std::string name_grad_B = "MatMul_Grad_B";
    if (!t_a && !t_b) {
      TF_RETURN_IF_ERROR(MatMul(ctx, upstream_grad, B.get(), &matmul_A_output,
                                /*transpose_a = */ false,
                                /*transpose_b = */ true, name_grad_A.c_str()));

      TF_RETURN_IF_ERROR(MatMul(ctx, A.get(), upstream_grad, &matmul_B_output,
                                /*transpose_a = */ true,

module {
  // TODO(b/260020937): Support transpose_a, transpose_b for matmul.
  func.func @matmul(%arg0: tensor<2x12xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<12x2xf32>} : () -> tensor<12x2xf32>

  %out_matrix = "tf.MatMul"(%in_matrix, %matrix)  : (tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<3x3xf32>
  %in_matrix1 = "tf.TensorArrayReadV3"(%handle_2, %loop_count, %flow_in_2) : (tensor<?x!tf_type.resource>, tensor<i32>, tensor<*xf32>) -> tensor<3x3xf32>
  %out_matrix1 = "tf.MatMul"(%out_matrix, %matrix_2)  : (tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<3x3xf32>

        out = math_ops.matmul(input_tensor, self.filters, name='sample/matmul')
        if bias_fn is not None:
          out = bias_fn(out, self.bias)
        if activation_fn is not None:
          out = activation_fn(out)
        return {'output': out}

    model = MatmulModel(weight_shape)
    saved_model_save.save(
        model,
        saved_model_path,
        signatures=model.matmul.get_concrete_function(

// CHECK-LABEL: private @composite_matmul_with_bias_and_relu6_fn_1
// CHECK-NEXT: %[[matmul:.*]] = "tf.MatMul"(%arg0, %arg1)
// CHECK-SAME: attr_map = "0:transpose_a,1:transpose_b"
// CHECK-NEXT: tf.BiasAdd
// CHECK-NEXT: tf.Relu6
// CHECK-NEXT: return

// CHECK-LABEL: private @composite_matmul_with_bias_and_relu_fn_1
// CHECK-NEXT: tf.MatMul"(%arg0, %arg1)
// CHECK-SAME: attr_map = "0:transpose_a,1:transpose_b"

              absl::Span<AbstractTensorHandle* const> inputs,
              absl::Span<AbstractTensorHandle*> outputs) -> Status {
        return ops::MatMul(ctx, inputs[0], inputs[1], &outputs[0], transpose_a,
                           transpose_b, "MatMul");
      };
      ASSERT_NO_FATAL_FAILURE(CompareNumericalAndAutodiffGradients(
          MatMulModel, BuildGradModel(MatMulModel, registry_),

module {
  // TODO(b/260020937): Support transpose_a, transpose_b for matmul.
  func.func @matmul(%arg0: tensor<2x12xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<12x2xf32>} : () -> tensor<12x2xf32>

  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<512x512xf32>} : () -> tensor<512x512xf32>
  %out_1 = "tf.MatMul"(%arg0, %cst) {
    device = "", transpose_a = false, transpose_b = false
  } : (tensor<1x12x12x512xf32>, tensor<512x512xf32>) -> tensor<*xf32>
  %out_2 = "tf.MatMul"(%arg0, %arg0) {
    device = "", transpose_a = false, transpose_b = true
  } : (tensor<1x12x12x512xf32>, tensor<1x12x12x512xf32>) -> tensor<*xf32>

    (TF_SubOp $beta, (TF_MulOp $m, $mul)))>;

class TFi32<int v> : ConstantAttr<I32ElementsAttr, !cast<string>(v)>;

// Matmul without transpose on b to matmul with explicit transpose op and
// transposed b.
def ConvertMatmulWithoutTransposeToWithTranspose :
      Pat<(TF_MatMulOp $a, $b, ConstBoolAttrFalse:$at, ConstBoolAttrFalse, $grad_a, $grad_b),

//   outer dimension of "b" (after being transposed if transposed_b is
//   true).
//
//   *Note*: The default kernel implementation for MatMul on GPUs uses
//   cublas.
Status MatMul(AbstractContext* ctx, AbstractTensorHandle* const a,
              AbstractTensorHandle* const b, AbstractTensorHandle** product,
              bool transpose_a, bool transpose_b, const char* name,
              const char* raw_device_name) {