diff options
| author | Ondrej Jirman | 2026-07-15 00:14:03 +0200 |
|---|---|---|
| committer | Ondrej Jirman | 2026-07-15 00:14:03 +0200 |
| commit | cac6ece0e23ac1f944dd3af266680aff0c3184df (patch) | |
| tree | 2e225924d3a75f16b34f55cedd337bc0e87710ae | |
| parent | 318ffd1d5ee20c36e4435a6038dafc45dd9ef68e (diff) | |
| -rw-r--r-- | README | 9 |
1 files changed, 9 insertions, 0 deletions
@@ -241,6 +241,15 @@ Design have completed. This is the same guarantee FFmpeg's hwaccel gets structurally from its AVBufferPool refcount. +- Reuse also waits on any external consumer of the buffer's exported + dma-buf. The kernel offers no implicit synchronisation against a + decode overwriting a buffer a compositor or GPU is still sampling + (e.g. mpv --vo=gpu holding it as an EGL image), which would tear the + picture. POLLOUT on an exported dma-buf completes only once every + fence on its reservation - including a GPU read fence - has signalled, + so the driver polls it before handing the buffer back to the decoder. + Buffers never exported carry no fd and skip this. + - Multi-slice H.264/HEVC frames use V4L2_BUF_FLAG_M2M_HOLD_CAPTURE_BUF; H.264 queues one request per slice, HEVC batches up to the SLICE_PARAMS control dimension. Both honour the driver's decode mode (slice/frame |