split avx from avx2

- refactor all simd functions, just like FP32 - add openblas for ubuntu-aarch deps Signed-off-by: LHT129 <tianlan.lht@antgroup.com>
antgroup · Dec 24, 2024 · 59a72c4 · 59a72c4
1 parent 2698dcb
commit 59a72c4
Show file tree

Hide file tree

Showing 32 changed files with 1,193 additions and 1,106 deletions.
diff --git a/scripts/deps/install_deps_ubuntu.sh b/scripts/deps/install_deps_ubuntu.sh
@@ -5,7 +5,7 @@ if [[ "$arch" == "x86_64" ]]; then
     apt update && apt install -y gfortran python3-dev libomp-15-dev lcov intel-mkl gcc make cmake g++
 elif [[ "$arch" == "aarch64" ]]; then
     echo "Executing apt install for aarch64"
-    apt update && apt install -y gfortran python3-dev libomp-15-dev gcc make cmake g++ lcov
+    apt update && apt install -y gfortran python3-dev libomp-15-dev gcc make cmake g++ lcov libopenblas-dev
 else
     echo "Unknown architecture: $arch"
 fi
diff --git a/src/simd/CMakeLists.txt b/src/simd/CMakeLists.txt
@@ -1,63 +1,64 @@
-set (SIMD_SRCS
+set(SIMD_SRCS
         generic.cpp
+        sse.cpp
+        avx.cpp
+        avx2.cpp
+        avx512.cpp
         simd.cpp
+        basic_func.cpp
         fp32_simd.cpp
         sq8_simd.cpp
         sq4_simd.cpp
         sq4_uniform_simd.cpp
         sq8_uniform_simd.cpp
-        sse.cpp
-        avx.cpp
-        avx512.cpp
         normalize.cpp
 )
 if (DIST_CONTAINS_SSE)
-    set (SIMD_SRCS ${SIMD_SRCS} sse.cpp)
-    set_source_files_properties (
+    set(SIMD_SRCS ${SIMD_SRCS} sse.cpp)
+    set_source_files_properties(
             sse.cpp
             PROPERTIES COMPILE_FLAGS
             "-msse -msse2 -msse3 -mssse3 -msse4 -msse4a -msse4.1 -msse4.2")
 endif ()
 if (DIST_CONTAINS_AVX)
-    set (SIMD_SRCS ${SIMD_SRCS} avx.cpp)
-    set_source_files_properties (avx.cpp PROPERTIES COMPILE_FLAGS "-mavx")
+    set(SIMD_SRCS ${SIMD_SRCS} avx.cpp)
+    set_source_files_properties(avx.cpp PROPERTIES COMPILE_FLAGS "-mavx")
+endif ()
+if (DIST_CONTAINS_AVX2)
+    set_source_files_properties(avx2.cpp PROPERTIES COMPILE_FLAGS "-mavx2 -mfma")
 endif ()
-# FIXME(LHT): cause illegal instruction on platform which has avx only
-#if (DIST_CONTAINS_AVX2)
-#    set_source_files_properties (avx.cpp PROPERTIES COMPILE_FLAGS "-mavx2 -mfma")
-#endif ()
 if (DIST_CONTAINS_AVX512)
-    set (SIMD_SRCS ${SIMD_SRCS} avx512.cpp
+    set(SIMD_SRCS ${SIMD_SRCS} avx512.cpp
             normalize.cpp)
-    set_source_files_properties (
+    set_source_files_properties(
             avx512.cpp
             PROPERTIES
             COMPILE_FLAGS
             "-mavx512f -mavx512pf -mavx512er -mavx512cd -mavx512vl -mavx512bw -mavx512dq -mavx512ifma -mavx512vbmi"
     )
 endif ()
 
-set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
-set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -ftree-vectorize")
-set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fno-builtin-malloc")
-set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fno-builtin-calloc")
-set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fno-builtin-realloc")
-set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fno-builtin-free")
-set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fopenmp -fopenmp-simd")
-set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -funroll-loops")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -ftree-vectorize")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fno-builtin-malloc")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fno-builtin-calloc")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fno-builtin-realloc")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fno-builtin-free")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fopenmp -fopenmp-simd")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -funroll-loops")
 
-add_library (simd ${SIMD_SRCS})
+add_library(simd ${SIMD_SRCS})
 
-macro (simd_add_definitions flag1 flag2)
+macro(simd_add_definitions flag1 flag2)
     if (${flag1})
-        target_compile_definitions (simd PUBLIC ${flag2})
+        target_compile_definitions(simd PUBLIC ${flag2})
     endif ()
-endmacro ()
+endmacro()
 
-simd_add_definitions (DIST_CONTAINS_SSE -DENABLE_SSE=1)
-simd_add_definitions (DIST_CONTAINS_AVX -DENABLE_AVX=1)
-#simd_add_definitions (DIST_CONTAINS_AVX2 -DENABLE_AVX2=1)
-simd_add_definitions (DIST_CONTAINS_AVX512 -DENABLE_AVX512=1)
+simd_add_definitions(DIST_CONTAINS_SSE -DENABLE_SSE=1)
+simd_add_definitions(DIST_CONTAINS_AVX -DENABLE_AVX=1)
+simd_add_definitions(DIST_CONTAINS_AVX2 -DENABLE_AVX2=1)
+simd_add_definitions(DIST_CONTAINS_AVX512 -DENABLE_AVX512=1)
 
-target_link_libraries (simd PRIVATE cpuinfo)
-install (TARGETS simd ARCHIVE DESTINATION lib)
+target_link_libraries(simd PRIVATE cpuinfo)
+install(TARGETS simd ARCHIVE DESTINATION lib)