feat(native-cpu): cross-arch CI matrix + MSVC/Clang portability (PR 4 of 5)

.github/workflows/native-cpu-multiarch.yml

@@ -0,0 +1,117 @@
+name: Native CPU multi-arch build
+
+# Cross-arch CI for the FFM native kernel provider
+# (skainet-backends/skainet-backend-native-cpu). The local Gradle build
+# only produces a host-arch .so/.dylib/.dll; this workflow proves the
+# CMake + Kotlin pipeline works on every supported host so consumers
+# on Apple Silicon, ARM Linux, and Windows aren't silently broken when
+# they pull a published JAR built on x86_64 Linux.
+#
+# Each matrix job runs the native module's jvmTest end-to-end (CMake
+# configure + build + bundle into JAR resources + parity tests via
+# FFM downcall). The built shared library is uploaded as an artifact
+# so a later "fat-JAR" aggregation step (deferred to a follow-up PR)
+# can stage all four arches into one publishable artifact.
+
+on:
+  push:
+    branches: [main, develop]
+    paths:
+      - 'skainet-backends/skainet-backend-native-cpu/**'
+      - 'skainet-backends/skainet-backend-api/src/jvmMain/kotlin/sk/ainet/backend/api/kernel/**'
+      - '.github/workflows/native-cpu-multiarch.yml'
+  pull_request:
+    paths:
+      - 'skainet-backends/skainet-backend-native-cpu/**'
+      - 'skainet-backends/skainet-backend-api/src/jvmMain/kotlin/sk/ainet/backend/api/kernel/**'
+      - '.github/workflows/native-cpu-multiarch.yml'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  native-build-test:
+    name: ${{ matrix.arch_label }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - os: ubuntu-latest
+            arch_label: linux-x86_64
+            lib_name: libskainet_kernels.so
+          - os: ubuntu-24.04-arm
+            arch_label: linux-arm64
+            lib_name: libskainet_kernels.so
+          - os: macos-14
+            arch_label: macos-arm64
+            lib_name: libskainet_kernels.dylib
+          - os: windows-latest
+            arch_label: windows-x86_64
+            lib_name: skainet_kernels.dll
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 30
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+
+      - name: Copy CI gradle.properties (Unix)
+        if: runner.os != 'Windows'
+        run: |
+          mkdir -p ~/.gradle
+          cp .github/ci-gradle.properties ~/.gradle/gradle.properties
+
+      - name: Copy CI gradle.properties (Windows)
+        if: runner.os == 'Windows'
+        shell: pwsh
+        run: |
+          New-Item -ItemType Directory -Force -Path "$HOME\.gradle" | Out-Null
+          Copy-Item .github\ci-gradle.properties "$HOME\.gradle\gradle.properties"
+
+      - name: Set up JDK 25
+        uses: actions/setup-java@v5
+        with:
+          distribution: 'zulu'
+          java-version: 25
+
+      - name: Verify cmake
+        run: cmake --version
+
+      - name: Build + test native module (Unix)
+        if: runner.os != 'Windows'
+        env:
+          GRADLE_OPTS: -Dorg.gradle.jvmargs=-Xmx4g -Dfile.encoding=UTF-8
+        run: |
+          ./gradlew --no-daemon --stacktrace \
+            :skainet-backends:skainet-backend-native-cpu:jvmTest \
+            :skainet-backends:skainet-backend-native-cpu:jvmJar
+
+      - name: Build + test native module (Windows)
+        if: runner.os == 'Windows'
+        shell: pwsh
+        env:
+          GRADLE_OPTS: -Dorg.gradle.jvmargs=-Xmx4g -Dfile.encoding=UTF-8
+        run: |
+          .\gradlew.bat --no-daemon --stacktrace `
+            :skainet-backends:skainet-backend-native-cpu:jvmTest `
+            :skainet-backends:skainet-backend-native-cpu:jvmJar
+
+      - name: Upload native library
+        if: success()
+        uses: actions/upload-artifact@v7
+        with:
+          name: libskainet_kernels-${{ matrix.arch_label }}
+          path: skainet-backends/skainet-backend-native-cpu/build/native/resources/native/${{ matrix.arch_label }}/${{ matrix.lib_name }}
+          if-no-files-found: error
+          retention-days: 14
+
+      - name: Upload test reports
+        if: always()
+        uses: actions/upload-artifact@v7
+        with:
+          name: native-cpu-test-reports-${{ matrix.arch_label }}
+          path: |
+            skainet-backends/skainet-backend-native-cpu/build/reports/tests/**
+            skainet-backends/skainet-backend-native-cpu/build/test-results/**
+          retention-days: 14

skainet-backends/skainet-backend-native-cpu/native/CMakeLists.txt

@@ -24,8 +24,11 @@ if(WIN32)
     set_target_properties(skainet_kernels PROPERTIES PREFIX "")
 endif()
 
-# Hide non-exported symbols on ELF / Mach-O for a smaller surface area
-# and let the compiler auto-vectorize the Q4_K hot loop.
+# Per-compiler tuning. The Q4_K kernel hot loop is straight-line FP
+# arithmetic that auto-vectorizes cleanly under aggressive optimization
+# (AVX2 on x86_64, NEON on ARM64). Visibility is also handled here on
+# ELF / Mach-O; on Windows the SKAINET_API macro adds dllexport so we
+# don't need /VISIBILITY flags.
 if(CMAKE_C_COMPILER_ID MATCHES "Clang|GNU")
     target_compile_options(skainet_kernels PRIVATE
         -fvisibility=hidden
@@ -35,4 +38,10 @@ if(CMAKE_C_COMPILER_ID MATCHES "Clang|GNU")
         -funroll-loops
     )
     set_target_properties(skainet_kernels PROPERTIES C_VISIBILITY_PRESET hidden)
+elseif(CMAKE_C_COMPILER_ID MATCHES "MSVC")
+    target_compile_options(skainet_kernels PRIVATE
+        /O2
+        /fp:fast
+        /W3
+    )
 endif()

skainet-backends/skainet-backend-native-cpu/native/include/skainet_kernels.h

@@ -11,6 +11,17 @@
 #  define SKAINET_API
 #endif
 
+/* Portable "restrict" qualifier: GNU/Clang accept __restrict__,
+ * MSVC accepts __restrict, and the C99 keyword `restrict` is
+ * unreliable across compiler modes. */
+#if defined(__GNUC__) || defined(__clang__)
+#  define SKAINET_RESTRICT __restrict__
+#elif defined(_MSC_VER)
+#  define SKAINET_RESTRICT __restrict
+#else
+#  define SKAINET_RESTRICT
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif

skainet-backends/skainet-backend-native-cpu/native/src/q4k_matmul.c

@@ -76,13 +76,13 @@ static inline void skainet_q4k_decode_scales(
  * codeSum/inputSum accumulators on AVX2/NEON.
  */
 SKAINET_API void skainet_q4k_matmul(
-    const float* __restrict__ input,
+    const float* SKAINET_RESTRICT input,
     int32_t input_offset,
-    const uint8_t* __restrict__ weight,
+    const uint8_t* SKAINET_RESTRICT weight,
     int32_t weight_byte_offset,
     int32_t input_dim,
     int32_t output_dim,
-    float* __restrict__ output,
+    float* SKAINET_RESTRICT output,
     int32_t output_offset
 ) {
     if (output_dim <= 0 || input_dim <= 0) return;