我试图在映射内存上使用thrust :: copy_if函数。但是,当我得到一个运行时错误并且我无法找到它时,在花费大量时间进行调试之前,我想确认它实际上是允许将指针传递给映射内存的事实push :: device_ptr包装器的位置。
这是我的意思的一个例子:
int size=1024;
int* v_locked;
int* v_device;
int* stencil_device;
device_ptr<int> v_wrapper;
device_ptr<int> v_wrapper_end;
device_ptr<int> stencil_wrapper;
cudaHostAlloc((void**)&v_locked, size*sizeof(int), cudaHostAllocMapped));
cudaHostGetDevicePointer(&v_device, &v_locked, 0);
cudaMalloc((void**)&stencil_device, size*sizeof(int));
/*
kernel assigning stencil_device elements ...
*/
v_wrapper = device_pointer_cast(v_device);
stencil_wrapper = device_pointer_cast(stencil_device);
v_wrapper_end = copy_if(make_counting_iterator<int>(0), make_counting_iterator<int>(size), stencil_wrapper, v_wrapper, _1 == 1);
这是使用推力库的映射内存的正确用法吗?
谢谢。
答案 0 :(得分:4)
是的,这是可能的。
我相信您的代码存在一些问题。
cudaHostGetDevicePointer
的调用似乎编译正确,但它们未正确设置。< / LI>
cudaHostGetDevicePointer()
的来电未正确设置。第二个指针参数作为单个指针(*
)传递,而不是双指针(**
)。请参阅documentation此编写的调用会抛出您可以捕获的cuda运行时错误。cudaHostAlloc
来电之前,您应该使用对enable this feature的cudaSetDeviceFlags(cudaDeviceMapHost);
来电。 以下示例代码似乎对我有效,并修复了上述问题:
$ cat t281.cu
#include <iostream>
#include <thrust/device_vector.h>
#include <thrust/device_ptr.h>
#include <thrust/copy.h>
#define cudaCheckErrors(msg) \
do { \
cudaError_t __err = cudaGetLastError(); \
if (__err != cudaSuccess) { \
fprintf(stderr, "Fatal error: %s (%s at %s:%d)\n", \
msg, cudaGetErrorString(__err), \
__FILE__, __LINE__); \
fprintf(stderr, "*** FAILED - ABORTING\n"); \
exit(1); \
} \
} while (0)
template<typename T>
struct is_one : thrust::unary_function<T, bool>
{
__host__ __device__
bool operator()(const T &x)
{
return (x==1);
}
};
int main(){
int size=1024;
int* v_locked;
int* v_device;
int* stencil_locked;
int* stencil_device;
cudaSetDeviceFlags(cudaDeviceMapHost);
cudaCheckErrors("cudaSetDeviceFlags");
cudaHostAlloc((void**)&v_locked, size*sizeof(int), cudaHostAllocMapped);
cudaCheckErrors("cudaHostAlloc 1");
cudaHostGetDevicePointer(&v_device, v_locked, 0);
cudaCheckErrors("cudaHostGetDevicePointer 1");
cudaHostAlloc((void**)&stencil_locked, size*sizeof(int), cudaHostAllocMapped);
cudaCheckErrors("cudaHostAlloc 2");
cudaHostGetDevicePointer(&stencil_device, stencil_locked, 0);
cudaCheckErrors("cudaHostGetDevicePointer 2");
for (int i = 0; i < size; i++){
v_locked[i] = i;
stencil_locked[i] = i%2;}
thrust::device_ptr<int> v_wrapper = thrust::device_pointer_cast(v_device);
thrust::device_ptr<int> stencil_wrapper = thrust::device_pointer_cast(stencil_device);
thrust::device_ptr<int> v_wrapper_end = v_wrapper + size;
thrust::device_vector<int> result(size);
thrust::device_vector<int>::iterator result_end = copy_if(v_wrapper, v_wrapper_end, stencil_wrapper, result.begin(), is_one<int>());
int result_size = result_end - result.begin();
thrust::host_vector<int> h_result(result_size);
thrust::copy_n(result.begin(), result_size, h_result.begin());
thrust::copy_n(h_result.begin(), 10, std::ostream_iterator<int>(std::cout, " "));
std::cout << std::endl;
return 0;
}
$ nvcc -arch=sm_20 -o t281 t281.cu
$ ./t281
1 3 5 7 9 11 13 15 17 19
$