我在CUDA C ++ API中调用cudaMemcpyFrom(To)Symbol函数时遇到问题。欢迎在块之间存储一些参数的替代方案。下面,我提供了一些(示例)代码,这些代码无法正常工作。
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <stdio.h>
#include <cstdlib>
#include <ctime>
int avgHost(int*, int);
cudaError_t cudaError;
__device__ int getGlobalIdx()
{
int blockId = blockIdx.x + blockIdx.y * gridDim.x + gridDim.x * gridDim.y * blockIdx.z;
int threadId = blockId * (blockDim.x * blockDim.y * blockDim.z) + (threadIdx.z * (blockDim.x * blockDim.y)) + (threadIdx.y * blockDim.x) + threadIdx.x;
return threadId;
}
__device__ int avg;
__device__ int count;
__device__ int sum;
__global__ void avgKernel(const int *arr)
{
auto idx = getGlobalIdx();
count++;
sum += arr[idx];
avg = sum / count;
}
int main(int argc, char ** argv)
{
const int arraySize = 128;
auto arr1 = (int*)__vcrt_malloc_normal(arraySize * sizeof(int));
srand(time(NULL));
for (size_t i = 0; i < arraySize; i++)
{
arr1[i] = rand() % 100;
}
auto hostAvg = avgHost(arr1, arraySize);
fprintf_s(stdout, "AVG: %d", hostAvg);
free(arr1);
return 0;
}
int avgHost(int* arr, int arraySize)
{
int *dArray = nullptr;
cudaSetDevice(0);
cudaError = cudaMemcpyToSymbol((void *)count, (void*)0, sizeof(int), 0, cudaMemcpyHostToDevice);
if (cudaError)
{
fprintf_s(stderr, "%s\t%d\n", cudaGetErrorString(cudaError), __LINE__);
return -1;
}
cudaMalloc((void**)&dArray, arraySize * sizeof(int));
cudaMemcpy(dArray, arr, arraySize * sizeof(int), cudaMemcpyKind::cudaMemcpyHostToDevice);
avgKernel <<<1, arraySize>>> (dArray);
cudaDeviceSynchronize();
int hostResult = -1;
cudaError = cudaMemcpyFromSymbol(&hostResult, (void *)avg, sizeof(int), 0, cudaMemcpyDeviceToHost);
if (cudaError)
{
fprintf_s(stderr, "%s\t%d\n", cudaGetErrorString(cudaError), __LINE__);
}
cudaFree(dArray);
return hostResult;
}
当我启动打印代码系统时
无效的设备符号55 AVG:-1
C:\ Users \ Administrator \ source \ repos \ CudaTests \ x64 \ Debug \ cudabase.exe(进程18152)以代码0退出。
有什么想法吗?
删除此行,因为启动内核之前设备变量“ count”始终为0。