Prepare 0.22.0

[michalharakal]

Summary

Release prep for SKaiNET 0.22.0, mirroring the 0.21.0 pattern (5839da98 — release branch merged into develop, then tagged from develop). Closes the M5 milestone of the JVM inference performance roadmap with a priority-100 native (FFM) CPU kernel provider.

Headline numbers (Linux x86_64, JDK 21.0.10, gcc 13.3 -O3 -ffast-math)

Kernel	shapes	speedup vs Panama
Native Q4_K matmul	1024² / 2048² / 4096²	5.87× / 4.71× / 4.17×
Native FP32 SGEMM	256³ / 512³ / 1024³	1.77× / 1.58× / 1.55×
MemSeg zero-copy Q4_K	4096²	+20% vs heap-copy

Native is single-threaded scalar C with -O3 -ffast-math; gcc/clang auto-vec emit AVX2 / NEON FMA without hand-tuned intrinsics. Surfaces an interesting incidental finding: native single-threaded beats Panama's parallelChunks multi-threaded path on every measured Q4_K shape — parallelChunks dispatch overhead seems to dominate.

What's in this release

• Native (FFM) CPU kernel provider (5-PR rollout): feat(native-cpu): scaffold FFM kernel provider module (PR 1 of 5) #571 scaffolding · feat(native-cpu): native FFM Q4_K matmul kernel (PR 2 of 5) #572 Q4_K matmul · feat(native-cpu): zero-copy Q4_K MemSeg kernel + SPI sibling (PR 3 of 5) #573 zero-copy MemSeg sibling · feat(native-cpu): cross-arch CI matrix + MSVC/Clang portability (PR 4 of 5) #574 cross-arch CI · feat(native-cpu): native FFM FP32 SGEMM kernel (PR 5 of 5) #575 FP32 SGEMM
• Publishing + BOM: build(native-cpu): add publishing config + register in BOM #576 wires vanniktech.mavenPublish and registers the new module in skainet-bom
• CI cleanup: ci(native-cpu): drop linux-arm64 runner from multiarch matrix #577 drops linux-arm64 from the cross-arch matrix (Kotlin/Native host limitation)
• Documentation: docs(native-cpu): document JVM-only consumption + shadow-jar caveat #579 NativeKernelProvider consumption kdoc

Multi-arch publishing — Windows AMD64 supported out of the box

publish.yml is now a two-phase flow:

1. build-native matrix on ubuntu-latest, macos-14, windows-latest — each builds its host's libskainet_kernels and uploads as an artifact.
2. publish on macOS — downloads all three artifacts, stages them into the native module's resources tree, runs ./gradlew publish. Resulting JAR ships native libs for linux-x86_64, macos-arm64, and windows-x86_64. Windows AMD64 devs get a working native path with no manual side-loading.

Limitations called out in CHANGELOG

• Linux ARM64 native lib not in the published JAR — Kotlin/Native plugin 2.3.21 doesn't support linux aarch64 as a HOST target. Linux ARM64 consumers fall back cleanly to Panama priority-50.
• Shadow-jar consumers on com.gradleup.shadow:9.4.x need the doLast workaround from SKaiNET-transformers PR Bump com.android.library from 8.12.1 to 8.12.2 #88 (mergeServiceFiles() bug). Spring Boot apps consuming via Maven are unaffected.

After merge

Following the same flow as 0.21.0:

1. Tag 0.22.0 on develop at the merge commit → publish.yml fires on tag push, runs the matrix + publishes signed artifacts to Maven Central.
2. Bump VERSION_NAME to 0.23.0-SNAPSHOT on develop (separate trivial commit, mirroring 4a3758af "Increment next version" after 0.21.0).

🤖 Generated with Claude Code