// Use identity op to avoid the filter being constant-folded.
    %identity = "tf.Identity"(%filter) : (tensor<*xi8>) -> tensor<*xi8>
    %2 = "tf.Cast"(%identity) {Truncate = false} : (tensor<*xi8>) -> tensor<*xf32>
    %3 = "tf.MatMul"(%input, %2) {
      attr_map = "transpose_a:0,transpose_b:1"
    } : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
    func.return %3 : tensor<*xf32>
  }

  func.func private @internal_conv2d_fn(

  TFE_TensorHandle* m = TestMatrixTensorHandle(ctx);
  TFE_Op* matmul = MatMulOp(ctx, m, m);
  TFE_TensorHandle* retvals[2] = {nullptr, nullptr};
  int num_retvals = 2;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);
  EXPECT_EQ(1, num_retvals);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteOp(matmul);
  TFE_DeleteTensorHandle(m);

// 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 {

    %zp_fp32 = "tf.Cast"(%zp_from_max) : (tensor<1xf64>) -> tensor<1xf32>
    %zp = "tf.Cast"(%zp_fp32) : (tensor<1xf32>) -> tensor<1xi32>

    func.return %scale, %zp : tensor<1xf32>, tensor<1xi32>
  }

  // Matmul with int32 accumulation
  func.func private @internal_matmul_fn(
                         %input : tensor<*xi8>, %filter : tensor<*xi8>,
                         %input_scale : tensor<*xf32>, %input_zp : tensor<*xi32>,

std::unique_ptr<OperationPass<mlir::func::FuncOp>> CreateOptimizePass();

// Creates an instance of the ReplaceCastHacksWithTFXLAOpsPass, which will
// replace mixed-type convolution and matmul cast hacks by XLA Conv2DOp and
// MatmulOp.
std::unique_ptr<OperationPass<func::FuncOp>>
CreateReplaceCastHacksWithTFXLAOpsPass();

// Creates a pass that moves & merges initializer function's ops into the @main

(TODO)

## Authoring Op Composition in Python

The composable TF provides a single API to define a new op with its composition
at the same time. For example, the following code defines a new
`FusedFullyConnected` op, which have `MatMul`, `Add` and some
`activation function` (specified by an op attribute) fused.


```python
import tensorflow as tf

@Composite(
    'FusedFullyConnected',
    inputs=['input_: T', 'filter_: T', 'bias: T'],

///     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

                   absl::Span<AbstractTensorHandle* const> inputs,
                   absl::Span<AbstractTensorHandle*> outputs) {
  return ops::MatMul(ctx, inputs[0], inputs[1], &outputs[0],
                     /*transpose_a=*/false,
                     /*transpose_b=*/false, "MatMul");
}

Status MulModel(AbstractContext* ctx,
                absl::Span<AbstractTensorHandle* const> inputs,

                           std::vector<Output>* grad_outputs);

/// GradOpRegistry maintains a static registry of gradient functions.
/// Gradient functions are indexed in the registry by the forward op name (i.e.
/// "MatMul" -> MatMulGrad func).
class GradOpRegistry {
 public:
  /// Registers 'func' as the gradient function for 'op'.
  /// Returns true if registration was successful, check fails otherwise.