Bugzilla – Bug 2457
Windows: WAV Export runs 6 to 7 times slower than on 2.3.3
Last modified: 2020-06-11 10:27:16 UTC
Created attachment 983 [details] WAV and MP3 Export speed tests On Windows WAV Export runs 6 to 7 times slower than on 2.3.3 Thus a regression on 2.3.3 But there was lao a regression 2.2.2 to 2.3.9 where the WAV export time doubled. So now with 2.4.x we Export WAV files 12 to 14 times slower than our track record. I have attached a spreadsheet to show my timing tests. Note that over time in contrast we have significantly improved MP3 exports on Windows. But also note that on my MacBook Catalina WAV exports are blindingly fast compared to my W10 laptop - twice as fast as out best ever on Windows/ This issue was first reported by Forum user fortchagos: https://forum.audacityteam.org/viewtopic.php?f=50&t=111219&p=396135
My results with 1 hour stereo track: Linux (budget i7 laptop with SSD): Shaped dither: 16 seconds No dither: 4 seconds macOS (i7 Mac Mini with SSD): Shaped dither: 7 seconds No dither: 3.5 seconds macOS is extraordinarily fast applying shaped dither. Must be some good optimizations happening.
(In reply to Steve Daulton from comment #1) OK my results - one hour stereo chirp W10: Shaped dither: 198 seconds No dither: 88 seconds macOS Catalina 10.15.4: Shaped dither: 7 seconds No dither: 3 seconds so it looks like the dither takes up roughly half the time om both platforms - but the Mac is blindingly fast compared to the W10-PC and both are expensive SSD models. But dither is recommended - and is our default setting.
Wow...I thought there just HAD to be something wrong with Peter's machine. :-) But, I can confirm: Windows (i9 @ 5GHz with SSD #1), shaped dither: 2.4.2: 120 seconds 2.2.2: 12 seconds Export and temp files location set to SSD #1 (Windows boot drive) Windows (i9 @ 5GHz with SSD #2), shaped dither: 2.4.2: 118 seconds 2.2.2: 12 seconds Export and temp files location set to SSD #2 Windows (i9 @ 5GHz with SSD #1 and #2), shaped dither: 2.4.2: 120 seconds 2.2.2: 12 seconds Temp files location set to SSD #1 and export to SSD #2 Now, here's the fun stuff... :-) Linux (VM running under above Windows machine on SSD #2), shaped dither: 2.4.2: 13 seconds
Yep, just as I suspected. The old build process gets the timings back to normal. I'll compare all options between the cmake and old build processes.
Fixed in: https://github.com/audacity/audacity/commit/cd38e0e https://github.com/audacity/audacity/actions/runs/118494952
(In reply to Leland Lucius from comment #5) Testing on W10 with Audacity 2.4.2 ef98e5d I see no improvement at all with this latest W10 alpha - exactly the same time as yesterdays' test 5 mins 15 seconds = 195 seconds on exporting a one hour stereo chirp. Accordingly I have to re-open this
Round #2... https://github.com/audacity/audacity/commit/5bede69 https://github.com/audacity/audacity/actions/runs/119044507
(In reply to Leland Lucius from comment #7) Testing on W10 with Audacity 2.4.2 5bede69 We have now got it back to 2,3.3 performance - exporting a one-hour stereo WAV in31 seconds (versus 230 in 2.3.3) close enough. But we're still twice as slow as we were in 2.2 and earlier where the same export takes 15 seconds (but unlike the recent performance change 2.3=>2.4 no-one noticed or complained about the 2.2=>2.3 performance hit). And its still much faster on Mac Catalina a mere 7 seconds. Curiously both the Chirp generation and the MP3 export both took a little longer with this build a) Chirp 29 seconds versus 22 seconds b)MP3 export 57 seconds versus 54 seconds So my question is are we happy to close this bug now that we can at least match 2.3.x performance - or do we want to go further and aim for 2.2.x performance ?
Yes. Close it. The P1 bug is gone. It does not hold up 2.4.2. Performance generally should be dealt with by test scripts that time.
It looks like this may have also affected AC3 exports too - but also now cleared up. See this Forum thread: https://forum.audacityteam.org/viewtopic.php?f=46&t=111496&p=397389
