Merge pull request #101 from xianyi/develop

rebase

Author: martin-frbg

Makefile.x86

@@ -54,3 +54,19 @@ LIBATLAS	= -L$(ATLASPATH)/32 -lcblas -lf77blas -latlas -lm
 else
 LIBATLAS	= -L$(ATLASPATH)/32 -lptf77blas -lptatlas -lpthread -lm
 endif
+
+ifdef HAVE_SSE3
+ifndef DYNAMIC_ARCH
+CCOMMON_OPT += -msse3
+FCOMMON_OPT += -msse3
+ifdef HAVE_SSSE3
+CCOMMON_OPT += -mssse3
+FCOMMON_OPT += -mssse3
+endif
+ifdef HAVE_SSE4_1
+CCOMMON_OPT += -msse4.1
+FCOMMON_OPT += -msse4.1
+endif
+endif
+endif
+

cmake/cc.cmake

@@ -124,10 +124,19 @@ if (NOT DYNAMIC_ARCH)
 	if (HAVE_AVX)
         set (CCOMMON_OPT  "${CCOMMON_OPT} -mavx")
 	endif ()
+	if (HAVE_SSE)
+	set (CCOMMON_OPT  "${CCOMMON_OPT} -msse")
+	endif ()
+	if (HAVE_SSE2)
+        set (CCOMMON_OPT  "${CCOMMON_OPT} -msse2")
+	endif ()
 	if (HAVE_SSE3)
         set (CCOMMON_OPT  "${CCOMMON_OPT} -msse3")
 	endif ()
 	if (HAVE_SSSE3)
         set (CCOMMON_OPT  "${CCOMMON_OPT} -mssse3")
 	endif ()
+	if (HAVE_SSE4_1)
+	set (CCOMMON_OPT  "${CCOMMON_OPT} -msse4.1")
+	endif ()
 endif()

cmake/system.cmake

@@ -70,9 +70,21 @@ if (DEFINED TARGET)
       set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mavx2")
     endif()
   endif()
+  if (DEFINED HAVE_SSE)
+    set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -msse")
+  endif()
+  if (DEFINED HAVE_SSE2)
+    set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -msse2")
+  endif()
   if (DEFINED HAVE_SSE3)
     set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -msse3")
   endif()
+    if (DEFINED HAVE_SSSE3)
+    set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mssse3")
+  endif()
+    if (DEFINED HAVE_SSE4_1)
+    set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -msse4.1")
+  endif()
 endif()
 
 if (DEFINED TARGET)

kernel/Makefile

@@ -45,7 +45,10 @@ endif
 
 ifdef TARGET_CORE
 	ifeq ($(TARGET_CORE), $(filter $(TARGET_CORE),PRESCOTT CORE2 PENRYN DUNNINGTON ATOM NANO SANDYBRIDGE HASWELL NEHALEM ZEN BARCELONA BOBCAT BULLDOZER PILEDRIVER EXCAVATOR STEAMROLLER OPTERON_SSE3))
-	override CFLAGS += -msse3 -mssse3 -msse4.1
+	override CFLAGS += -msse -msse2 -msse3 -mssse3 -msse4.1
+endif
+	ifeq ($(TARGET_CORE), $(filter $(TARGET_CORE),KATMAI COPPERMINE BANIAS NORTHWOOD ATHLON OPTERON))
+	override CFLAGS += -msse -msse2
 endif
 ifeq ($(TARGET_CORE), COOPERLAKE)
  override CFLAGS += -DBUILD_KERNEL -DTABLE_NAME=gotoblas_$(TARGET_CORE)

kernel/setparam-ref.c

@@ -1164,7 +1164,7 @@ static void init_parameter(void) {
   TABLE_NAME.xgemm3m_q = QGEMM_DEFAULT_Q;
 #endif
 
-#if (CORE_KATMAI)  || (CORE_COPPERMINE) || (CORE_BANIAS) || (CORE_YONAH) || (CORE_ATHLON)
+#if defined(CORE_KATMAI)  || defined(CORE_COPPERMINE) || defined(CORE_BANIAS) || defined(CORE_YONAH) || defined(CORE_ATHLON)
 
 #ifdef DEBUG
   fprintf(stderr, "Katmai, Coppermine, Banias, Athlon\n");

kernel/x86_64/iamax_sse.S

@@ -51,6 +51,8 @@
 #define MAXPS	maxps
 #define MAXSS	maxss
 #ifdef USE_MIN
+#undef MAXPS
+#undef MAXSS
 #define MAXPS	minps
 #define MAXSS	minss
 #endif