[vlc-commits] Run-time CPU detection for ARM (meaning NEON)

[Rémi Denis-Courmont]

vlc | branch: master | Rémi Denis-Courmont <remi at remlab.net> | Mon Jul  4 21:14:18 2011 +0300| [f25efcaf2a7d3c2d9aefeb4974d225a25525fc21] | committer: Rémi Denis-Courmont

Run-time CPU detection for ARM (meaning NEON)

Unfortunately, we cannot emit NEON opcodes when NEON is not explicitly
enabled (-mfpu=neon), contrary to MMX & SSE on x86. As a consequence,
this will not work for inline assembler in a non-optimized plugin,
namely the deinterlacer.

There is also a(n hopefully theoretical) bug whereby the compiler would
emit NEON instructions in the descriptor or activation callback of a
NEON plugin. This could then crash if NEON is not supported, even
before the NEON run-time check is reached.

> http://git.videolan.org/gitweb.cgi/vlc.git/?a=commit;h=f25efcaf2a7d3c2d9aefeb4974d225a25525fc21
---

 configure.ac                                   |   10 ++++--
 modules/Makefile.am                            |    2 +-
 modules/arm_neon/Modules.am                    |    2 +-
 modules/video_filter/deinterlace/deinterlace.c |    2 +-
 src/misc/cpu.c                                 |   36 ++++++++++++++++++++++-
 5 files changed, 44 insertions(+), 8 deletions(-)

diff --git a/configure.ac b/configure.ac
index 14c729c..831f94f 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1489,12 +1489,16 @@ asm volatile("ssat r0, #1, r0":::"r0"); /* assume ARMv6 */
     ])
     CFLAGS="${CFLAGS_save}"
   ])
-  ARM_NEON_CFLAGS="$ac_cv_neon_inline"
+  AS_IF([test "$ac_cv_neon_inline" != "no"], [
+    NEON_CFLAGS="$ac_cv_neon_inline"
+    AC_DEFINE([CAN_COMPILE_NEON], 1,
+      [Define to 1 if NEON (and ARMv6) assembly is available with NEON_CFLAGS.])
+  ])
 ], [
   ac_cv_neon_inline="no"
 ])
-AC_SUBST(ARM_NEON_CFLAGS)
-AM_CONDITIONAL(HAVE_ARM_NEON, [test "${ac_cv_neon_inline}" != "no"])
+AC_SUBST(NEON_CFLAGS)
+AM_CONDITIONAL(HAVE_NEON, [test "${ac_cv_neon_inline}" != "no"])
 
 
 AC_ARG_ENABLE(altivec,
diff --git a/modules/Makefile.am b/modules/Makefile.am
index b84f2d9..4473820 100644
--- a/modules/Makefile.am
+++ b/modules/Makefile.am
@@ -51,7 +51,7 @@ endif
 if HAVE_ALTIVEC
 SUBDIRS += altivec
 endif
-if HAVE_ARM_NEON
+if HAVE_NEON
 SUBDIRS += arm_neon
 endif
 if BUILD_LUA
diff --git a/modules/arm_neon/Modules.am b/modules/arm_neon/Modules.am
index 1842561..8fa5b04 100644
--- a/modules/arm_neon/Modules.am
+++ b/modules/arm_neon/Modules.am
@@ -2,7 +2,7 @@
 # without this. (This is the case with iOS).
 LIBTOOL=@LIBTOOL@ --tag=CC
 
-AM_CFLAGS += $(ARM_NEON_CFLAGS)
+AM_CFLAGS += $(NEON_CFLAGS)
 
 libaudio_format_neon_plugin_la_SOURCES = \
 	s32_s16.S \
diff --git a/modules/video_filter/deinterlace/deinterlace.c b/modules/video_filter/deinterlace/deinterlace.c
index fa1a00b..60b5f45 100644
--- a/modules/video_filter/deinterlace/deinterlace.c
+++ b/modules/video_filter/deinterlace/deinterlace.c
@@ -682,7 +682,7 @@ int Open( vlc_object_t *p_this )
     }
     else
 #endif
-#if defined __ARM_NEON__
+#if defined __ARM_NEON__ // FIXME: runtime detect support
     if( vlc_CPU() & CPU_CAPABILITY_NEON )
     {
         p_sys->pf_merge = MergeNEON;
diff --git a/src/misc/cpu.c b/src/misc/cpu.c
index ad0c0b6..81f6948 100644
--- a/src/misc/cpu.c
+++ b/src/misc/cpu.c
@@ -247,10 +247,42 @@ uint32_t CPUCapabilities( void )
     }
 out:
 
-#elif defined( __arm__ )
-#   if defined( __ARM_NEON__ )
+#elif defined (__arm__)
+
+# if defined (__ARM_NEON__)
     i_capabilities |= CPU_CAPABILITY_NEON;
+# elif defined (CAN_COMPILE_NEON)
+#  define NEED_RUNTIME_CPU_CHECK 1
+# endif
+
+# ifdef NEED_RUNTIME_CPU_CHECK
+#  if defined (__linux__)
+    FILE *info = fopen ("/proc/cpuinfo", "rt");
+    if (info != NULL)
+    {
+        char *line = NULL;
+        size_t linelen = 0;
+
+        while (getline (&line, &linelen, info) != -1)
+        {
+             const char *cap;
+
+             if (strncmp (line, "Features\t:", 10))
+                 continue;
+#   if defined (CAN_COMPILE_NEON) && !defined (__ARM_NEON__)
+             cap = strstr (line + 10, " neon");
+             if (cap != NULL && (cap[5] == '\0' || cap[5] == ' '))
+                 i_capabilities |= CPU_CAPABILITY_NEON;
 #   endif
+             break;
+        }
+        fclose (info);
+        free (line);
+    }
+#  else
+#   warning Run-time CPU detection missing: optimizations disabled!
+#  endif
+# endif
 
 #elif defined( __powerpc__ ) || defined( __ppc__ ) || defined( __powerpc64__ ) \
     || defined( __ppc64__ )