stablehlo.return %2 : tensor<f32>
  }) {padding = dense<[[0, 0], [1, 1], [1, 1], [0, 0]]> : tensor<4x2xi64>, window_dimensions = array<i64: 1, 3, 3, 1>} : (tensor<2x3x1x3xf32>, tensor<f32>) -> tensor<2x3x1x3xf32>
  return %1 : tensor<2x3x1x3xf32>
}

// -----

// CHECK-LABEL: @bitcast_convert_i32_f32(%arg0: tensor<1x256128xi32>) -> tensor<1x256128xbf16>

  %cst = arith.constant dense<1.0> : tensor<4xf32>
  %cst_3 = arith.constant dense<2.0> : tensor<4x3x3x3xf32>
  %0 = "tfl.conv_2d"(%arg0, %cst_3, %cst) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "RELU6", padding = "VALID", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<?x19x19x3xf32>, tensor<4x3x3x3xf32>, tensor<4xf32>) -> tensor<?x9x9x4xf32>
  func.return %0 : tensor<?x9x9x4xf32>

  }

  func.func private @composite_conv_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32> attributes {_from_xla_call_module} {
    %0 = stablehlo.convolution(%arg0, %arg1) dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f], window = {pad = [[0, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32>
    return %0 : tensor<1x3x4x2xf32>
  }

func.func @FoldTrivialReshapeIntoTranspose(%arg: tensor<2x1x3x3xf32>) -> tensor<1x3x3x2xf32> {
  %cst = arith.constant dense<[1, 3, 3, 2]> : tensor<4xi32>
  %cst_2 = arith.constant dense<[2, 3, 0, 1]> : tensor<4xi32>
  %2 = "tf.Transpose"(%arg, %cst_2) : (tensor<2x1x3x3xf32>, tensor<4xi32>) -> tensor<3x3x2x1xf32>
  %3 = "tf.Reshape"(%2, %cst) : (tensor<3x3x2x1xf32>, tensor<4xi32>) -> tensor<1x3x3x2xf32>
  return %3: tensor<1x3x3x2xf32>

    %16 = stablehlo.subtract %10, %15 : tensor<1x3x3x4xf32>
    %17 = stablehlo.broadcast_in_dim %4, dims = [0, 1, 2, 3] : (tensor<1x1x1x4xf32>) -> tensor<1x3x3x4xf32>
    %18 = stablehlo.multiply %16, %17 : tensor<1x3x3x4xf32>

    %0 = "quantfork.stats"(%arg0) {layerStats = dense<[1.27501142, 149.824783]> : tensor<2xf32>} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32>

        func.return %2: tensor<1x3x3x4xf32>
      }

      func.func @conv_1_fn(%arg0: tensor<1x3x3x4xf32>) -> tensor<1x3x3x4xf32> {
        %0 = stablehlo.constant dense<2.000000e+00> : tensor<3x3x4x4xf32>

  %q_weight = "quantfork.qcast"(%weight) : (tensor<2x3x3x2xf32>) -> tensor<2x3x3x2x!quant.uniform<i8:f32, 0.074855112561992565:-1>>
  %dq_weight = "quantfork.dcast"(%q_weight) : (tensor<2x3x3x2x!quant.uniform<i8:f32, 0.074855112561992565:-1>>) -> tensor<2x3x3x2xf32>

  func.return %0 : tensor<256x32x32x16xf32>
}

func.func @testConv2DGPU(tensor<256x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<256x32x32x16xf32> {
^bb0(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<16x3x3x3xf32>, %arg2: tensor<16xf32>):
  // CHECK: tac.cost = 0x4C300000

  %cst_0 = arith.constant dense<1.000000e+00> : tensor<f32>
  %cst_1 = arith.constant dense<0.000000e+00> : tensor<8xf32>
  %cst_2 = arith.constant dense<0.000000e+00> : tensor<8x3x3x3xf32>
  %0 = "tfl.sub"(%arg0, %cst_0) {fused_activation_function = "NONE"} : (tensor<3x15x14x3xf32>, tensor<f32>) -> tensor<3x15x14x3xf32>