Multithread the memory-bound level-1/2 operations (OpenMP)#943
Draft
jeffhammond wants to merge 1 commit into
Draft
Multithread the memory-bound level-1/2 operations (OpenMP)#943jeffhammond wants to merge 1 commit into
jeffhammond wants to merge 1 commit into
Conversation
BLIS runs level-1 and level-2 single-threaded, so large vector/matrix-vector
ops are capped at one core's memory bandwidth. Add optional OpenMP threading
(opt-in via BLIS_ENABLE_L1_OPENMP) for the memory-bound ops:
- level-1v: axpyv, dotv, scal2v, scalv/invscalv/setv, copyv/addv/subv
(disjoint chunks; dotv uses a partial-sum-then-combine reduction);
- util: asumv, and nrm2 (normfv) threaded while preserving its
overflow-safe scaling via a LAPACK-style pairwise (scale,sumsq) combine;
- level-2: gemv (both transposes -- row-parallel axpyf / output-parallel
dotxf) and ger (both variants -- column/row parallel).
Only fires for large operands, when not already nested in a parallel region.
Reduction combines are correct for real and complex types. Default builds
(macro undefined) keep the original single-threaded behavior, so no
regression. On a 10-core cluster this is a ~2-9x speedup for these ops
(e.g. axpy 4->11, gemv 8->22 GFLOP/s, asum 15->90, nrm2 7->69 GB/s).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
jeffhammond
marked this pull request as draft
July 16, 2026 15:25
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
BLIS runs level-1 and level-2 single-threaded, so large vector/matrix-vector ops are capped at one core's memory bandwidth. Add optional OpenMP threading (opt-in via BLIS_ENABLE_L1_OPENMP) for the memory-bound ops:
Only fires for large operands, when not already nested in a parallel region. Reduction combines are correct for real and complex types. Default builds (macro undefined) keep the original single-threaded behavior, so no regression. On a 10-core cluster this is a ~2-9x speedup for these ops (e.g. axpy 4->11, gemv 8->22 GFLOP/s, asum 15->90, nrm2 7->69 GB/s).