我正在编写 CUDA 程序来处理图像。我遇到了分段错误问题,我完全不知道为什么。我可能只是想念一些小事,但经过数小时的尝试纠正自己,我无法让它运行。
根据我硬件上的 deviceQuery(GeForce310M 总 shmem:16384,每个块的最大线程数:512 和最大块暗淡 521),我正在设置正确的网格、块和切碎内存值(至少我是这么认为的)。这是我在分段错误之前得到的输出:
Cols and Rows:
256
384
total:98304 // rows*cols
Block Size:512
Grid Size:192
shared mem:2048
下面是内核代码
__global__ void reduce_min(float *minVar,float* d_logLuminance)
{
extern __shared__ float s_data[]; //shared memeory
unsigned int tId = threadIdx.x;
unsigned int global_id = blockIdx.x*blockDim.x + tId;
//Copy to shared mem
s_data[tId] = d_logLuminance[global_id];
__syncthreads();
for(unsigned int i = 1;i<blockDim.x;i*=2)
{
if(tId%(2*i) == 0)
{
s_data[tId] = min(s_data[tId],s_data[tId+i]);
}
__syncthreads();
}
if(tId == 0)
d_logLuminance[blockIdx.x] = s_data[0];
}
以及我分配内存和调用内核的部分
size_t size = 512;
size_t sizeOfgrid = (numRows*numCols)/size;
const dim3 blockSize(size,1,1);
const dim3 gridSize(sizeOfgrid,1,1);
unsigned int sharedMem =(sizeof(float)*size); //Shared memory per block
printf("%s\n%d\n%d\n%s%d\n","Cols and Rows:",numCols,numRows,"total:",numRows*numCols);
printf("%s%d\n%s%d\n%s%d\n","Block Size:",size,"Grid Size:",sizeOfgrid,"shared mem:",sharedMem);
float* d_lum;
float* outData;
checkCudaErrors(cudaMalloc(&d_lum,sizeof(float)*numRows*numCols));
checkCudaErrors(cudaMemcpy(d_lum,d_logLuminance,sizeof(float)*numRows*numCols,cudaMemcpyDeviceToDevice));
checkCudaErrors(cudaMalloc(&outData,sizeof(float)*numRows*numCols));
reduce_min<<<gridSize,blockSize,sharedMem>>>(outData,d_lum);
cudaDeviceSynchronize();
checkCudaErrors(cudaGetLastError());
checkCudaErrors(cudaMemcpy(&min_logLum,outData,sizeof(float)*numRows*numCols,cudaMemcpyDeviceToHost));
提前感谢您的帮助。