critical 使用:
如果程式碼片段只需要一個thread執行,可以使用single標明。如果程式碼片段需要ID為0的thread執行,使用master標明。
那麼,如果程式碼片段需要各個CPU互斥執行,也就是要求每個CPU都執行一次,但任何時候只用一個CPU在執行。這種
情況可以使用critical。從宏觀上看,該程式碼片段被依次在各個CPU上被執行,各個CPU在執行該代碼的時序上沒有重疊。
測試代碼:
void testCritical()
{
omp_set_num_threads( 4 );
clock_t t1 = clock();
#pragma omp parallel
{
printf("test OpenMP/n");
#pragma omp critical
{
clock_t t11 = clock();
printf("test OpenMP critical/n");
printf("execute thread ID is %d/n", omp_get_thread_num());
for (int i = 0; i < 100000000; i++)
{
int a = i+1;
}
clock_t t22 = clock();
printf("thread time = %d/n", t22 - t11);
}
}
clock_t t2 = clock();
printf("total time = %d/n", t2 - t1);
}
運行結果:
test OpenMP
test OpenMP
test OpenMP
test OpenMP
test OpenMP critical
execute thread ID is 0
thread time = 291
test OpenMP critical
execute thread ID is 2
thread time = 290
test OpenMP critical
execute thread ID is 1
thread time = 294
test OpenMP critical
execute thread ID is 3
thread time = 286
total time = 1164
可以看出,整個啟動並執行時間為1164ms,四個thread不是同步執行critical程式碼片段,而是依次執行的。
為了對比,注釋掉critical,
void testCritical()
{
omp_set_num_threads( 4 );
clock_t t1 = clock();
#pragma omp parallel
{
printf("test OpenMP/n");
//#pragma omp critical
{
clock_t t11 = clock();
printf("test OpenMP critical/n");
printf("execute thread ID is %d/n", omp_get_thread_num());
for (int i = 0; i < 100000000; i++)
{
int a = i+1;
}
clock_t t22 = clock();
printf("thread time = %d/n", t22 - t11);
}
}
clock_t t2 = clock();
printf("total time = %d/n", t2 - t1);
}
再運行,結果為:
test OpenMP
test OpenMP critical
execute thread ID is 0
test OpenMP
test OpenMP critical
execute thread ID is 2
test OpenMP
test OpenMP critical
execute thread ID is 3
test OpenMP
test OpenMP critical
execute thread ID is 1
thread time = 224
thread time = 228
thread time = 237
thread time = 248
total time = 250
此時,total time基本上等於一個thread的運行所耗時間。