%cst_3 = "tf.Const"() {value = dense<[[[[1.400000e+01]], [[-2.800000e+01]], [[4.200000e+01]]], [[[-5.600000e+01]], [[7.100000e+01]], [[-8.500000e+01]]], [[[9.900000e+01]], [[-1.130000e+02]], [[1.270000e+02]]]]> : tensor<3x3x1x1xf32>} : () -> tensor<3x3x1x1xf32>
  %cst_4 = "tf.Const"() {value = dense<-1.280000e+02> : tensor<f32>} : () -> tensor<f32>
  %cst_5 = "tf.Const"() {value = dense<0.00118110236> : tensor<1xf32>} : () -> tensor<1xf32>

  %cst_0 = arith.constant dense<[1.0, 2.0, 3.0]> : tensor<3xf32>
  %w = arith.constant dense<2.0> : tensor<3x3x3x3xf32>
  %q = "tfl.quantize"(%w) {qtype = tensor<3x3x3x3x!quant.uniform<i8:f32, 0.1:1>>} : (tensor<3x3x3x3xf32>) -> tensor<3x3x3x3x!quant.uniform<i8:f32, 0.1:1>>
  %dq = "tfl.dequantize"(%q) : (tensor<3x3x3x3x!quant.uniform<i8:f32, 0.1:1>>) -> tensor<3x3x3x3xf32>

      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>

  func.return %rst : tensor<256x8x7x16xf32>

// CHECK: %[[CONSTANT0:.*]] = "tf.Const"() <{value = dense<0.000000e+00> : tensor<3x3x3x16xf32>}>
// CHECK: %[[QUANTIZE:.*]] = "quantfork.qcast"(%[[CONSTANT0]]) : (tensor<3x3x3x16xf32>) -> tensor<3x3x3x16x!quant.uniform<i8:f32, 1.000000e+00:-128>>

    %0 = "tf.DepthwiseConv2dNative"(%arg0, %cst_0) {data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 2, 2, 1]} : (tensor<1x3x4x3xf32>, tensor<2x3x3x1xf32>) -> tensor<1x2x2x3xf32>

// CHECK-LABEL: conv2d_backprop_input_with_add
func.func @conv2d_backprop_input_with_add(%arg0: tensor<4xi32>, %arg1: tensor<3x3x1x32xf32>, %arg2: tensor<15x14x14x32xf32>) -> tensor<15x28x28x1xf32> {
  %0 = "tf.Conv2DBackpropInput"(%arg0, %arg1, %arg2) {strides = [1, 2, 2, 1], padding="SAME", dilations=[1, 1, 1, 1]}: (tensor<4xi32>, tensor<3x3x1x32xf32>, tensor<15x14x14x32xf32>) -> tensor<15x28x28x1xf32>

// CHECK:  [[VAL_1:%.*]] = "tfl.reshape"(%2, %[[CST]]) {tac.device = "GPU",  tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>
// CHECK:  [[VAL_2:%.*]] = "tfl.concatenation"([[VAL_0]], [[VAL_1]]) <{axis = 3 : i32, fused_activation_function = "NONE"}> {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x2xf32>

func.func @depthwise_conv_2d_bf16(%arg0 : tensor<256x32x32x3xbf16>, %arg1 : tensor<3x3x3x4xf32>, %arg2 : tensor<256x3x32x32xf32>) -> tensor<256x30x30x12xbf16> {
  %0 = "tf.DepthwiseConv2dNative"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xbf16>, tensor<3x3x3x4xf32>) -> tensor<256x30x30x12xbf16>
  func.return %0 : tensor<256x30x30x12xbf16>

  %3 = "tf.ReadVariableOp"(%arg2) : (tensor<!tf_type.resource<tensor<5xf32>>>) -> tensor<5xf32>
  %4 = "tf.ReadVariableOp"(%arg1) : (tensor<!tf_type.resource<tensor<3x3x1x5xf32>>>) -> tensor<3x3x1x5xf32>
  %5 = "tf.ReadVariableOp"(%arg3) : (tensor<!tf_type.resource<tensor<3920x10xf32>>>) -> tensor<3920x10xf32>

// CHECK: return %1 : tensor<5x2x3x4xf32>

// -----

// CHECK-LABEL: pushTposeBcastNoChange
func.func @pushTposeBcastNoChange(%arg0: tensor<2x3x4x1xf32>) -> tensor<5x2x3x4xf32> {
  %perm = arith.constant dense<[3, 0, 1, 2]> : tensor<4xi32>
  %0 = "tfl.transpose"(%arg0, %perm) : (tensor<2x3x4x1xf32>, tensor<4xi32>) -> tensor<1x2x3x4xf32>
  %cst = arith.constant dense<1.0> : tensor<5x2x3x4xf32>