OpenMP locks instead of busy-waiting with NUM_PARALLEL

Makefile.rule

@@ -307,4 +307,4 @@ COMMON_PROF = -pg
 # BUILD_COMPLEX16 = 1
 #
 #  End of user configuration
-#
+#

Makefile.system

@@ -1707,11 +1707,7 @@ endif
 
 LIBDLLNAME   = $(LIBPREFIX).dll
 IMPLIBNAME   = lib$(LIBNAMEBASE).dll.a
-ifneq ($(OSNAME), AIX)
 LIBSONAME    = $(LIBNAME:.$(LIBSUFFIX)=.so)
-else
-LIBSONAME    = $(LIBNAME:.$(LIBSUFFIX)=.a)
-endif
 LIBDYNNAME   = $(LIBNAME:.$(LIBSUFFIX)=.dylib)
 LIBDEFNAME   = $(LIBNAME:.$(LIBSUFFIX)=.def)
 LIBEXPNAME   = $(LIBNAME:.$(LIBSUFFIX)=.exp)
@@ -1859,4 +1855,4 @@ ifneq ($(OSNAME), SunOS)
 SUNPATH		= /opt/sunstudio12.1
 else
 SUNPATH		= /opt/SUNWspro
-endif
+endif

cmake/system.cmake

@@ -799,4 +799,4 @@ endif ()
 #  export CUCC
 #  export CUFLAGS
 #  export CULIB
-#endif
+#endif

common.h

@@ -940,4 +940,4 @@ extern int gotoblas_profile;
 
 #endif  /* __cplusplus */
 
-#endif
+#endif

driver/level3/level3_thread.c

@@ -540,18 +540,34 @@ static int round_up(int remainder, int width, int multiple)
 	return width;
 }
 
-
 static int gemm_driver(blas_arg_t *args, BLASLONG *range_m, BLASLONG
 		       *range_n, IFLOAT *sa, IFLOAT *sb,
                        BLASLONG nthreads_m, BLASLONG nthreads_n) {
 
-#ifndef USE_OPENMP
-#ifndef OS_WINDOWS
-static pthread_mutex_t  level3_lock    = PTHREAD_MUTEX_INITIALIZER;
+#ifdef USE_OPENMP
+  static omp_lock_t level3_lock, critical_section_lock;
+  static volatile BLASLONG init_lock = 0, omp_lock_initialized = 0,
+                  parallel_section_left = MAX_PARALLEL_NUMBER;
+
+  // Lock initialization; Todo : Maybe this part can be moved to blas_init() in blas_server_omp.c
+  while(omp_lock_initialized == 0)
+  {
+    blas_lock(&init_lock);
+    {
+      if(omp_lock_initialized == 0)
+      {
+        omp_init_lock(&level3_lock);
+        omp_init_lock(&critical_section_lock);
+        omp_lock_initialized = 1;
+      }
+    blas_unlock(&init_lock);
+    }
+  }
+#elif defined(OS_WINDOWS)
+  CRITICAL_SECTION level3_lock;
+  InitializeCriticalSection((PCRITICAL_SECTION)&level3_lock);
 #else
-CRITICAL_SECTION level3_lock;
-InitializeCriticalSection((PCRITICAL_SECTION)&level3_lock);
-#endif
+  static pthread_mutex_t  level3_lock    = PTHREAD_MUTEX_INITIALIZER;
 #endif
 
   blas_arg_t newarg;
@@ -599,12 +615,21 @@ InitializeCriticalSection((PCRITICAL_SECTION)&level3_lock);
 #endif
 #endif
 
-#ifndef USE_OPENMP
-#ifndef OS_WINDOWS
-pthread_mutex_lock(&level3_lock);
+#ifdef USE_OPENMP
+  omp_set_lock(&level3_lock);
+  omp_set_lock(&critical_section_lock);
+
+  parallel_section_left--;
+
+  if(parallel_section_left != 0) 
+    omp_unset_lock(&level3_lock);
+
+  omp_unset_lock(&critical_section_lock);
+
+#elif defined(OS_WINDOWS)
+  EnterCriticalSection((PCRITICAL_SECTION)&level3_lock);
 #else
-EnterCriticalSection((PCRITICAL_SECTION)&level3_lock);
-#endif
+  pthread_mutex_lock(&level3_lock);
 #endif
 
 #ifdef USE_ALLOC_HEAP
@@ -732,12 +757,19 @@ EnterCriticalSection((PCRITICAL_SECTION)&level3_lock);
   free(job);
 #endif
 
-#ifndef USE_OPENMP
-#ifndef OS_WINDOWS
-  pthread_mutex_unlock(&level3_lock);
-#else
+#ifdef USE_OPENMP
+  omp_set_lock(&critical_section_lock);
+  parallel_section_left++;
+
+  if(parallel_section_left == 1)
+    omp_unset_lock(&level3_lock);
+
+  omp_unset_lock(&critical_section_lock);
+
+#elif defined(OS_WINDOWS)
   LeaveCriticalSection((PCRITICAL_SECTION)&level3_lock);
-#endif
+#else
+  pthread_mutex_unlock(&level3_lock);
 #endif
 
   return 0;

driver/others/blas_server_omp.c

@@ -407,7 +407,6 @@ int exec_blas(BLASLONG num, blas_queue_t *queue){
   }
 #endif
 
-  while(true) {
     for(i=0; i < MAX_PARALLEL_NUMBER; i++) {
 #ifdef HAVE_C11
       _Bool inuse = false;
@@ -420,9 +419,6 @@ int exec_blas(BLASLONG num, blas_queue_t *queue){
         break;
       }
     }
-    if(i != MAX_PARALLEL_NUMBER)
-      break;
-  }
 
 if (openblas_omp_adaptive_env() != 0) {
 #pragma omp parallel for num_threads(num) schedule(OMP_SCHED)
@@ -455,4 +451,4 @@ if (openblas_omp_adaptive_env() != 0) {
   return 0;
 }
 
-#endif
+#endif