[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CAKGDhHWHqkbsmE=48otsH5=mZX431R2gVDo86kyFbHpyP==K_A@mail.gmail.com>
Date: Sat, 6 Jun 2015 14:33:55 +0200
From: Agnieszka Bielec <bielecagnieszka8@...il.com>
To: john-dev@...ts.openwall.com
Subject: Re: PHC: Lyra2 on CPU
2015-06-06 14:09 GMT+02:00 Solar Designer <solar@...nwall.com>:
> On Sat, Jun 06, 2015 at 12:41:10PM +0200, Agnieszka Bielec wrote:
>> but it looks like omp has problems with barriers or there is something
>> I don't know . I wrote simple program in C
>
> In this program, your threads share an stdio buffer for stdout. That's
> risky. You're lucky stdout is line-buffered when output is to a tty, so
> the buffer is flushed after each printf(), but still you have your
> threads work on this shared data structure.
>
> I think it's wrong for you to use this to draw conclusions about OpenMP
> barriers, given that your debugging output itself may be out of order.
>
> You could want to try write(2) (syscall, no buffering in userspace)
> instead of printf(3) (libc function, buffering in userspace). For
> "production use", write(2) has its own drawbacks (may return after
> sending only partial output, so needs to be put in a loop), but for your
> current experiment it's a fine choice.
#include <stdio.h>
#include <omp.h>
static void func()
{
char t[30];
sprintf(t,"%d %d\n",omp_get_num_threads(),omp_get_thread_num());
write(0,t,10);
write(0,"checkpoint 1\n",13);
#pragma omp barrier
write(0,"checkpoint 2\n",13);
#pragma omp barrier
write(0,"checkpoint 3\n",13);
#pragma omp barrier
write(0,"checkpoint 4\n",13);
}
int main()
{
int i;
#pragma omp parallel for
for(i=0;i<2;i++)
{
func();
}
}
output:
none@...e ~/Desktop $ ./omp
8 0
checkpoint 1
8 1
checkpoint 1
checkpoint 2
checkpoint 2
[blocks]
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
