Adding a GPU Op in Tensorflow

Question

I am trying to add a new op to TensorFlow loosely following this document. The difference being that I am trying to implement a GPU based op. The op I'm trying to add is the cuda op from here (cuda_op.py, cuda_op_kernel.cc, cuda_op_kernel.cu.cc). I am trying to compile these outside of tensorflow and the use tf.load_op_library to pull them in. I have made some changes so here are my files:

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<int32>();

    // 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<int32>();

    // 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 <cuda.h>
#include <stdio.h>

__global__ void AddOneKernel(const int* in, const int N, int* out) {
  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N;
       i += blockDim.x * gridDim.x) {
    out[i] = in[i] + 1;
  }
}

void AddOneKernelLauncher(const int* in, const int N, int* out) {
  AddOneKernel<<<32, 256>>>(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())

I am able to compile and run test.py successfully, but the output is always [0 0 0 0 0]. If I replace AddOneKernel<<<32, 256>>>(in, N, out); with for (int i = 0; i < N; i++) out[i] = in[i] + 1; and DEVICE_GPU with DEVICE_CPU, the op outputs the right values [6 5 4 3 2] (with exact same CMakeList.txt).

Any idea how to get the correct values to be returned?

Answer 1

I don't fully remember where I found the cmake stuff for CUDA, but the options were messing the compilation up somehow. Replacing cuda_add_library in the CMakeLists.txt to be the following fixed the problem.

#no options needed
cuda_add_library(
    cuda_op SHARED
    src/cuda_op_kernel.cu
    src/cuda_op_kernel.cc)