Merge pull request #7 from MarioSieg/develop

Runtime CPU detection and blas specializations.
MarioSieg · Jan 24, 2025 · 552ebe5 · 552ebe5
2 parents eaf8b70 + 8a5adb5
commit 552ebe5
Show file tree

Hide file tree

Showing 31 changed files with 760 additions and 523 deletions.
diff --git a/.github/workflows/cmake-multi-platform.yml b/.github/workflows/cmake-multi-platform.yml
@@ -13,7 +13,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        os: [ubuntu-latest, windows-latest, macos-latest]
+        os: [ubuntu-latest, windows-latest, macos-latest, ubuntu-24.04-arm]
         build_type: [Release]
         c_compiler: [gcc, clang, cl]
         include:
@@ -26,6 +26,12 @@ jobs:
           - os: ubuntu-latest
             c_compiler: clang
             cpp_compiler: clang++
+          - os: ubuntu-24.04-arm
+            c_compiler: gcc
+            cpp_compiler: g++
+          - os: ubuntu-24.04-arm
+            c_compiler: clang
+            cpp_compiler: clang++
           - os: macos-latest
             c_compiler: clang
             cpp_compiler: clang++
@@ -36,6 +42,8 @@ jobs:
             c_compiler: clang
           - os: ubuntu-latest
             c_compiler: cl
+          - os: ubuntu-24.04-arm
+            c_compiler: cl
           - os: macos-latest
             c_compiler: gcc
           - os: macos-latest
@@ -103,4 +111,4 @@ jobs:
         shell: bash
         run: |
           . .venv/bin/activate
-          python -m pytest ${{ github.workspace }}/python/tests/tests.py
+          python -m pytest ${{ github.workspace }}/python/tests/*
diff --git a/.gitignore b/.gitignore
@@ -33,3 +33,4 @@ out/
 .tmp
 CMakeSettings.json
 magnetron_chat/.idea
+/build/
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -35,7 +35,7 @@ endif()
 
 include(cmake/arch.cmake)
 include(cmake/lib.cmake)
-include(cmake/comflags.cmake)
+include(cmake/compiler_config.cmake)
 
 if (${MAGNETRON_ENABLE_CUDA})
     include(cmake/cuda.cmake)

diff --git a/cmake/blas_tune.cmake b/cmake/blas_tune.cmake
@@ -10,25 +10,28 @@ function(set_blas_spec_arch filename posix_arch msvc_arch)
 endfunction()
 
 set(MAGNETRON_BLAS_SPEC_AMD64_SOURCES
-    magnetron/magnetron_cpu_blas_amd64_sse42.c
-    magnetron/magnetron_cpu_blas_amd64_avx.c
-    magnetron/magnetron_cpu_blas_amd64_avx2.c
-    magnetron/magnetron_cpu_blas_amd64_avx512f.c
-    magnetron/magnetron_cpu_blas_amd64_znver4.c
+    magnetron/magnetron_cpu_blas_amd64_v2.c
+    magnetron/magnetron_cpu_blas_amd64_v2_5.c
+    magnetron/magnetron_cpu_blas_amd64_v3.c
+    magnetron/magnetron_cpu_blas_amd64_v4.c
+    magnetron/magnetron_cpu_blas_amd64_v4_5.c
 )
 
 set(MAGNETRON_BLAS_SPEC_ARM64_SOURCES
-    magnetron/magnetron_cpu_blas_arm64_82.c
+    magnetron/magnetron_cpu_blas_arm64_v8_2.c
+    magnetron/magnetron_cpu_blas_arm64_v9.c
 )
 
 if (${IS_AMD64}) # x86-64 specific compilation options
     set(MAGNETRON_SOURCES ${MAGNETRON_SOURCES} ${MAGNETRON_BLAS_SPEC_AMD64_SOURCES})
-    set_blas_spec_arch("magnetron_cpu_blas_amd64_sse42.c" "-mtune=nehalem -msse4.2"  "/arch:SSE4.2")
-    set_blas_spec_arch("magnetron_cpu_blas_amd64_avx.c" "-mtune=sandybridge -mavx"  "/arch:AVX")
-    set_blas_spec_arch("magnetron_cpu_blas_amd64_avx2.c" "-mtune=skylake -mavx -mavx2 -mfma -mf16c"  "/arch:AVX2")
-    set_blas_spec_arch("magnetron_cpu_blas_amd64_avx512f.c" "-mtune=cannonlake -mavx -mavx2 -mfma -mf16c -mavx512f" "/arch:AVX512")
-    set_blas_spec_arch("magnetron_cpu_blas_amd64_znver4.c" "-mavx -mavx2 -mfma -mf16c -mavx512f -mavx512vl -mavx512vnni -mavx512bf16 -mavx512bw -mavx512dq" "/arch:AVX512")
+    set_blas_spec_arch("magnetron_cpu_blas_amd64_v2.c" "-mtune=nehalem -mcx16 -mpopcnt -msse3 -mssse3 -msse4.1 -msse4.2"  "/arch:SSE4.2")
+    set_blas_spec_arch("magnetron_cpu_blas_amd64_v2_5.c" "-mtune=ivybridge -mavx -mno-avx2 -mcx16 -mpopcnt -msse3 -mssse3 -msse4.1 -msse4.2"  "/arch:AVX")
+    set_blas_spec_arch("magnetron_cpu_blas_amd64_v3.c" "-mtune=haswell -mavx -mavx2 -mbmi -mbmi2 -mf16c -mfma -mlzcnt -mmovbe"
+        "/arch:AVX2 /D__BMI__=1 /D__BMI2__=1 /D__F16C__=1  /D__FMA__=1") # MSVC is just annoying
+    set_blas_spec_arch("magnetron_cpu_blas_amd64_v4.c" "-mtune=cannonlake -mavx512f -mavx512bw -mavx512vl -mavx512dq -mavx -mavx2 -mbmi -mbmi2 -mf16c -mfma -mlzcnt -mmovbe"  "/arch:AVX512")
+    set_blas_spec_arch("magnetron_cpu_blas_amd64_v4_5.c" "-mtune=generic -mavx512f -mavx512bw -mavx512vl -mavx512dq -mavx512vnni -mavx512bf16 -mavx -mavx2 -mbmi -mbmi2 -mf16c -mfma -mlzcnt -mmovbe" "/arch:AVX512")
  elseif(${IS_ARM64})
     set(MAGNETRON_SOURCES ${MAGNETRON_SOURCES} ${MAGNETRON_BLAS_SPEC_ARM64_SOURCES})
-    set_blas_spec_arch("magnetron_cpu_blas_arm64_82.c" "-march=armv8.2-a+dotprod+fp16" "")
+    set_blas_spec_arch("magnetron_cpu_blas_arm64_v8_2.c" "-march=armv8.2-a+dotprod+fp16" "")
+    set_blas_spec_arch("magnetron_cpu_blas_arm64_v9.c" "-march=armv9-a+sve+sve2" "")
  endif()
diff --git a/cmake/comflags.cmake → cmake/compiler_config.cmake b/cmake/comflags.cmake → cmake/compiler_config.cmake
@@ -27,6 +27,7 @@ set(MAG_CLANG_COMPILE_FLAGS
 set(MAG_CLANG_RELEASE_COMPILE_FLAGS
     -O3
     -flto
+    -fomit-frame-pointer
 )
 set(MAG_CLANG_LINK_OPTIONS "")
 set(MAG_CLANG_RELEASE_LINK_OPTIONS -flto)
@@ -50,6 +51,14 @@ set(MAG_GCC_RELEASE_COMPILE_FLAGS
 set(MAG_GCC_LINK_OPTIONS "")
 set(MAG_GCC_RELEASE_LINK_OPTIONS -flto)
 
+if (${IS_ARM64})
+    set(MAG_CLANG_COMPILE_FLAGS ${MAG_CLANG_COMPILE_FLAGS} -march=armv8-a)
+    set(MAG_GCC_COMPILE_FLAGS ${MAG_CLANG_COMPILE_FLAGS} -march=armv8-a)
+elseif (${IS_AMD64})
+    set(MAG_CLANG_COMPILE_FLAGS ${MAG_CLANG_COMPILE_FLAGS} -msse -msse2)
+    set(MAG_GCC_COMPILE_FLAGS ${MAG_CLANG_COMPILE_FLAGS} -msse -msse2)
+endif()
+
 if (WIN32) # Windows (MSVC) specific config
     target_compile_options(magnetron PRIVATE ${MAG_MSVC_COMPILE_FLAGS})
     target_link_options(magnetron PRIVATE ${MAG_MSVC_LINK_OPTIONS})
-Original file line number
+Diff line change
@@ Expand Up / @@ -33,3 +33,4 @@ out/ @@
     .tmp
     CMakeSettings.json
     magnetron_chat/.idea
+    /build/