}
    // FusedMatMul kernel does not support grad_a/grad_b attrs
    if ((matmul->hasAttr("grad_a") &&
         mlir::cast<BoolAttr>(matmul->getAttr("grad_a")).getValue()) ||
        (matmul->hasAttr("grad_b") &&
         mlir::cast<BoolAttr>(matmul->getAttr("grad_b")).getValue())) {
      (void)rewriter.notifyMatchFailure(matmul, [&](Diagnostic &diag) {
        diag << "FusedMatMul kernel does not support grad_a/grad_b attrs";

// CHECK: %[[MATMUL:.*]] = "tf.MatMul"(%arg0, %[[DEQUANTIZED]]) <{transpose_a = false, transpose_b = false}> {attr_map = "0:transpose_a,1:transpose_a", device = ""} : (tensor<1x2x2x2xf32>, tensor<2x1024xf32>) -> tensor<*xf32>
// CHECK: return %[[MATMUL]] : tensor<*xf32>

// CHECK-LABEL: func.func private @composite_dequantize_uniform(%arg0: tensor<*xi8>) -> tensor<*xf32>
// CHECK-DAG: %[[SCALE:.*]] = "tf.Const"() <{value = dense<0.0157480314> : tensor<f32>

    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,

// RUN: tf-quant-opt %s -split-input-file -quant-lift-quantizable-spots-as-functions -quant-prepare-quantize-drq -quant-quantize='weight-quantization=true' -verify-each=false | FileCheck %s

// -----

module {
  func.func @matmul(%arg0: tensor<1x2x2x3xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2x1024xf32>} : () -> tensor<2x1024xf32>

  %0 = corert.get_op_handler %arg0 "cpu"
  // CHECK: %[[RESULT:.*]] = corert.executeop(%[[ARG_0:.*]]) "tf.MatMul"(%[[ARG_1:.*]], %[[ARG_1]]) {T = f32, transpose_a = false, transpose_b = false} : 1
  %1 = corert.executeop(%0) "tf.MatMul"(%arg1, %arg1) {T = f32, device = "cpu", transpose_a = false, transpose_b = false} : 1
  tfrt.return %arg0, %1 : !tfrt.chain, !corert.tensorhandle

# RUN: tf_tfl_translate -tf-input-arrays=Placeholder,Placeholder_1 -tf-input-shapes=2,5,3:3,7 -tf-input-data-types=DT_FLOAT,DT_FLOAT -tf-output-arrays=MatMul -unfold_batchmatmul=true -output-mlir %s -o - 2>&1 | FileCheck %s

node {
  name: "Placeholder"
  op: "Placeholder"
  attr {
    key: "dtype"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "shape"
    value {
      shape {
        dim {
          size: 2
        }

// CHECK: %[[IDENTITY:.*]] = "tf.Identity"(%cst) : (tensor<2x1024xi8>) -> tensor<2x1024xi8>
// CHECK: %[[MATMUL:.*]] = "tf.XlaDotV2"(%arg0, %[[IDENTITY]]) <{dimension_numbers = "\12\01\00\0A\01\03", precision_config = ""}> {device = ""} : (tensor<1x2x2x2xbf16>, tensor<2x1024xi8>) -> tensor<1x2x2x1024xbf16>

  // CHECK: tfrt_fallback_async.executeop key(2) cost({{.*}}) device("/device:CPU:0") "tf.MatMul"
  %0 = "tf.ReadVariableOp"(%arg1) {device = "/device:CPU:0", dtype = f32} : (tensor<!tf_type.resource<tensor<1x3xf32>>>) -> tensor<1x3xf32>
  %1 = "tf.MatMul"(%arg0, %0) {T = f32, device = "/device:CPU:0", transpose_a = false, transpose_b = false} : (tensor<3x1xf32>, tensor<1x3xf32>) -> tensor<3x3xf32>

  %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>

  %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>