1849806 - Reduce number of exp2 calls for detune effect calculations

Status: RESOLVED FIXED

(Core :: Web Audio, enhancement)

Description

[Karl Tomlinson (:karlt)]

Bug 1358149 comment 55 indicates that transcendental function evaluation is having a significant effect on performance. exp2() is a commonly called such function. Skipping the calls in common unnecessary cases can improve performance.

Comment 1

[Karl Tomlinson (:karlt)]

When recomputing the AudioBufferSourceNode output rate only when necessary, the geometric mean results improve by ~1%, as many of the webaudio benchmarks use AudioBufferSourceNode. Granular synthesis speed improves by 5% or so.

Comment 2

[Karl Tomlinson (:karlt)]

OscillatorNode already caches its computed frequency so only "Periodic Wave with Automation" and "Substractive synth" benchmarks are repeatedly calculating the detune effect because the frequency is automated. Caching the detune calculation improves speeds by 11-24% and 7-13%. The linked results are the combined effect of skipping ABSN playback rate calculations and OscillatorNode detune calculations, but "Periodic Wave with Automation" and "Substractive synth" do not use ABSN.

Comment 3

[Karl Tomlinson (:karlt)]

I haven't cached the result of the detune calculation in AudioBufferSourceNode when the playbackRate is changing. AudioBufferSourceNode would recalculate detune a maximum of once per render block, and I expect resampler re-initialization time to swamp detune effect calculation time.

Comment 4

[Karl Tomlinson (:karlt)]

I considered switching from exp2 to the lower precision exp2f in OscillatorNode, but this function can amplify rounding error. Consider an input a x where x is close to 1 but contains rounding error. For f(x) = 2^(ax), f'(x)/f(x) = a ln 2. Rounding error is attenuated when |detune/1200| < 1 / ln 2. i.e. |detune| ⪅ 1700, so exp2f would be fine for typical moderate values. For a large detune of 10 octaves, however, error is amplified by 10 ln 2 ≅ 7. i.e. almost 3 of the 24 bits of precision would be lost.

Comment 5

[Karl Tomlinson (:karlt)]

Attachment: Bug 1849806 cache detune arithmetic in OscillatorNode r?padenot

Comment 6

[Karl Tomlinson (:karlt)]

Attachment: Bug 1849806 remove an unnecessary division in phase increment calculation r?padenot

Depends on D186686

Comment 7

[Karl Tomlinson (:karlt)]

Attachment: Bug 1849806 recalculate phase increment only when frequency changes r?padenot

Depends on D186687

Comment 8

[Karl Tomlinson (:karlt)]

Attachment: Bug 1849806 recompute buffer source output rate only when necessary r?padenot

to avoid the exp2f() call, which was showing up in benchmarks.

Depends on D186688

Comment 9

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/b5371614cd5a
cache detune arithmetic in OscillatorNode r=padenot

Comment 10

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/c668c58af537
remove an unnecessary division in phase increment calculation r=padenot

Comment 11

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/9a1dfe1dd89a
recalculate phase increment only when frequency changes r=padenot

Comment 12

[Cristian Tuns]

https://hg.mozilla.org/mozilla-central/rev/b5371614cd5a

Comment 13

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/27b0c1032942
recompute buffer source output rate only when necessary r=padenot

Comment 14

[Natalia Csoregi [:nataliaCs]]

https://hg.mozilla.org/mozilla-central/rev/c668c58af537
https://hg.mozilla.org/mozilla-central/rev/9a1dfe1dd89a
https://hg.mozilla.org/mozilla-central/rev/27b0c1032942

Comment 15

[Karl Tomlinson (:karlt)]

Something surprising in the OscillatorNode automation benchmark measurements was that they showed a noticeable difference from removing the unnecessary division. Numbers on 64-bit WINNT were

	Base		New	Delta	Confidence	Total Runs
Periodic Wave with Automation	182.14 ± 1.12	>	176.71 ± 0.78	-2.98%	5.84 (high)	7 / 7
Substractive synth	288.71 ± 0.56	>	283.71 ± 0.63	-1.73%	5.49 (high)	7 / 7

Numbers of other platforms were similar with some variability.

So I experimented with removing another division, but didn't get further improvements.