Threadwell

Recent Notes

JohannesVuorine
Aug 26, 2026
Threads
Aug 26, 2026
- archive
iquilezles
Aug 26, 2026

iquilezles

36 items with this tag.

Aug 26, 2026
iquilezles
Feb 09, 2025
Hey, don't "optimize" conditional moves with mix()+step().
Jul 30, 2025
@iquilezles I suspect the reason for the poor advice stems from the beginning of programmable shaders on mobile.
Jun 01, 2025
@iquilezles @james_ay_em_dee Read about this from ‘GPU Gems 2’, ‘Chapter 34.
May 08, 2025
@iquilezles As a shader newb, this is helpful.
Feb 12, 2025
@iquilezles Is there a use case for step()? I always forget if it's a < or > comparison.
Feb 10, 2025
@iquilezles I don't know much about shader optimization, but I do wonder how many more of these witty optimization gimmicks are just consumed like tricks of the trade, but nobody ever tests their assumptions, they just replicate as coding memes.
Feb 10, 2025
@iquilezles This comes from common misunderstanding of divergence on GPU ( magic thinking of hiding all supposedly evil IF by something that may be costlier )
Feb 10, 2025
@iquilezles x86 has had conditional move instructions since Pentium Pro (1995).
Feb 10, 2025
@iquilezles In addition, mix() can cause NaN because GPU can calculate 0*NaN=NaN.
Feb 10, 2025
@iquilezles compilers are smart enough to take care of conditional logic without branching(or conditional jump) in many cases, so don't pretend to be smart and just use if else.
Feb 10, 2025
@iquilezles okay..
Feb 09, 2025
@iquilezles Conditional moves are also a CPU thing, and compilers knows about them !
Feb 09, 2025
@iquilezles @iquilezles :
Feb 09, 2025
@iquilezles I feel half of the problem is that people don't quite understand when branching actually occurs, and then get convinced by the whole "ternary/if-else is bad because branching".
Feb 09, 2025
@iquilezles tl dr, compilers will generally fix this kind of issue for you
Feb 09, 2025
@iquilezles I think we need deeper write-ups on this subject as well.
Feb 09, 2025
@iquilezles You didn’t show the compiler assembly for the mix alternative.
Feb 09, 2025
@iquilezles Is this also good advice when programming shaders for mobile hardware?
Feb 09, 2025
@iquilezles This greatly depends.
Feb 09, 2025
@iquilezles But if I use that doesn’t that result in a time when the compiler uses a separate instance to do the effects that would have been done in a different way otherwise? Hypothetically, if you consider the end effects
Feb 09, 2025
@iquilezles At what point does conditional branching occur though? Larger blocks of code?
Feb 09, 2025
@iquilezles I'd say ALWAYS look at the generated assembly when you care about optimizations.
Feb 09, 2025
@iquilezles @SebAaltonen Thanks! But if I understand you correctly, there are other cases where if-statements should be avoided.
Feb 09, 2025
@iquilezles Why would someone do the BAD one?
Feb 09, 2025
@iquilezles What about conditions from sampled textures? If every pixel in a warp takes a different branch, ?: causes extreme divergence.
Feb 09, 2025
@iquilezles Is this true for ARM GPUs as well? I don’t have arm shader disassembler on hand.
Feb 09, 2025
@iquilezles The fear of branching is still so strong.
Feb 09, 2025
@iquilezles What about the classic:
Feb 09, 2025
@iquilezles Thanks for that.
Feb 09, 2025
@iquilezles also valid for any decent modern CPU.
Feb 09, 2025
@iquilezles Maybe the article should further mention that the true evil of if statements in shaders only comes when the neighbouring shader invocations take different paths, as both instances must execute both paths and then throw one away
Feb 09, 2025
@iquilezles Good advice! In HLSL you can also use the select() function to do the ternary which makes it explicit that you want a ternary, not sure if GLSL has a similar thing
Feb 09, 2025
@iquilezles it’s still faster even if the condition is not true for all threads in a group? :o
Feb 09, 2025
@iquilezles This is so infuriating.
Feb 09, 2025
@iquilezles Thanks for this! I learned using step was the right way but it makes code ugly and harder to read.

Created with Quartz v5.0.0 © 2026

GitHub