how do you "turn off denormals"

[be@floss.social]

@rygorous @aeva oh that's an interesting trick

[rygorous@mastodon.gamedev.place]

@be @aeva you can tweak the number as necessary (depending on what multiplicative factors you have in filters something like 1e-20 might be better) but basically a very slight DC offset to not center exactly on 0 but slightly off it is generally all you need

[be@floss.social]

@rygorous @aeva Did you come up with this trick? Or did you learn it from somewhere?

[rygorous@mastodon.gamedev.place]

@be @aeva come up with it, but I'm sure it's been independently invented a million times

[sonnybonds@mastodon.gamedev.place]

@rygorous @aeva We're in no way, shape or form the gold standard for how to do anything, but as a data point we use FTZ at Kilohearts and do it exactly like that; in a scope guard around our process calls.

[glyph@mastodon.social]

@aeva have you seen https://www.youtube.com/watch?v=y-NOz94ZEOA

[aeva@mastodon.gamedev.place]

@glyph nope. is it good

[glyph@mastodon.social]

@aeva yeah. it's where I learned about this problem. I'd experienced a few side effects of it before but never really understood the underlying cause. (sometimes my code runs *near* signal processing but I've never done it directly)

also this is a banger too, just to appreciate a totally different and much wackier way that denormals can really hurt you https://issues.chromium.org/issues/382005099#comment10

[aeva@mastodon.gamedev.place]

@glyph I realized it was probably the problem when I was digging through @TomF 's blog the other day looking for his blog post imploring people to not do the thing I'm doing in another project to find the tolerances on the thing and found a different blog on a completely different topic and it obliquely mentioned shutting off denormals to improve perf and i was like wait i remember hearing something about that somewheres

[aeva@mastodon.gamedev.place]

@glyph oh god i just started reading this. i hate the "we must make computers unusably slow at all costs. for security." people so so very much

[tomf@mastodon.gamedev.place]

@aeva @glyph You can also just check if it's between +FLT_MIN and -FLT_MIN and set it to zero yourself. Saves having to screw around with CPU modes.

[tomf@mastodon.gamedev.place]

@aeva @glyph
> looking for his blog post imploring people to not do the thing I'm doing in another project

Do I even want to know? Maybe not. <replaces rock carefully>

[aeva@mastodon.gamedev.place]

@TomF @glyph :3

[aeva@mastodon.gamedev.place]

@TomF @glyph it's a good blog post tom

[oblomov@sociale.network]

@TomF @aeva @glyph that doesn't save you from the cost of having to handle denormals in intermediate ops unless you split and check everything though?

[tomf@mastodon.gamedev.place]

@oblomov @aeva @glyph Sure, but usually normal audio processing doesn't produce denormals. The denormals arise because there's a feedback with a sub-1.0 gain, which eventually drains to very small numbers, and then denormals. If you flush the output, as soon as it gets there it just goes to zero immediately.

[aeva@mastodon.gamedev.place]

@TomF @oblomov @glyph i do like that switching on DAZ+FTZ makes this not a problem though. seems preferable to math hax or random branching?

[aeva@mastodon.gamedev.place]

@TomF @glyph specifically it was the one about using fixed point coordinates for objects. I'm still considering how to make that work, but I'd like to use jolt and jolt only speaks floats, and I am not expecting GLES2 GLSL to provide everything I need to do what I was doing before on the rendering side either. I'm probably going to use tile relative system instead on the shader side, and whatever will require as little translation on the CPU side.

[aeva@mastodon.gamedev.place]

@TomF @glyph or I NIH the physics in which case sure why not use fixed point

[giuseppebilotta@fediscience.org]

@aeva @TomF @glyph is this where I mention the int + float option, which is what we did for GPUSPH? ;-)

(We did it for performance reasons because int on GPUs was SLOW at the time, but has the upside of being “library friendly” once you're in relative coordinates, but that may be domain-specific)