GitLab

While reverse playback the number of buffer has to increase since the whole GOP has to be allocated before rendering. This might bring a lot of memory pressure, mainly in hardware accelerated decoding.

For example, in order to make more efficient the dmabuf-based memory allocation, va has an allocator's pool that stores all the created and released memories. But, after reverse playback (returning to normal playback) a lot of memories aren't required any more. It should be nice to find a mechanism to invalidate that memory cache.