Question

看看我自己编写的cuda内核。我有一个很大的内核，但它返回给我错误信息。然后我简化了它，发现它在一个循环中失败了。我简化了这个循环，发现如果我使用int值或常量值来填充循环中的数据[threadIdx.x]，它可以正常工作。但是，如果我使用double类型值，则返回错误。

建议：如果您没有正确地将数据从主机复制到设备，则可以在使用Nsight时收到“警告：检测到Cuda API错误：cudaLaunch返回（0x7）”消息，或者您可以从终端运行应用程序时出现分段错误错误

__global__ void sumSeries(double* dSum,int* totalThreadNumber){
    volatile  __shared__ double data[768];
    double var=0;
    data[threadIdx.x]=0;
    for ( int i = 10 ; i < 20 ;++i){
        var=i;
        data[threadIdx.x] += (var)/(var*var+1);
        __syncthreads();
    }


}

为什么不起作用？

int main() {

    int threadsPerBlock=768;
    int blockCount=8;

    int *hostThreadNumber=new int ;
        *hostThreadNumber=threadsPerBlock*blockCount;
    int* deviceThreadNumber=NULL;

    double* deviceSum=NULL;
    double* hostSum=(double*)malloc(blockCount);


    cudaError_t cuerr=cudaDeviceSetCacheConfig(cudaFuncCachePreferShared);
    if (cuerr != cudaSuccess){
        std::cout<<"Cant SetCacheConfig: "<<cudaGetErrorString(cuerr)<<std::endl;
        return -1;
    }

    cuerr=cudaMalloc(&deviceSum,blockCount*sizeof(double));//размер дабла*число блоков
    if (cuerr != cudaSuccess){
        std::cout<<"Cant allocate memory for deviceSum: "<<cudaGetErrorString(cuerr)<<std::endl;
        return -1;
    }
    cuerr=cudaMalloc(&deviceThreadNumber,sizeof(int));
    if (cuerr != cudaSuccess){
        std::cout<<"Cant allocate memory for deviceThreadNumber: "<<cudaGetErrorString(cuerr)<<std::endl;
        return -1;
    }

    cuerr = cudaMemcpy(deviceSum,hostSum,blockCount*sizeof(double),cudaMemcpyHostToDevice);
    if (cuerr != cudaSuccess){
        std::cout<<"Can not copy hostSum to device: "<<cudaGetErrorString(cuerr)<<std::endl;
        return -1;
    }
    cuerr = cudaMemcpy(deviceThreadNumber,hostThreadNumber,sizeof(int),cudaMemcpyHostToDevice);
    if (cuerr != cudaSuccess){
        std::cout<<"Can not copy hostThreadNumber to device: "<<cudaGetErrorString(cuerr)<<std::endl;
        return -1;
    }

    sumSeries<<<dim3(blockCount),dim3(threadsPerBlock)>>>(deviceSum,deviceThreadNumber);
    cuerr=cudaGetLastError();
    if (cuerr != cudaSuccess){
        std::cout<<"Cuda kernel error: "<<cudaGetErrorString(cuerr)<<std::endl;
        return -1;
    }
    cuerr= cudaDeviceSynchronize();
    if (cuerr != cudaSuccess){
        std::cout<<"Can not synchronize cuda kernel : "<<cudaGetErrorString(cuerr)<<std::endl;
        return -1;
    }
    cuerr= cudaMemcpy(hostSum,deviceSum,blockCount*sizeof(double),cudaMemcpyDeviceToHost);
    if (cuerr != cudaSuccess){
        std::cout<<"Can not copy data to host: "<<cudaGetErrorString(cuerr)<<std::endl;
        return -1;
    }
    cudaFree(deviceSum);
    cudaFree(deviceThreadNumber);
    return 0;
}

Answer 1

您刚刚为hostSum

分配了8个字节的内存

double* hostSum=(double*)malloc(blockCount)

如果我假设您要为其分配blockCount * sizeof(double)个字节，那就错了，因为您为deviceSum分配了这个内存量，并将其用于主机和设备之间的内存复制。

cuerr = cudaMalloc(&deviceSum,blockCount*sizeof(double));

cuerr = cudaMemcpy(deviceSum,hostSum,blockCount*sizeof(double),cudaMemcpyHostToDevice);

cuerr= cudaMemcpy(hostSum,deviceSum,blockCount*sizeof(double),cudaMemcpyDeviceToHost);

无法运行CUDA内核：请求启动的资源太多

1 个答案: