# CHECK-NEXT: %[[READ_VAR_0:.*]] = "tf.ReadVariableOp"(%[[ARG_2]]) {{{.*}}} : (tensor<!tf_type.resource<tensor<1x3xf32>>>) -> tensor<1x3xf32>
# CHECK-NEXT: %[[MATMUL_0:.*]] = "tf.MatMul"(%[[ARG_1]], %[[READ_VAR_0]]) <{{{.*}}}> {{{.*}}} : (tensor<3x1xf32>, tensor<1x3xf32>) -> tensor<3x3xf32>
# CHECK-NEXT: return %[[MATMUL_0]] : tensor<3x3xf32>


def Test():
  x = tf.constant([[1.0], [1.0], [1.0]])
  y = tf.compat.v1.get_variable(
      name='y',

  // 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]

  // CHECK: %[[MATMUL_6:.*]] = "tf.MatMul"(%[[LHS_6]], %[[RHS_6]]) <{grad_a = false, grad_b = false, transpose_a = false, transpose_b = false}> : (tensor<4x5xf32>, tensor<5x6xf32>) -> tensor<4x6xf32>

      auto dv = Const(scope, {{1.0, 1.0}, {1.0, 1.0}});
      auto dt = MatMul(scope, dv, u, MatMul::TransposeB(true));
      auto du = MatMul(scope, t, dv, MatMul::TransposeA(true));

      auto dz = Const(scope, {{1.0, 1.0}, {1.0, 1.0}});
      auto dx = MatMul(scope, dz, y, MatMul::TransposeB(true));
      auto dy = MatMul(scope, x, dz, MatMul::TransposeA(true));
    } else {
      // Call AddSymbolicGradients.

// TF-DAG: %[[pc_6:.*]] = "tf.PartitionedCall"(%[[pc_2]], %[[cst_0]]) <{config = "", config_proto = "", executor_type = "", f = @composite_matmul_fn_1_0}

// 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

// CHECK: %[[MATMUL_3:.*]] = "tf.MatMul"(%arg0, %[[ORIGINAL_IDENTITY]]) <{transpose_a = false, transpose_b = false}> {attr_map = "0:transpose_a,1:transpose_a", device = ""} : (tensor<1x2x2x2xf32>, tensor<2x1024xf32>) -> tensor<*xf32>
// CHECK: return %[[MATMUL_1]], %[[MATMUL_2]], %[[MATMUL_3]] : tensor<*xf32>, tensor<*xf32>, tensor<*xf32>

// 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()) {

  num_retvals = 1;
  TFE_Execute(matmul.get(), &retval, &num_retvals, status.get());
  ASSERT_NE(TF_OK, TF_GetCode(status.get()));
  ASSERT_TRUE(absl::StrContains(TF_Message(status.get()), custom0));
  ASSERT_TRUE(absl::StrContains(TF_Message(status.get()), custom1));

  // Custom device: mix of custom/physical places the op on the custom device.