Implement LinearAlgebra's storage-specific `mul!` methods by maleadt · Pull Request #3257 · JuliaGPU/CUDA.jl

maleadt · 2026-09-03T12:37:43Z

Julia 1.13-rc4 bumped LinearAlgebra to a version including JuliaLang/LinearAlgebra.jl#1671, which renamed the storage-level generic_matvecmul!/generic_matmatmul! entry points into 6- and 7-argument mul! methods. LinearAlgebra's strided BlasFloat mul! is more specific than the fallback that still routes to the old names, so our generic_matvecmul! overload was never reached on 1.13 and matrix-vector products ended up in CPU BLAS.

This implements the products as the new mul! methods and keeps the legacy names only as forwarding shims for Julia < 1.13 (one @static block per file, to be dropped when 1.13 becomes the floor). It requires GPUArrays 11.5.14 (JuliaGPU/GPUArrays.jl#775, now registered), whose generic_matmatmul_wrapper! hooks route through mul! on 1.13.

Julia 1.13-rc4 bumped LinearAlgebra to a version that includes JuliaLang/LinearAlgebra.jl#1671, which renamed the storage-level `generic_matvecmul!`/`generic_matmatmul!` entry points into 6- and 7-argument `mul!` methods, with the old names kept only as a fallback behind the generic `mul!` method. For matrix-vector products this breaks cuBLAS dispatch: LinearAlgebra's strided BlasFloat `mul!` method is more specific than that fallback, so it wins over our `generic_matvecmul!` overload and calls into CPU BLAS, failing with ArgumentError: Illegal conversion of a CUDACore.DeviceMemory to a Ptr{Float32} Implement the cuBLAS and cuSPARSE products as the new `mul!` methods, and keep the legacy names only as forwarding shims for Julia before 1.13 so that they can be dropped together once 1.13 becomes the floor. Require GPUArrays 11.5.14, whose `generic_matmatmul_wrapper!` hooks (used for element types cuBLAS does not cover) route through `mul!` rather than the legacy name, which on 1.13 would bypass these methods.

codecov · 2026-09-03T18:55:11Z

Codecov Report

❌ Patch coverage is 52.00000% with 12 lines in your changes missing coverage. Please review.
✅ Project coverage is 92.25%. Comparing base (413be53) to head (6e6c88c).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
lib/cusparse/src/interfaces.jl	29.41%	12 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3257      +/-   ##
==========================================
+ Coverage   92.18%   92.25%   +0.06%     
==========================================
  Files         184      184              
  Lines       16291    16289       -2     
==========================================
+ Hits        15018    15027       +9     
+ Misses       1273     1262      -11

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

github-actions · 2026-09-03T18:55:36Z

CUDA.jl Benchmarks

Details

Benchmark suite	Current: `6e6c88c`	Previous: `413be53`	Ratio
`array/accumulate/Float32/1d`	`98008` ns	`98229` ns	`1.00`
`array/accumulate/Float32/dims=1`	`71850` ns	`72403` ns	`0.99`
`array/accumulate/Float32/dims=1L`	`1599638` ns	`1599427` ns	`1.00`
`array/accumulate/Float32/dims=2`	`137251` ns	`138275` ns	`0.99`
`array/accumulate/Float32/dims=2L`	`660028` ns	`660231` ns	`1.00`
`array/accumulate/Int64/1d`	`117893` ns	`118416` ns	`1.00`
`array/accumulate/Int64/dims=1`	`75923` ns	`75920` ns	`1.00`
`array/accumulate/Int64/dims=1L`	`1715740` ns	`1715471` ns	`1.00`
`array/accumulate/Int64/dims=2`	`148381` ns	`148942` ns	`1.00`
`array/accumulate/Int64/dims=2L`	`986966` ns	`987603` ns	`1.00`
`array/broadcast`	`18240` ns	`17937` ns	`1.02`
`array/broadcast launch`	`9392` ns	`9013` ns	`1.04`
`array/construct`	`921.1071428571429` ns	`884.7551020408164` ns	`1.04`
`array/copy`	`16511` ns	`16497` ns	`1.00`
`array/copyto!/cpu_to_gpu`	`208227` ns	`208624` ns	`1.00`
`array/copyto!/gpu_to_cpu`	`241868` ns	`241688` ns	`1.00`
`array/copyto!/gpu_to_gpu`	`8914` ns	`8784` ns	`1.01`
`array/iteration/findall/bool`	`132939` ns	`133723` ns	`0.99`
`array/iteration/findall/int`	`148289` ns	`147552` ns	`1.00`
`array/iteration/findfirst/bool`	`69408` ns	`69312` ns	`1.00`
`array/iteration/findfirst/int`	`70256` ns	`70611` ns	`0.99`
`array/iteration/findmin/1d`	`70497` ns	`71647` ns	`0.98`
`array/iteration/findmin/2d`	`99943` ns	`99851` ns	`1.00`
`array/iteration/logical`	`190688` ns	`191816` ns	`0.99`
`array/iteration/scalar`	`63372` ns	`63027` ns	`1.01`
`array/permutedims/2d`	`49063` ns	`49063` ns	`1`
`array/permutedims/3d`	`50797` ns	`50390` ns	`1.01`
`array/permutedims/4d`	`50426` ns	`49946` ns	`1.01`
`array/random/rand/Float32`	`11384` ns	`10926` ns	`1.04`
`array/random/rand/Int64`	`23138` ns	`21140` ns	`1.09`
`array/random/rand!/Float32`	`7982.25` ns	`7777.75` ns	`1.03`
`array/random/rand!/Int64`	`20499` ns	`20510` ns	`1.00`
`array/random/randn/Float32`	`33667` ns	`33460` ns	`1.01`
`array/random/randn!/Float32`	`28844` ns	`23484` ns	`1.23`
`array/reductions/mapreduce/Float32/1d`	`33238` ns	`32875` ns	`1.01`
`array/reductions/mapreduce/Float32/dims=1`	`37500` ns	`37986` ns	`0.99`
`array/reductions/mapreduce/Float32/dims=1L`	`50590` ns	`50664` ns	`1.00`
`array/reductions/mapreduce/Float32/dims=2`	`54979` ns	`55130` ns	`1.00`
`array/reductions/mapreduce/Float32/dims=2L`	`67103` ns	`67684` ns	`0.99`
`array/reductions/mapreduce/Int64/1d`	`40463` ns	`40586` ns	`1.00`
`array/reductions/mapreduce/Int64/dims=1`	`40529` ns	`40894` ns	`0.99`
`array/reductions/mapreduce/Int64/dims=1L`	`88568` ns	`88686` ns	`1.00`
`array/reductions/mapreduce/Int64/dims=2`	`57559` ns	`57559` ns	`1`
`array/reductions/mapreduce/Int64/dims=2L`	`83732` ns	`83485` ns	`1.00`
`array/reductions/reduce/Float32/1d`	`32752` ns	`33402` ns	`0.98`
`array/reductions/reduce/Float32/dims=1`	`37599` ns	`37970` ns	`0.99`
`array/reductions/reduce/Float32/dims=1L`	`50413` ns	`50711` ns	`0.99`
`array/reductions/reduce/Float32/dims=2`	`54980` ns	`55262` ns	`0.99`
`array/reductions/reduce/Float32/dims=2L`	`68481` ns	`69402` ns	`0.99`
`array/reductions/reduce/Int64/1d`	`39979` ns	`40219` ns	`0.99`
`array/reductions/reduce/Int64/dims=1`	`40330` ns	`40569` ns	`0.99`
`array/reductions/reduce/Int64/dims=1L`	`88549` ns	`88630` ns	`1.00`
`array/reductions/reduce/Int64/dims=2`	`57115` ns	`57679` ns	`0.99`
`array/reductions/reduce/Int64/dims=2L`	`83091` ns	`83192` ns	`1.00`
`array/reverse/1d`	`16999` ns	`17277` ns	`0.98`
`array/reverse/1dL`	`69739` ns	`70117` ns	`0.99`
`array/reverse/1dL_inplace`	`67570` ns	`67420` ns	`1.00`
`array/reverse/1d_inplace`	`8630` ns	`8525.666666666666` ns	`1.01`
`array/reverse/2d`	`20172` ns	`20268` ns	`1.00`
`array/reverse/2dL`	`73600` ns	`73702` ns	`1.00`
`array/reverse/2dL_inplace`	`67299` ns	`67234` ns	`1.00`
`array/reverse/2d_inplace`	`9957` ns	`9987` ns	`1.00`
`array/sorting/1d`	`2647684` ns	`2658165` ns	`1.00`
`array/sorting/2d`	`1028673` ns	`1029672` ns	`1.00`
`array/sorting/by`	`3182607` ns	`3182246` ns	`1.00`
`cuda/synchronization/context/auto`	`1008.1` ns	`1024.7` ns	`0.98`
`cuda/synchronization/context/blocking`	`771.1509433962265` ns	`780.2424242424242` ns	`0.99`
`cuda/synchronization/context/nonblocking`	`5967.833333333333` ns	`5857` ns	`1.02`
`cuda/synchronization/stream/auto`	`863.7719298245614` ns	`874.9387755102041` ns	`0.99`
`cuda/synchronization/stream/blocking`	`665.1132075471698` ns	`674.7012987012987` ns	`0.99`
`cuda/synchronization/stream/nonblocking`	`5857.428571428572` ns	`5722.857142857143` ns	`1.02`
`integration/byval/reference`	`147877` ns	`147920` ns	`1.00`
`integration/byval/slices=1`	`149120` ns	`148937` ns	`1.00`
`integration/byval/slices=2`	`291726` ns	`292148` ns	`1.00`
`integration/byval/slices=3`	`434777` ns	`434674` ns	`1.00`
`integration/cudadevrt`	`105280` ns	`105237` ns	`1.00`
`integration/volumerhs`	`9139844` ns	`9139215` ns	`1.00`
`kernel/indexing`	`13009` ns	`13092` ns	`0.99`
`kernel/indexing_checked`	`13732` ns	`13675` ns	`1.00`
`kernel/launch`	`2204.777777777778` ns	`2261.3333333333335` ns	`0.97`
`kernel/occupancy`	`716.2785714285715` ns	`792.40625` ns	`0.90`
`kernel/rand`	`14940` ns	`17535` ns	`0.85`
`latency/import`	`4229849868` ns	`4118475955` ns	`1.03`
`latency/precompile`	`4986576090` ns	`4881109694` ns	`1.02`
`latency/ttfp`	`4671639575` ns	`4573842559` ns	`1.02`

This comment was automatically generated by workflow using github-action-benchmark.

maleadt force-pushed the tb/linalg-mul-1.13 branch 3 times, most recently from 1f0ac8e to 48cc607 Compare September 3, 2026 13:24

maleadt marked this pull request as ready for review September 3, 2026 14:33

maleadt force-pushed the tb/linalg-mul-1.13 branch from 48cc607 to 6e6c88c Compare September 3, 2026 14:47

maleadt merged commit 7ab5356 into main Sep 3, 2026
1 check was pending

maleadt deleted the tb/linalg-mul-1.13 branch September 3, 2026 18:16

vchuravy mentioned this pull request Sep 24, 2026

Attach the GPU matmul rules to generic_matmatmul! on Julia 1.13 too EnzymeAD/Enzyme.jl#3648

Merged

aalquwayfili mentioned this pull request Sep 30, 2026

Strided SubArray views of CuArray no longer reach cuBLAS in mul! since v11.5.11 JuliaGPU/GPUArrays.jl#799

Open

Navigation Menu

Repository navigation

Implement LinearAlgebra's storage-specific `mul!` methods - #3257

Implement LinearAlgebra's storage-specific `mul!` methods#3257
maleadt merged 1 commit into
mainfrom
tb/linalg-mul-1.13

maleadt commented Sep 3, 2026 •

edited

Loading

Uh oh!

Uh oh!

codecov Bot commented Sep 3, 2026

Uh oh!

github-actions Bot commented Sep 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

maleadt commented Sep 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

codecov Bot commented Sep 3, 2026

Codecov Report

Uh oh!

github-actions Bot commented Sep 3, 2026

CUDA.jl Benchmarks

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

maleadt commented Sep 3, 2026 •

edited

Loading