Add support for AVX-VNNI and update SIMD implementation

Author: bab2min

CMakeLists.txt

@@ -38,6 +38,13 @@ if(NOT KIWI_CPU_ARCH)
   set(KIWI_CPU_ARCH "${KIWI_CPU_ARCH}" PARENT_SCOPE)
 endif()
 
+set( AVX_VNNI_SUPPORTED (KIWI_USE_CPUINFO AND 
+  (MSVC OR 
+   (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 11) OR 
+   (CMAKE_CXX_COMPILER_ID STREQUAL "Clang" AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 11)
+  )
+))
+
 if(APPLE)
   set(CMAKE_OSX_ARCHITECTURES "${KIWI_CPU_ARCH}")
 endif()
@@ -120,6 +127,11 @@ if(KIWI_USE_CPUINFO)
   )
 endif()
 
+if (AVX_VNNI_SUPPORTED)
+  message(STATUS "AVX-VNNI is supported")
+  set ( ADDITIONAL_FLAGS ${ADDITIONAL_FLAGS} "-DKIWI_AVX_VNNI_SUPPORTED" )
+endif()
+
 if(MSVC)
   set ( CMAKE_C_FLAGS_DEBUG "-DDEBUG -DC_FLAGS -Zi -Od /utf-8 /bigobj" )
   set ( CMAKE_CXX_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG}" )
@@ -161,7 +173,6 @@ if (KIWI_CPU_ARCH MATCHES "x86_64")
       src/archImpl/avx512vnni.cpp
     )
     # If AVX-VNNI is supported (MSVC, GCC 11+ or Clang 11+)
-    set ( AVX_VNNI_SUPPORTED (MSVC OR (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 11) OR (CMAKE_CXX_COMPILER_ID STREQUAL "Clang" AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 11)))
     if (AVX_VNNI_SUPPORTED)
       set( CORE_SRCS 
         ${CORE_SRCS}

src/ArchAvailable.h

@@ -14,7 +14,9 @@ namespace kiwi
 #if CPUINFO_ARCH_X86_64
 		static_cast<std::ptrdiff_t>(ArchType::avx512vnni),
 		static_cast<std::ptrdiff_t>(ArchType::avx512bw),
+#ifdef KIWI_AVX_VNNI_SUPPORTED
 		static_cast<std::ptrdiff_t>(ArchType::avx_vnni),
+#endif
 		static_cast<std::ptrdiff_t>(ArchType::avx2),
 		static_cast<std::ptrdiff_t>(ArchType::sse4_1),
 #endif
@@ -28,7 +30,9 @@ namespace kiwi
 #ifdef KIWI_ARCH_X86_64
 		static_cast<std::ptrdiff_t>(ArchType::avx512vnni),
 		static_cast<std::ptrdiff_t>(ArchType::avx512bw),
+#ifdef KIWI_AVX_VNNI_SUPPORTED
 		static_cast<std::ptrdiff_t>(ArchType::avx_vnni),
+#endif
 		static_cast<std::ptrdiff_t>(ArchType::avx2),
 		static_cast<std::ptrdiff_t>(ArchType::sse4_1),
 #endif
@@ -48,7 +52,9 @@ namespace kiwi
 #if CPUINFO_ARCH_X86_64
 		static_cast<std::ptrdiff_t>(ArchType::avx512vnni),
 		static_cast<std::ptrdiff_t>(ArchType::avx512bw),
+#ifdef KIWI_AVX_VNNI_SUPPORTED
 		static_cast<std::ptrdiff_t>(ArchType::avx_vnni),
+#endif
 		static_cast<std::ptrdiff_t>(ArchType::avx2),
 		static_cast<std::ptrdiff_t>(ArchType::sse4_1)
 #endif
@@ -59,7 +65,9 @@ namespace kiwi
 #ifdef KIWI_ARCH_X86_64
 		static_cast<std::ptrdiff_t>(ArchType::avx512vnni),
 		static_cast<std::ptrdiff_t>(ArchType::avx512bw),
+#ifdef KIWI_AVX_VNNI_SUPPORTED
 		static_cast<std::ptrdiff_t>(ArchType::avx_vnni),
+#endif
 		static_cast<std::ptrdiff_t>(ArchType::avx2),
 		static_cast<std::ptrdiff_t>(ArchType::sse4_1)
 #endif

src/SIMD.hpp

@@ -600,7 +600,7 @@ namespace kiwi
 				// reduce sum of eight int32_t to one int32_t
 				__m256i sum = _mm256_hadd_epi32(acc, acc);
 				sum = _mm256_hadd_epi32(sum, sum);
-                return _mm256_cvtsi256_si32(sum) + _mm256_extract_epi32(sum, 4);
+                return _mm_cvtsi128_si32(_mm256_castsi256_si128(sum)) + _mm256_extract_epi32(sum, 4);
             }
 		};