topic Simple matrix vector OpenMP problem in Software Archive

Simple matrix vector OpenMP problem

Mikhail_Posypkin — Fri, 21 May 2010 16:33:43 GMT

Consider the following very simple OpenMP program

#define n 5000
double a, b, c;

int _tmain (int argc, char *argv[])
{
omp_set_num_threads (1);
double start = omp_get_wtime ();

#pragma omp parallel for
for (int i = 0; i < n; i++)
{
c = 0;
for (int k = 0; k < n; k++)
{
c += a * b;
}

}
printf ("time = %lf (%d)\\n", omp_get_wtime () - start, omp_get_num_threads());
}

on my Intel Core 2 Quad it runs 0.17 s with 1 thread, 0.18 s with 2 threads, 0.18 s with 4 threads
So, no speedup.

Intel Parallel Amplifier shows bad utilization, but doesn't show the reason. How can I discover the reason with Intel tools integrated in MS Visual Studio?

Simple matrix vector OpenMP problem

afd_lml — Sat, 22 May 2010 01:34:06 GMT

please check your program in debug version or release version ?

Simple matrix vector OpenMP problem

Mikhail_Posypkin — Sat, 22 May 2010 08:17:15 GMT

Debug target is selected.

Simple matrix vector OpenMP problem

Mikhail_Posypkin — Sat, 22 May 2010 11:00:51 GMT

I selected Release Win32 - it solved problem! Many thanks!

Now (for N=1000) it works with 4 times faster with 4 threads. However I
the code is incorrect from the semantics point of view (indexes should not be shared).
Why it speeds up so poorly with Debug tagret selected?

Simple matrix vector OpenMP problem

JenniferJ — Tue, 25 May 2010 17:47:12 GMT

For inner loop, auto-vectorization from /arch:SSE2|SSE3|SSE4 or /QxSSE2... will help performance a lot.

but it's disabled in "Debug". You can use /Qvec-report[1|2|3|4|5] to show the details.
for openmp, use /Qopenmp-report[1|2]

>>icl /O2 /Qopenmp /Ob2 /Qvec-report3 /Qopenmp-report:2 u.cpp
Intel C++ Compiler for applications running on IA-32, Version 12.0.0.024 Beta Build 20100415
Copyright (C) 1985-2010 Intel Corporation. All rights reserved.

u.cpp
C:\temp\u.cpp(12): (col. 1) remark: OpenMP DEFINED LOOP WAS PARALLELIZED.
C:\temp\u.cpp(13): (col. 3) remark: loop was not vectorized: not inner loop.
C:\temp\u.cpp(16): (col. 7) remark: LOOP WAS VECTORIZED.
Microsoft Incremental Linker Version 9.00.30729.01
Copyright (C) Microsoft Corporation. All rights reserved.

You can add /Qvec-report3 /Qopenmp-report2 to the project property C/C++->Advanced->Additional option.

Jennifer