Inconsistency between openblas_get_num_threads and openblas_set_num_threads

[luc-j-bourhis]

Consider the following C code:

#include <cblas.h>
#include <stdio.h>

int main() {
  int initial_thread_count = openblas_get_num_threads();
  if(initial_thread_count >= 2) {
    int new_thread_count = initial_thread_count/2;
    openblas_set_num_threads(new_thread_count);
    printf("Initially: #threads = %d\n", initial_thread_count);
    printf("Setting #threads to %d: ", new_thread_count);
    int reported_threads = openblas_get_num_threads();
    if(reported_threads == new_thread_count) {
      printf("passed!\n");
    }
    else {
      printf("failed (#threads = %d)\n", reported_threads);
    }
  }
}

On my MacPro, it prints

OpenBLAS build configuration:
DYNAMIC_ARCH NO_AFFINITY Nehalem
Initially: #threads = 6
Setting #threads to 3: failed (#threads = 6)

So openblas_set_num_threads failed to reduce the number of threads. My machine is a MacPro 2010 with 6 physical cores and I compiled OpenBLAS with

make CC=clang FC=gfortran USE_THREAD=1 NUM_THREADS=16 DYNAMIC_ARCH=1 NO_STATIC=1

I specified NUM_THREADS because I wanted libopenblas.dylib to be usable on other machines with more cores. Here are the versions of the compilers:

~> clang --version
Apple LLVM version 7.0.2 (clang-700.1.81)
Target: x86_64-apple-darwin15.3.0
Thread model: posix

~> gfortran --version
GNU Fortran (MacPorts gcc5 5.3.0_0) 5.3.0

Either I am missing something badly or this is a bug!

[jeromerobert]

I see the same thing. I would say there is a bug in goto_set_num_threads (blas_server.c) because it only set the blas_num_threads global if num_threads > blas_num_threads.

[jeromerobert]

I was wrong. We have 2 globals:

• blas_num_cpu: the number of threads used
• blas_num_threads: the number of created threads

blas_num_threads can only grow (threads may be dormant). So we want openblas_get_num_threads to return blas_num_cpu not blas_num_threads.

There is also a openblas_get_num_procs function but it does not return blas_num_cpu. It return the physical number of cpu on the machine.