// -----// IR Dump Before LLVMGPUSelectLoweringStrategy (iree-llvmgpu-select-lowering-strategy) //----- //
module {
  func.func @main$async_dispatch_146_attention_2x20x1024x64xf16_generic() attributes {translation_info = #iree_codegen.translation_info<TransformDialectCodegen codegen_spec = @__attention_main, {llvm_func_attrs = {"amdgpu-waves-per-eu" = "2", "denormal-fp-math-f32" = "preserve-sign"}}>} {
    %cst = arith.constant 1.270000e+02 : f16
    %cst_0 = arith.constant -1.280000e+02 : f16
    %cst_1 = arith.constant 1.250000e-01 : f16
    %c0 = arith.constant 0 : index
    %0 = hal.interface.constant.load[0] : i32
    %1 = hal.interface.constant.load[1] : i32
    %2 = hal.interface.constant.load[2] : i32
    %3 = hal.interface.constant.load[3] : i32
    %4 = arith.index_castui %0 : i32 to index
    %5 = arith.index_castui %1 : i32 to index
    %6 = arith.index_castui %2 : i32 to index
    %7 = arith.index_castui %3 : i32 to index
    %8 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%4) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<2x20x1024x64xf16>>
    %9 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%5) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<2x20x1024x64xf16>>
    %10 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%6) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<2x20x1024x64xf16>>
    %11 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<20x64xf16>>
    %12 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<f32>>
    %13 = hal.interface.binding.subspan set(0) binding(3) type(storage_buffer) alignment(64) offset(%7) : !flow.dispatch.tensor<writeonly:tensor<2x1024x20x64xi8>>
    %14 = flow.dispatch.tensor.load %8, offsets = [0, 0, 0, 0], sizes = [2, 20, 1024, 64], strides = [1, 1, 1, 1] : !flow.dispatch.tensor<readonly:tensor<2x20x1024x64xf16>> -> tensor<2x20x1024x64xf16>
    %15 = flow.dispatch.tensor.load %9, offsets = [0, 0, 0, 0], sizes = [2, 20, 1024, 64], strides = [1, 1, 1, 1] : !flow.dispatch.tensor<readonly:tensor<2x20x1024x64xf16>> -> tensor<2x20x1024x64xf16>
    %16 = flow.dispatch.tensor.load %10, offsets = [0, 0, 0, 0], sizes = [2, 20, 1024, 64], strides = [1, 1, 1, 1] : !flow.dispatch.tensor<readonly:tensor<2x20x1024x64xf16>> -> tensor<2x20x1024x64xf16>
    %17 = flow.dispatch.tensor.load %11, offsets = [0, 0], sizes = [20, 64], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<20x64xf16>> -> tensor<20x64xf16>
    %18 = flow.dispatch.tensor.load %12, offsets = [], sizes = [], strides = [] : !flow.dispatch.tensor<readonly:tensor<f32>> -> tensor<f32>
    %19 = tensor.empty() : tensor<2x1024x20x64xi8>
    %20 = tensor.empty() : tensor<2x20x1024x64xf16>
    %21 = iree_linalg_ext.attention {indexing_maps = [affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1, d4, d3)>, affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1, d4, d5)>, affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1, d2, d5)>]} ins(%14, %15, %16, %cst_1 : tensor<2x20x1024x64xf16>, tensor<2x20x1024x64xf16>, tensor<2x20x1024x64xf16>, f16) outs(%20 : tensor<2x20x1024x64xf16>) -> tensor<2x20x1024x64xf16>
    %22 = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d1, d3)>, affine_map<(d0, d1, d2, d3) -> ()>, affine_map<(d0, d1, d2, d3) -> (d0, d2, d1, d3)>], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%21, %17, %18 : tensor<2x20x1024x64xf16>, tensor<20x64xf16>, tensor<f32>) outs(%19 : tensor<2x1024x20x64xi8>) {
    ^bb0(%in: f16, %in_2: f16, %in_3: f32, %out: i8):
      %23 = arith.mulf %in, %in_2 : f16
      %24 = arith.truncf %in_3 : f32 to f16
      %25 = arith.divf %23, %24 : f16
      %26 = math.roundeven %25 : f16
      %27 = arith.cmpf ult, %26, %cst_0 : f16
      %28 = arith.select %27, %cst_0, %26 : f16
      %29 = arith.cmpf ugt, %28, %cst : f16
      %30 = arith.select %29, %cst, %28 : f16
      %31 = arith.fptosi %30 : f16 to i8
      linalg.yield %31 : i8
    } -> tensor<2x1024x20x64xi8>
    flow.dispatch.tensor.store %22, %13, offsets = [0, 0, 0, 0], sizes = [2, 1024, 20, 64], strides = [1, 1, 1, 1] : tensor<2x1024x20x64xi8> -> !flow.dispatch.tensor<writeonly:tensor<2x1024x20x64xi8>>
    return
  }
}

// -----// IR Dump After LLVMGPUSelectLoweringStrategy (iree-llvmgpu-select-lowering-strategy) //----- //
module {
  func.func @main$async_dispatch_146_attention_2x20x1024x64xf16_generic() attributes {translation_info = #iree_codegen.translation_info<TransformDialectCodegen codegen_spec = @__attention_main, {llvm_func_attrs = {"amdgpu-waves-per-eu" = "2", "denormal-fp-math-f32" = "preserve-sign"}}>} {
    %cst = arith.constant 1.270000e+02 : f16
    %cst_0 = arith.constant -1.280000e+02 : f16
    %cst_1 = arith.constant 1.250000e-01 : f16
    %c0 = arith.constant 0 : index
    %0 = hal.interface.constant.load[0] : i32
    %1 = hal.interface.constant.load[1] : i32
    %2 = hal.interface.constant.load[2] : i32
    %3 = hal.interface.constant.load[3] : i32
    %4 = arith.index_castui %0 : i32 to index
    %5 = arith.index_castui %1 : i32 to index
    %6 = arith.index_castui %2 : i32 to index
    %7 = arith.index_castui %3 : i32 to index
    %8 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%4) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<2x20x1024x64xf16>>
    %9 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%5) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<2x20x1024x64xf16>>
    %10 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%6) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<2x20x1024x64xf16>>
    %11 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<20x64xf16>>
    %12 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<f32>>
    %13 = hal.interface.binding.subspan set(0) binding(3) type(storage_buffer) alignment(64) offset(%7) : !flow.dispatch.tensor<writeonly:tensor<2x1024x20x64xi8>>
    %14 = flow.dispatch.tensor.load %8, offsets = [0, 0, 0, 0], sizes = [2, 20, 1024, 64], strides = [1, 1, 1, 1] : !flow.dispatch.tensor<readonly:tensor<2x20x1024x64xf16>> -> tensor<2x20x1024x64xf16>
    %15 = flow.dispatch.tensor.load %9, offsets = [0, 0, 0, 0], sizes = [2, 20, 1024, 64], strides = [1, 1, 1, 1] : !flow.dispatch.tensor<readonly:tensor<2x20x1024x64xf16>> -> tensor<2x20x1024x64xf16>
    %16 = flow.dispatch.tensor.load %10, offsets = [0, 0, 0, 0], sizes = [2, 20, 1024, 64], strides = [1, 1, 1, 1] : !flow.dispatch.tensor<readonly:tensor<2x20x1024x64xf16>> -> tensor<2x20x1024x64xf16>
    %17 = flow.dispatch.tensor.load %11, offsets = [0, 0], sizes = [20, 64], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<20x64xf16>> -> tensor<20x64xf16>
    %18 = flow.dispatch.tensor.load %12, offsets = [], sizes = [], strides = [] : !flow.dispatch.tensor<readonly:tensor<f32>> -> tensor<f32>
    %19 = tensor.empty() : tensor<2x1024x20x64xi8>
    %20 = tensor.empty() : tensor<2x20x1024x64xf16>
    %21 = iree_linalg_ext.attention {indexing_maps = [affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1, d4, d3)>, affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1, d4, d5)>, affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1, d2, d5)>]} ins(%14, %15, %16, %cst_1 : tensor<2x20x1024x64xf16>, tensor<2x20x1024x64xf16>, tensor<2x20x1024x64xf16>, f16) outs(%20 : tensor<2x20x1024x64xf16>) -> tensor<2x20x1024x64xf16>
    %22 = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d1, d3)>, affine_map<(d0, d1, d2, d3) -> ()>, affine_map<(d0, d1, d2, d3) -> (d0, d2, d1, d3)>], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%21, %17, %18 : tensor<2x20x1024x64xf16>, tensor<20x64xf16>, tensor<f32>) outs(%19 : tensor<2x1024x20x64xi8>) {
    ^bb0(%in: f16, %in_2: f16, %in_3: f32, %out: i8):
      %23 = arith.mulf %in, %in_2 : f16
      %24 = arith.truncf %in_3 : f32 to f16
      %25 = arith.divf %23, %24 : f16
      %26 = math.roundeven %25 : f16
      %27 = arith.cmpf ult, %26, %cst_0 : f16
      %28 = arith.select %27, %cst_0, %26 : f16
      %29 = arith.cmpf ugt, %28, %cst : f16
      %30 = arith.select %29, %cst, %28 : f16
      %31 = arith.fptosi %30 : f16 to i8
      linalg.yield %31 : i8
    } -> tensor<2x1024x20x64xi8>
    flow.dispatch.tensor.store %22, %13, offsets = [0, 0, 0, 0], sizes = [2, 1024, 20, 64], strides = [1, 1, 1, 1] : tensor<2x1024x20x64xi8> -> !flow.dispatch.tensor<writeonly:tensor<2x1024x20x64xi8>>
    return
  }
}

// -----// IR Dump Before LLVMGPULowerExecutableTarget (iree-llvmgpu-lower-executable-target) //----- //
func.func @main$async_dispatch_146_attention_2x20x1024x64xf16_generic() attributes {translation_info = #iree_codegen.translation_info<None workgroup_size = [64, 4, 1] subgroup_size = 64, {llvm_func_attrs = {"amdgpu-waves-per-eu" = "2", "denormal-fp-math-f32" = "preserve-sign"}}>} {
  %c992 = arith.constant 992 : index
  %cst = arith.constant 0.000000e+00 : f16
  %cst_0 = arith.constant dense<0.000000e+00> : vector<1x1x16xf32>
  %c0 = arith.constant 0 : index
  %cst_1 = arith.constant dense<1.270000e+02> : vector<2x1x16xf16>
  %cst_2 = arith.constant dense<-1.280000e+02> : vector<2x1x16xf16>
  %cst_3 = arith.constant dense<1.000000e+00> : vector<2x1x16xf32>
  %cst_4 = arith.constant dense<-6.550400e+04> : vector<4x1x4xf32>
  %cst_5 = arith.constant dense<6.550400e+04> : vector<4x1x4xf32>
  %cst_6 = arith.constant dense<1.802980e-01> : vector<1x4x8xf16>
  %cst_7 = arith.constant dense<0.000000e+00> : vector<1xf32>
  %cst_8 = arith.constant dense<-3.40282347E+38> : vector<1xf32>
  %cst_9 = arith.constant dense<0.000000e+00> : vector<2x1x16xf32>
  %c32 = arith.constant 32 : index
  %cst_10 = arith.constant dense<0.000000e+00> : vector<1x4x8xf16>
  %c32_i32 = arith.constant 32 : i32
  %c64_i32 = arith.constant 64 : i32
  %cst_11 = arith.constant dense<0.000000e+00> : vector<4x1x4xf32>
  %cst_12 = arith.constant dense<0.000000e+00> : vector<2x4x4xf16>
  %cst_13 = arith.constant dense<0.000000e+00> : vector<2x1x16xf16>
  %cst_14 = arith.constant dense<0.000000e+00> : vector<16xf32>
  %thread_id_x = gpu.thread_id  x
  %thread_id_y = gpu.thread_id  y
  %thread_id_z = gpu.thread_id  z
  %alloc = memref.alloc() : memref<1x1x32x68xf16, #gpu.address_space<workgroup>>
  %subview = memref.subview %alloc[0, 0, 0, 0] [1, 1, 32, 64] [1, 1, 1, 1] : memref<1x1x32x68xf16, #gpu.address_space<workgroup>> to memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>
  %alloc_15 = memref.alloc() : memref<1x1x32x68xf16, #gpu.address_space<workgroup>>
  %subview_16 = memref.subview %alloc_15[0, 0, 0, 0] [1, 1, 32, 64] [1, 1, 1, 1] : memref<1x1x32x68xf16, #gpu.address_space<workgroup>> to memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>
  %0 = hal.interface.constant.load[0] : i32
  %1 = hal.interface.constant.load[1] : i32
  %2 = hal.interface.constant.load[2] : i32
  %3 = hal.interface.constant.load[3] : i32
  %4 = arith.index_castui %0 : i32 to index
  %5 = arith.index_castui %1 : i32 to index
  %6 = arith.index_castui %2 : i32 to index
  %7 = arith.index_castui %3 : i32 to index
  %8 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%4) flags(ReadOnly) : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>
  memref.assume_alignment %8, 1 : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>
  %9 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%5) flags(ReadOnly) : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>
  memref.assume_alignment %9, 1 : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>
  %10 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%6) flags(ReadOnly) : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>
  memref.assume_alignment %10, 1 : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>
  %11 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>
  memref.assume_alignment %11, 64 : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>
  %12 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : memref<f32, #hal.descriptor_type<storage_buffer>>
  memref.assume_alignment %12, 64 : memref<f32, #hal.descriptor_type<storage_buffer>>
  %13 = hal.interface.binding.subspan set(0) binding(3) type(storage_buffer) alignment(64) offset(%7) : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>
  memref.assume_alignment %13, 1 : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_id_z = hal.interface.workgroup.id[2] : index
  %14 = affine.apply affine_map<()[s0, s1, s2] -> (s0 * 128 + s2 * 32 + (s1 floordiv 64) * 32)>()[%workgroup_id_z, %thread_id_x, %thread_id_y]
  %15 = affine.apply affine_map<()[s0] -> (s0 mod 32)>()[%thread_id_x]
  %16 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 8)>()[%thread_id_x]
  %17 = arith.addi %14, %15 : index
  %18 = vector.load %8[%workgroup_id_x, %workgroup_id_y, %17, %16] : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<8xf16>
  %19 = vector.insert_strided_slice %18, %cst_10 {offsets = [0, 0, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
  %20 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 8 + 16)>()[%thread_id_x]
  %21 = vector.load %8[%workgroup_id_x, %workgroup_id_y, %17, %20] : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<8xf16>
  %22 = vector.insert_strided_slice %21, %19 {offsets = [0, 1, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
  %23 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 8 + 32)>()[%thread_id_x]
  %24 = vector.load %8[%workgroup_id_x, %workgroup_id_y, %17, %23] : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<8xf16>
  %25 = vector.insert_strided_slice %24, %22 {offsets = [0, 2, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
  %26 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 8 + 48)>()[%thread_id_x]
  %27 = vector.load %8[%workgroup_id_x, %workgroup_id_y, %17, %26] : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<8xf16>
  %28 = vector.insert_strided_slice %27, %25 {offsets = [0, 3, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
  %29 = arith.mulf %28, %cst_6 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHX,  LANEX], [1, 32]>, <[ BATCHY,  LANEY,  VECTORX], [4, 2, 8]>>, #iree_vector_ext.layout<<[ BATCHX,  LANEX], [1, 32]>, <[ BATCHY,  LANEY,  VECTORX], [4, 2, 8]>>], [#iree_vector_ext.layout<<[ BATCHX,  LANEX], [1, 32]>, <[ BATCHY,  LANEY,  VECTORX], [4, 2, 8]>>]], __vector_layout_test_anchor_result_0 = #iree_vector_ext.layout<<[ BATCHX,  LANEX], [1, 32]>, <[ BATCHY,  LANEY,  VECTORX], [4, 2, 8]>>} : vector<1x4x8xf16>
  %30 = affine.apply affine_map<()[s0, s1, s2] -> (s1 * 8 + s2 * 32 + s0 floordiv 8)>()[%thread_id_x, %thread_id_y, %thread_id_z]
  %31 = affine.apply affine_map<()[s0] -> (s0 * 8 - (s0 floordiv 8) * 64)>()[%thread_id_x]
  %32 = vector.transfer_read %9[%workgroup_id_x, %workgroup_id_y, %30, %31], %cst {in_bounds = [true, true, true, true]} : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<1x1x1x8xf16>
  %33 = vector.transfer_read %10[%workgroup_id_x, %workgroup_id_y, %30, %31], %cst {in_bounds = [true, true, true, true]} : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<1x1x1x8xf16>
  vector.transfer_write %32, %subview_16[%c0, %c0, %30, %31] {in_bounds = [true, true, true, true]} : vector<1x1x1x8xf16>, memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>
  vector.transfer_write %33, %subview[%c0, %c0, %30, %31] {in_bounds = [true, true, true, true]} : vector<1x1x1x8xf16>, memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>
  %34:3 = scf.for %arg0 = %c0 to %c992 step %c32 iter_args(%arg1 = %cst_8, %arg2 = %cst_7, %arg3 = %cst_9) -> (vector<1xf32>, vector<1xf32>, vector<2x1x16xf32>) {
    %533 = arith.addi %arg0, %c32 : index
    %534 = affine.apply affine_map<()[s0, s1, s2, s3] -> (s0 + s2 * 8 + s3 * 32 + s1 floordiv 8)>()[%533, %thread_id_x, %thread_id_y, %thread_id_z]
    %535 = vector.transfer_read %9[%workgroup_id_x, %workgroup_id_y, %534, %31], %cst {in_bounds = [true, true, true, true]} : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<1x1x1x8xf16>
    %536 = vector.transfer_read %10[%workgroup_id_x, %workgroup_id_y, %534, %31], %cst {in_bounds = [true, true, true, true]} : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<1x1x1x8xf16>
    gpu.barrier
    %537 = vector.load %subview_16[%c0, %c0, %15, %16] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<8xf16>
    %538 = vector.insert_strided_slice %537, %cst_10 {offsets = [0, 0, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
    %539 = vector.load %subview_16[%c0, %c0, %15, %20] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<8xf16>
    %540 = vector.insert_strided_slice %539, %538 {offsets = [0, 1, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
    %541 = vector.load %subview_16[%c0, %c0, %15, %23] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<8xf16>
    %542 = vector.insert_strided_slice %541, %540 {offsets = [0, 2, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
    %543 = vector.load %subview_16[%c0, %c0, %15, %26] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<8xf16>
    %544 = vector.insert_strided_slice %543, %542 {offsets = [0, 3, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
    %545 = vector.extract_strided_slice %544 {offsets = [0, 0, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %546 = vector.extract_strided_slice %544 {offsets = [0, 0, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %547 = vector.extract_strided_slice %544 {offsets = [0, 1, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %548 = vector.extract_strided_slice %544 {offsets = [0, 1, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %549 = vector.extract_strided_slice %544 {offsets = [0, 2, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %550 = vector.extract_strided_slice %544 {offsets = [0, 2, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %551 = vector.extract_strided_slice %544 {offsets = [0, 3, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %552 = vector.extract_strided_slice %544 {offsets = [0, 3, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %553 = vector.extract_strided_slice %29 {offsets = [0, 0, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %554 = vector.extract_strided_slice %29 {offsets = [0, 0, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %555 = vector.extract_strided_slice %29 {offsets = [0, 1, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %556 = vector.extract_strided_slice %29 {offsets = [0, 1, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %557 = vector.extract_strided_slice %29 {offsets = [0, 2, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %558 = vector.extract_strided_slice %29 {offsets = [0, 2, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %559 = vector.extract_strided_slice %29 {offsets = [0, 3, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %560 = vector.extract_strided_slice %29 {offsets = [0, 3, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %561 = vector.extract %545[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %562 = vector.extract %553[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %563 = amdgpu.mfma %561 * %562 + %cst_14 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %564 = vector.extract %546[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %565 = vector.extract %554[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %566 = amdgpu.mfma %564 * %565 + %563 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %567 = vector.extract %547[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %568 = vector.extract %555[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %569 = amdgpu.mfma %567 * %568 + %566 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %570 = vector.extract %548[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %571 = vector.extract %556[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %572 = amdgpu.mfma %570 * %571 + %569 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %573 = vector.extract %549[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %574 = vector.extract %557[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %575 = amdgpu.mfma %573 * %574 + %572 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %576 = vector.extract %550[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %577 = vector.extract %558[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %578 = amdgpu.mfma %576 * %577 + %575 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %579 = vector.extract %551[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %580 = vector.extract %559[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %581 = amdgpu.mfma %579 * %580 + %578 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %582 = vector.extract %552[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %583 = vector.extract %560[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %584 = amdgpu.mfma %582 * %583 + %581 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %585 = vector.broadcast %584 : vector<16xf32> to vector<1x1x16xf32>
    %586 = vector.extract %arg1[0] : f32 from vector<1xf32>
    %587 = vector.extract %584[0] : f32 from vector<16xf32>
    %588 = vector.insert %587, %cst_7 [0] : f32 into vector<1xf32>
    %589 = vector.extract %584[1] : f32 from vector<16xf32>
    %590 = vector.insert %589, %588 [0] : f32 into vector<1xf32>
    %591 = arith.maximumf %588, %590 : vector<1xf32>
    %592 = vector.extract %584[2] : f32 from vector<16xf32>
    %593 = vector.insert %592, %590 [0] : f32 into vector<1xf32>
    %594 = arith.maximumf %591, %593 : vector<1xf32>
    %595 = vector.extract %584[3] : f32 from vector<16xf32>
    %596 = vector.insert %595, %593 [0] : f32 into vector<1xf32>
    %597 = arith.maximumf %594, %596 : vector<1xf32>
    %598 = vector.extract %584[4] : f32 from vector<16xf32>
    %599 = vector.insert %598, %596 [0] : f32 into vector<1xf32>
    %600 = arith.maximumf %597, %599 : vector<1xf32>
    %601 = vector.extract %584[5] : f32 from vector<16xf32>
    %602 = vector.insert %601, %599 [0] : f32 into vector<1xf32>
    %603 = arith.maximumf %600, %602 : vector<1xf32>
    %604 = vector.extract %584[6] : f32 from vector<16xf32>
    %605 = vector.insert %604, %602 [0] : f32 into vector<1xf32>
    %606 = arith.maximumf %603, %605 : vector<1xf32>
    %607 = vector.extract %584[7] : f32 from vector<16xf32>
    %608 = vector.insert %607, %605 [0] : f32 into vector<1xf32>
    %609 = arith.maximumf %606, %608 : vector<1xf32>
    %610 = vector.extract %584[8] : f32 from vector<16xf32>
    %611 = vector.insert %610, %608 [0] : f32 into vector<1xf32>
    %612 = arith.maximumf %609, %611 : vector<1xf32>
    %613 = vector.extract %584[9] : f32 from vector<16xf32>
    %614 = vector.insert %613, %611 [0] : f32 into vector<1xf32>
    %615 = arith.maximumf %612, %614 : vector<1xf32>
    %616 = vector.extract %584[10] : f32 from vector<16xf32>
    %617 = vector.insert %616, %614 [0] : f32 into vector<1xf32>
    %618 = arith.maximumf %615, %617 : vector<1xf32>
    %619 = vector.extract %584[11] : f32 from vector<16xf32>
    %620 = vector.insert %619, %617 [0] : f32 into vector<1xf32>
    %621 = arith.maximumf %618, %620 : vector<1xf32>
    %622 = vector.extract %584[12] : f32 from vector<16xf32>
    %623 = vector.insert %622, %620 [0] : f32 into vector<1xf32>
    %624 = arith.maximumf %621, %623 : vector<1xf32>
    %625 = vector.extract %584[13] : f32 from vector<16xf32>
    %626 = vector.insert %625, %623 [0] : f32 into vector<1xf32>
    %627 = arith.maximumf %624, %626 : vector<1xf32>
    %628 = vector.extract %584[14] : f32 from vector<16xf32>
    %629 = vector.insert %628, %626 [0] : f32 into vector<1xf32>
    %630 = arith.maximumf %627, %629 : vector<1xf32>
    %631 = vector.extract %584[15] : f32 from vector<16xf32>
    %632 = vector.insert %631, %629 [0] : f32 into vector<1xf32>
    %633 = arith.maximumf %630, %632 : vector<1xf32>
    %634 = vector.bitcast %633 : vector<1xf32> to vector<1xi32>
    %635 = vector.extract %634[0] : i32 from vector<1xi32>
    %shuffleResult_19, %valid_20 = gpu.shuffle  xor %635, %c32_i32, %c64_i32 : i32
    %636 = vector.broadcast %shuffleResult_19 : i32 to vector<1xi32>
    %637 = vector.bitcast %636 : vector<1xi32> to vector<1xf32>
    %638 = arith.maximumf %637, %633 : vector<1xf32>
    %639 = vector.extract %638[0] : f32 from vector<1xf32>
    %640 = arith.maximumf %639, %586 : f32
    %641 = vector.insert %640, %cst_7 [0] : f32 into vector<1xf32>
    %642 = vector.insert %640, %cst_0 [0, 0, 0] : f32 into vector<1x1x16xf32>
    %643 = vector.insert %640, %642 [0, 0, 1] : f32 into vector<1x1x16xf32>
    %644 = vector.insert %640, %643 [0, 0, 2] : f32 into vector<1x1x16xf32>
    %645 = vector.insert %640, %644 [0, 0, 3] : f32 into vector<1x1x16xf32>
    %646 = vector.insert %640, %645 [0, 0, 4] : f32 into vector<1x1x16xf32>
    %647 = vector.insert %640, %646 [0, 0, 5] : f32 into vector<1x1x16xf32>
    %648 = vector.insert %640, %647 [0, 0, 6] : f32 into vector<1x1x16xf32>
    %649 = vector.insert %640, %648 [0, 0, 7] : f32 into vector<1x1x16xf32>
    %650 = vector.insert %640, %649 [0, 0, 8] : f32 into vector<1x1x16xf32>
    %651 = vector.insert %640, %650 [0, 0, 9] : f32 into vector<1x1x16xf32>
    %652 = vector.insert %640, %651 [0, 0, 10] : f32 into vector<1x1x16xf32>
    %653 = vector.insert %640, %652 [0, 0, 11] : f32 into vector<1x1x16xf32>
    %654 = vector.insert %640, %653 [0, 0, 12] : f32 into vector<1x1x16xf32>
    %655 = vector.insert %640, %654 [0, 0, 13] : f32 into vector<1x1x16xf32>
    %656 = vector.insert %640, %655 [0, 0, 14] : f32 into vector<1x1x16xf32>
    %657 = vector.insert %640, %656 [0, 0, 15] : f32 into vector<1x1x16xf32>
    %658 = arith.subf %585, %657 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>]]} : vector<1x1x16xf32>
    %659 = math.exp2 %658 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>]]} : vector<1x1x16xf32>
    %660 = arith.subf %arg1, %641 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>]]} : vector<1xf32>
    %661 = math.exp2 %660 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>]]} : vector<1xf32>
    %662 = arith.mulf %661, %arg2 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>]]} : vector<1xf32>
    %663 = vector.extract %662[0] : f32 from vector<1xf32>
    %664 = vector.extract %659[0, 0, 0] : f32 from vector<1x1x16xf32>
    %665 = vector.insert %664, %cst_7 [0] : f32 into vector<1xf32>
    %666 = vector.extract %659[0, 0, 1] : f32 from vector<1x1x16xf32>
    %667 = vector.insert %666, %665 [0] : f32 into vector<1xf32>
    %668 = arith.addf %665, %667 : vector<1xf32>
    %669 = vector.extract %659[0, 0, 2] : f32 from vector<1x1x16xf32>
    %670 = vector.insert %669, %667 [0] : f32 into vector<1xf32>
    %671 = arith.addf %668, %670 : vector<1xf32>
    %672 = vector.extract %659[0, 0, 3] : f32 from vector<1x1x16xf32>
    %673 = vector.insert %672, %670 [0] : f32 into vector<1xf32>
    %674 = arith.addf %671, %673 : vector<1xf32>
    %675 = vector.extract %659[0, 0, 4] : f32 from vector<1x1x16xf32>
    %676 = vector.insert %675, %673 [0] : f32 into vector<1xf32>
    %677 = arith.addf %674, %676 : vector<1xf32>
    %678 = vector.extract %659[0, 0, 5] : f32 from vector<1x1x16xf32>
    %679 = vector.insert %678, %676 [0] : f32 into vector<1xf32>
    %680 = arith.addf %677, %679 : vector<1xf32>
    %681 = vector.extract %659[0, 0, 6] : f32 from vector<1x1x16xf32>
    %682 = vector.insert %681, %679 [0] : f32 into vector<1xf32>
    %683 = arith.addf %680, %682 : vector<1xf32>
    %684 = vector.extract %659[0, 0, 7] : f32 from vector<1x1x16xf32>
    %685 = vector.insert %684, %682 [0] : f32 into vector<1xf32>
    %686 = arith.addf %683, %685 : vector<1xf32>
    %687 = vector.extract %659[0, 0, 8] : f32 from vector<1x1x16xf32>
    %688 = vector.insert %687, %685 [0] : f32 into vector<1xf32>
    %689 = arith.addf %686, %688 : vector<1xf32>
    %690 = vector.extract %659[0, 0, 9] : f32 from vector<1x1x16xf32>
    %691 = vector.insert %690, %688 [0] : f32 into vector<1xf32>
    %692 = arith.addf %689, %691 : vector<1xf32>
    %693 = vector.extract %659[0, 0, 10] : f32 from vector<1x1x16xf32>
    %694 = vector.insert %693, %691 [0] : f32 into vector<1xf32>
    %695 = arith.addf %692, %694 : vector<1xf32>
    %696 = vector.extract %659[0, 0, 11] : f32 from vector<1x1x16xf32>
    %697 = vector.insert %696, %694 [0] : f32 into vector<1xf32>
    %698 = arith.addf %695, %697 : vector<1xf32>
    %699 = vector.extract %659[0, 0, 12] : f32 from vector<1x1x16xf32>
    %700 = vector.insert %699, %697 [0] : f32 into vector<1xf32>
    %701 = arith.addf %698, %700 : vector<1xf32>
    %702 = vector.extract %659[0, 0, 13] : f32 from vector<1x1x16xf32>
    %703 = vector.insert %702, %700 [0] : f32 into vector<1xf32>
    %704 = arith.addf %701, %703 : vector<1xf32>
    %705 = vector.extract %659[0, 0, 14] : f32 from vector<1x1x16xf32>
    %706 = vector.insert %705, %703 [0] : f32 into vector<1xf32>
    %707 = arith.addf %704, %706 : vector<1xf32>
    %708 = vector.extract %659[0, 0, 15] : f32 from vector<1x1x16xf32>
    %709 = vector.insert %708, %706 [0] : f32 into vector<1xf32>
    %710 = arith.addf %707, %709 : vector<1xf32>
    %711 = vector.bitcast %710 : vector<1xf32> to vector<1xi32>
    %712 = vector.extract %711[0] : i32 from vector<1xi32>
    %shuffleResult_21, %valid_22 = gpu.shuffle  xor %712, %c32_i32, %c64_i32 : i32
    %713 = vector.broadcast %shuffleResult_21 : i32 to vector<1xi32>
    %714 = vector.bitcast %713 : vector<1xi32> to vector<1xf32>
    %715 = arith.addf %714, %710 : vector<1xf32>
    %716 = vector.extract %715[0] : f32 from vector<1xf32>
    %717 = arith.addf %716, %663 : f32
    %718 = vector.insert %717, %cst_7 [0] : f32 into vector<1xf32>
    %719 = vector.extract_strided_slice %659 {offsets = [0, 0, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x1x16xf32> to vector<1x1x4xf32>
    %720 = vector.insert_strided_slice %719, %cst_11 {offsets = [0, 0, 0], strides = [1, 1, 1]} : vector<1x1x4xf32> into vector<4x1x4xf32>
    %721 = vector.extract_strided_slice %659 {offsets = [0, 0, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x1x16xf32> to vector<1x1x4xf32>
    %722 = vector.insert_strided_slice %721, %720 {offsets = [1, 0, 0], strides = [1, 1, 1]} : vector<1x1x4xf32> into vector<4x1x4xf32>
    %723 = vector.extract_strided_slice %659 {offsets = [0, 0, 8], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x1x16xf32> to vector<1x1x4xf32>
    %724 = vector.insert_strided_slice %723, %722 {offsets = [2, 0, 0], strides = [1, 1, 1]} : vector<1x1x4xf32> into vector<4x1x4xf32>
    %725 = vector.extract_strided_slice %659 {offsets = [0, 0, 12], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x1x16xf32> to vector<1x1x4xf32>
    %726 = vector.insert_strided_slice %725, %724 {offsets = [3, 0, 0], strides = [1, 1, 1]} : vector<1x1x4xf32> into vector<4x1x4xf32>
    %727 = arith.cmpf ogt, %726, %cst_5 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xf32>
    %728 = arith.cmpf olt, %726, %cst_4 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xf32>
    %729 = arith.select %727, %cst_5, %726 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xi1>, vector<4x1x4xf32>
    %730 = arith.select %728, %cst_4, %729 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xi1>, vector<4x1x4xf32>
    %731 = arith.truncf %730 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xf32> to vector<4x1x4xf16>
    %732 = vector.extract %661[0] : f32 from vector<1xf32>
    %733 = vector.insert %732, %cst_9 [0, 0, 0] : f32 into vector<2x1x16xf32>
    %734 = vector.insert %732, %733 [0, 0, 1] : f32 into vector<2x1x16xf32>
    %735 = vector.insert %732, %734 [0, 0, 2] : f32 into vector<2x1x16xf32>
    %736 = vector.insert %732, %735 [0, 0, 3] : f32 into vector<2x1x16xf32>
    %737 = vector.insert %732, %736 [0, 0, 4] : f32 into vector<2x1x16xf32>
    %738 = vector.insert %732, %737 [0, 0, 5] : f32 into vector<2x1x16xf32>
    %739 = vector.insert %732, %738 [0, 0, 6] : f32 into vector<2x1x16xf32>
    %740 = vector.insert %732, %739 [0, 0, 7] : f32 into vector<2x1x16xf32>
    %741 = vector.insert %732, %740 [0, 0, 8] : f32 into vector<2x1x16xf32>
    %742 = vector.insert %732, %741 [0, 0, 9] : f32 into vector<2x1x16xf32>
    %743 = vector.insert %732, %742 [0, 0, 10] : f32 into vector<2x1x16xf32>
    %744 = vector.insert %732, %743 [0, 0, 11] : f32 into vector<2x1x16xf32>
    %745 = vector.insert %732, %744 [0, 0, 12] : f32 into vector<2x1x16xf32>
    %746 = vector.insert %732, %745 [0, 0, 13] : f32 into vector<2x1x16xf32>
    %747 = vector.insert %732, %746 [0, 0, 14] : f32 into vector<2x1x16xf32>
    %748 = vector.insert %732, %747 [0, 0, 15] : f32 into vector<2x1x16xf32>
    %749 = vector.insert %732, %748 [1, 0, 0] : f32 into vector<2x1x16xf32>
    %750 = vector.insert %732, %749 [1, 0, 1] : f32 into vector<2x1x16xf32>
    %751 = vector.insert %732, %750 [1, 0, 2] : f32 into vector<2x1x16xf32>
    %752 = vector.insert %732, %751 [1, 0, 3] : f32 into vector<2x1x16xf32>
    %753 = vector.insert %732, %752 [1, 0, 4] : f32 into vector<2x1x16xf32>
    %754 = vector.insert %732, %753 [1, 0, 5] : f32 into vector<2x1x16xf32>
    %755 = vector.insert %732, %754 [1, 0, 6] : f32 into vector<2x1x16xf32>
    %756 = vector.insert %732, %755 [1, 0, 7] : f32 into vector<2x1x16xf32>
    %757 = vector.insert %732, %756 [1, 0, 8] : f32 into vector<2x1x16xf32>
    %758 = vector.insert %732, %757 [1, 0, 9] : f32 into vector<2x1x16xf32>
    %759 = vector.insert %732, %758 [1, 0, 10] : f32 into vector<2x1x16xf32>
    %760 = vector.insert %732, %759 [1, 0, 11] : f32 into vector<2x1x16xf32>
    %761 = vector.insert %732, %760 [1, 0, 12] : f32 into vector<2x1x16xf32>
    %762 = vector.insert %732, %761 [1, 0, 13] : f32 into vector<2x1x16xf32>
    %763 = vector.insert %732, %762 [1, 0, 14] : f32 into vector<2x1x16xf32>
    %764 = vector.insert %732, %763 [1, 0, 15] : f32 into vector<2x1x16xf32>
    %765 = arith.mulf %764, %arg3 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf32>
    %766 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4)>()[%thread_id_x]
    %767 = vector.load %subview[%c0, %c0, %766, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %768 = vector.insert_strided_slice %767, %cst_12 {offsets = [0, 0, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %769 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 1)>()[%thread_id_x]
    %770 = vector.load %subview[%c0, %c0, %769, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %771 = vector.insert_strided_slice %770, %768 {offsets = [0, 0, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %772 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 2)>()[%thread_id_x]
    %773 = vector.load %subview[%c0, %c0, %772, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %774 = vector.insert_strided_slice %773, %771 {offsets = [0, 0, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %775 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 3)>()[%thread_id_x]
    %776 = vector.load %subview[%c0, %c0, %775, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %777 = vector.insert_strided_slice %776, %774 {offsets = [0, 0, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %778 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 + 2) floordiv 8)>()[%thread_id_x]
    %779 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 - (((s0 mod 64) floordiv 32 + 2) floordiv 8) * 8 + 2) floordiv 8)>()[%thread_id_x]
    %780 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - (((s0 mod 64) floordiv 32 + 2) floordiv 8) * 32 + 8)>()[%thread_id_x]
    %781 = vector.load %subview[%778, %779, %780, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %782 = vector.insert_strided_slice %781, %777 {offsets = [0, 1, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %783 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 9) floordiv 32)>()[%thread_id_x]
    %784 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 9) floordiv 32) * 32 + 9) floordiv 32)>()[%thread_id_x]
    %785 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 9) floordiv 32) * 32 + 9)>()[%thread_id_x]
    %786 = vector.load %subview[%783, %784, %785, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %787 = vector.insert_strided_slice %786, %782 {offsets = [0, 1, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %788 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 + 5) floordiv 16)>()[%thread_id_x]
    %789 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 - ((((s0 mod 64) floordiv 32) * 2 + 5) floordiv 16) * 16 + 5) floordiv 16)>()[%thread_id_x]
    %790 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 2 + 5) floordiv 16) * 32 + 10)>()[%thread_id_x]
    %791 = vector.load %subview[%788, %789, %790, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %792 = vector.insert_strided_slice %791, %787 {offsets = [0, 1, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %793 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 11) floordiv 32)>()[%thread_id_x]
    %794 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 11) floordiv 32) * 32 + 11) floordiv 32)>()[%thread_id_x]
    %795 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 11) floordiv 32) * 32 + 11)>()[%thread_id_x]
    %796 = vector.load %subview[%793, %794, %795, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %797 = vector.insert_strided_slice %796, %792 {offsets = [0, 1, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %798 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 + 4) floordiv 8)>()[%thread_id_x]
    %799 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 - (((s0 mod 64) floordiv 32 + 4) floordiv 8) * 8 + 4) floordiv 8)>()[%thread_id_x]
    %800 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - (((s0 mod 64) floordiv 32 + 4) floordiv 8) * 32 + 16)>()[%thread_id_x]
    %801 = vector.load %subview[%798, %799, %800, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %802 = vector.insert_strided_slice %801, %797 {offsets = [0, 2, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %803 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 17) floordiv 32)>()[%thread_id_x]
    %804 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 17) floordiv 32) * 32 + 17) floordiv 32)>()[%thread_id_x]
    %805 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 17) floordiv 32) * 32 + 17)>()[%thread_id_x]
    %806 = vector.load %subview[%803, %804, %805, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %807 = vector.insert_strided_slice %806, %802 {offsets = [0, 2, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %808 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 + 9) floordiv 16)>()[%thread_id_x]
    %809 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 - ((((s0 mod 64) floordiv 32) * 2 + 9) floordiv 16) * 16 + 9) floordiv 16)>()[%thread_id_x]
    %810 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 2 + 9) floordiv 16) * 32 + 18)>()[%thread_id_x]
    %811 = vector.load %subview[%808, %809, %810, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %812 = vector.insert_strided_slice %811, %807 {offsets = [0, 2, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %813 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 19) floordiv 32)>()[%thread_id_x]
    %814 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 19) floordiv 32) * 32 + 19) floordiv 32)>()[%thread_id_x]
    %815 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 19) floordiv 32) * 32 + 19)>()[%thread_id_x]
    %816 = vector.load %subview[%813, %814, %815, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %817 = vector.insert_strided_slice %816, %812 {offsets = [0, 2, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %818 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 + 6) floordiv 8)>()[%thread_id_x]
    %819 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 - (((s0 mod 64) floordiv 32 + 6) floordiv 8) * 8 + 6) floordiv 8)>()[%thread_id_x]
    %820 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - (((s0 mod 64) floordiv 32 + 6) floordiv 8) * 32 + 24)>()[%thread_id_x]
    %821 = vector.load %subview[%818, %819, %820, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %822 = vector.insert_strided_slice %821, %817 {offsets = [0, 3, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %823 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 25) floordiv 32)>()[%thread_id_x]
    %824 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 25) floordiv 32) * 32 + 25) floordiv 32)>()[%thread_id_x]
    %825 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 25) floordiv 32) * 32 + 25)>()[%thread_id_x]
    %826 = vector.load %subview[%823, %824, %825, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %827 = vector.insert_strided_slice %826, %822 {offsets = [0, 3, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %828 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 + 13) floordiv 16)>()[%thread_id_x]
    %829 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 - ((((s0 mod 64) floordiv 32) * 2 + 13) floordiv 16) * 16 + 13) floordiv 16)>()[%thread_id_x]
    %830 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 2 + 13) floordiv 16) * 32 + 26)>()[%thread_id_x]
    %831 = vector.load %subview[%828, %829, %830, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %832 = vector.insert_strided_slice %831, %827 {offsets = [0, 3, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %833 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 27) floordiv 32)>()[%thread_id_x]
    %834 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 27) floordiv 32) * 32 + 27) floordiv 32)>()[%thread_id_x]
    %835 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 27) floordiv 32) * 32 + 27)>()[%thread_id_x]
    %836 = vector.load %subview[%833, %834, %835, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %837 = vector.insert_strided_slice %836, %832 {offsets = [0, 3, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %838 = affine.apply affine_map<()[s0] -> (s0 mod 32 + 32)>()[%thread_id_x]
    %839 = vector.load %subview[%c0, %c0, %766, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %840 = vector.insert_strided_slice %839, %837 {offsets = [1, 0, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %841 = vector.load %subview[%c0, %c0, %769, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %842 = vector.insert_strided_slice %841, %840 {offsets = [1, 0, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %843 = vector.load %subview[%c0, %c0, %772, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %844 = vector.insert_strided_slice %843, %842 {offsets = [1, 0, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %845 = vector.load %subview[%c0, %c0, %775, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %846 = vector.insert_strided_slice %845, %844 {offsets = [1, 0, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %847 = vector.load %subview[%778, %779, %780, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %848 = vector.insert_strided_slice %847, %846 {offsets = [1, 1, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %849 = vector.load %subview[%783, %784, %785, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %850 = vector.insert_strided_slice %849, %848 {offsets = [1, 1, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %851 = vector.load %subview[%788, %789, %790, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %852 = vector.insert_strided_slice %851, %850 {offsets = [1, 1, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %853 = vector.load %subview[%793, %794, %795, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %854 = vector.insert_strided_slice %853, %852 {offsets = [1, 1, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %855 = vector.load %subview[%798, %799, %800, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %856 = vector.insert_strided_slice %855, %854 {offsets = [1, 2, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %857 = vector.load %subview[%803, %804, %805, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %858 = vector.insert_strided_slice %857, %856 {offsets = [1, 2, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %859 = vector.load %subview[%808, %809, %810, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %860 = vector.insert_strided_slice %859, %858 {offsets = [1, 2, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %861 = vector.load %subview[%813, %814, %815, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %862 = vector.insert_strided_slice %861, %860 {offsets = [1, 2, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %863 = vector.load %subview[%818, %819, %820, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %864 = vector.insert_strided_slice %863, %862 {offsets = [1, 3, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %865 = vector.load %subview[%823, %824, %825, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %866 = vector.insert_strided_slice %865, %864 {offsets = [1, 3, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %867 = vector.load %subview[%828, %829, %830, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %868 = vector.insert_strided_slice %867, %866 {offsets = [1, 3, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %869 = vector.load %subview[%833, %834, %835, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %870 = vector.insert_strided_slice %869, %868 {offsets = [1, 3, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %871 = vector.extract %765[0, 0] : vector<16xf32> from vector<2x1x16xf32>
    %872 = vector.extract %870[0, 0] : vector<4xf16> from vector<2x4x4xf16>
    %873 = vector.extract %731[0, 0] : vector<4xf16> from vector<4x1x4xf16>
    %874 = amdgpu.mfma %872 * %873 + %871 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %875 = vector.extract %870[0, 1] : vector<4xf16> from vector<2x4x4xf16>
    %876 = vector.extract %731[1, 0] : vector<4xf16> from vector<4x1x4xf16>
    %877 = amdgpu.mfma %875 * %876 + %874 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %878 = vector.extract %870[0, 2] : vector<4xf16> from vector<2x4x4xf16>
    %879 = vector.extract %731[2, 0] : vector<4xf16> from vector<4x1x4xf16>
    %880 = amdgpu.mfma %878 * %879 + %877 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %881 = vector.extract %870[0, 3] : vector<4xf16> from vector<2x4x4xf16>
    %882 = vector.extract %731[3, 0] : vector<4xf16> from vector<4x1x4xf16>
    %883 = amdgpu.mfma %881 * %882 + %880 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %884 = vector.insert %883, %cst_9 [0, 0] : vector<16xf32> into vector<2x1x16xf32>
    %885 = vector.extract %765[1, 0] : vector<16xf32> from vector<2x1x16xf32>
    %886 = vector.extract %870[1, 0] : vector<4xf16> from vector<2x4x4xf16>
    %887 = amdgpu.mfma %886 * %873 + %885 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %888 = vector.extract %870[1, 1] : vector<4xf16> from vector<2x4x4xf16>
    %889 = amdgpu.mfma %888 * %876 + %887 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %890 = vector.extract %870[1, 2] : vector<4xf16> from vector<2x4x4xf16>
    %891 = amdgpu.mfma %890 * %879 + %889 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %892 = vector.extract %870[1, 3] : vector<4xf16> from vector<2x4x4xf16>
    %893 = amdgpu.mfma %892 * %882 + %891 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %894 = vector.insert %893, %884 [1, 0] : vector<16xf32> into vector<2x1x16xf32>
    gpu.barrier
    vector.transfer_write %535, %subview_16[%c0, %c0, %30, %31] {in_bounds = [true, true, true, true]} : vector<1x1x1x8xf16>, memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>
    vector.transfer_write %536, %subview[%c0, %c0, %30, %31] {in_bounds = [true, true, true, true]} : vector<1x1x1x8xf16>, memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>
    scf.yield %641, %718, %894 : vector<1xf32>, vector<1xf32>, vector<2x1x16xf32>
  }
  gpu.barrier
  %35 = vector.load %subview_16[%c0, %c0, %15, %16] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<8xf16>
  %36 = vector.insert_strided_slice %35, %cst_10 {offsets = [0, 0, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
  %37 = vector.load %subview_16[%c0, %c0, %15, %20] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<8xf16>
  %38 = vector.insert_strided_slice %37, %36 {offsets = [0, 1, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
  %39 = vector.load %subview_16[%c0, %c0, %15, %23] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<8xf16>
  %40 = vector.insert_strided_slice %39, %38 {offsets = [0, 2, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
  %41 = vector.load %subview_16[%c0, %c0, %15, %26] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<8xf16>
  %42 = vector.insert_strided_slice %41, %40 {offsets = [0, 3, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
  %43 = vector.extract_strided_slice %42 {offsets = [0, 0, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %44 = vector.extract_strided_slice %42 {offsets = [0, 0, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %45 = vector.extract_strided_slice %42 {offsets = [0, 1, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %46 = vector.extract_strided_slice %42 {offsets = [0, 1, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %47 = vector.extract_strided_slice %42 {offsets = [0, 2, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %48 = vector.extract_strided_slice %42 {offsets = [0, 2, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %49 = vector.extract_strided_slice %42 {offsets = [0, 3, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %50 = vector.extract_strided_slice %42 {offsets = [0, 3, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %51 = vector.extract_strided_slice %29 {offsets = [0, 0, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %52 = vector.extract_strided_slice %29 {offsets = [0, 0, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %53 = vector.extract_strided_slice %29 {offsets = [0, 1, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %54 = vector.extract_strided_slice %29 {offsets = [0, 1, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %55 = vector.extract_strided_slice %29 {offsets = [0, 2, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %56 = vector.extract_strided_slice %29 {offsets = [0, 2, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %57 = vector.extract_strided_slice %29 {offsets = [0, 3, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %58 = vector.extract_strided_slice %29 {offsets = [0, 3, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %59 = vector.extract %43[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %60 = vector.extract %51[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %61 = amdgpu.mfma %59 * %60 + %cst_14 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %62 = vector.extract %44[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %63 = vector.extract %52[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %64 = amdgpu.mfma %62 * %63 + %61 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %65 = vector.extract %45[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %66 = vector.extract %53[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %67 = amdgpu.mfma %65 * %66 + %64 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %68 = vector.extract %46[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %69 = vector.extract %54[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %70 = amdgpu.mfma %68 * %69 + %67 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %71 = vector.extract %47[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %72 = vector.extract %55[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %73 = amdgpu.mfma %71 * %72 + %70 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %74 = vector.extract %48[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %75 = vector.extract %56[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %76 = amdgpu.mfma %74 * %75 + %73 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %77 = vector.extract %49[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %78 = vector.extract %57[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %79 = amdgpu.mfma %77 * %78 + %76 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %80 = vector.extract %50[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %81 = vector.extract %58[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %82 = amdgpu.mfma %80 * %81 + %79 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %83 = vector.broadcast %82 : vector<16xf32> to vector<1x1x16xf32>
  %84 = vector.extract %34#0[0] : f32 from vector<1xf32>
  %85 = vector.extract %82[0] : f32 from vector<16xf32>
  %86 = vector.insert %85, %cst_7 [0] : f32 into vector<1xf32>
  %87 = vector.extract %82[1] : f32 from vector<16xf32>
  %88 = vector.insert %87, %86 [0] : f32 into vector<1xf32>
  %89 = arith.maximumf %86, %88 : vector<1xf32>
  %90 = vector.extract %82[2] : f32 from vector<16xf32>
  %91 = vector.insert %90, %88 [0] : f32 into vector<1xf32>
  %92 = arith.maximumf %89, %91 : vector<1xf32>
  %93 = vector.extract %82[3] : f32 from vector<16xf32>
  %94 = vector.insert %93, %91 [0] : f32 into vector<1xf32>
  %95 = arith.maximumf %92, %94 : vector<1xf32>
  %96 = vector.extract %82[4] : f32 from vector<16xf32>
  %97 = vector.insert %96, %94 [0] : f32 into vector<1xf32>
  %98 = arith.maximumf %95, %97 : vector<1xf32>
  %99 = vector.extract %82[5] : f32 from vector<16xf32>
  %100 = vector.insert %99, %97 [0] : f32 into vector<1xf32>
  %101 = arith.maximumf %98, %100 : vector<1xf32>
  %102 = vector.extract %82[6] : f32 from vector<16xf32>
  %103 = vector.insert %102, %100 [0] : f32 into vector<1xf32>
  %104 = arith.maximumf %101, %103 : vector<1xf32>
  %105 = vector.extract %82[7] : f32 from vector<16xf32>
  %106 = vector.insert %105, %103 [0] : f32 into vector<1xf32>
  %107 = arith.maximumf %104, %106 : vector<1xf32>
  %108 = vector.extract %82[8] : f32 from vector<16xf32>
  %109 = vector.insert %108, %106 [0] : f32 into vector<1xf32>
  %110 = arith.maximumf %107, %109 : vector<1xf32>
  %111 = vector.extract %82[9] : f32 from vector<16xf32>
  %112 = vector.insert %111, %109 [0] : f32 into vector<1xf32>
  %113 = arith.maximumf %110, %112 : vector<1xf32>
  %114 = vector.extract %82[10] : f32 from vector<16xf32>
  %115 = vector.insert %114, %112 [0] : f32 into vector<1xf32>
  %116 = arith.maximumf %113, %115 : vector<1xf32>
  %117 = vector.extract %82[11] : f32 from vector<16xf32>
  %118 = vector.insert %117, %115 [0] : f32 into vector<1xf32>
  %119 = arith.maximumf %116, %118 : vector<1xf32>
  %120 = vector.extract %82[12] : f32 from vector<16xf32>
  %121 = vector.insert %120, %118 [0] : f32 into vector<1xf32>
  %122 = arith.maximumf %119, %121 : vector<1xf32>
  %123 = vector.extract %82[13] : f32 from vector<16xf32>
  %124 = vector.insert %123, %121 [0] : f32 into vector<1xf32>
  %125 = arith.maximumf %122, %124 : vector<1xf32>
  %126 = vector.extract %82[14] : f32 from vector<16xf32>
  %127 = vector.insert %126, %124 [0] : f32 into vector<1xf32>
  %128 = arith.maximumf %125, %127 : vector<1xf32>
  %129 = vector.extract %82[15] : f32 from vector<16xf32>
  %130 = vector.insert %129, %127 [0] : f32 into vector<1xf32>
  %131 = arith.maximumf %128, %130 : vector<1xf32>
  %132 = vector.bitcast %131 : vector<1xf32> to vector<1xi32>
  %133 = vector.extract %132[0] : i32 from vector<1xi32>
  %shuffleResult, %valid = gpu.shuffle  xor %133, %c32_i32, %c64_i32 : i32
  %134 = vector.broadcast %shuffleResult : i32 to vector<1xi32>
  %135 = vector.bitcast %134 : vector<1xi32> to vector<1xf32>
  %136 = arith.maximumf %135, %131 : vector<1xf32>
  %137 = vector.extract %136[0] : f32 from vector<1xf32>
  %138 = arith.maximumf %137, %84 : f32
  %139 = vector.insert %138, %cst_7 [0] : f32 into vector<1xf32>
  %140 = vector.insert %138, %cst_0 [0, 0, 0] : f32 into vector<1x1x16xf32>
  %141 = vector.insert %138, %140 [0, 0, 1] : f32 into vector<1x1x16xf32>
  %142 = vector.insert %138, %141 [0, 0, 2] : f32 into vector<1x1x16xf32>
  %143 = vector.insert %138, %142 [0, 0, 3] : f32 into vector<1x1x16xf32>
  %144 = vector.insert %138, %143 [0, 0, 4] : f32 into vector<1x1x16xf32>
  %145 = vector.insert %138, %144 [0, 0, 5] : f32 into vector<1x1x16xf32>
  %146 = vector.insert %138, %145 [0, 0, 6] : f32 into vector<1x1x16xf32>
  %147 = vector.insert %138, %146 [0, 0, 7] : f32 into vector<1x1x16xf32>
  %148 = vector.insert %138, %147 [0, 0, 8] : f32 into vector<1x1x16xf32>
  %149 = vector.insert %138, %148 [0, 0, 9] : f32 into vector<1x1x16xf32>
  %150 = vector.insert %138, %149 [0, 0, 10] : f32 into vector<1x1x16xf32>
  %151 = vector.insert %138, %150 [0, 0, 11] : f32 into vector<1x1x16xf32>
  %152 = vector.insert %138, %151 [0, 0, 12] : f32 into vector<1x1x16xf32>
  %153 = vector.insert %138, %152 [0, 0, 13] : f32 into vector<1x1x16xf32>
  %154 = vector.insert %138, %153 [0, 0, 14] : f32 into vector<1x1x16xf32>
  %155 = vector.insert %138, %154 [0, 0, 15] : f32 into vector<1x1x16xf32>
  %156 = arith.subf %83, %155 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>]]} : vector<1x1x16xf32>
  %157 = math.exp2 %156 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>]]} : vector<1x1x16xf32>
  %158 = arith.subf %34#0, %139 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>]]} : vector<1xf32>
  %159 = math.exp2 %158 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>]]} : vector<1xf32>
  %160 = arith.mulf %159, %34#1 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>]]} : vector<1xf32>
  %161 = vector.extract %160[0] : f32 from vector<1xf32>
  %162 = vector.extract %157[0, 0, 0] : f32 from vector<1x1x16xf32>
  %163 = vector.insert %162, %cst_7 [0] : f32 into vector<1xf32>
  %164 = vector.extract %157[0, 0, 1] : f32 from vector<1x1x16xf32>
  %165 = vector.insert %164, %163 [0] : f32 into vector<1xf32>
  %166 = arith.addf %163, %165 : vector<1xf32>
  %167 = vector.extract %157[0, 0, 2] : f32 from vector<1x1x16xf32>
  %168 = vector.insert %167, %165 [0] : f32 into vector<1xf32>
  %169 = arith.addf %166, %168 : vector<1xf32>
  %170 = vector.extract %157[0, 0, 3] : f32 from vector<1x1x16xf32>
  %171 = vector.insert %170, %168 [0] : f32 into vector<1xf32>
  %172 = arith.addf %169, %171 : vector<1xf32>
  %173 = vector.extract %157[0, 0, 4] : f32 from vector<1x1x16xf32>
  %174 = vector.insert %173, %171 [0] : f32 into vector<1xf32>
  %175 = arith.addf %172, %174 : vector<1xf32>
  %176 = vector.extract %157[0, 0, 5] : f32 from vector<1x1x16xf32>
  %177 = vector.insert %176, %174 [0] : f32 into vector<1xf32>
  %178 = arith.addf %175, %177 : vector<1xf32>
  %179 = vector.extract %157[0, 0, 6] : f32 from vector<1x1x16xf32>
  %180 = vector.insert %179, %177 [0] : f32 into vector<1xf32>
  %181 = arith.addf %178, %180 : vector<1xf32>
  %182 = vector.extract %157[0, 0, 7] : f32 from vector<1x1x16xf32>
  %183 = vector.insert %182, %180 [0] : f32 into vector<1xf32>
  %184 = arith.addf %181, %183 : vector<1xf32>
  %185 = vector.extract %157[0, 0, 8] : f32 from vector<1x1x16xf32>
  %186 = vector.insert %185, %183 [0] : f32 into vector<1xf32>
  %187 = arith.addf %184, %186 : vector<1xf32>
  %188 = vector.extract %157[0, 0, 9] : f32 from vector<1x1x16xf32>
  %189 = vector.insert %188, %186 [0] : f32 into vector<1xf32>
  %190 = arith.addf %187, %189 : vector<1xf32>
  %191 = vector.extract %157[0, 0, 10] : f32 from vector<1x1x16xf32>
  %192 = vector.insert %191, %189 [0] : f32 into vector<1xf32>
  %193 = arith.addf %190, %192 : vector<1xf32>
  %194 = vector.extract %157[0, 0, 11] : f32 from vector<1x1x16xf32>
  %195 = vector.insert %194, %192 [0] : f32 into vector<1xf32>
  %196 = arith.addf %193, %195 : vector<1xf32>
  %197 = vector.extract %157[0, 0, 12] : f32 from vector<1x1x16xf32>
  %198 = vector.insert %197, %195 [0] : f32 into vector<1xf32>
  %199 = arith.addf %196, %198 : vector<1xf32>
  %200 = vector.extract %157[0, 0, 13] : f32 from vector<1x1x16xf32>
  %201 = vector.insert %200, %198 [0] : f32 into vector<1xf32>
  %202 = arith.addf %199, %201 : vector<1xf32>
  %203 = vector.extract %157[0, 0, 14] : f32 from vector<1x1x16xf32>
  %204 = vector.insert %203, %201 [0] : f32 into vector<1xf32>
  %205 = arith.addf %202, %204 : vector<1xf32>
  %206 = vector.extract %157[0, 0, 15] : f32 from vector<1x1x16xf32>
  %207 = vector.insert %206, %204 [0] : f32 into vector<1xf32>
  %208 = arith.addf %205, %207 : vector<1xf32>
  %209 = vector.bitcast %208 : vector<1xf32> to vector<1xi32>
  %210 = vector.extract %209[0] : i32 from vector<1xi32>
  %shuffleResult_17, %valid_18 = gpu.shuffle  xor %210, %c32_i32, %c64_i32 : i32
  %211 = vector.broadcast %shuffleResult_17 : i32 to vector<1xi32>
  %212 = vector.bitcast %211 : vector<1xi32> to vector<1xf32>
  %213 = arith.addf %212, %208 : vector<1xf32>
  %214 = vector.extract %213[0] : f32 from vector<1xf32>
  %215 = arith.addf %214, %161 : f32
  %216 = vector.extract_strided_slice %157 {offsets = [0, 0, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x1x16xf32> to vector<1x1x4xf32>
  %217 = vector.insert_strided_slice %216, %cst_11 {offsets = [0, 0, 0], strides = [1, 1, 1]} : vector<1x1x4xf32> into vector<4x1x4xf32>
  %218 = vector.extract_strided_slice %157 {offsets = [0, 0, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x1x16xf32> to vector<1x1x4xf32>
  %219 = vector.insert_strided_slice %218, %217 {offsets = [1, 0, 0], strides = [1, 1, 1]} : vector<1x1x4xf32> into vector<4x1x4xf32>
  %220 = vector.extract_strided_slice %157 {offsets = [0, 0, 8], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x1x16xf32> to vector<1x1x4xf32>
  %221 = vector.insert_strided_slice %220, %219 {offsets = [2, 0, 0], strides = [1, 1, 1]} : vector<1x1x4xf32> into vector<4x1x4xf32>
  %222 = vector.extract_strided_slice %157 {offsets = [0, 0, 12], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x1x16xf32> to vector<1x1x4xf32>
  %223 = vector.insert_strided_slice %222, %221 {offsets = [3, 0, 0], strides = [1, 1, 1]} : vector<1x1x4xf32> into vector<4x1x4xf32>
  %224 = arith.cmpf ogt, %223, %cst_5 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xf32>
  %225 = arith.cmpf olt, %223, %cst_4 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xf32>
  %226 = arith.select %224, %cst_5, %223 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xi1>, vector<4x1x4xf32>
  %227 = arith.select %225, %cst_4, %226 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xi1>, vector<4x1x4xf32>
  %228 = arith.truncf %227 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xf32> to vector<4x1x4xf16>
  %229 = vector.extract %159[0] : f32 from vector<1xf32>
  %230 = vector.insert %229, %cst_9 [0, 0, 0] : f32 into vector<2x1x16xf32>
  %231 = vector.insert %229, %230 [0, 0, 1] : f32 into vector<2x1x16xf32>
  %232 = vector.insert %229, %231 [0, 0, 2] : f32 into vector<2x1x16xf32>
  %233 = vector.insert %229, %232 [0, 0, 3] : f32 into vector<2x1x16xf32>
  %234 = vector.insert %229, %233 [0, 0, 4] : f32 into vector<2x1x16xf32>
  %235 = vector.insert %229, %234 [0, 0, 5] : f32 into vector<2x1x16xf32>
  %236 = vector.insert %229, %235 [0, 0, 6] : f32 into vector<2x1x16xf32>
  %237 = vector.insert %229, %236 [0, 0, 7] : f32 into vector<2x1x16xf32>
  %238 = vector.insert %229, %237 [0, 0, 8] : f32 into vector<2x1x16xf32>
  %239 = vector.insert %229, %238 [0, 0, 9] : f32 into vector<2x1x16xf32>
  %240 = vector.insert %229, %239 [0, 0, 10] : f32 into vector<2x1x16xf32>
  %241 = vector.insert %229, %240 [0, 0, 11] : f32 into vector<2x1x16xf32>
  %242 = vector.insert %229, %241 [0, 0, 12] : f32 into vector<2x1x16xf32>
  %243 = vector.insert %229, %242 [0, 0, 13] : f32 into vector<2x1x16xf32>
  %244 = vector.insert %229, %243 [0, 0, 14] : f32 into vector<2x1x16xf32>
  %245 = vector.insert %229, %244 [0, 0, 15] : f32 into vector<2x1x16xf32>
  %246 = vector.insert %229, %245 [1, 0, 0] : f32 into vector<2x1x16xf32>
  %247 = vector.insert %229, %246 [1, 0, 1] : f32 into vector<2x1x16xf32>
  %248 = vector.insert %229, %247 [1, 0, 2] : f32 into vector<2x1x16xf32>
  %249 = vector.insert %229, %248 [1, 0, 3] : f32 into vector<2x1x16xf32>
  %250 = vector.insert %229, %249 [1, 0, 4] : f32 into vector<2x1x16xf32>
  %251 = vector.insert %229, %250 [1, 0, 5] : f32 into vector<2x1x16xf32>
  %252 = vector.insert %229, %251 [1, 0, 6] : f32 into vector<2x1x16xf32>
  %253 = vector.insert %229, %252 [1, 0, 7] : f32 into vector<2x1x16xf32>
  %254 = vector.insert %229, %253 [1, 0, 8] : f32 into vector<2x1x16xf32>
  %255 = vector.insert %229, %254 [1, 0, 9] : f32 into vector<2x1x16xf32>
  %256 = vector.insert %229, %255 [1, 0, 10] : f32 into vector<2x1x16xf32>
  %257 = vector.insert %229, %256 [1, 0, 11] : f32 into vector<2x1x16xf32>
  %258 = vector.insert %229, %257 [1, 0, 12] : f32 into vector<2x1x16xf32>
  %259 = vector.insert %229, %258 [1, 0, 13] : f32 into vector<2x1x16xf32>
  %260 = vector.insert %229, %259 [1, 0, 14] : f32 into vector<2x1x16xf32>
  %261 = vector.insert %229, %260 [1, 0, 15] : f32 into vector<2x1x16xf32>
  %262 = arith.mulf %261, %34#2 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf32>
  %263 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4)>()[%thread_id_x]
  %264 = vector.load %subview[%c0, %c0, %263, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %265 = vector.insert_strided_slice %264, %cst_12 {offsets = [0, 0, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %266 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 1)>()[%thread_id_x]
  %267 = vector.load %subview[%c0, %c0, %266, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %268 = vector.insert_strided_slice %267, %265 {offsets = [0, 0, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %269 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 2)>()[%thread_id_x]
  %270 = vector.load %subview[%c0, %c0, %269, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %271 = vector.insert_strided_slice %270, %268 {offsets = [0, 0, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %272 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 3)>()[%thread_id_x]
  %273 = vector.load %subview[%c0, %c0, %272, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %274 = vector.insert_strided_slice %273, %271 {offsets = [0, 0, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %275 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 + 2) floordiv 8)>()[%thread_id_x]
  %276 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 - (((s0 mod 64) floordiv 32 + 2) floordiv 8) * 8 + 2) floordiv 8)>()[%thread_id_x]
  %277 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - (((s0 mod 64) floordiv 32 + 2) floordiv 8) * 32 + 8)>()[%thread_id_x]
  %278 = vector.load %subview[%275, %276, %277, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %279 = vector.insert_strided_slice %278, %274 {offsets = [0, 1, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %280 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 9) floordiv 32)>()[%thread_id_x]
  %281 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 9) floordiv 32) * 32 + 9) floordiv 32)>()[%thread_id_x]
  %282 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 9) floordiv 32) * 32 + 9)>()[%thread_id_x]
  %283 = vector.load %subview[%280, %281, %282, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %284 = vector.insert_strided_slice %283, %279 {offsets = [0, 1, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %285 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 + 5) floordiv 16)>()[%thread_id_x]
  %286 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 - ((((s0 mod 64) floordiv 32) * 2 + 5) floordiv 16) * 16 + 5) floordiv 16)>()[%thread_id_x]
  %287 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 2 + 5) floordiv 16) * 32 + 10)>()[%thread_id_x]
  %288 = vector.load %subview[%285, %286, %287, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %289 = vector.insert_strided_slice %288, %284 {offsets = [0, 1, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %290 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 11) floordiv 32)>()[%thread_id_x]
  %291 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 11) floordiv 32) * 32 + 11) floordiv 32)>()[%thread_id_x]
  %292 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 11) floordiv 32) * 32 + 11)>()[%thread_id_x]
  %293 = vector.load %subview[%290, %291, %292, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %294 = vector.insert_strided_slice %293, %289 {offsets = [0, 1, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %295 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 + 4) floordiv 8)>()[%thread_id_x]
  %296 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 - (((s0 mod 64) floordiv 32 + 4) floordiv 8) * 8 + 4) floordiv 8)>()[%thread_id_x]
  %297 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - (((s0 mod 64) floordiv 32 + 4) floordiv 8) * 32 + 16)>()[%thread_id_x]
  %298 = vector.load %subview[%295, %296, %297, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %299 = vector.insert_strided_slice %298, %294 {offsets = [0, 2, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %300 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 17) floordiv 32)>()[%thread_id_x]
  %301 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 17) floordiv 32) * 32 + 17) floordiv 32)>()[%thread_id_x]
  %302 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 17) floordiv 32) * 32 + 17)>()[%thread_id_x]
  %303 = vector.load %subview[%300, %301, %302, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %304 = vector.insert_strided_slice %303, %299 {offsets = [0, 2, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %305 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 + 9) floordiv 16)>()[%thread_id_x]
  %306 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 - ((((s0 mod 64) floordiv 32) * 2 + 9) floordiv 16) * 16 + 9) floordiv 16)>()[%thread_id_x]
  %307 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 2 + 9) floordiv 16) * 32 + 18)>()[%thread_id_x]
  %308 = vector.load %subview[%305, %306, %307, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %309 = vector.insert_strided_slice %308, %304 {offsets = [0, 2, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %310 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 19) floordiv 32)>()[%thread_id_x]
  %311 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 19) floordiv 32) * 32 + 19) floordiv 32)>()[%thread_id_x]
  %312 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 19) floordiv 32) * 32 + 19)>()[%thread_id_x]
  %313 = vector.load %subview[%310, %311, %312, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %314 = vector.insert_strided_slice %313, %309 {offsets = [0, 2, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %315 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 + 6) floordiv 8)>()[%thread_id_x]
  %316 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 - (((s0 mod 64) floordiv 32 + 6) floordiv 8) * 8 + 6) floordiv 8)>()[%thread_id_x]
  %317 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - (((s0 mod 64) floordiv 32 + 6) floordiv 8) * 32 + 24)>()[%thread_id_x]
  %318 = vector.load %subview[%315, %316, %317, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %319 = vector.insert_strided_slice %318, %314 {offsets = [0, 3, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %320 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 25) floordiv 32)>()[%thread_id_x]
  %321 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 25) floordiv 32) * 32 + 25) floordiv 32)>()[%thread_id_x]
  %322 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 25) floordiv 32) * 32 + 25)>()[%thread_id_x]
  %323 = vector.load %subview[%320, %321, %322, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %324 = vector.insert_strided_slice %323, %319 {offsets = [0, 3, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %325 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 + 13) floordiv 16)>()[%thread_id_x]
  %326 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 - ((((s0 mod 64) floordiv 32) * 2 + 13) floordiv 16) * 16 + 13) floordiv 16)>()[%thread_id_x]
  %327 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 2 + 13) floordiv 16) * 32 + 26)>()[%thread_id_x]
  %328 = vector.load %subview[%325, %326, %327, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %329 = vector.insert_strided_slice %328, %324 {offsets = [0, 3, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %330 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 27) floordiv 32)>()[%thread_id_x]
  %331 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 27) floordiv 32) * 32 + 27) floordiv 32)>()[%thread_id_x]
  %332 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 27) floordiv 32) * 32 + 27)>()[%thread_id_x]
  %333 = vector.load %subview[%330, %331, %332, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %334 = vector.insert_strided_slice %333, %329 {offsets = [0, 3, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %335 = affine.apply affine_map<()[s0] -> (s0 mod 32 + 32)>()[%thread_id_x]
  %336 = vector.load %subview[%c0, %c0, %263, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %337 = vector.insert_strided_slice %336, %334 {offsets = [1, 0, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %338 = vector.load %subview[%c0, %c0, %266, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %339 = vector.insert_strided_slice %338, %337 {offsets = [1, 0, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %340 = vector.load %subview[%c0, %c0, %269, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %341 = vector.insert_strided_slice %340, %339 {offsets = [1, 0, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %342 = vector.load %subview[%c0, %c0, %272, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %343 = vector.insert_strided_slice %342, %341 {offsets = [1, 0, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %344 = vector.load %subview[%275, %276, %277, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %345 = vector.insert_strided_slice %344, %343 {offsets = [1, 1, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %346 = vector.load %subview[%280, %281, %282, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %347 = vector.insert_strided_slice %346, %345 {offsets = [1, 1, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %348 = vector.load %subview[%285, %286, %287, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %349 = vector.insert_strided_slice %348, %347 {offsets = [1, 1, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %350 = vector.load %subview[%290, %291, %292, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %351 = vector.insert_strided_slice %350, %349 {offsets = [1, 1, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %352 = vector.load %subview[%295, %296, %297, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %353 = vector.insert_strided_slice %352, %351 {offsets = [1, 2, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %354 = vector.load %subview[%300, %301, %302, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %355 = vector.insert_strided_slice %354, %353 {offsets = [1, 2, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %356 = vector.load %subview[%305, %306, %307, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %357 = vector.insert_strided_slice %356, %355 {offsets = [1, 2, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %358 = vector.load %subview[%310, %311, %312, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %359 = vector.insert_strided_slice %358, %357 {offsets = [1, 2, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %360 = vector.load %subview[%315, %316, %317, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %361 = vector.insert_strided_slice %360, %359 {offsets = [1, 3, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %362 = vector.load %subview[%320, %321, %322, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %363 = vector.insert_strided_slice %362, %361 {offsets = [1, 3, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %364 = vector.load %subview[%325, %326, %327, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %365 = vector.insert_strided_slice %364, %363 {offsets = [1, 3, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %366 = vector.load %subview[%330, %331, %332, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %367 = vector.insert_strided_slice %366, %365 {offsets = [1, 3, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %368 = vector.extract %262[0, 0] : vector<16xf32> from vector<2x1x16xf32>
  %369 = vector.extract %367[0, 0] : vector<4xf16> from vector<2x4x4xf16>
  %370 = vector.extract %228[0, 0] : vector<4xf16> from vector<4x1x4xf16>
  %371 = amdgpu.mfma %369 * %370 + %368 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %372 = vector.extract %367[0, 1] : vector<4xf16> from vector<2x4x4xf16>
  %373 = vector.extract %228[1, 0] : vector<4xf16> from vector<4x1x4xf16>
  %374 = amdgpu.mfma %372 * %373 + %371 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %375 = vector.extract %367[0, 2] : vector<4xf16> from vector<2x4x4xf16>
  %376 = vector.extract %228[2, 0] : vector<4xf16> from vector<4x1x4xf16>
  %377 = amdgpu.mfma %375 * %376 + %374 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %378 = vector.extract %367[0, 3] : vector<4xf16> from vector<2x4x4xf16>
  %379 = vector.extract %228[3, 0] : vector<4xf16> from vector<4x1x4xf16>
  %380 = amdgpu.mfma %378 * %379 + %377 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %381 = vector.insert %380, %cst_9 [0, 0] : vector<16xf32> into vector<2x1x16xf32>
  %382 = vector.extract %262[1, 0] : vector<16xf32> from vector<2x1x16xf32>
  %383 = vector.extract %367[1, 0] : vector<4xf16> from vector<2x4x4xf16>
  %384 = amdgpu.mfma %383 * %370 + %382 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %385 = vector.extract %367[1, 1] : vector<4xf16> from vector<2x4x4xf16>
  %386 = amdgpu.mfma %385 * %373 + %384 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %387 = vector.extract %367[1, 2] : vector<4xf16> from vector<2x4x4xf16>
  %388 = amdgpu.mfma %387 * %376 + %386 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %389 = vector.extract %367[1, 3] : vector<4xf16> from vector<2x4x4xf16>
  %390 = amdgpu.mfma %389 * %379 + %388 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %391 = vector.insert %390, %381 [1, 0] : vector<16xf32> into vector<2x1x16xf32>
  %392 = vector.insert %215, %cst_9 [0, 0, 0] : f32 into vector<2x1x16xf32>
  %393 = vector.insert %215, %392 [0, 0, 1] : f32 into vector<2x1x16xf32>
  %394 = vector.insert %215, %393 [0, 0, 2] : f32 into vector<2x1x16xf32>
  %395 = vector.insert %215, %394 [0, 0, 3] : f32 into vector<2x1x16xf32>
  %396 = vector.insert %215, %395 [0, 0, 4] : f32 into vector<2x1x16xf32>
  %397 = vector.insert %215, %396 [0, 0, 5] : f32 into vector<2x1x16xf32>
  %398 = vector.insert %215, %397 [0, 0, 6] : f32 into vector<2x1x16xf32>
  %399 = vector.insert %215, %398 [0, 0, 7] : f32 into vector<2x1x16xf32>
  %400 = vector.insert %215, %399 [0, 0, 8] : f32 into vector<2x1x16xf32>
  %401 = vector.insert %215, %400 [0, 0, 9] : f32 into vector<2x1x16xf32>
  %402 = vector.insert %215, %401 [0, 0, 10] : f32 into vector<2x1x16xf32>
  %403 = vector.insert %215, %402 [0, 0, 11] : f32 into vector<2x1x16xf32>
  %404 = vector.insert %215, %403 [0, 0, 12] : f32 into vector<2x1x16xf32>
  %405 = vector.insert %215, %404 [0, 0, 13] : f32 into vector<2x1x16xf32>
  %406 = vector.insert %215, %405 [0, 0, 14] : f32 into vector<2x1x16xf32>
  %407 = vector.insert %215, %406 [0, 0, 15] : f32 into vector<2x1x16xf32>
  %408 = vector.insert %215, %407 [1, 0, 0] : f32 into vector<2x1x16xf32>
  %409 = vector.insert %215, %408 [1, 0, 1] : f32 into vector<2x1x16xf32>
  %410 = vector.insert %215, %409 [1, 0, 2] : f32 into vector<2x1x16xf32>
  %411 = vector.insert %215, %410 [1, 0, 3] : f32 into vector<2x1x16xf32>
  %412 = vector.insert %215, %411 [1, 0, 4] : f32 into vector<2x1x16xf32>
  %413 = vector.insert %215, %412 [1, 0, 5] : f32 into vector<2x1x16xf32>
  %414 = vector.insert %215, %413 [1, 0, 6] : f32 into vector<2x1x16xf32>
  %415 = vector.insert %215, %414 [1, 0, 7] : f32 into vector<2x1x16xf32>
  %416 = vector.insert %215, %415 [1, 0, 8] : f32 into vector<2x1x16xf32>
  %417 = vector.insert %215, %416 [1, 0, 9] : f32 into vector<2x1x16xf32>
  %418 = vector.insert %215, %417 [1, 0, 10] : f32 into vector<2x1x16xf32>
  %419 = vector.insert %215, %418 [1, 0, 11] : f32 into vector<2x1x16xf32>
  %420 = vector.insert %215, %419 [1, 0, 12] : f32 into vector<2x1x16xf32>
  %421 = vector.insert %215, %420 [1, 0, 13] : f32 into vector<2x1x16xf32>
  %422 = vector.insert %215, %421 [1, 0, 14] : f32 into vector<2x1x16xf32>
  %423 = vector.insert %215, %422 [1, 0, 15] : f32 into vector<2x1x16xf32>
  %424 = vector.load %11[%workgroup_id_y, %263] : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>, vector<4xf16>
  %425 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 8)>()[%thread_id_x]
  %426 = vector.load %11[%workgroup_id_y, %425] : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>, vector<4xf16>
  %427 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 16)>()[%thread_id_x]
  %428 = vector.load %11[%workgroup_id_y, %427] : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>, vector<4xf16>
  %429 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 24)>()[%thread_id_x]
  %430 = vector.load %11[%workgroup_id_y, %429] : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>, vector<4xf16>
  %431 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 32)>()[%thread_id_x]
  %432 = vector.load %11[%workgroup_id_y, %431] : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>, vector<4xf16>
  %433 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 40)>()[%thread_id_x]
  %434 = vector.load %11[%workgroup_id_y, %433] : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>, vector<4xf16>
  %435 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 48)>()[%thread_id_x]
  %436 = vector.load %11[%workgroup_id_y, %435] : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>, vector<4xf16>
  %437 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 56)>()[%thread_id_x]
  %438 = vector.load %11[%workgroup_id_y, %437] : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>, vector<4xf16>
  %439 = vector.extract %424[0] : f16 from vector<4xf16>
  %440 = vector.insert %439, %cst_13 [0, 0, 0] : f16 into vector<2x1x16xf16>
  %441 = vector.extract %424[1] : f16 from vector<4xf16>
  %442 = vector.insert %441, %440 [0, 0, 1] : f16 into vector<2x1x16xf16>
  %443 = vector.extract %424[2] : f16 from vector<4xf16>
  %444 = vector.insert %443, %442 [0, 0, 2] : f16 into vector<2x1x16xf16>
  %445 = vector.extract %424[3] : f16 from vector<4xf16>
  %446 = vector.insert %445, %444 [0, 0, 3] : f16 into vector<2x1x16xf16>
  %447 = vector.extract %426[0] : f16 from vector<4xf16>
  %448 = vector.insert %447, %446 [0, 0, 4] : f16 into vector<2x1x16xf16>
  %449 = vector.extract %426[1] : f16 from vector<4xf16>
  %450 = vector.insert %449, %448 [0, 0, 5] : f16 into vector<2x1x16xf16>
  %451 = vector.extract %426[2] : f16 from vector<4xf16>
  %452 = vector.insert %451, %450 [0, 0, 6] : f16 into vector<2x1x16xf16>
  %453 = vector.extract %426[3] : f16 from vector<4xf16>
  %454 = vector.insert %453, %452 [0, 0, 7] : f16 into vector<2x1x16xf16>
  %455 = vector.extract %428[0] : f16 from vector<4xf16>
  %456 = vector.insert %455, %454 [0, 0, 8] : f16 into vector<2x1x16xf16>
  %457 = vector.extract %428[1] : f16 from vector<4xf16>
  %458 = vector.insert %457, %456 [0, 0, 9] : f16 into vector<2x1x16xf16>
  %459 = vector.extract %428[2] : f16 from vector<4xf16>
  %460 = vector.insert %459, %458 [0, 0, 10] : f16 into vector<2x1x16xf16>
  %461 = vector.extract %428[3] : f16 from vector<4xf16>
  %462 = vector.insert %461, %460 [0, 0, 11] : f16 into vector<2x1x16xf16>
  %463 = vector.extract %430[0] : f16 from vector<4xf16>
  %464 = vector.insert %463, %462 [0, 0, 12] : f16 into vector<2x1x16xf16>
  %465 = vector.extract %430[1] : f16 from vector<4xf16>
  %466 = vector.insert %465, %464 [0, 0, 13] : f16 into vector<2x1x16xf16>
  %467 = vector.extract %430[2] : f16 from vector<4xf16>
  %468 = vector.insert %467, %466 [0, 0, 14] : f16 into vector<2x1x16xf16>
  %469 = vector.extract %430[3] : f16 from vector<4xf16>
  %470 = vector.insert %469, %468 [0, 0, 15] : f16 into vector<2x1x16xf16>
  %471 = vector.extract %432[0] : f16 from vector<4xf16>
  %472 = vector.insert %471, %470 [1, 0, 0] : f16 into vector<2x1x16xf16>
  %473 = vector.extract %432[1] : f16 from vector<4xf16>
  %474 = vector.insert %473, %472 [1, 0, 1] : f16 into vector<2x1x16xf16>
  %475 = vector.extract %432[2] : f16 from vector<4xf16>
  %476 = vector.insert %475, %474 [1, 0, 2] : f16 into vector<2x1x16xf16>
  %477 = vector.extract %432[3] : f16 from vector<4xf16>
  %478 = vector.insert %477, %476 [1, 0, 3] : f16 into vector<2x1x16xf16>
  %479 = vector.extract %434[0] : f16 from vector<4xf16>
  %480 = vector.insert %479, %478 [1, 0, 4] : f16 into vector<2x1x16xf16>
  %481 = vector.extract %434[1] : f16 from vector<4xf16>
  %482 = vector.insert %481, %480 [1, 0, 5] : f16 into vector<2x1x16xf16>
  %483 = vector.extract %434[2] : f16 from vector<4xf16>
  %484 = vector.insert %483, %482 [1, 0, 6] : f16 into vector<2x1x16xf16>
  %485 = vector.extract %434[3] : f16 from vector<4xf16>
  %486 = vector.insert %485, %484 [1, 0, 7] : f16 into vector<2x1x16xf16>
  %487 = vector.extract %436[0] : f16 from vector<4xf16>
  %488 = vector.insert %487, %486 [1, 0, 8] : f16 into vector<2x1x16xf16>
  %489 = vector.extract %436[1] : f16 from vector<4xf16>
  %490 = vector.insert %489, %488 [1, 0, 9] : f16 into vector<2x1x16xf16>
  %491 = vector.extract %436[2] : f16 from vector<4xf16>
  %492 = vector.insert %491, %490 [1, 0, 10] : f16 into vector<2x1x16xf16>
  %493 = vector.extract %436[3] : f16 from vector<4xf16>
  %494 = vector.insert %493, %492 [1, 0, 11] : f16 into vector<2x1x16xf16>
  %495 = vector.extract %438[0] : f16 from vector<4xf16>
  %496 = vector.insert %495, %494 [1, 0, 12] : f16 into vector<2x1x16xf16>
  %497 = vector.extract %438[1] : f16 from vector<4xf16>
  %498 = vector.insert %497, %496 [1, 0, 13] : f16 into vector<2x1x16xf16>
  %499 = vector.extract %438[2] : f16 from vector<4xf16>
  %500 = vector.insert %499, %498 [1, 0, 14] : f16 into vector<2x1x16xf16>
  %501 = vector.extract %438[3] : f16 from vector<4xf16>
  %502 = vector.insert %501, %500 [1, 0, 15] : f16 into vector<2x1x16xf16>
  %503 = memref.load %12[] : memref<f32, #hal.descriptor_type<storage_buffer>>
  %504 = arith.divf %cst_3, %423 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>, <[ BATCHY,  LANEX], [1, 32]>>, #iree_vector_ext.layout<<[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>, <[ BATCHY,  LANEX], [1, 32]>>], [#iree_vector_ext.layout<<[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>, <[ BATCHY,  LANEX], [1, 32]>>]]} : vector<2x1x16xf32>
  %505 = arith.mulf %504, %391 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf32>
  %506 = arith.truncf %505 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf32> to vector<2x1x16xf16>
  %507 = arith.mulf %506, %502 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf16>
  %508 = arith.truncf %503 : f32 to f16
  %509 = vector.broadcast %508 : f16 to vector<2x1x16xf16>
  %510 = arith.divf %507, %509 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf16>
  %511 = math.roundeven %510 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf16>
  %512 = arith.cmpf ult, %511, %cst_2 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf16>
  %513 = arith.select %512, %cst_2, %511 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xi1>, vector<2x1x16xf16>
  %514 = arith.cmpf ugt, %513, %cst_1 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf16>
  %515 = arith.select %514, %cst_1, %513 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xi1>, vector<2x1x16xf16>
  %516 = arith.fptosi %515 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf16> to vector<2x1x16xi8>
  %517 = vector.extract_strided_slice %516 {offsets = [0, 0, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<2x1x16xi8> to vector<1x1x4xi8>
  %518 = vector.extract %517[0, 0] : vector<4xi8> from vector<1x1x4xi8>
  vector.store %518, %13[%workgroup_id_x, %17, %workgroup_id_y, %263] : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<4xi8>
  %519 = vector.extract_strided_slice %516 {offsets = [0, 0, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<2x1x16xi8> to vector<1x1x4xi8>
  %520 = vector.extract %519[0, 0] : vector<4xi8> from vector<1x1x4xi8>
  vector.store %520, %13[%workgroup_id_x, %17, %workgroup_id_y, %425] : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<4xi8>
  %521 = vector.extract_strided_slice %516 {offsets = [0, 0, 8], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<2x1x16xi8> to vector<1x1x4xi8>
  %522 = vector.extract %521[0, 0] : vector<4xi8> from vector<1x1x4xi8>
  vector.store %522, %13[%workgroup_id_x, %17, %workgroup_id_y, %427] : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<4xi8>
  %523 = vector.extract_strided_slice %516 {offsets = [0, 0, 12], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<2x1x16xi8> to vector<1x1x4xi8>
  %524 = vector.extract %523[0, 0] : vector<4xi8> from vector<1x1x4xi8>
  vector.store %524, %13[%workgroup_id_x, %17, %workgroup_id_y, %429] : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<4xi8>
  %525 = vector.extract_strided_slice %516 {offsets = [1, 0, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<2x1x16xi8> to vector<1x1x4xi8>
  %526 = vector.extract %525[0, 0] : vector<4xi8> from vector<1x1x4xi8>
  vector.store %526, %13[%workgroup_id_x, %17, %workgroup_id_y, %431] : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<4xi8>
  %527 = vector.extract_strided_slice %516 {offsets = [1, 0, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<2x1x16xi8> to vector<1x1x4xi8>
  %528 = vector.extract %527[0, 0] : vector<4xi8> from vector<1x1x4xi8>
  vector.store %528, %13[%workgroup_id_x, %17, %workgroup_id_y, %433] : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<4xi8>
  %529 = vector.extract_strided_slice %516 {offsets = [1, 0, 8], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<2x1x16xi8> to vector<1x1x4xi8>
  %530 = vector.extract %529[0, 0] : vector<4xi8> from vector<1x1x4xi8>
  vector.store %530, %13[%workgroup_id_x, %17, %workgroup_id_y, %435] : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<4xi8>
  %531 = vector.extract_strided_slice %516 {offsets = [1, 0, 12], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<2x1x16xi8> to vector<1x1x4xi8>
  %532 = vector.extract %531[0, 0] : vector<4xi8> from vector<1x1x4xi8>
  vector.store %532, %13[%workgroup_id_x, %17, %workgroup_id_y, %437] : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<4xi8>
  return
}

// -----// IR Dump After LLVMGPULowerExecutableTarget (iree-llvmgpu-lower-executable-target) //----- //
func.func @main$async_dispatch_146_attention_2x20x1024x64xf16_generic() attributes {translation_info = #iree_codegen.translation_info<None workgroup_size = [64, 4, 1] subgroup_size = 64, {llvm_func_attrs = {"amdgpu-waves-per-eu" = "2", "denormal-fp-math-f32" = "preserve-sign"}}>} {
  %c992 = arith.constant 992 : index
  %cst = arith.constant 0.000000e+00 : f16
  %cst_0 = arith.constant dense<0.000000e+00> : vector<1x1x16xf32>
  %c0 = arith.constant 0 : index
  %cst_1 = arith.constant dense<1.270000e+02> : vector<2x1x16xf16>
  %cst_2 = arith.constant dense<-1.280000e+02> : vector<2x1x16xf16>
  %cst_3 = arith.constant dense<1.000000e+00> : vector<2x1x16xf32>
  %cst_4 = arith.constant dense<-6.550400e+04> : vector<4x1x4xf32>
  %cst_5 = arith.constant dense<6.550400e+04> : vector<4x1x4xf32>
  %cst_6 = arith.constant dense<1.802980e-01> : vector<1x4x8xf16>
  %cst_7 = arith.constant dense<0.000000e+00> : vector<1xf32>
  %cst_8 = arith.constant dense<-3.40282347E+38> : vector<1xf32>
  %cst_9 = arith.constant dense<0.000000e+00> : vector<2x1x16xf32>
  %c32 = arith.constant 32 : index
  %cst_10 = arith.constant dense<0.000000e+00> : vector<1x4x8xf16>
  %c32_i32 = arith.constant 32 : i32
  %c64_i32 = arith.constant 64 : i32
  %cst_11 = arith.constant dense<0.000000e+00> : vector<4x1x4xf32>
  %cst_12 = arith.constant dense<0.000000e+00> : vector<2x4x4xf16>
  %cst_13 = arith.constant dense<0.000000e+00> : vector<2x1x16xf16>
  %cst_14 = arith.constant dense<0.000000e+00> : vector<16xf32>
  %thread_id_x = gpu.thread_id  x
  %thread_id_y = gpu.thread_id  y
  %thread_id_z = gpu.thread_id  z
  %alloc = memref.alloc() : memref<1x1x32x68xf16, #gpu.address_space<workgroup>>
  %subview = memref.subview %alloc[0, 0, 0, 0] [1, 1, 32, 64] [1, 1, 1, 1] : memref<1x1x32x68xf16, #gpu.address_space<workgroup>> to memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>
  %alloc_15 = memref.alloc() : memref<1x1x32x68xf16, #gpu.address_space<workgroup>>
  %subview_16 = memref.subview %alloc_15[0, 0, 0, 0] [1, 1, 32, 64] [1, 1, 1, 1] : memref<1x1x32x68xf16, #gpu.address_space<workgroup>> to memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>
  %0 = hal.interface.constant.load[0] : i32
  %1 = hal.interface.constant.load[1] : i32
  %2 = hal.interface.constant.load[2] : i32
  %3 = hal.interface.constant.load[3] : i32
  %4 = arith.index_castui %0 : i32 to index
  %5 = arith.index_castui %1 : i32 to index
  %6 = arith.index_castui %2 : i32 to index
  %7 = arith.index_castui %3 : i32 to index
  %8 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%4) flags(ReadOnly) : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>
  memref.assume_alignment %8, 1 : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>
  %9 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%5) flags(ReadOnly) : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>
  memref.assume_alignment %9, 1 : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>
  %10 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%6) flags(ReadOnly) : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>
  memref.assume_alignment %10, 1 : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>
  %11 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>
  memref.assume_alignment %11, 64 : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>
  %12 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : memref<f32, #hal.descriptor_type<storage_buffer>>
  memref.assume_alignment %12, 64 : memref<f32, #hal.descriptor_type<storage_buffer>>
  %13 = hal.interface.binding.subspan set(0) binding(3) type(storage_buffer) alignment(64) offset(%7) : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>
  memref.assume_alignment %13, 1 : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_id_z = hal.interface.workgroup.id[2] : index
  %14 = affine.apply affine_map<()[s0, s1, s2] -> (s0 * 128 + s2 * 32 + (s1 floordiv 64) * 32)>()[%workgroup_id_z, %thread_id_x, %thread_id_y]
  %15 = affine.apply affine_map<()[s0] -> (s0 mod 32)>()[%thread_id_x]
  %16 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 8)>()[%thread_id_x]
  %17 = arith.addi %14, %15 : index
  %18 = vector.load %8[%workgroup_id_x, %workgroup_id_y, %17, %16] : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<8xf16>
  %19 = vector.insert_strided_slice %18, %cst_10 {offsets = [0, 0, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
  %20 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 8 + 16)>()[%thread_id_x]
  %21 = vector.load %8[%workgroup_id_x, %workgroup_id_y, %17, %20] : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<8xf16>
  %22 = vector.insert_strided_slice %21, %19 {offsets = [0, 1, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
  %23 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 8 + 32)>()[%thread_id_x]
  %24 = vector.load %8[%workgroup_id_x, %workgroup_id_y, %17, %23] : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<8xf16>
  %25 = vector.insert_strided_slice %24, %22 {offsets = [0, 2, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
  %26 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 8 + 48)>()[%thread_id_x]
  %27 = vector.load %8[%workgroup_id_x, %workgroup_id_y, %17, %26] : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<8xf16>
  %28 = vector.insert_strided_slice %27, %25 {offsets = [0, 3, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
  %29 = arith.mulf %28, %cst_6 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHX,  LANEX], [1, 32]>, <[ BATCHY,  LANEY,  VECTORX], [4, 2, 8]>>, #iree_vector_ext.layout<<[ BATCHX,  LANEX], [1, 32]>, <[ BATCHY,  LANEY,  VECTORX], [4, 2, 8]>>], [#iree_vector_ext.layout<<[ BATCHX,  LANEX], [1, 32]>, <[ BATCHY,  LANEY,  VECTORX], [4, 2, 8]>>]], __vector_layout_test_anchor_result_0 = #iree_vector_ext.layout<<[ BATCHX,  LANEX], [1, 32]>, <[ BATCHY,  LANEY,  VECTORX], [4, 2, 8]>>} : vector<1x4x8xf16>
  %30 = affine.apply affine_map<()[s0, s1, s2] -> (s1 * 8 + s2 * 32 + s0 floordiv 8)>()[%thread_id_x, %thread_id_y, %thread_id_z]
  %31 = affine.apply affine_map<()[s0] -> (s0 * 8 - (s0 floordiv 8) * 64)>()[%thread_id_x]
  %32 = vector.transfer_read %9[%workgroup_id_x, %workgroup_id_y, %30, %31], %cst {in_bounds = [true, true, true, true]} : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<1x1x1x8xf16>
  %33 = vector.transfer_read %10[%workgroup_id_x, %workgroup_id_y, %30, %31], %cst {in_bounds = [true, true, true, true]} : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<1x1x1x8xf16>
  vector.transfer_write %32, %subview_16[%c0, %c0, %30, %31] {in_bounds = [true, true, true, true]} : vector<1x1x1x8xf16>, memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>
  vector.transfer_write %33, %subview[%c0, %c0, %30, %31] {in_bounds = [true, true, true, true]} : vector<1x1x1x8xf16>, memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>
  %34:3 = scf.for %arg0 = %c0 to %c992 step %c32 iter_args(%arg1 = %cst_8, %arg2 = %cst_7, %arg3 = %cst_9) -> (vector<1xf32>, vector<1xf32>, vector<2x1x16xf32>) {
    %533 = arith.addi %arg0, %c32 : index
    %534 = affine.apply affine_map<()[s0, s1, s2, s3] -> (s0 + s2 * 8 + s3 * 32 + s1 floordiv 8)>()[%533, %thread_id_x, %thread_id_y, %thread_id_z]
    %535 = vector.transfer_read %9[%workgroup_id_x, %workgroup_id_y, %534, %31], %cst {in_bounds = [true, true, true, true]} : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<1x1x1x8xf16>
    %536 = vector.transfer_read %10[%workgroup_id_x, %workgroup_id_y, %534, %31], %cst {in_bounds = [true, true, true, true]} : memref<2x20x1024x64xf16, strided<[1310720, 65536, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<1x1x1x8xf16>
    gpu.barrier
    %537 = vector.load %subview_16[%c0, %c0, %15, %16] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<8xf16>
    %538 = vector.insert_strided_slice %537, %cst_10 {offsets = [0, 0, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
    %539 = vector.load %subview_16[%c0, %c0, %15, %20] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<8xf16>
    %540 = vector.insert_strided_slice %539, %538 {offsets = [0, 1, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
    %541 = vector.load %subview_16[%c0, %c0, %15, %23] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<8xf16>
    %542 = vector.insert_strided_slice %541, %540 {offsets = [0, 2, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
    %543 = vector.load %subview_16[%c0, %c0, %15, %26] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<8xf16>
    %544 = vector.insert_strided_slice %543, %542 {offsets = [0, 3, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
    %545 = vector.extract_strided_slice %544 {offsets = [0, 0, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %546 = vector.extract_strided_slice %544 {offsets = [0, 0, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %547 = vector.extract_strided_slice %544 {offsets = [0, 1, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %548 = vector.extract_strided_slice %544 {offsets = [0, 1, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %549 = vector.extract_strided_slice %544 {offsets = [0, 2, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %550 = vector.extract_strided_slice %544 {offsets = [0, 2, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %551 = vector.extract_strided_slice %544 {offsets = [0, 3, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %552 = vector.extract_strided_slice %544 {offsets = [0, 3, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %553 = vector.extract_strided_slice %29 {offsets = [0, 0, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %554 = vector.extract_strided_slice %29 {offsets = [0, 0, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %555 = vector.extract_strided_slice %29 {offsets = [0, 1, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %556 = vector.extract_strided_slice %29 {offsets = [0, 1, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %557 = vector.extract_strided_slice %29 {offsets = [0, 2, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %558 = vector.extract_strided_slice %29 {offsets = [0, 2, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %559 = vector.extract_strided_slice %29 {offsets = [0, 3, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %560 = vector.extract_strided_slice %29 {offsets = [0, 3, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
    %561 = vector.extract %545[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %562 = vector.extract %553[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %563 = amdgpu.mfma %561 * %562 + %cst_14 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %564 = vector.extract %546[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %565 = vector.extract %554[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %566 = amdgpu.mfma %564 * %565 + %563 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %567 = vector.extract %547[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %568 = vector.extract %555[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %569 = amdgpu.mfma %567 * %568 + %566 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %570 = vector.extract %548[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %571 = vector.extract %556[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %572 = amdgpu.mfma %570 * %571 + %569 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %573 = vector.extract %549[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %574 = vector.extract %557[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %575 = amdgpu.mfma %573 * %574 + %572 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %576 = vector.extract %550[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %577 = vector.extract %558[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %578 = amdgpu.mfma %576 * %577 + %575 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %579 = vector.extract %551[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %580 = vector.extract %559[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %581 = amdgpu.mfma %579 * %580 + %578 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %582 = vector.extract %552[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %583 = vector.extract %560[0, 0] : vector<4xf16> from vector<1x1x4xf16>
    %584 = amdgpu.mfma %582 * %583 + %581 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %585 = vector.broadcast %584 : vector<16xf32> to vector<1x1x16xf32>
    %586 = vector.extract %arg1[0] : f32 from vector<1xf32>
    %587 = vector.extract %584[0] : f32 from vector<16xf32>
    %588 = vector.insert %587, %cst_7 [0] : f32 into vector<1xf32>
    %589 = vector.extract %584[1] : f32 from vector<16xf32>
    %590 = vector.insert %589, %588 [0] : f32 into vector<1xf32>
    %591 = arith.maximumf %588, %590 : vector<1xf32>
    %592 = vector.extract %584[2] : f32 from vector<16xf32>
    %593 = vector.insert %592, %590 [0] : f32 into vector<1xf32>
    %594 = arith.maximumf %591, %593 : vector<1xf32>
    %595 = vector.extract %584[3] : f32 from vector<16xf32>
    %596 = vector.insert %595, %593 [0] : f32 into vector<1xf32>
    %597 = arith.maximumf %594, %596 : vector<1xf32>
    %598 = vector.extract %584[4] : f32 from vector<16xf32>
    %599 = vector.insert %598, %596 [0] : f32 into vector<1xf32>
    %600 = arith.maximumf %597, %599 : vector<1xf32>
    %601 = vector.extract %584[5] : f32 from vector<16xf32>
    %602 = vector.insert %601, %599 [0] : f32 into vector<1xf32>
    %603 = arith.maximumf %600, %602 : vector<1xf32>
    %604 = vector.extract %584[6] : f32 from vector<16xf32>
    %605 = vector.insert %604, %602 [0] : f32 into vector<1xf32>
    %606 = arith.maximumf %603, %605 : vector<1xf32>
    %607 = vector.extract %584[7] : f32 from vector<16xf32>
    %608 = vector.insert %607, %605 [0] : f32 into vector<1xf32>
    %609 = arith.maximumf %606, %608 : vector<1xf32>
    %610 = vector.extract %584[8] : f32 from vector<16xf32>
    %611 = vector.insert %610, %608 [0] : f32 into vector<1xf32>
    %612 = arith.maximumf %609, %611 : vector<1xf32>
    %613 = vector.extract %584[9] : f32 from vector<16xf32>
    %614 = vector.insert %613, %611 [0] : f32 into vector<1xf32>
    %615 = arith.maximumf %612, %614 : vector<1xf32>
    %616 = vector.extract %584[10] : f32 from vector<16xf32>
    %617 = vector.insert %616, %614 [0] : f32 into vector<1xf32>
    %618 = arith.maximumf %615, %617 : vector<1xf32>
    %619 = vector.extract %584[11] : f32 from vector<16xf32>
    %620 = vector.insert %619, %617 [0] : f32 into vector<1xf32>
    %621 = arith.maximumf %618, %620 : vector<1xf32>
    %622 = vector.extract %584[12] : f32 from vector<16xf32>
    %623 = vector.insert %622, %620 [0] : f32 into vector<1xf32>
    %624 = arith.maximumf %621, %623 : vector<1xf32>
    %625 = vector.extract %584[13] : f32 from vector<16xf32>
    %626 = vector.insert %625, %623 [0] : f32 into vector<1xf32>
    %627 = arith.maximumf %624, %626 : vector<1xf32>
    %628 = vector.extract %584[14] : f32 from vector<16xf32>
    %629 = vector.insert %628, %626 [0] : f32 into vector<1xf32>
    %630 = arith.maximumf %627, %629 : vector<1xf32>
    %631 = vector.extract %584[15] : f32 from vector<16xf32>
    %632 = vector.insert %631, %629 [0] : f32 into vector<1xf32>
    %633 = arith.maximumf %630, %632 : vector<1xf32>
    %634 = vector.bitcast %633 : vector<1xf32> to vector<1xi32>
    %635 = vector.extract %634[0] : i32 from vector<1xi32>
    %shuffleResult_19, %valid_20 = gpu.shuffle  xor %635, %c32_i32, %c64_i32 : i32
    %636 = vector.broadcast %shuffleResult_19 : i32 to vector<1xi32>
    %637 = vector.bitcast %636 : vector<1xi32> to vector<1xf32>
    %638 = arith.maximumf %637, %633 : vector<1xf32>
    %639 = vector.extract %638[0] : f32 from vector<1xf32>
    %640 = arith.maximumf %639, %586 : f32
    %641 = vector.insert %640, %cst_7 [0] : f32 into vector<1xf32>
    %642 = vector.insert %640, %cst_0 [0, 0, 0] : f32 into vector<1x1x16xf32>
    %643 = vector.insert %640, %642 [0, 0, 1] : f32 into vector<1x1x16xf32>
    %644 = vector.insert %640, %643 [0, 0, 2] : f32 into vector<1x1x16xf32>
    %645 = vector.insert %640, %644 [0, 0, 3] : f32 into vector<1x1x16xf32>
    %646 = vector.insert %640, %645 [0, 0, 4] : f32 into vector<1x1x16xf32>
    %647 = vector.insert %640, %646 [0, 0, 5] : f32 into vector<1x1x16xf32>
    %648 = vector.insert %640, %647 [0, 0, 6] : f32 into vector<1x1x16xf32>
    %649 = vector.insert %640, %648 [0, 0, 7] : f32 into vector<1x1x16xf32>
    %650 = vector.insert %640, %649 [0, 0, 8] : f32 into vector<1x1x16xf32>
    %651 = vector.insert %640, %650 [0, 0, 9] : f32 into vector<1x1x16xf32>
    %652 = vector.insert %640, %651 [0, 0, 10] : f32 into vector<1x1x16xf32>
    %653 = vector.insert %640, %652 [0, 0, 11] : f32 into vector<1x1x16xf32>
    %654 = vector.insert %640, %653 [0, 0, 12] : f32 into vector<1x1x16xf32>
    %655 = vector.insert %640, %654 [0, 0, 13] : f32 into vector<1x1x16xf32>
    %656 = vector.insert %640, %655 [0, 0, 14] : f32 into vector<1x1x16xf32>
    %657 = vector.insert %640, %656 [0, 0, 15] : f32 into vector<1x1x16xf32>
    %658 = arith.subf %585, %657 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>]]} : vector<1x1x16xf32>
    %659 = math.exp2 %658 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>]]} : vector<1x1x16xf32>
    %660 = arith.subf %arg1, %641 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>]]} : vector<1xf32>
    %661 = math.exp2 %660 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>]]} : vector<1xf32>
    %662 = arith.mulf %661, %arg2 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>]]} : vector<1xf32>
    %663 = vector.extract %662[0] : f32 from vector<1xf32>
    %664 = vector.extract %659[0, 0, 0] : f32 from vector<1x1x16xf32>
    %665 = vector.insert %664, %cst_7 [0] : f32 into vector<1xf32>
    %666 = vector.extract %659[0, 0, 1] : f32 from vector<1x1x16xf32>
    %667 = vector.insert %666, %665 [0] : f32 into vector<1xf32>
    %668 = arith.addf %665, %667 : vector<1xf32>
    %669 = vector.extract %659[0, 0, 2] : f32 from vector<1x1x16xf32>
    %670 = vector.insert %669, %667 [0] : f32 into vector<1xf32>
    %671 = arith.addf %668, %670 : vector<1xf32>
    %672 = vector.extract %659[0, 0, 3] : f32 from vector<1x1x16xf32>
    %673 = vector.insert %672, %670 [0] : f32 into vector<1xf32>
    %674 = arith.addf %671, %673 : vector<1xf32>
    %675 = vector.extract %659[0, 0, 4] : f32 from vector<1x1x16xf32>
    %676 = vector.insert %675, %673 [0] : f32 into vector<1xf32>
    %677 = arith.addf %674, %676 : vector<1xf32>
    %678 = vector.extract %659[0, 0, 5] : f32 from vector<1x1x16xf32>
    %679 = vector.insert %678, %676 [0] : f32 into vector<1xf32>
    %680 = arith.addf %677, %679 : vector<1xf32>
    %681 = vector.extract %659[0, 0, 6] : f32 from vector<1x1x16xf32>
    %682 = vector.insert %681, %679 [0] : f32 into vector<1xf32>
    %683 = arith.addf %680, %682 : vector<1xf32>
    %684 = vector.extract %659[0, 0, 7] : f32 from vector<1x1x16xf32>
    %685 = vector.insert %684, %682 [0] : f32 into vector<1xf32>
    %686 = arith.addf %683, %685 : vector<1xf32>
    %687 = vector.extract %659[0, 0, 8] : f32 from vector<1x1x16xf32>
    %688 = vector.insert %687, %685 [0] : f32 into vector<1xf32>
    %689 = arith.addf %686, %688 : vector<1xf32>
    %690 = vector.extract %659[0, 0, 9] : f32 from vector<1x1x16xf32>
    %691 = vector.insert %690, %688 [0] : f32 into vector<1xf32>
    %692 = arith.addf %689, %691 : vector<1xf32>
    %693 = vector.extract %659[0, 0, 10] : f32 from vector<1x1x16xf32>
    %694 = vector.insert %693, %691 [0] : f32 into vector<1xf32>
    %695 = arith.addf %692, %694 : vector<1xf32>
    %696 = vector.extract %659[0, 0, 11] : f32 from vector<1x1x16xf32>
    %697 = vector.insert %696, %694 [0] : f32 into vector<1xf32>
    %698 = arith.addf %695, %697 : vector<1xf32>
    %699 = vector.extract %659[0, 0, 12] : f32 from vector<1x1x16xf32>
    %700 = vector.insert %699, %697 [0] : f32 into vector<1xf32>
    %701 = arith.addf %698, %700 : vector<1xf32>
    %702 = vector.extract %659[0, 0, 13] : f32 from vector<1x1x16xf32>
    %703 = vector.insert %702, %700 [0] : f32 into vector<1xf32>
    %704 = arith.addf %701, %703 : vector<1xf32>
    %705 = vector.extract %659[0, 0, 14] : f32 from vector<1x1x16xf32>
    %706 = vector.insert %705, %703 [0] : f32 into vector<1xf32>
    %707 = arith.addf %704, %706 : vector<1xf32>
    %708 = vector.extract %659[0, 0, 15] : f32 from vector<1x1x16xf32>
    %709 = vector.insert %708, %706 [0] : f32 into vector<1xf32>
    %710 = arith.addf %707, %709 : vector<1xf32>
    %711 = vector.bitcast %710 : vector<1xf32> to vector<1xi32>
    %712 = vector.extract %711[0] : i32 from vector<1xi32>
    %shuffleResult_21, %valid_22 = gpu.shuffle  xor %712, %c32_i32, %c64_i32 : i32
    %713 = vector.broadcast %shuffleResult_21 : i32 to vector<1xi32>
    %714 = vector.bitcast %713 : vector<1xi32> to vector<1xf32>
    %715 = arith.addf %714, %710 : vector<1xf32>
    %716 = vector.extract %715[0] : f32 from vector<1xf32>
    %717 = arith.addf %716, %663 : f32
    %718 = vector.insert %717, %cst_7 [0] : f32 into vector<1xf32>
    %719 = vector.extract_strided_slice %659 {offsets = [0, 0, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x1x16xf32> to vector<1x1x4xf32>
    %720 = vector.insert_strided_slice %719, %cst_11 {offsets = [0, 0, 0], strides = [1, 1, 1]} : vector<1x1x4xf32> into vector<4x1x4xf32>
    %721 = vector.extract_strided_slice %659 {offsets = [0, 0, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x1x16xf32> to vector<1x1x4xf32>
    %722 = vector.insert_strided_slice %721, %720 {offsets = [1, 0, 0], strides = [1, 1, 1]} : vector<1x1x4xf32> into vector<4x1x4xf32>
    %723 = vector.extract_strided_slice %659 {offsets = [0, 0, 8], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x1x16xf32> to vector<1x1x4xf32>
    %724 = vector.insert_strided_slice %723, %722 {offsets = [2, 0, 0], strides = [1, 1, 1]} : vector<1x1x4xf32> into vector<4x1x4xf32>
    %725 = vector.extract_strided_slice %659 {offsets = [0, 0, 12], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x1x16xf32> to vector<1x1x4xf32>
    %726 = vector.insert_strided_slice %725, %724 {offsets = [3, 0, 0], strides = [1, 1, 1]} : vector<1x1x4xf32> into vector<4x1x4xf32>
    %727 = arith.cmpf ogt, %726, %cst_5 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xf32>
    %728 = arith.cmpf olt, %726, %cst_4 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xf32>
    %729 = arith.select %727, %cst_5, %726 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xi1>, vector<4x1x4xf32>
    %730 = arith.select %728, %cst_4, %729 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xi1>, vector<4x1x4xf32>
    %731 = arith.truncf %730 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xf32> to vector<4x1x4xf16>
    %732 = vector.extract %661[0] : f32 from vector<1xf32>
    %733 = vector.insert %732, %cst_9 [0, 0, 0] : f32 into vector<2x1x16xf32>
    %734 = vector.insert %732, %733 [0, 0, 1] : f32 into vector<2x1x16xf32>
    %735 = vector.insert %732, %734 [0, 0, 2] : f32 into vector<2x1x16xf32>
    %736 = vector.insert %732, %735 [0, 0, 3] : f32 into vector<2x1x16xf32>
    %737 = vector.insert %732, %736 [0, 0, 4] : f32 into vector<2x1x16xf32>
    %738 = vector.insert %732, %737 [0, 0, 5] : f32 into vector<2x1x16xf32>
    %739 = vector.insert %732, %738 [0, 0, 6] : f32 into vector<2x1x16xf32>
    %740 = vector.insert %732, %739 [0, 0, 7] : f32 into vector<2x1x16xf32>
    %741 = vector.insert %732, %740 [0, 0, 8] : f32 into vector<2x1x16xf32>
    %742 = vector.insert %732, %741 [0, 0, 9] : f32 into vector<2x1x16xf32>
    %743 = vector.insert %732, %742 [0, 0, 10] : f32 into vector<2x1x16xf32>
    %744 = vector.insert %732, %743 [0, 0, 11] : f32 into vector<2x1x16xf32>
    %745 = vector.insert %732, %744 [0, 0, 12] : f32 into vector<2x1x16xf32>
    %746 = vector.insert %732, %745 [0, 0, 13] : f32 into vector<2x1x16xf32>
    %747 = vector.insert %732, %746 [0, 0, 14] : f32 into vector<2x1x16xf32>
    %748 = vector.insert %732, %747 [0, 0, 15] : f32 into vector<2x1x16xf32>
    %749 = vector.insert %732, %748 [1, 0, 0] : f32 into vector<2x1x16xf32>
    %750 = vector.insert %732, %749 [1, 0, 1] : f32 into vector<2x1x16xf32>
    %751 = vector.insert %732, %750 [1, 0, 2] : f32 into vector<2x1x16xf32>
    %752 = vector.insert %732, %751 [1, 0, 3] : f32 into vector<2x1x16xf32>
    %753 = vector.insert %732, %752 [1, 0, 4] : f32 into vector<2x1x16xf32>
    %754 = vector.insert %732, %753 [1, 0, 5] : f32 into vector<2x1x16xf32>
    %755 = vector.insert %732, %754 [1, 0, 6] : f32 into vector<2x1x16xf32>
    %756 = vector.insert %732, %755 [1, 0, 7] : f32 into vector<2x1x16xf32>
    %757 = vector.insert %732, %756 [1, 0, 8] : f32 into vector<2x1x16xf32>
    %758 = vector.insert %732, %757 [1, 0, 9] : f32 into vector<2x1x16xf32>
    %759 = vector.insert %732, %758 [1, 0, 10] : f32 into vector<2x1x16xf32>
    %760 = vector.insert %732, %759 [1, 0, 11] : f32 into vector<2x1x16xf32>
    %761 = vector.insert %732, %760 [1, 0, 12] : f32 into vector<2x1x16xf32>
    %762 = vector.insert %732, %761 [1, 0, 13] : f32 into vector<2x1x16xf32>
    %763 = vector.insert %732, %762 [1, 0, 14] : f32 into vector<2x1x16xf32>
    %764 = vector.insert %732, %763 [1, 0, 15] : f32 into vector<2x1x16xf32>
    %765 = arith.mulf %764, %arg3 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf32>
    %766 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4)>()[%thread_id_x]
    %767 = vector.load %subview[%c0, %c0, %766, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %768 = vector.insert_strided_slice %767, %cst_12 {offsets = [0, 0, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %769 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 1)>()[%thread_id_x]
    %770 = vector.load %subview[%c0, %c0, %769, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %771 = vector.insert_strided_slice %770, %768 {offsets = [0, 0, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %772 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 2)>()[%thread_id_x]
    %773 = vector.load %subview[%c0, %c0, %772, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %774 = vector.insert_strided_slice %773, %771 {offsets = [0, 0, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %775 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 3)>()[%thread_id_x]
    %776 = vector.load %subview[%c0, %c0, %775, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %777 = vector.insert_strided_slice %776, %774 {offsets = [0, 0, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %778 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 + 2) floordiv 8)>()[%thread_id_x]
    %779 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 - (((s0 mod 64) floordiv 32 + 2) floordiv 8) * 8 + 2) floordiv 8)>()[%thread_id_x]
    %780 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - (((s0 mod 64) floordiv 32 + 2) floordiv 8) * 32 + 8)>()[%thread_id_x]
    %781 = vector.load %subview[%778, %779, %780, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %782 = vector.insert_strided_slice %781, %777 {offsets = [0, 1, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %783 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 9) floordiv 32)>()[%thread_id_x]
    %784 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 9) floordiv 32) * 32 + 9) floordiv 32)>()[%thread_id_x]
    %785 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 9) floordiv 32) * 32 + 9)>()[%thread_id_x]
    %786 = vector.load %subview[%783, %784, %785, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %787 = vector.insert_strided_slice %786, %782 {offsets = [0, 1, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %788 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 + 5) floordiv 16)>()[%thread_id_x]
    %789 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 - ((((s0 mod 64) floordiv 32) * 2 + 5) floordiv 16) * 16 + 5) floordiv 16)>()[%thread_id_x]
    %790 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 2 + 5) floordiv 16) * 32 + 10)>()[%thread_id_x]
    %791 = vector.load %subview[%788, %789, %790, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %792 = vector.insert_strided_slice %791, %787 {offsets = [0, 1, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %793 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 11) floordiv 32)>()[%thread_id_x]
    %794 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 11) floordiv 32) * 32 + 11) floordiv 32)>()[%thread_id_x]
    %795 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 11) floordiv 32) * 32 + 11)>()[%thread_id_x]
    %796 = vector.load %subview[%793, %794, %795, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %797 = vector.insert_strided_slice %796, %792 {offsets = [0, 1, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %798 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 + 4) floordiv 8)>()[%thread_id_x]
    %799 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 - (((s0 mod 64) floordiv 32 + 4) floordiv 8) * 8 + 4) floordiv 8)>()[%thread_id_x]
    %800 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - (((s0 mod 64) floordiv 32 + 4) floordiv 8) * 32 + 16)>()[%thread_id_x]
    %801 = vector.load %subview[%798, %799, %800, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %802 = vector.insert_strided_slice %801, %797 {offsets = [0, 2, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %803 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 17) floordiv 32)>()[%thread_id_x]
    %804 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 17) floordiv 32) * 32 + 17) floordiv 32)>()[%thread_id_x]
    %805 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 17) floordiv 32) * 32 + 17)>()[%thread_id_x]
    %806 = vector.load %subview[%803, %804, %805, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %807 = vector.insert_strided_slice %806, %802 {offsets = [0, 2, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %808 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 + 9) floordiv 16)>()[%thread_id_x]
    %809 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 - ((((s0 mod 64) floordiv 32) * 2 + 9) floordiv 16) * 16 + 9) floordiv 16)>()[%thread_id_x]
    %810 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 2 + 9) floordiv 16) * 32 + 18)>()[%thread_id_x]
    %811 = vector.load %subview[%808, %809, %810, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %812 = vector.insert_strided_slice %811, %807 {offsets = [0, 2, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %813 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 19) floordiv 32)>()[%thread_id_x]
    %814 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 19) floordiv 32) * 32 + 19) floordiv 32)>()[%thread_id_x]
    %815 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 19) floordiv 32) * 32 + 19)>()[%thread_id_x]
    %816 = vector.load %subview[%813, %814, %815, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %817 = vector.insert_strided_slice %816, %812 {offsets = [0, 2, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %818 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 + 6) floordiv 8)>()[%thread_id_x]
    %819 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 - (((s0 mod 64) floordiv 32 + 6) floordiv 8) * 8 + 6) floordiv 8)>()[%thread_id_x]
    %820 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - (((s0 mod 64) floordiv 32 + 6) floordiv 8) * 32 + 24)>()[%thread_id_x]
    %821 = vector.load %subview[%818, %819, %820, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %822 = vector.insert_strided_slice %821, %817 {offsets = [0, 3, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %823 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 25) floordiv 32)>()[%thread_id_x]
    %824 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 25) floordiv 32) * 32 + 25) floordiv 32)>()[%thread_id_x]
    %825 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 25) floordiv 32) * 32 + 25)>()[%thread_id_x]
    %826 = vector.load %subview[%823, %824, %825, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %827 = vector.insert_strided_slice %826, %822 {offsets = [0, 3, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %828 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 + 13) floordiv 16)>()[%thread_id_x]
    %829 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 - ((((s0 mod 64) floordiv 32) * 2 + 13) floordiv 16) * 16 + 13) floordiv 16)>()[%thread_id_x]
    %830 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 2 + 13) floordiv 16) * 32 + 26)>()[%thread_id_x]
    %831 = vector.load %subview[%828, %829, %830, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %832 = vector.insert_strided_slice %831, %827 {offsets = [0, 3, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %833 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 27) floordiv 32)>()[%thread_id_x]
    %834 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 27) floordiv 32) * 32 + 27) floordiv 32)>()[%thread_id_x]
    %835 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 27) floordiv 32) * 32 + 27)>()[%thread_id_x]
    %836 = vector.load %subview[%833, %834, %835, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %837 = vector.insert_strided_slice %836, %832 {offsets = [0, 3, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %838 = affine.apply affine_map<()[s0] -> (s0 mod 32 + 32)>()[%thread_id_x]
    %839 = vector.load %subview[%c0, %c0, %766, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %840 = vector.insert_strided_slice %839, %837 {offsets = [1, 0, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %841 = vector.load %subview[%c0, %c0, %769, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %842 = vector.insert_strided_slice %841, %840 {offsets = [1, 0, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %843 = vector.load %subview[%c0, %c0, %772, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %844 = vector.insert_strided_slice %843, %842 {offsets = [1, 0, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %845 = vector.load %subview[%c0, %c0, %775, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %846 = vector.insert_strided_slice %845, %844 {offsets = [1, 0, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %847 = vector.load %subview[%778, %779, %780, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %848 = vector.insert_strided_slice %847, %846 {offsets = [1, 1, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %849 = vector.load %subview[%783, %784, %785, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %850 = vector.insert_strided_slice %849, %848 {offsets = [1, 1, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %851 = vector.load %subview[%788, %789, %790, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %852 = vector.insert_strided_slice %851, %850 {offsets = [1, 1, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %853 = vector.load %subview[%793, %794, %795, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %854 = vector.insert_strided_slice %853, %852 {offsets = [1, 1, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %855 = vector.load %subview[%798, %799, %800, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %856 = vector.insert_strided_slice %855, %854 {offsets = [1, 2, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %857 = vector.load %subview[%803, %804, %805, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %858 = vector.insert_strided_slice %857, %856 {offsets = [1, 2, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %859 = vector.load %subview[%808, %809, %810, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %860 = vector.insert_strided_slice %859, %858 {offsets = [1, 2, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %861 = vector.load %subview[%813, %814, %815, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %862 = vector.insert_strided_slice %861, %860 {offsets = [1, 2, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %863 = vector.load %subview[%818, %819, %820, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %864 = vector.insert_strided_slice %863, %862 {offsets = [1, 3, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %865 = vector.load %subview[%823, %824, %825, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %866 = vector.insert_strided_slice %865, %864 {offsets = [1, 3, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %867 = vector.load %subview[%828, %829, %830, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %868 = vector.insert_strided_slice %867, %866 {offsets = [1, 3, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %869 = vector.load %subview[%833, %834, %835, %838] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
    %870 = vector.insert_strided_slice %869, %868 {offsets = [1, 3, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
    %871 = vector.extract %765[0, 0] : vector<16xf32> from vector<2x1x16xf32>
    %872 = vector.extract %870[0, 0] : vector<4xf16> from vector<2x4x4xf16>
    %873 = vector.extract %731[0, 0] : vector<4xf16> from vector<4x1x4xf16>
    %874 = amdgpu.mfma %872 * %873 + %871 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %875 = vector.extract %870[0, 1] : vector<4xf16> from vector<2x4x4xf16>
    %876 = vector.extract %731[1, 0] : vector<4xf16> from vector<4x1x4xf16>
    %877 = amdgpu.mfma %875 * %876 + %874 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %878 = vector.extract %870[0, 2] : vector<4xf16> from vector<2x4x4xf16>
    %879 = vector.extract %731[2, 0] : vector<4xf16> from vector<4x1x4xf16>
    %880 = amdgpu.mfma %878 * %879 + %877 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %881 = vector.extract %870[0, 3] : vector<4xf16> from vector<2x4x4xf16>
    %882 = vector.extract %731[3, 0] : vector<4xf16> from vector<4x1x4xf16>
    %883 = amdgpu.mfma %881 * %882 + %880 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %884 = vector.insert %883, %cst_9 [0, 0] : vector<16xf32> into vector<2x1x16xf32>
    %885 = vector.extract %765[1, 0] : vector<16xf32> from vector<2x1x16xf32>
    %886 = vector.extract %870[1, 0] : vector<4xf16> from vector<2x4x4xf16>
    %887 = amdgpu.mfma %886 * %873 + %885 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %888 = vector.extract %870[1, 1] : vector<4xf16> from vector<2x4x4xf16>
    %889 = amdgpu.mfma %888 * %876 + %887 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %890 = vector.extract %870[1, 2] : vector<4xf16> from vector<2x4x4xf16>
    %891 = amdgpu.mfma %890 * %879 + %889 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %892 = vector.extract %870[1, 3] : vector<4xf16> from vector<2x4x4xf16>
    %893 = amdgpu.mfma %892 * %882 + %891 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
    %894 = vector.insert %893, %884 [1, 0] : vector<16xf32> into vector<2x1x16xf32>
    gpu.barrier
    vector.transfer_write %535, %subview_16[%c0, %c0, %30, %31] {in_bounds = [true, true, true, true]} : vector<1x1x1x8xf16>, memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>
    vector.transfer_write %536, %subview[%c0, %c0, %30, %31] {in_bounds = [true, true, true, true]} : vector<1x1x1x8xf16>, memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>
    scf.yield %641, %718, %894 : vector<1xf32>, vector<1xf32>, vector<2x1x16xf32>
  }
  gpu.barrier
  %35 = vector.load %subview_16[%c0, %c0, %15, %16] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<8xf16>
  %36 = vector.insert_strided_slice %35, %cst_10 {offsets = [0, 0, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
  %37 = vector.load %subview_16[%c0, %c0, %15, %20] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<8xf16>
  %38 = vector.insert_strided_slice %37, %36 {offsets = [0, 1, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
  %39 = vector.load %subview_16[%c0, %c0, %15, %23] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<8xf16>
  %40 = vector.insert_strided_slice %39, %38 {offsets = [0, 2, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
  %41 = vector.load %subview_16[%c0, %c0, %15, %26] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<8xf16>
  %42 = vector.insert_strided_slice %41, %40 {offsets = [0, 3, 0], strides = [1]} : vector<8xf16> into vector<1x4x8xf16>
  %43 = vector.extract_strided_slice %42 {offsets = [0, 0, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %44 = vector.extract_strided_slice %42 {offsets = [0, 0, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %45 = vector.extract_strided_slice %42 {offsets = [0, 1, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %46 = vector.extract_strided_slice %42 {offsets = [0, 1, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %47 = vector.extract_strided_slice %42 {offsets = [0, 2, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %48 = vector.extract_strided_slice %42 {offsets = [0, 2, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %49 = vector.extract_strided_slice %42 {offsets = [0, 3, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %50 = vector.extract_strided_slice %42 {offsets = [0, 3, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %51 = vector.extract_strided_slice %29 {offsets = [0, 0, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %52 = vector.extract_strided_slice %29 {offsets = [0, 0, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %53 = vector.extract_strided_slice %29 {offsets = [0, 1, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %54 = vector.extract_strided_slice %29 {offsets = [0, 1, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %55 = vector.extract_strided_slice %29 {offsets = [0, 2, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %56 = vector.extract_strided_slice %29 {offsets = [0, 2, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %57 = vector.extract_strided_slice %29 {offsets = [0, 3, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %58 = vector.extract_strided_slice %29 {offsets = [0, 3, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x4x8xf16> to vector<1x1x4xf16>
  %59 = vector.extract %43[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %60 = vector.extract %51[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %61 = amdgpu.mfma %59 * %60 + %cst_14 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %62 = vector.extract %44[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %63 = vector.extract %52[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %64 = amdgpu.mfma %62 * %63 + %61 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %65 = vector.extract %45[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %66 = vector.extract %53[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %67 = amdgpu.mfma %65 * %66 + %64 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %68 = vector.extract %46[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %69 = vector.extract %54[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %70 = amdgpu.mfma %68 * %69 + %67 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %71 = vector.extract %47[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %72 = vector.extract %55[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %73 = amdgpu.mfma %71 * %72 + %70 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %74 = vector.extract %48[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %75 = vector.extract %56[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %76 = amdgpu.mfma %74 * %75 + %73 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %77 = vector.extract %49[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %78 = vector.extract %57[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %79 = amdgpu.mfma %77 * %78 + %76 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %80 = vector.extract %50[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %81 = vector.extract %58[0, 0] : vector<4xf16> from vector<1x1x4xf16>
  %82 = amdgpu.mfma %80 * %81 + %79 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %83 = vector.broadcast %82 : vector<16xf32> to vector<1x1x16xf32>
  %84 = vector.extract %34#0[0] : f32 from vector<1xf32>
  %85 = vector.extract %82[0] : f32 from vector<16xf32>
  %86 = vector.insert %85, %cst_7 [0] : f32 into vector<1xf32>
  %87 = vector.extract %82[1] : f32 from vector<16xf32>
  %88 = vector.insert %87, %86 [0] : f32 into vector<1xf32>
  %89 = arith.maximumf %86, %88 : vector<1xf32>
  %90 = vector.extract %82[2] : f32 from vector<16xf32>
  %91 = vector.insert %90, %88 [0] : f32 into vector<1xf32>
  %92 = arith.maximumf %89, %91 : vector<1xf32>
  %93 = vector.extract %82[3] : f32 from vector<16xf32>
  %94 = vector.insert %93, %91 [0] : f32 into vector<1xf32>
  %95 = arith.maximumf %92, %94 : vector<1xf32>
  %96 = vector.extract %82[4] : f32 from vector<16xf32>
  %97 = vector.insert %96, %94 [0] : f32 into vector<1xf32>
  %98 = arith.maximumf %95, %97 : vector<1xf32>
  %99 = vector.extract %82[5] : f32 from vector<16xf32>
  %100 = vector.insert %99, %97 [0] : f32 into vector<1xf32>
  %101 = arith.maximumf %98, %100 : vector<1xf32>
  %102 = vector.extract %82[6] : f32 from vector<16xf32>
  %103 = vector.insert %102, %100 [0] : f32 into vector<1xf32>
  %104 = arith.maximumf %101, %103 : vector<1xf32>
  %105 = vector.extract %82[7] : f32 from vector<16xf32>
  %106 = vector.insert %105, %103 [0] : f32 into vector<1xf32>
  %107 = arith.maximumf %104, %106 : vector<1xf32>
  %108 = vector.extract %82[8] : f32 from vector<16xf32>
  %109 = vector.insert %108, %106 [0] : f32 into vector<1xf32>
  %110 = arith.maximumf %107, %109 : vector<1xf32>
  %111 = vector.extract %82[9] : f32 from vector<16xf32>
  %112 = vector.insert %111, %109 [0] : f32 into vector<1xf32>
  %113 = arith.maximumf %110, %112 : vector<1xf32>
  %114 = vector.extract %82[10] : f32 from vector<16xf32>
  %115 = vector.insert %114, %112 [0] : f32 into vector<1xf32>
  %116 = arith.maximumf %113, %115 : vector<1xf32>
  %117 = vector.extract %82[11] : f32 from vector<16xf32>
  %118 = vector.insert %117, %115 [0] : f32 into vector<1xf32>
  %119 = arith.maximumf %116, %118 : vector<1xf32>
  %120 = vector.extract %82[12] : f32 from vector<16xf32>
  %121 = vector.insert %120, %118 [0] : f32 into vector<1xf32>
  %122 = arith.maximumf %119, %121 : vector<1xf32>
  %123 = vector.extract %82[13] : f32 from vector<16xf32>
  %124 = vector.insert %123, %121 [0] : f32 into vector<1xf32>
  %125 = arith.maximumf %122, %124 : vector<1xf32>
  %126 = vector.extract %82[14] : f32 from vector<16xf32>
  %127 = vector.insert %126, %124 [0] : f32 into vector<1xf32>
  %128 = arith.maximumf %125, %127 : vector<1xf32>
  %129 = vector.extract %82[15] : f32 from vector<16xf32>
  %130 = vector.insert %129, %127 [0] : f32 into vector<1xf32>
  %131 = arith.maximumf %128, %130 : vector<1xf32>
  %132 = vector.bitcast %131 : vector<1xf32> to vector<1xi32>
  %133 = vector.extract %132[0] : i32 from vector<1xi32>
  %shuffleResult, %valid = gpu.shuffle  xor %133, %c32_i32, %c64_i32 : i32
  %134 = vector.broadcast %shuffleResult : i32 to vector<1xi32>
  %135 = vector.bitcast %134 : vector<1xi32> to vector<1xf32>
  %136 = arith.maximumf %135, %131 : vector<1xf32>
  %137 = vector.extract %136[0] : f32 from vector<1xf32>
  %138 = arith.maximumf %137, %84 : f32
  %139 = vector.insert %138, %cst_7 [0] : f32 into vector<1xf32>
  %140 = vector.insert %138, %cst_0 [0, 0, 0] : f32 into vector<1x1x16xf32>
  %141 = vector.insert %138, %140 [0, 0, 1] : f32 into vector<1x1x16xf32>
  %142 = vector.insert %138, %141 [0, 0, 2] : f32 into vector<1x1x16xf32>
  %143 = vector.insert %138, %142 [0, 0, 3] : f32 into vector<1x1x16xf32>
  %144 = vector.insert %138, %143 [0, 0, 4] : f32 into vector<1x1x16xf32>
  %145 = vector.insert %138, %144 [0, 0, 5] : f32 into vector<1x1x16xf32>
  %146 = vector.insert %138, %145 [0, 0, 6] : f32 into vector<1x1x16xf32>
  %147 = vector.insert %138, %146 [0, 0, 7] : f32 into vector<1x1x16xf32>
  %148 = vector.insert %138, %147 [0, 0, 8] : f32 into vector<1x1x16xf32>
  %149 = vector.insert %138, %148 [0, 0, 9] : f32 into vector<1x1x16xf32>
  %150 = vector.insert %138, %149 [0, 0, 10] : f32 into vector<1x1x16xf32>
  %151 = vector.insert %138, %150 [0, 0, 11] : f32 into vector<1x1x16xf32>
  %152 = vector.insert %138, %151 [0, 0, 12] : f32 into vector<1x1x16xf32>
  %153 = vector.insert %138, %152 [0, 0, 13] : f32 into vector<1x1x16xf32>
  %154 = vector.insert %138, %153 [0, 0, 14] : f32 into vector<1x1x16xf32>
  %155 = vector.insert %138, %154 [0, 0, 15] : f32 into vector<1x1x16xf32>
  %156 = arith.subf %83, %155 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>]]} : vector<1x1x16xf32>
  %157 = math.exp2 %156 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [1, 4, 2, 4]>>]]} : vector<1x1x16xf32>
  %158 = arith.subf %34#0, %139 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>]]} : vector<1xf32>
  %159 = math.exp2 %158 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>]]} : vector<1xf32>
  %160 = arith.mulf %159, %34#1 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>>]]} : vector<1xf32>
  %161 = vector.extract %160[0] : f32 from vector<1xf32>
  %162 = vector.extract %157[0, 0, 0] : f32 from vector<1x1x16xf32>
  %163 = vector.insert %162, %cst_7 [0] : f32 into vector<1xf32>
  %164 = vector.extract %157[0, 0, 1] : f32 from vector<1x1x16xf32>
  %165 = vector.insert %164, %163 [0] : f32 into vector<1xf32>
  %166 = arith.addf %163, %165 : vector<1xf32>
  %167 = vector.extract %157[0, 0, 2] : f32 from vector<1x1x16xf32>
  %168 = vector.insert %167, %165 [0] : f32 into vector<1xf32>
  %169 = arith.addf %166, %168 : vector<1xf32>
  %170 = vector.extract %157[0, 0, 3] : f32 from vector<1x1x16xf32>
  %171 = vector.insert %170, %168 [0] : f32 into vector<1xf32>
  %172 = arith.addf %169, %171 : vector<1xf32>
  %173 = vector.extract %157[0, 0, 4] : f32 from vector<1x1x16xf32>
  %174 = vector.insert %173, %171 [0] : f32 into vector<1xf32>
  %175 = arith.addf %172, %174 : vector<1xf32>
  %176 = vector.extract %157[0, 0, 5] : f32 from vector<1x1x16xf32>
  %177 = vector.insert %176, %174 [0] : f32 into vector<1xf32>
  %178 = arith.addf %175, %177 : vector<1xf32>
  %179 = vector.extract %157[0, 0, 6] : f32 from vector<1x1x16xf32>
  %180 = vector.insert %179, %177 [0] : f32 into vector<1xf32>
  %181 = arith.addf %178, %180 : vector<1xf32>
  %182 = vector.extract %157[0, 0, 7] : f32 from vector<1x1x16xf32>
  %183 = vector.insert %182, %180 [0] : f32 into vector<1xf32>
  %184 = arith.addf %181, %183 : vector<1xf32>
  %185 = vector.extract %157[0, 0, 8] : f32 from vector<1x1x16xf32>
  %186 = vector.insert %185, %183 [0] : f32 into vector<1xf32>
  %187 = arith.addf %184, %186 : vector<1xf32>
  %188 = vector.extract %157[0, 0, 9] : f32 from vector<1x1x16xf32>
  %189 = vector.insert %188, %186 [0] : f32 into vector<1xf32>
  %190 = arith.addf %187, %189 : vector<1xf32>
  %191 = vector.extract %157[0, 0, 10] : f32 from vector<1x1x16xf32>
  %192 = vector.insert %191, %189 [0] : f32 into vector<1xf32>
  %193 = arith.addf %190, %192 : vector<1xf32>
  %194 = vector.extract %157[0, 0, 11] : f32 from vector<1x1x16xf32>
  %195 = vector.insert %194, %192 [0] : f32 into vector<1xf32>
  %196 = arith.addf %193, %195 : vector<1xf32>
  %197 = vector.extract %157[0, 0, 12] : f32 from vector<1x1x16xf32>
  %198 = vector.insert %197, %195 [0] : f32 into vector<1xf32>
  %199 = arith.addf %196, %198 : vector<1xf32>
  %200 = vector.extract %157[0, 0, 13] : f32 from vector<1x1x16xf32>
  %201 = vector.insert %200, %198 [0] : f32 into vector<1xf32>
  %202 = arith.addf %199, %201 : vector<1xf32>
  %203 = vector.extract %157[0, 0, 14] : f32 from vector<1x1x16xf32>
  %204 = vector.insert %203, %201 [0] : f32 into vector<1xf32>
  %205 = arith.addf %202, %204 : vector<1xf32>
  %206 = vector.extract %157[0, 0, 15] : f32 from vector<1x1x16xf32>
  %207 = vector.insert %206, %204 [0] : f32 into vector<1xf32>
  %208 = arith.addf %205, %207 : vector<1xf32>
  %209 = vector.bitcast %208 : vector<1xf32> to vector<1xi32>
  %210 = vector.extract %209[0] : i32 from vector<1xi32>
  %shuffleResult_17, %valid_18 = gpu.shuffle  xor %210, %c32_i32, %c64_i32 : i32
  %211 = vector.broadcast %shuffleResult_17 : i32 to vector<1xi32>
  %212 = vector.bitcast %211 : vector<1xi32> to vector<1xf32>
  %213 = arith.addf %212, %208 : vector<1xf32>
  %214 = vector.extract %213[0] : f32 from vector<1xf32>
  %215 = arith.addf %214, %161 : f32
  %216 = vector.extract_strided_slice %157 {offsets = [0, 0, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x1x16xf32> to vector<1x1x4xf32>
  %217 = vector.insert_strided_slice %216, %cst_11 {offsets = [0, 0, 0], strides = [1, 1, 1]} : vector<1x1x4xf32> into vector<4x1x4xf32>
  %218 = vector.extract_strided_slice %157 {offsets = [0, 0, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x1x16xf32> to vector<1x1x4xf32>
  %219 = vector.insert_strided_slice %218, %217 {offsets = [1, 0, 0], strides = [1, 1, 1]} : vector<1x1x4xf32> into vector<4x1x4xf32>
  %220 = vector.extract_strided_slice %157 {offsets = [0, 0, 8], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x1x16xf32> to vector<1x1x4xf32>
  %221 = vector.insert_strided_slice %220, %219 {offsets = [2, 0, 0], strides = [1, 1, 1]} : vector<1x1x4xf32> into vector<4x1x4xf32>
  %222 = vector.extract_strided_slice %157 {offsets = [0, 0, 12], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<1x1x16xf32> to vector<1x1x4xf32>
  %223 = vector.insert_strided_slice %222, %221 {offsets = [3, 0, 0], strides = [1, 1, 1]} : vector<1x1x4xf32> into vector<4x1x4xf32>
  %224 = arith.cmpf ogt, %223, %cst_5 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xf32>
  %225 = arith.cmpf olt, %223, %cst_4 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xf32>
  %226 = arith.select %224, %cst_5, %223 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xi1>, vector<4x1x4xf32>
  %227 = arith.select %225, %cst_4, %226 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xi1>, vector<4x1x4xf32>
  %228 = arith.truncf %227 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  LANEY,  VECTORX], [4, 2, 4]>>]]} : vector<4x1x4xf32> to vector<4x1x4xf16>
  %229 = vector.extract %159[0] : f32 from vector<1xf32>
  %230 = vector.insert %229, %cst_9 [0, 0, 0] : f32 into vector<2x1x16xf32>
  %231 = vector.insert %229, %230 [0, 0, 1] : f32 into vector<2x1x16xf32>
  %232 = vector.insert %229, %231 [0, 0, 2] : f32 into vector<2x1x16xf32>
  %233 = vector.insert %229, %232 [0, 0, 3] : f32 into vector<2x1x16xf32>
  %234 = vector.insert %229, %233 [0, 0, 4] : f32 into vector<2x1x16xf32>
  %235 = vector.insert %229, %234 [0, 0, 5] : f32 into vector<2x1x16xf32>
  %236 = vector.insert %229, %235 [0, 0, 6] : f32 into vector<2x1x16xf32>
  %237 = vector.insert %229, %236 [0, 0, 7] : f32 into vector<2x1x16xf32>
  %238 = vector.insert %229, %237 [0, 0, 8] : f32 into vector<2x1x16xf32>
  %239 = vector.insert %229, %238 [0, 0, 9] : f32 into vector<2x1x16xf32>
  %240 = vector.insert %229, %239 [0, 0, 10] : f32 into vector<2x1x16xf32>
  %241 = vector.insert %229, %240 [0, 0, 11] : f32 into vector<2x1x16xf32>
  %242 = vector.insert %229, %241 [0, 0, 12] : f32 into vector<2x1x16xf32>
  %243 = vector.insert %229, %242 [0, 0, 13] : f32 into vector<2x1x16xf32>
  %244 = vector.insert %229, %243 [0, 0, 14] : f32 into vector<2x1x16xf32>
  %245 = vector.insert %229, %244 [0, 0, 15] : f32 into vector<2x1x16xf32>
  %246 = vector.insert %229, %245 [1, 0, 0] : f32 into vector<2x1x16xf32>
  %247 = vector.insert %229, %246 [1, 0, 1] : f32 into vector<2x1x16xf32>
  %248 = vector.insert %229, %247 [1, 0, 2] : f32 into vector<2x1x16xf32>
  %249 = vector.insert %229, %248 [1, 0, 3] : f32 into vector<2x1x16xf32>
  %250 = vector.insert %229, %249 [1, 0, 4] : f32 into vector<2x1x16xf32>
  %251 = vector.insert %229, %250 [1, 0, 5] : f32 into vector<2x1x16xf32>
  %252 = vector.insert %229, %251 [1, 0, 6] : f32 into vector<2x1x16xf32>
  %253 = vector.insert %229, %252 [1, 0, 7] : f32 into vector<2x1x16xf32>
  %254 = vector.insert %229, %253 [1, 0, 8] : f32 into vector<2x1x16xf32>
  %255 = vector.insert %229, %254 [1, 0, 9] : f32 into vector<2x1x16xf32>
  %256 = vector.insert %229, %255 [1, 0, 10] : f32 into vector<2x1x16xf32>
  %257 = vector.insert %229, %256 [1, 0, 11] : f32 into vector<2x1x16xf32>
  %258 = vector.insert %229, %257 [1, 0, 12] : f32 into vector<2x1x16xf32>
  %259 = vector.insert %229, %258 [1, 0, 13] : f32 into vector<2x1x16xf32>
  %260 = vector.insert %229, %259 [1, 0, 14] : f32 into vector<2x1x16xf32>
  %261 = vector.insert %229, %260 [1, 0, 15] : f32 into vector<2x1x16xf32>
  %262 = arith.mulf %261, %34#2 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf32>
  %263 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4)>()[%thread_id_x]
  %264 = vector.load %subview[%c0, %c0, %263, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %265 = vector.insert_strided_slice %264, %cst_12 {offsets = [0, 0, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %266 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 1)>()[%thread_id_x]
  %267 = vector.load %subview[%c0, %c0, %266, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %268 = vector.insert_strided_slice %267, %265 {offsets = [0, 0, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %269 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 2)>()[%thread_id_x]
  %270 = vector.load %subview[%c0, %c0, %269, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %271 = vector.insert_strided_slice %270, %268 {offsets = [0, 0, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %272 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 3)>()[%thread_id_x]
  %273 = vector.load %subview[%c0, %c0, %272, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %274 = vector.insert_strided_slice %273, %271 {offsets = [0, 0, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %275 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 + 2) floordiv 8)>()[%thread_id_x]
  %276 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 - (((s0 mod 64) floordiv 32 + 2) floordiv 8) * 8 + 2) floordiv 8)>()[%thread_id_x]
  %277 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - (((s0 mod 64) floordiv 32 + 2) floordiv 8) * 32 + 8)>()[%thread_id_x]
  %278 = vector.load %subview[%275, %276, %277, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %279 = vector.insert_strided_slice %278, %274 {offsets = [0, 1, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %280 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 9) floordiv 32)>()[%thread_id_x]
  %281 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 9) floordiv 32) * 32 + 9) floordiv 32)>()[%thread_id_x]
  %282 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 9) floordiv 32) * 32 + 9)>()[%thread_id_x]
  %283 = vector.load %subview[%280, %281, %282, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %284 = vector.insert_strided_slice %283, %279 {offsets = [0, 1, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %285 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 + 5) floordiv 16)>()[%thread_id_x]
  %286 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 - ((((s0 mod 64) floordiv 32) * 2 + 5) floordiv 16) * 16 + 5) floordiv 16)>()[%thread_id_x]
  %287 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 2 + 5) floordiv 16) * 32 + 10)>()[%thread_id_x]
  %288 = vector.load %subview[%285, %286, %287, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %289 = vector.insert_strided_slice %288, %284 {offsets = [0, 1, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %290 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 11) floordiv 32)>()[%thread_id_x]
  %291 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 11) floordiv 32) * 32 + 11) floordiv 32)>()[%thread_id_x]
  %292 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 11) floordiv 32) * 32 + 11)>()[%thread_id_x]
  %293 = vector.load %subview[%290, %291, %292, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %294 = vector.insert_strided_slice %293, %289 {offsets = [0, 1, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %295 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 + 4) floordiv 8)>()[%thread_id_x]
  %296 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 - (((s0 mod 64) floordiv 32 + 4) floordiv 8) * 8 + 4) floordiv 8)>()[%thread_id_x]
  %297 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - (((s0 mod 64) floordiv 32 + 4) floordiv 8) * 32 + 16)>()[%thread_id_x]
  %298 = vector.load %subview[%295, %296, %297, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %299 = vector.insert_strided_slice %298, %294 {offsets = [0, 2, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %300 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 17) floordiv 32)>()[%thread_id_x]
  %301 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 17) floordiv 32) * 32 + 17) floordiv 32)>()[%thread_id_x]
  %302 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 17) floordiv 32) * 32 + 17)>()[%thread_id_x]
  %303 = vector.load %subview[%300, %301, %302, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %304 = vector.insert_strided_slice %303, %299 {offsets = [0, 2, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %305 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 + 9) floordiv 16)>()[%thread_id_x]
  %306 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 - ((((s0 mod 64) floordiv 32) * 2 + 9) floordiv 16) * 16 + 9) floordiv 16)>()[%thread_id_x]
  %307 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 2 + 9) floordiv 16) * 32 + 18)>()[%thread_id_x]
  %308 = vector.load %subview[%305, %306, %307, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %309 = vector.insert_strided_slice %308, %304 {offsets = [0, 2, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %310 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 19) floordiv 32)>()[%thread_id_x]
  %311 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 19) floordiv 32) * 32 + 19) floordiv 32)>()[%thread_id_x]
  %312 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 19) floordiv 32) * 32 + 19)>()[%thread_id_x]
  %313 = vector.load %subview[%310, %311, %312, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %314 = vector.insert_strided_slice %313, %309 {offsets = [0, 2, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %315 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 + 6) floordiv 8)>()[%thread_id_x]
  %316 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32 - (((s0 mod 64) floordiv 32 + 6) floordiv 8) * 8 + 6) floordiv 8)>()[%thread_id_x]
  %317 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - (((s0 mod 64) floordiv 32 + 6) floordiv 8) * 32 + 24)>()[%thread_id_x]
  %318 = vector.load %subview[%315, %316, %317, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %319 = vector.insert_strided_slice %318, %314 {offsets = [0, 3, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %320 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 25) floordiv 32)>()[%thread_id_x]
  %321 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 25) floordiv 32) * 32 + 25) floordiv 32)>()[%thread_id_x]
  %322 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 25) floordiv 32) * 32 + 25)>()[%thread_id_x]
  %323 = vector.load %subview[%320, %321, %322, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %324 = vector.insert_strided_slice %323, %319 {offsets = [0, 3, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %325 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 + 13) floordiv 16)>()[%thread_id_x]
  %326 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 2 - ((((s0 mod 64) floordiv 32) * 2 + 13) floordiv 16) * 16 + 13) floordiv 16)>()[%thread_id_x]
  %327 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 2 + 13) floordiv 16) * 32 + 26)>()[%thread_id_x]
  %328 = vector.load %subview[%325, %326, %327, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %329 = vector.insert_strided_slice %328, %324 {offsets = [0, 3, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %330 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 + 27) floordiv 32)>()[%thread_id_x]
  %331 = affine.apply affine_map<()[s0] -> ((((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 27) floordiv 32) * 32 + 27) floordiv 32)>()[%thread_id_x]
  %332 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 - ((((s0 mod 64) floordiv 32) * 4 + 27) floordiv 32) * 32 + 27)>()[%thread_id_x]
  %333 = vector.load %subview[%330, %331, %332, %15] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %334 = vector.insert_strided_slice %333, %329 {offsets = [0, 3, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %335 = affine.apply affine_map<()[s0] -> (s0 mod 32 + 32)>()[%thread_id_x]
  %336 = vector.load %subview[%c0, %c0, %263, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %337 = vector.insert_strided_slice %336, %334 {offsets = [1, 0, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %338 = vector.load %subview[%c0, %c0, %266, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %339 = vector.insert_strided_slice %338, %337 {offsets = [1, 0, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %340 = vector.load %subview[%c0, %c0, %269, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %341 = vector.insert_strided_slice %340, %339 {offsets = [1, 0, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %342 = vector.load %subview[%c0, %c0, %272, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %343 = vector.insert_strided_slice %342, %341 {offsets = [1, 0, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %344 = vector.load %subview[%275, %276, %277, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %345 = vector.insert_strided_slice %344, %343 {offsets = [1, 1, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %346 = vector.load %subview[%280, %281, %282, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %347 = vector.insert_strided_slice %346, %345 {offsets = [1, 1, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %348 = vector.load %subview[%285, %286, %287, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %349 = vector.insert_strided_slice %348, %347 {offsets = [1, 1, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %350 = vector.load %subview[%290, %291, %292, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %351 = vector.insert_strided_slice %350, %349 {offsets = [1, 1, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %352 = vector.load %subview[%295, %296, %297, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %353 = vector.insert_strided_slice %352, %351 {offsets = [1, 2, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %354 = vector.load %subview[%300, %301, %302, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %355 = vector.insert_strided_slice %354, %353 {offsets = [1, 2, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %356 = vector.load %subview[%305, %306, %307, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %357 = vector.insert_strided_slice %356, %355 {offsets = [1, 2, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %358 = vector.load %subview[%310, %311, %312, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %359 = vector.insert_strided_slice %358, %357 {offsets = [1, 2, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %360 = vector.load %subview[%315, %316, %317, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %361 = vector.insert_strided_slice %360, %359 {offsets = [1, 3, 0], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %362 = vector.load %subview[%320, %321, %322, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %363 = vector.insert_strided_slice %362, %361 {offsets = [1, 3, 1], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %364 = vector.load %subview[%325, %326, %327, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %365 = vector.insert_strided_slice %364, %363 {offsets = [1, 3, 2], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %366 = vector.load %subview[%330, %331, %332, %335] : memref<1x1x32x64xf16, strided<[2176, 2176, 68, 1]>, #gpu.address_space<workgroup>>, vector<1xf16>
  %367 = vector.insert_strided_slice %366, %365 {offsets = [1, 3, 3], strides = [1]} : vector<1xf16> into vector<2x4x4xf16>
  %368 = vector.extract %262[0, 0] : vector<16xf32> from vector<2x1x16xf32>
  %369 = vector.extract %367[0, 0] : vector<4xf16> from vector<2x4x4xf16>
  %370 = vector.extract %228[0, 0] : vector<4xf16> from vector<4x1x4xf16>
  %371 = amdgpu.mfma %369 * %370 + %368 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %372 = vector.extract %367[0, 1] : vector<4xf16> from vector<2x4x4xf16>
  %373 = vector.extract %228[1, 0] : vector<4xf16> from vector<4x1x4xf16>
  %374 = amdgpu.mfma %372 * %373 + %371 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %375 = vector.extract %367[0, 2] : vector<4xf16> from vector<2x4x4xf16>
  %376 = vector.extract %228[2, 0] : vector<4xf16> from vector<4x1x4xf16>
  %377 = amdgpu.mfma %375 * %376 + %374 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %378 = vector.extract %367[0, 3] : vector<4xf16> from vector<2x4x4xf16>
  %379 = vector.extract %228[3, 0] : vector<4xf16> from vector<4x1x4xf16>
  %380 = amdgpu.mfma %378 * %379 + %377 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %381 = vector.insert %380, %cst_9 [0, 0] : vector<16xf32> into vector<2x1x16xf32>
  %382 = vector.extract %262[1, 0] : vector<16xf32> from vector<2x1x16xf32>
  %383 = vector.extract %367[1, 0] : vector<4xf16> from vector<2x4x4xf16>
  %384 = amdgpu.mfma %383 * %370 + %382 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %385 = vector.extract %367[1, 1] : vector<4xf16> from vector<2x4x4xf16>
  %386 = amdgpu.mfma %385 * %373 + %384 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %387 = vector.extract %367[1, 2] : vector<4xf16> from vector<2x4x4xf16>
  %388 = amdgpu.mfma %387 * %376 + %386 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %389 = vector.extract %367[1, 3] : vector<4xf16> from vector<2x4x4xf16>
  %390 = amdgpu.mfma %389 * %379 + %388 {blocks = 1 : i32, k = 8 : i32, m = 32 : i32, n = 32 : i32} blgp =  none : vector<4xf16>, vector<4xf16>, vector<16xf32>
  %391 = vector.insert %390, %381 [1, 0] : vector<16xf32> into vector<2x1x16xf32>
  %392 = vector.insert %215, %cst_9 [0, 0, 0] : f32 into vector<2x1x16xf32>
  %393 = vector.insert %215, %392 [0, 0, 1] : f32 into vector<2x1x16xf32>
  %394 = vector.insert %215, %393 [0, 0, 2] : f32 into vector<2x1x16xf32>
  %395 = vector.insert %215, %394 [0, 0, 3] : f32 into vector<2x1x16xf32>
  %396 = vector.insert %215, %395 [0, 0, 4] : f32 into vector<2x1x16xf32>
  %397 = vector.insert %215, %396 [0, 0, 5] : f32 into vector<2x1x16xf32>
  %398 = vector.insert %215, %397 [0, 0, 6] : f32 into vector<2x1x16xf32>
  %399 = vector.insert %215, %398 [0, 0, 7] : f32 into vector<2x1x16xf32>
  %400 = vector.insert %215, %399 [0, 0, 8] : f32 into vector<2x1x16xf32>
  %401 = vector.insert %215, %400 [0, 0, 9] : f32 into vector<2x1x16xf32>
  %402 = vector.insert %215, %401 [0, 0, 10] : f32 into vector<2x1x16xf32>
  %403 = vector.insert %215, %402 [0, 0, 11] : f32 into vector<2x1x16xf32>
  %404 = vector.insert %215, %403 [0, 0, 12] : f32 into vector<2x1x16xf32>
  %405 = vector.insert %215, %404 [0, 0, 13] : f32 into vector<2x1x16xf32>
  %406 = vector.insert %215, %405 [0, 0, 14] : f32 into vector<2x1x16xf32>
  %407 = vector.insert %215, %406 [0, 0, 15] : f32 into vector<2x1x16xf32>
  %408 = vector.insert %215, %407 [1, 0, 0] : f32 into vector<2x1x16xf32>
  %409 = vector.insert %215, %408 [1, 0, 1] : f32 into vector<2x1x16xf32>
  %410 = vector.insert %215, %409 [1, 0, 2] : f32 into vector<2x1x16xf32>
  %411 = vector.insert %215, %410 [1, 0, 3] : f32 into vector<2x1x16xf32>
  %412 = vector.insert %215, %411 [1, 0, 4] : f32 into vector<2x1x16xf32>
  %413 = vector.insert %215, %412 [1, 0, 5] : f32 into vector<2x1x16xf32>
  %414 = vector.insert %215, %413 [1, 0, 6] : f32 into vector<2x1x16xf32>
  %415 = vector.insert %215, %414 [1, 0, 7] : f32 into vector<2x1x16xf32>
  %416 = vector.insert %215, %415 [1, 0, 8] : f32 into vector<2x1x16xf32>
  %417 = vector.insert %215, %416 [1, 0, 9] : f32 into vector<2x1x16xf32>
  %418 = vector.insert %215, %417 [1, 0, 10] : f32 into vector<2x1x16xf32>
  %419 = vector.insert %215, %418 [1, 0, 11] : f32 into vector<2x1x16xf32>
  %420 = vector.insert %215, %419 [1, 0, 12] : f32 into vector<2x1x16xf32>
  %421 = vector.insert %215, %420 [1, 0, 13] : f32 into vector<2x1x16xf32>
  %422 = vector.insert %215, %421 [1, 0, 14] : f32 into vector<2x1x16xf32>
  %423 = vector.insert %215, %422 [1, 0, 15] : f32 into vector<2x1x16xf32>
  %424 = vector.load %11[%workgroup_id_y, %263] : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>, vector<4xf16>
  %425 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 8)>()[%thread_id_x]
  %426 = vector.load %11[%workgroup_id_y, %425] : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>, vector<4xf16>
  %427 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 16)>()[%thread_id_x]
  %428 = vector.load %11[%workgroup_id_y, %427] : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>, vector<4xf16>
  %429 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 24)>()[%thread_id_x]
  %430 = vector.load %11[%workgroup_id_y, %429] : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>, vector<4xf16>
  %431 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 32)>()[%thread_id_x]
  %432 = vector.load %11[%workgroup_id_y, %431] : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>, vector<4xf16>
  %433 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 40)>()[%thread_id_x]
  %434 = vector.load %11[%workgroup_id_y, %433] : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>, vector<4xf16>
  %435 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 48)>()[%thread_id_x]
  %436 = vector.load %11[%workgroup_id_y, %435] : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>, vector<4xf16>
  %437 = affine.apply affine_map<()[s0] -> (((s0 mod 64) floordiv 32) * 4 + 56)>()[%thread_id_x]
  %438 = vector.load %11[%workgroup_id_y, %437] : memref<20x64xf16, #hal.descriptor_type<storage_buffer>>, vector<4xf16>
  %439 = vector.extract %424[0] : f16 from vector<4xf16>
  %440 = vector.insert %439, %cst_13 [0, 0, 0] : f16 into vector<2x1x16xf16>
  %441 = vector.extract %424[1] : f16 from vector<4xf16>
  %442 = vector.insert %441, %440 [0, 0, 1] : f16 into vector<2x1x16xf16>
  %443 = vector.extract %424[2] : f16 from vector<4xf16>
  %444 = vector.insert %443, %442 [0, 0, 2] : f16 into vector<2x1x16xf16>
  %445 = vector.extract %424[3] : f16 from vector<4xf16>
  %446 = vector.insert %445, %444 [0, 0, 3] : f16 into vector<2x1x16xf16>
  %447 = vector.extract %426[0] : f16 from vector<4xf16>
  %448 = vector.insert %447, %446 [0, 0, 4] : f16 into vector<2x1x16xf16>
  %449 = vector.extract %426[1] : f16 from vector<4xf16>
  %450 = vector.insert %449, %448 [0, 0, 5] : f16 into vector<2x1x16xf16>
  %451 = vector.extract %426[2] : f16 from vector<4xf16>
  %452 = vector.insert %451, %450 [0, 0, 6] : f16 into vector<2x1x16xf16>
  %453 = vector.extract %426[3] : f16 from vector<4xf16>
  %454 = vector.insert %453, %452 [0, 0, 7] : f16 into vector<2x1x16xf16>
  %455 = vector.extract %428[0] : f16 from vector<4xf16>
  %456 = vector.insert %455, %454 [0, 0, 8] : f16 into vector<2x1x16xf16>
  %457 = vector.extract %428[1] : f16 from vector<4xf16>
  %458 = vector.insert %457, %456 [0, 0, 9] : f16 into vector<2x1x16xf16>
  %459 = vector.extract %428[2] : f16 from vector<4xf16>
  %460 = vector.insert %459, %458 [0, 0, 10] : f16 into vector<2x1x16xf16>
  %461 = vector.extract %428[3] : f16 from vector<4xf16>
  %462 = vector.insert %461, %460 [0, 0, 11] : f16 into vector<2x1x16xf16>
  %463 = vector.extract %430[0] : f16 from vector<4xf16>
  %464 = vector.insert %463, %462 [0, 0, 12] : f16 into vector<2x1x16xf16>
  %465 = vector.extract %430[1] : f16 from vector<4xf16>
  %466 = vector.insert %465, %464 [0, 0, 13] : f16 into vector<2x1x16xf16>
  %467 = vector.extract %430[2] : f16 from vector<4xf16>
  %468 = vector.insert %467, %466 [0, 0, 14] : f16 into vector<2x1x16xf16>
  %469 = vector.extract %430[3] : f16 from vector<4xf16>
  %470 = vector.insert %469, %468 [0, 0, 15] : f16 into vector<2x1x16xf16>
  %471 = vector.extract %432[0] : f16 from vector<4xf16>
  %472 = vector.insert %471, %470 [1, 0, 0] : f16 into vector<2x1x16xf16>
  %473 = vector.extract %432[1] : f16 from vector<4xf16>
  %474 = vector.insert %473, %472 [1, 0, 1] : f16 into vector<2x1x16xf16>
  %475 = vector.extract %432[2] : f16 from vector<4xf16>
  %476 = vector.insert %475, %474 [1, 0, 2] : f16 into vector<2x1x16xf16>
  %477 = vector.extract %432[3] : f16 from vector<4xf16>
  %478 = vector.insert %477, %476 [1, 0, 3] : f16 into vector<2x1x16xf16>
  %479 = vector.extract %434[0] : f16 from vector<4xf16>
  %480 = vector.insert %479, %478 [1, 0, 4] : f16 into vector<2x1x16xf16>
  %481 = vector.extract %434[1] : f16 from vector<4xf16>
  %482 = vector.insert %481, %480 [1, 0, 5] : f16 into vector<2x1x16xf16>
  %483 = vector.extract %434[2] : f16 from vector<4xf16>
  %484 = vector.insert %483, %482 [1, 0, 6] : f16 into vector<2x1x16xf16>
  %485 = vector.extract %434[3] : f16 from vector<4xf16>
  %486 = vector.insert %485, %484 [1, 0, 7] : f16 into vector<2x1x16xf16>
  %487 = vector.extract %436[0] : f16 from vector<4xf16>
  %488 = vector.insert %487, %486 [1, 0, 8] : f16 into vector<2x1x16xf16>
  %489 = vector.extract %436[1] : f16 from vector<4xf16>
  %490 = vector.insert %489, %488 [1, 0, 9] : f16 into vector<2x1x16xf16>
  %491 = vector.extract %436[2] : f16 from vector<4xf16>
  %492 = vector.insert %491, %490 [1, 0, 10] : f16 into vector<2x1x16xf16>
  %493 = vector.extract %436[3] : f16 from vector<4xf16>
  %494 = vector.insert %493, %492 [1, 0, 11] : f16 into vector<2x1x16xf16>
  %495 = vector.extract %438[0] : f16 from vector<4xf16>
  %496 = vector.insert %495, %494 [1, 0, 12] : f16 into vector<2x1x16xf16>
  %497 = vector.extract %438[1] : f16 from vector<4xf16>
  %498 = vector.insert %497, %496 [1, 0, 13] : f16 into vector<2x1x16xf16>
  %499 = vector.extract %438[2] : f16 from vector<4xf16>
  %500 = vector.insert %499, %498 [1, 0, 14] : f16 into vector<2x1x16xf16>
  %501 = vector.extract %438[3] : f16 from vector<4xf16>
  %502 = vector.insert %501, %500 [1, 0, 15] : f16 into vector<2x1x16xf16>
  %503 = memref.load %12[] : memref<f32, #hal.descriptor_type<storage_buffer>>
  %504 = arith.divf %cst_3, %423 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>, <[ BATCHY,  LANEX], [1, 32]>>, #iree_vector_ext.layout<<[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>, <[ BATCHY,  LANEX], [1, 32]>>], [#iree_vector_ext.layout<<[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>, <[ BATCHY,  LANEX], [1, 32]>>]]} : vector<2x1x16xf32>
  %505 = arith.mulf %504, %391 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf32>
  %506 = arith.truncf %505 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf32> to vector<2x1x16xf16>
  %507 = arith.mulf %506, %502 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf16>
  %508 = arith.truncf %503 : f32 to f16
  %509 = vector.broadcast %508 : f16 to vector<2x1x16xf16>
  %510 = arith.divf %507, %509 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf16>
  %511 = math.roundeven %510 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf16>
  %512 = arith.cmpf ult, %511, %cst_2 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf16>
  %513 = arith.select %512, %cst_2, %511 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xi1>, vector<2x1x16xf16>
  %514 = arith.cmpf ugt, %513, %cst_1 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf16>
  %515 = arith.select %514, %cst_1, %513 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>, #iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xi1>, vector<2x1x16xf16>
  %516 = arith.fptosi %515 {__vector_layout_fetcher_storage = [[#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>], [#iree_vector_ext.layout<<[ BATCHY,  LANEX], [1, 32]>, <[ BATCHX,  VECTORY,  LANEY,  VECTORX], [2, 4, 2, 4]>>]]} : vector<2x1x16xf16> to vector<2x1x16xi8>
  %517 = vector.extract_strided_slice %516 {offsets = [0, 0, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<2x1x16xi8> to vector<1x1x4xi8>
  %518 = vector.extract %517[0, 0] : vector<4xi8> from vector<1x1x4xi8>
  vector.store %518, %13[%workgroup_id_x, %17, %workgroup_id_y, %263] : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<4xi8>
  %519 = vector.extract_strided_slice %516 {offsets = [0, 0, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<2x1x16xi8> to vector<1x1x4xi8>
  %520 = vector.extract %519[0, 0] : vector<4xi8> from vector<1x1x4xi8>
  vector.store %520, %13[%workgroup_id_x, %17, %workgroup_id_y, %425] : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<4xi8>
  %521 = vector.extract_strided_slice %516 {offsets = [0, 0, 8], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<2x1x16xi8> to vector<1x1x4xi8>
  %522 = vector.extract %521[0, 0] : vector<4xi8> from vector<1x1x4xi8>
  vector.store %522, %13[%workgroup_id_x, %17, %workgroup_id_y, %427] : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<4xi8>
  %523 = vector.extract_strided_slice %516 {offsets = [0, 0, 12], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<2x1x16xi8> to vector<1x1x4xi8>
  %524 = vector.extract %523[0, 0] : vector<4xi8> from vector<1x1x4xi8>
  vector.store %524, %13[%workgroup_id_x, %17, %workgroup_id_y, %429] : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<4xi8>
  %525 = vector.extract_strided_slice %516 {offsets = [1, 0, 0], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<2x1x16xi8> to vector<1x1x4xi8>
  %526 = vector.extract %525[0, 0] : vector<4xi8> from vector<1x1x4xi8>
  vector.store %526, %13[%workgroup_id_x, %17, %workgroup_id_y, %431] : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<4xi8>
  %527 = vector.extract_strided_slice %516 {offsets = [1, 0, 4], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<2x1x16xi8> to vector<1x1x4xi8>
  %528 = vector.extract %527[0, 0] : vector<4xi8> from vector<1x1x4xi8>
  vector.store %528, %13[%workgroup_id_x, %17, %workgroup_id_y, %433] : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<4xi8>
  %529 = vector.extract_strided_slice %516 {offsets = [1, 0, 8], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<2x1x16xi8> to vector<1x1x4xi8>
  %530 = vector.extract %529[0, 0] : vector<4xi8> from vector<1x1x4xi8>
  vector.store %530, %13[%workgroup_id_x, %17, %workgroup_id_y, %435] : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<4xi8>
  %531 = vector.extract_strided_slice %516 {offsets = [1, 0, 12], sizes = [1, 1, 4], strides = [1, 1, 1]} : vector<2x1x16xi8> to vector<1x1x4xi8>
  %532 = vector.extract %531[0, 0] : vector<4xi8> from vector<1x1x4xi8>
  vector.store %532, %13[%workgroup_id_x, %17, %workgroup_id_y, %437] : memref<2x1024x20x64xi8, strided<[1310720, 1280, 64, 1], offset: ?>, #hal.descriptor_type<storage_buffer>>, vector<4xi8>
  return
}