// 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]

      state.padding = rewriter.getStringAttr("VALID");
    }
    auto conv_op = static_cast<const ConcreteType *>(this)->createTFLOp(
        &state, rewriter, op->getLoc(), tf_op.getType(), input, filter, bias);

    rewriter.replaceOp(op, conv_op.getResult());
    return success();
  }

  const IntegerAttr intAttrOne;

 private:

  }
  func.func private @composite_conv2d_with_bias_and_relu6_fn_1(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>, %arg2: tensor<2xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {

  }
  func.func private @composite_conv2d_with_bias_and_relu6_fn_2(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>, %arg2: tensor<2xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {

//
//
//   SpaceToBatchND -> Expand -> Conv2D -> Squeeze -> BatchToSpaceND -> BiasAdd
//
//   SpaceToBatchND -> Expand -> Conv2D -> Squeeze -> Pad -> BatchToSpaceND ->
//   BiasAdd
//
//   SpaceToBatchND -> Expand -> Conv2D -> Squeeze -> BiasAdd -> BatchToSpaceND
//
//   SpaceToBatchND -> Conv2D -> Pad -> BatchToSpaceND -> BiasAdd
//
//   SpaceToBatchND -> Conv2D -> BatchToSpaceND -> BiasAdd
//
//

    }
    func.func private @composite_fn_1(%arg0: tensor<1x4x4x3xf32>, %arg1: tensor<3x1x1x3xf32>, %arg2: tensor<3xf32>) -> tensor<1x4x4x3xf32> attributes {tf_quant.composite_function} {

// RUN: tf-opt %s -tf-fused-kernel-matcher | FileCheck %s

//===----------------------------------------------------------------------===//
// Conv2D + BiasAdd + <Activation> fusions.
//===----------------------------------------------------------------------===//

// CHECK-LABEL: conv2DBiasAdd_noActivation
func.func @conv2DBiasAdd_noActivation(%arg0: tensor<128xf32>, %arg1: tensor<1x1x3x128xf32>, %arg2: tensor<8x32x32x3xf32>) -> (tensor<*xf32>) {

  %filter = arith.constant dense<2.0> : tensor<3x3x3x16xf32>
  %bias = arith.constant dense<3.0> : tensor<16xf32>
  %value = arith.constant dense<4.0> : tensor<16xf32>
  %0 = "tf.Conv2D"(%arg, %filter) {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>

  // CHECK: %[[ARG_PERM:.*]] = "tf.Const"() <{value = dense<[0, 2, 3, 1]> : 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 = "NHWC"
  // CHECK-SAME: dilations = [1, 3, 4, 2]
  // CHECK-SAME: explicit_paddings = [1, 2, 5, 6, 7, 8, 3, 4]
  // CHECK-SAME: padding = "EXPLICIT"

      type: DT_INT32
    }
  }
  attr {
    key: "keep_dims"
    value {
      b: false
    }
  }
}
node {
  name: "conv1_1/Conv2D"
  op: "Conv2D"
  input: "conv1_pad_1/Pad"
  input: "conv1_1/Conv2D/Cast"
  device: "/job:localhost/replica:0/task:0/device:GPU:0"
  attr {
    key: "T"
    value {
      type: DT_HALF
    }
  }
  attr {