1945313 - Testcase convolving two sine waves N times is 1.7x slower (and may be getting slower as N increases)

Status: NEW

(Core :: JavaScript Engine: JIT, enhancement, P3)

Description

[Mayank Bansal]

Attachment: Audio convolve.HTML

Open attached testcase
Enter 500 or 1000
Click on "Start convolution)

500
Nightly: https://share.firefox.dev/40UV2KI (26s)
Chrome: https://share.firefox.dev/3EumpSR (17s)
1.5x slower

1000
Nightly: https://share.firefox.dev/3EnsmBm (59s)
Chrome: https://share.firefox.dev/4hkVKXe (34s)
1.7x slower

Testcase generated by chatpgpt on my prompt. I doubt this is a realistic testcase, but hopefully i can highlight some hotspots.
As usual, feel free to INVALID/WONTFIX.

ni? :padenot.

Comment 1

[Paul Adenot (:padenot)]

Yeah I'd file this under "general spidermonkey optimization bug", there's almost 0% in Web Audio (although this is DSP code, it is done in JavaScript). So we can rename this to be something like: particular math heavy code is slower in SpiderMonkey.

In the grand scheme of things, you wouldn't do a convolution like this anyways, you would do something like: https://en.wikipedia.org/wiki/Convolution#Fast_convolution_algorithms. This is what the ConvolverNode do under the hood (and it it multithreaded and all that).

Comment 2

[Iain Ireland [:iain]]

Similarly to bug 1944837, we're unlikely to spend much time examining this, but we can throw it in a bucket of testcases to measure future improvements against.