cuda奇怪的內存訪問錯誤共享內存

我被困在這一整天。以下程序將給出「超出範圍的共享或本地地址」錯誤。評論這條線將解決這個問題。cuda奇怪的內存訪問錯誤共享內存

hist[tidx] = 0;

但是，我不認爲分配大小爲88 * 4字節的共享內存將是任何問題。

註釋掉此行也將解決這個問題

NVMatrix Acts(acts, true);

看來，如果我分配在全局存儲器的行爲矩陣，共享內存的行爲異常。任何想法？

int main(int argc, char ** argv) 
{ 
    float * act = new float[2985984]; 
    for (int i=0; i<2985984; i++) 
     act[i] = 0.0001*(i+1); 

    Matrix acts(act, 23328, 128); // use act as the data to initialize the 23328x128, matrix in cpu 

    NVMatrix Acts(acts, true);  // create a Acts Matrix which uses GPU global memory, and copies the value from CPU to GPU 
            // If comment out this line, there is no problem to execute the program 

    float cost = Calculate(); 

} 

float Calculate() 
{ 
    dim3 blocks(4,96); 
    dim3 threads(32,8); 

    cudaFuncSetCacheConfig(createShare<8, 32>, cudaFuncCachePreferShared); 

    int numLabels = 88; 

    createShare<8, 32><<<blocks, threads, numLabels>>>(numLabels); 

    return 0; 
} 

template <int B_Y, int B_X> 
__global__ void createShare(int numLabels) 
{ 
    extern __shared__ float hist[]; 

    int tidx = threadIdx.y * B_X + threadIdx.x; 
    if (tidx<numLabels) { 
     printf("block %d %d %d\n", blockIdx.x, blockIdx.y, tidx); 
     hist[tidx] = 0; 
    } 
}

來源

2013-07-22 Min Lin

更改此：

createShare<8, 32><<<blocks, threads, numLabels>>>(numLabels);

這樣：

createShare<8, 32><<<blocks, threads, numLabels*sizeof(float)>>>(numLabels);

要傳遞給內核，動態共享分配的大小是字節。您需要分配足夠的字節以涵蓋88 float數量。

來源

2013-07-22 16:01:29

cuda奇怪的內存訪問錯誤共享內存

回答

相關問題