Repository navigation
Fix the ordered reduction's vector loads after the SIMD.jl removal - #175
Merged
Merged
Conversation
#165 and #168 were merged on the same day. #165 added a vector load to the ordered reduction kernel with SIMD.jl's `vloada`, which #168 had just replaced with an `llvmcall`-based `_vloada` and stopped importing. The kernel thus failed to compile for dense sources with `commutative=false` on every GPU back-end, which breaks 0.5.8 and the existing vector-load tests in test/generic/reduce.jl.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Fix the ordered reduction's vector loads after the SIMD.jl removal
#165 and #168 were merged minutes apart. #165 added a vector load to
_mapreduce_ordered!using SIMD.jl'svloada(Vec{W,T}, ptr), but #168 had just replaced that with thellvmcall-based_vloada(Val(W), ptr)and dropped theusing SIMD. Each PR passed CI on its own; the combination doesn't compile. With 0.5.8, every GPU back-end therefore failsreduce/mapreducewithcommutative=falseon dense sources whose vector width dividesitems_per_thread:This is why main is red (OpenCL, KA host kernels, Buildkite), and it also breaks CI on #137, #140 and JuliaGPU/GPUArrays.jl#822.
The existing vector-load tests in
test/generic/reduce.jl(line 1596 onwards) cover this, so no new test is needed. Locally on Julia 1.12,cuda/reduce+opencl/reduce(PoCL) hit 14 errors on main and pass all 36900 tests with this change.The second commit bumps the version to 0.5.9, so the fix can be released right away.