auto y = Sub(scope_, x1, x2);
  RunTest({x1, x2}, {x1_shape, x2_shape}, {y}, {x1_shape});
}

TEST_F(NaryGradTest, Mul) {
  TensorShape x1_shape({3, 2, 5});
  TensorShape x2_shape({2, 5});
  auto x1 = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x1_shape));
  auto x2 = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x2_shape));
  auto y = Mul(scope_, x1, x2);

TEST_F(DataFlowGradTest, DynamicStitchGrad) {
  TensorShape d1_shape({2});
  TensorShape d2_shape({2, 2});
  std::vector<Output> indices = {Const(scope_, 2), Const(scope_, {1, 0})};
  std::vector<Output> data = {
      Placeholder(scope_, DT_FLOAT, Placeholder::Shape(d1_shape)),
      Placeholder(scope_, DT_FLOAT, Placeholder::Shape(d2_shape))};
  auto y = DynamicStitch(scope_, indices, data);
  TensorShape y_shape({3, 2});

  auto y = Einsum(scope_, {x}, "ij->ji");
  TensorShape y_shape({3, 2});
  RunTest({x}, {x_shape}, {y}, {y_shape});
}

TEST_F(LinalgGradTest, Einsum_TransposeBroadcast) {
  TensorShape x_shape({3, 2, 3});
  Output x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x_shape));
  auto y = Einsum(scope_, {x}, "...ij->...ji");
  TensorShape y_shape({3, 3, 2});
  RunTest({x}, {x_shape}, {y}, {y_shape});
}

  xs.push_back(Placeholder(scope, DT_DOUBLE, Placeholder::Shape(x_shape)));
  xs.push_back(Placeholder(scope, DT_DOUBLE, Placeholder::Shape(x_shape)));
  auto y = Stack(scope, xs, Stack::Axis(0));
  TensorShape y_shape({2, 1, 2, 3});
  double max_error;
  TF_ASSERT_OK((ComputeGradientError<double, double, double>(
      scope, xs, {x_shape, x_shape}, {y}, {y_shape}, &max_error)));
  EXPECT_LT(max_error, 1e-10);
}

  RunTest(x, x_shape, y, y_shape);
}

TEST_F(ArrayGradTest, DiagPartGrad) {
  TensorShape x_shape({5, 2, 5, 2});
  auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x_shape));
  auto y = DiagPart(scope_, x);
  TensorShape y_shape({5, 2});
  RunTest(x, x_shape, y, y_shape);
}

TEST_F(ArrayGradTest, MatrixDiagGrad) {
  TensorShape x_shape({5, 2});

  auto y = AvgPool3D(scope_, x, ksize, strides, "SAME");
  RunTest(x, x_shape, y, y_shape);
}

TEST_F(NNGradTest, LRN) {
  TensorShape x_shape({1, 1, 2, 1});
  auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x_shape));
  auto y = LRN(scope_, x);
  RunTest(x, x_shape, y, x_shape);
}

TEST_F(NNGradTest, SoftplusGrad) {
  TensorShape shape({3, 7});

  FunctionGradTest() : scope_(Scope::NewRootScope()) {}

  void RunTest(const Output& x, const TensorShape& x_shape, const Output& y,
               const TensorShape& y_shape) {
    TF_ASSERT_OK(scope_.status());
    float max_error;
    auto result = (ComputeGradientError<float, float, float>(
        scope_, {x}, {x_shape}, {y}, {y_shape}, &max_error));
    TF_CHECK_OK(result);
    TF_ASSERT_OK(result);
    EXPECT_LT(max_error, 1e-3);
  }

#endif

// An opaque type corresponding to a shape in tensorflow. In the future,
// we may expose the ABI of TF_Shape for performance reasons.
typedef struct TF_Shape TF_Shape;

// Return a new, unknown rank shape object. The caller is responsible for
// calling TF_DeleteShape to deallocate and destroy the returned shape.
TF_CAPI_EXPORT extern TF_Shape* TF_NewShape();

// Returns the rank of `shape`. If `shape` has unknown rank, returns -1.

#include "tensorflow/c/tf_shape.h"

#include <stdint.h>

#include "tensorflow/c/tf_shape_internal.h"
#include "tensorflow/core/framework/tensor_shape.h"

extern "C" {

TF_Shape* TF_NewShape() {
  return tensorflow::wrap(new tensorflow::PartialTensorShape());
}

int TF_ShapeDims(const TF_Shape* shape) {
  return tensorflow::unwrap(shape)->dims();
}

  // https://www.tensorflow.org/api_docs/cc/class/tensorflow/ops/batch-mat-mul
  int64_t x_row_dim = x_shape[x_shape.size() - 2];
  int64_t x_col_dim = x_shape[x_shape.size() - 1];
  int64_t y_row_dim = y_shape[y_shape.size() - 2];
  int64_t y_col_dim = y_shape[y_shape.size() - 1];
  int64_t out_row_dim = output_shape[output_shape.size() - 2];
  int64_t out_col_dim = output_shape[output_shape.size() - 1];