Closed danielsan901998 closed 2 months ago
nuomi2021
commented
2 months ago @danielsan901998 thank you for the patch. Are you sure the performance introduced by the ready/unready queue? We refacted the code, and every task in the queue is a ready task. https://github.com/ffvvc/FFmpeg/blob/e81b6d78fc2ddf8edd53a6a052713354ef8d27c2/libavcodec/vvc/vvc_thread.c#L199C9-L199C17
danielsan901998
commented
2 months ago | After running the benchmark this time i observed the reverse result, confirming that it was not a real improvement: | video | before | after |
|---|---|---|---|
| BQTerrace_1920x1080_60_10_420_22_RA.vvc | 81.7 | 80.0 | |
| NovosobornayaSquare_1920x1080.bin | 151.3 | 148.0 | |
| RitualDance_1920x1080_60_10_420_37_RA.266 | 123.0 | 124.0 | |
| Tango2_3840x2160_60_10_420_27_LD.266 | 31.0 | 31.0 | |
| RitualDance_1920x1080_60_10_420_32_LD.266 | 142.3 | 135.7 | |
| Chimera_8bit_1080P_1000_frames.vvc | 143.7 | 141.0 |
Running multiple times the benchmark makes me think that this must have been statistical noise, i use a laptop and maybe background task influence the result, that explain why i did not see a difference in performance when calling the update in av_executor_execute, when i did the last change i must have misinterpreted what might have been a random noise that increased the FPS in the result with an improvement in performance.
Since now every task is a ready task, does that mean that dividing the queue in ready and unready task is no longer desired? Should I close the pull request?
nuomi2021
commented
2 months ago Laptop will reduce frequency when the workload is high. Still thank you for help on this. Please help check other issues. thank you
In the current form when adding a task first it is checked if it is ready, and depending on the result it decide to which queue to add it, and only notify a worker when adding it to the ready queue, then in the worker, when calling run_one_task in the case there is no ready task it call update_tasks to try to find a ready task to execute.
I tried to only call the update_tasks function in av_executor_execute after adding the task to the unready queue, but it did not improve the performance unlike this method of only calling update_tasks when a worker do not find ready tasks, i think this is because in this case the unready queue is only checked when a worker does not find a task, but in the case of executing update_tasks after any task added it does unnecessary work checking the unready queue even when the ready queue is still not empty and all the workers are executing tasks.