CUDA內存管理/類問題指針

Question

我正在努力解決一些內存管理問題。將結果複製到主機時，我一直收到「未指定的啓動失敗」。

我的代碼很簡單 - 它在每個線程中生成兩個提示並將它們相乘。 我有類提供一個隨機數：

class CuRandCuRandomNumberProvider : 
{ 
public: 
    CuRandCuRandomNumberProvider(dim3 numBlocks, dim3 threadsPerBlock); 
    CuRandCuRandomNumberProvider(dim3 numBlocks, dim3 threadsPerBlock, unsigned int seed); 
    __device__ unsigned int GetRandomNumber(); 
    ~CuRandCuRandomNumberProvider(); 
protected: 
    curandState * states; 
    __device__ bool IsPrime(unsigned int number); 
}; 

CuRandCuRandomNumberProvider::CuRandCuRandomNumberProvider(dim3 numBlocks, dim3 threadsPerBlock) 
{ 
    int numberOfThreads = threadsPerBlock.x * threadsPerBlock.y * numBlocks.x * numBlocks.y; 
    std::cout << numberOfThreads << std::endl; 
    cudaMalloc (&this->states, numberOfThreads*sizeof(curandState)); 
    setup_kernel <<< numBlocks, threadsPerBlock >>> (this->states, time(NULL)); 
} 

__device__ unsigned int CuRandCuRandomNumberProvider::GetRandomNumber() 
{ 
    int x = threadIdx.x + blockIdx.x * blockDim.x; 
    int y = threadIdx.y + blockIdx.y * blockDim.y; 
    int offset = x + y * blockDim.x * gridDim.x; 
    register float r = curand_uniform(&this->states[offset]); 
    return 0 + ((double)UINT_MAX) * r; 
} 

setup_kernel存儲在頭文件，看起來像這樣：

__global__ void setup_kernel (curandState * state, unsigned long seed) 
{ 
    int x = threadIdx.x + blockIdx.x * blockDim.x; 
    int y = threadIdx.y + blockIdx.y * blockDim.y; 
    int offset = x + y * blockDim.x * gridDim.x; 
    curand_init (seed, offset, 0, &state[offset]); 
} 

我的主要核心是非常簡單的，看起來像這樣：

​​

最後cudaMemcpy導致問題的主執行是：

uint3 * pqnD; 

uint3 * pqnH = (uint3*)malloc(sizeof(uint3) * numberOfThreads); 
memset(pqnH,0,sizeof(uint3) * numberOfThreads); 

HANDLE_ERROR(cudaMalloc((void**)&pqnD, sizeof(uint3) * numberOfThreads)); 

CuRandCuRandomNumberProvider * provider = new CuRandCuRandomNumberProvider(numBlocks, threadsPerBlock); 

InitKernel<<<numBlocks, threadsPerBlock>>>(pqnD, provider); 

HANDLE_ERROR(cudaMemcpy(pqnH, pqnD, sizeof(uint3) * numberOfThreads, cudaMemcpyDeviceToHost)); // this line causes error 

HANDLE_ERROR(cudaFree(pqnD)); 

如果我做的一切explicily，如：

uint3 * pqnD; 

uint3 * pqnH = (uint3*)malloc(sizeof(uint3) * numberOfThreads); 

memset(pqnH,0,sizeof(uint3) * numberOfThreads); 

HANDLE_ERROR(cudaMalloc((void**)&pqnD, sizeof(uint3) * numberOfThreads)); 

curandState * states; 

cudaMalloc (&states, numberOfThreads*sizeof(curandState)); 

setup_kernel <<< numBlocks, threadsPerBlock >>> (states, time(NULL)); 

CuRandCuRandomNumberProvider * provider = new CuRandCuRandomNumberProvider(numBlocks, threadsPerBlock, states); 


InitKernel2<<<numBlocks, threadsPerBlock>>>(pqnD, states); 

HANDLE_ERROR(cudaMemcpy(pqnH, pqnD, sizeof(uint3) * numberOfThreads, cudaMemcpyDeviceToHost)); 

HANDLE_ERROR(cudaFree(pqnD)); 

哪裏setup_kernel是完全一樣的，並InitKernel2樣子：

__global__ void InitKernel2(uint3 * ptr, curandState * states) 
{ 
    int x = threadIdx.x + blockIdx.x * blockDim.x; 
    int y = threadIdx.y + blockIdx.y * blockDim.y; 
    int offset = x + y * blockDim.x * gridDim.x; 

    ptr[offset].x = GetRandomNumber(states); 
    ptr[offset].y = GetRandomNumber(states); 
    ptr[offset].z =  ptr[offset].x *  ptr[offset].y; 
} 

和getRandomNumber的是：

__device__ unsigned int GetRandomNumber(curandState * states) 
{ 
    int x = threadIdx.x + blockIdx.x * blockDim.x; 
    int y = threadIdx.y + blockIdx.y * blockDim.y; 
    int offset = x + y * blockDim.x * gridDim.x; 
    register float r = curand_uniform(&states[offset]); 
    return 0 + ((double)UINT_MAX) * r; 

} 

一切正常作爲魅力。有沒有人有線索我做錯了什麼？我一直在掙扎幾個小時。我的事情可能是內存管理或指針傳遞的東西，但我不知道它會是什麼。

請幫忙:)！

Answer 1

這是非法的：

CuRandCuRandomNumberProvider * provider = new CuRandCuRandomNumberProvider(numBlocks, threadsPerBlock); 

InitKernel<<<numBlocks, threadsPerBlock>>>(pqnD, provider); 

provider是，你分配在主機上的變量。順便指出指針裝置和設備代碼解引用它：

ptr[offset].x = provider->GetRandomNumber(); 

（最終導致:)

register float r = curand_uniform(&this->states[offset]); 

是非法的。

由於您似乎想要在主機上設置對象（類CuRandCuRandomNumberProvider）並將其傳遞給設備，因此一種可能的解決方法是按值而不是指針傳遞對象。這將需要一些變化，主要：

CuRandCuRandomNumberProvider provider(numBlocks, threadsPerBlock); 

在InitKernel

：

__global__ void InitKernel(uint3 * ptr, CuRandCuRandomNumberProvider provider) // change 
{ 
    int x = threadIdx.x + blockIdx.x * blockDim.x; 
    int y = threadIdx.y + blockIdx.y * blockDim.y; 
    int offset = x + y * blockDim.x * gridDim.x; 

    ptr[offset].x = provider.GetRandomNumber(); // change 
    ptr[offset].y = provider.GetRandomNumber(); // change 
    ptr[offset].z = ptr[offset].x * ptr[offset].y; 
} 

在CuRandCuRandomNumberProvider

:: getRandomNumber的（）：

__device__ unsigned int CuRandCuRandomNumberProvider::GetRandomNumber() 
{ 
    int x = threadIdx.x + blockIdx.x * blockDim.x; 
    int y = threadIdx.y + blockIdx.y * blockDim.y; 
    int offset = x + y * blockDim.x * gridDim.x; 
    register float r = curand_uniform(&(states[offset])); // change 
    return 0 + ((double)UINT_MAX) * r; 
} 

（我刪除的析構函數原型太，因爲它正在妨礙你。）