Repository navigation
Add NEON optimizations of SynetDynamicQuantizedInnerProduct16bGemmV0 - #1130
Merged
Merged
Conversation
…GemmV0. Port the SSE4.1 GemmV0 kernels to NEON with a 12-row microkernel, and dispatch them on ARM. 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Adds a NEON implementation of
SynetDynamicQuantizedInnerProduct16bGemmV0, following the SSE4.1 kernels, and dispatches it on ARM.Madd4<true>) with the M template extended to 12. The tail dequantizes, adds bias, and applies the activation, writing FP32 or BF16.SynetDynamicQuantizedInnerProduct16bGemmV0optimizations. The VS2022 Neon project includes the new sources. The 16b auto test runs the NEON path.Test
Cross-built with
aarch64-linux-gnu-g++and run under qemu-aarch64:./Test -r=.. -fi=SynetDynamicQuantizedInnerProduct16bForward -tt=1 -ts=1All tests finished successfully. On 333×555×444, NEON was about 1.5×–1.6× faster than Base under emulation (common time 84.9 ms vs 131.1 ms). An additional Base-vs-NEON sweep (partial N and K, M up to 340, K past the L1 tile, BF16 and int8 B, several activations) matched within the auto-test tolerance.