CUDA:将数组复制到 GPU

Question

我正在尝试创建一个函数，该函数将值分配并复制到设备上的数组中。由于某种原因，该代码内联工作正常，但当我将其放入函数中时却无法正常工作，如下所示。我传递的参数是否正确？

我的 CUDA 程序仍然执行(没有错误)，但是如果我读回我写入的值，它们似乎是未定义的，即。不是 HostArray 中的原始值。

void CopyArrayToGPU(double *DeviceArray, double *HostArray, int NumElements)
{
    int bytes = sizeof(double) * NumElements;

    // Allocate memory on the GPU for array
    if (cudaMalloc((void**)&DeviceArray, bytes) != cudaSuccess)
    {
        printf("CopyArrayToGPU(): Couldn't allocate mem for array on GPU.");
    }

    // Copy the contents of the host array to the GPU
    if (cudaMemcpy(DeviceArray, HostArray, bytes, cudaMemcpyHostToDevice) != cudaSuccess)
    {
        printf("CopyArrayToGPU(): Couldn't copy host array to GPU.");
    }
}

// Declare device side array
static double *d_Inputs;

// Allocate host side array and plug in some test values
double *testArray = (double*)malloc(ImageSize);
testArray[0] = 0.6455696203; testArray[1] = 0.7954545455; testArray[2] = 0.2028985507; testArray[3] = 0.08;

CopyArrayToGPU(d_Inputs, testArray, NUM_INPUTS * NUM_ROWS);

Answer 1

您的变量 DeviceArray 是您的函数的区域设置。

您必须采用double **DeviceArray作为参数或返回cuda分配的值。

返回值版本:

double *CopyArrayToGPU(double *HostArray, int NumElements)
{
    int bytes = sizeof(double) * NumElements;
    void *DeviceArray;

    // Allocate memory on the GPU for array
    if (cudaMalloc(&DeviceArray, bytes) != cudaSuccess)
    {
        printf("CopyArrayToGPU(): Couldn't allocate mem for array on GPU.");
        return NULL;
    }

    // Copy the contents of the host array to the GPU
    if (cudaMemcpy(DeviceArray, HostArray, bytes, cudaMemcpyHostToDevice) != cudaSuccess)
    {
        printf("CopyArrayToGPU(): Couldn't copy host array to GPU.");
        cudaFree(DeviceArray);
        return NULL;
    }

    return DeviceArray;
}

所以你必须这样调用它:

d_Inputs = CopyArrayToGPU(testArray, NUM_INPUTS * NUM_ROWS);

<小时/>

double **DeviceArray 作为参数版本:

int CopyArrayToGPU(double **DeviceArray, double *HostArray, int NumElements)
{
    int bytes = sizeof(double) * NumElements;

    // Allocate memory on the GPU for array
    if (cudaMalloc((void **)DeviceArray, bytes) != cudaSuccess)
    {
        printf("CopyArrayToGPU(): Couldn't allocate mem for array on GPU.");
        return 1;
    }

    // Copy the contents of the host array to the GPU
    if (cudaMemcpy(*DeviceArray, HostArray, bytes, cudaMemcpyHostToDevice) != cudaSuccess)
    {
        printf("CopyArrayToGPU(): Couldn't copy host array to GPU.");
        cudaFree(*DeviceArray);
        return 1;
    }

    return 0;
}

所以你必须这样调用它:

CopyArrayToGPU(&d_Inputs, testArray, NUM_INPUTS * NUM_ROWS);