c++ - cudaMemcpyFromSymbol和cudaMemcpyToSymbol始终返回cudaErrorInvalidSymbol(13)错误

Question

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，这些代码无法正常运行。

#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

invalid device symbol   55 AVG: -1

C:\Users\Administrator\source\repos\CudaTests\x64\Debug\cudabase.exe
(process 18152) exited with code 0.

有任何想法吗？

Answer 1

删除此行，因为在启动内核之前，设备变量“count”始终为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;
    }

编辑

cudaError = cudaMemcpyFromSymbol(&hostResult, (void *)avg, sizeof(int), 0, CudaMemcpyDeviceToHost);

TO

cudaError = cudaMemcpyFromSymbol(&hostResult, avg, sizeof(int), 0, cudaMemcpyDeviceToHost);

(从参数“avg”中删除强制转换)。智能提供错误，但是代码可以编译并运行成功。