shape %collapsed [[0], [1, 2], [3]] : tensor<1x512x4096xf16> into tensor<1x32x16x4096xf16>
    %7 = tensor.empty() : tensor<1x32x16x4096xf32>
    %8 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded : tensor<1x32x16x4096xf16>) outs(%7 : tensor<1x32x16x4096xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x4096xf32>
    %9 = tensor.empty() : tensor<1x32x16x4096xf64>
    %10 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%8 : tensor<1x32x16x4096xf32>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %11 = tensor.empty() : tensor<1x32x1x1xf64>
    %12 = linalg.fill ins(%cst_143 : f64) outs(%11 : tensor<1x32x1x1xf64>) -> tensor<1x32x1x1xf64>
    %13 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%10 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %14 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%13 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %15 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%10, %14 : tensor<1x32x16x4096xf64>, tensor<1x32x1x1xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %16 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%15, %15 : tensor<1x32x16x4096xf64>, tensor<1x32x16x4096xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %17 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%16 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %18 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%17 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %19 = tensor.empty() : tensor<1x32x1x1xf32>
    %20 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%18 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %21 = linalg.fill ins(%cst_140 : f32) outs(%19 : tensor<1x32x1x1xf32>) -> tensor<1x32x1x1xf32>
    %22 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%8 : tensor<1x32x16x4096xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %23 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%22 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_148 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %24 = tensor.empty() : tensor<1x32x1x1xf16>
    %25 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%20 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %26 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%25 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %27 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%23 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %28 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%27 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %29 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%26 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %30 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%29 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %31 = tensor.empty() : tensor<1x32x16x4096xf16>
    %32 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded, %28 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %33 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%32, %30 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %collapsed_160 = tensor.collapse_shape %33 [[0], [1, 2], [3]] : tensor<1x32x16x4096xf16> into tensor<1x512x4096xf16>
    %expanded_161 = tensor.expand_shape %collapsed_160 [[0], [1], [2, 3]] : tensor<1x512x4096xf16> into tensor<1x512x64x64xf16>
    %expanded_162 = tensor.expand_shape %cst_134 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_163 = tensor.expand_shape %cst_133 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %34 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_161, %expanded_163 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %35 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%34, %expanded_162 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %36 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%35 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x512x64x64xf16>
    %37 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%36, %35 : tensor<1x512x64x64xf16>, tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %padded_164 = tensor.pad %37 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x64x64xf16> to tensor<1x512x66x66xf16>
    %38 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_131 : tensor<512xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %39 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_164, %cst_132 : tensor<1x512x66x66xf16>, tensor<512x512x3x3xf16>) outs(%38 : tensor<1x512x64x64xf16>) -> tensor<1x512x64x64xf16>
    %collapsed_165 = tensor.collapse_shape %39 [[0], [1], [2, 3]] : tensor<1x512x64x64xf16> into tensor<1x512x4096xf16>
    %expanded_166 = tensor.expand_shape %collapsed_165 [[0], [1, 2], [3]] : tensor<1x512x4096xf16> into tensor<1x32x16x4096xf16>
    %40 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_166 : tensor<1x32x16x4096xf16>) outs(%7 : tensor<1x32x16x4096xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x4096xf32>
    %41 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%40 : tensor<1x32x16x4096xf32>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %42 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%41 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %43 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%42 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %44 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%41, %43 : tensor<1x32x16x4096xf64>, tensor<1x32x1x1xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %45 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%44, %44 : tensor<1x32x16x4096xf64>, tensor<1x32x16x4096xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %46 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%45 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %47 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%46 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %48 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%47 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %49 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%40 : tensor<1x32x16x4096xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %50 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%49 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_148 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %51 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%48 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %52 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%51 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %53 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%50 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %54 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%53 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %55 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%52 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %56 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%55 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %57 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_166, %54 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %58 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%57, %56 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %collapsed_167 = tensor.collapse_shape %58 [[0], [1, 2], [3]] : tensor<1x32x16x4096xf16> into tensor<1x512x4096xf16>
    %expanded_168 = tensor.expand_shape %collapsed_167 [[0], [1], [2, 3]] : tensor<1x512x4096xf16> into tensor<1x512x64x64xf16>
    %expanded_169 = tensor.expand_shape %cst_130 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_170 = tensor.expand_shape %cst_129 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %59 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_168, %expanded_170 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %60 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%59, %expanded_169 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %61 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%60 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x512x64x64xf16>
    %62 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%61, %60 : tensor<1x512x64x64xf16>, tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %padded_171 = tensor.pad %62 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x64x64xf16> to tensor<1x512x66x66xf16>
    %63 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_127 : tensor<512xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %64 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_171, %cst_128 : tensor<1x512x66x66xf16>, tensor<512x512x3x3xf16>) outs(%63 : tensor<1x512x64x64xf16>) -> tensor<1x512x64x64xf16>
    %65 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%6, %64 : tensor<1x512x64x64xf16>, tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %66 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%65 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %collapsed_172 = tensor.collapse_shape %66 [[0], [1], [2, 3]] : tensor<1x512x64x64xf16> into tensor<1x512x4096xf16>
    %expanded_173 = tensor.expand_shape %collapsed_172 [[0], [1, 2], [3]] : tensor<1x512x4096xf16> into tensor<1x32x16x4096xf16>
    %67 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_173 : tensor<1x32x16x4096xf16>) outs(%7 : tensor<1x32x16x4096xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x4096xf32>
    %68 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%67 : tensor<1x32x16x4096xf32>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %69 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%68 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %70 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%69 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %71 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%68, %70 : tensor<1x32x16x4096xf64>, tensor<1x32x1x1xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %72 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%71, %71 : tensor<1x32x16x4096xf64>, tensor<1x32x16x4096xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %73 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%72 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %74 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%73 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %75 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%74 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %76 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%67 : tensor<1x32x16x4096xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %77 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%76 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_148 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %78 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%75 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %79 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%78 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %80 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%77 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %81 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%80 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %82 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%79 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %83 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%82 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %84 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_173, %81 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %85 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%84, %83 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %collapsed_174 = tensor.collapse_shape %85 [[0], [1, 2], [3]] : tensor<1x32x16x4096xf16> into tensor<1x512x4096xf16>
    %expanded_175 = tensor.expand_shape %collapsed_174 [[0], [1], [2, 3]] : tensor<1x512x4096xf16> into tensor<1x512x64x64xf16>
    %expanded_176 = tensor.expand_shape %cst_126 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_177 = tensor.expand_shape %cst_125 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %86 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_175, %expanded_177 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %87 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%86, %expanded_176 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %collapsed_178 = tensor.collapse_shape %87 [[0], [1], [2, 3]] : tensor<1x512x64x64xf16> into tensor<1x512x4096xf16>
    %88 = tensor.empty() : tensor<1x4096x512xf16>
    %89 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_178 : tensor<1x512x4096xf16>) outs(%88 : tensor<1x4096x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x4096x512xf16>
    %90 = tensor.empty() : tensor<512x512xf16>
    %91 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_124 : tensor<512x512xf16>) outs(%90 : tensor<512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<512x512xf16>
    %collapsed_179 = tensor.collapse_shape %89 [[0, 1], [2]] : tensor<1x4096x512xf16> into tensor<4096x512xf16>
    %92 = tensor.empty() : tensor<4096x512xf16>
    %93 = linalg.fill ins(%cst_139 : f16) outs(%92 : tensor<4096x512xf16>) -> tensor<4096x512xf16>
    %94 = linalg.matmul ins(%collapsed_179, %91 : tensor<4096x512xf16>, tensor<512x512xf16>) outs(%93 : tensor<4096x512xf16>) -> tensor<4096x512xf16>
    %expanded_180 = tensor.expand_shape %94 [[0, 1], [2]] : tensor<4096x512xf16> into tensor<1x4096x512xf16>
    %95 = linalg.generic {indexing_maps = [#map9, #map10, #map5], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_180, %cst_123 : tensor<1x4096x512xf16>, tensor<512xf16>) outs(%88 : tensor<1x4096x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x4096x512xf16>
    %96 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_122 : tensor<512x512xf16>) outs(%90 : tensor<512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<512x512xf16>
    %97 = linalg.matmul ins(%collapsed_179, %96 : tensor<4096x512xf16>, tensor<512x512xf16>) outs(%93 : tensor<4096x512xf16>) -> tensor<4096x512xf16>
    %expanded_181 = tensor.expand_shape %97 [[0, 1], [2]] : tensor<4096x512xf16> into tensor<1x4096x512xf16>
    %98 = linalg.generic {indexing_maps = [#map9, #map10, #map5], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_181, %cst_121 : tensor<1x4096x512xf16>, tensor<512xf16>) outs(%88 : tensor<1x4096x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x4096x512xf16>
    %99 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_120 : tensor<512x512xf16>) outs(%90 : tensor<512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<512x512xf16>
    %100 = linalg.matmul ins(%collapsed_179, %99 : tensor<4096x512xf16>, tensor<512x512xf16>) outs(%93 : tensor<4096x512xf16>) -> tensor<4096x512xf16>
    %expanded_182 = tensor.expand_shape %100 [[0, 1], [2]] : tensor<4096x512xf16> into tensor<1x4096x512xf16>
    %101 = linalg.generic {indexing_maps = [#map9, #map10, #map5], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_182, %cst_119 : tensor<1x4096x512xf16>, tensor<512xf16>) outs(%88 : tensor<1x4096x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x4096x512xf16>
    %expanded_183 = tensor.expand_shape %95 [[0], [1], [2, 3]] : tensor<1x4096x512xf16> into tensor<1x4096x1x512xf16>
    %102 = tensor.empty() : tensor<1x1x4096x512xf16>
    %103 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_183 : tensor<1x4096x1x512xf16>) outs(%102 : tensor<1x1x4096x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x1x4096x512xf16>
    %collapsed_184 = tensor.collapse_shape %103 [[0], [1, 2], [3]] : tensor<1x1x4096x512xf16> into tensor<1x4096x512xf16>
    %expanded_185 = tensor.expand_shape %98 [[0], [1], [2, 3]] : tensor<1x4096x512xf16> into tensor<1x4096x1x512xf16>
    %104 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_185 : tensor<1x4096x1x512xf16>) outs(%102 : tensor<1x1x4096x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x1x4096x512xf16>
    %collapsed_186 = tensor.collapse_shape %104 [[0], [1, 2], [3]] : tensor<1x1x4096x512xf16> into tensor<1x4096x512xf16>
    %expanded_187 = tensor.expand_shape %101 [[0], [1], [2, 3]] : tensor<1x4096x512xf16> into tensor<1x4096x1x512xf16>
    %105 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_187 : tensor<1x4096x1x512xf16>) outs(%102 : tensor<1x1x4096x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x1x4096x512xf16>
    %collapsed_188 = tensor.collapse_shape %105 [[0], [1, 2], [3]] : tensor<1x1x4096x512xf16> into tensor<1x4096x512xf16>
    %106 = tensor.empty() : tensor<1x4096x4096xf16>
    %107 = linalg.generic {indexing_maps = [#map9, #map5], iterator_types = ["parallel", "parallel", "parallel"]} ins(%106 : tensor<1x4096x4096xf16>) outs(%106 : tensor<1x4096x4096xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %cst_139 : f16
    } -> tensor<1x4096x4096xf16>
    %108 = tensor.empty() : tensor<1x512x4096xf16>
    %109 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_186 : tensor<1x4096x512xf16>) outs(%108 : tensor<1x512x4096xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x4096xf16>
    %110 = linalg.fill ins(%cst_139 : f16) outs(%106 : tensor<1x4096x4096xf16>) -> tensor<1x4096x4096xf16>
    %111 = linalg.batch_matmul ins(%collapsed_184, %109 : tensor<1x4096x512xf16>, tensor<1x512x4096xf16>) outs(%110 : tensor<1x4096x4096xf16>) -> tensor<1x4096x4096xf16>
    %112 = linalg.generic {indexing_maps = [#map9, #map5], iterator_types = ["parallel", "parallel", "parallel"]} ins(%111 : tensor<1x4096x4096xf16>) outs(%106 : tensor<1x4096x4096xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_146 : f64 to f16
      %879 = arith.mulf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x4096x4096xf16>
    %113 = linalg.generic {indexing_maps = [#map9, #map9, #map5], iterator_types = ["parallel", "parallel", "parallel"]} ins(%112, %107 : tensor<1x4096x4096xf16>, tensor<1x4096x4096xf16>) outs(%106 : tensor<1x4096x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in_385, %cst_139 : f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x4096x4096xf16>
    %114 = tensor.empty() : tensor<1x4096x1xi64>
    %115 = linalg.fill ins(%c0_i64 : i64) outs(%114 : tensor<1x4096x1xi64>) -> tensor<1x4096x1xi64>
    %116 = tensor.empty() : tensor<1x4096x1xf16>
    %117 = linalg.fill ins(%cst_142 : f16) outs(%116 : tensor<1x4096x1xf16>) -> tensor<1x4096x1xf16>
    %118:2 = linalg.generic {indexing_maps = [#map5, #map12, #map12], iterator_types = ["parallel", "parallel", "reduction"]} ins(%113 : tensor<1x4096x4096xf16>) outs(%117, %115 : tensor<1x4096x1xf16>, tensor<1x4096x1xi64>) {
    ^bb0(%in: f16, %out: f16, %out_385: i64):
      %878 = linalg.index 2 : index
      %879 = arith.index_cast %878 : index to i64
      %880 = arith.maxf %in, %out : f16
      %881 = arith.cmpf ogt, %in, %out : f16
      %882 = arith.select %881, %879, %out_385 : i64
      linalg.yield %880, %882 : f16, i64
    } -> (tensor<1x4096x1xf16>, tensor<1x4096x1xi64>)
    %119 = linalg.generic {indexing_maps = [#map9, #map13, #map5], iterator_types = ["parallel", "parallel", "parallel"]} ins(%113, %118#0 : tensor<1x4096x4096xf16>, tensor<1x4096x1xf16>) outs(%106 : tensor<1x4096x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x4096x4096xf16>
    %120 = linalg.generic {indexing_maps = [#map9, #map5], iterator_types = ["parallel", "parallel", "parallel"]} ins(%119 : tensor<1x4096x4096xf16>) outs(%106 : tensor<1x4096x4096xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.exp %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x4096x4096xf16>
    %121 = linalg.fill ins(%cst_139 : f16) outs(%116 : tensor<1x4096x1xf16>) -> tensor<1x4096x1xf16>
    %122 = linalg.generic {indexing_maps = [#map5, #map12], iterator_types = ["parallel", "parallel", "reduction"]} ins(%120 : tensor<1x4096x4096xf16>) outs(%121 : tensor<1x4096x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.addf %in, %out : f16
      linalg.yield %878 : f16
    } -> tensor<1x4096x1xf16>
    %123 = linalg.generic {indexing_maps = [#map9, #map13, #map5], iterator_types = ["parallel", "parallel", "parallel"]} ins(%120, %122 : tensor<1x4096x4096xf16>, tensor<1x4096x1xf16>) outs(%106 : tensor<1x4096x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.divf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x4096x4096xf16>
    %124 = linalg.fill ins(%cst_139 : f16) outs(%88 : tensor<1x4096x512xf16>) -> tensor<1x4096x512xf16>
    %125 = linalg.batch_matmul ins(%123, %collapsed_188 : tensor<1x4096x4096xf16>, tensor<1x4096x512xf16>) outs(%124 : tensor<1x4096x512xf16>) -> tensor<1x4096x512xf16>
    %expanded_189 = tensor.expand_shape %125 [[0], [1, 2], [3]] : tensor<1x4096x512xf16> into tensor<1x1x4096x512xf16>
    %126 = tensor.empty() : tensor<1x4096x1x512xf16>
    %127 = linalg.generic {indexing_maps = [#map1, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_189 : tensor<1x1x4096x512xf16>) outs(%126 : tensor<1x4096x1x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x4096x1x512xf16>
    %128 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel"]} ins(%cst_118 : tensor<512x512xf16>) outs(%90 : tensor<512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<512x512xf16>
    %collapsed_190 = tensor.collapse_shape %127 [[0, 1], [2, 3]] : tensor<1x4096x1x512xf16> into tensor<4096x512xf16>
    %129 = linalg.matmul ins(%collapsed_190, %128 : tensor<4096x512xf16>, tensor<512x512xf16>) outs(%93 : tensor<4096x512xf16>) -> tensor<4096x512xf16>
    %130 = linalg.generic {indexing_maps = [#map14, #map7, #map7], iterator_types = ["parallel", "parallel"]} ins(%cst_117, %129 : tensor<512xf16>, tensor<4096x512xf16>) outs(%92 : tensor<4096x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<4096x512xf16>
    %expanded_191 = tensor.expand_shape %130 [[0, 1], [2]] : tensor<4096x512xf16> into tensor<1x4096x512xf16>
    %131 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_191 : tensor<1x4096x512xf16>) outs(%108 : tensor<1x512x4096xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x4096xf16>
    %expanded_192 = tensor.expand_shape %131 [[0], [1], [2, 3]] : tensor<1x512x4096xf16> into tensor<1x512x64x64xf16>
    %132 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_192, %66 : tensor<1x512x64x64xf16>, tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %133 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%132 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %134 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%133 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %collapsed_193 = tensor.collapse_shape %134 [[0], [1], [2, 3]] : tensor<1x512x64x64xf16> into tensor<1x512x4096xf16>
    %expanded_194 = tensor.expand_shape %collapsed_193 [[0], [1, 2], [3]] : tensor<1x512x4096xf16> into tensor<1x32x16x4096xf16>
    %135 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_194 : tensor<1x32x16x4096xf16>) outs(%7 : tensor<1x32x16x4096xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x4096xf32>
    %136 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%135 : tensor<1x32x16x4096xf32>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %137 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%136 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %138 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%137 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %139 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%136, %138 : tensor<1x32x16x4096xf64>, tensor<1x32x1x1xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %140 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%139, %139 : tensor<1x32x16x4096xf64>, tensor<1x32x16x4096xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %141 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%140 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %142 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%141 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %143 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%142 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %144 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%135 : tensor<1x32x16x4096xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %145 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%144 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_148 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %146 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%143 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %147 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%146 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %148 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%145 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %149 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%148 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %150 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%147 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %151 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%150 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %152 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_194, %149 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %153 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%152, %151 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %collapsed_195 = tensor.collapse_shape %153 [[0], [1, 2], [3]] : tensor<1x32x16x4096xf16> into tensor<1x512x4096xf16>
    %expanded_196 = tensor.expand_shape %collapsed_195 [[0], [1], [2, 3]] : tensor<1x512x4096xf16> into tensor<1x512x64x64xf16>
    %expanded_197 = tensor.expand_shape %cst_116 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_198 = tensor.expand_shape %cst_115 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %154 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_196, %expanded_198 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %155 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%154, %expanded_197 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %156 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%155 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x512x64x64xf16>
    %157 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%156, %155 : tensor<1x512x64x64xf16>, tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %padded_199 = tensor.pad %157 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x64x64xf16> to tensor<1x512x66x66xf16>
    %158 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_113 : tensor<512xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %159 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_199, %cst_114 : tensor<1x512x66x66xf16>, tensor<512x512x3x3xf16>) outs(%158 : tensor<1x512x64x64xf16>) -> tensor<1x512x64x64xf16>
    %collapsed_200 = tensor.collapse_shape %159 [[0], [1], [2, 3]] : tensor<1x512x64x64xf16> into tensor<1x512x4096xf16>
    %expanded_201 = tensor.expand_shape %collapsed_200 [[0], [1, 2], [3]] : tensor<1x512x4096xf16> into tensor<1x32x16x4096xf16>
    %160 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_201 : tensor<1x32x16x4096xf16>) outs(%7 : tensor<1x32x16x4096xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x4096xf32>
    %161 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%160 : tensor<1x32x16x4096xf32>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %162 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%161 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %163 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%162 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %164 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%161, %163 : tensor<1x32x16x4096xf64>, tensor<1x32x1x1xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %165 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%164, %164 : tensor<1x32x16x4096xf64>, tensor<1x32x16x4096xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %166 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%165 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %167 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%166 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %168 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%167 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %169 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%160 : tensor<1x32x16x4096xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %170 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%169 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_148 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %171 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%168 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %172 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%171 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %173 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%170 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %174 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%173 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %175 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%172 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %176 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%175 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %177 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_201, %174 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %178 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%177, %176 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %collapsed_202 = tensor.collapse_shape %178 [[0], [1, 2], [3]] : tensor<1x32x16x4096xf16> into tensor<1x512x4096xf16>
    %expanded_203 = tensor.expand_shape %collapsed_202 [[0], [1], [2, 3]] : tensor<1x512x4096xf16> into tensor<1x512x64x64xf16>
    %expanded_204 = tensor.expand_shape %cst_112 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_205 = tensor.expand_shape %cst_111 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %179 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_203, %expanded_205 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %180 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%179, %expanded_204 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %181 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%180 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x512x64x64xf16>
    %182 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%181, %180 : tensor<1x512x64x64xf16>, tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %padded_206 = tensor.pad %182 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x64x64xf16> to tensor<1x512x66x66xf16>
    %183 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_109 : tensor<512xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %184 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_206, %cst_110 : tensor<1x512x66x66xf16>, tensor<512x512x3x3xf16>) outs(%183 : tensor<1x512x64x64xf16>) -> tensor<1x512x64x64xf16>
    %185 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%133, %184 : tensor<1x512x64x64xf16>, tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %186 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%185 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %187 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%186 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %collapsed_207 = tensor.collapse_shape %187 [[0], [1], [2, 3]] : tensor<1x512x64x64xf16> into tensor<1x512x4096xf16>
    %expanded_208 = tensor.expand_shape %collapsed_207 [[0], [1, 2], [3]] : tensor<1x512x4096xf16> into tensor<1x32x16x4096xf16>
    %188 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_208 : tensor<1x32x16x4096xf16>) outs(%7 : tensor<1x32x16x4096xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x4096xf32>
    %189 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%188 : tensor<1x32x16x4096xf32>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %190 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%189 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %191 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%190 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %192 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%189, %191 : tensor<1x32x16x4096xf64>, tensor<1x32x1x1xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %193 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%192, %192 : tensor<1x32x16x4096xf64>, tensor<1x32x16x4096xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %194 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%193 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %195 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%194 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %196 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%195 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %197 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%188 : tensor<1x32x16x4096xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %198 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%197 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_148 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %199 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%196 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %200 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%199 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %201 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%198 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %202 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%201 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %203 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%200 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %204 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%203 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %205 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_208, %202 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %206 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%205, %204 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %collapsed_209 = tensor.collapse_shape %206 [[0], [1, 2], [3]] : tensor<1x32x16x4096xf16> into tensor<1x512x4096xf16>
    %expanded_210 = tensor.expand_shape %collapsed_209 [[0], [1], [2, 3]] : tensor<1x512x4096xf16> into tensor<1x512x64x64xf16>
    %expanded_211 = tensor.expand_shape %cst_108 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_212 = tensor.expand_shape %cst_107 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %207 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_210, %expanded_212 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %208 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%207, %expanded_211 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %209 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%208 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x512x64x64xf16>
    %210 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%209, %208 : tensor<1x512x64x64xf16>, tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %padded_213 = tensor.pad %210 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x64x64xf16> to tensor<1x512x66x66xf16>
    %211 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_105 : tensor<512xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %212 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_213, %cst_106 : tensor<1x512x66x66xf16>, tensor<512x512x3x3xf16>) outs(%211 : tensor<1x512x64x64xf16>) -> tensor<1x512x64x64xf16>
    %collapsed_214 = tensor.collapse_shape %212 [[0], [1], [2, 3]] : tensor<1x512x64x64xf16> into tensor<1x512x4096xf16>
    %expanded_215 = tensor.expand_shape %collapsed_214 [[0], [1, 2], [3]] : tensor<1x512x4096xf16> into tensor<1x32x16x4096xf16>
    %213 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_215 : tensor<1x32x16x4096xf16>) outs(%7 : tensor<1x32x16x4096xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x4096xf32>
    %214 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%213 : tensor<1x32x16x4096xf32>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %215 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%214 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %216 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%215 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %217 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%214, %216 : tensor<1x32x16x4096xf64>, tensor<1x32x1x1xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %218 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%217, %217 : tensor<1x32x16x4096xf64>, tensor<1x32x16x4096xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %219 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%218 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %220 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%219 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %221 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%220 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %222 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%213 : tensor<1x32x16x4096xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %223 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%222 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_148 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %224 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%221 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %225 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%224 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %226 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%223 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %227 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%226 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %228 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%225 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %229 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%228 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %230 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_215, %227 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %231 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%230, %229 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %collapsed_216 = tensor.collapse_shape %231 [[0], [1, 2], [3]] : tensor<1x32x16x4096xf16> into tensor<1x512x4096xf16>
    %expanded_217 = tensor.expand_shape %collapsed_216 [[0], [1], [2, 3]] : tensor<1x512x4096xf16> into tensor<1x512x64x64xf16>
    %expanded_218 = tensor.expand_shape %cst_104 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_219 = tensor.expand_shape %cst_103 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %232 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_217, %expanded_219 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %233 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%232, %expanded_218 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %234 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%233 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x512x64x64xf16>
    %235 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%234, %233 : tensor<1x512x64x64xf16>, tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %padded_220 = tensor.pad %235 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x64x64xf16> to tensor<1x512x66x66xf16>
    %236 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_101 : tensor<512xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %237 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_220, %cst_102 : tensor<1x512x66x66xf16>, tensor<512x512x3x3xf16>) outs(%236 : tensor<1x512x64x64xf16>) -> tensor<1x512x64x64xf16>
    %238 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%186, %237 : tensor<1x512x64x64xf16>, tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %239 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%238 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %240 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%239 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %collapsed_221 = tensor.collapse_shape %240 [[0], [1], [2, 3]] : tensor<1x512x64x64xf16> into tensor<1x512x4096xf16>
    %expanded_222 = tensor.expand_shape %collapsed_221 [[0], [1, 2], [3]] : tensor<1x512x4096xf16> into tensor<1x32x16x4096xf16>
    %241 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_222 : tensor<1x32x16x4096xf16>) outs(%7 : tensor<1x32x16x4096xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x4096xf32>
    %242 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%241 : tensor<1x32x16x4096xf32>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %243 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%242 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %244 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%243 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %245 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%242, %244 : tensor<1x32x16x4096xf64>, tensor<1x32x1x1xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %246 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%245, %245 : tensor<1x32x16x4096xf64>, tensor<1x32x16x4096xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %247 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%246 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %248 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%247 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %249 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%248 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %250 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%241 : tensor<1x32x16x4096xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %251 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%250 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_148 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %252 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%249 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %253 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%252 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %254 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%251 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %255 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%254 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %256 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%253 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %257 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%256 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %258 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_222, %255 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %259 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%258, %257 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %collapsed_223 = tensor.collapse_shape %259 [[0], [1, 2], [3]] : tensor<1x32x16x4096xf16> into tensor<1x512x4096xf16>
    %expanded_224 = tensor.expand_shape %collapsed_223 [[0], [1], [2, 3]] : tensor<1x512x4096xf16> into tensor<1x512x64x64xf16>
    %expanded_225 = tensor.expand_shape %cst_100 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_226 = tensor.expand_shape %cst_99 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %260 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_224, %expanded_226 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %261 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%260, %expanded_225 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %262 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%261 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x512x64x64xf16>
    %263 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%262, %261 : tensor<1x512x64x64xf16>, tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %padded_227 = tensor.pad %263 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x64x64xf16> to tensor<1x512x66x66xf16>
    %264 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_97 : tensor<512xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %265 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_227, %cst_98 : tensor<1x512x66x66xf16>, tensor<512x512x3x3xf16>) outs(%264 : tensor<1x512x64x64xf16>) -> tensor<1x512x64x64xf16>
    %collapsed_228 = tensor.collapse_shape %265 [[0], [1], [2, 3]] : tensor<1x512x64x64xf16> into tensor<1x512x4096xf16>
    %expanded_229 = tensor.expand_shape %collapsed_228 [[0], [1, 2], [3]] : tensor<1x512x4096xf16> into tensor<1x32x16x4096xf16>
    %266 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_229 : tensor<1x32x16x4096xf16>) outs(%7 : tensor<1x32x16x4096xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x4096xf32>
    %267 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%266 : tensor<1x32x16x4096xf32>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %268 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%267 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %269 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%268 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %270 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%267, %269 : tensor<1x32x16x4096xf64>, tensor<1x32x1x1xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %271 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%270, %270 : tensor<1x32x16x4096xf64>, tensor<1x32x16x4096xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %272 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%271 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %273 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%272 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %274 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%273 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %275 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%266 : tensor<1x32x16x4096xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %276 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%275 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_148 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %277 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%274 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %278 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%277 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %279 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%276 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %280 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%279 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %281 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%278 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %282 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%281 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %283 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_229, %280 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %284 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%283, %282 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %collapsed_230 = tensor.collapse_shape %284 [[0], [1, 2], [3]] : tensor<1x32x16x4096xf16> into tensor<1x512x4096xf16>
    %expanded_231 = tensor.expand_shape %collapsed_230 [[0], [1], [2, 3]] : tensor<1x512x4096xf16> into tensor<1x512x64x64xf16>
    %expanded_232 = tensor.expand_shape %cst_96 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_233 = tensor.expand_shape %cst_95 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %285 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_231, %expanded_233 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %286 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%285, %expanded_232 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %287 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%286 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x512x64x64xf16>
    %288 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%287, %286 : tensor<1x512x64x64xf16>, tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %padded_234 = tensor.pad %288 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x64x64xf16> to tensor<1x512x66x66xf16>
    %289 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_93 : tensor<512xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %290 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_234, %cst_94 : tensor<1x512x66x66xf16>, tensor<512x512x3x3xf16>) outs(%289 : tensor<1x512x64x64xf16>) -> tensor<1x512x64x64xf16>
    %291 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%239, %290 : tensor<1x512x64x64xf16>, tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %292 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%291 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %293 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%292 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %collapsed_235 = tensor.collapse_shape %293 [[0], [1], [2, 3]] : tensor<1x512x64x64xf16> into tensor<1x512x4096xf16>
    %expanded_236 = tensor.expand_shape %collapsed_235 [[0], [1, 2], [3]] : tensor<1x512x4096xf16> into tensor<1x32x16x4096xf16>
    %294 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_236 : tensor<1x32x16x4096xf16>) outs(%7 : tensor<1x32x16x4096xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x4096xf32>
    %295 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%294 : tensor<1x32x16x4096xf32>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %296 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%295 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %297 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%296 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %298 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%295, %297 : tensor<1x32x16x4096xf64>, tensor<1x32x1x1xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %299 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%298, %298 : tensor<1x32x16x4096xf64>, tensor<1x32x16x4096xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %300 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%299 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %301 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%300 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %302 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%301 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %303 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%294 : tensor<1x32x16x4096xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %304 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%303 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_148 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %305 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%302 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %306 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%305 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %307 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%304 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %308 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%307 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %309 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%306 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %310 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%309 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %311 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_236, %308 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %312 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%311, %310 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %collapsed_237 = tensor.collapse_shape %312 [[0], [1, 2], [3]] : tensor<1x32x16x4096xf16> into tensor<1x512x4096xf16>
    %expanded_238 = tensor.expand_shape %collapsed_237 [[0], [1], [2, 3]] : tensor<1x512x4096xf16> into tensor<1x512x64x64xf16>
    %expanded_239 = tensor.expand_shape %cst_92 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_240 = tensor.expand_shape %cst_91 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %313 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_238, %expanded_240 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %314 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%313, %expanded_239 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %315 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%314 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x512x64x64xf16>
    %316 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%315, %314 : tensor<1x512x64x64xf16>, tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %padded_241 = tensor.pad %316 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x64x64xf16> to tensor<1x512x66x66xf16>
    %317 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_89 : tensor<512xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %318 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_241, %cst_90 : tensor<1x512x66x66xf16>, tensor<512x512x3x3xf16>) outs(%317 : tensor<1x512x64x64xf16>) -> tensor<1x512x64x64xf16>
    %collapsed_242 = tensor.collapse_shape %318 [[0], [1], [2, 3]] : tensor<1x512x64x64xf16> into tensor<1x512x4096xf16>
    %expanded_243 = tensor.expand_shape %collapsed_242 [[0], [1, 2], [3]] : tensor<1x512x4096xf16> into tensor<1x32x16x4096xf16>
    %319 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_243 : tensor<1x32x16x4096xf16>) outs(%7 : tensor<1x32x16x4096xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x4096xf32>
    %320 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%319 : tensor<1x32x16x4096xf32>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %321 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%320 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %322 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%321 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %323 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%320, %322 : tensor<1x32x16x4096xf64>, tensor<1x32x1x1xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %324 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%323, %323 : tensor<1x32x16x4096xf64>, tensor<1x32x16x4096xf64>) outs(%9 : tensor<1x32x16x4096xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x4096xf64>
    %325 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%324 : tensor<1x32x16x4096xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %326 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%325 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_147 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %327 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%326 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %328 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%319 : tensor<1x32x16x4096xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %329 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%328 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_148 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %330 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%327 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %331 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%330 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %332 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%329 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %333 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%332 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %334 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%331 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %335 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%334 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %336 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_243, %333 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %337 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%336, %335 : tensor<1x32x16x4096xf16>, tensor<1x32x1x1xf16>) outs(%31 : tensor<1x32x16x4096xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x4096xf16>
    %collapsed_244 = tensor.collapse_shape %337 [[0], [1, 2], [3]] : tensor<1x32x16x4096xf16> into tensor<1x512x4096xf16>
    %expanded_245 = tensor.expand_shape %collapsed_244 [[0], [1], [2, 3]] : tensor<1x512x4096xf16> into tensor<1x512x64x64xf16>
    %expanded_246 = tensor.expand_shape %cst_88 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_247 = tensor.expand_shape %cst_87 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %338 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_245, %expanded_247 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %339 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%338, %expanded_246 : tensor<1x512x64x64xf16>, tensor<1x512x1x1xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %340 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%339 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x512x64x64xf16>
    %341 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%340, %339 : tensor<1x512x64x64xf16>, tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %padded_248 = tensor.pad %341 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x64x64xf16> to tensor<1x512x66x66xf16>
    %342 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_85 : tensor<512xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %343 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_248, %cst_86 : tensor<1x512x66x66xf16>, tensor<512x512x3x3xf16>) outs(%342 : tensor<1x512x64x64xf16>) -> tensor<1x512x64x64xf16>
    %344 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%292, %343 : tensor<1x512x64x64xf16>, tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x64x64xf16>
    %345 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%344 : tensor<1x512x64x64xf16>) outs(%4 : tensor<1x512x64x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x64x64xf16>
    %346 = tensor.empty() : tensor<1x512x128x128xf16>
    %347 = linalg.generic {indexing_maps = [#map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%out: f16):
      %878 = linalg.index 0 : index
      %879 = linalg.index 1 : index
      %880 = linalg.index 2 : index
      %881 = linalg.index 3 : index
      %882 = arith.floordivsi %880, %c2 : index
      %883 = arith.floordivsi %881, %c2 : index
      %extracted = tensor.extract %345[%878, %879, %882, %883] : tensor<1x512x64x64xf16>
      linalg.yield %extracted : f16
    } -> tensor<1x512x128x128xf16>
    %padded_249 = tensor.pad %347 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x128x128xf16> to tensor<1x512x130x130xf16>
    %348 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_83 : tensor<512xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x128x128xf16>
    %349 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_249, %cst_84 : tensor<1x512x130x130xf16>, tensor<512x512x3x3xf16>) outs(%348 : tensor<1x512x128x128xf16>) -> tensor<1x512x128x128xf16>
    %collapsed_250 = tensor.collapse_shape %349 [[0], [1], [2, 3]] : tensor<1x512x128x128xf16> into tensor<1x512x16384xf16>
    %expanded_251 = tensor.expand_shape %collapsed_250 [[0], [1, 2], [3]] : tensor<1x512x16384xf16> into tensor<1x32x16x16384xf16>
    %350 = tensor.empty() : tensor<1x32x16x16384xf32>
    %351 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_251 : tensor<1x32x16x16384xf16>) outs(%350 : tensor<1x32x16x16384xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x16384xf32>
    %352 = tensor.empty() : tensor<1x32x16x16384xf64>
    %353 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%351 : tensor<1x32x16x16384xf32>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %354 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%353 : tensor<1x32x16x16384xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %355 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%354 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_149 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %356 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%353, %355 : tensor<1x32x16x16384xf64>, tensor<1x32x1x1xf64>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %357 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%356, %356 : tensor<1x32x16x16384xf64>, tensor<1x32x16x16384xf64>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %358 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%357 : tensor<1x32x16x16384xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %359 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%358 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_149 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %360 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%359 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %361 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%351 : tensor<1x32x16x16384xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %362 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%361 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_150 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %363 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%360 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %364 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%363 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %365 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%362 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %366 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%365 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %367 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%364 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %368 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%367 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %369 = tensor.empty() : tensor<1x32x16x16384xf16>
    %370 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_251, %366 : tensor<1x32x16x16384xf16>, tensor<1x32x1x1xf16>) outs(%369 : tensor<1x32x16x16384xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x16384xf16>
    %371 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%370, %368 : tensor<1x32x16x16384xf16>, tensor<1x32x1x1xf16>) outs(%369 : tensor<1x32x16x16384xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x16384xf16>
    %collapsed_252 = tensor.collapse_shape %371 [[0], [1, 2], [3]] : tensor<1x32x16x16384xf16> into tensor<1x512x16384xf16>
    %expanded_253 = tensor.expand_shape %collapsed_252 [[0], [1], [2, 3]] : tensor<1x512x16384xf16> into tensor<1x512x128x128xf16>
    %expanded_254 = tensor.expand_shape %cst_82 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_255 = tensor.expand_shape %cst_81 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %372 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_253, %expanded_255 : tensor<1x512x128x128xf16>, tensor<1x512x1x1xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %373 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%372, %expanded_254 : tensor<1x512x128x128xf16>, tensor<1x512x1x1xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %374 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%373 : tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x512x128x128xf16>
    %375 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%374, %373 : tensor<1x512x128x128xf16>, tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %padded_256 = tensor.pad %375 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x128x128xf16> to tensor<1x512x130x130xf16>
    %376 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_79 : tensor<512xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x128x128xf16>
    %377 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_256, %cst_80 : tensor<1x512x130x130xf16>, tensor<512x512x3x3xf16>) outs(%376 : tensor<1x512x128x128xf16>) -> tensor<1x512x128x128xf16>
    %collapsed_257 = tensor.collapse_shape %377 [[0], [1], [2, 3]] : tensor<1x512x128x128xf16> into tensor<1x512x16384xf16>
    %expanded_258 = tensor.expand_shape %collapsed_257 [[0], [1, 2], [3]] : tensor<1x512x16384xf16> into tensor<1x32x16x16384xf16>
    %378 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_258 : tensor<1x32x16x16384xf16>) outs(%350 : tensor<1x32x16x16384xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x16384xf32>
    %379 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%378 : tensor<1x32x16x16384xf32>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %380 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%379 : tensor<1x32x16x16384xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %381 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%380 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_149 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %382 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%379, %381 : tensor<1x32x16x16384xf64>, tensor<1x32x1x1xf64>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %383 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%382, %382 : tensor<1x32x16x16384xf64>, tensor<1x32x16x16384xf64>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %384 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%383 : tensor<1x32x16x16384xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %385 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%384 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_149 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %386 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%385 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %387 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%378 : tensor<1x32x16x16384xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %388 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%387 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_150 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %389 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%386 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %390 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%389 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %391 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%388 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %392 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%391 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %393 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%390 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %394 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%393 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %395 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_258, %392 : tensor<1x32x16x16384xf16>, tensor<1x32x1x1xf16>) outs(%369 : tensor<1x32x16x16384xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x16384xf16>
    %396 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%395, %394 : tensor<1x32x16x16384xf16>, tensor<1x32x1x1xf16>) outs(%369 : tensor<1x32x16x16384xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x16384xf16>
    %collapsed_259 = tensor.collapse_shape %396 [[0], [1, 2], [3]] : tensor<1x32x16x16384xf16> into tensor<1x512x16384xf16>
    %expanded_260 = tensor.expand_shape %collapsed_259 [[0], [1], [2, 3]] : tensor<1x512x16384xf16> into tensor<1x512x128x128xf16>
    %expanded_261 = tensor.expand_shape %cst_78 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_262 = tensor.expand_shape %cst_77 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %397 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_260, %expanded_262 : tensor<1x512x128x128xf16>, tensor<1x512x1x1xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %398 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%397, %expanded_261 : tensor<1x512x128x128xf16>, tensor<1x512x1x1xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %399 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%398 : tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x512x128x128xf16>
    %400 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%399, %398 : tensor<1x512x128x128xf16>, tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %padded_263 = tensor.pad %400 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x128x128xf16> to tensor<1x512x130x130xf16>
    %401 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_75 : tensor<512xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x128x128xf16>
    %402 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_263, %cst_76 : tensor<1x512x130x130xf16>, tensor<512x512x3x3xf16>) outs(%401 : tensor<1x512x128x128xf16>) -> tensor<1x512x128x128xf16>
    %403 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%349, %402 : tensor<1x512x128x128xf16>, tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %404 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%403 : tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x128x128xf16>
    %collapsed_264 = tensor.collapse_shape %404 [[0], [1], [2, 3]] : tensor<1x512x128x128xf16> into tensor<1x512x16384xf16>
    %expanded_265 = tensor.expand_shape %collapsed_264 [[0], [1, 2], [3]] : tensor<1x512x16384xf16> into tensor<1x32x16x16384xf16>
    %405 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_265 : tensor<1x32x16x16384xf16>) outs(%350 : tensor<1x32x16x16384xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x16384xf32>
    %406 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%405 : tensor<1x32x16x16384xf32>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %407 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%406 : tensor<1x32x16x16384xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %408 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%407 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_149 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %409 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%406, %408 : tensor<1x32x16x16384xf64>, tensor<1x32x1x1xf64>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %410 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%409, %409 : tensor<1x32x16x16384xf64>, tensor<1x32x16x16384xf64>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %411 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%410 : tensor<1x32x16x16384xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %412 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%411 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_149 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %413 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%412 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %414 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%405 : tensor<1x32x16x16384xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %415 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%414 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_150 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %416 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%413 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %417 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%416 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %418 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%415 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %419 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%418 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %420 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%417 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %421 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%420 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %422 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_265, %419 : tensor<1x32x16x16384xf16>, tensor<1x32x1x1xf16>) outs(%369 : tensor<1x32x16x16384xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x16384xf16>
    %423 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%422, %421 : tensor<1x32x16x16384xf16>, tensor<1x32x1x1xf16>) outs(%369 : tensor<1x32x16x16384xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x16384xf16>
    %collapsed_266 = tensor.collapse_shape %423 [[0], [1, 2], [3]] : tensor<1x32x16x16384xf16> into tensor<1x512x16384xf16>
    %expanded_267 = tensor.expand_shape %collapsed_266 [[0], [1], [2, 3]] : tensor<1x512x16384xf16> into tensor<1x512x128x128xf16>
    %expanded_268 = tensor.expand_shape %cst_74 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_269 = tensor.expand_shape %cst_73 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %424 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_267, %expanded_269 : tensor<1x512x128x128xf16>, tensor<1x512x1x1xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %425 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%424, %expanded_268 : tensor<1x512x128x128xf16>, tensor<1x512x1x1xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %426 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%425 : tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x512x128x128xf16>
    %427 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%426, %425 : tensor<1x512x128x128xf16>, tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %padded_270 = tensor.pad %427 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x128x128xf16> to tensor<1x512x130x130xf16>
    %428 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_71 : tensor<512xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x128x128xf16>
    %429 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_270, %cst_72 : tensor<1x512x130x130xf16>, tensor<512x512x3x3xf16>) outs(%428 : tensor<1x512x128x128xf16>) -> tensor<1x512x128x128xf16>
    %collapsed_271 = tensor.collapse_shape %429 [[0], [1], [2, 3]] : tensor<1x512x128x128xf16> into tensor<1x512x16384xf16>
    %expanded_272 = tensor.expand_shape %collapsed_271 [[0], [1, 2], [3]] : tensor<1x512x16384xf16> into tensor<1x32x16x16384xf16>
    %430 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_272 : tensor<1x32x16x16384xf16>) outs(%350 : tensor<1x32x16x16384xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x16384xf32>
    %431 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%430 : tensor<1x32x16x16384xf32>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %432 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%431 : tensor<1x32x16x16384xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %433 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%432 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_149 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %434 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%431, %433 : tensor<1x32x16x16384xf64>, tensor<1x32x1x1xf64>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %435 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%434, %434 : tensor<1x32x16x16384xf64>, tensor<1x32x16x16384xf64>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %436 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%435 : tensor<1x32x16x16384xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %437 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%436 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_149 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %438 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%437 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %439 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%430 : tensor<1x32x16x16384xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %440 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%439 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_150 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %441 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%438 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %442 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%441 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %443 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%440 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %444 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%443 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %445 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%442 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %446 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%445 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %447 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_272, %444 : tensor<1x32x16x16384xf16>, tensor<1x32x1x1xf16>) outs(%369 : tensor<1x32x16x16384xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x16384xf16>
    %448 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%447, %446 : tensor<1x32x16x16384xf16>, tensor<1x32x1x1xf16>) outs(%369 : tensor<1x32x16x16384xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x16384xf16>
    %collapsed_273 = tensor.collapse_shape %448 [[0], [1, 2], [3]] : tensor<1x32x16x16384xf16> into tensor<1x512x16384xf16>
    %expanded_274 = tensor.expand_shape %collapsed_273 [[0], [1], [2, 3]] : tensor<1x512x16384xf16> into tensor<1x512x128x128xf16>
    %expanded_275 = tensor.expand_shape %cst_70 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_276 = tensor.expand_shape %cst_69 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %449 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_274, %expanded_276 : tensor<1x512x128x128xf16>, tensor<1x512x1x1xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %450 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%449, %expanded_275 : tensor<1x512x128x128xf16>, tensor<1x512x1x1xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %451 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%450 : tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x512x128x128xf16>
    %452 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%451, %450 : tensor<1x512x128x128xf16>, tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %padded_277 = tensor.pad %452 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x128x128xf16> to tensor<1x512x130x130xf16>
    %453 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_67 : tensor<512xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x128x128xf16>
    %454 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_277, %cst_68 : tensor<1x512x130x130xf16>, tensor<512x512x3x3xf16>) outs(%453 : tensor<1x512x128x128xf16>) -> tensor<1x512x128x128xf16>
    %455 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%404, %454 : tensor<1x512x128x128xf16>, tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %456 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%455 : tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x128x128xf16>
    %collapsed_278 = tensor.collapse_shape %456 [[0], [1], [2, 3]] : tensor<1x512x128x128xf16> into tensor<1x512x16384xf16>
    %expanded_279 = tensor.expand_shape %collapsed_278 [[0], [1, 2], [3]] : tensor<1x512x16384xf16> into tensor<1x32x16x16384xf16>
    %457 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_279 : tensor<1x32x16x16384xf16>) outs(%350 : tensor<1x32x16x16384xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x16384xf32>
    %458 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%457 : tensor<1x32x16x16384xf32>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %459 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%458 : tensor<1x32x16x16384xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %460 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%459 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_149 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %461 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%458, %460 : tensor<1x32x16x16384xf64>, tensor<1x32x1x1xf64>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %462 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%461, %461 : tensor<1x32x16x16384xf64>, tensor<1x32x16x16384xf64>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %463 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%462 : tensor<1x32x16x16384xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %464 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%463 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_149 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %465 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%464 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %466 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%457 : tensor<1x32x16x16384xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %467 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%466 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_150 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %468 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%465 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %469 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%468 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %470 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%467 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %471 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%470 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %472 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%469 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %473 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%472 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %474 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_279, %471 : tensor<1x32x16x16384xf16>, tensor<1x32x1x1xf16>) outs(%369 : tensor<1x32x16x16384xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x16384xf16>
    %475 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%474, %473 : tensor<1x32x16x16384xf16>, tensor<1x32x1x1xf16>) outs(%369 : tensor<1x32x16x16384xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x16384xf16>
    %collapsed_280 = tensor.collapse_shape %475 [[0], [1, 2], [3]] : tensor<1x32x16x16384xf16> into tensor<1x512x16384xf16>
    %expanded_281 = tensor.expand_shape %collapsed_280 [[0], [1], [2, 3]] : tensor<1x512x16384xf16> into tensor<1x512x128x128xf16>
    %expanded_282 = tensor.expand_shape %cst_66 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_283 = tensor.expand_shape %cst_65 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %476 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_281, %expanded_283 : tensor<1x512x128x128xf16>, tensor<1x512x1x1xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %477 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%476, %expanded_282 : tensor<1x512x128x128xf16>, tensor<1x512x1x1xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %478 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%477 : tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x512x128x128xf16>
    %479 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%478, %477 : tensor<1x512x128x128xf16>, tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %padded_284 = tensor.pad %479 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x128x128xf16> to tensor<1x512x130x130xf16>
    %480 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_63 : tensor<512xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x128x128xf16>
    %481 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_284, %cst_64 : tensor<1x512x130x130xf16>, tensor<512x512x3x3xf16>) outs(%480 : tensor<1x512x128x128xf16>) -> tensor<1x512x128x128xf16>
    %collapsed_285 = tensor.collapse_shape %481 [[0], [1], [2, 3]] : tensor<1x512x128x128xf16> into tensor<1x512x16384xf16>
    %expanded_286 = tensor.expand_shape %collapsed_285 [[0], [1, 2], [3]] : tensor<1x512x16384xf16> into tensor<1x32x16x16384xf16>
    %482 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_286 : tensor<1x32x16x16384xf16>) outs(%350 : tensor<1x32x16x16384xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x16384xf32>
    %483 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%482 : tensor<1x32x16x16384xf32>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %484 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%483 : tensor<1x32x16x16384xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %485 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%484 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_149 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %486 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%483, %485 : tensor<1x32x16x16384xf64>, tensor<1x32x1x1xf64>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %487 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%486, %486 : tensor<1x32x16x16384xf64>, tensor<1x32x16x16384xf64>) outs(%352 : tensor<1x32x16x16384xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x16384xf64>
    %488 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%487 : tensor<1x32x16x16384xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %489 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%488 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_149 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %490 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%489 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %491 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%482 : tensor<1x32x16x16384xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %492 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%491 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_150 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %493 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%490 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %494 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%493 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %495 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%492 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %496 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%495 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %497 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%494 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %498 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%497 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %499 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_286, %496 : tensor<1x32x16x16384xf16>, tensor<1x32x1x1xf16>) outs(%369 : tensor<1x32x16x16384xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x16384xf16>
    %500 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%499, %498 : tensor<1x32x16x16384xf16>, tensor<1x32x1x1xf16>) outs(%369 : tensor<1x32x16x16384xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x16384xf16>
    %collapsed_287 = tensor.collapse_shape %500 [[0], [1, 2], [3]] : tensor<1x32x16x16384xf16> into tensor<1x512x16384xf16>
    %expanded_288 = tensor.expand_shape %collapsed_287 [[0], [1], [2, 3]] : tensor<1x512x16384xf16> into tensor<1x512x128x128xf16>
    %expanded_289 = tensor.expand_shape %cst_62 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_290 = tensor.expand_shape %cst_61 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %501 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_288, %expanded_290 : tensor<1x512x128x128xf16>, tensor<1x512x1x1xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %502 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%501, %expanded_289 : tensor<1x512x128x128xf16>, tensor<1x512x1x1xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %503 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%502 : tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x512x128x128xf16>
    %504 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%503, %502 : tensor<1x512x128x128xf16>, tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %padded_291 = tensor.pad %504 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x128x128xf16> to tensor<1x512x130x130xf16>
    %505 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_59 : tensor<512xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x128x128xf16>
    %506 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_291, %cst_60 : tensor<1x512x130x130xf16>, tensor<512x512x3x3xf16>) outs(%505 : tensor<1x512x128x128xf16>) -> tensor<1x512x128x128xf16>
    %507 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%456, %506 : tensor<1x512x128x128xf16>, tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x128x128xf16>
    %508 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%507 : tensor<1x512x128x128xf16>) outs(%346 : tensor<1x512x128x128xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x128x128xf16>
    %509 = tensor.empty() : tensor<1x512x256x256xf16>
    %510 = linalg.generic {indexing_maps = [#map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} outs(%509 : tensor<1x512x256x256xf16>) {
    ^bb0(%out: f16):
      %878 = linalg.index 0 : index
      %879 = linalg.index 1 : index
      %880 = linalg.index 2 : index
      %881 = linalg.index 3 : index
      %882 = arith.floordivsi %880, %c2 : index
      %883 = arith.floordivsi %881, %c2 : index
      %extracted = tensor.extract %508[%878, %879, %882, %883] : tensor<1x512x128x128xf16>
      linalg.yield %extracted : f16
    } -> tensor<1x512x256x256xf16>
    %padded_292 = tensor.pad %510 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x256x256xf16> to tensor<1x512x258x258xf16>
    %511 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_57 : tensor<512xf16>) outs(%509 : tensor<1x512x256x256xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x512x256x256xf16>
    %512 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_292, %cst_58 : tensor<1x512x258x258xf16>, tensor<512x512x3x3xf16>) outs(%511 : tensor<1x512x256x256xf16>) -> tensor<1x512x256x256xf16>
    %collapsed_293 = tensor.collapse_shape %512 [[0], [1], [2, 3]] : tensor<1x512x256x256xf16> into tensor<1x512x65536xf16>
    %expanded_294 = tensor.expand_shape %collapsed_293 [[0], [1, 2], [3]] : tensor<1x512x65536xf16> into tensor<1x32x16x65536xf16>
    %513 = tensor.empty() : tensor<1x32x16x65536xf32>
    %514 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_294 : tensor<1x32x16x65536xf16>) outs(%513 : tensor<1x32x16x65536xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x16x65536xf32>
    %515 = tensor.empty() : tensor<1x32x16x65536xf64>
    %516 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%514 : tensor<1x32x16x65536xf32>) outs(%515 : tensor<1x32x16x65536xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x65536xf64>
    %517 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%516 : tensor<1x32x16x65536xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %518 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%517 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_151 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %519 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%516, %518 : tensor<1x32x16x65536xf64>, tensor<1x32x1x1xf64>) outs(%515 : tensor<1x32x16x65536xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x65536xf64>
    %520 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%519, %519 : tensor<1x32x16x65536xf64>, tensor<1x32x16x65536xf64>) outs(%515 : tensor<1x32x16x65536xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x16x65536xf64>
    %521 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%520 : tensor<1x32x16x65536xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %522 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%521 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_151 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %523 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%522 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %524 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%514 : tensor<1x32x16x65536xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %525 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%524 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_152 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %526 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%523 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %527 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%526 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %528 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%525 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %529 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%528 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %530 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%527 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %531 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%530 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %532 = tensor.empty() : tensor<1x32x16x65536xf16>
    %533 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_294, %529 : tensor<1x32x16x65536xf16>, tensor<1x32x1x1xf16>) outs(%532 : tensor<1x32x16x65536xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x65536xf16>
    %534 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%533, %531 : tensor<1x32x16x65536xf16>, tensor<1x32x1x1xf16>) outs(%532 : tensor<1x32x16x65536xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x16x65536xf16>
    %collapsed_295 = tensor.collapse_shape %534 [[0], [1, 2], [3]] : tensor<1x32x16x65536xf16> into tensor<1x512x65536xf16>
    %expanded_296 = tensor.expand_shape %collapsed_295 [[0], [1], [2, 3]] : tensor<1x512x65536xf16> into tensor<1x512x256x256xf16>
    %expanded_297 = tensor.expand_shape %cst_56 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %expanded_298 = tensor.expand_shape %cst_55 [[0, 1, 2, 3]] : tensor<512xf16> into tensor<1x512x1x1xf16>
    %535 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_296, %expanded_298 : tensor<1x512x256x256xf16>, tensor<1x512x1x1xf16>) outs(%509 : tensor<1x512x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x256x256xf16>
    %536 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%535, %expanded_297 : tensor<1x512x256x256xf16>, tensor<1x512x1x1xf16>) outs(%509 : tensor<1x512x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x256x256xf16>
    %537 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%536 : tensor<1x512x256x256xf16>) outs(%509 : tensor<1x512x256x256xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x512x256x256xf16>
    %538 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%537, %536 : tensor<1x512x256x256xf16>, tensor<1x512x256x256xf16>) outs(%509 : tensor<1x512x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x512x256x256xf16>
    %padded_299 = tensor.pad %538 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x512x256x256xf16> to tensor<1x512x258x258xf16>
    %539 = tensor.empty() : tensor<1x256x256x256xf16>
    %540 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_53 : tensor<256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x256x256x256xf16>
    %541 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_299, %cst_54 : tensor<1x512x258x258xf16>, tensor<256x512x3x3xf16>) outs(%540 : tensor<1x256x256x256xf16>) -> tensor<1x256x256x256xf16>
    %collapsed_300 = tensor.collapse_shape %541 [[0], [1], [2, 3]] : tensor<1x256x256x256xf16> into tensor<1x256x65536xf16>
    %expanded_301 = tensor.expand_shape %collapsed_300 [[0], [1, 2], [3]] : tensor<1x256x65536xf16> into tensor<1x32x8x65536xf16>
    %542 = tensor.empty() : tensor<1x32x8x65536xf32>
    %543 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_301 : tensor<1x32x8x65536xf16>) outs(%542 : tensor<1x32x8x65536xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x8x65536xf32>
    %544 = tensor.empty() : tensor<1x32x8x65536xf64>
    %545 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%543 : tensor<1x32x8x65536xf32>) outs(%544 : tensor<1x32x8x65536xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x65536xf64>
    %546 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%545 : tensor<1x32x8x65536xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %547 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%546 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_153 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %548 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%545, %547 : tensor<1x32x8x65536xf64>, tensor<1x32x1x1xf64>) outs(%544 : tensor<1x32x8x65536xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x65536xf64>
    %549 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%548, %548 : tensor<1x32x8x65536xf64>, tensor<1x32x8x65536xf64>) outs(%544 : tensor<1x32x8x65536xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x65536xf64>
    %550 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%549 : tensor<1x32x8x65536xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %551 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%550 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_153 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %552 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%551 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %553 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%543 : tensor<1x32x8x65536xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %554 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%553 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_154 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %555 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%552 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %556 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%555 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %557 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%554 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %558 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%557 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %559 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%556 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %560 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%559 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %561 = tensor.empty() : tensor<1x32x8x65536xf16>
    %562 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_301, %558 : tensor<1x32x8x65536xf16>, tensor<1x32x1x1xf16>) outs(%561 : tensor<1x32x8x65536xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x8x65536xf16>
    %563 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%562, %560 : tensor<1x32x8x65536xf16>, tensor<1x32x1x1xf16>) outs(%561 : tensor<1x32x8x65536xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x8x65536xf16>
    %collapsed_302 = tensor.collapse_shape %563 [[0], [1, 2], [3]] : tensor<1x32x8x65536xf16> into tensor<1x256x65536xf16>
    %expanded_303 = tensor.expand_shape %collapsed_302 [[0], [1], [2, 3]] : tensor<1x256x65536xf16> into tensor<1x256x256x256xf16>
    %expanded_304 = tensor.expand_shape %cst_52 [[0, 1, 2, 3]] : tensor<256xf16> into tensor<1x256x1x1xf16>
    %expanded_305 = tensor.expand_shape %cst_51 [[0, 1, 2, 3]] : tensor<256xf16> into tensor<1x256x1x1xf16>
    %564 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_303, %expanded_305 : tensor<1x256x256x256xf16>, tensor<1x256x1x1xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %565 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%564, %expanded_304 : tensor<1x256x256x256xf16>, tensor<1x256x1x1xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %566 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%565 : tensor<1x256x256x256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x256x256x256xf16>
    %567 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%566, %565 : tensor<1x256x256x256xf16>, tensor<1x256x256x256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %padded_306 = tensor.pad %567 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x256x256x256xf16> to tensor<1x256x258x258xf16>
    %568 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_49 : tensor<256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x256x256x256xf16>
    %569 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_306, %cst_50 : tensor<1x256x258x258xf16>, tensor<256x256x3x3xf16>) outs(%568 : tensor<1x256x256x256xf16>) -> tensor<1x256x256x256xf16>
    %570 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_47 : tensor<256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x256x256x256xf16>
    %571 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%512, %cst_48 : tensor<1x512x256x256xf16>, tensor<256x512x1x1xf16>) outs(%570 : tensor<1x256x256x256xf16>) -> tensor<1x256x256x256xf16>
    %572 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%571, %569 : tensor<1x256x256x256xf16>, tensor<1x256x256x256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %573 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%572 : tensor<1x256x256x256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x256x256x256xf16>
    %collapsed_307 = tensor.collapse_shape %573 [[0], [1], [2, 3]] : tensor<1x256x256x256xf16> into tensor<1x256x65536xf16>
    %expanded_308 = tensor.expand_shape %collapsed_307 [[0], [1, 2], [3]] : tensor<1x256x65536xf16> into tensor<1x32x8x65536xf16>
    %574 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_308 : tensor<1x32x8x65536xf16>) outs(%542 : tensor<1x32x8x65536xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x8x65536xf32>
    %575 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%574 : tensor<1x32x8x65536xf32>) outs(%544 : tensor<1x32x8x65536xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x65536xf64>
    %576 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%575 : tensor<1x32x8x65536xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %577 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%576 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_153 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %578 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%575, %577 : tensor<1x32x8x65536xf64>, tensor<1x32x1x1xf64>) outs(%544 : tensor<1x32x8x65536xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x65536xf64>
    %579 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%578, %578 : tensor<1x32x8x65536xf64>, tensor<1x32x8x65536xf64>) outs(%544 : tensor<1x32x8x65536xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x65536xf64>
    %580 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%579 : tensor<1x32x8x65536xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %581 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%580 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_153 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %582 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%581 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %583 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%574 : tensor<1x32x8x65536xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %584 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%583 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_154 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %585 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%582 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %586 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%585 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %587 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%584 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %588 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%587 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %589 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%586 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %590 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%589 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %591 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_308, %588 : tensor<1x32x8x65536xf16>, tensor<1x32x1x1xf16>) outs(%561 : tensor<1x32x8x65536xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x8x65536xf16>
    %592 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%591, %590 : tensor<1x32x8x65536xf16>, tensor<1x32x1x1xf16>) outs(%561 : tensor<1x32x8x65536xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x8x65536xf16>
    %collapsed_309 = tensor.collapse_shape %592 [[0], [1, 2], [3]] : tensor<1x32x8x65536xf16> into tensor<1x256x65536xf16>
    %expanded_310 = tensor.expand_shape %collapsed_309 [[0], [1], [2, 3]] : tensor<1x256x65536xf16> into tensor<1x256x256x256xf16>
    %expanded_311 = tensor.expand_shape %cst_46 [[0, 1, 2, 3]] : tensor<256xf16> into tensor<1x256x1x1xf16>
    %expanded_312 = tensor.expand_shape %cst_45 [[0, 1, 2, 3]] : tensor<256xf16> into tensor<1x256x1x1xf16>
    %593 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_310, %expanded_312 : tensor<1x256x256x256xf16>, tensor<1x256x1x1xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %594 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%593, %expanded_311 : tensor<1x256x256x256xf16>, tensor<1x256x1x1xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %595 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%594 : tensor<1x256x256x256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x256x256x256xf16>
    %596 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%595, %594 : tensor<1x256x256x256xf16>, tensor<1x256x256x256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %padded_313 = tensor.pad %596 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x256x256x256xf16> to tensor<1x256x258x258xf16>
    %597 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_43 : tensor<256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x256x256x256xf16>
    %598 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_313, %cst_44 : tensor<1x256x258x258xf16>, tensor<256x256x3x3xf16>) outs(%597 : tensor<1x256x256x256xf16>) -> tensor<1x256x256x256xf16>
    %collapsed_314 = tensor.collapse_shape %598 [[0], [1], [2, 3]] : tensor<1x256x256x256xf16> into tensor<1x256x65536xf16>
    %expanded_315 = tensor.expand_shape %collapsed_314 [[0], [1, 2], [3]] : tensor<1x256x65536xf16> into tensor<1x32x8x65536xf16>
    %599 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_315 : tensor<1x32x8x65536xf16>) outs(%542 : tensor<1x32x8x65536xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x8x65536xf32>
    %600 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%599 : tensor<1x32x8x65536xf32>) outs(%544 : tensor<1x32x8x65536xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x65536xf64>
    %601 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%600 : tensor<1x32x8x65536xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %602 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%601 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_153 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %603 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%600, %602 : tensor<1x32x8x65536xf64>, tensor<1x32x1x1xf64>) outs(%544 : tensor<1x32x8x65536xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x65536xf64>
    %604 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%603, %603 : tensor<1x32x8x65536xf64>, tensor<1x32x8x65536xf64>) outs(%544 : tensor<1x32x8x65536xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x65536xf64>
    %605 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%604 : tensor<1x32x8x65536xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %606 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%605 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_153 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %607 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%606 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %608 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%599 : tensor<1x32x8x65536xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %609 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%608 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_154 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %610 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%607 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %611 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%610 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %612 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%609 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %613 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%612 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %614 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%611 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %615 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%614 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %616 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_315, %613 : tensor<1x32x8x65536xf16>, tensor<1x32x1x1xf16>) outs(%561 : tensor<1x32x8x65536xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x8x65536xf16>
    %617 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%616, %615 : tensor<1x32x8x65536xf16>, tensor<1x32x1x1xf16>) outs(%561 : tensor<1x32x8x65536xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x8x65536xf16>
    %collapsed_316 = tensor.collapse_shape %617 [[0], [1, 2], [3]] : tensor<1x32x8x65536xf16> into tensor<1x256x65536xf16>
    %expanded_317 = tensor.expand_shape %collapsed_316 [[0], [1], [2, 3]] : tensor<1x256x65536xf16> into tensor<1x256x256x256xf16>
    %expanded_318 = tensor.expand_shape %cst_42 [[0, 1, 2, 3]] : tensor<256xf16> into tensor<1x256x1x1xf16>
    %expanded_319 = tensor.expand_shape %cst_41 [[0, 1, 2, 3]] : tensor<256xf16> into tensor<1x256x1x1xf16>
    %618 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_317, %expanded_319 : tensor<1x256x256x256xf16>, tensor<1x256x1x1xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %619 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%618, %expanded_318 : tensor<1x256x256x256xf16>, tensor<1x256x1x1xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %620 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%619 : tensor<1x256x256x256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x256x256x256xf16>
    %621 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%620, %619 : tensor<1x256x256x256xf16>, tensor<1x256x256x256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %padded_320 = tensor.pad %621 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x256x256x256xf16> to tensor<1x256x258x258xf16>
    %622 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_39 : tensor<256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x256x256x256xf16>
    %623 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_320, %cst_40 : tensor<1x256x258x258xf16>, tensor<256x256x3x3xf16>) outs(%622 : tensor<1x256x256x256xf16>) -> tensor<1x256x256x256xf16>
    %624 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%573, %623 : tensor<1x256x256x256xf16>, tensor<1x256x256x256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %625 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%624 : tensor<1x256x256x256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x256x256x256xf16>
    %collapsed_321 = tensor.collapse_shape %625 [[0], [1], [2, 3]] : tensor<1x256x256x256xf16> into tensor<1x256x65536xf16>
    %expanded_322 = tensor.expand_shape %collapsed_321 [[0], [1, 2], [3]] : tensor<1x256x65536xf16> into tensor<1x32x8x65536xf16>
    %626 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_322 : tensor<1x32x8x65536xf16>) outs(%542 : tensor<1x32x8x65536xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x8x65536xf32>
    %627 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%626 : tensor<1x32x8x65536xf32>) outs(%544 : tensor<1x32x8x65536xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x65536xf64>
    %628 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%627 : tensor<1x32x8x65536xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %629 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%628 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_153 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %630 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%627, %629 : tensor<1x32x8x65536xf64>, tensor<1x32x1x1xf64>) outs(%544 : tensor<1x32x8x65536xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x65536xf64>
    %631 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%630, %630 : tensor<1x32x8x65536xf64>, tensor<1x32x8x65536xf64>) outs(%544 : tensor<1x32x8x65536xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x65536xf64>
    %632 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%631 : tensor<1x32x8x65536xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %633 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%632 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_153 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %634 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%633 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %635 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%626 : tensor<1x32x8x65536xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %636 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%635 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_154 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %637 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%634 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %638 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%637 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %639 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%636 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %640 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%639 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %641 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%638 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %642 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%641 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %643 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_322, %640 : tensor<1x32x8x65536xf16>, tensor<1x32x1x1xf16>) outs(%561 : tensor<1x32x8x65536xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x8x65536xf16>
    %644 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%643, %642 : tensor<1x32x8x65536xf16>, tensor<1x32x1x1xf16>) outs(%561 : tensor<1x32x8x65536xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x8x65536xf16>
    %collapsed_323 = tensor.collapse_shape %644 [[0], [1, 2], [3]] : tensor<1x32x8x65536xf16> into tensor<1x256x65536xf16>
    %expanded_324 = tensor.expand_shape %collapsed_323 [[0], [1], [2, 3]] : tensor<1x256x65536xf16> into tensor<1x256x256x256xf16>
    %expanded_325 = tensor.expand_shape %cst_38 [[0, 1, 2, 3]] : tensor<256xf16> into tensor<1x256x1x1xf16>
    %expanded_326 = tensor.expand_shape %cst_37 [[0, 1, 2, 3]] : tensor<256xf16> into tensor<1x256x1x1xf16>
    %645 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_324, %expanded_326 : tensor<1x256x256x256xf16>, tensor<1x256x1x1xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %646 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%645, %expanded_325 : tensor<1x256x256x256xf16>, tensor<1x256x1x1xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %647 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%646 : tensor<1x256x256x256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x256x256x256xf16>
    %648 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%647, %646 : tensor<1x256x256x256xf16>, tensor<1x256x256x256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %padded_327 = tensor.pad %648 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x256x256x256xf16> to tensor<1x256x258x258xf16>
    %649 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_35 : tensor<256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x256x256x256xf16>
    %650 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_327, %cst_36 : tensor<1x256x258x258xf16>, tensor<256x256x3x3xf16>) outs(%649 : tensor<1x256x256x256xf16>) -> tensor<1x256x256x256xf16>
    %collapsed_328 = tensor.collapse_shape %650 [[0], [1], [2, 3]] : tensor<1x256x256x256xf16> into tensor<1x256x65536xf16>
    %expanded_329 = tensor.expand_shape %collapsed_328 [[0], [1, 2], [3]] : tensor<1x256x65536xf16> into tensor<1x32x8x65536xf16>
    %651 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_329 : tensor<1x32x8x65536xf16>) outs(%542 : tensor<1x32x8x65536xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x8x65536xf32>
    %652 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%651 : tensor<1x32x8x65536xf32>) outs(%544 : tensor<1x32x8x65536xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x65536xf64>
    %653 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%652 : tensor<1x32x8x65536xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %654 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%653 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_153 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %655 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%652, %654 : tensor<1x32x8x65536xf64>, tensor<1x32x1x1xf64>) outs(%544 : tensor<1x32x8x65536xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x65536xf64>
    %656 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%655, %655 : tensor<1x32x8x65536xf64>, tensor<1x32x8x65536xf64>) outs(%544 : tensor<1x32x8x65536xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x65536xf64>
    %657 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%656 : tensor<1x32x8x65536xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %658 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%657 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_153 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %659 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%658 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %660 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%651 : tensor<1x32x8x65536xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %661 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%660 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_154 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %662 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%659 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %663 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%662 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %664 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%661 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %665 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%664 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %666 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%663 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %667 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%666 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %668 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_329, %665 : tensor<1x32x8x65536xf16>, tensor<1x32x1x1xf16>) outs(%561 : tensor<1x32x8x65536xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x8x65536xf16>
    %669 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%668, %667 : tensor<1x32x8x65536xf16>, tensor<1x32x1x1xf16>) outs(%561 : tensor<1x32x8x65536xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x8x65536xf16>
    %collapsed_330 = tensor.collapse_shape %669 [[0], [1, 2], [3]] : tensor<1x32x8x65536xf16> into tensor<1x256x65536xf16>
    %expanded_331 = tensor.expand_shape %collapsed_330 [[0], [1], [2, 3]] : tensor<1x256x65536xf16> into tensor<1x256x256x256xf16>
    %expanded_332 = tensor.expand_shape %cst_34 [[0, 1, 2, 3]] : tensor<256xf16> into tensor<1x256x1x1xf16>
    %expanded_333 = tensor.expand_shape %cst_33 [[0, 1, 2, 3]] : tensor<256xf16> into tensor<1x256x1x1xf16>
    %670 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_331, %expanded_333 : tensor<1x256x256x256xf16>, tensor<1x256x1x1xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %671 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%670, %expanded_332 : tensor<1x256x256x256xf16>, tensor<1x256x1x1xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %672 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%671 : tensor<1x256x256x256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x256x256x256xf16>
    %673 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%672, %671 : tensor<1x256x256x256xf16>, tensor<1x256x256x256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %padded_334 = tensor.pad %673 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x256x256x256xf16> to tensor<1x256x258x258xf16>
    %674 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_31 : tensor<256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x256x256x256xf16>
    %675 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_334, %cst_32 : tensor<1x256x258x258xf16>, tensor<256x256x3x3xf16>) outs(%674 : tensor<1x256x256x256xf16>) -> tensor<1x256x256x256xf16>
    %676 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%625, %675 : tensor<1x256x256x256xf16>, tensor<1x256x256x256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x256x256xf16>
    %677 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%676 : tensor<1x256x256x256xf16>) outs(%539 : tensor<1x256x256x256xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x256x256x256xf16>
    %678 = tensor.empty() : tensor<1x256x512x512xf16>
    %679 = linalg.generic {indexing_maps = [#map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} outs(%678 : tensor<1x256x512x512xf16>) {
    ^bb0(%out: f16):
      %878 = linalg.index 0 : index
      %879 = linalg.index 1 : index
      %880 = linalg.index 2 : index
      %881 = linalg.index 3 : index
      %882 = arith.floordivsi %880, %c2 : index
      %883 = arith.floordivsi %881, %c2 : index
      %extracted = tensor.extract %677[%878, %879, %882, %883] : tensor<1x256x256x256xf16>
      linalg.yield %extracted : f16
    } -> tensor<1x256x512x512xf16>
    %padded_335 = tensor.pad %679 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x256x512x512xf16> to tensor<1x256x514x514xf16>
    %680 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_29 : tensor<256xf16>) outs(%678 : tensor<1x256x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x256x512x512xf16>
    %681 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_335, %cst_30 : tensor<1x256x514x514xf16>, tensor<256x256x3x3xf16>) outs(%680 : tensor<1x256x512x512xf16>) -> tensor<1x256x512x512xf16>
    %collapsed_336 = tensor.collapse_shape %681 [[0], [1], [2, 3]] : tensor<1x256x512x512xf16> into tensor<1x256x262144xf16>
    %expanded_337 = tensor.expand_shape %collapsed_336 [[0], [1, 2], [3]] : tensor<1x256x262144xf16> into tensor<1x32x8x262144xf16>
    %682 = tensor.empty() : tensor<1x32x8x262144xf32>
    %683 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_337 : tensor<1x32x8x262144xf16>) outs(%682 : tensor<1x32x8x262144xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x8x262144xf32>
    %684 = tensor.empty() : tensor<1x32x8x262144xf64>
    %685 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%683 : tensor<1x32x8x262144xf32>) outs(%684 : tensor<1x32x8x262144xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x262144xf64>
    %686 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%685 : tensor<1x32x8x262144xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %687 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%686 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_155 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %688 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%685, %687 : tensor<1x32x8x262144xf64>, tensor<1x32x1x1xf64>) outs(%684 : tensor<1x32x8x262144xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x262144xf64>
    %689 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%688, %688 : tensor<1x32x8x262144xf64>, tensor<1x32x8x262144xf64>) outs(%684 : tensor<1x32x8x262144xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x8x262144xf64>
    %690 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%689 : tensor<1x32x8x262144xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %691 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%690 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_155 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %692 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%691 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %693 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%683 : tensor<1x32x8x262144xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %694 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%693 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_156 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %695 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%692 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %696 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%695 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %697 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%694 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %698 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%697 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %699 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%696 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %700 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%699 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %701 = tensor.empty() : tensor<1x32x8x262144xf16>
    %702 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_337, %698 : tensor<1x32x8x262144xf16>, tensor<1x32x1x1xf16>) outs(%701 : tensor<1x32x8x262144xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x8x262144xf16>
    %703 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%702, %700 : tensor<1x32x8x262144xf16>, tensor<1x32x1x1xf16>) outs(%701 : tensor<1x32x8x262144xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x8x262144xf16>
    %collapsed_338 = tensor.collapse_shape %703 [[0], [1, 2], [3]] : tensor<1x32x8x262144xf16> into tensor<1x256x262144xf16>
    %expanded_339 = tensor.expand_shape %collapsed_338 [[0], [1], [2, 3]] : tensor<1x256x262144xf16> into tensor<1x256x512x512xf16>
    %expanded_340 = tensor.expand_shape %cst_28 [[0, 1, 2, 3]] : tensor<256xf16> into tensor<1x256x1x1xf16>
    %expanded_341 = tensor.expand_shape %cst_27 [[0, 1, 2, 3]] : tensor<256xf16> into tensor<1x256x1x1xf16>
    %704 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_339, %expanded_341 : tensor<1x256x512x512xf16>, tensor<1x256x1x1xf16>) outs(%678 : tensor<1x256x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x512x512xf16>
    %705 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%704, %expanded_340 : tensor<1x256x512x512xf16>, tensor<1x256x1x1xf16>) outs(%678 : tensor<1x256x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x512x512xf16>
    %706 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%705 : tensor<1x256x512x512xf16>) outs(%678 : tensor<1x256x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x256x512x512xf16>
    %707 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%706, %705 : tensor<1x256x512x512xf16>, tensor<1x256x512x512xf16>) outs(%678 : tensor<1x256x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x256x512x512xf16>
    %padded_342 = tensor.pad %707 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x256x512x512xf16> to tensor<1x256x514x514xf16>
    %708 = tensor.empty() : tensor<1x128x512x512xf16>
    %709 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_25 : tensor<128xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x128x512x512xf16>
    %710 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_342, %cst_26 : tensor<1x256x514x514xf16>, tensor<128x256x3x3xf16>) outs(%709 : tensor<1x128x512x512xf16>) -> tensor<1x128x512x512xf16>
    %collapsed_343 = tensor.collapse_shape %710 [[0], [1], [2, 3]] : tensor<1x128x512x512xf16> into tensor<1x128x262144xf16>
    %expanded_344 = tensor.expand_shape %collapsed_343 [[0], [1, 2], [3]] : tensor<1x128x262144xf16> into tensor<1x32x4x262144xf16>
    %711 = tensor.empty() : tensor<1x32x4x262144xf32>
    %712 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_344 : tensor<1x32x4x262144xf16>) outs(%711 : tensor<1x32x4x262144xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x4x262144xf32>
    %713 = tensor.empty() : tensor<1x32x4x262144xf64>
    %714 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%712 : tensor<1x32x4x262144xf32>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %715 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%714 : tensor<1x32x4x262144xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %716 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%715 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_151 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %717 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%714, %716 : tensor<1x32x4x262144xf64>, tensor<1x32x1x1xf64>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %718 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%717, %717 : tensor<1x32x4x262144xf64>, tensor<1x32x4x262144xf64>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %719 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%718 : tensor<1x32x4x262144xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %720 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%719 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_151 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %721 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%720 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %722 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%712 : tensor<1x32x4x262144xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %723 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%722 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_152 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %724 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%721 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %725 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%724 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %726 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%723 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %727 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%726 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %728 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%725 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %729 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%728 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %730 = tensor.empty() : tensor<1x32x4x262144xf16>
    %731 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_344, %727 : tensor<1x32x4x262144xf16>, tensor<1x32x1x1xf16>) outs(%730 : tensor<1x32x4x262144xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x4x262144xf16>
    %732 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%731, %729 : tensor<1x32x4x262144xf16>, tensor<1x32x1x1xf16>) outs(%730 : tensor<1x32x4x262144xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x4x262144xf16>
    %collapsed_345 = tensor.collapse_shape %732 [[0], [1, 2], [3]] : tensor<1x32x4x262144xf16> into tensor<1x128x262144xf16>
    %expanded_346 = tensor.expand_shape %collapsed_345 [[0], [1], [2, 3]] : tensor<1x128x262144xf16> into tensor<1x128x512x512xf16>
    %expanded_347 = tensor.expand_shape %cst_24 [[0, 1, 2, 3]] : tensor<128xf16> into tensor<1x128x1x1xf16>
    %expanded_348 = tensor.expand_shape %cst_23 [[0, 1, 2, 3]] : tensor<128xf16> into tensor<1x128x1x1xf16>
    %733 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_346, %expanded_348 : tensor<1x128x512x512xf16>, tensor<1x128x1x1xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %734 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%733, %expanded_347 : tensor<1x128x512x512xf16>, tensor<1x128x1x1xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %735 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%734 : tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x128x512x512xf16>
    %736 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%735, %734 : tensor<1x128x512x512xf16>, tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %padded_349 = tensor.pad %736 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x128x512x512xf16> to tensor<1x128x514x514xf16>
    %737 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_21 : tensor<128xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x128x512x512xf16>
    %738 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_349, %cst_22 : tensor<1x128x514x514xf16>, tensor<128x128x3x3xf16>) outs(%737 : tensor<1x128x512x512xf16>) -> tensor<1x128x512x512xf16>
    %739 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_19 : tensor<128xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x128x512x512xf16>
    %740 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%681, %cst_20 : tensor<1x256x512x512xf16>, tensor<128x256x1x1xf16>) outs(%739 : tensor<1x128x512x512xf16>) -> tensor<1x128x512x512xf16>
    %741 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%740, %738 : tensor<1x128x512x512xf16>, tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %742 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%741 : tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x128x512x512xf16>
    %collapsed_350 = tensor.collapse_shape %742 [[0], [1], [2, 3]] : tensor<1x128x512x512xf16> into tensor<1x128x262144xf16>
    %expanded_351 = tensor.expand_shape %collapsed_350 [[0], [1, 2], [3]] : tensor<1x128x262144xf16> into tensor<1x32x4x262144xf16>
    %743 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_351 : tensor<1x32x4x262144xf16>) outs(%711 : tensor<1x32x4x262144xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x4x262144xf32>
    %744 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%743 : tensor<1x32x4x262144xf32>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %745 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%744 : tensor<1x32x4x262144xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %746 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%745 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_151 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %747 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%744, %746 : tensor<1x32x4x262144xf64>, tensor<1x32x1x1xf64>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %748 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%747, %747 : tensor<1x32x4x262144xf64>, tensor<1x32x4x262144xf64>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %749 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%748 : tensor<1x32x4x262144xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %750 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%749 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_151 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %751 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%750 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %752 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%743 : tensor<1x32x4x262144xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %753 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%752 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_152 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %754 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%751 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %755 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%754 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %756 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%753 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %757 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%756 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %758 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%755 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %759 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%758 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %760 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_351, %757 : tensor<1x32x4x262144xf16>, tensor<1x32x1x1xf16>) outs(%730 : tensor<1x32x4x262144xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x4x262144xf16>
    %761 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%760, %759 : tensor<1x32x4x262144xf16>, tensor<1x32x1x1xf16>) outs(%730 : tensor<1x32x4x262144xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x4x262144xf16>
    %collapsed_352 = tensor.collapse_shape %761 [[0], [1, 2], [3]] : tensor<1x32x4x262144xf16> into tensor<1x128x262144xf16>
    %expanded_353 = tensor.expand_shape %collapsed_352 [[0], [1], [2, 3]] : tensor<1x128x262144xf16> into tensor<1x128x512x512xf16>
    %expanded_354 = tensor.expand_shape %cst_18 [[0, 1, 2, 3]] : tensor<128xf16> into tensor<1x128x1x1xf16>
    %expanded_355 = tensor.expand_shape %cst_17 [[0, 1, 2, 3]] : tensor<128xf16> into tensor<1x128x1x1xf16>
    %762 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_353, %expanded_355 : tensor<1x128x512x512xf16>, tensor<1x128x1x1xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %763 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%762, %expanded_354 : tensor<1x128x512x512xf16>, tensor<1x128x1x1xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %764 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%763 : tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x128x512x512xf16>
    %765 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%764, %763 : tensor<1x128x512x512xf16>, tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %padded_356 = tensor.pad %765 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x128x512x512xf16> to tensor<1x128x514x514xf16>
    %766 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_15 : tensor<128xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x128x512x512xf16>
    %767 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_356, %cst_16 : tensor<1x128x514x514xf16>, tensor<128x128x3x3xf16>) outs(%766 : tensor<1x128x512x512xf16>) -> tensor<1x128x512x512xf16>
    %collapsed_357 = tensor.collapse_shape %767 [[0], [1], [2, 3]] : tensor<1x128x512x512xf16> into tensor<1x128x262144xf16>
    %expanded_358 = tensor.expand_shape %collapsed_357 [[0], [1, 2], [3]] : tensor<1x128x262144xf16> into tensor<1x32x4x262144xf16>
    %768 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_358 : tensor<1x32x4x262144xf16>) outs(%711 : tensor<1x32x4x262144xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x4x262144xf32>
    %769 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%768 : tensor<1x32x4x262144xf32>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %770 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%769 : tensor<1x32x4x262144xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %771 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%770 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_151 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %772 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%769, %771 : tensor<1x32x4x262144xf64>, tensor<1x32x1x1xf64>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %773 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%772, %772 : tensor<1x32x4x262144xf64>, tensor<1x32x4x262144xf64>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %774 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%773 : tensor<1x32x4x262144xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %775 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%774 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_151 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %776 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%775 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %777 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%768 : tensor<1x32x4x262144xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %778 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%777 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_152 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %779 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%776 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %780 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%779 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %781 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%778 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %782 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%781 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %783 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%780 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %784 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%783 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %785 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_358, %782 : tensor<1x32x4x262144xf16>, tensor<1x32x1x1xf16>) outs(%730 : tensor<1x32x4x262144xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x4x262144xf16>
    %786 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%785, %784 : tensor<1x32x4x262144xf16>, tensor<1x32x1x1xf16>) outs(%730 : tensor<1x32x4x262144xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x4x262144xf16>
    %collapsed_359 = tensor.collapse_shape %786 [[0], [1, 2], [3]] : tensor<1x32x4x262144xf16> into tensor<1x128x262144xf16>
    %expanded_360 = tensor.expand_shape %collapsed_359 [[0], [1], [2, 3]] : tensor<1x128x262144xf16> into tensor<1x128x512x512xf16>
    %expanded_361 = tensor.expand_shape %cst_14 [[0, 1, 2, 3]] : tensor<128xf16> into tensor<1x128x1x1xf16>
    %expanded_362 = tensor.expand_shape %cst_13 [[0, 1, 2, 3]] : tensor<128xf16> into tensor<1x128x1x1xf16>
    %787 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_360, %expanded_362 : tensor<1x128x512x512xf16>, tensor<1x128x1x1xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %788 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%787, %expanded_361 : tensor<1x128x512x512xf16>, tensor<1x128x1x1xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %789 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%788 : tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x128x512x512xf16>
    %790 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%789, %788 : tensor<1x128x512x512xf16>, tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %padded_363 = tensor.pad %790 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x128x512x512xf16> to tensor<1x128x514x514xf16>
    %791 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_11 : tensor<128xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x128x512x512xf16>
    %792 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_363, %cst_12 : tensor<1x128x514x514xf16>, tensor<128x128x3x3xf16>) outs(%791 : tensor<1x128x512x512xf16>) -> tensor<1x128x512x512xf16>
    %793 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%742, %792 : tensor<1x128x512x512xf16>, tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %794 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%793 : tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x128x512x512xf16>
    %collapsed_364 = tensor.collapse_shape %794 [[0], [1], [2, 3]] : tensor<1x128x512x512xf16> into tensor<1x128x262144xf16>
    %expanded_365 = tensor.expand_shape %collapsed_364 [[0], [1, 2], [3]] : tensor<1x128x262144xf16> into tensor<1x32x4x262144xf16>
    %795 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_365 : tensor<1x32x4x262144xf16>) outs(%711 : tensor<1x32x4x262144xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x4x262144xf32>
    %796 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%795 : tensor<1x32x4x262144xf32>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %797 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%796 : tensor<1x32x4x262144xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %798 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%797 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_151 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %799 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%796, %798 : tensor<1x32x4x262144xf64>, tensor<1x32x1x1xf64>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %800 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%799, %799 : tensor<1x32x4x262144xf64>, tensor<1x32x4x262144xf64>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %801 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%800 : tensor<1x32x4x262144xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %802 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%801 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_151 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %803 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%802 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %804 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%795 : tensor<1x32x4x262144xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %805 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%804 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_152 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %806 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%803 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %807 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%806 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %808 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%805 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %809 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%808 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %810 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%807 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %811 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%810 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %812 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_365, %809 : tensor<1x32x4x262144xf16>, tensor<1x32x1x1xf16>) outs(%730 : tensor<1x32x4x262144xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x4x262144xf16>
    %813 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%812, %811 : tensor<1x32x4x262144xf16>, tensor<1x32x1x1xf16>) outs(%730 : tensor<1x32x4x262144xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x4x262144xf16>
    %collapsed_366 = tensor.collapse_shape %813 [[0], [1, 2], [3]] : tensor<1x32x4x262144xf16> into tensor<1x128x262144xf16>
    %expanded_367 = tensor.expand_shape %collapsed_366 [[0], [1], [2, 3]] : tensor<1x128x262144xf16> into tensor<1x128x512x512xf16>
    %expanded_368 = tensor.expand_shape %cst_10 [[0, 1, 2, 3]] : tensor<128xf16> into tensor<1x128x1x1xf16>
    %expanded_369 = tensor.expand_shape %cst_9 [[0, 1, 2, 3]] : tensor<128xf16> into tensor<1x128x1x1xf16>
    %814 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_367, %expanded_369 : tensor<1x128x512x512xf16>, tensor<1x128x1x1xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %815 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%814, %expanded_368 : tensor<1x128x512x512xf16>, tensor<1x128x1x1xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %816 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%815 : tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x128x512x512xf16>
    %817 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%816, %815 : tensor<1x128x512x512xf16>, tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %padded_370 = tensor.pad %817 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x128x512x512xf16> to tensor<1x128x514x514xf16>
    %818 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_7 : tensor<128xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x128x512x512xf16>
    %819 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_370, %cst_8 : tensor<1x128x514x514xf16>, tensor<128x128x3x3xf16>) outs(%818 : tensor<1x128x512x512xf16>) -> tensor<1x128x512x512xf16>
    %collapsed_371 = tensor.collapse_shape %819 [[0], [1], [2, 3]] : tensor<1x128x512x512xf16> into tensor<1x128x262144xf16>
    %expanded_372 = tensor.expand_shape %collapsed_371 [[0], [1, 2], [3]] : tensor<1x128x262144xf16> into tensor<1x32x4x262144xf16>
    %820 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_372 : tensor<1x32x4x262144xf16>) outs(%711 : tensor<1x32x4x262144xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x4x262144xf32>
    %821 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%820 : tensor<1x32x4x262144xf32>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %822 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%821 : tensor<1x32x4x262144xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %823 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%822 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_151 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %824 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%821, %823 : tensor<1x32x4x262144xf64>, tensor<1x32x1x1xf64>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %825 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%824, %824 : tensor<1x32x4x262144xf64>, tensor<1x32x4x262144xf64>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %826 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%825 : tensor<1x32x4x262144xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %827 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%826 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_151 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %828 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%827 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %829 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%820 : tensor<1x32x4x262144xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %830 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%829 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_152 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %831 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%828 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %832 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%831 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %833 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%830 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %834 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%833 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %835 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%832 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %836 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%835 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %837 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_372, %834 : tensor<1x32x4x262144xf16>, tensor<1x32x1x1xf16>) outs(%730 : tensor<1x32x4x262144xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x4x262144xf16>
    %838 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%837, %836 : tensor<1x32x4x262144xf16>, tensor<1x32x1x1xf16>) outs(%730 : tensor<1x32x4x262144xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x4x262144xf16>
    %collapsed_373 = tensor.collapse_shape %838 [[0], [1, 2], [3]] : tensor<1x32x4x262144xf16> into tensor<1x128x262144xf16>
    %expanded_374 = tensor.expand_shape %collapsed_373 [[0], [1], [2, 3]] : tensor<1x128x262144xf16> into tensor<1x128x512x512xf16>
    %expanded_375 = tensor.expand_shape %cst_6 [[0, 1, 2, 3]] : tensor<128xf16> into tensor<1x128x1x1xf16>
    %expanded_376 = tensor.expand_shape %cst_5 [[0, 1, 2, 3]] : tensor<128xf16> into tensor<1x128x1x1xf16>
    %839 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_374, %expanded_376 : tensor<1x128x512x512xf16>, tensor<1x128x1x1xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %840 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%839, %expanded_375 : tensor<1x128x512x512xf16>, tensor<1x128x1x1xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %841 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%840 : tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x128x512x512xf16>
    %842 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%841, %840 : tensor<1x128x512x512xf16>, tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %padded_377 = tensor.pad %842 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x128x512x512xf16> to tensor<1x128x514x514xf16>
    %843 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_3 : tensor<128xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x128x512x512xf16>
    %844 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_377, %cst_4 : tensor<1x128x514x514xf16>, tensor<128x128x3x3xf16>) outs(%843 : tensor<1x128x512x512xf16>) -> tensor<1x128x512x512xf16>
    %845 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%794, %844 : tensor<1x128x512x512xf16>, tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %846 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%845 : tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x128x512x512xf16>
    %collapsed_378 = tensor.collapse_shape %846 [[0], [1], [2, 3]] : tensor<1x128x512x512xf16> into tensor<1x128x262144xf16>
    %expanded_379 = tensor.expand_shape %collapsed_378 [[0], [1, 2], [3]] : tensor<1x128x262144xf16> into tensor<1x32x4x262144xf16>
    %847 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_379 : tensor<1x32x4x262144xf16>) outs(%711 : tensor<1x32x4x262144xf32>) {
    ^bb0(%in: f16, %out: f32):
      %878 = arith.extf %in : f16 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x4x262144xf32>
    %848 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%847 : tensor<1x32x4x262144xf32>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f32, %out: f64):
      %878 = arith.extf %in : f32 to f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %849 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%848 : tensor<1x32x4x262144xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %850 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%849 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_151 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %851 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%848, %850 : tensor<1x32x4x262144xf64>, tensor<1x32x1x1xf64>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.subf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %852 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%851, %851 : tensor<1x32x4x262144xf64>, tensor<1x32x4x262144xf64>) outs(%713 : tensor<1x32x4x262144xf64>) {
    ^bb0(%in: f64, %in_385: f64, %out: f64):
      %878 = arith.mulf %in, %in_385 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x4x262144xf64>
    %853 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%852 : tensor<1x32x4x262144xf64>) outs(%12 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.addf %in, %out : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %854 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%853 : tensor<1x32x1x1xf64>) outs(%11 : tensor<1x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %878 = arith.divf %in, %cst_151 : f64
      linalg.yield %878 : f64
    } -> tensor<1x32x1x1xf64>
    %855 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%854 : tensor<1x32x1x1xf64>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %878 = arith.truncf %in : f64 to f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %856 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%847 : tensor<1x32x4x262144xf32>) outs(%21 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.addf %in, %out : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %857 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%856 : tensor<1x32x1x1xf32>) outs(%19 : tensor<1x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %878 = arith.divf %in, %cst_152 : f32
      linalg.yield %878 : f32
    } -> tensor<1x32x1x1xf32>
    %858 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%855 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %859 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%858 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %860 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%857 : tensor<1x32x1x1xf32>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %878 = arith.truncf %in : f32 to f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %861 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%860 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x32x1x1xf16>
    %862 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%859 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.truncf %cst_145 : f64 to f16
      %879 = arith.addf %in, %878 : f16
      linalg.yield %879 : f16
    } -> tensor<1x32x1x1xf16>
    %863 = linalg.generic {indexing_maps = [#map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%862 : tensor<1x32x1x1xf16>) outs(%24 : tensor<1x32x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.rsqrt %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x1x1xf16>
    %864 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_379, %861 : tensor<1x32x4x262144xf16>, tensor<1x32x1x1xf16>) outs(%730 : tensor<1x32x4x262144xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.subf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x4x262144xf16>
    %865 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%864, %863 : tensor<1x32x4x262144xf16>, tensor<1x32x1x1xf16>) outs(%730 : tensor<1x32x4x262144xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x32x4x262144xf16>
    %collapsed_380 = tensor.collapse_shape %865 [[0], [1, 2], [3]] : tensor<1x32x4x262144xf16> into tensor<1x128x262144xf16>
    %expanded_381 = tensor.expand_shape %collapsed_380 [[0], [1], [2, 3]] : tensor<1x128x262144xf16> into tensor<1x128x512x512xf16>
    %expanded_382 = tensor.expand_shape %cst_2 [[0, 1, 2, 3]] : tensor<128xf16> into tensor<1x128x1x1xf16>
    %expanded_383 = tensor.expand_shape %cst_1 [[0, 1, 2, 3]] : tensor<128xf16> into tensor<1x128x1x1xf16>
    %866 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_381, %expanded_383 : tensor<1x128x512x512xf16>, tensor<1x128x1x1xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %867 = linalg.generic {indexing_maps = [#map, #map4, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%866, %expanded_382 : tensor<1x128x512x512xf16>, tensor<1x128x1x1xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.addf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %868 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%867 : tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.negf %in : f16
      %879 = math.exp %878 : f16
      %880 = arith.addf %879, %cst_141 : f16
      %881 = arith.divf %cst_141, %880 : f16
      linalg.yield %881 : f16
    } -> tensor<1x128x512x512xf16>
    %869 = linalg.generic {indexing_maps = [#map, #map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%868, %867 : tensor<1x128x512x512xf16>, tensor<1x128x512x512xf16>) outs(%708 : tensor<1x128x512x512xf16>) {
    ^bb0(%in: f16, %in_385: f16, %out: f16):
      %878 = arith.mulf %in, %in_385 : f16
      linalg.yield %878 : f16
    } -> tensor<1x128x512x512xf16>
    %padded_384 = tensor.pad %869 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
      tensor.yield %cst_139 : f16
    } : tensor<1x128x512x512xf16> to tensor<1x128x514x514xf16>
    %870 = tensor.empty() : tensor<1x3x512x512xf16>
    %871 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst : tensor<3xf16>) outs(%870 : tensor<1x3x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1x3x512x512xf16>
    %872 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_384, %cst_0 : tensor<1x128x514x514xf16>, tensor<3x128x3x3xf16>) outs(%871 : tensor<1x3x512x512xf16>) -> tensor<1x3x512x512xf16>
    %873 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%872 : tensor<1x3x512x512xf16>) outs(%870 : tensor<1x3x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.divf %in, %cst_157 : f16
      linalg.yield %878 : f16
    } -> tensor<1x3x512x512xf16>
    %874 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%873 : tensor<1x3x512x512xf16>) outs(%870 : tensor<1x3x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.addf %in, %cst_158 : f16
      linalg.yield %878 : f16
    } -> tensor<1x3x512x512xf16>
    %875 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%874 : tensor<1x3x512x512xf16>) outs(%870 : tensor<1x3x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.cmpf ult, %in, %cst_139 : f16
      %879 = arith.select %878, %cst_139, %in : f16
      %880 = arith.cmpf ugt, %879, %cst_141 : f16
      %881 = arith.select %880, %cst_141, %879 : f16
      linalg.yield %881 : f16
    } -> tensor<1x3x512x512xf16>
    %876 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%875 : tensor<1x3x512x512xf16>) outs(%870 : tensor<1x3x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = arith.mulf %in, %cst_159 : f16
      linalg.yield %878 : f16
    } -> tensor<1x3x512x512xf16>
    %877 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%876 : tensor<1x3x512x512xf16>) outs(%870 : tensor<1x3x512x512xf16>) {
    ^bb0(%in: f16, %out: f16):
      %878 = math.roundeven %in : f16
      linalg.yield %878 : f16
    } -> tensor<1x3x512x512xf16>
    return %877 : tensor<1x3x512x512xf16>
  }
}