From 570a4b05aef2f95471b46ed56267d7c674ffcef2 Mon Sep 17 00:00:00 2001
From: Jeremy Martinon <jeremy.martinon@gmail.com>
Date: Fri, 16 May 2025 16:32:05 -0400
Subject: [PATCH] CMake: support Apple ARM + GCC 14; fix typo ARM micro kernel

---
 CMakeLists.txt    | 44 +++++++++++++++++++++++++++++++++++++-------
 src/transpose.cpp |  2 +-
 2 files changed, 38 insertions(+), 8 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 582ada3..f489c48 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,5 +1,5 @@
 
-cmake_minimum_required(VERSION 3.7 FATAL_ERROR)
+cmake_minimum_required(VERSION 3.14 FATAL_ERROR)
 project (HPTT C CXX)
 
 set(CMAKE_CXX_STANDARD 11)
@@ -16,6 +16,8 @@ if(CMAKE_CXX_COMPILER_ID STREQUAL "Intel")
 elseif(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
   if(ENABLE_IBM)
     set(HPTT_CXX_FLAGS ${HPTT_CXX_FLAGS} -fopenmp)
+  elseif(APPLE)
+	  set(HPTT_CXX_FLAGS ${HPTT_CXX_FLAGS} -fopenmp -mtune=native)
   else()
     set(HPTT_CXX_FLAGS ${HPTT_CXX_FLAGS} -fopenmp -march=native -mtune=native)
   endif()
@@ -35,18 +37,46 @@ elseif(ENABLE_IBM)
   set(HPTT_CXX_FLAGS ${HPTT_CXX_FLAGS} -mtune=native -DHPTT_ARCH_IBM -maltivec -mabi=altivec)
 endif()
 
+if(APPLE)
+    if(CMAKE_SYSTEM_PROCESSOR MATCHES "arm64")
+        message(STATUS "Running on Apple Silicon (M-series)")
+		set(HPTT_CXX_FLAGS ${HPTT_CXX_FLAGS} -O3 -ffast-math -funroll-loops -ftree-vectorize
+	 -mcpu=native -DHPTT_ARCH_ARM)
+    elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64")
+        message(STATUS "Running on Intel-based macOS")
+		set(HPTT_CXX_FLAGS ${HPTT_CXX_FLAGS} -O3 -ffast-math -funroll-loops -ftree-vectorize -mavx -DHPTT_ARCH_AVX)
+    endif()
+endif()
+
 set(HPTT_SRCS src/hptt.cpp src/plan.cpp src/transpose.cpp src/utils.cpp)
 
-add_library(hptt STATIC ${HPTT_SRCS})
-target_compile_features(hptt PUBLIC cxx_std_11)
-target_include_directories(hptt PUBLIC ${PROJECT_SOURCE_DIR}/include)
-#target_compile_definitions(hptt PRIVATE ${HPTT_CXX_COMPILE_DEFS})
-target_compile_options(hptt PUBLIC ${HPTT_CXX_FLAGS})
+add_library(hptt_static STATIC ${HPTT_SRCS})
+target_compile_features(hptt_static PUBLIC cxx_std_11)
+set_target_properties(hptt_static PROPERTIES OUTPUT_NAME hptt)
+target_include_directories(hptt_static PUBLIC ${PROJECT_SOURCE_DIR}/include)
+target_compile_options(hptt_static PUBLIC ${HPTT_CXX_FLAGS})
+
+add_library(hptt_dyn SHARED ${HPTT_SRCS})
+target_compile_features(hptt_dyn PUBLIC cxx_std_11)
+set_target_properties(hptt_dyn PROPERTIES OUTPUT_NAME hptt)
+target_include_directories(hptt_dyn PUBLIC ${PROJECT_SOURCE_DIR}/include)
+target_compile_options(hptt_dyn PUBLIC ${HPTT_CXX_FLAGS})
+
+
+find_package(OpenMP REQUIRED)
+target_link_libraries(hptt_static PUBLIC OpenMP::OpenMP_CXX)
+target_link_libraries(hptt_dyn PUBLIC OpenMP::OpenMP_CXX)
 
-install(TARGETS hptt
+
+install(TARGETS hptt_static
         LIBRARY DESTINATION ${CMAKE_INSTALL_PREFIX}/lib
         ARCHIVE DESTINATION ${CMAKE_INSTALL_PREFIX}/lib)
 
+install(TARGETS hptt_dyn
+        LIBRARY DESTINATION ${CMAKE_INSTALL_PREFIX}/lib
+        ARCHIVE DESTINATION ${CMAKE_INSTALL_PREFIX}/lib)
+
+
 set(HPTT_INCLUDES 
     include/compute_node.h 
     include/hptt_types.h 
diff --git a/src/transpose.cpp b/src/transpose.cpp
index f77cd5b..a5de997 100644
--- a/src/transpose.cpp
+++ b/src/transpose.cpp
@@ -244,7 +244,7 @@ static INLINE void prefetch(const floatType* A, const int lda) { }
 #include <arm_neon.h>
 
 template <int betaIsZero, bool conjA>
-struct micro_kernel<float, betaIsZero>
+struct micro_kernel<float, betaIsZero, conjA>
 {
     static void execute(const float* __restrict__ A, const size_t lda, float* __restrict__ B, const size_t ldb, const float alpha ,const float beta)
     {