// IntPerLayer-DAG: "tf.DumpTensor"(%[[m0_1]]) <{enabled = true, file_name = "unquantized_tensor_data.pb", func_name = "matmul2", log_dir_path = "/tmp/dumps/composite_matmul_fn_2", node_name = "MatMul"}> : (tensor<2x2xf32>) -> ()

                                                /*transpose_b=*/op.getAdjY());
    matmuls.emplace_back(matmul.getProduct());
  }

  // Combine the result of each individual MatMul into a rank-3 tensor.
  Type packed_type = RankedTensorType::get(
      {bcast.output_batch_size(), rows, cols}, element_type);
  const auto axis = rewriter.getI64IntegerAttr(0);
  auto pack_op =

    matmul.arg0(1, 0) = 4;
    matmul.arg0(1, 1) = 5;
    matmul.arg0(1, 2) = 6;

    matmul.arg1(0, 0) = 7;
    matmul.arg1(0, 1) = 8;
    matmul.arg1(1, 0) = 9;
    matmul.arg1(1, 1) = 10;
    matmul.arg1(2, 0) = 11;
    matmul.arg1(2, 1) = 12;

    EXPECT_TRUE(matmul.Run());
    EXPECT_EQ(matmul.error_msg(), "");
    const float results[4] = {58, 64, 139, 154};
    for (int i = 0; i < 4; ++i) {

   (IsInt8ElementType $weight),
   (IsConstTensor $weight),
   (IsInt32ElementType $matmul),
   (HasStaticShapeConstraint $weight)],
  [], (addBenefit 10)>;

// Convert Matmul with hybrid inputs (f32 activation/int8 weight) to XlaDotV2
def ConvertTFMatMulToXLADotV2OpWeightOnly : Pat<
  (TF_MatMulOp:$matmul
    $input,
    (TF_MulOp (TF_CastOp (TF_IdentityOp $weight), $truncate1), $scale),

    }
    // 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,

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(