From 5b8e3ecf388d9f8921412822049ef90565553067 Mon Sep 17 00:00:00 2001 From: Jocelyn Turcotte Date: Thu, 28 Aug 2014 19:43:36 +0200 Subject: Wait on sync point fences instead of creating a new fence Posting a runnable on the GPU thread's message loop and creating a GL fence sync there assumes that there is a current GL context, and that synchronizing with this context will make sure that we wait for all GL context producing any of our consumed texture mailboxes. This is however not always the case like when: - The current GL context on the GPU thread is destroyed right before our runnable is handled, displaying errors on the console that glFlush needs a current context. - The GL driver will do extra scheduling and let the scene graph thread synchronize its GL command stream only with the GL context in which the fence sync was created. To remedy the situation, make sure that Chromium creates a fence sync for every sync points associated with a mailbox that we consume and do so directly in the GL context associated with the originating glInsertSyncPointCHROMIUM call. Wait for all those syncs on the Qt side afterward. This might also help with a few erratic behaviors noticed on some embedded GL drivers. Change-Id: I5fc60fcf51497477b2e1b3a535d0a141954fc6e5 Reviewed-by: Andras Becsi --- src/core/delegated_frame_node.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'src/core/delegated_frame_node.h') diff --git a/src/core/delegated_frame_node.h b/src/core/delegated_frame_node.h index 98861eefa..a031a464f 100644 --- a/src/core/delegated_frame_node.h +++ b/src/core/delegated_frame_node.h @@ -75,7 +75,7 @@ private: QExplicitlySharedDataPointer m_data; QList > m_renderPassTextures; int m_numPendingSyncPoints; - FenceSync m_mailboxesGLFence; + QMap m_mailboxGLFences; QWaitCondition m_mailboxesFetchedWaitCond; QMutex m_mutex; -- cgit v1.2.3