}

// CHECK: func @cast_bf16_matmul_to_fp32
// CHECK-DAG: %[[cst:.*]] = "tf.Const"() <{value = dense<{{.*}}> : tensor<10x2xf32>}> : () -> tensor<10x2xf32>
// CHECK: %[[matmul:.*]] = "tf.MatMul"(%arg0, %[[cst]])
// CHECK: %[[identity:.*]] = "tf.IdentityN"(%[[matmul]])
// CHECK: return %[[identity]] : tensor<1x2xf32>

func.func @cast_bf16_depthwise_conv_to_fp32(%arg0: tensor<1x3x4x3xf32>) -> (tensor<1x2x2x6xf32>) {

// CHECK:    return
//
// CHECK:  func.func @_ifrt_program__func(%arg0: tensor<1x3xf32>, %arg1: tensor<3x1xf32>) -> tensor<1x1xf32>
// CHECK-SAME:      tfrt_ifrt_serving.program_id = [[PROGRAM_ID]] : i64
// CHECK-NEXT:     %0 = "tf.MatMul"(%arg0, %arg1)
// CHECK:          return

    %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>

  // 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>

namespace tensorflow {
namespace ops {

//
Status Neg(AbstractContext* ctx, AbstractTensorHandle* const x, AbstractTensorHandle** y, const char* name = nullptr, const char* raw_device_name = nullptr);

//

// RUN: tf-opt %s -tf-batch-matmul-to-tf-einsum | FileCheck %s

func.func @test_batch_matmul_to_einsum(%arg0: tensor<1x2x3xf32>, %arg1: tensor<3x4xf32>) -> tensor<1x2x4xf32> {
  // CHECK-LABEL: test_batch_matmul_to_einsum
  // CHECK: "tf.Einsum"(%arg0, %arg1) <{equation = "...mk,...kn->...mn"}> : (tensor<1x2x3xf32>, tensor<3x4xf32>) -> tensor<1x2x4xf32>

def Test():

  x = tf.constant([[1.0], [1.0], [1.0]])
  y = tf.compat.v1.get_variable(
      name='y',
      shape=(1, 3),
      initializer=tf.random_normal_initializer(),
      trainable=True)
  r = tf.matmul(x, y)

  tensor_info_x = tf.compat.v1.saved_model.utils.build_tensor_info(x)
  tensor_info_r = tf.compat.v1.saved_model.utils.build_tensor_info(r)

  return {

// 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>

def Test():

  x = tf.constant([[1.0], [1.0], [1.0]])
  y = tf.get_variable(
      name='y',
      shape=(1, 3),
      initializer=tf.random_normal_initializer(),
      trainable=True)
  r = tf.matmul(x, y)

  tensor_info_x = tf.saved_model.utils.build_tensor_info(x)
  tensor_info_r = tf.saved_model.utils.build_tensor_info(r)

  signature_def = tf.saved_model.signature_def_utils.build_signature_def(