v1.5.0

Metal v1.5.0

Diff since v1.4.2

Metal.jl 1.5 is a relatively minor release, which the most important change being behind the scenes: GPUArrays.jl v11 has switched to KernelAbstractions.jl (#461).

There is also one (technically) breaking change: code_agx and @device_code_agx have been removed (#512) because of the heavy Python dependency, and conflicts with PythonCall.jl. This functionality did not support recent M GPUs anyway, so it is unlikely to affect many users.

Features

• Improve performance of shared storage copies: #445
• Add an is_m4 function: #498
• #499

Bug fixes

• Fix fill: #496

Merged pull requests:

• Add more tests to api validation testing (#447) (@christiangnrd)
• Adapt to GPUArrays.jl transition to KernelAbstractions.jl. (#461) (@maleadt)
• Switch CI to 1.11. (#462) (@maleadt)
• Remove old code and test cleanup (#464) (@christiangnrd)
• Adapt to JuliaGPU/GPUArrays.jl#567. (#475) (@maleadt)
• Bump LLVM downgrader (#479) (@maleadt)
• Store more debug files when encountering compilation errors. (#482) (@maleadt)
• Use OncePerProcess in 1.12+ (#483) (@christiangnrd)
• Don't run benchmarks from fork (#485) (@christiangnrd)
• Still run GH Action when merged (#486) (@christiangnrd)
• Bump IR downgrader (#489) (@maleadt)
• Move MTL tests and add a few (#491) (@christiangnrd)
• Generate MTL and MPS structs and enums with Clang.jl (#492) (@christiangnrd)
• Fix copy tests (#493) (@christiangnrd)
• Simplify benchmark runner and pipelines (#494) (@maleadt)
• Fix global linear indexing (fill!) (#496) (@christiangnrd)
• Couple typos and is_m4 function (#498) (@christiangnrd)
• Initial support for MPSNDArray (#499) (@christiangnrd)
• Tweak benchmark CI job (#501) (@maleadt)
• Fix MPSNDArrayDescriptor wrapper (#502) (@christiangnrd)
• Metal library parsing: using CodecBzip2 feature to ignore padding. (#504) (@maleadt)
• Followup to #492: Enable C function wrapping (#505) (@christiangnrd)
• Rerun random tests with chance of false negative once. (#506) (@christiangnrd)
• Bump LLVM downgrader (#507) (@maleadt)
• Test loading of package on unsupported platforms (#509) (@christiangnrd)
• Remove device_code_agx (#512) (@christiangnrd)
• Fix typo in random tests (#514) (@christiangnrd)
• Fix Documenter failures (#515) (@christiangnrd)

Closed issues:

• KernelAbstractions: add Atomix back-end (#218)
• @device_code_agx errors when Metal Shader Validation is enabled (#463)
• fill broken after KA integration (#466)
• Compilation to native code failed: NSError: Undefined symbols (#480)
• ObjectiveC.Foundation.NSErrorInstance(ObjectiveC.id{ObjectiveC.Foundation.NSError}(0x000000014cb8bd90)) (#487)
• phi-related IR downgrade issue (#488)
• Circular dependency when precompiling (#495)
• Bad interaction between PyCall and Metal (#500)
• Add github actions CI for linux, windows and non-functional macOS to ensure that precompilation and loading works (#508)