| 1 | FFmpeg multithreading methods |
| 2 | ============================================== |
| 3 | |
| 4 | FFmpeg provides two methods for multithreading codecs. |
| 5 | |
| 6 | Slice threading decodes multiple parts of a frame at the same time, using |
| 7 | AVCodecContext execute() and execute2(). |
| 8 | |
| 9 | Frame threading decodes multiple frames at the same time. |
| 10 | It accepts N future frames and delays decoded pictures by N-1 frames. |
| 11 | The later frames are decoded in separate threads while the user is |
| 12 | displaying the current one. |
| 13 | |
| 14 | Restrictions on clients |
| 15 | ============================================== |
| 16 | |
| 17 | Slice threading - |
| 18 | * The client's draw_horiz_band() must be thread-safe according to the comment |
| 19 | in avcodec.h. |
| 20 | |
| 21 | Frame threading - |
| 22 | * Restrictions with slice threading also apply. |
| 23 | * For best performance, the client should set thread_safe_callbacks if it |
| 24 | provides a thread-safe get_buffer() callback. |
| 25 | * There is one frame of delay added for every thread beyond the first one. |
| 26 | Clients must be able to handle this; the pkt_dts and pkt_pts fields in |
| 27 | AVFrame will work as usual. |
| 28 | |
| 29 | Restrictions on codec implementations |
| 30 | ============================================== |
| 31 | |
| 32 | Slice threading - |
| 33 | None except that there must be something worth executing in parallel. |
| 34 | |
| 35 | Frame threading - |
| 36 | * Codecs can only accept entire pictures per packet. |
| 37 | * Codecs similar to ffv1, whose streams don't reset across frames, |
| 38 | will not work because their bitstreams cannot be decoded in parallel. |
| 39 | |
| 40 | * The contents of buffers must not be read before ff_thread_await_progress() |
| 41 | has been called on them. reget_buffer() and buffer age optimizations no longer work. |
| 42 | * The contents of buffers must not be written to after ff_thread_report_progress() |
| 43 | has been called on them. This includes draw_edges(). |
| 44 | |
| 45 | Porting codecs to frame threading |
| 46 | ============================================== |
| 47 | |
| 48 | Find all context variables that are needed by the next frame. Move all |
| 49 | code changing them, as well as code calling get_buffer(), up to before |
| 50 | the decode process starts. Call ff_thread_finish_setup() afterwards. If |
| 51 | some code can't be moved, have update_thread_context() run it in the next |
| 52 | thread. |
| 53 | |
| 54 | If the codec allocates writable tables in its init(), add an init_thread_copy() |
| 55 | which re-allocates them for other threads. |
| 56 | |
| 57 | Add CODEC_CAP_FRAME_THREADS to the codec capabilities. There will be very little |
| 58 | speed gain at this point but it should work. |
| 59 | |
| 60 | If there are inter-frame dependencies, so the codec calls |
| 61 | ff_thread_report/await_progress(), set AVCodecInternal.allocate_progress. The |
| 62 | frames must then be freed with ff_thread_release_buffer(). |
| 63 | Otherwise leave it at zero and decode directly into the user-supplied frames. |
| 64 | |
| 65 | Call ff_thread_report_progress() after some part of the current picture has decoded. |
| 66 | A good place to put this is where draw_horiz_band() is called - add this if it isn't |
| 67 | called anywhere, as it's useful too and the implementation is trivial when you're |
| 68 | doing this. Note that draw_edges() needs to be called before reporting progress. |
| 69 | |
| 70 | Before accessing a reference frame or its MVs, call ff_thread_await_progress(). |