[FFmpeg-devel] [PATCH] NVENC Surface Allocation Reduction
timo at rothenpieler.org
Wed Apr 26 12:54:07 EEST 2017
Thanks, patch looks fine to me now.
While testing, I came across some very weird behavior:
The following commandline:
./ffmpeg -hwaccel cuvid -c:v h264_cuvid -i test.mkv -an -sn -c:v h264_nvenc -preset slow -qp 22 -bf 0 -f null -
Sometimes, it runs into the following error:
[h264_nvenc @ 0x3f7d8c0] Failed locking bitstream buffer: invalid param (8)
Omitting "-hwaccel cuvid" prevents it from happening.
And, that's the weird parts, omitting "-bf 0" also prevents it from happening.
The weird thing about that is, 0 is the default. And some quickly added debug printfs show that indeed the affected values do not change.
avctx->max_b_frames is 0 in both cases.
Reverting this patch also prevents it from happening, but I somehow doubt it's the fault of this patch, specially as just running that command over and over again makes it work in ~50% of the cases.
Also, the same command line, but with -bf 1 or any higher number causes:
[h264_nvenc @ 0x2c788c0] EncodePicture failed!: no encode device (1)
This is not a new issue, it is happening ever since. It only happens when -hwaccel cuvid is specified.
It's not an issue with the CUDA Frame Input code in nvenc either, as passing cuda frames via -vf hwupload_cuda works flawlessly.
It only happens in the combination of direct cuda frame input from cuvid to nvenc.
Like I said, this is not a regression from this patch, just wanted to bring it to attention, as it somehow feels like
a driver issue to me.
With this new -bf related issue, I'm not so sure about that anymore though, and I'm wondering if something in ffmpeg corrupts memory somewhere, somehow, when -bf is set.
More information about the ffmpeg-devel