[FFmpeg-user] Failed creating CUDA context for NVENC

Yugandhar Veeramachaneni hello at yugandhar.me
Thu Jul 19 17:03:40 EEST 2018 

> I patched the Nvidia driver by forcing FFmpeg link to patched nvidia
library.

May I know how you did this? Is there some documentation that I can follow?
Or is this the same as the one at https://github.com/keylase/nvidia-patch/ ?

On Thu, Jul 19, 2018 at 8:40 AM Yang Zhang <yzfedora at gmail.com> wrote:

> Also, I noticed a change on Nvidia Matrix website. I remember previous
> matrix says 1070ti only 1 NVENC chip. but now it says some 1060ti and
> 1070ti has 2 NVENC chips. and the strangest thing is: the old model of GTX
> 1070ti used much less GPU RAM than new GTX 1070ti. for example: old 1070ti
> used about 80 MB for full gpu based SD channels transcoding. but the new
> model used 130 MB.
>
> On Thu, Jul 19, 2018 at 9:36 PM, Yang Zhang <yzfedora at gmail.com> wrote:
>
> > Sure,
> >
> > I am using the latest 390.24 driver. I build the ffmpeg by myself. I
> > patched the Nvidia driver by forcing FFmpeg link to patched nvidia
> library.
> > so this can let you use any version driver also without limits. the
> command
> > is generate by software. for example:
> >
> > ffmpeg -hwaccel cuvid -hwaccel_device 0 -fflags +discardcorrupt -c:v
> > h264_cuvid -gpu 0 -probesize 10M -analyzeduration 10M -thread_queue_size
> > 128 -i /home/test/Videos/boku_no_hero_academia_16.mp4 -filter_complex
> > hwdownload,format=pix_fmts=nv12,split=outputs=1[hwupload:0];
> > [hwupload:0]hwupload=extra_hw_frames=1[map:v:0];
> > asplit=outputs=1[map:a:0] -map [map:v:0] -c:v h264_nvenc -flags:v
> > +global_header+cgop -preset:v hp -profile:v main -g 250 -gpu 0 -b:v:0
> 3000k
> > -maxrate:v:0 3500k -bufsize:v:0 7000k -map [map:a:0] -c:a libfdk_aac -ac
> 2
> > -ar 44100 -b:a:0 128k -f tee [f=hls:hls_time=2:hls_list_
> > size=5:hls_flags=delete_segments:select=\'v:0,a:0\':
> > bsfs/v=dump_extra=freq=keyframe]../cache/hls/13/highest.m3u8 -hide_banner
> > -loglevel info
> >
> > To transcoding 60 SD channels the bottleneck is at GPU Memory side. above
> > command is using full gpu based hardware acceleration. so the GPU RAM
> > overhead is about 80MB for 1 channel. actually I expect GTX 1070TI 8GB to
> > transcoding 80 SD channels.  but also, same problem with yours. GPU
> > actually used 6900 MB. but can not transcoding more.
> >
> > On Thu, Jul 19, 2018 at 9:29 PM, Yugandhar Veeramachaneni <
> > hello at yugandhar.me> wrote:
> >
> >> Hi Zhang,
> >>
> >> Are you using the driver published by NVIDIA or did you modify anything
> in
> >> it? Also, can you please share your full ffmpeg command string that you
> >> use
> >> for transcoding 60 SD channels for the sake of the community?
> >>
> >> Regards,
> >>
> >> Yugandhar
> >>
> >> On Thu, Jul 19, 2018 at 8:19 AM Yang Zhang <yzfedora at gmail.com> wrote:
> >>
> >> > I got this error before. the error code 0x2 is indicate no enough gpu
> >> > memory.  perhaps you can see the output of nvidia-smi tool indicate
> >> > the gpu is still left 1GB or more. perhaps nvidia-smi's output is not
> >> > exactly. or the nvidia driver limited it. my previous test is GTX
> >> > 1070ti 8GB can transcoding 60 SD channels.
> >> >
> >> > On 7/19/18, Yugandhar Veeramachaneni <hello at yugandhar.me> wrote:
> >> > > I see unrestricted in the # of concurrent sessions column -
> >> > >
> https://screenshots.firefox.com/ZTBoEcMVKO336dh8/developer.nvidia.com
> >> > which
> >> > > is why I bought this GPU in the first place.
> >> > >
> >> > > Your solution to override the hard-coded limits seems promising at
> the
> >> > > first glance. I'll try that soon and report back my findings.
> >> > >
> >> > > Thank you for your help.
> >> > >
> >> > > - Yugandhar
> >> > >
> >> > > On Wed, Jul 18, 2018 at 2:46 PM Dennis Mungai <dmngaie at gmail.com>
> >> wrote:
> >> > >
> >> > >> Also, see this list:
> >> > >>
> https://developer.nvidia.com/video-encode-decode-gpu-support-matrix
> >> > >>
> >> > >> You have the Quadro P4000, which is artificially crippled to ~2
> >> > concurrent
> >> > >> sessions according to the documentation above.
> >> > >>
> >> > >> On 18 July 2018 at 22:34, Dennis Mungai <dmngaie at gmail.com> wrote:
> >> > >>
> >> > >> > NVENC runs on a discrete silicon IP core (SIP) block in the GPU,
> >> and
> >> > not
> >> > >> > on the GPU's shaders.
> >> > >> > That SIP block has hard-coded limits (set in firmware) to ensure
> >> that
> >> > >> > these who need more than the artificial limit buy Tesla-grade
> >> > hardware.
> >> > >> >
> >> > >> > Marketing shills by NVIDIA.
> >> > >> >
> >> > >> > You can override this limit by using Keylase's nvidia patcher
> here:
> >> > >> > https://github.com/keylase/nvidia-patch
> >> > >> >
> >> > >> > On 18 July 2018 at 22:00, Yugandhar Veeramachaneni <
> >> > hello at yugandhar.me>
> >> > >> > wrote:
> >> > >> >
> >> > >> >> I'm a little confused here because the GPU resource utilization
> is
> >> > well
> >> > >> >> below the hard limits.
> >> > >> >>
> >> > >> >> When I run nvidia-smi, this is what I get -
> >> > >> >>
> >> > >> >> yugandharv at eclairs:~$ nvidia-smi
> >> > >> >> Thu Jul 19 00:29:45 2018
> >> > >> >> +-----------------------------------------------------------
> >> > >> >> ------------------+
> >> > >> >> | NVIDIA-SMI 390.67                 Driver Version:
> >> > >> >> 390.67                    |
> >> > >> >> |-------------------------------+----------------------+----
> >> > >> >> ------------------+
> >> > >> >> | GPU  Name        Persistence-M| Bus-Id        Disp.A |
> Volatile
> >> > >> Uncorr.
> >> > >> >> ECC |
> >> > >> >> | Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage |
> GPU-Util
> >> > >> Compute
> >> > >> >> M. |
> >> > >> >> |===============================+======================+====
> >> > >> >> ==================|
> >> > >> >> |   0  Quadro P4000        Off  | 00000000:02:00.0 Off |
> >> > >> >> N/A |
> >> > >> >> | 52%   55C    P0    38W / 105W |   6646MiB /  8117MiB |     25%
> >> > >> >> Default |
> >> > >> >> +-------------------------------+----------------------+----
> >> > >> >> ------------------+
> >> > >> >>
> >> > >> >>
> >> > >> >> +-----------------------------------------------------------
> >> > >> >> ------------------+
> >> > >> >> | Processes:
> >> >  GPU
> >> > >> >> Memory |
> >> > >> >> |  GPU       PID   Type   Process name
> >> > >> >> Usage      |
> >> > >> >> |===========================================================
> >> > >> >> ==================|
> >> > >> >> |    0       983      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 229MiB |
> >> > >> >> |    0       995      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 229MiB |
> >> > >> >> |    0      1010      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 229MiB |
> >> > >> >> |    0      1012      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 215MiB |
> >> > >> >> |    0      1017      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 235MiB |
> >> > >> >> |    0      1986      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 224MiB |
> >> > >> >> |    0      8484      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 229MiB |
> >> > >> >> |    0      8497      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 229MiB |
> >> > >> >> |    0      8538      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 225MiB |
> >> > >> >> |    0     11097      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 193MiB |
> >> > >> >> |    0     11109      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 193MiB |
> >> > >> >> |    0     11121      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 193MiB |
> >> > >> >> |    0     11134      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 229MiB |
> >> > >> >> |    0     11169      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 225MiB |
> >> > >> >> |    0     11175      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 193MiB |
> >> > >> >> |    0     11186      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 229MiB |
> >> > >> >> |    0     11195      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 229MiB |
> >> > >> >> |    0     11207      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 224MiB |
> >> > >> >> |    0     11222      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 229MiB |
> >> > >> >> |    0     11226      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 225MiB |
> >> > >> >> |    0     11231      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 229MiB |
> >> > >> >> |    0     11234      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 225MiB |
> >> > >> >> |    0     11235      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 229MiB |
> >> > >> >> |    0     11237      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 229MiB |
> >> > >> >> |    0     11240      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 229MiB |
> >> > >> >> |    0     11252      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 229MiB |
> >> > >> >> |    0     11300      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 229MiB |
> >> > >> >> |    0     13264      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 193MiB |
> >> > >> >> |    0     14176      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 193MiB |
> >> > >> >> |    0     30258      C   /home/yugandharv/bin/ffmpeg
> >> > >> >> 229MiB |
> >> > >> >> +-----------------------------------------------------------
> >> > >> >> ------------------+
> >> > >> >>
> >> > >> >>
> >> > >> >> Thanks,
> >> > >> >>
> >> > >> >> Yugandhar
> >> > >> >>
> >> > >> >> On Wed, Jul 18, 2018 at 11:55 AM Dennis Mungai <
> dmngaie at gmail.com
> >> >
> >> > >> wrote:
> >> > >> >>
> >> > >> >> > You mentioned 30 streams. And more cause this to "fail".
> >> > >> >> >
> >> > >> >> > Perhaps the Quadro line has a maximum simultaneous encoder
> >> limit of
> >> > >> 30,
> >> > >> >> > similar to the GeForce's limit of 2?
> >> > >> >> >
> >> > >> >> > On Wed, Jul 18, 2018, 19:28 Yugandhar Veeramachaneni <
> >> > >> >> hello at yugandhar.me>
> >> > >> >> > wrote:
> >> > >> >> >
> >> > >> >> > > Hello everyone,
> >> > >> >> > >
> >> > >> >> > > I'm currently using an NVIDIA Quadro P4000 GPU to transcode
> >> > >> multicast
> >> > >> >> UDP
> >> > >> >> > > streams to RTMP streams. I have quite about 30 streams
> >> running in
> >> > >> >> > parallel
> >> > >> >> > > on this GPU and I have no problems so far. I tried adding
> more
> >> > >> >> > > today
> >> > >> >> and
> >> > >> >> > > many of them are failing with this error.
> >> > >> >> > >
> >> > >> >> > > [h264_nvenc @ 0x55ef24458040] Loaded Nvenc version 8.1
> >> > >> >> > > [h264_nvenc @ 0x55ef24458040] Nvenc initialized successfully
> >> > >> >> > > [h264_nvenc @ 0x55ef24458040] 1 CUDA capable devices found
> >> > >> >> > > [h264_nvenc @ 0x55ef24458040] [ GPU #0 - < Quadro P4000 >
> has
> >> > >> Compute
> >> > >> >> SM
> >> > >> >> > > 6.1 ]
> >> > >> >> > > [h264_nvenc @ 0x55ef24458040] Failed creating CUDA context
> for
> >> > >> NVENC:
> >> > >> >> 0x2
> >> > >> >> > > [h264_nvenc @ 0x55ef24458040] No NVENC capable devices found
> >> > >> >> > > [h264_nvenc @ 0x55ef24458040] Nvenc unloaded
> >> > >> >> > >
> >> > >> >> > > Full log is pasted at
> >> > >> >> > >
> >> > >>
> https://gist.github.com/yugandhar91/47a1c30482d1e89a47f7b6fb6dd420ca
> >> > >> >> > >
> >> > >> >> > > Can you please point me to my mistake?
> >> > >> >> > >
> >> > >> >> > > Thanks,
> >> > >> >> > >
> >> > >> >> > > Yugandhar
> >> > >> >> > > _______________________________________________
> >> > >> >> > > ffmpeg-user mailing list
> >> > >> >> > > ffmpeg-user at ffmpeg.org
> >> > >> >> > > http://ffmpeg.org/mailman/listinfo/ffmpeg-user
> >> > >> >> > >
> >> > >> >> > > To unsubscribe, visit link above, or email
> >> > >> >> > > ffmpeg-user-request at ffmpeg.org with subject "unsubscribe".
> >> > >> >> > _______________________________________________
> >> > >> >> > ffmpeg-user mailing list
> >> > >> >> > ffmpeg-user at ffmpeg.org
> >> > >> >> > http://ffmpeg.org/mailman/listinfo/ffmpeg-user
> >> > >> >> >
> >> > >> >> > To unsubscribe, visit link above, or email
> >> > >> >> > ffmpeg-user-request at ffmpeg.org with subject "unsubscribe".
> >> > >> >> _______________________________________________
> >> > >> >> ffmpeg-user mailing list
> >> > >> >> ffmpeg-user at ffmpeg.org
> >> > >> >> http://ffmpeg.org/mailman/listinfo/ffmpeg-user
> >> > >> >>
> >> > >> >> To unsubscribe, visit link above, or email
> >> > >> >> ffmpeg-user-request at ffmpeg.org with subject "unsubscribe".
> >> > >> >>
> >> > >> >
> >> > >> >
> >> > >> _______________________________________________
> >> > >> ffmpeg-user mailing list
> >> > >> ffmpeg-user at ffmpeg.org
> >> > >> http://ffmpeg.org/mailman/listinfo/ffmpeg-user
> >> > >>
> >> > >> To unsubscribe, visit link above, or email
> >> > >> ffmpeg-user-request at ffmpeg.org with subject "unsubscribe".
> >> > > _______________________________________________
> >> > > ffmpeg-user mailing list
> >> > > ffmpeg-user at ffmpeg.org
> >> > > http://ffmpeg.org/mailman/listinfo/ffmpeg-user
> >> > >
> >> > > To unsubscribe, visit link above, or email
> >> > > ffmpeg-user-request at ffmpeg.org with subject "unsubscribe".
> >> > _______________________________________________
> >> > ffmpeg-user mailing list
> >> > ffmpeg-user at ffmpeg.org
> >> > http://ffmpeg.org/mailman/listinfo/ffmpeg-user
> >> >
> >> > To unsubscribe, visit link above, or email
> >> > ffmpeg-user-request at ffmpeg.org with subject "unsubscribe".
> >> _______________________________________________
> >> ffmpeg-user mailing list
> >> ffmpeg-user at ffmpeg.org
> >> http://ffmpeg.org/mailman/listinfo/ffmpeg-user
> >>
> >> To unsubscribe, visit link above, or email
> >> ffmpeg-user-request at ffmpeg.org with subject "unsubscribe".
> >>
> >
> >
> _______________________________________________
> ffmpeg-user mailing list
> ffmpeg-user at ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-user
>
> To unsubscribe, visit link above, or email
> ffmpeg-user-request at ffmpeg.org with subject "unsubscribe".

More information about the ffmpeg-user mailing list