cuda + openmp 적용 예제

어제 했던 소스에서 대충 openmp 적용해봄
2014/01/16 - [Programming/openCL / CUDA] - cuda 1차원 데이터, 2차원 처리 예제

CUDA Runtime API - Host - Extra C++ Options 에서 /openmp 를 추가하면 된다.

[링크 : https://devtalk.nvidia.com/.../vs2010-cuda4rc2-howto-compile-openmp-host-code...]

물론 #inclued <omp.h>를 추가하지 않으면
컴파일 시에는 문제가 없으나 실행시에 이런 문제가 발생한다.

[링크 : http://blog.naver.com/changfull7/70110120004]

openmp 적용

cpu Time : 0.048000 gpu Time : 0.001000

openmp 미적용

cpu Time : 0.131000 gpu Time : 0.001000

#include < stdio.h >
#include < stdlib.h >
#include < time.h >

#include < omp.h >

#include "cuda_runtime.h"

#define BLOCK_WID	128
#define THREAD_WID	32
#define ARRAY_SIZE	 (THREAD_WID * THREAD_WID * BLOCK_WID * BLOCK_WID)

__global__ void kernel_test(int *a, int *b, int *c)
{
	int idx = threadIdx.x +blockIdx.x * blockDim.x + (gridDim.x * blockDim.x) * (blockIdx.y * blockDim.y + threadIdx.y);
	c[idx] = a[idx] + b[idx];
}

void main()
{
	clock_t start_time, end_time;
	int *a, *b, *c, *res;
	int *dev_a,*dev_b,*dev_c;
	int idx = 0;
	dim3 block(BLOCK_WID,BLOCK_WID);
	dim3 thread(THREAD_WID,THREAD_WID);

	a = (int *)malloc(ARRAY_SIZE * sizeof(int));
	b = (int *)malloc(ARRAY_SIZE * sizeof(int));
	c = (int *)malloc(ARRAY_SIZE * sizeof(int));
	res = (int *)malloc(ARRAY_SIZE * sizeof(int));

	// initialize
	srand (time(NULL));

#pragma omp parallel for
	for(idx = 0;idx < ARRAY_SIZE ; idx++)
	{
		a[idx] = rand() & 0xFFFF;
		b[idx] = rand() & 0xFFFF;
		c[idx] = 0;
	}

	start_time = clock();

#pragma omp parallel for
	for(idx = 0;idx < ARRAY_SIZE ; idx++)
	{
		res[idx] = a[idx] + b[idx];
	}
	end_time = clock();
	printf("cpu Time : %f\n", ((double)(end_time-start_time)) / CLOCKS_PER_SEC); 

	cudaMalloc(&dev_a, ARRAY_SIZE * sizeof(int));
	cudaMalloc(&dev_b, ARRAY_SIZE * sizeof(int));
	cudaMalloc(&dev_c, ARRAY_SIZE * sizeof(int));

	cudaMemcpy(dev_a, a, ARRAY_SIZE * sizeof(int), cudaMemcpyHostToDevice);
	cudaMemcpy(dev_b, b, ARRAY_SIZE * sizeof(int), cudaMemcpyHostToDevice);

	start_time = clock();
	kernel_test<<<block,thread>>>(dev_a,dev_b,dev_c);
	end_time = clock();
	printf("gpu Time : %f\n", ((double)(end_time-start_time)) / CLOCKS_PER_SEC); 

	cudaMemcpy(c, dev_c, ARRAY_SIZE * sizeof(int), cudaMemcpyDeviceToHost);

	for(idx = 0;idx < ARRAY_SIZE ; idx++)
	{
		if(res[idx] != c[idx])
		{
			printf("%5d a:%5d b:%5d c:%5d", idx, a[idx], b[idx], c[idx]);
			if(res[idx] != c[idx])
					printf(" != ");
			else	printf(" == ");
			printf("r:%5d\n",res[idx]);
			break;
		}
	}

	cudaFree(dev_a);
	cudaFree(dev_b);
	cudaFree(dev_c);

	free(a);
	free(b);
	free(c);
	free(res);
}