我正在尝试在CUDA 7.5中实现并行归约求和。我一直在尝试遵循 NVIDIA 的PDF,其中逐步介绍了最初的算法,然后是逐渐优化的版本。我当前正在创建一个数组,该数组中每个位置的值都为1,以便我可以检查输出是否正确,但是对于大小为64的数组,我得到了-842159451的值。我期望内核代码是正确的,因为我已经按照NVIDIA的完全相同的代码进行了操作,以下是我的内核:
__global__ void reduce0(int *input, int *output) {
extern __shared__ int sdata[];
unsigned int tid = threadIdx.x;
unsigned int i = blockIdx.x * blockDim.x + threadIdx.x;
sdata[tid] = input[i];
__syncthreads();
for (unsigned int s = 1; s < blockDim.x; s *= 2) {
if (tid % (2 * s) == 0) {
sdata[tid] += sdata[tid + s];
}
__syncthreads();
}
if (tid == 0) output[blockIdx.x] = sdata[0];
}
这是我的调用内核的代码,我认为我的问题就在这里:
int main()
{
int numThreadsPerBlock = 1024;
int *hostInput;
int *hostOutput;
int *deviceInput;
int *deviceOutput;
int numInputElements = 64;
int numOutputElements; // number of elements in the output list, initialised below
numOutputElements = numInputElements / (numThreadsPerBlock / 2);
if (numInputElements % (numThreadsPerBlock / 2)) {
numOutputElements++;
}
hostInput = (int *)malloc(numInputElements * sizeof(int));
hostOutput = (int *)malloc(numOutputElements * sizeof(int));
for (int i = 0; i < numInputElements; ++i) {
hostInput[i] = 1;
}
const dim3 blockSize(numThreadsPerBlock, 1, 1);
const dim3 gridSize(numOutputElements, 1, 1);
cudaMalloc((void **)&deviceInput, numInputElements * sizeof(int));
cudaMalloc((void **)&deviceOutput, numOutputElements * sizeof(int));
cudaMemcpy(deviceInput, hostInput, numInputElements * sizeof(int), cudaMemcpyHostToDevice);
reduce0 << <gridSize, blockSize >> >(deviceInput, deviceOutput);
cudaMemcpy(hostOutput, deviceOutput, numOutputElements * sizeof(int), cudaMemcpyDeviceToHost);
for (int ii = 1; ii < numOutputElements; ii++) {
hostOutput[0] += hostOutput[ii]; //accumulates the sum in the first element
}
int sumGPU = hostOutput[0];
printf("GPU Result: %d\n", sumGPU);
std::string wait;
std::cin >> wait;
return 0;
}
我还尝试了更大和更小的数据输入数组,但无论数组大小如何,都得到一个非常大的负值。