[libpng16] Rearranged the ARM-NEON optimizations: Isolated the machine specific

code to the hardware subdirectory and added comments to pngrutil.c so that
implementors of other optimizations know what to do.

ANNOUNCE

@@ -1,5 +1,5 @@
 
-Libpng 1.6.0beta33 - December 10, 2012
+Libpng 1.6.0beta33 - December 13, 2012
 
 This is not intended to be a public release.  It will be replaced
 within a few weeks by a public version or by another test version.
@@ -552,12 +552,15 @@ Version 1.6.0beta32 [November 25, 2012]
   Fixed error checking in the simplified write API (Olaf van der Spek)
   Made png_user_version_check() ok to use with libpng version 1.10.x and later.
 
-Version 1.6.0beta33 [December 10, 2012]
+Version 1.6.0beta33 [December 13, 2012]
   Fixed typo in png.c (PNG_SET_CHUNK_MALLOC_MAX should be PNG_CHUNK_MALLOC_MAX)
     that causes the MALLOC_MAX limit not to work (John Bowler)
   Change png_warning() to png_app_error() in pngwrite.c and comment the
     fall-through condition.
   Change png_warning() to png_app_warning() in png_write_tRNS().
+  Rearranged the ARM-NEON optimizations: Isolated the machine specific code
+    to the hardware subdirectory and added comments to pngrutil.c so that
+    implementors of other optimizations know what to do.
 
 Send comments/corrections/commendations to png-mng-implement at lists.sf.net
 (subscription required; visit

CHANGES

@@ -4304,12 +4304,15 @@ Version 1.6.0beta32 [November 25, 2012]
   Fixed error checking in the simplified write API (Olaf van der Spek)
   Made png_user_version_check() ok to use with libpng version 1.10.x and later.
 
-Version 1.6.0beta33 [December 10, 2012]
+Version 1.6.0beta33 [December 13, 2012]
   Fixed typo in png.c (PNG_SET_CHUNK_MALLOC_MAX should be PNG_CHUNK_MALLOC_MAX)
     that causes the MALLOC_MAX limit not to work (John Bowler)
   Change png_warning() to png_app_error() in pngwrite.c and comment the
     fall-through condition.
   Change png_warning() to png_app_warning() in png_write_tRNS().
+  Rearranged the ARM-NEON optimizations: Isolated the machine specific code
+    to the hardware subdirectory and added comments to pngrutil.c so that
+    implementors of other optimizations know what to do.
 
 Send comments/corrections/commendations to png-mng-implement at lists.sf.net
 (subscription required; visit

Makefile.am

@@ -43,7 +43,8 @@ libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@_la_SOURCES = png.c pngerror.c\
 	png.h pngconf.h pngdebug.h pnginfo.h pngpriv.h pngstruct.h pngusr.dfa
 
 if PNG_ARM_NEON
-libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@_la_SOURCES += arm/filter_neon.S
+libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@_la_SOURCES += arm/arm_init.c\
+	arm/filter_neon.S
 endif
 
 nodist_libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@_la_SOURCES = pnglibconf.h

arm/arm_init.c

@@ -0,0 +1,74 @@
+
+/* filter_neon.S - NEON optimised filter functions
+ *
+ * Copyright (c) 2011 Glenn Randers-Pehrson
+ * Written by Mans Rullgard, 2011.
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+#include "../pngpriv.h"
+
+/* __arm__ is defined by GCC, MSVC defines _M_ARM to the ARM version number */
+#if defined __linux__ && defined __arm__
+#include <stdio.h>
+#include <elf.h>
+#include <asm/hwcap.h>
+
+static int png_have_hwcap(unsigned cap)
+{
+   FILE *f = fopen("/proc/self/auxv", "r");
+   Elf32_auxv_t aux;
+   int have_cap = 0;
+
+   if (!f)
+      return 0;
+
+   while (fread(&aux, sizeof(aux), 1, f) > 0)
+   {
+      if (aux.a_type == AT_HWCAP &&
+          aux.a_un.a_val & cap)
+      {
+         have_cap = 1;
+         break;
+      }
+   }
+
+   fclose(f);
+
+   return have_cap;
+}
+#endif /* __linux__ && __arm__ */
+
+void
+png_init_filter_functions_neon(png_structp pp, unsigned int bpp)
+{
+#ifdef __arm__
+#ifdef __linux__
+   if (!png_have_hwcap(HWCAP_NEON))
+      return;
+#endif
+
+   pp->read_filter[PNG_FILTER_VALUE_UP-1] = png_read_filter_row_up_neon;
+
+   if (bpp == 3)
+   {
+      pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub3_neon;
+      pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg3_neon;
+      pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
+         png_read_filter_row_paeth3_neon;
+   }
+
+   else if (bpp == 4)
+   {
+      pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub4_neon;
+      pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg4_neon;
+      pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
+          png_read_filter_row_paeth4_neon;
+   }
+#else
+   PNG_UNUSED(pp)
+   PNG_UNUSED(bpp)
+#endif
+}

arm/filter_neon.S

@@ -9,6 +9,7 @@
  * and license in png.h
  */
 
+#ifdef __arm__
 #if defined(__linux__) && defined(__ELF__)
 .section .note.GNU-stack,"",%progbits /* mark stack as non-executable */
 #endif
@@ -223,3 +224,4 @@ func    png_read_filter_row_paeth3_neon, export=1
 
         pop             {r4,pc}
 endfunc
+#endif

configure.ac

@@ -241,8 +241,11 @@ AC_SUBST([AM_CCASFLAGS], [-Wa,--noexecstack])
 AC_ARG_ENABLE([arm-neon],
    AS_HELP_STRING([[[--enable-arm-neon]]], [Enable ARM NEON optimizations]),
    [if test "${enableval}" = "yes"; then
-      AC_DEFINE([PNG_ARM_NEON], [1], [Enable ARM NEON optimizations])
+      AC_DEFINE([PNG_FILTER_OPTIMIZATIONS],
-      AC_DEFINE([PNG_ALIGNED_MEMORY_SUPPORTED], [1], [Align row buffers])
+                [png_init_filter_functions_neon],
+                [ARM NEON filter initialization function])
+      AC_DEFINE([PNG_ALIGNED_MEMORY_SUPPORTED], [1],
+                [Align row buffers])
     fi])
 AM_CONDITIONAL([PNG_ARM_NEON], [test "${enable_arm_neon:-no}" = yes])
 

pngpriv.h

@@ -1862,6 +1862,15 @@ PNG_INTERNAL_FUNCTION(void, png_image_free, (png_imagep image), PNG_EMPTY);
 
 #endif /* SIMPLIFIED READ/WRITE */
 
+#ifdef PNG_FILTER_OPTIMIZATIONS
+PNG_INTERNAL_FUNCTION(void, PNG_FILTER_OPTIMIZATIONS, (png_structp png_ptr, unsigned int bpp), PNG_EMPTY);
+   /* This is the initialization function for hardware specific optimizations,
+    * one implementation (for ARM NEON machines) is contained in
+    * arm/filter_neon.c.  It need not be defined - the generic code will be used
+    * if not.
+    */
+#endif
+
 /* Maintainer: Put new private prototypes here ^ */
 
 #include "pngdebug.h"

pngrutil.c

@@ -3863,66 +3863,6 @@ png_read_filter_row_paeth_multibyte_pixel(png_row_infop row_info, png_bytep row,
    }
 }
 
-#ifdef PNG_ARM_NEON
-
-#ifdef __linux__
-#include <stdio.h>
-#include <elf.h>
-#include <asm/hwcap.h>
-
-static int png_have_hwcap(unsigned cap)
-{
-   FILE *f = fopen("/proc/self/auxv", "r");
-   Elf32_auxv_t aux;
-   int have_cap = 0;
-
-   if (!f)
-      return 0;
-
-   while (fread(&aux, (sizeof aux), 1, f) > 0)
-   {
-      if (aux.a_type == AT_HWCAP &&
-          aux.a_un.a_val & cap)
-      {
-         have_cap = 1;
-         break;
-      }
-   }
-
-   fclose(f);
-
-   return have_cap;
-}
-#endif /* __linux__ */
-
-static void
-png_init_filter_functions_neon(png_structrp pp, unsigned int bpp)
-{
-#ifdef __linux__
-   if (!png_have_hwcap(HWCAP_NEON))
-      return;
-#endif
-
-   pp->read_filter[PNG_FILTER_VALUE_UP-1] = png_read_filter_row_up_neon;
-
-   if (bpp == 3)
-   {
-      pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub3_neon;
-      pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg3_neon;
-      pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
-         png_read_filter_row_paeth3_neon;
-   }
-
-   else if (bpp == 4)
-   {
-      pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub4_neon;
-      pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg4_neon;
-      pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
-          png_read_filter_row_paeth4_neon;
-   }
-}
-#endif /* PNG_ARM_NEON */
-
 static void
 png_init_filter_functions(png_structrp pp)
 {
@@ -3938,8 +3878,16 @@ png_init_filter_functions(png_structrp pp)
       pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
          png_read_filter_row_paeth_multibyte_pixel;
 
-#ifdef PNG_ARM_NEON
+#ifdef PNG_FILTER_OPTIMIZATIONS
-   png_init_filter_functions_neon(pp, bpp);
+   /* To use this define PNG_FILTER_OPTIMIZATIONS as the name of a function to
+    * call to install hardware optimizations for the above functions; simply
+    * replace whatever elements of the pp->read_filter[] array with a hardware
+    * specific (or, for that matter, generic) optimization.
+    *
+    * To see an example of this examine what configure.ac does when
+    * --enable-arm-neon is specified on the command line.
+    */
+   PNG_FILTER_OPTIMIZATIONS(pp, bpp);
 #endif
 }