Re: AVX512 code now committed, some more testing needed

Hi all,

same gcc-14.2.0 running on same Debian (linux 6.1), cpu i5 1035G1
running with power cord connected, lame frontend built with
--disable-dynamic-frontends

jsons attached, command line output is here:


topolinik@faye:~/svn/lame-svn-r6751-trunk-lame$ export
LD_LIBRARY_PATH=$PWD/libmp3lame/.libs
IN=~/svn/lame-svn-r6751-trunk-lame/audio_10m.wav L=$PWD/frontend/lame

   taskset -c 2 hyperfine --warmup 1 --runs 25 --style basic \
     --export-json fwd-V2.json \
     -n "avx2 -V 2" \
       "$L -S --vector avx2 -V 2 $IN /dev/null" \
     -n "avx512 -V 2" \
       "$L -S --vector avx512 -V 2 $IN /dev/null" \
     -n "avx512+table -V 2" \
       "env LAME_AVX512_CHOOSE_TABLE=1 $L -S --vector avx512 -V 2 $IN
   /dev/null" \ -n "avx512+both -V 2" \
       "env LAME_AVX512_CHOOSE_TABLE=1 LAME_AVX512_SFB_NOISE=1 $L -S
   --vector avx512 -V 2 $IN /dev/null"

   taskset -c 2 hyperfine --warmup 1 --runs 25 --style basic \
     --export-json rev-V2.json \
     -n "avx512+both -V 2" \
       "env LAME_AVX512_CHOOSE_TABLE=1 LAME_AVX512_SFB_NOISE=1 $L -S
   --vector avx512 -V 2 $IN /dev/null" \ -n "avx512+table -V 2" \
       "env LAME_AVX512_CHOOSE_TABLE=1 $L -S --vector avx512 -V 2 $IN
   /dev/null" \ -n "avx512 -V 2" \
       "$L -S --vector avx512 -V 2 $IN /dev/null" \
     -n "avx2 -V 2" \
       "$L -S --vector avx2 -V 2 $IN /dev/null"
Benchmark 1: avx2 -V 2
  Time (mean ± σ):      6.361 s ±  0.064 s    [User: 6.338 s, System:
   0.023 s] Range (min … max):    6.280 s …  6.495 s    25 runs
 
Benchmark 2: avx512 -V 2
  Time (mean ± σ):      6.366 s ±  0.070 s    [User: 6.343 s, System:
   0.022 s] Range (min … max):    6.249 s …  6.507 s    25 runs
 
Benchmark 3: avx512+table -V 2
  Time (mean ± σ):      6.455 s ±  0.152 s    [User: 6.430 s, System:
   0.024 s] Range (min … max):    6.396 s …  7.176 s    25 runs
 
  Warning: Statistical outliers were detected. Consider re-running this
   benchmark on a quiet system without any interferences from other
   programs. It might help to use the '--warmup' or '--prepare'
   options. Benchmark 4: avx512+both -V 2 Time (mean ± σ):      6.791 s
   ±  0.033 s    [User: 6.766 s, System: 0.024 s] Range (min … max):
   6.764 s …  6.935 s    25 runs 
Summary
  avx2 -V 2 ran
    1.00 ± 0.01 times faster than avx512 -V 2
    1.01 ± 0.03 times faster than avx512+table -V 2
    1.07 ± 0.01 times faster than avx512+both -V 2
Benchmark 1: avx512+both -V 2
  Time (mean ± σ):      6.775 s ±  0.014 s    [User: 6.750 s, System:
   0.024 s] Range (min … max):    6.737 s …  6.806 s    25 runs
 
Benchmark 2: avx512+table -V 2
  Time (mean ± σ):      6.420 s ±  0.012 s    [User: 6.397 s, System:
   0.023 s] Range (min … max):    6.402 s …  6.454 s    25 runs
 
Benchmark 3: avx512 -V 2
  Time (mean ± σ):      6.382 s ±  0.057 s    [User: 6.357 s, System:
   0.024 s] Range (min … max):    6.305 s …  6.490 s    25 runs
 
Benchmark 4: avx2 -V 2
  Time (mean ± σ):      6.373 s ±  0.066 s    [User: 6.348 s, System:
   0.025 s] Range (min … max):    6.270 s …  6.483 s    25 runs
 
Summary
  avx2 -V 2 ran
    1.00 ± 0.01 times faster than avx512 -V 2
    1.01 ± 0.01 times faster than avx512+table -V 2
    1.06 ± 0.01 times faster than avx512+both -V 2
topolinik@faye:~/svn/lame-svn-r6751-trunk-lame$ 

Elio



Il giorno Tue, 04 Aug 2026 19:28:54 +0200
Alexander Leidinger via Lame-dev <[email protected]> ha
scritto:

> Same file as last time, mains power, an otherwise quiet system and
> the two JSON exports are all I need back.


_______________________________________________
Lame-dev mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/lame-dev