使用 C 和 OpenMP 進行並行編程

Question

我正在嘗試將其轉換為 openMP，但我無法正確轉換它，任何幫助將不勝感激。 我現在可以將簡單循環轉換為 openMp，但不能使用函數。 所以試圖了解它是如何工作的。 謝謝！

#include <omp.h>
#include <stdio.h>
#include <stdlib.h>
#include <sys/time.h>

#define FILTERLENGTH 960
#define TRACE_LENGTH 16384
#define TRACE_COUNT 32
#define TRACE_SPACE (TRACE_LENGTH + FILTERLENGTH)

float f[FILTERLENGTH];
float d[TRACE_COUNT][TRACE_SPACE];
float out[TRACE_COUNT][TRACE_LENGTH];

void filter(int FL, float filt[FL], int NT, int TL, float traces[NT][TL + FL], float out[NT][TL]);

int main(int argc, char *argv[])
{
    #pragma omp parallel
    {
        int i, j, k, nc, id;
        struct timeval start, stop, elapse;

        float fmax = (float)RAND_MAX;
        #pragma omp parallel for
        for (k = 0; k < FILTERLENGTH; k++)
            f[k] = k - (FILTERLENGTH - 1) / 2.0;
        for (j = 0; j < TRACE_COUNT; j++)
        {
            #pragma omp parallel for
            for (i = 0; i < FILTERLENGTH / 2; i++)
                d[j][i] = 0;
            for (i = FILTERLENGTH / 2; i < TRACE_LENGTH + FILTERLENGTH / 2; i++)
                d[j][i] = rand() / fmax;
            for (i = TRACE_LENGTH + FILTERLENGTH / 2; i < TRACE_SPACE; i++)
                d[j][i] = 0;
        }
        gettimeofday(&start, NULL);
        filter(FILTERLENGTH, f, TRACE_COUNT, TRACE_LENGTH, d, out);
        gettimeofday(&stop, NULL);
        timersub(&stop, &start, &elapse);

        fprintf(stderr, "Filter length\t%d\n", FILTERLENGTH);
        fprintf(stderr, "Trace length\t%d\n", TRACE_LENGTH);
        fprintf(stderr, "Trace count\t%d\n", TRACE_COUNT);
        fprintf(stderr, "Elapse time\t%g\n", elapse.tv_sec + 0.000001 * elapse.tv_usec);
    }
}

void filter(int FL, float filt[FL], int NT, int TL, float traces[NT][TL + FL], float out[NT][TL])
{
    int i, j, k, nc;
    for (j = 0; j < NT; j++)
        for (i = 0; i < TL; i++)
            for (k = 0; k < FL; k++)
                out[j][i] = filt[k] * traces[j][i + k];
}

Answer 1

您的代碼存在一些問題，在 function main中，循環的迭代沒有按照您的意願分配給線程。 因為您再次添加了與#pragma omp for parallel的子句，並假設您已禁用嵌套並行性，默認情況下，在外部並行區域中創建的每個線程都將“按順序”執行該區域內的代碼。 有關它的更多詳細信息，請閱讀此SO 線程。

除此之外，這段代碼：

        for (int j = 0; j < TRACE_COUNT; j++)
        {
            #pragma omp for
            for (int i = 0; i < FILTERLENGTH / 2; i++)
                d[j][i] = 0;

            for (int i = FILTERLENGTH / 2; i < TRACE_LENGTH + FILTERLENGTH / 2; i++)
                d[j][i] = rand() / fmax;
            for (i = TRACE_LENGTH + FILTERLENGTH / 2; i < TRACE_SPACE; i++)
                d[j][i] = 0;
        }

可以通過將#pragma omp for移到外循環來改進：

        #pragma omp for
        for (int j = 0; j < TRACE_COUNT; j++)
        {
            for (int i = 0; i < FILTERLENGTH / 2; i++)
                d[j][i] = 0;
            for (int i = FILTERLENGTH / 2; i < TRACE_LENGTH + FILTERLENGTH / 2; i++)
                d[j][i] = rand() / fmax;
            for (int i = TRACE_LENGTH + FILTERLENGTH / 2; i < TRACE_SPACE; i++)
                d[j][i] = 0;
        }

並且應該減少並行區域的scope。 一切都放在一起：

#include <omp.h>
#include <stdio.h>
#include <stdlib.h>
#include <sys/time.h>

#define FILTERLENGTH 960
#define TRACE_LENGTH 16384
#define TRACE_COUNT 32
#define TRACE_SPACE (TRACE_LENGTH + FILTERLENGTH)

float f[FILTERLENGTH];
float d[TRACE_COUNT][TRACE_SPACE];
float out[TRACE_COUNT][TRACE_LENGTH];

void filter(int FL, float filt[FL], int NT, int TL, float traces[NT][TL + FL], float out[NT][TL]);

int main(int argc, char *argv[])
{
    #pragma omp parallel
    {
        float fmax = (float)RAND_MAX;
        #pragma omp for nowait
        for (int k = 0; k < FILTERLENGTH; k++)
            f[k] = k - (FILTERLENGTH - 1) / 2.0;

        #pragma omp for nowait
        for (int j = 0; j < TRACE_COUNT; j++)
        {    
            for (int i = 0; i < FILTERLENGTH / 2; i++)
                d[j][i] = 0;
            for (int i = FILTERLENGTH / 2; i < TRACE_LENGTH + FILTERLENGTH / 2; i++)
                d[j][i] = rand() / fmax;
            for (int i = TRACE_LENGTH + FILTERLENGTH / 2; i < TRACE_SPACE; i++)
                d[j][i] = 0;
        }
     }
     struct timeval start, stop, elapse;
     gettimeofday(&start, NULL);
     filter(FILTERLENGTH, f, TRACE_COUNT, TRACE_LENGTH, d, out);
     gettimeofday(&stop, NULL);
     timersub(&stop, &start, &elapse);

     fprintf(stderr, "Filter length\t%d\n", FILTERLENGTH);
     fprintf(stderr, "Trace length\t%d\n", TRACE_LENGTH);
     fprintf(stderr, "Trace count\t%d\n", TRACE_COUNT);
     fprintf(stderr, "Elapse time\t%g\n", elapse.tv_sec + 0.000001 * elapse.tv_usec);

}

您仍然可以嘗試並行化 function filter 。