在Tensorflow中添加GPUOp

我正在尝试在本文档之后宽松地向TensorFlow添加新操作。不同之处在于我正在尝试实现基于GPU的操作。我要添加的操作是此处的cuda操作（cuda_op.py，cuda_op_kernel.cc，cuda_op_kernel.cu.cc）。我正在尝试在tensorflow之外编译这些文件，并使用tf.load_op_library它们将它们拉入。我进行了一些更改，所以这里是我的文件：

cuda_op_kernel.cc

#include "tensorflow/core/framework/op.h"
#include "tensorflow/core/framework/shape_inference.h"
#include "tensorflow/core/framework/op_kernel.h"

using namespace tensorflow;  // NOLINT(build/namespaces)

REGISTER_OP("AddOne")
    .Input("input: int32")
    .Output("output: int32")
    .SetShapeFn([](::tensorflow::shape_inference::InferenceContext* c) {
      c->set_output(0, c->input(0));
      return Status::OK();
    });

void AddOneKernelLauncher(const int* in, const int N, int* out);

class AddOneOp : public OpKernel {
 public:
  explicit AddOneOp(OpKernelConstruction* context) : OpKernel(context) {}

  void Compute(OpKernelContext* context) override {
    // Grab the input tensor
    const Tensor& input_tensor = context->input(0);
    auto input = input_tensor.flat();

    // Create an output tensor
    Tensor* output_tensor = NULL;
    OP_REQUIRES_OK(context, context->allocate_output(0, input_tensor.shape(),
                                                     &output_tensor));
    auto output = output_tensor->template flat();

    // Set all but the first element of the output tensor to 0.
    const int N = input.size();
    // Call the cuda kernel launcher
    AddOneKernelLauncher(input.data(), N, output.data());

  }
};

REGISTER_KERNEL_BUILDER(Name("AddOne").Device(DEVICE_GPU), AddOneOp);

cuda_op_kernel.cu

#define EIGEN_USE_GPU
#include 
#include 

__global__ void AddOneKernel(const int* in, const int N, int* out) {
  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i >>(in, N, out);

  cudaError_t cudaerr = cudaDeviceSynchronize();
  if (cudaerr != cudaSuccess)
    printf("kernel launch failed with error \"%s\".\n", cudaGetErrorString(cudaerr));
}

CMakeLists.txt

cmake_minimum_required(VERSION 3.5)

#found from running python -c 'import tensorflow as tf; print(tf.sysconfig.get_include())'
include_directories(/usr/local/lib/python3.5/dist-packages/tensorflow/include)

find_package(CUDA)

#set flags based on tutorial
set (CMAKE_CXX_FLAGS "--std=c++11 -fPIC -O2 -D_GLIBCXX_USE_CXX11_ABI=0")

#pass flags to c++ compiler
SET(CUDA_PROPAGATE_HOST_FLAGS ON)

#create library
cuda_add_library(
    cuda_op SHARED
    src/cuda_op_kernel.cu
    src/cuda_op_kernel.cc
    OPTIONS -gencode=arch=compute_20,code=sm_20)

#copy test file to build folder
configure_file(src/test.py test.py COPYONLY)

test.py

import tensorflow as tf
mod = tf.load_op_library('./libcuda_op.so')
with tf.Session() as sess:
    start = [5,4,3,2,1]
    print(start)
    print(mod.add_one(start).eval())

我能够编译并test.py成功运行，但是输出始终为[0 0 0 0 0]。如果我更换AddOneKernel<<<32, 256>>>(in, N, out);用for (int i = 0; i 和DEVICE_GPU与DEVICE_CPU，运算输出右值[6 5 4 3 2]（具有完全一样的CMakeList.txt）。