v1.4.2

Metal v1.4.2

Diff since v1.4.1

Merged pull requests:

• Fix loading on unsupported platforms (#459) (@christiangnrd)

Closed issues:

• Relax package requirements (#22)
• [windows:] Metal does not precompile anymore when installation not functional (#457)
• [MacOS:] Metal.functional() wrongly returns true despite no GPUs available (#458)

v1.4.1

Metal v1.4.1

Diff since v1.4.0

Merged pull requests:

• Update Readme (#444) (@christiangnrd)
• Use CPU copy with SharedStorage (#445) (@christiangnrd)
• Disable nightly CI and fix invalid Metal API usage (#448) (@christiangnrd)
• Don't report benchmarks on main branch commits (#450) (@christiangnrd)
• Fix #451 and a couple other fixes (#452) (@christiangnrd)
• Only load BFloat16s extension on Apple systems (#454) (@christiangnrd)
• CompatHelper: bump compat for GPUCompiler to 1, (keep existing compat) (#455) (@github-actions[bot])

Closed issues:

• Don't run benchmarks on the master branch? (#449)
• unsafe_wrap(Array, ...) of a view does not preserve offset information (#451)
• Metal does not load any more without error when installation not functional (#453)

v1.4.0

Metal v1.4.0

Diff since v1.3.0

Merged pull requests:

• Use unified memory for scalar indexing of permutation matrices (#313) (@tgymnich)
• Add MPSMatrixRandom (#321) (@christiangnrd)
• [.gitignore] Also ignore versioned Manifests (#410) (@christiangnrd)
• Remove broken link in Docs (#413) (@christiangnrd)
• Remove unused [extras] section in Project.toml (#415) (@christiangnrd)
• Small fix and typos (#417) (@christiangnrd)
• Add Benchmarking CI (#420) (@christiangnrd)
• [NFC] Fix warning in topk docstrings (#421) (@christiangnrd)
• Allow initialisation of MTLSize with tuples of different integer types (#425) (@tgymnich)
• Add CI for macOS 15 (#426) (@christiangnrd)
• Simplify versioninfo() and report more packages. (#429) (@maleadt)
• Allow controlling compilation target versions. (#430) (@maleadt)
• Add a missing memory fence to a SIMD test. (#432) (@maleadt)
• Fix MPS.synchronize_state (#434) (@christiangnrd)
• Make lu results have same storage mode as input (#435) (@christiangnrd)
• Fix benchmarking CI and benchmark Shared and Private storage modes (#437) (@christiangnrd)
• NFC tweak to MPSMatrixCopy tests (#439) (@christiangnrd)
• Get more descriptive errors from flaky test (#440) (@christiangnrd)

Closed issues:

• Port the opportunistic synchronization from CUDA.jl (#317)
• Control flow-related miscompilation: (#401)
• More sporadic 1.11 hangs (#412)
• Support for LinearAlgebra.kron (#422)
• Can't use gemm! methods with Metal (#423)
• Error for thread/group size with different integer types (#424)
• README example broken (#427)
• Intermittent load_store_tg test failure (#428)

v1.3.0

Metal v1.3.0

Diff since v1.2.0

Merged pull requests:

• Fix typo in docs (#384) (@christiangnrd)
• Bump minimal Julia requirement to v1.10. (#385) (@maleadt)
• Remove Requires dependency (#386) (@christiangnrd)
• Reflection: Figure out kernel names by looking at metallib section. (#390) (@maleadt)
• Add tests for broadcasting minimum and maximum (#391) (@tgymnich)
• Don't export MTL (#392) (@christiangnrd)
• Add erfinv (#394) (@tgymnich)
• Add expm1 (#395) (@tgymnich)
• Cleanup some imports (#398) (@christiangnrd)
• Remove type-pirated function (#399) (@christiangnrd)
• Unexport some high-level MPS functionality from MPS (#400) (@christiangnrd)
• Adapt to new REPL precompile changes (JuliaLang/julia#55210) (#403) (@christiangnrd)
• Bump GPUCompiler. (#404) (@maleadt)
• Bump LLVM compat (#407) (@maleadt)
• Make 1.11 CI success mandatory. (#408) (@maleadt)

Closed issues:

• Audit exports/public symbols (#359)
• Compilation failure on 1.11 (#370)
• MTLBinaryArchive (#387)
• Metal.code_agx() failing in MacOS 15 Beta 3 (#388)
• Test for min / max broadcasting issue (#389)
• Type piracy (#396)
• Potentially unused code in gpuarrays.jl (#397)
• Shared vs SharedStorage in examples/unified_memory (#405)
• Unsuported call to an unknown function when calling Distributions (#406)

v1.2.0

Metal v1.2.0

Diff since v1.1.0

Merged pull requests:

• Avoid constructing MulAddMuls on Julia v1.12+ (#295) (@dkarrasch)
• Trigger the runtime profiler when a test times out. (#330) (@maleadt)
• Add MPSMatrixSoftMax (#333) (@christiangnrd)
• Reorganize and add some MPS tests (#335) (@christiangnrd)
• Typo fix (#336) (#337) (@101001000)
• Add error message for running Metal.jl under Rosetta (#339) (@tgymnich)
• Add MPSCommandBuffer (#340) (@christiangnrd)
• Bump julia-actions/setup-julia from 1 to 2 (#341) (@dependabot[bot])
• Revert error message for Rosetta (#342) (@tgymnich)
• Update to ObjectiveC.jl v3. (#343) (@maleadt)
• Add autoreleasepools to MPS interface methods. (#344) (@maleadt)
• Don't redundantly return the cmdbuf from commit methods. (#345) (@maleadt)
• Whitespace fixes (#346) (@christiangnrd)
• CompatHelper: bump compat for LLVM to 7, (keep existing compat) (#347) (@github-actions[bot])
• CompatHelper: add new compat entry for SpecialFunctions in [weakdeps] at version 2, (keep existing compat) (#352) (@github-actions[bot])
• [NFC] Fix indentation (#353) (@christiangnrd)
• Bump LLVM downgrader (#354) (@maleadt)
• Don't export non-existent contents (#356) (@christiangnrd)
• Remove/fix unused exports (#357) (@christiangnrd)
• Unexport SimpleVersion and AS (#360) (@christiangnrd)
• Add support for opaque pointers (#361) (@maleadt)
• Docstrings (#362) (@christiangnrd)
• Initial MacOS 15 support (#365) (@christiangnrd)
• Replace current_device() with device() (#366) (@christiangnrd)
• Support reading metallib v1.2.8 files from macOS 15. (#367) (@maleadt)
• Add metallib (dis)assembly helper scripts. (#368) (@maleadt)
• Simplify testing of examples. (#369) (@maleadt)
• Temporarily allow 1.11 to fail. (#371) (@maleadt)
• CompatHelper: add new compat entry for PrecompileTools at version 1, (keep existing compat) (#372) (@github-actions[bot])
• Define complex sqrt (#374) (@mtfishman)
• Check the macOS version during initialization. (#375) (@maleadt)
• CompatHelper: bump compat for LLVM to 8, (keep existing compat) (#376) (@github-actions[bot])
• Add accumulate implementation (#377) (@chengchingwen)
• fix derived device array (#378) (@chengchingwen)
• avoid ReshapedArray using Int128 in metal kernel (#379) (@chengchingwen)
• improve type stability of derived array (#380) (@chengchingwen)
• add findall implementation (#382) (@zhenwu0728)
• Bump version (#383) (@christiangnrd)

Closed issues:

• Tests sporadically timing out on 1.11 (#329)
• ReshapedArray indexing broken because of Int128 operation (#332)
• KernelAbstractions copyto! typo (#336)
• Segmentation Faults (#338)
• Port accmulate! and findall from CUDA.jl (#348)
• Tests failing with GPUCompiler v0.26.5 and LLVM v7.1 (#350)
• downgrades LLVM (#355)
• sqrt(::Complex) unsupported due to conversion exceptions (#364)

v1.1.0

Metal v1.1.0

Diff since v1.0.0

Merged pull requests:

• Add resize! (#279) (@mtfishman)
• Initial MTLTexture support (#280) (@christiangnrd)
• Avoid redundant pointer conversions for threadgroup memory. (#283) (@maleadt)
• Re-implement metallib generation in Julia. (#284) (@maleadt)
• CompatHelper: add new compat entry for SHA at version 0.7, (keep existing compat) (#286) (@github-actions[bot])
• Support more of the metallib format (#288) (@maleadt)
• Address potentiallly buggy mtl behaviour. (#290) (@christiangnrd)
• CompatHelper: add new compat entry for CodecBzip2 at version 0.8, (keep existing compat) (#292) (@github-actions[bot])
• Remove an unneeded pointer method. (#293) (@maleadt)
• Use NSAutoreleasePool to clean up memory. (#294) (@maleadt)
• adapt_storage-related improvements (#296) (@christiangnrd)
• CompatHelper: bump compat for ObjectiveC to 2, (keep existing compat) (#297) (@github-actions[bot])
• Add support for signposts (#300) (@maleadt)
• Retain NSError we rethrow to avoid an UAF. (#302) (@maleadt)
• Minor mapreduce improvements (#303) (@maleadt)
• Specialize broadcast to avoid integer divisions. (#304) (@maleadt)
• Better Support for Unified Memory (#305) (@tgymnich)
• Add 1.11 CI (#306) (@christiangnrd)
• Remove unused files (#307) (@tgymnich)
• Skip profiling tests on macOS 14.4/M1. (#310) (@maleadt)
• Increase test timeout limit to accomodate 1.8 (#311) (@christiangnrd)
• Test all storage modes (#314) (@christiangnrd)
• Fix doctests (#315) (@christiangnrd)
• Fix KernelAbstractions for Unified Memory (#316) (@tgymnich)
• CompatHelper: add new compat entry for Preferences at version 1, (keep existing compat) (#318) (@github-actions[bot])
• Minor cleanup (#319) (@christiangnrd)
• Create MtlArray using memory allocated by Array (#320) (@christiangnrd)
• Re-enable profiling tests on M1/14.4 when using Xcode 15.3. (#322) (@maleadt)
• Small typo and doc fixup (#325) (@christiangnrd)
• BFloat16s.jl extension and related improvements (#326) (@christiangnrd)
• Support for Julia 1.11 (#327) (@maleadt)

Closed issues:

• Validation-related back-end crash on macOS Ventura (#34)
• slow broadcast copy in 2D (#41)
• Poor performance of mapreduce (#46)
• Multiplication with SubArrays (#47)
• Add support to creating MtlArray using a memory allocated by Array (#62)
• Improve use of unified memory (#86)
• Use Autoreleasepools with Metal (#103)
• Unknown RFLT tag generated by macOS 13 Metal compiler (#167)
• mapreduce allocates a lot on the CPU (#211)
• Legalization errors with vectorized code (#257)
• Compilation Failure due to undefined symbols (#276)
• resize!, append! not defined (#277)
• tag new version (#278)
• Panic during profiling tests on 14.4 beta (#281)
• M3 backend cannot handle atomics with complicated pointer conversions (#282)
• Int128 does not compile (#287)
• Two suspicious mtl-related behaviours (#289)
• LU factorization: add allowsingular keyword argument (#299)
• Autorelease changes lead to use after free with errors (#301)
• Reductions don't work on Shared Arrays (#312)

v1.0.0

Metal v1.0.0

Diff since v0.5.1

Merged pull requests:

• Matrix batches (#158) (@tgymnich)
• Add 1.10 CI. (#256) (@maleadt)
• Update manifest (#258) (@github-actions[bot])
• CompatHelper: bump compat for GPUCompiler to 0.25, (keep existing compat) (#259) (@github-actions[bot])
• Bump actions/checkout from 3 to 4 (#260) (@dependabot[bot])
• Update manifest (#261) (@github-actions[bot])
• CompatHelper: bump compat for CEnum to 0.5, (keep existing compat) (#262) (@github-actions[bot])
• Update manifest (#263) (@github-actions[bot])
• CompatHelper: add new compat entry for Artifacts at version 1, (keep existing compat) (#264) (@github-actions[bot])
• Reduce launch overhead by generating code to encode arguments. (#265) (@maleadt)
• Remove unused function argument (#266) (@tgymnich)
• Introduce application tracing profiler (#267) (@maleadt)
• Remove content(::MTLBuffer), use convert intead. (#268) (@maleadt)
• Allow more kwargs syntax with kernel launches (#269) (@maleadt)
• Don't re-use the IO object when shelling out to Python. (#271) (@maleadt)
• Preserve storage mode when broadcasting. (#273) (@maleadt)

Closed issues:

• Support for macOS Sonoma (#201)
• Error with Julia 1.10 (#274)

v0.5.1

Metal v0.5.1

Diff since v0.5.0

Merged pull requests:

• MPSMatrix improvements (#157) (@tgymnich)
• Update manifest (#221) (@github-actions[bot])
• Update manifest (#222) (@github-actions[bot])
• Update manifest (#224) (@github-actions[bot])
• Update manifest (#227) (@github-actions[bot])
• CompatHelper: bump compat for ObjectiveC to 1, (keep existing compat) (#228) (@github-actions[bot])
• Update manifest (#230) (@github-actions[bot])
• Fix argument types in sincos (#232) (@fjebaker)
• Update manifest (#233) (@github-actions[bot])
• Improve docs (#235) (@christiangnrd)
• Remove linear algebra section of MPS docs (#237) (@christiangnrd)
• CompatHelper: bump compat for GPUCompiler to 0.22, (keep existing compat) (#238) (@github-actions[bot])
• Port openlibm log1pf as log1p (#239) (@sotlampr)
• Port openlibm erf (#240) (@tgymnich)
• Remove 1.6-era override mechanism. (#241) (@maleadt)
• CompatHelper: add new compat entry for Requires at version 1, (keep existing compat) (#242) (@github-actions[bot])
• Update manifest (#243) (@github-actions[bot])
• enable dependabot for GitHub actions (#244) (@ranocha)
• Bump actions/checkout from 2 to 3 (#245) (@dependabot[bot])
• Bump peter-evans/create-pull-request from 3 to 5 (#246) (@dependabot[bot])
• Show METAL_CAPTURE_ENABLED in Metal.versioninfo() when the environment variable is set (#248) (@christiangnrd)
• Update manifest (#249) (@github-actions[bot])
• Adapt to GPUCompiler.jl, and other small updates. (#250) (@maleadt)
• Switch to GPUArrays buffer management. (#251) (@maleadt)
• Update manifest (#252) (@github-actions[bot])
• Update manifest (#253) (@github-actions[bot])
• Bump GPUCompiler (#255) (@maleadt)

Closed issues:

• Random access indexing into MtlArray views cause scalar indexing (#149)
• Q: How to debug kernels - KA.@print? (#223)
• Crash during MTLDispatchListApply (#225)
• Unable to compile trig functions through ForwardDiff (#229)
• symbol multiply defined! Bug/crash on Julia master, fine on 1.10 (#231)
• log1p fails on MtlArray{Float32} (#234)
• When precompiling, UndefVarError: CompilerConfig not defined (#247)

v0.5.0

Metal v0.5.0

Diff since v0.4.1

Metal.jl 0.5 is a feature release, bringing initial support for atomic operations (#168).
Low-level atomics that mimic Metal C are supported (atomic_store_explicit,
atomic_load_explicit, etc), as well as a higher-level Metal.@atomic that can be used to
update array values similar to how CUDA.jl's @atomic works. This uses native atomics when
supported, and falls back to a compare-exchange loop otherwise.

Minor changes include an update for the @device_code_agx disassembler, the addition of a
type variable to MtlArray encoding the storage mode (#194), and support for MPSVector
(#199) which should accelerate matrix/vector multiplications.

Also note that Metal.jl now disallows the construction of Float64 arrays, as these are not
support by the Metal libraries.

Closed issues:

• Support for atomics (#79)
• Make MtlArray storage mode a type parameter (#190)
• Long stacktrace when trying to create Float64 rand arrays (#205)
• allowscalar equivalent for Metal.jl (#206)
• Define map! ? (#219)

Merged pull requests:

• Implement atomics using compiler intrinsics (#168) (@maleadt)
• Parameterize MtlArray storage mode (#194) (@christiangnrd)
• Implement MPSVector (#199) (@tgymnich)
• Update manifest (#200) (@github-actions[bot])
• Add Metal 3.1 to MTLLanguageVersion (#202) (@christiangnrd)
• Update manifest (#203) (@github-actions[bot])
• CompatHelper: bump compat for GPUCompiler to 0.21, (keep existing compat) (#204) (@github-actions[bot])
• Update manifest (#207) (@github-actions[bot])
• Disallow Float64 arrays entirely. (#209) (@maleadt)
• Adapt to LLVM.jl 6. (#213) (@maleadt)
• Update manifest (#215) (@github-actions[bot])
• Bump disassembler. (#216) (@maleadt)

v0.4.1

Metal v0.4.1

Diff since v0.4.0

Closed issues:

• Command buffer callbacks can cause bus error during thread adoption (#138)
• how to set up Project.toml (#185)
• Metal.rand() creates a CPU array (#187)
• fill! for Int8 errors when the value is negative (#192)

Merged pull requests:

• Refactor matmatmul code for faster load time (#186) (@dkarrasch)
• Add *.DS_Store to .gitignore (#188) (@christiangnrd)
• Add GPUArrays out-of-place random methods (#189) (@tgymnich)
• Revert "Don't rely on thread adoption for command buffer callbacks." (#191) (@maleadt)
• Fix fill! with negative Int8 values (#193) (@christiangnrd)
• disambiguate gemm_wrapper! with LinAlg.jl (#195) (@dkarrasch)
• Add type annotations for character args in matmatmul (#196) (@dkarrasch)
• Handle missing adjoint case. (#197) (@maleadt)
• Fix transposed matmul. (#198) (@maleadt)