[Mesa-dev] [PATCH kmscube 2/2] video: fencing to avoid passing frame back to decoder too early

Rob Clark robdclark at gmail.com
Thu May 4 13:39:16 UTC 2017


With atomic kms backend, userspace can get further ahead of the gpu.  In
the decoder, we unref the previous frame when retrieving the current
frame.  If userspace gets too far ahead, this can happen while the gpu
is still sampling from the previous frame.

Simple solution is add a fence.

Signed-off-by: Rob Clark <robdclark at gmail.com>
---
 cube-video.c | 15 ++++++++++++++-
 1 file changed, 14 insertions(+), 1 deletion(-)

diff --git a/cube-video.c b/cube-video.c
index 095544d..03bfb80 100644
--- a/cube-video.c
+++ b/cube-video.c
@@ -49,6 +49,8 @@ struct {
 	struct decoder *decoder;
 	int filenames_count, idx;
 	const char *filenames[32];
+
+	EGLSyncKHR last_fence;
 } gl;
 
 static const struct egl *egl = &gl.egl;
@@ -223,6 +225,12 @@ static void draw_cube_video(unsigned i)
 	ESMatrix modelview;
 	EGLImage frame;
 
+	if (gl.last_fence) {
+		egl->eglClientWaitSyncKHR(egl->display, gl.last_fence, 0, EGL_FOREVER_KHR);
+		egl->eglDestroySyncKHR(egl->display, gl.last_fence);
+		gl.last_fence = NULL;
+	}
+
 	frame = video_frame(gl.decoder);
 	if (!frame) {
 		/* end of stream */
@@ -289,6 +297,8 @@ static void draw_cube_video(unsigned i)
 	glDrawArrays(GL_TRIANGLE_STRIP, 12, 4);
 	glDrawArrays(GL_TRIANGLE_STRIP, 16, 4);
 	glDrawArrays(GL_TRIANGLE_STRIP, 20, 4);
+
+	gl.last_fence = egl->eglCreateSyncKHR(egl->display, EGL_SYNC_FENCE_KHR, NULL);
 }
 
 const struct egl * init_cube_video(const struct gbm *gbm, const char *filenames)
@@ -300,7 +310,10 @@ const struct egl * init_cube_video(const struct gbm *gbm, const char *filenames)
 	if (ret)
 		return NULL;
 
-	if (egl_check(&gl.egl, glEGLImageTargetTexture2DOES))
+	if (egl_check(&gl.egl, glEGLImageTargetTexture2DOES) ||
+	    egl_check(egl, eglCreateSyncKHR) ||
+	    egl_check(egl, eglDestroySyncKHR) ||
+	    egl_check(egl, eglClientWaitSyncKHR))
 		return NULL;
 
 	fnames = strdup(filenames);
-- 
2.9.3



More information about the mesa-dev mailing list