Skip to content

[AMDGPU] Remove add/sub combines corrupting carry bit - #91

Closed
vosen wants to merge 2 commits into
ROCm:amd-stagingfrom
vosen:carry_fix
Closed

vosen wants to merge 2 commits into
ROCm:amd-stagingfrom
vosen:carry_fix

Conversation

@vosen

@vosen vosen commented May 20, 2024

Copy link
Copy Markdown

Isel would apply optimizations in the form of uaddo_carry (add x, y), 0, cc => uaddo_carry x, y, cc. This preserves the correct addition result, but not correct cc bit. For large x and y it could set cc = 0 on the lhs and cc = 1 on the rhs

vosen added 2 commits May 20, 2024 17:27
Isel would apply optimizations in the form of uaddo_carry (add x, y), 0, cc => uaddo_carry x, y, cc. This preserves the correct addition result, but not correct cc bit. For large x and y it could set cc = 0 on the lhs and cc = 1 on the rhs
@waarmond

waarmond commented Jun 5, 2024

Copy link
Copy Markdown

@searlmc1 Have you seen this?

@searlmc1
searlmc1 deleted the branch ROCm:amd-staging June 14, 2024 20:54
@searlmc1 searlmc1 closed this Jun 14, 2024
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants