Merge branch 'master' of github.com:flame/blis

common.mk

@@ -161,7 +161,7 @@ LDFLAGS      += -fopenmp
 endif
 ifeq ($(THREADING_MODEL),pthreads)
 CTHREADFLAGS := -pthread -DBLIS_ENABLE_PTHREADS
-LDFLAGS      += -pthread
+LDFLAGS      += -lpthread
 endif
 endif
 
@@ -175,7 +175,7 @@ LDFLAGS      += -openmp
 endif
 ifeq ($(THREADING_MODEL),pthreads)
 CTHREADFLAGS := -pthread -DBLIS_ENABLE_PTHREADS
-LDFLAGS      += -pthread
+LDFLAGS      += -lpthread
 endif
 endif
 
@@ -188,7 +188,7 @@ $(error OpenMP is not supported with Clang.)
 endif
 ifeq ($(THREADING_MODEL),pthreads)
 CTHREADFLAGS := -pthread -DBLIS_ENABLE_PTHREADS
-LDFLAGS      += -pthread
+LDFLAGS      += -lpthread
 endif
 endif
 

frame/base/bli_threading_pthreads.c

@@ -36,11 +36,52 @@
 
 #ifdef BLIS_ENABLE_PTHREADS
 
+#ifndef _POSIX_BARRIERS
+
+int pthread_barrier_init(pthread_barrier_t *barrier, const pthread_barrierattr_t *attr, unsigned int count)
+{
+    if( barrier == NULL ) return 0;
+    barrier->n_threads = count;
+    barrier->sense = 0;
+    barrier->threads_arrived = 0;
+    pthread_mutex_init( &barrier->mutex, NULL );
+    return 0;
+}
+
+int pthread_barrier_destroy(pthread_barrier_t *barrier)
+{
+    if( barrier == NULL ) return 0;
+    pthread_mutex_destroy( &barrier->mutex );
+    return 0;
+}
+
+int pthread_barrier_wait(pthread_barrier_t *barrier)
+{
+    if(barrier == NULL || barrier->n_threads == 1) return 0;
+    bool_t my_sense = barrier->sense;
+    dim_t my_threads_arrived;
+
+    pthread_mutex_lock( &barrier->mutex );
+    my_threads_arrived = ++(barrier->threads_arrived);
+    pthread_mutex_unlock( &barrier->mutex );
+
+    if( my_threads_arrived == barrier->n_threads ) {
+        barrier->threads_arrived = 0;
+        barrier->sense = !barrier->sense;
+    }
+    else {
+        volatile bool_t* listener = &barrier->sense;
+        while( *listener == my_sense ) {}
+    }
+    return 0;
+}
+#endif
+
 void* thread_decorator_helper( void* data_void );
 
 typedef struct thread_data
 {
-  level3_int_t func;
+  l3_int_t func;
   obj_t* alpha; 
   obj_t* a; 
   obj_t* b; 
@@ -85,11 +126,9 @@ void bli_level3_thread_decorator
      )
 {
     pthread_t* pthreads = (pthread_t*) bli_malloc(sizeof(pthread_t) * n_threads);
-    //Saying "datas" is kind of like saying "all y'all"
     thread_data_t* datas = (thread_data_t*) bli_malloc(sizeof(thread_data_t) * n_threads);
-    //pthread_attr_t* attr = (pthread_attr_t*) bli_malloc(sizeof(pthread_attr_t) * n_threads);
 
-    for( int i = 0; i < n_threads; i++ )
+    for( int i = 1; i < n_threads; i++ )
     {
         //Setup the thread data
         datas[i].func = func;
@@ -105,7 +144,9 @@ void bli_level3_thread_decorator
         pthread_create( &pthreads[i], NULL, &thread_decorator_helper, &datas[i] );
     }
 
-    for( int i = 0; i < n_threads; i++)
+    func( alpha, a, b, beta, c, cntx, cntl, thread[0] );
+
+    for( int i = 1; i < n_threads; i++)
     {
         pthread_join( pthreads[i], NULL );
     }

frame/base/bli_threading_pthreads.h

@@ -39,15 +39,17 @@
 
 #include <pthread.h>
 
-#ifdef __APPLE__
+#if !_POSIX_BARRIER
 typedef int pthread_barrierattr_t;
-typedef struct
+
+struct pthread_barrier_s
 {
     pthread_mutex_t mutex;
-    pthread_cond_t cond;
-    int count;
-    int tripCount;
-} pthread_barrier_t;
+    bool_t  sense;
+    dim_t   threads_arrived;
+    dim_t   n_threads;
+};
+typedef struct pthread_barrier_s pthread_barrier_t;
 
 
 int pthread_barrier_init(pthread_barrier_t *barrier, const pthread_barrierattr_t *attr, unsigned int count);