CUDA学习系列之cuda-runtime-api(一)

cudaMalloc  分配GPU内存到setDevice指定的设备上

cudaMallocHost  分配page locked memory，也叫做pinned momory，页锁定内存

•         页锁定内存是主机内存，CPU可直接访问

cudaMemcpy   实现Device To Host 或 Host To Device

cudaStreamCreate   创建流，是异步控制的主要手段

int main(){
    
    int device_id = 0;
    cudaSetDevice(device_id);

    float* memory_device = nullptr;
    cudaMalloc(&memory_device, 100 * sizeof(float));
    
    float* momory_host = new float[100];
    memory_host[2] = 250.0;
    cudaMemcpy(memory_device, memory_host, sizeof(float) * 100, cudaMemcpyHostToDevice);
    
    float* memory_page_locked = nullptr;
    cudaMallocHost(&momory_page_locked, 100 * sizeof(float)); // 返回的是开辟的pin memory地址
    cudaMemcpy(memory_page_locked, memory_device, sizeof(float) * 100,     cudaMemcpyDeviceToHost);

    printf("%f\n", memory_page_locked[2]);

   
    cudaFreeHost(memory_page_locked);
    delete [] memory_host;
    cudaFree(memory_device);

    return 0;

}

1、在gpu上开辟空间，将地址记录在mem_device；

2、在cpu上开辟空间，将地址记录在mem_host上，并修改该地址指向区域第二个值；

3、把mem_host所指向区域数据都复制到mem_device的所指区域；

4、在cpu上开辟一块空间，把地址记录在mem_page_locked；

5、把mem_device所指区域的数据复制回cpu上的mem_page_locked。

int main(){

    int device_id = 0;
    cudaSetDevice(device_id);

    cudaStream_t stream = nullptr;
    cudaStreamCreate(&stream);

    float* memory_device = nullptr;
    cudaMalloc(&memory_device, 100 * sizeof(float));

    float* memory_host = new float[100];
    memory_host[2] = 200.2;
    cudaMemcpyAsync(memory_device, memory_host, sizeof(float) * 100, cudaMemcpyHostToDevice, stream);

    float* memory_page_locked = nullptr;
    cudaMallocHost(& memory_page_locked, 100 * sizeof(float));
    cudaMemcpyAsync(memory_page_locked, memory_device, sizeof(float) * 100, cudaMemcpyDeviceToHost, stream);
    cudaStreamSynchronize(stream);

    printf("%f\n", memory_page_locked[2]);

    cudaFreeHost(memory_page_locked);
    cudaFree(memory_device);
    cudaStreamDestory(stream);
    delete [] memory_host;
    
    return 0;

}