};

TEST_P(GradientCheckerTest, TestMatMul) {
  float A_vals[] = {1.0f, 2.0f, 3.0f, 4.0f};
  int64_t A_dims[] = {2, 2};
  AbstractTensorHandlePtr A;
  {
    AbstractTensorHandle* A_raw;
    Status s = TestTensorHandleWithDims<float, TF_FLOAT>(ctx_.get(), A_vals,
                                                         A_dims, 2, &A_raw);

  auto ReluGradModel = BuildGradModel(ReluModel, registry_);

  float X_vals[] = {1.0f, 2.0f, 3.0f, -5.0f, -4.0f, -3.0f, 2.0f, 10.0f, -1.0f};
  int64_t X_dims[] = {3, 3};
  AbstractTensorHandlePtr X;
  {
    AbstractTensorHandle* X_raw;
    status_ = TestTensorHandleWithDims<float, TF_FLOAT>(
        immediate_execution_ctx_.get(), X_vals, X_dims, 2, &X_raw);
    ASSERT_EQ(errors::OK, status_.code()) << status_.message();

         TF_TensorByteSize(result_tensor));

  // Build expected result & verify.
  float e_vals[] = {19.0f, 22.0f, 43.0f, 50.0f};

  int data_len = 4;  // length of e_vals
  for (int i = 0; i < data_len; i++) {
    EXPECT_EQ(result_data[i], e_vals[i]);
  }

  TF_DeleteTensor(result_tensor);
  TF_DeleteAbstractTensor(result);
  TF_DeleteOutputList(o);

  AbstractTensorHandlePtr sum_dims;
  {
    vector<int32_t> vals(num_dims_out);
    int64_t vals_shape[] = {num_dims_out};
    Range(&vals, 0, num_dims_out);
    AbstractTensorHandle* sum_dims_raw = nullptr;
    TF_RETURN_IF_ERROR(TestTensorHandleWithDims<int32_t, TF_INT32>(
        ctx, vals.data(), vals_shape, 1, &sum_dims_raw));
    sum_dims.reset(sum_dims_raw);
  }

  bool transpose_a_vals[] = {false, false, true, true};
  bool transpose_b_vals[] = {false, true, false, true};
  float dA_vals[4][9] = {{24, 15, 6, 24, 15, 6, 24, 15, 6},
                         {18, 15, 12, 18, 15, 12, 18, 15, 12},
                         {24, 24, 24, 15, 15, 15, 6, 6, 6},
                         {18, 18, 18, 15, 15, 15, 12, 12, 12}};
  float dB_vals[4][9] = {{12, 12, 12, 15, 15, 15, 18, 18, 18},