v1.9.1

Metal v1.9.1

Diff since v1.9.0

This release contains a few bug fixes, including the potentially nasty #710 where gpu-gpu copies failed for buffers >4GiB.

Merged pull requests:

• Check max threadgroup size before launching mapreduce kernel (#698) (@christiangnrd)
• fix rand(::AbstractRange{Int64}) (#700) (@simeonschaub)
• Skip all broken rand tests (#709) (@christiangnrd)
• Fix large (>4GiB) GPU -> GPU copies (#712) (@christiangnrd)

Closed issues:

• Array test failing for seemingly no reason. (#442)
• Occasional reduction errors (#616)
• MtlArray transfer silently fails for arrays larger than 4GB (2^32 bytes) (#710)
• trunc(Int, ...) doesn't work; neither does convert(MtlVector{Int64}, ...) (#707)

v1.9.0

Metal v1.9.0

Diff since v1.8.2

Merged pull requests:

• Use Julia 1.12 for CI (#674) (@christiangnrd)
• Switch to ParallelTestRunner.jl. (#676) (@maleadt)
• Enable MPS rand for more array sizes (#677) (@christiangnrd)
• Adapt to nightly ccall changes (#678) (@christiangnrd)
• Fix mul_hi for 1.13 (#679) (@christiangnrd)
• Device-side RNG (#657) (@simeonschaub)
• Fix launch overhead regression (#669) (@maleadt)

Closed issues:

• MTL.append_copy! silently ignores Metal documentation restriction (#349)
• Basic scalar rand() support (#481)

v1.8.2

Metal v1.8.2

Diff since v1.8.1

This release adds support for Julia 1.12.

Merged pull requests:

• Updates for Julia 1.12 (#576) (@christiangnrd)
• Switch to macos-15-intel for x64 load tests (#660) (@christiangnrd)
• Update test runner for Julia 1.13. (#663) (@maleadt)
• Fix type instability in KA.allocate (#665) (@christiangnrd)
• CompatHelper: bump compat for BFloat16s to 0.6, (keep existing compat) (#667) (@github-actions[bot])
• GPUToolbox compat (#668) (@christiangnrd

Closed issues:

• 1.12 Support (#597)

v1.8.1

Metal v1.8.1

Diff since v1.8.0

Merged pull requests:

• Fix serial reductions (#662) (@christiangnrd)

v1.8.0

Metal v1.8.0

Diff since v1.7.0

Merged pull requests:

• Add and test shuffle and fill intrinsics (#555) (@christiangnrd)
• More MPSGraphs API (#624) (@christiangnrd)
• Port mapreduce optimization from CUDA (#627) (@christiangnrd)
• Synchronize using MTLSharedEvents (#633) (@christiangnrd)
• Various documentation fixes. (#641) (@christiangnrd)
• Revert "Synchronize using MTLSharedEvents" (#645) (@christiangnrd)
• Remove old renamed/removed benchmarks (#646) (@christiangnrd)
• Remove support for MtlArrays backed by managed storage. (#649) (@christiangnrd)
• Re-add old valid benchmarks and remove problematic trailing commas (#651) (@christiangnrd)
• Improve GPU core # retrieval (#652) (@christiangnrd)
• Bump actions/checkout from 4 to 5 (#654) (@dependabot[bot])
• Don't transpose empty MtlArrays (#658) (@christiangnrd)
• Support for macOS 26 and misc fixes (#659) (@christiangnrd)

Closed issues:

• Problem using @atomic in a kernel (#619)
• Remove support for managed storage (#642)
• LinearAlgebra.transpose! with MtlArrays with size 0 along an axis cause a Metal error (#656)

v1.7.0

Metal v1.7.0

Diff since v1.6.4

Merged pull requests:

• Add function to retrieve # of gpu cores in system (#626) (@christiangnrd)
• Support KA unified memory (#630) (@christiangnrd)
• Add GPUToolbox 0.3 compat (#639) (@christiangnrd)
• Return the old value from atomic_fetch_op_explicit. (#640) (@maleadt)

v1.6.4

Metal v1.6.4

Diff since v1.6.3

Merged pull requests:

• typo in MPSMatrixMultiplication comment (#622) (@jandrej)
• Remove unnecessary OS signposts (#623) (@christiangnrd)
• Accept alternate filename as optional argument (#629) (@christiangnrd)
• Support Float32 threadgroup atomics by bitcasting. (#636) (@maleadt)

Closed issues:

• @signpost_events make the code awfully slow (#621)

v1.6.3

Metal v1.6.3

Diff since v1.6.2

Merged pull requests:

• More accumulation and reduction benchmarks (#614) (@christiangnrd)
• Remove the unnecessary reshape during mapreduce (#615) (@christiangnrd)
• Synchronize resources before cpu access of ManagedStorage resource (#617) (@christiangnrd)
• Fix linalg tests for MPS and MPSGraph (#618) (@christiangnrd)
• Don't warn on macOS 26 and bump version (#620) (@christiangnrd)

v1.6.2

Metal v1.6.2

Diff since v1.6.1

Merged pull requests:

• Handle broadcasting when storage types are different (#605) (@limarta)
• Add JLD2 to test env (#606) (@christiangnrd)
• Tahoe versions (#607) (@christiangnrd)
• Add MemoryFlagDevice to KA.jl's synchronization primitive. (#609) (@maleadt)
• Update wrappers (#610) (@christiangnrd)
• Bump version (#611) (@christiangnrd)

Closed issues:

• KA.@synchronize -- threadgroup_barrier semantics (#608)

v1.6.1

Metal v1.6.1

Diff since v1.6.0

Merged pull requests:

• Adding definition for KA.functional (#598) (@astrozot)
• Update requirements (#599) (@christiangnrd)
• Fix findall with empty MtlArray of Bool (#601) (@christiangnrd)
• [NFC] Typo (#602) (@christiangnrd)
• Add bare minimum for macOS 26 Tahoe (#604) (@christiangnrd)

Closed issues:

• Warnings when precompiling Metal with Julia 1.12 (#594)