Support MSA SIMD for MIPS

Makefile.am

@@ -94,6 +94,11 @@ NEON_LIBS = dft/simd/neon/libdft_neon_codelets.la	\
 rdft/simd/neon/librdft_neon_codelets.la
 endif
 
+if HAVE_MSA
+MSA_LIBS = dft/simd/msa/libdft_msa_codelets.la	\
+rdft/simd/msa/librdft_msa_codelets.la
+endif
+
 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
@@ -125,7 +130,7 @@ libfftw3@PREC_SUFFIX@_la_LIBADD =			\
 	reodft/libreodft.la				\
 	api/libapi.la					\
         $(SIMD_LIBS) $(SSE2_LIBS) $(AVX_LIBS) $(AVX_128_FMA_LIBS) \
-        $(AVX2_LIBS) $(ALTIVEC_LIBS) \
+        $(AVX2_LIBS) $(ALTIVEC_LIBS) $(MSA_LIBS) \
         $(VSX_LIBS) $(NEON_LIBS) $(KCVI_LIBS) $(AVX512_LIBS) \
         $(GENERIC_SIMD128_LIBS) $(GENERIC_SIMD256_LIBS) \
 	$(COMBINED_THREADLIBS)

api/version.c

@@ -77,6 +77,10 @@ const char X(version)[] = PACKAGE "-" PACKAGE_VERSION
    "-neon"
 #endif
 
+#if HAVE_MSA
+   "-msa"
+#endif
+
 #if defined(HAVE_GENERIC_SIMD128)
    "-generic_simd128"
 #endif

cmake.config.h.in

@@ -202,6 +202,9 @@
 /* Define to enable ARM NEON optimizations. */
 /* #undef HAVE_NEON */
 
+/* Define to enable MIPS MSA optimizations. */
+/* #undef HAVE_MSA */
+
 /* Define if OpenMP is enabled */
 #cmakedefine HAVE_OPENMP
 

configure.ac

@@ -234,6 +234,11 @@ if test "$have_generic_simd256" = "yes"; then
 fi
 AM_CONDITIONAL(HAVE_GENERIC_SIMD256, test "$have_generic_simd256" = "yes")
 
+AC_ARG_ENABLE(msa, [AC_HELP_STRING([--enable-msa],[enable MIPS MSA optimizations])], have_msa=$enableval, have_msa=no)
+if test "$have_msa" = "yes"; then
+        AC_DEFINE(HAVE_MSA,1,[Define to enable MIPS MSA optimizations.])
+fi
+AM_CONDITIONAL(HAVE_MSA, test "$have_msa" = "yes")
 
 dnl FIXME:
 dnl AC_ARG_ENABLE(mips-ps, [AC_HELP_STRING([--enable-mips-ps],[enable MIPS pair-single optimizations])], have_mips_ps=$enableval, have_mips_ps=no)
@@ -359,9 +364,12 @@ case "${ax_cv_c_compiler_vendor}" in
         fi
 
         # AVX2
+        # gcc-4.8 works with -march=core-avx2, but -mavx2 is not enough.
+        # Later versions seem to happy with -mavx2, so try the arch one first.
         if test "$have_avx2" = "yes" -a "x$AVX2_CFLAGS" = x; then
-            AX_CHECK_COMPILER_FLAGS(-mavx2, [AVX2_CFLAGS="-mavx2"],
-                [AC_MSG_ERROR([Need a version of gcc with -mavx2])])
+            AX_CHECK_COMPILER_FLAGS(-march=core-avx2, [AVX2_CFLAGS="-march=core-avx2"],
+                [AX_CHECK_COMPILER_FLAGS(-mavx2, [AVX2_CFLAGS="-mavx2"],
+                    [AC_MSG_ERROR([Need a version of gcc with either -march=core-avx2 or -mavx2])])])
             AX_CHECK_COMPILER_FLAGS(-mfma, [AVX2_CFLAGS="$AVX2_CFLAGS -mfma"],
                 [AC_MSG_WARN([Need a version of gcc with -mfma (harmless for icc)])])
         fi
@@ -411,6 +419,11 @@ case "${ax_cv_c_compiler_vendor}" in
                 [AC_MSG_ERROR([Need a version of gcc with -mvsx])])
         fi
 
+        if test "$have_msa" = "yes" -a "x$MSA_CFLAGS" = x; then
+            AX_CHECK_COMPILER_FLAGS(-mmsa, [MSA_CFLAGS="-mmsa"],
+                [AC_MSG_ERROR([Need a version of gcc with -mmsa])])
+        fi
+
     dnl FIXME:
     dnl elif test "$have_mips_ps" = "yes"; then
     dnl     # Just punt here and use only new 4.2 compiler :(
@@ -471,6 +484,7 @@ AC_SUBST(KCVI_CFLAGS)
 AC_SUBST(ALTIVEC_CFLAGS)
 AC_SUBST(VSX_CFLAGS)
 AC_SUBST(NEON_CFLAGS)
+AC_SUBST(MSA_CFLAGS)
 
 dnl add stack alignment CFLAGS if so requested
 if test "$with_incoming_stack_boundary"x != "no"x; then
@@ -766,6 +780,7 @@ AC_CONFIG_FILES([
    dft/simd/altivec/Makefile
    dft/simd/vsx/Makefile
    dft/simd/neon/Makefile
+   dft/simd/msa/Makefile
    dft/simd/generic-simd128/Makefile
    dft/simd/generic-simd256/Makefile
 
@@ -786,6 +801,7 @@ AC_CONFIG_FILES([
    rdft/simd/altivec/Makefile
    rdft/simd/vsx/Makefile
    rdft/simd/neon/Makefile
+   rdft/simd/msa/Makefile
    rdft/simd/generic-simd128/Makefile
    rdft/simd/generic-simd256/Makefile
 

dft/codelet-dft.h

@@ -106,6 +106,7 @@ extern const solvtab X(solvtab_dft_kcvi);
 extern const solvtab X(solvtab_dft_altivec);
 extern const solvtab X(solvtab_dft_vsx);
 extern const solvtab X(solvtab_dft_neon);
+extern const solvtab X(solvtab_dft_msa);
 extern const solvtab X(solvtab_dft_generic_simd128);
 extern const solvtab X(solvtab_dft_generic_simd256);
 

dft/conf.c

@@ -79,6 +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_MSA
+     if (X(have_simd_msa)())
+	  X(solvtab_exec)(X(solvtab_dft_msa), p);
+#endif
 #if HAVE_GENERIC_SIMD128
      X(solvtab_exec)(X(solvtab_dft_generic_simd128), p);
 #endif

dft/simd/Makefile.am

@@ -1,4 +1,4 @@
 AM_CPPFLAGS = -I $(top_srcdir)
-SUBDIRS = common sse2 avx avx-128-fma avx2 avx2-128 avx512 kcvi altivec vsx neon generic-simd128 generic-simd256
+SUBDIRS = common sse2 avx avx-128-fma avx2 avx2-128 avx512 kcvi altivec vsx neon msa generic-simd128 generic-simd256
 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

dft/simd/msa/Makefile.am

@@ -0,0 +1,13 @@
+AM_CFLAGS = $(MSA_CFLAGS)
+SIMD_HEADER=simd-support/simd-msa.h
+
+include $(top_srcdir)/dft/simd/codlist.mk
+include $(top_srcdir)/dft/simd/simd.mk
+
+if HAVE_MSA
+
+BUILT_SOURCES = $(EXTRA_DIST)
+noinst_LTLIBRARIES = libdft_msa_codelets.la
+libdft_msa_codelets_la_SOURCES = $(BUILT_SOURCES)
+
+endif

doc/install.texi

@@ -199,6 +199,7 @@ of the time).  @xref{Cycle Counters}.
 @code{--enable-altivec} (single),
 @code{--enable-vsx} (single, double),
 @code{--enable-neon} (single, double on aarch64),
+@code{--enable-msa} (single, double on mips),
 @code{--enable-generic-simd128},
 and
 @code{--enable-generic-simd256}:

doc/intro.texi

@@ -18,8 +18,7 @@ transform (DFT) and various special cases thereof.
 
 @item  FFTW supports arbitrary multi-dimensional data.
 
-@item  FFTW supports the SSE, SSE2, AVX, AVX2, AVX512, KCVI, Altivec, VSX, and
-       NEON vector instruction sets.
+@item  FFTW supports the SSE, SSE2, AVX, AVX2, AVX512, KCVI, Altivec, VSX, NEON and MSA vector instruction sets.
 
 @item  FFTW includes parallel (multi-threaded) transforms
        for shared-memory systems.

doc/other.texi

@@ -16,8 +16,9 @@ special operations supported by some processors to perform a single
 operation on several numbers (usually 2 or 4) simultaneously.  SIMD
 floating-point instructions are available on several popular CPUs:
 SSE/SSE2/AVX/AVX2/AVX512/KCVI on some x86/x86-64 processors, AltiVec and
-VSX on some POWER/PowerPCs, NEON on some ARM models.  FFTW can be
-compiled to support the SIMD instructions on any of these systems.
+VSX on some POWER/PowerPCs, NEON on some ARM models, MSA on some MIPS
+models.  FFTW can be compiled to support the SIMD instructions on any of
+these systems.
 @cindex SIMD
 @cindex SSE
 @cindex SSE2

kernel/ifftw.h

@@ -103,7 +103,7 @@ extern void X(extract_reim)(int sign, R *c, R **r, R **i);
       defined(HAVE_AVX2) || defined(HAVE_AVX512) || \
       defined(HAVE_KCVI) || \
       defined(HAVE_ALTIVEC) || defined(HAVE_VSX) || \
-      defined(HAVE_MIPS_PS) || \
+      defined(HAVE_MIPS_PS) || defined(HAVE_MSA) || \
       defined(HAVE_GENERIC_SIMD128) || defined(HAVE_GENERIC_SIMD256)
 #define HAVE_SIMD 1
 #else
@@ -119,6 +119,7 @@ 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_msa)(void);
 
 /* forward declarations */
 typedef struct problem_s problem;

rdft/codelet-rdft.h

@@ -145,6 +145,7 @@ extern const solvtab X(solvtab_rdft_kcvi);
 extern const solvtab X(solvtab_rdft_altivec);
 extern const solvtab X(solvtab_rdft_vsx);
 extern const solvtab X(solvtab_rdft_neon);
+extern const solvtab X(solvtab_rdft_msa);
 extern const solvtab X(solvtab_rdft_generic_simd128);
 extern const solvtab X(solvtab_rdft_generic_simd256);
 

rdft/conf.c

@@ -96,6 +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_MSA
+     if (X(have_simd_msa)())
+	  X(solvtab_exec)(X(solvtab_rdft_msa), p);
+#endif
 #if HAVE_GENERIC_SIMD128
      X(solvtab_exec)(X(solvtab_rdft_generic_simd128), p);
 #endif

rdft/simd/Makefile.am

@@ -1,4 +1,4 @@
 
 AM_CPPFLAGS = -I $(top_srcdir)
-SUBDIRS = common sse2 avx avx-128-fma avx2 avx2-128 avx512 kcvi altivec vsx neon generic-simd128 generic-simd256
+SUBDIRS = common sse2 avx avx-128-fma avx2 avx2-128 avx512 kcvi altivec vsx neon msa generic-simd128 generic-simd256
 EXTRA_DIST = hc2cbv.h hc2cfv.h codlist.mk simd.mk

rdft/simd/msa/Makefile.am

@@ -0,0 +1,13 @@
+AM_CFLAGS = $(MSA_CFLAGS)
+SIMD_HEADER=simd-support/simd-msa.h
+
+include $(top_srcdir)/rdft/simd/codlist.mk
+include $(top_srcdir)/rdft/simd/simd.mk
+
+if HAVE_MSA
+
+noinst_LTLIBRARIES = librdft_msa_codelets.la
+BUILT_SOURCES = $(EXTRA_DIST)
+librdft_msa_codelets_la_SOURCES = $(BUILT_SOURCES)
+
+endif

simd-support/Makefile.am

@@ -11,5 +11,6 @@ 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 \
+msa.c simd-msa.h \
 simd-generic128.h simd-generic256.h
 

simd-support/msa.c

@@ -0,0 +1,73 @@
+/*
+ * 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
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
+ *
+ */
+
+
+#include "kernel/ifftw.h"
+
+#if HAVE_MSA
+
+/* check for an environment where signals are known to work */
+#if defined(unix) || defined(linux)
+  # include <signal.h>
+  # include <setjmp.h>
+
+  static jmp_buf jb;
+
+  static void sighandler(int x)
+  {
+    UNUSED(x);
+    longjmp(jb, 1);
+  }
+
+  static int msa_works(void)
+  {
+    void (*oldsig)(int);
+    oldsig = signal(SIGILL, sighandler);
+    if (setjmp(jb)) {
+      signal(SIGILL, oldsig);
+      return 0;
+    } else {
+      /* asm volatile ("xor.v $w0, $w0, $w0"); */
+      asm volatile (".long 0x7860001e");
+      signal(SIGILL, oldsig);
+      return 1;
+    }
+  }
+
+  int X(have_simd_msa)(void)
+  {
+    static int init = 0, res;
+
+    if (!init) {
+      res = msa_works();
+      init = 1;
+    }
+    return res;
+  }
+
+#else
+/* don't know how to autodetect MSA; assume it is present */
+  int X(have_simd_msa)(void)
+  {
+    return 1;
+  }
+#endif
+
+#endif

simd-support/simd-common.h

@@ -44,6 +44,9 @@
 #    define ALIGNMENT 16     /* Alignment for the LD/ST macros */
 #  endif
 #  define ALIGNMENTA 64   /* Alignment for the LDA/STA macros */
+#elif defined(HAVE_MSA)
+#  define ALIGNMENT  16    /* Alignment for the LD/ST macros */
+#  define ALIGNMENTA 16    /* Alignment for the LDA/STA macros */
 #elif defined(HAVE_GENERIC_SIMD256)
 #  if defined(FFTW_SINGLE)
 #    define ALIGNMENT 8