Repository navigation
Process two AVX2 columns in Dqip16bGemmV0 ConvertB - #1134
Merged
Merged
Conversation
…_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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
AVX2
Dqip16bGemmV0_ConvertBfandDqip16bGemmV0_ConvertBbnow walkjtwo SIMD widths at a time, matching the SSE4.1 kernels. Each iteration handles2 * Fcolumns (16), writing the second block atdst + aK * F, then falls through to the existing single-width and tail loops.Test plan
Base,SSE4.1,AVX2)./Test -r=.. -fi=SynetDynamicQuantizedInnerProduct16b -tt=1 -ts=1 -mt=100— all cases passed, including AVX2 vs scalar base, vs SSE4.1, and vsSynetInnerProduct32fShapes
333x555-444cover float (ConvertBf) and bfloat16 (ConvertBb) B, withKnot a multiple of 4 andNnot a multiple of 16, so the double-jloop, the single-Floop, and theNtail all run. AVX2 forward time is about 2.1× faster than SSE4.1 and about 21× faster than the scalar base. Against the previous single-jAVX2 kernel the full inner product stays within run-to-run noise, because the int8 GEMM dominates ConvertB.