// 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

// WholeModel-DAG: "tf.DumpTensor"(%[[matmul0_q]]) <{enabled = true, file_name = "unquantized_tensor_data.pb", func_name = "composite_dot_general_with_bias_and_relu6_dynamic_fn_2", log_dir_path = "/tmp/dumps/composite_dot_general_with_bias_and_relu6_dynamic_fn_2", node_name = "_empty_node"}> : (tensor<?x2xf32>) -> ()

  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()) {

// Performs a fusion of the following pattern(s), if possible:
//   MatMulOp + BiasAdd + <Activation> -> _FusedMatMulOp
class FuseMatMulBiasAdd
    : public FuseContractionWithBiasAdd<MatMulOp, _FusedMatMulOp> {
  using FuseContractionWithBiasAdd<MatMulOp,
                                   _FusedMatMulOp>::FuseContractionWithBiasAdd;

  bool AreFuseCompatible(MatMulOp matmul, BiasAddOp bias_add,

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) {

  }
  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,

        }

        const int x_row =
            matmul_op.getAdjX() ? shape_x.back() : *(shape_x.rbegin() + 1);
        const int x_col =
            !matmul_op.getAdjX() ? shape_x.back() : *(shape_x.rbegin() + 1);

        const int y_row =
            matmul_op.getAdjY() ? shape_y.back() : *(shape_y.rbegin() + 1);
        const int y_col =
            !matmul_op.getAdjY() ? shape_y.back() : *(shape_y.rbegin() + 1);