CUDA內存管理/類中的指針問題

Question

我正在努力解決一些內存管理問題。 將結果復制到主機時，我一直收到“未指定的啟動失敗”消息。

我的代碼非常簡單-它在每個線程中生成兩個uint並將它們相乘。 我有提供隨機數的課程：

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存儲在頭文件中，如下所示：

__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] );
}

我的主內核非常簡單，看起來像這樣：

__global__  void InitKernel(uint3 * ptr, CuRandCuRandomNumberProvider * provider)
{
    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();
    ptr[offset].y = provider->GetRandomNumber();
    ptr[offset].z = ptr[offset].x * ptr[offset].y;
}

最后一個cudaMemcpy引起問題的main中的執行是：

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 ) );

如果我一切都做得很出色，例如：

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 ) );

setup_kernel完全相同，而InitKernel2如下所示：

__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;

}

一切都是魅力。 有人知道我在做什么錯嗎？ 我已經為此努力了幾個小時。 我認為這可能與內存管理或指針傳遞有關，但我不知道會是什么。

請幫忙 ：）！

Answer 1

這是非法的：

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

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

provider是要在主機上分配的變量。 將該指針傳遞給設備並在設備代碼中取消引用：

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按指針傳遞對象。 這將需要在主要方面進行一些更改：

CuRandCuRandomNumberProvider provider(numBlocks, threadsPerBlock);

在InitKernel中：

__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（）中：

__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;
}

（並且我也刪除了析構函數原型，因為它正在妨礙。）