Make 128/256 bit generic simd separate options

These will only be used on esoteric and/or new architectures, which likely also miss cycle counters. In this case the widest simd would be picked automatically based on flops estimates, so to give the user more control it is better to provide separate options to enable/disable these two choices.
chiajeng · Apr 8, 2015 · 4b3dbf7 · 4b3dbf7
1 parent cbe2a4a
commit 4b3dbf7
Show file tree

Hide file tree

Showing 16 changed files with 47 additions and 63 deletions.
diff --git a/.gitignore b/.gitignore
@@ -20,6 +20,8 @@ rdft/simd/avx-128/*.c
 rdft/simd/avx2/*.c
 rdft/simd/avx2-128/*.c
 rdft/simd/avx512/*.c
+rdft/simd/generic-simd128/*.c
+rdft/simd/generic-simd256/*.c
 rdft/simd/common/*.c
 rdft/simd/kcvi/*.c
 rdft/simd/neon/*.c
@@ -34,6 +36,8 @@ dft/simd/avx-128/*.c
 dft/simd/avx2/*.c
 dft/simd/avx2-128/*.c
 dft/simd/avx512/*.c
+dft/simd/generic-simd128/*.c
+dft/simd/generic-simd256/*.c
 dft/simd/common/*.c
 dft/simd/kcvi/*.c
 dft/simd/neon/*.c

diff --git a/Makefile.am b/Makefile.am
@@ -82,13 +82,16 @@ NEON_LIBS = dft/simd/neon/libdft_neon_codelets.la	\
 rdft/simd/neon/librdft_neon_codelets.la
 endif
 
-if HAVE_GENERIC_SIMD
-GENERIC_SIMD_LIBS = dft/simd/generic-simd256/libdft_generic_simd256_codelets.la	\
-dft/simd/generic-simd128/libdft_generic_simd128_codelets.la \
-rdft/simd/generic-simd256/librdft_generic_simd256_codelets.la \
+if HAVE_GENERIC_SIMD128
+GENERIC_SIMD128_LIBS = dft/simd/generic-simd128/libdft_generic_simd128_codelets.la \
 rdft/simd/generic-simd128/librdft_generic_simd128_codelets.la
 endif
 
+if HAVE_GENERIC_SIMD256
+GENERIC_SIMD256_LIBS = dft/simd/generic-simd256/libdft_generic_simd256_codelets.la      \
+rdft/simd/generic-simd256/librdft_generic_simd256_codelets.la
+endif
+
 if THREADS
 if COMBINED_THREADS
 COMBINED_THREADLIBS=threads/libfftw3@PREC_SUFFIX@_threads.la
@@ -111,7 +114,7 @@ libfftw3@PREC_SUFFIX@_la_LIBADD =			\
 	api/libapi.la					\
         $(SIMD_LIBS) $(SSE2_LIBS) $(AVX_LIBS) $(AVX2_LIBS) $(ALTIVEC_LIBS) \
         $(VSX_LIBS) $(NEON_LIBS) $(KCVI_LIBS) $(AVX512_LIBS) \
-        $(GENERIC_SIMD_LIBS) \
+        $(GENERIC_SIMD128_LIBS) $(GENERIC_SIMD256_LIBS) \
 	$(COMBINED_THREADLIBS)
 
 if QUAD

diff --git a/api/version.c b/api/version.c
@@ -66,9 +66,12 @@ const char X(version)[] = PACKAGE "-" PACKAGE_VERSION
    "-neon"
 #endif
 
-#if defined(HAVE_GENERIC_SIMD)
-   "-generic_simd"
+#if defined(HAVE_GENERIC_SIMD128)
+   "-generic_simd128"
 #endif
 
+#if defined(HAVE_GENERIC_SIMD256)
+   "-generic_simd256"
+#endif
 
 ;
diff --git a/configure.ac b/configure.ac
@@ -194,11 +194,18 @@ if test "$have_armv8cyclecounter"x = "yes"x; then
 	AC_DEFINE(HAVE_ARMV8CC,1,[Define if you have enabled the cycle counter on ARMv8])
 fi
 
-AC_ARG_ENABLE(generic-simd, [AC_HELP_STRING([--enable-generic-simd],[enable generic (gcc) SIMD optimizations])], have_generic_simd=$enableval, have_generic_simd=no)
-if test "$have_generic_simd" = "yes"; then
-        AC_DEFINE(HAVE_GENERIC_SIMD,1,[Define to enable generic (gcc) SIMD float optimizations.])
+AC_ARG_ENABLE(generic-simd128, [AC_HELP_STRING([--enable-generic-simd128],[enable generic (gcc) 128-bit SIMD optimizations])], have_generic_simd128=$enableval, have_generic_simd128=no)
+if test "$have_generic_simd128" = "yes"; then
+        AC_DEFINE(HAVE_GENERIC_SIMD128,1,[Define to enable generic (gcc) 128-bit SIMD optimizations.])
 fi
-AM_CONDITIONAL(HAVE_GENERIC_SIMD, test "$have_generic_simd" = "yes")
+AM_CONDITIONAL(HAVE_GENERIC_SIMD128, test "$have_generic_simd" = "yes")
+
+AC_ARG_ENABLE(generic-simd256, [AC_HELP_STRING([--enable-generic-simd256],[enable generic (gcc) 256-bit SIMD optimizations])], have_generic_simd256=$enableval, have_generic_simd256=no)
+if test "$have_generic_simd256" = "yes"; then
+        AC_DEFINE(HAVE_GENERIC_SIMD256,1,[Define to enable generic (gcc) 256-bit SIMD optimizations.])
+fi
+AM_CONDITIONAL(HAVE_GENERIC_SIMD256, test "$have_generic_simd256" = "yes")
+
 
 dnl FIXME:
 dnl AC_ARG_ENABLE(mips-ps, [AC_HELP_STRING([--enable-mips-ps],[enable MIPS pair-single optimizations])], have_mips_ps=$enableval, have_mips_ps=no)

diff --git a/dft/conf.c b/dft/conf.c
@@ -79,11 +79,10 @@ void X(dft_conf_standard)(planner *p)
      if (X(have_simd_neon)())
 	  X(solvtab_exec)(X(solvtab_dft_neon), p);
 #endif
-#if HAVE_GENERIC_SIMD
-     if (X(have_simd_generic)())
-     {
-         X(solvtab_exec)(X(solvtab_dft_generic_simd128), p);
-         X(solvtab_exec)(X(solvtab_dft_generic_simd256), p);
-     }
+#if HAVE_GENERIC_SIMD128
+     X(solvtab_exec)(X(solvtab_dft_generic_simd128), p);
+#endif
+#if HAVE_GENERIC_SIMD256
+     X(solvtab_exec)(X(solvtab_dft_generic_simd256), p);
 #endif
 }
diff --git a/dft/simd/generic-simd128/Makefile.am b/dft/simd/generic-simd128/Makefile.am
@@ -3,7 +3,7 @@ SIMD_HEADER=simd-generic128.h
 include $(top_srcdir)/dft/simd/codlist.mk
 include $(top_srcdir)/dft/simd/simd.mk
 
-if HAVE_GENERIC_SIMD
+if HAVE_GENERIC_SIMD128
 
 BUILT_SOURCES = $(EXTRA_DIST)
 noinst_LTLIBRARIES = libdft_generic_simd128_codelets.la

diff --git a/dft/simd/generic-simd256/Makefile.am b/dft/simd/generic-simd256/Makefile.am
@@ -3,7 +3,7 @@ SIMD_HEADER=simd-generic256.h
 include $(top_srcdir)/dft/simd/codlist.mk
 include $(top_srcdir)/dft/simd/simd.mk
 
-if HAVE_GENERIC_SIMD
+if HAVE_GENERIC_SIMD256
 
 BUILT_SOURCES = $(EXTRA_DIST)
 noinst_LTLIBRARIES = libdft_generic_simd256_codelets.la

diff --git a/kernel/ifftw.h b/kernel/ifftw.h
@@ -102,7 +102,7 @@ extern void X(extract_reim)(int sign, R *c, R **r, R **i);
       defined(HAVE_KCVI) || \
       defined(HAVE_ALTIVEC) || defined(HAVE_VSX) || \
       defined(HAVE_MIPS_PS) || \
-      defined(HAVE_GENERIC128D) || defined(HAVE_GENERIC256D)
+      defined(HAVE_GENERIC_SIMD128) || defined(HAVE_GENERIC_SIMD256)
 #define HAVE_SIMD 1
 #else
 #define HAVE_SIMD 0
@@ -115,8 +115,6 @@ extern int X(have_simd_avx512)(void);
 extern int X(have_simd_altivec)(void);
 extern int X(have_simd_vsx)(void);
 extern int X(have_simd_neon)(void);
-extern int X(have_simd_generic128d)(void);
-extern int X(have_simd_generic256d)(void);
 
 /* forward declarations */
 typedef struct problem_s problem;

diff --git a/rdft/conf.c b/rdft/conf.c
@@ -96,11 +96,10 @@ void X(rdft_conf_standard)(planner *p)
      if (X(have_simd_neon)())
 	  X(solvtab_exec)(X(solvtab_rdft_neon), p);
 #endif
-#if HAVE_GENERIC_SIMD
-     if (X(have_simd_generic)())
-     {
-        X(solvtab_exec)(X(solvtab_rdft_generic_simd128), p);
-	X(solvtab_exec)(X(solvtab_rdft_generic_simd256), p);
-     }
+#if HAVE_GENERIC_SIMD128
+     X(solvtab_exec)(X(solvtab_rdft_generic_simd128), p);
+#endif
+#if HAVE_GENERIC_SIMD256
+     X(solvtab_exec)(X(solvtab_rdft_generic_simd256), p);
 #endif
 }
diff --git a/rdft/simd/generic-simd128/Makefile.am b/rdft/simd/generic-simd128/Makefile.am
@@ -3,7 +3,7 @@ SIMD_HEADER=simd-generic128.h
 include $(top_srcdir)/rdft/simd/codlist.mk
 include $(top_srcdir)/rdft/simd/simd.mk
 
-if HAVE_GENERIC_SIMD
+if HAVE_GENERIC_SIMD128
 
 noinst_LTLIBRARIES = librdft_generic_simd128_codelets.la
 BUILT_SOURCES = $(EXTRA_DIST)

diff --git a/rdft/simd/generic-simd256/Makefile.am b/rdft/simd/generic-simd256/Makefile.am
@@ -3,7 +3,7 @@ SIMD_HEADER=simd-generic256.h
 include $(top_srcdir)/rdft/simd/codlist.mk
 include $(top_srcdir)/rdft/simd/simd.mk
 
-if HAVE_GENERIC_SIMD
+if HAVE_GENERIC_SIMD256
 
 noinst_LTLIBRARIES = librdft_generic_simd256_codelets.la
 BUILT_SOURCES = $(EXTRA_DIST)

diff --git a/simd-support/Makefile.am b/simd-support/Makefile.am
@@ -10,7 +10,7 @@ avx512.c simd-avx512.h \
 kcvi.c simd-kcvi.h \
 altivec.c simd-altivec.h vsx.c simd-vsx.h \
 neon.c simd-neon.h \
-generic.c simd-generic128.h simd-generic256.h
+simd-generic128.h simd-generic256.h
 
 # sse2-nonportable.c needs SSE2_CFLAGS, but Automake does not support
 # per-object CFLAGS.  Thus we build a separate library.

diff --git a/simd-support/generic.c b/simd-support/generic.c
diff --git a/simd-support/simd-common.h b/simd-support/simd-common.h
@@ -52,7 +52,7 @@
 #    define ALIGNMENT 16     /* Alignment for the LD/ST macros */
 #  endif
 #  define ALIGNMENTA 64   /* Alignment for the LDA/STA macros */
-#elif defined(HAVE_GENERIC_SIMD)
+#elif defined(HAVE_GENERIC_SIMD128) || defined(HAVE_GENERIC_SIMD256)
 #    define ALIGNMENT 16
 #    define ALIGNMENTA sizeof(V)
 #endif

diff --git a/simd-support/simd-generic128.h b/simd-support/simd-generic128.h
@@ -25,7 +25,7 @@
 
 
 #if defined(FFTW_LDOUBLE) || defined(FFTW_QUAD)
-#  error "Generic simd only works in single or double precision"
+#  error "Generic simd128 only works in single or double precision"
 #endif
 
 #define SIMD_SUFFIX  _generic_simd128  /* for renaming */

diff --git a/simd-support/simd-generic256.h b/simd-support/simd-generic256.h
@@ -24,7 +24,7 @@
  */
 
 #if defined(FFTW_LDOUBLE) || defined(FFTW_QUAD)
-#  error "Generic simd only works in single or double precision"
+#  error "Generic simd256 only works in single or double precision"
 #endif
 
 #define SIMD_SUFFIX  _generic_simd256  /* for renaming */