一. 利用Makefile
- 安装cuda toolkit之后,会有samples文件夹,存储了cuda的一些demo,便于学习与理解。
- 编译步骤
- 在单个sample文件夹内,有 .cu 文件和Makefile,终端定位到文件夹,输入sudo make就可以编译了
2.运行,如./asyncAPI ,就可以看到输出的信息
[./asyncAPI] - Starting...
GPU Device 0: "Pascal" with compute capability 6.1
CUDA device [GeForce GTX 1080 Ti]
time spent executing by the GPU: 168.72
time spent by CPU in CUDA calls: 0.02
CPU executed 668184 iterations while waiting for GPU to finish
注意:
Makefile project only supported on Mac OS X and Linux Platforms)
如果在windows里面使用,配合VS则会很方便
二.利用CMakeList
使用这种方法可以在Clion中编译cuda,也是爽歪歪!可以在CLion中新建一个cuda project
这里给出一个简单的demo
- CMakeList.txt
其中include路径由安装时决定
cmake_minimum_required(VERSION 3.16)
project(untitled_cuda CUDA)
set(CMAKE_CUDA_STANDARD 14)
#include_directories(includes/math)
include_directories(/usr/local/cuda-10.2)
include_directories(/usr/local/cuda-10.2/samples/common/inc)
add_executable(untitled_cuda main.cu)
set_target_properties(
untitled_cuda
PROPERTIES
CUDA_SEPARABLE_COMPILATION ON
)
- main.cu
#include <iostream>
#include <cmath>
#include <cuda_runtime.h>
// includes, project
#include <helper_cuda.h>
#include <helper_functions.h> // helper utility functions
using namespace std;
// 加法kernel
// cuda kernel 里面无法使用cout!
__global__ void increment_kernel(int *g_data, int inc_value)
{
int idx = blockIdx.x * blockDim.x + threadIdx.x;
g_data[idx] = g_data[idx] + inc_value;
}
// cpu线程校验计算结果
bool correct_output(int *data, const int n, const int x)
{
for (int i = 0; i < n; i++)
if (data[i] != x)
{
printf("Error! data[%d] = %d, ref = %d\n", i, data[i], x);
return false;
}
return true;
}
int main(int argc, char *argv[]) {
std::cout << "Hello, World!" << std::endl;
int devID;
cudaDeviceProp deviceProps;
printf("[%s] - Starting...\n", argv[0]);
// This will pick the best possible CUDA capable device
devID = findCudaDevice(argc, (const char **)argv);
// get device name
checkCudaErrors(cudaGetDeviceProperties(&deviceProps, devID));
printf("CUDA device [%s]\n", deviceProps.name);
int n = 16 * 1024 * 1024;
int nbytes = n * sizeof(int);
int value = 26;
// allocate host memory
int *a = 0; // cpu host内存上的变量
checkCudaErrors(cudaMallocHost((void **)&a, nbytes));
memset(a, 0, nbytes);
// allocate device memory
int *d_a=0; // gpu设备显存上的变量
checkCudaErrors(cudaMalloc((void **)&d_a, nbytes));
checkCudaErrors(cudaMemset(d_a, 255, nbytes));
// set kernel launch configuration
dim3 threads = dim3(512, 1);
dim3 blocks = dim3(n / threads.x, 1);
// create cuda event handles
cudaEvent_t start, stop;
checkCudaErrors(cudaEventCreate(&start));
checkCudaErrors(cudaEventCreate(&stop));
StopWatchInterface *timer = NULL;
sdkCreateTimer(&timer);
sdkResetTimer(&timer);
checkCudaErrors(cudaDeviceSynchronize());
float gpu_time = 0.0f;
// asynchronously issue work to the GPU (all to stream 0)
sdkStartTimer(&timer);
cudaEventRecord(start, 0);
cudaMemcpyAsync(d_a, a, nbytes, cudaMemcpyHostToDevice, 0);
increment_kernel<<<blocks, threads, 0, 0>>>(d_a, value);
cudaMemcpyAsync(a, d_a, nbytes, cudaMemcpyDeviceToHost, 0);
cudaEventRecord(stop, 0);
sdkStopTimer(&timer);
// have CPU do some work while waiting for stage 1 to finish
unsigned long int counter=0;
while (cudaEventQuery(stop) == cudaErrorNotReady)
{
counter++;
}
checkCudaErrors(cudaEventElapsedTime(&gpu_time, start, stop));
// print the cpu and gpu times
printf("time spent executing by the GPU: %.2f\n", gpu_time);
printf("time spent by CPU in CUDA calls: %.2f\n", sdkGetTimerValue(&timer));
printf("CPU executed %lu iterations while waiting for GPU to finish\n", counter);
// check the output for correctness
bool bFinalResults = correct_output(a, n, value);
// release resources
checkCudaErrors(cudaEventDestroy(start));
checkCudaErrors(cudaEventDestroy(stop));
checkCudaErrors(cudaFreeHost(a));
checkCudaErrors(cudaFree(d_a));
exit(bFinalResults ? EXIT_SUCCESS : EXIT_FAILURE);
}