add

4 files changed, 52 insertions, 39 deletions

diff --git a/libclapi/cl_enqueue.c b/libclapi/cl_enqueue.c
index 2cffa30c..bbfa5c79 100644
--- a/libclapi/cl_enqueue.c
+++ b/libclapi/cl_enqueue.c
@@ -147,7 +147,7 @@ static void *worker_thread_function(void *Arg)
 
       /* We execute it without lock. */
       worker->in_exec = CL_TRUE;
-      pthread_mutex_unlock(&worker->mutex);
+      CL_MUTEX_UNLOCK(&worker->mutex);
 
       assert(is_ready != 0);
       if (is_ready < 0) // Error happend, just cancel.
@@ -283,8 +283,19 @@ LOCAL cl_int cl_enqueue_queue_work_item(cl_command_queue queue, cl_command_queue
   return ret;
 }
 
+/* This function will be called to set a newly created event, we do not protect the event. */
+LOCAL void cl_enqueue_set_work_item_event(cl_command_queue_work_item it, cl_event event)
+{
+  assert(it->event == NULL);
+  assert(event);
+
+  cl_retain_event(event);
+  assert(cl_event_get_status(event) == CL_QUEUED);
+  event->status = it->status;
+}
+
 LOCAL cl_command_queue_work_item cl_enqueue_create_work_item(cl_command_queue queue,
-    cl_uint num_events_in_wait_list, const cl_event *event_wait_list, cl_event event)
+    cl_uint num_events_in_wait_list, const cl_event *event_wait_list)
 {
   cl_command_queue_work_item it;
   int i;
@@ -315,12 +326,6 @@ LOCAL cl_command_queue_work_item cl_enqueue_create_work_item(cl_command_queue qu
     it->depend_event_num = num_events_in_wait_list;
   }
 
-  if (event) {
-    cl_retain_event(event);
-    it->event = event;
-    assert(cl_event_get_status(event) == CL_QUEUED);
-  }
-
   it->status = CL_QUEUED;
   it->queued = CL_FALSE;
   return it;
diff --git a/libclapi/cl_internals.h b/libclapi/cl_internals.h
index fa6942b8..2f651183 100644
--- a/libclapi/cl_internals.h
+++ b/libclapi/cl_internals.h
@@ -493,7 +493,8 @@ extern cl_int cl_enqueue_submit_work_item(cl_command_queue queue, cl_command_que
 extern cl_int cl_enqueue_run_work_item(cl_command_queue queue, cl_command_queue_work_item item);
 extern cl_int cl_enqueue_complete_work_item(cl_command_queue queue, cl_command_queue_work_item item);
 extern cl_command_queue_work_item cl_enqueue_create_work_item(cl_command_queue queue,
-    cl_uint num_events_in_wait_list, const cl_event *event_wait_list, cl_event event);
+    cl_uint num_events_in_wait_list, const cl_event *event_wait_list);
+extern void cl_enqueue_set_work_item_event(cl_command_queue_work_item it, cl_event event);
 extern void cl_enqueue_destroy_work_item(cl_command_queue queue, cl_command_queue_work_item item);
 extern cl_int cl_command_queue_worker_init(cl_command_queue queue);
 extern void cl_command_queue_worker_destroy(cl_command_queue queue);
diff --git a/libclapi/cl_kernel.c b/libclapi/cl_kernel.c
index b0853536..46b93fb8 100644
--- a/libclapi/cl_kernel.c
+++ b/libclapi/cl_kernel.c
@@ -776,19 +776,6 @@ static cl_int cl_command_queue_ND_range(cl_command_queue queue, cl_kernel kernel
   cl_command_queue_work_item it = NULL;
   cl_int err = CL_SUCCESS;
   uint32_t i;
-  cl_int ret_status;
-
-  if (event_ret) {
-    event = cl_create_event(queue->ctx, queue, CL_FALSE, num_events_in_wait_list, event_wait_list, &err);
-    if (event == NULL)
-      goto error;
-  }
-
-  it = cl_enqueue_create_work_item(queue, num_events_in_wait_list, event_wait_list, event);
-  if (it == NULL) {
-    err = CL_OUT_OF_HOST_MEMORY;
-    goto error;
-  }
 
   CL_MUTEX_LOCK(&kernel->lock);
   for (i = 0; i < kernel->arg_num; ++i)
@@ -799,26 +786,27 @@ static cl_int cl_command_queue_ND_range(cl_command_queue queue, cl_kernel kernel
     }
   CL_MUTEX_UNLOCK(&kernel->lock);
 
+  it = cl_enqueue_create_work_item(queue, num_events_in_wait_list, event_wait_list);
+  if (it == NULL) {
+    err = CL_OUT_OF_HOST_MEMORY;
+    goto error;
+  }
+
   err = queue->device->driver->enqueue_nd_range_kernel(queue, kernel, work_dim, global_wk_off,
         global_wk_sz, local_wk_sz, it);
   if (err != CL_SUCCESS) {
     goto error;
   }
 
-  /* If no events depend, we submit it immediately. */
-  if (event_wait_list == NULL) {
-    ret_status = cl_enqueue_submit_work_item(queue, it);
-    if (ret_status < 0) {
-      /* We consider it as a error and return fail. */
-      err = CL_OUT_OF_RESOURCES;
+  if (event_ret) {
+    event = cl_create_event(queue->ctx, queue, CL_FALSE, num_events_in_wait_list, event_wait_list, &err);
+    if (event == NULL)
       goto error;
-    }
-    if (ret_status > CL_COMPLETE) {
-      err = cl_enqueue_insert_work_item(queue, it);
-      if (err != CL_SUCCESS)
-        goto error;
-    }
-  } else {
+  }
+  cl_enqueue_set_work_item_event(it, event);
+
+
+  if (it->status > CL_COMPLETE) { // Still something todo
     err = cl_enqueue_insert_work_item(queue, it);
     if (err != CL_SUCCESS)
       goto error;
diff --git a/libclapi/cl_mem.c b/libclapi/cl_mem.c
index 0eebf6e8..441a7c3c 100644
--- a/libclapi/cl_mem.c
+++ b/libclapi/cl_mem.c
@@ -530,11 +530,34 @@ static void* cl_enqueue_map_buffer(cl_command_queue queue, cl_mem buffer, cl_boo
   void *mem_ptr = NULL;
   cl_int index;
 
+
+  if (blocking_map) {
+    /* According to spec, when in block mode, we need to ensure all the
+     commands in queue are flushed. */
+    err = cl_enqueue_wait_for_flush(queue);
+    if (err != CL_SUCCESS)
+      goto error;
+  }
+
+  it = cl_enqueue_create_work_item(queue, num_events, event_list);
+  if (it == NULL) {
+    err = CL_OUT_OF_HOST_MEMORY;
+    goto error;
+  }
+
+  err = queue->device->driver->enqueue_map_buffer(queue, buffer, &mem_ptr, blocking_map, map_flags,
+        offset, size, num_events, event_list, event);
+  if (err != CL_SUCCESS)
+    goto error;
+
   if (event_ret) {
     event = cl_create_event(buffer->ctx, queue, CL_FALSE, num_events, event_list, &err);
     if (event == NULL)
       goto error;
   }
+
+  cl_enqueue_set_work_item_event(it, event);
+
 #if 0
   if (blocking_map) {
     /* According to spec, when in block mode, we need to ensure all the
@@ -558,10 +581,6 @@ static void* cl_enqueue_map_buffer(cl_command_queue queue, cl_mem buffer, cl_boo
     goto error;
   }
 
-  err = queue->device->driver->enqueue_map_buffer(queue, buffer, &mem_ptr, blocking_map, map_flags,
-        offset, size, num_events, event_list, event);
-  if (err != CL_SUCCESS)
-    goto error;