// ResourceHandle during Session run where the TF_Tensor is converted to a
// Tensor.
// TFv2 does not depend on this conversion. There is no matching
// TF_TensorFromTensorV1 because the conversion to string is performed by the
// python side of Session.
static Status TF_TensorToTensorV1(const TF_Tensor* src, Tensor* dst) {
  Status status = TF_TensorToTensor(src, dst);
  if (!status.ok()) {
    return status;
  }

    // Create C++ Tensor
    Tensor src(tensorflow::DT_STRING, TensorShape(dims));
    for (int64_t i = 0; i < src.NumElements(); ++i) {
      src.flat<tstring>()(i) = data[i];
    }
    TF_Tensor* dst = TF_TensorFromTensor(src, &status);
    ASSERT_TRUE(status.ok()) << status.message();

    // Convert back to a C++ Tensor and ensure we get expected output.
    Tensor output;

  tfrt.return %1, %2, %3 : !tfrt_fallback.tf_tensor, !tfrt_fallback.tf_tensor, !tfrt_fallback.tf_tensor
}

// CHECK-LABEL: func @test_insert_copy_for_arg
// CHECK-SAME: ([[arg:%.*]]: !tfrt_fallback.tf_tensor
func.func @test_insert_copy_for_arg(%arg: !tfrt_fallback.tf_tensor) -> (!tfrt_fallback.tf_tensor, !tfrt_fallback.tf_tensor, !tfrt_fallback.tf_tensor) attributes {tfrt.cost_threshold = 1024} {

            test::AsTensor<int>({15, 18, 21, 24}, {2, 2}));
}

TEST_F(StopGradientSingleOutputMultiEdgeTest, StopGradFirstEdge) {
  CheckGrad({true, false, false},
            test::AsTensor<int>({14, 16, 18, 20}, {2, 2}));
}

TEST_F(StopGradientSingleOutputMultiEdgeTest, StopGradSecondEdge) {
  CheckGrad({false, true, false},
            test::AsTensor<int>({10, 12, 14, 16}, {2, 2}));
}

def TFR_CastOp : TFR_Op<"cast", [Pure]> {
  let description = [{
    The `cast` operation converts the operand with built-in tensor type to
    tfr.tensor type, or vice versa.

    Example:

    ```mlir
    %1 = tfr.cast(%0) : tensor<f32> -> !tfr.tensor
    %3 = tfr.cast(%1) : !tfr.tensor -> tensor<f32>
    ```
  }];

  let arguments = (ins TFR_singleTensorType:$arg);

    name = "tf_tensor_hdrs",
    hdrs = [
        "tf_tensor.h",
        "tf_tensor_helper.h",
    ],
    visibility = ["//visibility:public"],
    deps = [
        ":c_api_macros_hdrs",
        ":tf_datatype_hdrs",
        ":tf_status_headers",
        "//tensorflow/core/platform:status",
    ],
)

cc_library(
    name = "tf_tensor",
    srcs = ["tf_tensor.cc"],
    hdrs = [
        "tf_tensor.h",

  r_.AddUse(/*ioperation=*/1, /*itensor=*/0);
  r_.AddUse(/*ioperation=*/2, /*itensor=*/0);
  r_.AddUse(/*ioperation=*/2, /*itensor=*/1);
  r_.AddUse(/*ioperation=*/3, /*itensor=*/0);
  r_.AddUse(/*ioperation=*/3, /*itensor=*/1);
  r_.AddUse(/*ioperation=*/3, /*itensor=*/2);

  // /* before */
  // %0 = f1()
  // %1 = f2(%0)
  // %2 = f3(%0,%1)
  // %3 = f4(%0,%1,%2)
  // f5()

      %outputs_0, %control_1 = tf_executor.island wraps "tf.Identity"(%outputs) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : (tensor<!tf_type.f32ref>) -> tensor<*xf32>
      tf_executor.fetch %outputs_0 : tensor<*xf32>
    }
    func.return %0 : tensor<*xf32>
  }

    TF_OpKernelConstruction* ctx);

// Allocates Tensor for output at given index. Caller takes ownership of
// returned TF_Tensor and should deallocate it using TF_DeleteTensor(tensor).
//
// This function should be used to allocate outputs inside kernel
// compute function.
TF_CAPI_EXPORT TF_Tensor* TF_AllocateOutput(TF_OpKernelContext* context,

  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TF_Tensor* tensor =
      TF_AllocateTensor(TF_FLOAT, dims.data(), dims.size(), sizeof(float));
  float tensor_data[] = {1};
  memcpy(TF_TensorData(tensor), tensor_data, TF_TensorByteSize(tensor));
  TFE_TensorHandle* tensor_handle = TFE_NewTensorHandle(tensor, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);