CUDA的内存拷贝

CPU到GPU

  在CUDA程序中，将数据从CPU传输到GPU，或者从GPU传输到CPU的时候，需要调用底层的内存拷贝函数。当有很多不同类型的数据的时候，这个过程会非常繁琐。于是，我专门写了个内存拷贝的模板函数，使这个过程变的十分方便。

template <typename T> T* valueHostToDevice(T *value, const int &num = 1, bool isDelete = false){
    T *devValue;
    cudaMalloc((void**)&devValue, num * sizeof(T));
    cudaMemcpy(devValue, value, num * sizeof(T), cudaMemcpyHostToDevice);
    if(isDelete){
        if(num == 1){delete value;}
        else{delete[] value;}
    }
    return devValue;
}

  以上的函数能够将数据从CPU拷贝到GPU，而且能够支持任意类型的数据，比如整型int、浮点型float或自定义的类型，同时既能拷贝单个对象，也能拷贝一个对象数组，最后还可以销毁CPU内存数据，其调用方法也是非常方便

// 任意类型的数组指针
A *m_a = new A[3];
// 从CPU的m_a拷贝到GPU的a
A *a = valueHostToDevice(m_a, 3);

  以上是带有返回值的函数，也可以写成是不带返回值的函数

template <typename T> void valueHostToDevice(T &value, T **devValue){
    T *hostValue;
    cudaHostAlloc((void**)&hostValue, sizeof(T), cudaHostAllocDefault);
    cudaMalloc((void**)devValue, sizeof(T));
    *hostValue = value;
    cudaMemcpy(*devValue, hostValue, sizeof(T), cudaMemcpyHostToDevice);
    cudaFreeHost(hostValue);
}

// 调用
valueHostToDevice(m_a, &a);

  以上这个函数就有点繁琐了。为了GPU的开辟内存空间，需要使用cudaMalloc函数操作GPU指针a。cudaalloc函数需要传入指针a的地址，而直接传入a的话，这是作为形参的，所以函数里的指针是个临时变量，并不是原始的指针a的地址，所以在函数中直接传入指针a是无法操作的。所以需要定义二级指针，再让这个二级指针指向a指针，才能对其进行操作，所调用时需要传入的形参是&a。

GPU到CPU

  类似的还有从GPU传到CPU的过程，其调用方法也是类似的。

template <typename T> T* valueDeviceToHost(T *devValue, const int &num = 1){
    T *hostValue;
    cudaHostAlloc((void**)&hostValue, num * sizeof(T), cudaHostAllocDefault);
    cudaMemcpy(hostValue, devValue, num * sizeof(T), cudaMemcpyDeviceToHost);
    return hostValue;
}

template <typename T> void valueDeviceToHost(T *devValue, T &value){
    T *hostValue;
    cudaHostAlloc((void**)&hostValue, 1 * sizeof(T), cudaHostAllocDefault);
    cudaMemcpy(hostValue, devValue, 1 * sizeof(T), cudaMemcpyDeviceToHost);
    value = *hostValue;
    cudaFreeHost(hostValue);
}

二维数组

  以上只是拷贝一维数组，同时还可以拷贝二维数组。

template <typename T> T** array2DHostToDevice(T *hostData, T **devData, 
                                               const int &rowNum, const int &colNum){
    // 开辟内存
    T **hostArray;
    cudaHostAlloc((void**)&hostArray, rowNum * sizeof(T*), cudaHostAllocDefault);
    T **devArray;
    cudaMalloc((void**)&devArray, rowNum * sizeof(T*));
    size_t pitch;
    cudaMallocPitch(devData, &pitch, colNum * sizeof(T), rowNum);
    cudaMemset2D(*devData, pitch, 0, colNum * sizeof(T), rowNum);
    // 设置二维数组的指针
    for(int i = 0; i < rowNum; i++){
        hostArray[i] = (T*)((float*)(*devData) + i*pitch / sizeof(float));
    }
    // 复制数据
    if(hostData != NULL)
        cudaMemcpy2D(*devData, pitch, hostData, colNum * sizeof(T), colNum * sizeof(T), rowNum, cudaMemcpyHostToDevice);
    cudaMemcpy(devArray, hostArray, rowNum * sizeof(T*), cudaMemcpyHostToDevice);
    cudaFreeHost(hostArray);
    return devArray;
}

// 调用
vec **image2D = NULL, *gpu_image2DData = NULL;
vec *cpu_image2DData = NULL;
image2D = array2DHostToDevice(cpu_image2DData, &gpu_normalData, height, width);

template <typename T> void array2DHostToDevice(T ***devArray, 
                                               T *hostData, T **devData, 
                                               const int &rowNum, const int &colNum){
    // 开辟内存
    T **hostArray;
    cudaHostAlloc((void**)&hostArray, rowNum * sizeof(T*), cudaHostAllocDefault);
    cudaMalloc((void**)devArray, rowNum * sizeof(T*));
    size_t pitch;
    cudaMallocPitch(devData, &pitch, colNum * sizeof(T), rowNum);
    cudaMemset2D(*devData, pitch, 0, colNum * sizeof(T), rowNum);
    // 设置二维数组的指针
    for(int i = 0; i < rowNum; i++){
        hostArray[i] = (T*)((float*)(*devData) + i*pitch / sizeof(float));
    }
    // 复制数据
    cudaMemcpy2D(*devData, pitch, hostData, colNum * sizeof(T), colNum * sizeof(T), rowNum, cudaMemcpyHostToDevice);
    cudaMemcpy(*devArray, hostArray, rowNum * sizeof(T*), cudaMemcpyHostToDevice);
    cudaFreeHost(hostArray);
}

  封装成类之后，可以更加方便。

class MemCopy{
public:
    MemCopy(){}
    template <typename T> T* hostToDevice(T *value,
                                          const int &num = 1,
                                          bool isDelete = false);
    template <typename T> T** hostToDevice(T *hostData,
                                           T **devData,
                                           const int &rowNum,
                                           const int &colNum);
};
template <typename T> T* MemCopy::hostToDevice(T *value,
                                               const int &num,
                                               bool isDelete){
    // 任意类型的数组指针
    // A *m_a = new A[3];
    // 从CPU的m_a拷贝到GPU的a
    // A *a = valueHostToDevice(m_a, 3);

    T *devValue;
    cudaMalloc((void**)&devValue, num * sizeof(T));
    cudaMemcpy(devValue, value, num * sizeof(T), cudaMemcpyHostToDevice);
    if(isDelete){
        if(num == 1){delete value;}
        else{delete[] value;}
    }
    return devValue;
}
template <typename T> T** MemCopy::hostToDevice(T *hostData,
                                       T **devData,
                                       const int &rowNum,
                                       const int &colNum){
    // 调用
    // vec **image2D = NULL, *gpu_image2DData = NULL;
    // vec *cpu_image2DData = NULL;
    // image2D = array2DHostToDevice(cpu_image2DData, &gpu_normalData, height, width);

    // 开辟内存
    T **hostArray;
    cudaHostAlloc((void**)&hostArray, rowNum * sizeof(T*), cudaHostAllocDefault);
    T **devArray;
    cudaMalloc((void**)&devArray, rowNum * sizeof(T*));
    size_t pitch;
    cudaMallocPitch(devData, &pitch, colNum * sizeof(T), rowNum);
    cudaMemset2D(*devData, pitch, 0, colNum * sizeof(T), rowNum);
    // 设置二维数组的指针
    for(int i = 0; i < rowNum; i++){
        hostArray[i] = (T*)((float*)(*devData) + i*pitch / sizeof(float));
    }
    // 复制数据
    if(hostData != NULL)
        cudaMemcpy2D(*devData, pitch, hostData, colNum * sizeof(T), colNum * sizeof(T), rowNum, cudaMemcpyHostToDevice);
    cudaMemcpy(devArray, hostArray, rowNum * sizeof(T*), cudaMemcpyHostToDevice);
    cudaFreeHost(hostArray);
    return devArray;
}

int main(void){
    Timing t = Timing();
    int *c_a = new int[100000000];
    for(int i = 0; i < 10000; i++){
        for(int j = 0; j < 10000; j++){
            c_a[i*10000+j] = 10;
        }
    }
    MemCopy c = MemCopy();
    t.tic();
    // Do Something
    int *g_a;
    int **a = c.hostToDevice(c_a, &g_a, 10000, 10000);
    t.toc();
    delete[] c_a;
    cudaFree(g_a);
    cudaFree(a);
    return 0;
}