On 11/03/2026 15:07, Andrew Randrianasulu wrote:
ср, 11 мар. 2026 г., 16:17 Terje J. Hanssen <[email protected]>:
On 11/03/2026 13:17, Andrew Randrianasulu wrote: > On Wed, Mar 11, 2026 at 2:14 PM Terje J. Hanssen via Cin > <[email protected]> wrote: >> >> >> On 10/03/2026 01:34, Andrew Randrianasulu wrote: >> >> On Mon, Mar 9, 2026 at 7:22 PM Terje J. Hanssen <[email protected]> wrote: >> >> On 09/03/2026 11:28, Andrew Randrianasulu wrote: >> >> You need to set remappig manually in (for single-user buld) >> >> bin/ffmpeg/decode.opts >> >> like >> >> remap_video_decoder mpeg2video=mpeg2_qsv >> >> and other similar lines if you want to test other formats (hevc, h264) >> >> Thank you for quick response. >> After editing: >> >> ~/Applications/squashfs-root/usr/bin/ffmpeg # cat decode.opts >> # apply at init decode >> loglevel=fatal >> formatprobesize=5000000 >> scan_all_pmts=1 >> remap_video_decoder libaom-av1=libdav1d >> remap_video_decoder mpeg2video=mpeg2_qsv >> >> Compilation of my test results as follows: >> >> mpeg2 transcoded to h264_qsv.mp4 >> >> decoding in cpu/gpu >> Cingg: 265/300 fps >> ffmpeg: 548/562 fps >> >> >> mpeg2 transcoded to hevc_qsv.mp4 >> >> decoding in cpu/gpu >> Cingg: 290/294 fps >> ffmpeg: - /688 fps >> >> >> mpeg2 transcoded to av1_qsv.mp4 >> >> decoding in cpu/gpu >> Cingg: 251/325 fps >> ffmpeg: 631/698 fps >> >> >> As seen the Cingg fps results are much slower than my system ffmpeg results. >> Is there any way to verify that decoding is done in the gpu like with ffmpeg? >> >> Comment out remap line and retest? Watch nvtop or "intel gpu top" not >> sure if they discriminate between enc and decoding on Intel (it works >> on amd polaris12)? >> >> As I said earlier we are not even supposed to be as fast as ffmpeg in >> ideal case because we roundtrip uncompressed video via system >> memory/bus. >> >> >> I tried intel_gpu_top which has a video line (for total gpu use in % I think). >> >> At least Cingg HW use device seemingly will output in the terminal if it is not supported as for: >> >> Settings|Preferences|Performance|Use HW Device: vaapi > Does it work better if you set Use hw device to none ?
Yes, as the previous results above, decoding cpu/gpu (before remap/after remap) But no speed benefit to set remap, results just within variations from time to time
You also can probably watch with top/conky/gkrellm if cpu usage goes down with qsv decoding case?
Does decoding heavier (hevc?) examples change the picture?
RPM Cinelerra Infinity - built: Mar 1 2026 06:30:41 A bit slower fps and lower video% with remap hevc-> av1_qsv ============== # remap_video_decoder mpeg2video=mpeg2_qsv Settings|Preferences|Performance|Use HW Device: none intel_gpu_top Video: up to 25% ** rendered 5972 frames in 22.474 secs, 265.729 fps audio0 pad 64 0 (64) ---------------- remap_video_decoder mpeg2video=mpeg2_qsv Settings|Preferences|Performance|Use HW Device: none intel_gpu_top Video: up to 23% ** rendered 5972 frames in 26.549 secs, 224.943 fps audio0 pad 64 0 (64)
A renewed test example:
av1_qsv ========
Settings|Preferences|Performance|Use HW Device: none
Decoding in GPU after remap remap_video_decoder mpeg2video=mpeg2_qsv
239-291 fps
-----------
Settings|Preferences|Performance|Use HW Device: vaapi
Decoder mpeg2_qsv does not support device type vaapi. HW device init failed, using SW decode. file:/Videoklipp/AV1/hdv09_04.m2t err: Operation not permitted ......
Render::render_single: Session finished.
** rendered 5972 frames in 20.875 secs, 286.084 fps audio0 pad 64 0 (64) malloc_consolidate(): unaligned fastbin chunk detected Aborted (core dumped) cin
===============
av1_qsv ========
Settings|Preferences|Performance|Use HW Device: none
Decoding in CPU before remap # remap_video_decoder mpeg2video=mpeg2_qsv
Render::render_single: Session finished.
** rendered 5972 frames in 20.001 secs, 298.585 fps audio0 pad 64 0 (64)
----------
Settings|Preferences|Performance|Use HW Device: vaapi
Decoding in cpu before remap # remap_video_decoder mpeg2video=mpeg2_qsv
int FFVideoConvert::convert_picture_vframe(VFrame*, AVFrame*, AVFrame*): Error retrieving data from GPU to CPU file: /Videoklipp/AV1/hdv09_04.m2t
Render::render_single: Session finished.
** rendered 5972 frames in 30.432 secs, 196.241 fps audio0 pad 64 0 (64)
Decoder libdav1d does not support device type vaapi. HW device init failed, using SW decode. file:/Videoklipp/QSV/hdv09-04_cin_av1_qsv_8b420.mp4 err: Operation not permitted Decoder libdav1d does not support device type vaapi. HW device init failed, using SW decode. file:/Videoklipp/QSV/hdv09-04_cin_av1_qsv_8b420.mp4 err: Operation not permitted
> >> Decoder mpeg2_qsv does not support device type vaapi. >> HW device init failed, using SW decode. >> file:/Videoklipp/AV1/hdv09_04.m2t >> err: Operation not permitted >> ...... >> Render::render_single: Session finished. >> ** rendered 5972 frames in 20.875 secs, 286.084 fps >> audio0 pad 64 0 (64) >> malloc_consolidate(): unaligned fastbin chunk detected >> Aborted (core dumped) cin >> >> ------- >> >> Yeah, of course Cingg and ffmpeg are not comparable with regards to video format transcoding speed.. >> >> I guess their working principles, surely oversimplified, may be something like this example: >> >> Cingg: mpeg2 input --> decoding -->internal/raw format --> decoding?--> encoding output >> >> FFmpeg: mpeg2 input --> decoding --> encoding output >> >> >> Therefore it's confusing where "remap_video_decoder mpeg2video=mpeg2_qsv" decoding really should take place, from mpeg2 input --> internal/raw format rather than decoding before encoding output? > somewhere inside libavcodec ... > > >> >> >> >> >> On Sun, Mar 8, 2026 at 11:46 PM Terje J. Hanssen via Cin >> <[email protected]> wrote: >> >> I have tested transcoding mpeg2video to h264_qsv, hevc_qsv, vp9_qsv and >> av1_qsv HW GPU accelerated decoding+encoding with ffmpeg. >> Just by adding "-hwaccel qsv" before the input file, the correct qsv >> codec for decoding in GPU apparently is auto-detected. >> >> ffmpeg -hide_banner -hwaccel qsv -i hdv09_04.m2t -c:v hevc_qsv -y >> hdv09_04_ffmpeg_hevc_qsv.mp4 >> >> With a medium powerful (balanced) Intel Alder Lake (i7-12700KF) /DG2 >> (A750) the qsv.mp4 video output and playback are clean and faster with >> GPU decoding. >> For the legacy, limited SkyLake (h264_qsv) and Coffee Lake (hevc_qsv) >> igpu's, it is best to keep decoding in CPU, due to igpu buffer errors >> and some distortions on the playback. >> >> >> So I wonder if it manageable to test-implement an additional environment >> variable "CIN_HW_DEV=qsv" for GPU HW decoding that adds "-hwaccel qsv" >> on the input side? >> If it works, it may be added as a "Use HW Device: qsv" next. >> >> Terje J. H >> >> >> >> >> >> >> _______________________________________________ >> Cin mailing list -- [email protected] >> To unsubscribe send an email to [email protected] >> >> >> >> _______________________________________________ >> Cin mailing list -- [email protected] >> To unsubscribe send an email to [email protected]