VSCode之C++ & CUDA极简环境配置(更新跨平台支持)

1. 背景
   想要了解CUDA并行计算原理，同时针对深度学习中出现一些“不支持算子”可能需要手写的需要，配置一个简单的CUDA编译环境，探索CUDA编程的范式【注：CUDA环境配置略】。
2. 结果展示
   
3. 示例代码

#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <iostream>

__global__ void VecAdd(int* A, int* B, int* C)
{
	int i = threadIdx.x;
	C[i] = A[i] + B[i];
}

void test_cuda(){
   // define data
    const int size = 3;
	int a[size] = { 1,2,3 };
	int b[size] = { 10,20,30 };
	int c[size] = { 0 };

	// define device
	int* dev_a = 0;
	int* dev_b = 0;
	int* dev_c = 0;
	// set cuda state
	cudaError_t cudaStatus;
	// select cuda device
	cudaStatus = cudaSetDevice(0);
	if (cudaStatus != cudaSuccess) {
		fprintf(stderr, "GPU device error");
		return;
	}
	// allocate memory on cuda
	cudaStatus = cudaMalloc((void**)&dev_c, size * sizeof(int));
	if (cudaStatus != cudaSuccess)
		fprintf(stderr, "device_c allocate error");
	cudaStatus = cudaMalloc((void**)&dev_a, size * sizeof(int));
	if (cudaStatus != cudaSuccess)
		fprintf(stderr,  "device_a allocate error");
	cudaStatus = cudaMalloc((void**)&dev_b, size * sizeof(int));
	if (cudaStatus != cudaSuccess)
		fprintf(stderr,  "device_b allocate error");
	// copy data from cpu to gpu
	cudaStatus = cudaMemcpy(dev_a, a, size * sizeof(int), cudaMemcpyHostToDevice);
	if (cudaStatus != cudaSuccess) {
		fprintf(stderr,  "device_a copy error");
	}
	// copy data from cpu to gpu
	cudaStatus = cudaMemcpy(dev_b, b, size * sizeof(int), cudaMemcpyHostToDevice);
	if (cudaStatus != cudaSuccess) {
		fprintf(stderr, "device_b copy error");
	}
	// core gpu operation
	VecAdd <<<1, size>>> (dev_a, dev_b, dev_c);
	// get error state
	cudaStatus = cudaGetLastError();
	if (cudaStatus != cudaSuccess) {
		fprintf(stderr, "VecAdd call error: %s\n", cudaGetErrorString(cudaStatus));
	}

	cudaStatus = cudaDeviceSynchronize();
	if (cudaStatus != cudaSuccess) {
		fprintf(stderr, "cudaDeviceSynchronize not sucess %d!\n", cudaStatus);
	}
	// copy data from cuda to cpu
	cudaStatus = cudaMemcpy(c, dev_c, size * sizeof(int), cudaMemcpyDeviceToHost);
	if (cudaStatus != cudaSuccess) {
		fprintf(stderr, "copy result to host error");
	}
	printf("{1,2,3} + {10,20,30} = {%d,%d,%d}\n", c[0], c[1], c[2]);
	cudaFree(dev_a);
	cudaFree(dev_b);
	cudaFree(dev_c); 
}

int main() {
	test_cuda();
	return 0;
}

8. 小结

1. NVCC编译cuda命令与g++编译C++较为相似，从而借鉴引入对应的include，实现Windows下cmake编译CUDA代码；
2. 示例代码展示了从CPU读取数据，在GPU端进行计算，最终传输给CPU的过程，与深度学习数据加载过程类似，是较为通用的过程；
3. 理解C++到CUDA的过渡、预加载过程，进一步从底层了解CUDA。

9. 更新(同时支持Window和Linux)
   新增跨平台编译的CMakelists.txt

cmake_minimum_required(VERSION 3.0.0)
project(demo VERSION 0.1.0)

if(MSVC)
    set(CMAKE_CUDA_COMPILER "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v10.2/bin/nvcc.exe")
    set(CUDACXX "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v10.2/bin/nvcc.exe")
    enable_language(CUDA)
    # set (CMAKE_CUDA_ARCHITECTURES "60;72;75" )
    set (CUDA_INCLUDE_DIRS )
else()
    # set(CUDA_PATH /usr/local/cuda-11.3)
    # set(CMAKE_CUDA_COMPILER /usr/local/cuda-11.3/bin/nvcc)
    find_package(CUDA REQUIRED)
    enable_language(CUDA)
endif()


add_library(a SHARED a.cu)
# add_library(advance_reduce advance_reduce.cu)
add_executable(main main.cu)
target_link_libraries(main a)