CUDA

Question

我是 CUDA 的新手（在 C 中也是..），我正在尝试使用 int shared 作为标志来在设置 finish 时停止所有线程设备，但是当我将它复制回主机时，它永远不会更新，我可以使用 char * 来完成，但是在使用简单的 int

时它不起作用

最小代码示例：


__global__ void bingo(int * finish){

    __shared__ int shared;

    if(threadIdx.x == 5){
        printf("\nassign to finish %d",threadIdx.x);
        shared = threadIdx.x;
        finish = (int*) threadIdx.x;
        printf("GPU says: %d\n",*finish);
        return;
    }
    __syncthreads();
    if(shared != NULL){
        printf("\nreturn from thread: %d", threadIdx.x);
        return;
    }
}

int main() {

    int* threadBingo;
    cudaMalloc((void**)&threadBingo, sizeof( int));

    bingo<<<1,10>>>(threadBingo );
    cudaDeviceSynchronize();

    int* threadWhoMadeBingo = (int *) malloc(sizeof(int));
    
    cudaMemcpy(threadWhoMadeBingo, threadBingo, sizeof(int), cudaMemcpyDeviceToHost);
    printf("\n thread who made bingo %d\n", *threadWhoMadeBingo);


    cudaDeviceReset();
    cudaDeviceSynchronize();
    
    return 0;
}

并且输出：

assign to finish 5
GPU says: 5
return from thread: 0
return from thread: 1
return from thread: 2
return from thread: 3
return from thread: 4
return from thread: 6
return from thread: 7
return from thread: 8
return from thread: 9
 thread who made bingo 0

如您所见，最后一行应该是 5 而不是 0

Answer 1

好的，我找到了：

行 finish = (int*) threadIdx.x; 应该是 -> *finish = threadIdx.x;.

我会在两天内接受这个答案。

CUDA - 来自设备的 int 在将其复制到主机时未更新

CUDA - int from device get not updated while copying it to host

pointers