[07:42:36 CET] <cone-130> ffmpeg 03Gyan Doshi 07master:5282cbae61cc: doc/filters: mention input requirements for ebur128 [16:56:01 CET] <philipl> BtbN: https://github.com/philipl/FFmpeg/commit/c0b6e4cb6d6d41dbf2684891ed9dd43d9dd... [16:56:12 CET] <philipl> Removing an unnecessary stream sync [16:57:31 CET] <BtbN> Is it really unnecessary? There's also a few calls in cuvid, in the copy path, which I'm not sure if they are needed. [17:08:47 CET] <philipl> BtbN: well, none of the other cuda filters sync and kernels are inherently asynchronous [17:09:17 CET] <philipl> In practice, the next stage either doesn't care or syncs implicitly by not using async calls [17:09:36 CET] <philipl> Until you reach the encoder [17:10:07 CET] <philipl> It's also true that the copy out of cuvid is synced when it might not need to be. [17:22:55 CET] <BtbN> patch looks good to go though! [17:24:05 CET] <philipl> BtbN: thanks [17:25:47 CET] <cone-543> ffmpeg 03Philip Langdale 07master:c0b6e4cb6d6d: avfilter/vf_yadif_cuda: Remove unnecessary stream synchronisation [18:34:04 CET] <philipl> BtbN: so yeah, I don't think we need to the sync in cuvid - it's safe for a filter or nvenc to pick up the frame without the sync. And I see a 33% speedup in a dvd encoding pipeline (with yadif_cuda). [18:34:17 CET] <philipl> I also think the sync when doing hwupload is also unnecessary. [18:34:32 CET] <philipl> The one for download is surely necessary or the host will read incomplete data. [18:35:06 CET] <BtbN> Yeah, download to software is the only path where I'm sure it's neccesary. [18:35:25 CET] <philipl> Will have changes shortly [18:57:13 CET] <philipl> BtbN: not shortly. Have to run out for a bit. Will push changes later. [21:11:08 CET] <cone-543> ffmpeg 03Philip Langdale 07master:5d90d1e36ef3: avcodec/cuviddec: Remove unnecessary stream synchronisation [21:11:09 CET] <cone-543> ffmpeg 03Philip Langdale 07master:52d8f35b14bc: avutil/hcontext_cuda: Remove unnecessary stream synchronisation [00:00:00 CET] --- Sun Mar 31 2019