End cheacj SSE and AVX + Intel options on MAcOS X

CMakeLists.txt

@@ -59,6 +59,7 @@ OPTION( ScalFMM_USE_ASSERT           "Set to ON to enable safe tests during exec
 OPTION( ScalFMM_USE_MIC_NATIVE       "Set to ON to compile in native mode for MIC" OFF  )
 # Set scalfmm to default libraries
 SET(SCALFMM_LIBRARIES "")
+SET(ScaLFMM_CXX_FLAGS "")
 #
 # Compile option
 #
@@ -68,9 +69,12 @@ SET(ScaLFMM_CXX_FLAGS "-std=c++11 -Wall -Wshadow -Wpointer-arith -Wcast-qual -Wc
 #
 find_package (OpenMP) 
 if(OPENMP_FOUND)
-	set(CMAKE_C_FLAGS            "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}")
-	set(CMAKE_CXX_FLAGS          "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
-	set(CMAKE_EXE_LINKER_FLAGS   "${CMAKE_EXE_LINKER_FLAGS} ${OpenMP_EXE_LINKER_FLAGS}")
+	set(CMAKE_C_FLAGS                  "${CMAKE_C_FLAGS}     ${OpenMP_C_FLAGS}")
+	set(CMAKE_CXX_FLAGS              "${CMAKE_CXX_FLAGS}  ${OpenMP_CXX_FLAGS}")
+	set(CMAKE_EXE_LINKER_FLAGS   "${CMAKE_EXE_LINKER_FLAGS} ${OpenMP_EXE_LINKER_FLAGS}")	
+# 	 SET(ScaLFMM_CXX_FLAGS  "${ScaLFMM_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
+	else()
+		MESSAGE(WARNING "OPENMP NOT FOUND")
 endif(OPENMP_FOUND)
 MESSAGE(STATUS "ScalFMM_BUILD_DEBUG          = ${ScalFMM_BUILD_DEBUG}" )
 if(CMAKE_SIZEOF_VOID_P EQUAL 8)
@@ -186,6 +190,7 @@ IF(CMAKE_CXX_COMPILER_ID STREQUAL "Intel")
     if( ${NAME} GREATER  0)
       if(  ${INTEL_VERSION} EQUAL 15. OR  ${INTEL_VERSION} GREATER 15.0.0) 
         MESSAGE( "  Intel compiler is icl++ ( version >= 15.0.0)")	
+        SET(INTEL_ICL_COMPILER "ON")
       else()
         MESSAGE(FATAL_ERROR  "  Intel compiler should be icl++ ( version >= 15.0.0)")  
       endif()
@@ -212,16 +217,29 @@ ENDIF()
 
 MESSAGE( STATUS "ScalFMM_USE_SSE              = ${ScalFMM_USE_SSE}" )
 if( ScalFMM_USE_SSE )
-	TRY_COMPILE(COMPILE_RESULT_VAR  ${CMAKE_CURRENT_BINARY_DIR} 
- 					     ${CMAKE_MODULE_PATH}/checkSSEpe.cpp )
-  	IF( NOT ${COMPILE_RESULT_VAR})
-    	  SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D__SSEPE_INTEL_COMPILER")
-  	ENDIF()
-  	IF(CMAKE_CXX_COMPILER_ID STREQUAL "Intel")
-		    SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse -msse2 -msse3 -mfpmath=sse")
-		    MESSAGE(WARNING "INTEL CXX_FLAG__SSE ${CMAKE_CXX_FLAGS} ")
+	TRY_COMPILE(COMPILE_SSE  ${CMAKE_CURRENT_BINARY_DIR} 
+ 					     ${CMAKE_MODULE_PATH}/compileTestSse.cpp )
+   	IF(${COMPILE_SSE})			     
+		TRY_COMPILE(COMPILE_RESULT_VAR  ${CMAKE_CURRENT_BINARY_DIR} 
+ 						     ${CMAKE_MODULE_PATH}/checkSSEpe.cpp )
+  		IF( NOT ${COMPILE_RESULT_VAR})
+    	 		 SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D__SSEPE_INTEL_COMPILER")
+  		ENDIF()
+  		IF(CMAKE_CXX_COMPILER_ID STREQUAL "Intel")
+  			if(INTEL_ICL_COMPILER) 
+  #				 SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}  -xHost -mfpmath=sse")
+  				 SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}  -msse4.2 ")
+ 				MESSAGE(WARNING "Need to check for SSE2 and SSE3 ")
+  			else()
+			    SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse4.2")
+			ENDIF()
+		    	MESSAGE(WARNING "INTEL  SSE CXX FLAGS: ${CMAKE_CXX_FLAGS} ")
+  		ELSE()
+  		  	SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse -msse2 -msse3 -mfpmath=sse")
+  		ENDIF()
   	ELSE()
-  	  	SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse -msse2 -msse3 -mfpmath=sse")
+  		 MESSAGE(FATAL_ERROR "SSE NOT SUPPORTED ; Set ScalFMM_USE_SSE  to OFF")
+  		
   	ENDIF()
 endif()
 ##################################################################
@@ -230,19 +248,24 @@ endif()
 
 MESSAGE(STATUS "ScalFMM_USE_AVX               = ${ScalFMM_USE_AVX}")
 if(ScalFMM_USE_AVX)
-	TRY_COMPILE(COMPILE_RESULT_VAR  ${CMAKE_CURRENT_BINARY_DIR} 
- 					     ${CMAKE_MODULE_PATH}/checkAVXpe.cpp )
-	  	IF( NOT ${COMPILE_RESULT_VAR})
- 		   	  SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D__AVXPE_INTEL_COMPILER")
+	TRY_COMPILE(COMPILE_AVX  ${CMAKE_CURRENT_BINARY_DIR} 
+ 					     ${CMAKE_MODULE_PATH}/compileTestAvx.cpp )
+   	IF(${COMPILE_AVX})
+   	 		TRY_COMPILE(COMPILE_RESULT_AVSPE  ${CMAKE_CURRENT_BINARY_DIR} 
+ 						     ${CMAKE_MODULE_PATH}/checkAVXpe.cpp )
+	  		IF( NOT ${COMPILE_RESULT_AVSPE})
+ 		   		  SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D__AVXPE_INTEL_COMPILER")
+  			ENDIF()
+  		IF(CMAKE_CXX_COMPILER_ID STREQUAL "Intel")
+    			SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}  -mavx")
+		    MESSAGE(WARNING "INTEL AVX CXX FLAGS:  ${CMAKE_CXX_FLAGS} ")
+  		ELSE()
+    			SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse -msse2 -msse3 -mfpmath=sse -mavx")
   		ENDIF()
-  	IF(CMAKE_CXX_COMPILER_ID STREQUAL "Intel")
-    SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse -msse2 -msse3 -mfpmath=sse -mavx")
-		    MESSAGE(WARNING "INTEL CXX_FLAG__SSE ${CMAKE_CXX_FLAGS} ")
-
-  ELSE()
-    SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse -msse2 -msse3 -mfpmath=sse -mavx")
-  ENDIF()
-MESSAGE(STATUS ${CMAKE_CXX_FLAGS} )
+		MESSAGE(STATUS ${CMAKE_CXX_FLAGS} )
+  	ELSE()
+  		 MESSAGE(FATAL_ERROR "AVS NOT SUPPORTED ; Set ScalFMM_USE_AVX  to OFF")
+  	ENDIF()
 endif()
 
 

Modules/checkAVXpe.cpp

 
-#include "../src/Container/instrset.h"
+#include "immintrin.h"
 
 
 int main() {
-  int iset = instrset_detect();
-  bool hasAVX = iset >= 7;
-  printf("Has AVX? %s (iset = %d)\n", hasAVX ? "Yes!" : "No.", iset);
-  return iset;
+	__m256d tx, ty ;
+	tx += ty ;
+  return 0;
 }

Modules/compileTestAvx.cpp

+
+#include <x86intrin.h>
+#include <xmmintrin.h> // SSE
+#include <emmintrin.h> // SSE2
+#include <pmmintrin.h> // SSE3
+#include <tmmintrin.h> // SSSE3
+#include <smmintrin.h> // SSE4
+
+#include <immintrin.h> // AVX
+
+int main(){
+	{
+		__m256d res0d, res1d;
+		res0d = _mm256_hadd_pd(res0d, res1d);
+
+		__m256 res0, res1;
+		res0 = _mm256_hadd_ps(res0, res1);
+	}
+	{
+		__m128d res0d, res1d;
+		res0d = _mm_hadd_pd(res0d, res1d);
+
+		__m128 res0, res1;
+		res0 = _mm_hadd_ps(res0, res1);
+	}
+	return 0;
+}

Modules/compileTestSse.cpp

+
+#include <x86intrin.h>
+#include <xmmintrin.h> // SSE
+#include <emmintrin.h> // SSE2
+#include <pmmintrin.h> // SSE3
+#ifdef __SSSE3__
+#include <tmmintrin.h>  //SSSE3
+#endif
+#ifdef __SSSE4_1__
+#include <smmintrin.h> // SSE4
+#endif
+int main(){
+	__m128d res0d, res1d;
+	res0d = _mm_hadd_pd(res0d, res1d);
+
+	__m128 res0, res1;
+	res0 = _mm_hadd_ps(res0, res1);
+
+	return 0;
+}