}

func.func @depthwiseConv2D(tensor<256x32x32x3xf32>, tensor<3x3x3x4xf32>, tensor<256x3x32x32xf32>) -> (tensor<256x30x30x12xf32>, tensor<256x12x30x30xf32>, tensor<256x30x30x12xf32>, tensor<256x30x30x12xf32>) {
^bb0(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<3x3x3x4xf32>, %arg2: tensor<256x3x32x32xf32>) :
   // OK

  %1 = "tf.Sub"(%0, %cst_0) {data_format = "NHWC"} : (tensor<1x3x2x2xf32>, tensor<1x1x1x2xf32>) -> tensor<1x3x2x2xf32>
  %2 = "tf.Mul"(%1, %cst_1) : (tensor<1x3x2x2xf32>, tensor<1x1x1x2xf32>) -> tensor<1x3x2x2xf32>
  %3 = "tf.AddV2"(%2, %cst_2) : (tensor<1x3x2x2xf32>, tensor<1x1x1x2xf32>) -> tensor<1x3x2x2xf32>
  func.return %3 : tensor<1x3x2x2xf32>
}

    %2 = "tf.Relu"(%1) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
    %3 = "tf.FakeQuantWithMinMaxArgs"(%2) {device = "", max = 4.000000e-01 : f32, min = -3.000000e-01 : f32, narrow_range = false, num_bits = 8 : i64} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>

    %6 = "quantfork.qcast"(%5) : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2x!quant.uniform<i8:f32, 0.0054901962186775953:-19>>
    %7 = "quantfork.dcast"(%6) : (tensor<1x3x2x2x!quant.uniform<i8:f32, 0.0054901962186775953:-19>>) -> tensor<1x3x2x2xf32>
    return %7 : tensor<1x3x2x2xf32>
  }

  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 3, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>
  %rst = "tf.Conv2D"(%arg, %fq) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x16xf32>) -> tensor<256x8x7x16xf32>
  func.return %rst : tensor<256x8x7x16xf32>

    : (tensor<?x3x2x1xf32>, tensor<2x1x1x1xf32>) -> tensor<?x2x2x1xf32>
  %1 = chlo.broadcast_add %0, %zp_offset : (
      tensor<?x2x2x1xf32>, tensor<?x2x2x1xf32>) -> tensor<?x2x2x1xf32>
  %2 = chlo.broadcast_add %1, %bias : (
      tensor<?x2x2x1xf32>, tensor<1xf32>) ->tensor<?x2x2x1xf32>
  return %2 : tensor<?x2x2x1xf32>
}

// -----

// CHECK-LABEL: func @dot_general_add_add

  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 5, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>
  %rst = "tf.Conv2D"(%arg, %fq) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x16xf32>) -> tensor<256x8x7x16xf32>
  func.return %rst : tensor<256x8x7x16xf32>

// CHECK{LITERAL}: %cst_1 = arith.constant dense<[[[[1.000000e+00], [1.250000e+00]]]]> : tensor<1x1x2x1xf32>
// CHECK: %3 = tfl.mul %2, %cst_1 {fused_activation_function = "NONE"} : tensor<1x1x2x1xf32>
// CHECK: %cst_2 = arith.constant dense<[0, 3, 1, 2]> : tensor<4xi32>
// CHECK: %4 = "tfl.transpose"(%3, %cst_2) : (tensor<1x1x2x1xf32>, tensor<4xi32>) -> tensor<1x1x1x2xf32>

    has_token_input_output = false, module = "", platforms = [],
    version = 5 : i64
  } : (tensor<1x3x2x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x2x2x2xf32>
  return %0 : tensor<1x2x2x2xf32>
}

// CHECK-LABEL: func.func @qdq_for_conv_weight_empty
// CHECK-SAME: (%[[ARG_0:.+]]: tensor<1x3x2x3xf32>) -> tensor<1x2x2x2xf32>

  %0 = "tf.Less"(%arg0, %arg1) : (tensor<8x7x6x5x?x3x2x1xf32>, tensor<?x3x2x1xf32>) -> tensor<8x7x6x5x?x3x2x1xi1>
  %1 = "tf.SelectV2"(%0, %arg0, %arg1) : (tensor<8x7x6x5x?x3x2x1xi1>, tensor<8x7x6x5x?x3x2x1xf32>, tensor<?x3x2x1xf32>) -> tensor<8x7x6x5x?x3x2x1xf32>
  func.return %1 : tensor<8x7x6x5x?x3x2x1xf32>