%2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x3x2x2xbf16>) -> tensor<1x3x2x2xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  return %3 : tensor<1x3x2x2xf32>
// CHECK: tf.Conv2D
// CHECK-SAME: loc(callsite("Model/conv2d@conv2d_with_valid_loc"("Conv2D") at "QuantizationUnit({{.*}})"))
}

// CHECK: %[[CONV2D:.*]] = "tf.Conv2D"(%arg0, %[[CONST]]) <{data_format = "NHWC", dilations = [1, 1, 2, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true}> : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x2x2xf32>

    %1 = "tf.Conv2D"(%0, %cst) {data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true}
        : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x2x2xf32> loc(fused["Conv2D:", "Model/conv2d"])
    %2 = "tf.IdentityN"(%1) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>

  %11 = "tf.Conv2D"(%8, %arg4)
       {
         data_format = "NCHW",
         dilations = [1, 1, 1, 1],
         explicit_paddings = [],
         padding = "VALID",
         strides = [1, 1, 1, 1]
       } : (tensor<?x64x56x56xf32>, tensor<1x1x64x256xf32>) -> tensor<?x256x56x56xf32>

  // CHECK: %[[CONV2:[0-9]*]] = "tf.Conv2D"(%[[MAX_POOL]], %arg4)
  // CHECK-SAME: data_format = "NHWC"

    }
  }

  // Handle Conv2D input, stride and filter.
  HandleConv2DInput(conv2d, block_size);
  HandleConv2DStride(conv2d);
  HandleConv2DFilter(conv2d, block_size);

  // Book keeping new filter shape for backprop filter rewrite.
  // Filter shape is defined in HandleConv2DFilter, thus it is RankedTensorType.
  filter_shape =
      mlir::cast<RankedTensorType>(conv2d.getFilter().getType()).getShape();

  %4 = "tf.Transpose"(%2, %3) : (tensor<1x32x32x8xf32>, tensor<4xi32>) -> tensor<1x8x32x32xf32>

  // Check that Conv2D computed in NCHW format, and all redundant transpose
  // operations removed from the function.

  // CHECK: %[[CONV:[0-9]*]] = "tf.Conv2D"(%arg0, %arg1)
  // CHECK-SAME: data_format = "NCHW"
  // CHECK-SAME: -> tensor<1x8x32x32xf32>

  // CHECK: return %[[CONV]]

  %1 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_0) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>
  %2 = "tf.Conv2D"(%0, %arg1) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x3xf32>, tensor<5x5x3x8xf32>) -> tensor<4x64x64x8xf32>
  %3 = "tf.Conv2D"(%1, %arg1) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x3xf32>, tensor<5x5x3x8xf32>) -> tensor<4x64x64x8xf32>

    return %3 : tensor<?x?x?x2xf32>
  }

// CHECK-LABEL: func @conv_with_dynamic_shape
// The Conv2D should not be quantized since it has dynamic channel.
// CHECK: "tf.Conv2D"
// CHECK-SAME: (tensor<?x?x?x?xf32>, tensor<2x3x3x2xf32>) -> tensor<?x?x?x2xf32>

  // CHECK: %[[ARG_PERM:.*]] = "tf.Const"() <{value = dense<[0, 3, 1, 2]> : tensor<4xi64>}>
  // CHECK: %[[ARG_TRANSPOSE:[0-9]*]] = "tf.Transpose"(%arg0, %[[ARG_PERM]])

  // CHECK: %[[CONV2D:[0-9]*]] = "tf.Conv2D"(%[[ARG_TRANSPOSE]], %arg1)
  // CHECK-SAME: data_format = "NCHW"
  // CHECK-SAME: dilations = [1, 4, 2, 3]
  // CHECK-SAME: explicit_paddings = [1, 2, 7, 8, 3, 4, 5, 6]
  // CHECK-SAME: padding = "EXPLICIT"

  // cuDNN prefers NCHW data format for spatial convolutions.
  // CHECK: "tf.Conv2D"(%[[INPUT_TRANSPOSE:[0-9]*]], %arg1)
  // CHECK-SAME: data_format = "NCHW"
  %0 = "tf.Conv2D"(%input, %filter)
       {
         data_format = "NHWC",
         padding = "VALID",
         strides = [1, 1, 1, 1]