// done in parallel.
  //
  // This graph is:
  // (Const0, Const0) -> MatMul0
  // (Const1, Const1) -> MatMul1
  // (MatMul0, MatMul1) -> MatMulCombined
  //
  // Device0: [Const0, Const0, MatMul0]
  // Device1: [Const1, Const1, MatMul1, MatMulCombined]
  //
  // Cluster0: [Const0, Const0, MatMul0]
  // Cluster1: [Const1, Const1, MatMul1]
  // Cluster2: [MatMulCombined]

// WholeModel-DAG: "tf.DumpTensor"(%[[m1]]) <{enabled = true, file_name = "unquantized_tensor_data.pb", func_name = "matmul2", log_dir_path = "/tmp/dumps/composite_matmul_fn_1", node_name = "MatMul_1"}
// WholeModel-DAG: return %[[m1]]

// IntPerLayer-LABEL: func @matmul2

  std::vector<Value> sliced_rhs =
      sliceInput(input_rhs, bcast.y_batch_size(), loc, rewriter);

  // Compute (single batch) MatMul for each output batch.
  std::vector<Value> matmuls;
  matmuls.reserve(bcast.output_batch_size());
  for (int batch_idx : llvm::seq<int>(0, bcast.output_batch_size())) {
    int lhs_batch_idx, rhs_batch_idx;
    if (bcast.IsBroadcastingRequired()) {

    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";
      });
      return false;
    }

Christian Sigg <******@****.***> 1714640622 -0700

      "/job:localhost/replica:0/task:0/device:CPU:0");
  Output a = ops::Placeholder(root.WithOpName("A"), DT_FLOAT);
  for (int i = 0; i < 5; i++) {
    a = ops::MatMul(root.WithOpName(absl::StrCat("matmul_", i)), a, a);
    a = ops::Add(root.WithOpName(absl::StrCat("add_", i)), a, a);
  }

  GraphDef graph_def;
  root.graph()->ToGraphDef(&graph_def);
  return graph_def;
}

                                   "gradients/MatMul", false, true);
    TF_Operation* matmul2 = MatMul(expected_graph_, s_, const0, const3,
                                   "gradients/MatMul_1", true, false);
    expected_grad_outputs[0] = {matmul1, 0};
    expected_grad_outputs[1] = {matmul2, 0};
  }

  TF_Tensor* FloatTensor2x2(const float* values) {

# RUN: tf_tfl_translate -unfold_batchmatmul=false -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 -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

  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);

  }
  return ConstantFoldOpIfPossible(value.getDefiningOp()).front();
}

// Matches convolution op with "NHWC" data format or matmul op with false adj_y.
// The list of supported ops in this function is:
// - Conv2DOp
// - Conv3DOp
// - DepthwiseConv2dNativeOp
// - MatMulOp
// - BatchMatMulV2Op
LogicalResult MatchSupportedAffineOp(Operation* op, Value& binding_output,