%7 = "tf.Relu"(%6) : (tensor<?x64x112x112xf32>) -> tensor<?x64x112x112xf32>
  %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]

  TensorShape y_shape({1, 1, 1, 1});
  auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x_shape));
  // Setup window and strides so that we only do one MaxPool.
  const std::vector<int> ksize{1, 2, 2, 1};
  const std::vector<int> strides{1, 2, 2, 1};
  auto y = MaxPool(scope_, x, ksize, strides, "VALID");
  Tensor x_init_value = Tensor(DT_FLOAT, x_shape);
  SetRandomValuesForMaxPooling<float>(&x_init_value);

// CHECK: %[[maxpool:.*]] = "tf.MaxPool"(%[[conv_quant]]) <{data_format = "NHWC", ksize = [1, 2, 2, 1], padding = "VALID", strides = [1, 1, 1, 1]}> : (tensor<*xi8>) -> tensor<*xi8>
// CHECK: %[[dequantize:.*]] = "tf.PartitionedCall"(%[[maxpool]]
// CHECK-SAME: f = @dequantize_i8
// CHECK: return %[[dequantize]]

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):
  filter_width = op.get_attr('filter_width')

  %explicit_paddings = tfr.constant [] -> !tfr.attr
  %data_format = tfr.constant "NHWC" -> !tfr.attr
  %MaxPool = tfr.call @tf__max_pool(%input_, %stride, %filter, %padding, %explicit_paddings, %data_format) : (!tfr.tensor, !tfr.attr, !tfr.attr, !tfr.attr, !tfr.attr, !tfr.attr) -> (!tfr.tensor)
  tfr.return %MaxPool : !tfr.tensor
// CHECK: tf__max_pool
}

// CHECK-LABEL: @tf__cast_float

// CHECK:  %1 = "tf.MaxPool"(%arg0)
// CHECK:  %2 = "tf.MaxPool"(%arg0)
}

func.func @maxPool2DChannelFirst(%arg0: tensor<1x16x6x6xf32>) -> tensor<1x16x1x1xf32> {
  // OK

      scope, op.input(0), op.output(0), grad_inputs[0], ksize, strides, padding,
      internal::MaxPoolGrad::DataFormat(data_format));
  grad_outputs->push_back(dx);
  return scope.status();
}
REGISTER_GRADIENT_OP("MaxPool", MaxPoolGradHelper);

Status MaxPoolGradV2Helper(const Scope& scope, const Operation& op,
                           const std::vector<Output>& grad_inputs,
                           std::vector<Output>* grad_outputs) {

          " RED: All reduction operations."
          " MISC: Mixed operations."
          " PWRED: TF operations that get converted to PW+RED operation in XLA."
          " REDUCEWINDOW: TF operations like MaxPool/AvgPool that get "
          "converted to ReduceWindow in XLA."
          " REDUCEWINDOWPW: Operation that get converted to ReduceWindow + PW "
          "(LRN, LRNGrad)."
          " BN: TF FusedBatchNorm* operations."

// -----

// CHECK-LABEL: maxpool_explicit_padding
func.func @maxpool_explicit_padding(%arg0: tensor<2x12x20x7xi32>) -> tensor<2x3x5x7xi32> {
  // CHECK: tf.MaxPool
  // TODO(b/165938852): need to support explicit padding in max_pool.

  %0 = "tf.MaxPool"(%arg0) {data_format = "NHWC", ksize = [1, 2, 2, 1], padding = "EXPLICIT", strides = [1, 4, 4, 1]} : (tensor<2x12x20x7xi32>) -> tensor<2x3x5x7xi32>

      "MatrixDiagV3",
      "MatrixInverse",
      "MatrixSetDiag",
      "MatrixSetDiagV2",
      "MatrixSetDiagV3",
      "MatrixSolve",
      "MatrixTriangularSolve",
      "MaxPool",
      "MaxPool3D",
      "MaxPool3DGrad",
      "MaxPool3DGradGrad",
      "MaxPoolGrad",
      "MaxPoolGradGrad",
      "MaxPoolGradGradV2",
      "MaxPoolGradV2",
      "MaxPoolV2",