[vaapi] av_hwframe_map with AV_PIX_FMT_PRIME_FD calls vaExportSurfaceHandle, but not vaSyncSurface

[goffioul]

When mapping a VA-API frame with av_hwframe_map and AV_PIX_FMT_PRIME_FD, vaExportSurfaceHandle is called to create the file descriptor, but it does not call vaSyncSurface. As a result, mmap'ing the file descriptor and reading from it may not yield correct result and lead to tearing artefact.

Technically, the client app could probably call vaSyncSurface itself, but that would require accessing internals of ffmpeg, e.g. to access the VA display. It would be cleaner if the VA-API bits were entirely managed by ffmpeg.

[Fei Wang]

Could you verify if the patch fixed your issue? Thanks!
​https://github.com/intel-media-ci/ffmpeg/pull/626

[goffioul]

@wangfei: I'm sorry, I won't be able to confirm, my client code has moved on too much and is not mapping anymore the VA-API frame to a prime fd. I can however confirm that at the time I submitted this bug, I was using the following patch locally, which effectively avoided the tearing artefact. While not being identical, it should be functionally equivalent. The main difference is that vaSynSurface was called aftervaExportSurfaceHandle.

diff --git a/libavutil/hwcontext_vaapi.c b/libavutil/hwcontext_vaapi.c
index 01ac91c33b..e2d0217fe1 100644
--- a/libavutil/hwcontext_vaapi.c
+++ b/libavutil/hwcontext_vaapi.c
@@ -1306,6 +1306,7 @@ static int vaapi_map_to_drm_esh(AVHWFramesContext *hwfc, AVFrame *dst,
                "%d (%s).\n", surface_id, vas, vaErrorStr(vas));
         return AVERROR(EIO);
     }
+    vaSyncSurface(hwctx->display, surface_id);

     drm_desc = av_mallocz(sizeof(*drm_desc));
     if (!drm_desc) {

[Fei Wang]

Replying to goffioul:

@wangfei: I'm sorry, I won't be able to confirm, my client code has moved on too much and is not mapping anymore the VA-API frame to a prime fd. I can however confirm that at the time I submitted this bug, I was using the following patch locally, which effectively avoided the tearing artefact. While not being identical, it should be functionally equivalent. The main difference is that vaSynSurface was called aftervaExportSurfaceHandle.

diff --git a/libavutil/hwcontext_vaapi.c b/libavutil/hwcontext_vaapi.c
index 01ac91c33b..e2d0217fe1 100644
--- a/libavutil/hwcontext_vaapi.c
+++ b/libavutil/hwcontext_vaapi.c
@@ -1306,6 +1306,7 @@ static int vaapi_map_to_drm_esh(AVHWFramesContext *hwfc, AVFrame *dst,
                "%d (%s).\n", surface_id, vas, vaErrorStr(vas));
         return AVERROR(EIO);
     }
+    vaSyncSurface(hwctx->display, surface_id);

     drm_desc = av_mallocz(sizeof(*drm_desc));
     if (!drm_desc) {

Hi @goffioul, It should have same effect to sync surface before or after export its dma fd. Prefer to follow the description in libva to add it before export. I will submit my patch into ML. And thanks for your information.

[Fei Wang]

Patch upstreamed. Tks.