try at new threading model

CONTRIBUTORS.md

@@ -216,3 +216,6 @@ In chronological order:
   
 * Pablo Romero <https://github.com/pablorcum>
   * [2022-08] Fix building from sources for QNX
+
+* Mark Seminatore <https://github.com/mseminatore>
+  * [2023-11-09] Improve Windows threading performance scaling

common_thread.h

@@ -111,8 +111,9 @@ typedef struct blas_queue {
   struct blas_queue *next;
 
 #if defined( __WIN32__) || defined(__CYGWIN32__) || defined(_WIN32) || defined(__CYGWIN__)
-  CRITICAL_SECTION lock;
+  // CRITICAL_SECTION lock;
-  HANDLE finish;
+  // HANDLE finish;
+  volatile int finished;
 #else
   pthread_mutex_t	 lock;
   pthread_cond_t	 finished;

driver/others/blas_server_win32.c

@@ -51,15 +51,19 @@
 /* This is a thread implementation for Win32 lazy implementation */
 
 /* Thread server common information */
-typedef struct{
+//typedef struct{
-  CRITICAL_SECTION lock;
+//  CRITICAL_SECTION lock;
-  HANDLE filled;
+//  HANDLE filled;
-  HANDLE killed;
+//  HANDLE killed;
+//
+//  blas_queue_t	*queue;    /* Parameter Pointer */
+//  int		shutdown;  /* server shutdown flag */
+//
+//} blas_pool_t;
 
-  blas_queue_t	*queue;    /* Parameter Pointer */
+static blas_queue_t *work_queue = NULL;
-  int		shutdown;  /* server shutdown flag */
+static HANDLE kickoff_event = NULL;
-
+static CRITICAL_SECTION queue_lock;
-} blas_pool_t;
 
 /* We need this global for checking if initialization is finished.   */
 int blas_server_avail = 0;
@@ -67,7 +71,7 @@ int blas_server_avail = 0;
 /* Local Variables */
 static BLASULONG server_lock       = 0;
 
-static blas_pool_t   pool;
+//static blas_pool_t   pool;
 static HANDLE	    blas_threads   [MAX_CPU_NUMBER];
 static DWORD	    blas_threads_id[MAX_CPU_NUMBER];
 
@@ -209,7 +213,7 @@ static DWORD WINAPI blas_thread_server(void *arg){
   void *buffer, *sa, *sb;
   blas_queue_t	*queue;
   DWORD action;
-  HANDLE handles[] = {pool.filled, pool.killed};
+  //HANDLE handles[] = {pool.filled, pool.killed};
 
   /* Each server needs each buffer */
   buffer   = blas_memory_alloc(2);
@@ -225,29 +229,38 @@ static DWORD WINAPI blas_thread_server(void *arg){
 #ifdef SMP_DEBUG
     fprintf(STDERR, "Server[%2ld] Waiting for Queue.\n", cpu);
 #endif
-
+	// event raised when work is added to the queue
-    do {
+	WaitForSingleObject(kickoff_event, INFINITE);
-      action = WaitForMultipleObjects(2, handles, FALSE, INFINITE);
-    } while ((action != WAIT_OBJECT_0) && (action != WAIT_OBJECT_0 + 1));
-
-    if (action == WAIT_OBJECT_0 + 1) break;
 
 #ifdef SMP_DEBUG
     fprintf(STDERR, "Server[%2ld] Got it.\n", cpu);
 #endif
 
-    EnterCriticalSection(&pool.lock);
+#if 1
+    EnterCriticalSection(&queue_lock);
 
-    queue = pool.queue;
+    queue = work_queue;
-    if (queue) pool.queue = queue->next;
+    if (queue)
+        work_queue = work_queue->next;
 
-    LeaveCriticalSection(&pool.lock);
+    LeaveCriticalSection(&queue_lock);
+#else
+    volatile work_queue_t* queue_next;
+
+    INT_PTR prev_value;
+    do {
+        queue = (volatile work_queue_t*)work_queue;
+        if (!queue)
+            break;
+
+        queue_next = (volatile work_queue_t*)queue->next;
+        prev_value = WIN_CAS((INT_PTR*)&work_queue, (INT_PTR)queue_next, (INT_PTR)queue);
+    } while (prev_value != work_item);
+#endif
 
     if (queue)  {
       int (*routine)(blas_arg_t *, void *, void *, void *, void *, BLASLONG) = queue -> routine;
 
-      if (pool.queue) SetEvent(pool.filled);
-
       sa = queue -> sa;
       sb = queue -> sb;
 
@@ -331,14 +344,6 @@ static DWORD WINAPI blas_thread_server(void *arg){
 #ifdef SMP_DEBUG
     fprintf(STDERR, "Server[%2ld] Finished!\n", cpu);
 #endif
-
-    EnterCriticalSection(&queue->lock);
-
-    queue -> status = BLAS_STATUS_FINISHED;
-
-    LeaveCriticalSection(&queue->lock);
-
-    SetEvent(queue->finish);
   }
 
   /* Shutdown procedure */
@@ -366,13 +371,10 @@ int blas_thread_init(void){
 #endif
 
   if (!blas_server_avail){
+	// create the kickoff Event
+	kickoff_event = CreateEvent(NULL, TRUE, FALSE, NULL);
 
-    InitializeCriticalSection(&pool.lock);
+    InitializeCriticalSection(&queue_lock);
-    pool.filled = CreateEvent(NULL, FALSE, FALSE, NULL);
-    pool.killed = CreateEvent(NULL, TRUE,  FALSE, NULL);
-
-    pool.shutdown = 0;
-    pool.queue    = NULL;
 
     for(i = 0; i < blas_cpu_number - 1; i++){
       blas_threads[i] = CreateThread(NULL, 0,
@@ -409,8 +411,6 @@ int exec_blas_async(BLASLONG pos, blas_queue_t *queue){
   current = queue;
 
   while (current) {
-    InitializeCriticalSection(&current -> lock);
-    current -> finish = CreateEvent(NULL, FALSE, FALSE, NULL);
     current -> position = pos;
 
 #ifdef CONSISTENT_FPCSR
@@ -418,23 +418,32 @@ int exec_blas_async(BLASLONG pos, blas_queue_t *queue){
     __asm__ __volatile__ ("stmxcsr %0" : "=m" (current -> sse_mode));
 #endif
 
+	current->finished = 0;
     current = current -> next;
     pos ++;
   }
 
-  EnterCriticalSection(&pool.lock);
+  EnterCriticalSection(&queue_lock);
 
-  if (pool.queue) {
+  if (!work_queue)
-    current = pool.queue;
+  {
-    while (current -> next) current = current -> next;
+      work_queue = queue;
-    current -> next = queue;
+  }
-  } else {
+  else
-    pool.queue = queue;
+  {
+	  blas_queue_t *next_item = work_queue;
+
+      // find the end of the work queue
+      while (next_item)
+          next_item = next_item->next;
+
+      // add new work to the end
+      next_item = queue;
   }
 
-  LeaveCriticalSection(&pool.lock);
+  LeaveCriticalSection(&queue_lock);
 
-  SetEvent(pool.filled);
+  SetEvent(kickoff_event);
 
   return 0;
 }
@@ -449,21 +458,26 @@ int exec_blas_async_wait(BLASLONG num, blas_queue_t *queue){
 #ifdef SMP_DEBUG
     fprintf(STDERR, "Waiting Queue ..\n");
 #endif
+	  while (!queue->finished)
+		  YIELDING;
 
-      WaitForSingleObject(queue->finish, INFINITE);
+  	  queue = queue->next;
-
+      num--;
-      CloseHandle(queue->finish);
-      DeleteCriticalSection(&queue -> lock);
-
-      queue = queue -> next;
-      num --;
     }
 
 #ifdef SMP_DEBUG
     fprintf(STDERR, "Completely Done.\n\n");
 #endif
+	// if work was added to the queue after this batch we can't sleep the worker threads
+	// by resetting the event
+	EnterCriticalSection(&queue_lock);
 
-  return 0;
+	if (work_queue == NULL)
+		ResetEvent(kickoff_event);
+
+	LeaveCriticalSection(&queue_lock);
+
+	return 0;
 }
 
 /* Execute Threads */
@@ -512,8 +526,6 @@ int BLASFUNC(blas_thread_shutdown)(void){
 
   if (blas_server_avail){
 
-    SetEvent(pool.killed);
-
     for(i = 0; i < blas_num_threads - 1; i++){
       // Could also just use WaitForMultipleObjects
       DWORD wait_thread_value = WaitForSingleObject(blas_threads[i], 50);
@@ -528,9 +540,6 @@ int BLASFUNC(blas_thread_shutdown)(void){
       CloseHandle(blas_threads[i]);
     }
 
-    CloseHandle(pool.filled);
-    CloseHandle(pool.killed);
-
     blas_server_avail = 0;
   }
 
@@ -558,13 +567,11 @@ void goto_set_num_threads(int num_threads)
 
 		//increased_threads = 1;
 	    if (!blas_server_avail){
+			// create the kickoff Event
+			kickoff_event = CreateEvent(NULL, TRUE, FALSE, NULL);
 
-			InitializeCriticalSection(&pool.lock);
+			InitializeCriticalSection(&queue_lock);
-			pool.filled = CreateEvent(NULL, FALSE, FALSE, NULL);
-			pool.killed = CreateEvent(NULL, TRUE,  FALSE, NULL);
 
-			pool.shutdown = 0;
-			pool.queue    = NULL;
 			blas_server_avail = 1;
 		}