- Cin - Cinelerra-GG Mailing Lists

Tried ffv1_vulkan ....
by Andrew Randrianasulu 29 Apr '25

29 Apr '25

It looks like tonemapping ate encoder's lunch ;) RADV_PERFTEST=video_decode,video_encode time -p ./ffmpeg -init_hw_device vulkan=vulkan -filter_hw_device vulkan -hwaccel vulkan -i ~/K38_sdcard1/Documents/iPhone11_4K-recorder_59.940HDR10.mov -vf libplacebo=format=yuv444p:w=1920:h=1080:force_original_aspect_ratio=decrease:normalize_sar=true:upscaler=ewa_lanczos:downscaler=ewa_lanczos:colorspace=bt709:color_primaries=bt709:color_trc=bt709:range=tv -c:a copy -c:v ffv1_vulkan -slicecrc 0 -level 4 -strict -2 -async_depth 5 -slices_v 32 -slices_h 32 -f matroska -benchmark /dev/null frame= 1148 fps=4.6 q=-0.0 Lsize= 2302672KiB time=00:00:19.16 bitrate=984419.6kbits/s speed=0.0769x bench: utime=28.586s stime=20.721s rtime=249.150s bench: maxrss=587076KiB real 251.00 user 28.67 sys 20.79 ===== But may be faster GPU/pcie bus like in Andrea's machine will speed it up enough to be practical

3 5

xavc-intra class 480
by Andrew Randrianasulu 28 Apr '25

28 Apr '25

In attempt at speeding up hdr10 4k HEVC video playback I tried to re-encode it into AVC-intra using this guide: https://www.voukoder.org/forum/thread/1163-replicate-sony-s-xavc-i-s-format/ basically either ffmpeg -i input -c:v libx264 -avcintra-class 480 -pix_fmt yuv422p10le -preset ultrafast -x264opts avcintra-flavor=sony -c:a pcm_s24le -ar 48k output.mxf or for half size ffmpeg -i ~/K38_sdcard1/Documents/iPhone11_4K-recorder_59.940HDR10.mov -s 1920x1080 -c:v libx264 -avcintra-class 200 -pix_fmt yuv422p10le -preset ultrafast -x264opts avcintra-flavor=sony -c:a pcm_s24le -ar 48k output-half.mxf Note, line errors out when I tried class 480 for 1920x1080 frame size Both files load into cingg Original file (4k hdr hevc, 59.94 fps) plays at around 22 fps at output x11, project rgba-float. But seeking is sad. Not to mention lack of tonemapping. First 4k avc-intra 4k reencode plays at around 33 fps and seeks much better (no artefacts) 1920x1080 proxy plays at 60 fps, and if I check play every frame and try 2x speed forward - at 83 Fps. But trying two FullHD tracks referencing same "proxy" file set to different fade leves with project set to RGBA-float kills fps back to 2-3 :) Encoding at ultrafast preset was 4.5 fps for 4k version and 12 fps for FullHD version. Filesizes: du -h *.mxf 1.0G output-half.mxf 2.2G output.mxf this is just for frame= 1149 time=00:00:19.15 Well, I guess without some serious GPU upgrade this machine not gonna play HDR 4k :) Regular 8 bit/SDR 4k somewhat plays in firefox (if file just av1, 24 fps), at least on Linux where I can set GPU frequency a bit higher than default on nouveau/libre drivers.

3 6

New speed winner for hw tonemapping!
by Andrew Randrianasulu 28 Apr '25

28 Apr '25

RADV_PERFTEST=video_decode,video_encode time -p ./ffmpeg -init_hw_device vulkan=vulkan -filter_hw_device vulkan -hwaccel vulkan -i ~/K38_sdcard1/Documents/iPhone11_4K-recorder_59.940HDR10.mov -vf libplacebo=w=1920:h=1080:force_original_aspect_ratio=decrease:normalize_sar=true:upscaler=ewa_lanczos:downscaler=ewa_lanczos:colorspace=bt709:color_primaries=bt709:color_trc=bt709:range=tv -c:a copy -c:v libx264 -f mp4 -benchmark /dev/shm/ffmpeg-git-libplacebo-vulkan-2k.mp4 frame= 1148 fps= 10 q=-1.0 Lsize= 30435KiB time=00:00:19.13 bitrate=13029.3kbits/s speed=0.167x bench: utime=355.600s stime=25.878s rtime=114.342s so nearly 10 fps with scaled to FHD input! (otherwise 32bit x264 OOMs) ffmpeg git 8bb682d454990a1049a21f1f51442205ea3337e9 configured as ./configure --enable-opencl --disable-debug --enable-libx265 --enable-libx264 --enable-gpl --enable-libplacebo --enable-vulkan --enable-libshaderc --enable-libzimg --enable-libaom --enable-libdav1d --enable-libsoxr --enable-libfontconfig --enable-libfreetype --enable-libfribidi --enable-gnutls --enable-libass --enable-libbluray --enable-libcdio --enable-frei0r --enable-libgsm --enable-openal --enable-libopus --enable-librtmp --enable-libsnappy --enable-libspeex --enable-libssh --enable-libtheora --enable-libtwolame --enable-libv4l2 --enable-libvidstab --enable-libvorbis --enable-libvpx --enable-libwebp Cpu usage for just decoding also was lower than in vaapi case, so *some* use for Vulkan decode, contrary to my initial sceptecism!

3 7

I tried to make sws_scale go multithread ..
by Andrew Randrianasulu 28 Apr '25

28 Apr '25

BUT FAILED. in sense my mod apparently just leaks (swsContext?) so it can't play even 20 sec of 1440*1080 video to completion in termux, running on tablet with 3Gb ram. There seems to be difference if I set thread count manually (on file reading only for now) but it looks negative, in other words higher thread count -> slower total playback .. tried 8, 4,1,16 on 8 cpu tablet. Speed was going down from 18 fps for unaltered cingg to 13 with 8 threads and crash/oom with 16 ;) And if I try to use sws_scale_frame I just got all blank/black frames on timeline .... I think switching to sws_scale_frame is important if we want to try new colormanagement support in ffmpeg.git ... Resources used, apart from that Blender diff: https://github.com/mltframework/mlt/blob/master/src/modules/avformat/filter…

1 0

threaded swscale in blender
by Andrew Randrianasulu 27 Apr '25

27 Apr '25

Interesting diff, may be cingg can use something like this for speeding up swscale conversions? https://projects.blender.org/blender/blender/pulls/116309/commits/4647c80f5… But it seems that only upcoming ffmpeg version will allow to set transfer characteristics directly on avframe during scaling. So may be until this day our colorspace plugin should grow some knowledge about various gamma systems and also white points?

1 1

Can sws_scale() set trc?
by Andrew Randrianasulu 27 Apr '25

27 Apr '25

Or we need to go and set in->color_trc out->color_trc on avframes directly, like libavfilter/vf_scale.c does ?

1 0

libavutil/csp.h reading
by Andrew Randrianasulu 27 Apr '25

27 Apr '25

=== /** * Returns the ITU EOTF corresponding to a given TRC. This converts from the * signal level [0,1] to the raw output display luminance in nits (cd/m^2). * This is done per channel in RGB space, except for AVCOL_TRC_SMPTE428, which * assumes CIE XYZ in- and output. * * @return A pointer to the function implementing the given TRC, or NULL if no * such function is defined. * * @note In general, the resulting function is defined (wherever possible) for * out-of-range values, even though these values do not have a physical * meaning on the given display. Users should clamp inputs (or outputs) * if this behavior is not desired. * * This is also the case for functions like PQ, which are defined over an * absolute signal range independent of the target display capabilities. */ av_csp_eotf_function av_csp_itu_eotf(enum AVColorTransferCharacteristic trc); I guess this can be used for potential HDR waveform scope display?

1 0

Ah, no vaapi-vulkan interop on gfx8!
by Andrew Randrianasulu 26 Apr '25

26 Apr '25

Feature request: https://gitlab.freedesktop.org/mesa/mesa/-/issues/11074 https://gitlab.freedesktop.org/mesa/mesa/-/issues/5882 found via this post on reddit: https://www.reddit.com/r/jellyfin/comments/y7zns9/vulkan_filtering_on_amd_v… ==== However, there are still some limitations from upstream. First of all, the essential Vulkan extension *VK_EXT_image_drm_format_modifier* is currently *only available on Vega and newer cards/gfx9+*, which means that Polaris and older <https://gitlab.freedesktop.org/mesa/mesa/-/issues/5882> users will need to wait until upstream to complete the extension before they can get support for this. In addition, DRM changes are involved, so the Linux *kernel requires 5.15+*. ====

1 0

I wonder
by Haldun ALTAN 26 Apr '25

26 Apr '25

Hello, It's been a quite a while that I receive only the mails from Sam via Cin. Somethings wrong or this is the new normal ? :) Greetings and thanks for your work you all :) Haldun.

2 1

cingg faster than ffmpeg itself on pulling frames!?
by Andrew Randrianasulu 26 Apr '25

26 Apr '25

I tried to bench simple ffmpeg command (ffmpeg 7.1.1): time -p ./ffmpeg -hwaccel vaapi -i ~/K38_sdcard1/Documents/iPhone11_4K-recorder_59.940HDR10.mov -c:a copy -c:v rawvideo -f nut /dev/null this benches for 4k 60 fps HDR video at frame= 1148 fps= 10 q=-0.0 Lsize=27900063KiB time=00:00:19.16 bitrate=11927607.4kbits/s speed=0.173x real 111.77 user 42.61 sys 119.01 10 fps,150% cpu load. ffmpeg 4.4 was using just 85% cpu but only managed frame= 1148 fps=8.5 q=-0.0 Lsize=27900063kB time=00:00:19.15 bitrate=11933560.4kbits/s speed=0.142x 8.5 fps. But I am fairly sure I saw *15* fps in cingg on same file , if you set "play every frame" and vo driver to x11. Anyone want to repeat this experiment? Is it artefact of running 32bit program on top of 64bit kernel? At some point I might try latest Fedora 42 to see if full 64bit Linux will pull more frames out of same gpu decoder on same motherboard

2 4