multiarch experiment

Makefile.am

@@ -2,7 +2,7 @@ ACLOCAL_AMFLAGS = -I m4
 
 lib_LTLIBRARIES = libsharp.la
 
-src_sharp = \
+libsharp_la_SOURCES = \
   c_utils/c_utils.c \
   c_utils/c_utils.h \
   pocketfft/pocketfft.c \
@@ -10,11 +10,6 @@ src_sharp = \
   libsharp/sharp.c \
   libsharp/sharp_almhelpers.c \
   libsharp/sharp_core.c \
-  libsharp/sharp_core_avx.c \
-  libsharp/sharp_core_avx2.c \
-  libsharp/sharp_core_fma.c \
-  libsharp/sharp_core_fma4.c \
-  libsharp/sharp_core_avx512f.c \
   libsharp/sharp_geomhelpers.c \
   libsharp/sharp_legendre_roots.c \
   libsharp/sharp_ylmgen_c.c \
@@ -23,6 +18,16 @@ src_sharp = \
   libsharp/sharp_vecsupport.h \
   libsharp/sharp_ylmgen_c.h
 
+libavx_la_SOURCES = libsharp/sharp_core_avx.c
+libavx2_la_SOURCES = libsharp/sharp_core_avx2.c
+libfma_la_SOURCES = libsharp/sharp_core_fma.c
+libfma4_la_SOURCES = libsharp/sharp_core_fma4.c
+libavx512f_la_SOURCES = libsharp/sharp_core_avx512f.c
+
+noinst_LTLIBRARIES = libavx.la libavx2.la libfma.la libfma4.la libavx512f.la
+
+libsharp_la_LIBADD = libavx.la libavx2.la libfma.la libfma4.la libavx512f.la
+
 include_HEADERS = \
   libsharp/sharp.h \
   libsharp/sharp_geomhelpers.h \
@@ -33,8 +38,6 @@ EXTRA_DIST = \
   libsharp/sharp_core_inc.c \
   runtest.sh
 
-libsharp_la_SOURCES = $(src_sharp)
-
 check_PROGRAMS = sharp_testsuite
 sharp_testsuite_SOURCES = libsharp/sharp_testsuite.c c_utils/memusage.c c_utils/memusage.h c_utils/walltime_c.c c_utils/walltime_c.h
 sharp_testsuite_LDADD = libsharp.la
@@ -43,6 +46,12 @@ TESTS = runtest.sh
 
 AM_CFLAGS = -I$(top_srcdir)/c_utils -I$(top_srcdir)/libsharp @AM_CFLAGS@
 
+libavx_la_CFLAGS = ${AM_CFLAGS} -mavx
+libavx2_la_CFLAGS = ${AM_CFLAGS} -mavx2
+libfma_la_CFLAGS = ${AM_CFLAGS} -mfma
+libfma4_la_CFLAGS = ${AM_CFLAGS} -mfma4
+libavx512f_la_CFLAGS = ${AM_CFLAGS} -mavx512f
+
 pkgconfigdir = $(libdir)/pkgconfig
 nodist_pkgconfig_DATA = @PACKAGE_NAME@.pc
 

libsharp/sharp_core.c

@@ -18,7 +18,7 @@ static t_veclen veclen_ = NULL;
 static t_max_nvec max_nvec_ = NULL;
 static t_architecture architecture_ = NULL;
 
-#if defined(__GNUC__) && defined (__x86_64__) && (__GNUC__>=6)
+#ifdef MULTIARCH
 
 #define DECL(arch) \
 static int XCONCATX2(have,arch)(void) \
@@ -59,7 +59,7 @@ DECL(avx)
 
 static void assign_funcs(void)
   {
-#if defined(__GNUC__) && defined (__x86_64__) && (__GNUC__>=6)
+#ifdef MULTIARCH
 #define DECL2(arch) \
   if (XCONCATX2(have,arch)()) \
     { \

libsharp/sharp_core_avx.c

@@ -1,11 +1,8 @@
-#if (!defined(__AVX__)) && defined(__GNUC__) && defined (__x86_64__) && (__GNUC__>=6)
+#ifdef MULTIARCH
-
 #define XCONCATX(a,b) a##_##b
 #define XCONCATX2(a,b) XCONCATX(a,b)
 #define XARCH(a) XCONCATX2(a,ARCH)
 
 #define ARCH avx
-#pragma GCC target("avx")
 #include "sharp_core_inc.c"
-
 #endif

libsharp/sharp_core_avx2.c

@@ -1,11 +1,8 @@
-#if (!defined(__AVX2__)) && defined(__GNUC__) && defined (__x86_64__) && (__GNUC__>=6)
+#ifdef MULTIARCH
-
 #define XCONCATX(a,b) a##_##b
 #define XCONCATX2(a,b) XCONCATX(a,b)
 #define XARCH(a) XCONCATX2(a,ARCH)
 
 #define ARCH avx2
-#pragma GCC target("avx2")
 #include "sharp_core_inc.c"
-
 #endif

libsharp/sharp_core_avx512f.c

@@ -1,11 +1,8 @@
-#if (!defined(__AVX512F__)) && defined(__GNUC__) && defined (__x86_64__) && (__GNUC__>=6)
+#ifdef MULTIARCH
-
 #define XCONCATX(a,b) a##_##b
 #define XCONCATX2(a,b) XCONCATX(a,b)
 #define XARCH(a) XCONCATX2(a,ARCH)
 
 #define ARCH avx512f
-#pragma GCC target("avx512f")
 #include "sharp_core_inc.c"
-
 #endif

libsharp/sharp_core_fma.c

@@ -1,11 +1,8 @@
-#if (!defined(__FMA__)) && defined(__GNUC__) && defined (__x86_64__) && (__GNUC__>=6)
+#ifdef MULTIARCH
-
 #define XCONCATX(a,b) a##_##b
 #define XCONCATX2(a,b) XCONCATX(a,b)
 #define XARCH(a) XCONCATX2(a,ARCH)
 
 #define ARCH fma
-#pragma GCC target("fma")
 #include "sharp_core_inc.c"
-
 #endif

libsharp/sharp_core_fma4.c

@@ -1,11 +1,8 @@
-#if (!defined(__FMA4__)) && defined(__GNUC__) && defined (__x86_64__) && (__GNUC__>=6)
+#ifdef MULTIARCH
-
 #define XCONCATX(a,b) a##_##b
 #define XCONCATX2(a,b) XCONCATX(a,b)
 #define XARCH(a) XCONCATX2(a,ARCH)
 
 #define ARCH fma4
-#pragma GCC target("fma4")
 #include "sharp_core_inc.c"
-
 #endif