[OpenMP] Use half of available logical processors for collapse tests

[xingxue-ibm]

The new collapse test cases define MAX_THREADS to be 256 and use all available threads/logical processors on the system. This triples the testing time on an AIX machine that has 128 logical processors. This patch changes to use half of available logical processors to avoid over subscribing because there are other libomp tests running at the same time, including 2 other such collapse tests.

Slowest Tests:
--------------------------------------------------------------------------
971.92s: libomp :: worksharing/for/omp_collapse_many_LTLEGE_int.c
971.62s: libomp :: worksharing/for/omp_collapse_many_GTGEGT_int.c
965.29s: libomp :: worksharing/for/omp_collapse_many_GELTGT_int.c
332.97s: libomp :: tasking/omp_taskloop_num_tasks.c
233.26s: libomp :: worksharing/single/omp_single.c
207.42s: libomp :: tasking/omp_taskloop_grainsize.c
178.41s: libomp :: tasking/hidden_helper_task/depend.cpp
172.02s: libomp :: worksharing/for/omp_collapse_many_int.c
....

[jprotze]

The test should not use more than 128 threads on your system (default value returned by omp_get_max_threads is number of available cores). Since we usually run multiple test in parallel, I think it makes sense to modify the test to never use more than half of the available cores:

-  unsigned num_threads = omp_get_max_threads();
+  unsigned num_threads = omp_get_max_threads()/2;
  if (num_threads > MAX_THREADS)
    num_threads = MAX_THREADS;
  omp_set_num_threads(num_threads);

Furthermore, I suggest to move the omp_get_thread_num call out of the loop:

#pragma omp parallel num_threads(num_threads)
  {
+    unsigned gtid = omp_get_thread_num();
#pragma omp for collapse(3) private(i, j, k)
    LOOP {
      unsigned count;
-      unsigned gtid = omp_get_thread_num();

[xingxue-ibm]

The test should not use more than 128 threads on your system (default value returned by omp_get_max_threads is number of available cores). Since we usually run multiple test in parallel, I think it makes sense to modify the test to never use more than half of the available cores:

-  unsigned num_threads = omp_get_max_threads();
+  unsigned num_threads = omp_get_max_threads()/2;
  if (num_threads > MAX_THREADS)
    num_threads = MAX_THREADS;
  omp_set_num_threads(num_threads);

Furthermore, I suggest to move the omp_get_thread_num call out of the loop:

#pragma omp parallel num_threads(num_threads)
  {
+    unsigned gtid = omp_get_thread_num();
#pragma omp for collapse(3) private(i, j, k)
    LOOP {
      unsigned count;
-      unsigned gtid = omp_get_thread_num();

Good point, thanks @jprotze! It makes sense not to use more than half of the available logical processors to avoid over-subscribing, noting there are other libomp test cases run running in parallel, including 2 other collapse test cases using collapse_test.inc. Changed as per suggestion. Also removed omp_set_num_threads(num_threads); because it will make the call to omp_get_max_threads() in the next iteration return half of the original value.

[jprotze]

Did you push the changes?

[xingxue-ibm]

Did you push the changes?

Having a problem pushing the changes. Working on it...

[xingxue-ibm]

Did you push the changes?

Having a problem pushing the changes. Working on it...

Changes are pushed.

[jprotze]

Lgtm now