Skip to content
This repository has been archived by the owner on Sep 23, 2024. It is now read-only.

Replaced use of rcp() and rsqrt() functions with non-approximating re… #20

Open
wants to merge 1 commit into
base: master
Choose a base branch
from

Conversation

John-Whigham
Copy link

Replaced use of rcp() and rsqrt() functions with non-approximating reciprocal and reciprocal-sqrt as the approximating versions produce subtly different results on different CPU architectures causing non-deterministic output (Intel Xeon vs. AMD Ryzen was observed to differ)

Differing output can be a big problem for patching tools for example if building a game on one machine produces different binary data than on another resulting in patches that are considerably larger than they need to be.

I've left comments in the code so future developers don't see replacing them with approximating versions as a potential optimization.

…ciprocal and reciprocal-sqrt as the approximating versions produce subtly different results on different CPU architectures causing non-deterministic output (Intel Xeon vs. AMD Ryzen was observed to differ)
@ivanassen
Copy link

+1 for this - producing bit-identical output regardless of which machine performed the particular compression is important in asset build pipelines.

Note that fastmath should also be off for this to work.

Maybe the bit-exact behavior needs to be opt-in - I can imagine environments where it's not necessary, and the performance cost is not acceptable.

aras-p added a commit to aras-p/smol-compute that referenced this pull request Dec 15, 2020
@telecran-telecrit
Copy link

@jspohr
Copy link

jspohr commented Dec 27, 2023

I think this PR could be closed since this change on master applies the same fix.

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

4 participants