如何在一个CUDA代码中使用CUB和Thrust

Question

我正在尝试向我的“旧”推力代码中引入一些CUB，因此我从一个小示例开始比较thrust::reduce_by_key和cub::DeviceReduce::ReduceByKey，两者都适用于thrust::device_vectors。

代码的推力部分没有问题，但幼稚地使用通过thrust::raw_pointer_cast获得的原始指针的CUB部分，在CUB调用后崩溃。我放了一个cudaDeviceSynchronize()来尝试解决这个问题，但它没有帮助。代码中的CUB部分是从CUB网页中抄袭过来的。

在OSX上，运行时错误是：

libc++abi.dylib: terminate called throwing an exception
Abort trap: 6 

在Linux上，运行时错误是：

terminate called after throwing an instance of 'thrust::system::system_error'
what():  an illegal memory access was encountered

cuda-memcheck的前几行是：

========= CUDA-MEMCHECK
========= Invalid __global__ write of size 4
=========     at 0x00127010 in /home/sdettrick/codes/MCthrust/tests/../cub-1.3.2/cub/device/dispatch/../../block_range/block_range_reduce_by_key.cuh:1017:void cub::ReduceByKeyRegionKernel<cub::DeviceReduceByKeyDispatch<unsigned int*, unsigned int*, float*, float*, int*, cub::Equality, CustomSum, int>::PtxReduceByKeyPolicy, unsigned int*, unsigned int*, float*, float*, int*, cub::ReduceByKeyScanTileState<float, int, bool=1>, cub::Equality, CustomSum, int>(unsigned int*, float*, float*, int*, cub::Equality, CustomSum, int, cub::DeviceReduceByKeyDispatch<unsigned int*, unsigned int*, float*, float*, int*, cub::Equality, CustomSum, int>::PtxReduceByKeyPolicy, unsigned int*, int, cub::GridQueue<int>)
=========     by thread (0,0,0) in block (0,0,0)
=========     Address 0x7fff7dbb3e88 is out of bounds
=========     Saved host backtrace up to driver entry point at kernel launch time

不幸的是，我不太确定该怎么做。

任何帮助都将不胜感激。我在NVIDIA开发者专区尝试了一下，但没有得到任何响应。完整的示例代码如下。它应该使用CUDA 6.5和cub 1.3.2进行编译：

#include <iostream>
#include <thrust/sort.h>
#include <thrust/gather.h>
#include <thrust/device_vector.h>
#include <thrust/iterator/zip_iterator.h>
#include <thrust/iterator/permutation_iterator.h>
#include <thrust/iterator/discard_iterator.h>

#include <cub/cub.cuh>   // or equivalently <cub/device/device_radix_sort.cuh>

//========================================
// for CUB:
struct CustomSum
{
    template <typename T>
    CUB_RUNTIME_FUNCTION __host__ __device__ __forceinline__
    //__host__ __device__ __forceinline__
    T operator()(const T &a, const T &b) const {
        return b+a;
    }
};
//========================================

int main()
{
  const int Nkey=20;
  int Nseg=9;
  int ikey[Nkey] = {0, 0, 0, 6, 8, 0, 2, 4, 6, 8, 1, 3, 5, 7, 8, 1, 3, 5, 7, 8}; 

  thrust::device_vector<unsigned int> key(ikey,ikey+Nkey);
  thrust::device_vector<unsigned int> keysout(Nkey);

  // Let's reduce x, by key:

  float xval[Nkey];
  for (int i=0; i<Nkey; i++) xval[i]=ikey[i]+0.1f;

  thrust::device_vector<float> x(xval,xval+Nkey);

  // First, sort x by key:

  thrust::sort_by_key(key.begin(),key.end(),x.begin());

  //---------------------------------------------------------------------
  std::cout<<"=================================================================="<<std::endl
       <<" THRUST reduce_by_key:"<<std::endl
       <<"=================================================================="<<std::endl;

  thrust::device_vector<float> output(Nseg,0.0f);

  thrust::reduce_by_key(key.begin(),
            key.end(),
            x.begin(),
            keysout.begin(),
            output.begin());

  for (int i=0;i<Nkey;i++) std::cout << x[i] <<" ";  std::cout<<std::endl;
  for (int i=0;i<Nkey;i++) std::cout << key[i] <<" ";  std::cout<<std::endl;
  for (int i=0;i<Nseg;i++) std::cout << output[i] <<" ";  std::cout<<std::endl;

  float ototal=thrust::reduce(output.begin(),output.end());
  float xtotal=thrust::reduce(x.begin(),x.end());
  std::cout << "total="<< ototal <<", should be "<<xtotal<<std::endl;

  //---------------------------------------------------------------------
  std::cout<<"=================================================================="<<std::endl
       <<" CUB ReduceByKey:"<<std::endl
       <<"=================================================================="<<std::endl;


  unsigned int *d_keys_in   =thrust::raw_pointer_cast(&key[0]);
  float        *d_values_in =thrust::raw_pointer_cast(&x[0]);  
  unsigned int *d_keys_out  =thrust::raw_pointer_cast(&keysout[0]);
  float        *d_values_out=thrust::raw_pointer_cast(&output[0]);
  int          *d_num_segments=&Nseg;
  CustomSum   reduction_op;

  std::cout << "CUB input" << std::endl;
  for (int i=0; i<Nkey; ++i) std::cout << key[i]  << " ";  std::cout<<std::endl;
  for (int i=0; i<Nkey; ++i) std::cout << x[i] << " ";  std::cout<< std::endl;
  for (int i=0; i<Nkey; ++i) std::cout << keysout[i] << " ";  std::cout<< std::endl;
  for (int i=0; i<Nseg; ++i) std::cout << output[i] << " ";  std::cout<< std::endl;

  // Determine temporary device storage requirements
  void     *d_temp_storage = NULL;
  size_t   temp_storage_bytes = 0;
  cub::DeviceReduce::ReduceByKey(d_temp_storage, temp_storage_bytes, d_keys_in, d_keys_out, d_values_in, d_values_out, d_num_segments, reduction_op, Nkey);

  // Allocate temporary storage
  cudaMalloc(&d_temp_storage, temp_storage_bytes);
  std::cout << "temp_storage_bytes = " << temp_storage_bytes << std::endl;

  // Run reduce-by-key
  cub::DeviceReduce::ReduceByKey(d_temp_storage, temp_storage_bytes, d_keys_in, d_keys_out, d_values_in, d_values_out, d_num_segments, reduction_op, Nkey);
  cudaDeviceSynchronize();

  std::cout << "CUB output" << std::endl;

  std::cout<<Nkey<<" "<<Nseg<<std::endl;
  std::cout<<key.size() << " "<<x.size() << " "<<keysout.size() << " "<<output.size() << std::endl;

  // At this point onward it dies:
  //libc++abi.dylib: terminate called throwing an exception
  //Abort trap: 6  

  // If the next line is uncommented, it crashes the Mac!
  for (int i=0; i<Nkey; ++i) std::cout << key[i]  << " ";  std::cout<<std::endl;
  // for (int i=0; i<Nkey; ++i) std::cout << x[i] << " ";  std::cout<< std::endl;
  // for (int i=0; i<Nkey; ++i) std::cout << keysout[i] << " ";  std::cout<< std::endl;
  // for (int i=0; i<Nseg; ++i) std::cout << output[i] << " ";  std::cout<< std::endl;
  cudaFree(d_temp_storage);

  ototal=thrust::reduce(output.begin(),output.end());
  xtotal=thrust::reduce(x.begin(),x.end());
  std::cout << "total="<< ototal <<", should be "<<xtotal<<std::endl;
  return 1;
}

Answer 1

这是不合适的：

 int          *d_num_segments=&Nseg;

您不能将主机变量的地址用作设备指针。

取而代之的是：

int *d_num_segments;
cudaMalloc(&d_num_segments, sizeof(int));

这将根据数据大小分配设备上的空间( cub将写入的单个整数)，并将该分配的地址分配给d_num_segments变量。然后，这将成为有效的设备指针。

在(*普通，非UM) CUDA中，取消对设备代码中的主机地址或主机代码中的设备地址的引用是非法的。