outputs=['o: T'])
def _composite_max_pool(input_, stride_w, stride_h, filter_width, filter_height,
                        padding):
  ksize = [1, filter_width, filter_height, 1]
  strides = [1, stride_w, stride_h, 1]
  return tf.raw_ops.MaxPool(
      input=input_, ksize=ksize, strides=strides, padding=padding)


@tf.RegisterGradient('NewMaxPool')
def _max_pool_grad(op: ops.Operation, grad):

  // CHECK: %[[R1:.*]] = "tf.MaxPool"(%[[R0]]) <{data_format = "NHWC", explicit_paddings = [], ksize = [1, 3, 3, 1], padding = "SAME", strides = [1, 2, 2, 1]}>
  // CHECK: %[[CST_0:.*]] = "tf.Const"() <{value = dense<[0, 3, 1, 2]> : tensor<4xi64>}>
  // CHECK: "tf.Transpose"(%[[R1]], %[[CST_0]])
  %0 = "tf.MaxPool"(%arg0)
       {
         data_format = "NCHW",
         ksize = [1, 1, 3, 3],
         padding = "SAME",
         strides = [1, 1, 2, 2]

  %1 = "quantfork.dcast"(%0) : (tensor<*x!quant.uniform<i8:f32, 5.000000e-02:-10>>) -> tensor<*xf32>
  %2 = "tf.MaxPool"(%1) {data_format = "NHWC", device = "", explicit_paddings = [], ksize = [1, 2, 2, 1], padding = "VALID", strides = [1, 2, 2, 1]} : (tensor<*xf32>) -> tensor<*xf32>
  %3 = "quantfork.qcast"(%2) {volatile} : (tensor<*xf32>) -> tensor<*x!quant.uniform<i8:f32, 5.000000e-02:-10>>

    %0:2 = tf_executor.graph {
      %outputs:2, %control = tf_executor.island wraps "tf.MaxPoolWithArgmax"(%arg0) {T = f32, Targmax = i32, include_batch_in_index = false, ksize = [1, 2, 2, 1], padding = "SAME", strides = [1, 2, 2, 1]} : (tensor<2x4x2x2xf32>) -> (tensor<2x2x1x2xf32>, tensor<2x2x1x2xi32>)
      tf_executor.fetch %outputs#1, %outputs#0 : tensor<2x2x1x2xi32>, tensor<2x2x1x2xf32>
    }

  %8 = "tf.MaxPool"(%7)
       {
         data_format = "NCHW",
         ksize = [1, 1, 3, 3],
         padding = "SAME",
         strides = [1, 1, 2, 2]
       } : (tensor<?x64x112x112xf32>) -> tensor<?x64x56x56xf32>

  // CHECK: %[[MAX_POOL:[0-9]*]] = "tf.MaxPool"
  // CHECK-SAME: data_format = "NHWC"
  // CHECK-SAME: ksize = [1, 3, 3, 1]
  // CHECK-SAME: strides = [1, 2, 2, 1]

			t.Errorf("DecodeRune(%q) = %#04x, want %#04x", s[i:], r1, r)
			return
		}
		r2, size2 := DecodeRuneInString(s[i:])
		if r != r2 {
			t.Errorf("DecodeRuneInString(%q) = %#04x, want %#04x", s[i:], r2, r)
			return
		}
		if size1 != size2 {
			t.Errorf("DecodeRune/DecodeRuneInString(%q) size mismatch %d/%d", s[i:], size1, size2)
			return
		}
		si += size1
	}
	j--
	for si = len(s); si > 0; {

	var dum [4]byte
	size1 := 0
	for _, r := range a {
		size1 += encoderune(dum[:], r)
	}
	s, b := rawstringtmp(buf, size1+3)
	size2 := 0
	for _, r := range a {
		// check for race
		if size2 >= size1 {
			break
		}
		size2 += encoderune(b[size2:], r)
	}
	return s[:size2]
}

type stringStruct struct {
	str unsafe.Pointer
	len int
}

    %1 = "quantfork.dcast"(%0) : (tensor<*x!quant.uniform<i8:f32, 0.05:-10>>) -> tensor<*xf32>
    %2 = "tf.MaxPool"(%1) {
      data_format = "NHWC", device = "", explicit_paddings = [],
      ksize = [1, 2, 2, 1], padding = "VALID", strides = [1, 2, 2, 1]
    } : (tensor<*xf32>) -> tensor<*xf32>
    %3 = "tf.Reshape"(%2, %cst) {device = ""} : (tensor<*xf32>, tensor<2xi32>) -> tensor<*xf32>

  %1 = "quantfork.dcast"(%0) : (tensor<*x!quant.uniform<i8:f32, 5.000000e-02:-10>>) -> tensor<*xf32>
  %2 = "tf.MaxPool"(%1) {data_format = "NHWC", device = "", explicit_paddings = [], ksize = [1, 2, 2, 1], padding = "VALID", strides = [1, 2, 2, 1]} : (tensor<*xf32>) -> tensor<*xf32>
  %3 = "quantfork.qcast"(%2) {volatile} : (tensor<*xf32>) -> tensor<*x!quant.uniform<i8:f32, 5.000000e-02:-10>>

      CompileTimeConstantOperand<TF::MaxOp, 1>,  // $reduction_indices
      // $ksize, $strides
      CompileTimeConstantOperand<TF::MaxPoolGradGradV2Op, 3, 4>,
      // $ksize, $strides
      CompileTimeConstantOperand<TF::MaxPoolGradV2Op, 2, 3>,
      CompileTimeConstantOperand<TF::MaxPoolV2Op, 1, 2>,   // $ksize, $strides
      CompileTimeConstantOperand<TF::MeanOp, 1>,           // $reduction_indices