Repository navigation
Software HMMA: optional fast rounding for the Deck - #19
Open
NxtDoorActivity wants to merge 1 commit into
Open
NxtDoorActivity wants to merge 1 commit into
NxtDoorActivity wants to merge 1 commit into
Conversation
The software matrix path rounds to half precision after every K=8 group with a long integer sequence (8x per depth-16 HMMA). The 14 ms neural-network compute shader on non-matrix GPUs (Steam Deck RDNA2) runs through it. With the env var set the intermediate rounding is skipped and the final pack rounds once. Off by default. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RmvPX2tqfoEsDjDfWTAmSf
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Adds EDEN_NX2_TENSOR_FAST_ROUND=1 (off by default). Skips the per-K=8-group half rounding in the software matrix path used for the game's neural-network compute pass (the ~14 ms shader) on GPUs without cooperative matrices. The final pack still rounds once.
Test on the Deck: run with and without the variable, compare FPS, look for visual differences.
Generated by Claude Code