128-bit AVX2 SIMD support

Add 128 bit support for AVX2. Similar to AVX-128, this improves slightly on SSE2 due to more efficient instructions, and the shorter SIMD width is beneficial in some cases. Both 128- and 256-bit flavors will be built automatically with --enable-avx2, and the timing routines will chose the best one automatically.
chiajeng · Mar 28, 2015 · de81bfd · de81bfd
1 parent da988fa
commit de81bfd
Show file tree

Hide file tree

Showing 16 changed files with 395 additions and 43 deletions.
diff --git a/.gitignore b/.gitignore
@@ -18,6 +18,7 @@ rdft/simd/altivec/*.c
 rdft/simd/avx/*.c
 rdft/simd/avx-128/*.c
 rdft/simd/avx2/*.c
+rdft/simd/avx2-128/*.c
 rdft/simd/common/*.c
 rdft/simd/kcvi/*.c
 rdft/simd/neon/*.c
@@ -30,6 +31,7 @@ dft/simd/altivec/*.c
 dft/simd/avx/*.c
 dft/simd/avx-128/*.c
 dft/simd/avx2/*.c
+dft/simd/avx2-128/*.c
 dft/simd/common/*.c
 dft/simd/kcvi/*.c
 dft/simd/neon/*.c

diff --git a/Makefile.am b/Makefile.am
@@ -52,7 +52,9 @@ endif
 
 if HAVE_AVX2
 AVX2_LIBS = dft/simd/avx2/libdft_avx2_codelets.la	\
-rdft/simd/avx2/librdft_avx2_codelets.la
+dft/simd/avx2-128/libdft_avx2_128_codelets.la \
+rdft/simd/avx2/librdft_avx2_codelets.la \
+rdft/simd/avx2-128/librdft_avx2_128_codelets.la
 endif
 
 if HAVE_KCVI

diff --git a/configure.ac b/configure.ac
@@ -615,6 +615,7 @@ AC_CONFIG_FILES([
    dft/simd/avx/Makefile
    dft/simd/avx-128/Makefile
    dft/simd/avx2/Makefile
+   dft/simd/avx2-128/Makefile
    dft/simd/kcvi/Makefile
    dft/simd/altivec/Makefile
    dft/simd/neon/Makefile
@@ -630,6 +631,7 @@ AC_CONFIG_FILES([
    rdft/simd/avx/Makefile
    rdft/simd/avx-128/Makefile
    rdft/simd/avx2/Makefile
+   rdft/simd/avx2-128/Makefile
    rdft/simd/kcvi/Makefile
    rdft/simd/altivec/Makefile
    rdft/simd/neon/Makefile

diff --git a/dft/codelet-dft.h b/dft/codelet-dft.h
@@ -100,6 +100,7 @@ extern const solvtab X(solvtab_dft_sse2);
 extern const solvtab X(solvtab_dft_avx);
 extern const solvtab X(solvtab_dft_avx_128);
 extern const solvtab X(solvtab_dft_avx2);
+extern const solvtab X(solvtab_dft_avx2_128);
 extern const solvtab X(solvtab_dft_kcvi);
 extern const solvtab X(solvtab_dft_altivec);
 extern const solvtab X(solvtab_dft_neon);

diff --git a/dft/conf.c b/dft/conf.c
@@ -54,7 +54,10 @@ void X(dft_conf_standard)(planner *p)
 #endif
 #if HAVE_AVX2
      if (X(have_simd_avx2)())
-	  X(solvtab_exec)(X(solvtab_dft_avx2), p);
+     {
+         X(solvtab_exec)(X(solvtab_dft_avx2), p);
+         X(solvtab_exec)(X(solvtab_dft_avx2_128), p);
+     }
 #endif
 #if HAVE_KCVI
      if (X(have_simd_kcvi)())

diff --git a/dft/simd/Makefile.am b/dft/simd/Makefile.am
@@ -1,3 +1,3 @@
-SUBDIRS = common sse2 avx avx-128 avx2 kcvi altivec neon
+SUBDIRS = common sse2 avx avx-128 avx2 avx2-128 kcvi altivec neon
 EXTRA_DIST = n1b.h n1f.h n2b.h n2f.h n2s.h q1b.h q1f.h t1b.h t1bu.h	\
 t1f.h t1fu.h t2b.h t2f.h t3b.h t3f.h ts.h codlist.mk simd.mk
diff --git a/dft/simd/avx2-128/Makefile.am b/dft/simd/avx2-128/Makefile.am
@@ -0,0 +1,13 @@
+AM_CFLAGS = $(AVX2_CFLAGS)
+SIMD_HEADER=simd-avx2-128.h
+
+include $(top_srcdir)/dft/simd/codlist.mk
+include $(top_srcdir)/dft/simd/simd.mk
+
+if HAVE_AVX2
+
+BUILT_SOURCES = $(EXTRA_DIST)
+noinst_LTLIBRARIES = libdft_avx2_128_codelets.la
+libdft_avx2_128_codelets_la_SOURCES = $(BUILT_SOURCES)
+
+endif
diff --git a/kernel/ifftw.h b/kernel/ifftw.h
@@ -97,8 +97,11 @@ extern void X(extract_reim)(int sign, R *c, R **r, R **i);
 #define CIMPLIES(ante, post) (!(ante) || (post))
 
 /* define HAVE_SIMD if any simd extensions are supported */
-#if defined(HAVE_SSE) || defined(HAVE_SSE2) || defined(HAVE_ALTIVEC) || \
-     defined(HAVE_MIPS_PS) || defined(HAVE_AVX) || defined(HAVE_AVX2) || defined(HAVE_KCVI)
+#if defined(HAVE_SSE) || defined(HAVE_SSE2) || \
+      defined(HAVE_AVX) || defined(HAVE_AVX2) || \
+      defined(HAVE_KCVI) || \
+      defined(HAVE_ALTIVEC) || \
+      defined(HAVE_MIPS_PS)
 #define HAVE_SIMD 1
 #else
 #define HAVE_SIMD 0

diff --git a/rdft/codelet-rdft.h b/rdft/codelet-rdft.h
@@ -139,6 +139,7 @@ extern const solvtab X(solvtab_rdft_sse2);
 extern const solvtab X(solvtab_rdft_avx);
 extern const solvtab X(solvtab_rdft_avx_128);
 extern const solvtab X(solvtab_rdft_avx2);
+extern const solvtab X(solvtab_rdft_avx2_128);
 extern const solvtab X(solvtab_rdft_kcvi);
 extern const solvtab X(solvtab_rdft_altivec);
 extern const solvtab X(solvtab_rdft_neon);

diff --git a/rdft/conf.c b/rdft/conf.c
@@ -71,7 +71,10 @@ void X(rdft_conf_standard)(planner *p)
 #endif
 #if HAVE_AVX2
      if (X(have_simd_avx2)())
-	  X(solvtab_exec)(X(solvtab_rdft_avx2), p);
+     {
+         X(solvtab_exec)(X(solvtab_rdft_avx2), p);
+         X(solvtab_exec)(X(solvtab_rdft_avx2_128), p);
+     }
 #endif
 #if HAVE_KCVI
      if (X(have_simd_kcvi)())

diff --git a/rdft/simd/Makefile.am b/rdft/simd/Makefile.am
@@ -1,2 +1,2 @@
-SUBDIRS = common sse2 avx avx-128 avx2 kcvi altivec neon
+SUBDIRS = common sse2 avx avx-128 avx2 avx2-128 kcvi altivec neon
 EXTRA_DIST = hc2cbv.h hc2cfv.h codlist.mk simd.mk
diff --git a/rdft/simd/avx2-128/Makefile.am b/rdft/simd/avx2-128/Makefile.am
@@ -0,0 +1,15 @@
+AM_CFLAGS = $(AVX2_CFLAGS)
+SIMD_HEADER=simd-avx2-128.h
+
+include $(top_srcdir)/rdft/simd/codlist.mk
+include $(top_srcdir)/rdft/simd/simd.mk
+
+if HAVE_AVX2
+
+noinst_LTLIBRARIES = librdft_avx2_128_codelets.la
+BUILT_SOURCES = $(EXTRA_DIST)
+librdft_avx2_128_codelets_la_SOURCES = $(BUILT_SOURCES)
+
+endif
+
+
diff --git a/simd-support/Makefile.am b/simd-support/Makefile.am
@@ -5,9 +5,9 @@ libsimd_support_la_SOURCES = taint.c simd-common.h \
 x86-cpuid.h amd64-cpuid.h \
 simd-sse2.h sse2.c	\
 avx.c simd-avx.h simd-avx-128.h \
-avx2.c simd-avx2.h \
+avx2.c simd-avx2.h simd-avx2-128.h \
 kcvi.c simd-kcvi.h \
-altivec.c simd-altivec.h	\
+altivec.c simd-altivec.h \
 neon.c simd-neon.h
 
 # sse2-nonportable.c needs SSE2_CFLAGS, but Automake does not support

diff --git a/simd-support/avx2.c b/simd-support/avx2.c
@@ -1,6 +1,6 @@
 /*
- * Copyright (c) 2003, 2007-11 Matteo Frigo
- * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
+ * Copyright (c) 2003, 2007-14 Matteo Frigo
+ * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License as published by