Repository navigation
Use QuantizeB4 in SSE4.1 Dqip16bGemmV0 ConvertB - #1133
Merged
Merged
Conversation
…ip16bGemmV0_ConvertBb. Co-authored-by: igor.ermolaev <igor.ermolaev@altezza.org>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
SSE4.1
Dqip16bGemmV0_ConvertBfandDqip16bGemmV0_ConvertBbnow share the sameQuantizeB4helper already used by AVX2 (and AVX-512BW / NEON).The helper clamps four float vectors to int8, accumulates the negative sum, packs the lanes, and stores them with
K8_SFL_IDX. Aligned and tailK/Nloops call it instead of duplicating that sequence.Test plan
Base,SSE4.1,AVX2)./Test -r=.. -fi=SynetDynamicQuantizedInnerProduct16b -tt=1 -ts=1 -mt=100— all cases passed, including SSE4.1 vs the dispatched AVX2 path and vsSynetInnerProduct32fShapes
333x555-444cover both float (ConvertBf) and bfloat16 (ConvertBb) B, withKnot a multiple of 4. SSE4.1 runtime is about 2.1× AVX2 and about 10× faster than the scalar base, which matches a 4-wide vs 8-wide kernel.