%arg1: tensor<1x3xf32> {tf_saved_model.index_path = [0]})
      -> (tensor<3x3xf32> {tf_saved_model.index_path = []}) {
  // CHECK: {{%.*}} = corert.get_op_handler %arg0 "/device:GPU:0"
  %2 = "tf.MatMul"(%arg0, %arg1) {T = f32, _output_shapes = ["tfshape$dim { size: 3 } dim { size: 3 }"], device = "/device:GPU:0", transpose_a = false, transpose_b = false} : (tensor<3x1xf32>, tensor<1x3xf32>) -> tensor<3x3xf32>

      @def_function.function
      def matmul(self, input_tensor: core.Tensor) -> Mapping[str, core.Tensor]:
        """Performs a matrix multiplication.

        Depending on self.has_bias and self.activation_fn, it may add a bias
        term or
        go through the activaction function.

        Args:
          input_tensor: Input tensor to matmul with the filter.

        Returns:

  // CHECK: {{%.*}} = tfrt_fallback_async.executeop {{.*}} device("/device:CPU:0") "tf.MatMul"
  // CHECK-SAME: {T = f32, transpose_a = false, transpose_b = false}

  // CHECK-NEXT: [[ch1:%.*]], [[var:%.*]] = tfrt_fallback_async.executeop.seq([[in_chain]]) {{.*}} "tf.ReadVariableOp"([[arg1]])
  // CHECK-NEXT: [[r0:%.*]] = tfrt_fallback_async.executeop {{.*}} "tf.MatMul"([[arg0]], [[var]])
  %2 = "tf.MatMul"(%arg0, %1) {T = f32, _output_shapes = ["tfshape$dim { size: 3 } dim { size: 3 }"], device = "/device:CPU:0", transpose_a = false, transpose_b = false} : (tensor<3x1xf32>, tensor<1x3xf32>) -> tensor<3x3xf32>

  // CHECK: %[[TRANSPOSE:.*]] = "tf.Transpose"(%[[DEQUANT]], %[[CST]]) : (tensor<3x4xf32>, tensor<?xi32>) -> tensor<*xf32>
  // CHECK: %[[MATMUL:.*]] = "tf.MatMul"(%arg0, %[[TRANSPOSE]]) <{grad_a = false, grad_b = false, transpose_a = false, transpose_b = true}> : (tensor<2x3xf32>, tensor<*xf32>) -> tensor<2x4xf32>
  // CHECK: return %[[MATMUL]] : tensor<2x4xf32>
}

  auto matmul = rewriter.create<TFL::BatchMatMulOp>(
      loc, RankedTensorType::get(matmul_shape, result_type.getElementType()),
      lhs_flattend, rhs_flattend, /*adj_x*/ false_attr, /*adj_y*/ false_attr,
      /*asym_quant_input*/ false_attr);
  if (result_type.hasStaticShape()) {
    auto reshaped =
        rewriter.create<mhlo::ReshapeOp>(loc, result_type, matmul.getResult());
    return reshaped.getResult();

  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_Op* matmul = MatMulOp(ctx, h0_task1, h1_task1);
  TFE_OpSetDevice(matmul, remote_device_name, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_TensorHandle* retvals[1];
  int num_retvals = 1;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

///     int idx = 3;
///     auto b = Variable(linear.WithOpName("b_", idx),
///                       {2}, DT_FLOAT);
///     auto x = Const(linear, {...});  // name: "linear/Const"
///     auto m = MatMul(linear, x, W);  // name: "linear/MatMul"
///     auto r = BiasAdd(linear, m, b); // name: "linear/BiasAdd"
///
/// Scope lifetime:
///
/// A new scope is created by calling Scope::NewRootScope. This creates some

// Run optimize-batch-matmul pass only and check the results.
// RUN: tf-opt %s -tfl-optimize-batch-matmul | FileCheck %s

// CHECK-LABEL: FuseTransposeFCRhsToBatchMatmul
func.func @FuseTransposeFCRhsToBatchMatmul(%arg0: tensor<16x1024xf32>, %arg1: tensor<1024x128xf32>, %arg2: none) -> tensor<16x128xf32> {
  %cst = arith.constant dense<[1, 0]> : tensor<2xi32>
  %0 = "tfl.transpose"(%arg1, %cst) : (tensor<1024x128xf32>, tensor<2xi32>) -> tensor<128x1024xf32>

==============================================================================*/

// This transformation pass decomposes dense operations that assume
// support for hybrid quantization. These cases cover when a dense operation
// (e.g. matmul) has both quantized and unquantized inputs by dequantizing
// the quantized inputs, performing the operation in the expressed type, then
// requantizing if a quantized output is required.
//