Only @noinline error path in matmul_size_check
#1310
Merged
Conversation
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
jishnub
added
performance
jishnub
force-pushed
the
jishnub/mul_size_check
branch
from
f6922cb to
651edc8
Compare
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## master #1310 +/- ##
==========================================
- Coverage 93.73% 93.72% -0.02%
==========================================
Files 34 34
Lines 15682 15685 +3
==========================================
+ Hits 14699 14700 +1
- Misses 983 985 +2 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
jishnub
added a commit
that referenced
this pull request
May 12, 2025
Instead of `@noinline` on the entire size-check function, we now separate the error-throwing part into a separate function and mark it as `@noinline`. This way, the size check may still be evaluated inline, and only the error path will not be inlined. This improves performance for small matmul. ```julia julia> A = [1 2; 3 4]; julia> @Btime $A * $A; 53.361 ns (2 allocations: 112 bytes) # v"1.13.0-DEV.438" 47.504 ns (2 allocations: 112 bytes) # this PR ``` (cherry picked from commit e4e8c19)
jishnub
added a commit
that referenced
this pull request
May 26, 2025
Backported PRs: - [x] #1209 <!-- Remove `LinearAlgebra` qualifications in `cholesky.jl` --> - [x] #1230 <!-- Avoid materializing `diag` in `Diagonal` `kron` --> - [x] #1240 <!-- Reduce `stable_muladdmul` branches in `generic matvecmul!` --> - [x] #1247 <!-- fix dispatch to herk --> - [x] #1255 <!-- use smaller matrix size in `peakflops` on 32-bit --> - [x] #1310 <!-- Only `@noinline` error path in `matmul_size_check` --> - [x] #1267 <!-- Refine column ranges in `_isbanded_impl` --> - [x] #1320 <!-- Copy matrices in `triu`/`tril` if no zero exists for the `eltype` --> - [x] #1324 <!-- Fix empty `Tridiagonal` broadcast --> - [x] #1327 <!-- `iszero` check in hessenberg setindex --> - [x] #1326 <!-- Fix multiplication with empty `HessenbergQ` --> - [x] #1332 <!-- Unwrap triangular matrices in broadcast --> - [x] #1337 <!-- Change `1:size` to `axes` in bidiag mul --> - [x] #1342 <!-- `Char` uplo in `Bidiagonal` constructor --> - [x] #1344 <!-- Update the docstring of ldiv! --> - [x] #1335 <!-- Test: prune old LA based on ENV variable --> - [x] #1346 <!-- Fix scaling unit triangular matrices --> - [x] #1355 <!-- Add compat notice for `diagview` --> - [x] #1349 <!-- Prune `LinearAlgebra` module in ambiguity test --> Contains multiple commits, manual intervention needed: - [x] #1238 <!-- Ensure positive-definite matrix in lapack posv test --> - [x] #1298 <!-- Add `diagm` example --> - [x] #1312 <!-- WIP: Try use method deletion instead of custom sysimage --> - [x] #1333 <!-- Make `fillstored!` public --> - [x] #1331 <!-- Document SingularException throw for inv(::AbstractMatrix) --> - [x] #1350 <!-- Fix copy for partly initialized unit triangular --> Non-merged PRs with backport label: - [x] #1352 <!-- log for dense diagonal matrix with negative elements --> - [ ] #1305 <!-- Bounds-checking in triangular indexing branches --> --------- Co-authored-by: Mateus Araújo <[email protected]> Co-authored-by: Jeff Bezanson <[email protected]> Co-authored-by: Steven G. Johnson <[email protected]> Co-authored-by: WalterMadelim <[email protected]> Co-authored-by: Kristoffer Carlsson <[email protected]> Co-authored-by: Daniel Karrasch <[email protected]> Co-authored-by: Michael Abbott <[email protected]>
ViralBShah
added a commit
that referenced
this pull request
Jun 2, 2025
This PR adds a size check in the 2-argument `mul`, so that now the destination array is allocated only if the sizes of the arguments are compatible with matrix multiplication. This means that we don't allocate in case of an error anymore. The performance for small-matrix multiplication seems largely similar (it's comparable to #1310, and seems identical within the noise limit): ```julia julia> A = [1 2; 3 4]; julia> @Btime $A * $A; 42.304 ns (2 allocations: 112 bytes) # before this PR 44.203 ns (2 allocations: 112 bytes) # this PR ``` We also redirect the generic `mul` to `_mul` now, which is the function that defines the multiplication code. This allows us to reuse the `_mul` definition elsewhere without having to repeat code. Currently, this is mainly necessary in the `Bidiagonal`-triangular multiplications.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
Instead of
@noinlineon the entire size-check function, we now separate the error-throwing part into a separate function and mark it as@noinline. This way, the size check may still be evaluated inline, and only the error path will not be inlined.This improves performance for small matmul.