clover: Run the associated action before an event is signalled.

And define a method for other threads to wait until the action
function associated with an event has been executed to completion.

For hard events, this will mean waiting until the corresponding
command has been submitted to the pipe driver, without necessarily
flushing the pipe_context and waiting for the actual command to be
processed by the GPU (which is what hard_event::wait() already does).

This weaker kind of event wait will allow implementing blocking memory
transfers efficiently.

Acked-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Jan Vesely <jan.vesely@rutgers.edu>

src/gallium/state_trackers/clover/core/event.cpp

@@ -44,19 +44,16 @@ event::trigger_self() {
    if (!--_wait_count)
       std::swap(_chain, evs);
 
+   cv.notify_all();
    return evs;
 }
 
 void
 event::trigger() {
-   auto evs = trigger_self();
-
-   if (signalled()) {
+   if (wait_count() == 1)
       action_ok(*this);
-      cv.notify_all();
-   }
 
-   for (event &ev : evs)
+   for (event &ev : trigger_self())
       ev.trigger();
 }
 
@@ -73,11 +70,9 @@ event::abort_self(cl_int status) {
 
 void
 event::abort(cl_int status) {
-   auto evs = abort_self(status);
-
    action_fail(*this);
 
-   for (event &ev : evs)
+   for (event &ev : abort_self(status))
       ev.abort(status);
 }
 
@@ -111,13 +106,18 @@ event::chain(event &ev) {
    ev.deps.push_back(*this);
 }
 
+void
+event::wait_signalled() const {
+   std::unique_lock<std::mutex> lock(mutex);
+   cv.wait(lock, [=]{ return !_wait_count; });
+}
+
 void
 event::wait() const {
    for (event &ev : deps)
       ev.wait();
 
-   std::unique_lock<std::mutex> lock(mutex);
-   cv.wait(lock, [=]{ return !_wait_count; });
+   wait_signalled();
 }
 
 hard_event::hard_event(command_queue &q, cl_command_type command,

src/gallium/state_trackers/clover/core/event.hpp

@@ -69,6 +69,7 @@ namespace clover {
       virtual cl_int status() const;
       virtual command_queue *queue() const = 0;
       virtual cl_command_type command() const = 0;
+      void wait_signalled() const;
       virtual void wait() const;
 
       virtual struct pipe_fence_handle *fence() const {