performance problem

[wnpllrzodiac]

I use transform_v1 filter on linux redhat machine with 24 cpu cores.
But only got about 0.1x conversion speed,Not real time.
I found that cpu percentage is 100%, never exceed it. How to tune to get better performance?

[kcircnc]

The processing frame rate is proportional to resolution. What are the width/height?

[wnpllrzodiac]

resolution is 4k

[wnpllrzodiac]

input is 4k. output is 720p.I think threaddata is the key. Your filter cannot utilize all cpu core to do the job, or this is a limit because transform cannot do in multi thread?

[kcircnc]

Multithread is not supported yes. This project is not designed for supporting realtime applications.

That said, I just transformed a stereo 4K x 4K 60FPS equirect video to a 1080x1440 stereo cubemap with 17fps processing (0.277x) on my mac book. Perhaps something other than filtering is your bottleneck.

[wnpllrzodiac]

Input #0, flv, from 'vr.flv':
  Metadata:
    make            : Insta360
    model           : Insta360 Pro
    encoder         : Lavf57.66.102
  Duration: 00:10:00.04, start: 0.000000, bitrate: 6193 kb/s
    Stream #0:0: Audio: mp3, 48000 Hz, mono, s16p, 64 kb/s
    Stream #0:1: Video: flv1, yuv420p, 3840x2160, 200 kb/s, 29.97 fps, 29.97 tbr
, 1k tbn

[wnpllrzodiac]

sample clip:
https://iloveyaya.ga:9443/test/tmp/vr_sample.flv
for 30 sec

[wnpllrzodiac]

conversion command:
ffmpeg -i vr.flv -vf transform_v1="input_stereo_format=MONO:w_subdivisions=4:h_subdivisions=4:max_cube_edge_length=512:w=1280:h=720" -c:v libx264 -b:v 2000k -t 30 -c:a aac -ar 44100 -b:a 64k -f flv out.flv

[kcircnc]

With your file and command I got 20fps processing (0.651x) on my 8-core mac book.

[wnpllrzodiac]

I run it on win8.1 x64 platform.
Because vs2015 compile ERROR against macro SOFTWARE_PREFETCH_OPT
I comment out "#define SOFTWARE_PREFETCH_OPT" to build.

[wnpllrzodiac]

I modify config to "input_stereo_format=MONO:w_subdivisions=4:h_subdivisions=4:max_cube_edge_length=512"

Got 0.629x speed on linux test machine.