Ship native binaries for four RIDs

.github/workflows/embree-cmake.yml

@@ -19,6 +19,11 @@ on:
         required: false
         type: string
         default: 'v4.4.1'
+      only-rid:
+        description: 'Build a single RID (e.g. win-arm64). Empty builds all five.'
+        required: false
+        type: string
+        default: ''
 
 jobs:
   build:
@@ -27,55 +32,101 @@ jobs:
       fail-fast: false
       matrix:
         include:
+          # `parallel` is the compiler job count. An x64 build compiles the kernels once per
+          # ISA (SSE2, SSE4.2, AVX, AVX2) and Embree's templated translation units are heavy:
+          # four concurrent g++ processes exhaust the 16 GB of a standard runner and the OOM
+          # killer takes the runner down with them, which surfaces as a job that dies after
+          # 47 minutes with no logs at all rather than as a compiler error. The NEON legs
+          # build a single ISA and have room to spare.
           - os: ubuntu-24.04
             rid: linux-x64
             max-isa: AVX2
+            parallel: 2
           - os: ubuntu-24.04-arm
             rid: linux-arm64
             max-isa: NEON
+            parallel: 4
+          # `cmake-arch` feeds the Visual Studio generator's -A flag. Without it CMake picks
+          # the generator's default platform, which on the ARM64 runner is not ARM64: the
+          # build then compiled the x86 SIMD paths and immintrin.h rejected the target.
           - os: windows-latest
             rid: win-x64
             max-isa: AVX2
-          - os: windows-11-arm
-            rid: win-arm64
-            max-isa: NEON
+            cmake-arch: x64
+            parallel: 4
+          # There is no win-arm64 leg. Embree 4.4.1 does not build for Windows on ARM with
+          # either toolset available on the runner, and the two failures are different walls:
+          #
+          #   MSVC     — Embree gates its ARM path on __ARM_NEON and __aarch64__, GCC/Clang
+          #              predefines that MSVC does not define on ARM64, so
+          #              common/sys/intrinsics.h takes its x86 branch and includes
+          #              immintrin.h, which rejects the target.
+          #   clang-cl — clears that wall, then Embree's own CMake feeds -msse2 to an
+          #              arm64-pc-windows-msvc target and clang-cl refuses it.
+          #
+          # Both are upstream gaps in a configuration Embree neither supports nor tests.
+          # Getting past the second would mean patching Embree's CMake and carrying that
+          # patch, for a ray tracing kernel where a wrong SIMD path gives wrong intersections
+          # rather than a crash. See the tracking issue before trying again.
           - os: macos-latest
             rid: osx-arm64
             max-isa: NEON
+            parallel: 4
 
     steps:
       - name: Checkout Embree
+        if: inputs.only-rid == '' || inputs.only-rid == matrix.rid
         uses: actions/checkout@v4
         with:
           repository: RenderKit/embree
           ref: ${{ inputs.embree-ref || 'v4.4.1' }}
 
       - name: Install build dependencies (Linux)
-        if: startsWith(matrix.os, 'ubuntu')
+        if: (inputs.only-rid == '' || inputs.only-rid == matrix.rid) && startsWith(matrix.os, 'ubuntu')
         run: |
           sudo apt-get update
           sudo apt-get install -y build-essential cmake
 
       - name: Configure
+        if: inputs.only-rid == '' || inputs.only-rid == matrix.rid
         shell: bash
         run: |
+          set -o pipefail
           cmake -S . -B build \
+            ${{ matrix.cmake-arch && format('-A {0}', matrix.cmake-arch) || '' }} \
+            ${{ matrix.toolset && format('-T {0}', matrix.toolset) || '' }} \
             -DCMAKE_BUILD_TYPE=Release \
             -DEMBREE_TASKING_SYSTEM=INTERNAL \
             -DEMBREE_ISPC_SUPPORT=OFF \
             -DEMBREE_TUTORIALS=OFF \
             -DEMBREE_STATIC_LIB=OFF \
             -DEMBREE_GEOMETRY_INSTANCE_ARRAY=ON \
-            -DEMBREE_MAX_ISA=${{ matrix.max-isa }}
+            -DEMBREE_MAX_ISA=${{ matrix.max-isa }} 2>&1 | tee configure.log
+
+      # Fail early and loudly if an ARM leg did not end up on the ARM code path: an x86 SIMD
+      # build for an ARM target dies deep inside immintrin.h with an error that says nothing
+      # about the real cause.
+      #
+      # This greps Embree's own configure output rather than CMakeCache.txt. EMBREE_ARM is a
+      # plain variable set with SET(), not a cache entry, so it never reaches the cache file
+      # and an assertion against it fails on a perfectly good build.
+      - name: Check ARM detection
+        if: (inputs.only-rid == '' || inputs.only-rid == matrix.rid) && matrix.max-isa == 'NEON'
+        shell: bash
+        run: |
+          grep -qE "Building for (AArch64|Apple silicon)" configure.log \
+            || { echo "CMake did not take Embree's ARM path:"; grep -iE "Building for|CMAKE_GENERATOR_PLATFORM" configure.log; exit 1; }
 
       - name: Build
-        run: cmake --build build --config Release --parallel
+        if: inputs.only-rid == '' || inputs.only-rid == matrix.rid
+        run: cmake --build build --config Release --parallel ${{ matrix.parallel }}
 
       # On Linux and macOS CMake emits versioned names behind symlinks
       # (libembree4.so -> libembree4.so.4 -> libembree4.so.4.4.1). Symlinks do not survive
       # NuGet packaging, so the real file is copied out under the plain name that
       # [DllImport("embree4")] probes for.
       - name: Collect artifacts
+        if: inputs.only-rid == '' || inputs.only-rid == matrix.rid
         shell: bash
         run: |
           set -euo pipefail
@@ -97,6 +148,7 @@ jobs:
           ls -l out
 
       - name: Upload
+        if: inputs.only-rid == '' || inputs.only-rid == matrix.rid
         uses: actions/upload-artifact@v4
         with:
           name: embree-${{ matrix.rid }}

EmbreeGen/Headers/embree4/rtcore_config.h

@@ -1,24 +1,12 @@
 // Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
-// This file is normally generated by CMake from kernels/rtcore_config.h.in.
-// It is vendored here so the binding generator produces code that matches the
-// native binaries shipped in Evergine.Bindings.Embree/runtimes/.
-//
-// It corresponds to Embree v4.4.1 configured by .github/workflows/embree-cmake.yml:
-//   -DEMBREE_TASKING_SYSTEM=INTERNAL
-//   -DEMBREE_ISPC_SUPPORT=OFF
-//   -DEMBREE_TUTORIALS=OFF
-//   -DEMBREE_STATIC_LIB=OFF
-//   -DEMBREE_GEOMETRY_INSTANCE_ARRAY=ON
-// and the upstream defaults EMBREE_MAX_INSTANCE_LEVEL_COUNT=1, EMBREE_MIN_WIDTH=OFF,
-// EMBREE_SYCL_SUPPORT=OFF, EMBREE_API_NAMESPACE="".
-//
-// If any of those options change, regenerate this file from the build output and
-// re-run the generator, otherwise struct layouts silently drift from the binaries.
-
 #pragma once
 
+#if !defined(EMBREE_SYCL_SUPPORT)
+/* #undef EMBREE_SYCL_SUPPORT */
+#endif
+
 #define RTC_VERSION_MAJOR 4
 #define RTC_VERSION_MINOR 4
 #define RTC_VERSION_PATCH 1
@@ -37,11 +25,13 @@
 #define EMBREE_MIN_WIDTH 0
 #define RTC_MIN_WIDTH EMBREE_MIN_WIDTH
 
+#if !defined(EMBREE_STATIC_LIB)
 /* #undef EMBREE_STATIC_LIB */
+#endif
 /* #undef EMBREE_API_NAMESPACE */
 
 #if defined(EMBREE_API_NAMESPACE)
-#  define RTC_NAMESPACE
+#  define RTC_NAMESPACE 
 #  define RTC_NAMESPACE_BEGIN namespace  {
 #  define RTC_NAMESPACE_END }
 #  define RTC_NAMESPACE_USE using namespace ;
@@ -110,3 +100,4 @@
 #  define RTC_SYCL_INDIRECTLY_CALLABLE
 #  define RTC_SYCL_API RTC_API
 #endif
+

README.md

@@ -39,9 +39,13 @@ CPU ray tracing, lightmap baking, ambient occlusion and general geometric querie
 
 ## Supported Platforms
 
-- [x] Windows x64, ARM64
+- [x] Windows x64
 - [x] Linux x64, ARM64
 - [x] MacOS ARM64
+- [ ] Windows ARM64 — Embree 4.4.1 does not build for this target with either toolset on the
+  runner: MSVC does not define the `__ARM_NEON`/`__aarch64__` macros Embree gates its ARM path
+  on, and clang-cl gets past that only to be handed `-msse2` for an ARM64 target by Embree's
+  own CMake. Both are upstream gaps.
 
 ## Usage notes