Set vectorization

howetuft · Jan 4, 2025 · dbdabc3 · dbdabc3
1 parent 5623593
commit dbdabc3
Show file tree

Hide file tree

Showing 7 changed files with 23 additions and 8 deletions.
diff --git a/.github/workflows/wheels.yml b/.github/workflows/wheels.yml
@@ -28,6 +28,8 @@ jobs:
         #os: [ubuntu-latest]
         #python-minor: [ 11]
         #os: [windows-latest]
+        #python-minor: [ 11]
+        #os: [macos-13]
         exclude:
           - os: macos-14
             python-minor: 8

diff --git a/conan-profiles/conan-profile-Linux-X64 b/conan-profiles/conan-profile-Linux-X64
@@ -8,9 +8,11 @@ compiler.version=12
 compiler.cppstd={{ os.getenv("CXX_VERSION") }}
 compiler.libcxx=libstdc++11
 
-[conf]
-openvdb/*:tools.cmake.cmaketoolchain:extra_variables={ 'USE_CCACHE':'OFF' }
-
 [options]
 openvdb/*:simd=AVX
-embree3/*:avx512=True
+embree3/*:avx=True
+
+[conf]
+tools.build:cflags+=["-ftree-vectorize", "-march=x86-64-v3", "-O3"]
+tools.build:cxxflags+=["-ftree-vectorize", "-march=x86-64-v3", "-O3"]
+openvdb/*:tools.cmake.cmaketoolchain:extra_variables={ 'USE_CCACHE':'OFF' }
diff --git a/conan-profiles/conan-profile-Windows-X64 b/conan-profiles/conan-profile-Windows-X64
@@ -10,7 +10,8 @@ compiler.runtime=dynamic
 
 [conf]
 tools.env.virtualenv:powershell=False
-tools.build:cxxflags=["/openmp", "/MP", "/DEBUG:NONE"]
+tools.build:cxxflags=["/openmp", "/MP", "/DEBUG:NONE", "/arch:AVX", "/O2"]
+embree3/*:tools.build:cxxflags=["/openmp", "/MP", "/DEBUG:NONE", "/O2"]
 
 [options]
 openvdb/*:simd=AVX

diff --git a/conan-profiles/conan-profile-common b/conan-profiles/conan-profile-common
@@ -16,3 +16,4 @@ yasm/*
 
 [options]
 oidn/*:shared=True
+embree3/*:with_tbb=False
diff --git a/conan-profiles/conan-profile-macOS-ARM64 b/conan-profiles/conan-profile-macOS-ARM64
@@ -37,3 +37,4 @@ embree3/*:neon2x=True
 [buildenv]
 CXX=clang++
 CC=clang
+&:tools.build:cxxflags=["-O3"]
diff --git a/conan-profiles/conan-profile-macOS-X64 b/conan-profiles/conan-profile-macOS-X64
@@ -13,9 +13,12 @@ compiler.libcxx=libc++
 minizip-ng/*:with_zlib=True
 minizip-ng/*:with_libcomp=False
 openvdb/*:simd=AVX
-embree3/*:avx512=True
+embree3/*:avx=True
 embree3/*:shared=True
 
 [buildenv]
 CXX=clang++
 CC=clang
+
+[conf]
+&:tools.build:cxxflags+=["-ftree-vectorize", "-march=x86-64-v3", "-O3", "-mno-avx2"]
diff --git a/doc/notes.md b/doc/notes.md
@@ -91,7 +91,12 @@ author).
 tried to have it minimal and it's 10.13. For the delocation to succeed, we also
 have to set `MACOSX_DEPLOYMENT_TARGET` environment variable to 10.13.
 
+## Vectorization
+For Intel, we stick to x86-64-v3 (Haswell, 2013), no more. For all platform
+except MacOS Intel, we rely on "-O" option.
+For MacOS Intel: Github runner is not standard, we have to deactivate avx2
+extensions. Otherwise `pyluxcore.so` cannot be tested ("Illegal instruction").
+See https://github.com/ggerganov/whisper.cpp/issues/358 and `sysctl -a
+machdep.cpu`
 
 
-we 
-
Original file line number	Diff line number	Diff line change
Expand Up		@@ -16,3 +16,4 @@ yasm/*

		[options]
		oidn/*:shared=True
		embree3/*:with_tbb=False