Merge pull request #3703 from martin-frbg/omp_adaptive

Add env variable OMP_ADAPTIVE to control OMP threadpool behaviour

driver/others/blas_server_omp.c

@@ -403,6 +403,7 @@ int exec_blas(BLASLONG num, blas_queue_t *queue){
       break;
   }
 
+if (openblas_omp_adaptive_env() != 0) {
 #pragma omp parallel for num_threads(num) schedule(OMP_SCHED)
   for (i = 0; i < num; i ++) {
 
@@ -412,6 +413,17 @@ int exec_blas(BLASLONG num, blas_queue_t *queue){
 
     exec_threads(&queue[i], buf_index);
   }
+} else {
+#pragma omp parallel for schedule(OMP_SCHED)
+  for (i = 0; i < num; i ++) {
+
+#ifndef USE_SIMPLE_THREADED_LEVEL3
+    queue[i].position = i;
+#endif
+
+    exec_threads(&queue[i], buf_index);
+  }
+}
 
 #ifdef HAVE_C11
   atomic_store(&blas_buffer_inuse[buf_index], false);

driver/others/openblas_env.c

@@ -39,6 +39,7 @@ static int openblas_env_block_factor=0;
 static int openblas_env_openblas_num_threads=0;
 static int openblas_env_goto_num_threads=0;
 static int openblas_env_omp_num_threads=0;
+static int openblas_env_omp_adaptive=0;
 
 int openblas_verbose() { return openblas_env_verbose;}
 unsigned int openblas_thread_timeout() { return openblas_env_thread_timeout;}
@@ -46,6 +47,7 @@ int openblas_block_factor() { return openblas_env_block_factor;}
 int openblas_num_threads_env() { return openblas_env_openblas_num_threads;}
 int openblas_goto_num_threads_env() { return openblas_env_goto_num_threads;}
 int openblas_omp_num_threads_env() { return openblas_env_omp_num_threads;}
+int openblas_omp_adaptive_env() { return openblas_env_omp_adaptive;}
 
 void openblas_read_env() {
   int ret=0;
@@ -79,6 +81,11 @@ void openblas_read_env() {
   if(ret<0) ret=0;
   openblas_env_omp_num_threads=ret;
 
+  ret=0;
+  if (readenv(p,"OMP_ADAPTIVE")) ret = atoi(p);
+  if(ret<0) ret=0;
+  openblas_env_omp_adaptive=ret;
+
 }