Performance tracking

[paulfd]

So I've been working to automate the performance regression tests and it's going OK. I have a fairly simple process down now, I'll probably put it in another repository though.

I've ran it on a set of MIDI and SFZ files for the past months of merge commits.
The midi files are:

• Some uptempo swing drum track with alot of crashes (files called jump-)
• Chopin's "Butterfly" etude (files called butterfly-)
• Chopin's Fantaisie Impromptu (files called fantaisie-)
• Bach's Well Tempered Keyboard first prelude (files called wtc-) with specific CCs that change the amplitude of the stops used e.g. in Prelude

The SFZ I've considered are:

• AVL Drumkits (avl)
• GCSW Drumkits (gcsw)
• Naked Drums (naked)
• DSmolken's Big Rusty Drums (rusty)
• Ivy Piano in 162 (ivy)
• Little Lizard Piano (lizard)
• Prelude.lv2 organ (prelude) -- a wavetable-heavy file

The results are here, I'll try to keep them updated as much as possible:

https://box.ferrand.cc/sfizz-reports/history/butterfly-ivy-amd64.html
https://box.ferrand.cc/sfizz-reports/history/butterfly-lizard-amd64.html
https://box.ferrand.cc/sfizz-reports/history/fantaisie-ivy-amd64.html
https://box.ferrand.cc/sfizz-reports/history/fantaisie-lizard-amd64.html
https://box.ferrand.cc/sfizz-reports/history/jump-avl-amd64.html
https://box.ferrand.cc/sfizz-reports/history/jump-gcsw-amd64.html
https://box.ferrand.cc/sfizz-reports/history/jump-naked-amd64.html
https://box.ferrand.cc/sfizz-reports/history/jump-rusty-amd64.html
https://box.ferrand.cc/sfizz-reports/history/wtc-prelude-amd64.html

https://box.ferrand.cc/sfizz-reports/history/jump-avl-armv7.html
https://box.ferrand.cc/sfizz-reports/history/wtc-prelude-armv7.html
https://box.ferrand.cc/sfizz-reports/history/butterfly-lizard-armv7.html

• We can see that some heavy performance impact was due to the envelope follower, now reduced and further cancelled by Add runtime configs for loading in ram and voice stealing #353 where the envelope follower is disable unless requested by the stealing algorithm
• The default resampling had some impact, also reduced by Put default quality to 1 #405
• The modulation matrix has some hot code paths, mostly seen in modulation-heavy libraries such as Big Rusty Drums; I'm trying to track them down now

Hopefully we can use all this to make sfizz as performant as possible and avoid regressions over time 🙂

[paulfd]

Links update with a blocksize of 128 samples, to better reflect normal use as a plugin (before it was 1024). This makes it worse with respect to things like polyphony checks and the modulation matrix since these are costs independent of the block size.

[paulfd]

Added some armv7 runs with a raspi 3.