[Mosaic GPU] Optimize the computation of tcgen05.mma matrix descriptors #33653
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Previously we used a simple approach of computing the descriptors entirely using LLVM ops. This was convenient, but it turns out that there are two problems with it: 1. LLVM doesn't always fully constant fold properly and sometimes emits PTX that causes ptxas to generate lots of non-uniform operations. 2. LLVM is quite aggressive to hoist descriptor computation outside of loops, which blows up the register pressure. The alternative implemented here is to compute the descriptors in inline ptx, with manual constant folding, and right before the MMA operations. This seems to generate code that has extremely low register pressure and only very few uniform operations on 32-bit quantities. PiperOrigin-RevId: 839224185
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
[Mosaic GPU] Optimize the computation of tcgen05.mma matrix descriptors
Previously we used a simple approach of computing the descriptors entirely
using LLVM ops. This was convenient, but it turns out that there are two
problems with it:
PTX that causes ptxas to generate lots of non-uniform operations.
which blows up the register pressure.
The alternative implemented here is to compute the descriptors in inline ptx,
with manual constant folding, and right before the MMA operations. This seems
to generate code that has extremely low register pressure and only very few
uniform operations on 32-bit quantities.