add SSE4 support, start with integer multiplication

CMakeLists.txt

@@ -63,31 +63,43 @@ if(CMAKE_COMPILER_IS_GNUCXX)
 
     if(NOT EIGEN_TEST_LIB)
       set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pedantic")
-    endif(NOT EIGEN_TEST_LIB)
+    endif()
 
     option(EIGEN_TEST_SSE2 "Enable/Disable SSE2 in tests/examples" OFF)
     if(EIGEN_TEST_SSE2)
       set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse2")
       message("Enabling SSE2 in tests/examples")
-    endif(EIGEN_TEST_SSE2)
+    endif()
 
     option(EIGEN_TEST_SSE3 "Enable/Disable SSE3 in tests/examples" OFF)
     if(EIGEN_TEST_SSE3)
       set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse3")
       message("Enabling SSE3 in tests/examples")
-    endif(EIGEN_TEST_SSE3)
+    endif()
 
     option(EIGEN_TEST_SSSE3 "Enable/Disable SSSE3 in tests/examples" OFF)
     if(EIGEN_TEST_SSSE3)
       set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mssse3")
       message("Enabling SSSE3 in tests/examples")
-    endif(EIGEN_TEST_SSSE3)
+    endif()
+
+    option(EIGEN_TEST_SSE4_1 "Enable/Disable SSE4.1 in tests/examples" OFF)
+    if(EIGEN_TEST_SSE4_1)
+      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse4.1")
+      message("Enabling SSE4.1 in tests/examples")
+    endif()
+
+    option(EIGEN_TEST_SSE4_2 "Enable/Disable SSE4.2 in tests/examples" OFF)
+    if(EIGEN_TEST_SSE4_2)
+      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse4.2")
+      message("Enabling SSE4.2 in tests/examples")
+    endif()
 
     option(EIGEN_TEST_ALTIVEC "Enable/Disable altivec in tests/examples" OFF)
     if(EIGEN_TEST_ALTIVEC)
       set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -maltivec -mabi=altivec")
       message("Enabling AltiVec in tests/examples")
-    endif(EIGEN_TEST_ALTIVEC)
+    endif()
 
   endif(CMAKE_SYSTEM_NAME MATCHES Linux)
 endif(CMAKE_COMPILER_IS_GNUCXX)

Eigen/Core

@@ -67,6 +67,12 @@
     #ifdef __SSSE3__
       #include <tmmintrin.h>
     #endif
+    #ifdef __SSE4_1__
+      #include <smmintrin.h>
+    #endif
+    #ifdef __SSE4_2__
+      #include <nmmintrin.h>
+    #endif
   #elif defined __ALTIVEC__
     #define EIGEN_VECTORIZE
     #define EIGEN_VECTORIZE_ALTIVEC

Eigen/src/Core/arch/SSE/PacketMath.h

@@ -118,6 +118,9 @@ template<> EIGEN_STRONG_INLINE Packet4f ei_pmul<Packet4f>(const Packet4f& a, con
 template<> EIGEN_STRONG_INLINE Packet2d ei_pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_mul_pd(a,b); }
 template<> EIGEN_STRONG_INLINE Packet4i ei_pmul<Packet4i>(const Packet4i& a, const Packet4i& b)
 {
+#ifdef __SSE4_1__
+  return _mm_mullo_epi32(a,b);
+#else
   // this version is slightly faster than 4 scalar products
   return ei_vec4i_swizzle1(
             ei_vec4i_swizzle2(
@@ -126,6 +129,7 @@ template<> EIGEN_STRONG_INLINE Packet4i ei_pmul<Packet4i>(const Packet4i& a, con
                             ei_vec4i_swizzle1(b,1,0,3,2)),
               0,2,0,2),
             0,2,1,3);
+#endif
 }
 
 template<> EIGEN_STRONG_INLINE Packet4f ei_pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_div_ps(a,b); }

cmake/EigenTesting.cmake

@@ -147,33 +147,45 @@ macro(ei_testing_print_summary)
 
   if(EIGEN_TEST_SSE2)
     message("SSE2:              ON")
-  else(EIGEN_TEST_SSE2)
+  else()
     message("SSE2:              Using architecture defaults")
-  endif(EIGEN_TEST_SSE2)
+  endif()
 
   if(EIGEN_TEST_SSE3)
     message("SSE3:              ON")
-  else(EIGEN_TEST_SSE3)
+  else()
     message("SSE3:              Using architecture defaults")
-  endif(EIGEN_TEST_SSE3)
+  endif()
 
   if(EIGEN_TEST_SSSE3)
     message("SSSE3:             ON")
-  else(EIGEN_TEST_SSSE3)
+  else()
     message("SSSE3:             Using architecture defaults")
-  endif(EIGEN_TEST_SSSE3)
+  endif()
+
+  if(EIGEN_TEST_SSE4_1)
+    message("SSE4.1:            ON")
+  else()
+    message("SSE4.1:            Using architecture defaults")
+  endif()
+
+  if(EIGEN_TEST_SSE4_2)
+    message("SSE4.2:            ON")
+  else()
+    message("SSE4.2:            Using architecture defaults")
+  endif()
 
   if(EIGEN_TEST_ALTIVEC)
     message("Altivec:           Using architecture defaults")
-  else(EIGEN_TEST_ALTIVEC)
+  else()
     message("Altivec:           Using architecture defaults")
-  endif(EIGEN_TEST_ALTIVEC)
+  endif()
 
   if(EIGEN_TEST_NO_EXPLICIT_VECTORIZATION)
     message("Explicit vec:      OFF")
-  else(EIGEN_TEST_NO_EXPLICIT_VECTORIZATION)
+  else()
     message("Explicit vec:      Using architecture defaults")
-  endif(EIGEN_TEST_NO_EXPLICIT_VECTORIZATION)
+  endif()
 
   message("\n${EIGEN_TESTING_SUMMARY}")
   #   message("CXX:               ${CMAKE_CXX_COMPILER}")