metal : fix event synchronization (ggml-org#22260)

Author: ggerganov

ggml/src/ggml-metal/ggml-metal-device.m

@@ -931,29 +931,29 @@ void ggml_metal_device_rsets_keep_alive(ggml_metal_device_t dev) {
 }
 
 struct ggml_metal_event {
-    void * obj; // id<MTLEvent>
+    void * obj; // id<MTLSharedEvent>
 
     atomic_int value;
 };
 
 void ggml_metal_event_encode_signal(ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf_raw) {
-    id<MTLEvent> event = (id<MTLEvent>)ev->obj;
+    id<MTLSharedEvent> event = (id<MTLSharedEvent>)ev->obj;
 
     id<MTLCommandBuffer> cmd_buf = (id<MTLCommandBuffer>) cmd_buf_raw;
 
     [cmd_buf encodeSignalEvent:event value:atomic_fetch_add_explicit(&ev->value, 1, memory_order_relaxed) + 1];
 }
 
 void ggml_metal_event_encode_wait(ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf_raw) {
-    id<MTLEvent> event = (id<MTLEvent>)ev->obj;
+    id<MTLSharedEvent> event = (id<MTLSharedEvent>)ev->obj;
 
     id<MTLCommandBuffer> cmd_buf = (id<MTLCommandBuffer>) cmd_buf_raw;
 
     [cmd_buf encodeWaitForEvent:event value:atomic_load_explicit(&ev->value, memory_order_relaxed)];
 }
 
 ggml_metal_event_t ggml_metal_device_event_init(ggml_metal_device_t dev) {
-    id<MTLEvent> event = [dev->mtl_device newEvent];
+    id<MTLSharedEvent> event = [dev->mtl_device newSharedEvent];
 
     ggml_metal_event_t ev = calloc(1, sizeof(struct ggml_metal_event));
 
@@ -964,7 +964,7 @@ ggml_metal_event_t ggml_metal_device_event_init(ggml_metal_device_t dev) {
 }
 
 void ggml_metal_device_event_free(ggml_metal_device_t dev, ggml_metal_event_t ev) {
-    id<MTLEvent> event = ev->obj;
+    id<MTLSharedEvent> event = ev->obj;
     [event release];
 
     free(ev);
@@ -973,14 +973,13 @@ void ggml_metal_device_event_free(ggml_metal_device_t dev, ggml_metal_event_t ev
 }
 
 void ggml_metal_device_event_synchronize(ggml_metal_device_t dev, ggml_metal_event_t ev) {
-    @autoreleasepool {
-        id<MTLEvent> event = ev->obj;
-
-        id<MTLCommandBuffer> cmd_buf = [dev->mtl_queue commandBuffer];
-        [cmd_buf encodeWaitForEvent:event value:atomic_load_explicit(&ev->value, memory_order_relaxed)];
-        [cmd_buf commit];
-        [cmd_buf waitUntilCompleted];
+    id<MTLSharedEvent> event = ev->obj;
+    const bool res = [event waitUntilSignaledValue:atomic_load_explicit(&ev->value, memory_order_relaxed) timeoutMS:60000];
+    if (!res) {
+        GGML_ABORT("%s: failed to wait for event\n", __func__);
     }
+
+    GGML_UNUSED(dev);
 }
 
 void ggml_metal_device_get_memory(ggml_metal_device_t dev, size_t * free, size_t * total) {