[dependabot]

Updates the requirements on GPUArrays and KernelAbstractions to permit the latest version.
Updates GPUArrays to 12.1.1

Release notes

Sourced from GPUArrays's releases.

v12.1.1

GPUArrays v12.1.1

Diff since v12.1.0

Merged pull requests:

• Show the array type in 2-arg show of GPU arrays (#429) (@​mcabbott)
• Upload index arrays when viewing a wrapped GPU array (#777) (@​aalquwayfili)
• Give coo_type a base case so a missing method does not recurse (#781) (@​aalquwayfili)
• Forward strided-view matrix products to the storage-level mul! (#800) (@​aalquwayfili)
• [NFC] Reorganize some tests to make failures easier to parse (#805) (@​christiangnrd)
• Let reductions infer their result type (#809) (@​vchuravy)
• Test for closure boxes on Julia 1.14+ (#812) (@​KristofferC)
• Faster findfirst/findlast and reverse (#815) (@​maleadt)

Closed issues:

• Broadcasting on SubArray with inds of Vector{Int} fails with compiler issue (#724)
• coo_type recurses until the stack overflows instead of throwing MethodError (#780)
• Strided SubArray views of CuArray no longer reach cuBLAS in mul! since v11.5.11 (#799)

Changelog

Sourced from GPUArrays's changelog.

GPUArrays.jl release history

This document lists the noteworthy changes in GPUArrays.jl releases that back-ends act on, newest first, aimed at the authors of back-end packages. For the complete list of merged pull requests, see the GitHub releases.

v12.1

GPUArrays defines Base.dataids and Base.mightalias for every AbstractGPUArray. A back-end that implements the new GPUArrays.memory_location hook (the coordinate of an array's allocation, normally its address, and its byte offset into it) gets precise alias detection: disjoint views of one buffer don't alias, empty arrays alias nothing on every Julia version, and SubArrays of two such arrays are compared without converting their parents to pointers. Arrays wrapped in a SubArray or other wrapper are still compared by allocation, which is conservative. Back-ends should implement the hook and delete their own dataids and mightalias methods, which would keep winning dispatch.

v12.0

Base's reductions, sorting, scans, findall, logical indexing and reverse are now implemented once for every AnyGPUArray, on top of AcceleratedKernels.jl (AK) 0.5. A back-end gets all of them by defining nothing beyond the array interface and a KernelAbstractions back-end, and they follow Base's semantics: stable sorting by default, init applied once, no neutral element needed, Base's result types, empty results and errors.

Breaking changes for back-ends:

• The GPUArrays.mapreducedim! hook is removed and GPUArrays no longer picks neutral elements. Every Base reduction goes to AK directly. Back-ends must delete their methods for the hook, together with the reduction kernels only it used.
• GPUArrays.default_rng and the deprecated RNG(state::AbstractGPUArray) and seed!(rng, ::Vector{UInt32}) methods are removed. Construct the RNG with RNG{AT}() and seed it with an integer.
• KernelAbstractions 0.9.43 or later and Adapt 4.7.3 or later are required.

What back-ends should delete:

Methods narrower than GPUArrays' keep winning dispatch, so a back-end that keeps its own findall, to_index/to_indices, _accumulate!/accumulate, sort!/sortperm!/ partialsort!, or reverse! bypasses GPUArrays' Base conformance, and some of those methods are ambiguous with GPUArrays' (e.g. findall(f, ::MyArray) against findall(::Fix2{typeof(in)}, ::AnyGPUArray)). Delete them.

A vendor sort or scan (e.g. one built on MPSGraph) can stay as a fast path, as long as it falls back to GPUArrays' method for the cases it doesn't handle and accepts AK's algorithm objects (alg=AK.MergeSort() etc.). There is no extension point for vendor reductions; GPUArrays._ak_mapreducedim! is internal.

... (truncated)

Commits

• 6fa6490 Bump version.
• ea47064 Give coo_type a base case so a missing method does not recurse (#781)
• 7bffdc6 [NFC] Reorganize some tests to make failures easier to parse (#805)
• 9447b8b Show the array type in 2-arg show of GPU arrays (#429)
• 8d52285 Upload index arrays when viewing a wrapped GPU array (#777)
• 85ac86f Forward strided-view matrix products to the storage-level mul! (#800)
• 4f06c10 Track GPUArrays coverage in the downstream jobs
• 0e33752 Only run the SPIR-V half of OpenCL.jl's tests downstream
• 1fc62f0 Give the OpenCL.jl job two hours
• e0d5567 Test for closure boxes on Julia 1.14+ (#812)
• Additional commits viewable in compare view

Updates KernelAbstractions to 0.10.0

Release notes

Sourced from KernelAbstractions's releases.

v0.10.0

KernelAbstractions v0.10.0

Diff since v0.9.45

Breaking changes

See the 0.10 section of the changelog: JuliaGPU/KernelAbstractions.jl#0 In short: the CPU backend is now an OpenCL backend running on PoCL, CPU(; static=true), isgpu, @ka_code_llvm and @ka_code_typed were removed, GPU is deprecated, @private arrays are a PrivateArray, @Const arguments are constified at launch, and backends implement KernelInterface 0.4.

Merged pull requests:

• Add @groupreduce and @groupscan (#830) (@​vchuravy)
• KernelInterface: sub-group shuffles, votes, reductions and scans (#831) (@​vchuravy)
• Complete the 0.10 release notes and update the NUMA example (#842) (@​vchuravy)
• POCL: recommend work-group sizes with PoCL's own heuristic (#861) (@​vchuravy)
• Compute indices in Int on PoCL, and mark the index arithmetic nuw/nsw (#879) (@​vchuravy)
• Finalize 0.10: release notes, precompile warnings, compat and CI (#885) (@​maleadt)

Closed issues:

• CPU backend: the bounds mask makes kernel bodies up to 2.3x slower, even when the workgroups tile the ndrange (#845)
• CPU backend: the default workgroup size is the device maximum (4096), so launches of up to 4096 work-items run on one thread (#859)
• CPU backend: work-item loops with fewer than 16 iterations are not vectorized (runtime SCEV checks + tiny trip count) (#868)

Commits

• 61bd51c Finalize 0.10: release notes, precompile warnings, compat and CI (#885)
• daccb07 Add @groupreduce and @groupscan (#830)
• d6392b6 KernelInterface: sub-group shuffles, votes, reductions and scans (#831)
• 2c67ccb Bump version from 0.10.0-dev to 0.10.0
• a0a2e68 Complete the 0.10 release notes and update the NUMA example (#842)
• 4794aae Compute indices in Int on PoCL, and mark the index arithmetic nuw/nsw (#879)
• ad1a8b9 POCL: recommend work-group sizes with PoCL's own heuristic (#861)
• bc4014d DGSEM benchmarks: a local-memory layout per backend (#875)
• 457a1ed DGSEM benchmarks: local memory with the variable index last (#871)
• a587ec6 DGSEM benchmarks: double-buffered slab kernel (#870)
• Additional commits viewable in compare view