1 diff --git libpng-1.6.22-orig/configure.ac libpng-1.6.22/configure.ac 2 --- libpng-1.6.22-orig/configure.ac 2016-05-25 18:59:10.000000000 -0400 3 +++ libpng-1.6.22/configure.ac 2016-05-25 19:48:10.631751170 -0400 4 @@ -341,16 +341,50 @@ AC_ARG_ENABLE([arm-neon], 5 6 AM_CONDITIONAL([PNG_ARM_NEON], 7 [test "$enable_arm_neon" != 'no' && 8 case "$host_cpu" in 9 arm*|aarch64*) :;; 10 *) test "$enable_arm_neon" != '';; 11 esac]) 12 13 +# INTEL 14 +# ===== 15 +# 16 +# INTEL SSE (SIMD) support. 17 + 18 +AC_ARG_ENABLE([intel-sse], 19 + AS_HELP_STRING([[[--enable-intel-sse]]], 20 + [Enable Intel SSE optimizations: =no/off, yes/on:] 21 + [no/off: disable the optimizations;] 22 + [yes/on: enable the optimizations.] 23 + [If not specified: determined by the compiler.]), 24 + [case "$enableval" in 25 + no|off) 26 + # disable the default enabling: 27 + AC_DEFINE([PNG_INTEL_SSE_OPT], [0], 28 + [Disable Intel SSE optimizations]) 29 + # Prevent inclusion of the assembler files below: 30 + enable_intel_sse=no;; 31 + yes|on) 32 + AC_DEFINE([PNG_INTEL_SSE_OPT], [1], 33 + [Enable Intel SSE optimizations]);; 34 + *) 35 + AC_MSG_ERROR([--enable-intel-sse=${enable_intel_sse}: invalid value]) 36 + esac]) 37 + 38 +# Add Intel specific files to all builds where the host_cpu is Intel ('x86*') 39 +# or where Intel optimizations were explicitly requested (this allows a 40 +# fallback if a future host CPU does not match 'x86*') 41 +AM_CONDITIONAL([PNG_INTEL_SSE], 42 + [test "$enable_intel_sse" != 'no' && 43 + case "$host_cpu" in 44 + i?86|x86_64) :;; 45 + *) test "$enable_intel_sse" != '';; 46 + esac]) 47 AC_MSG_NOTICE([[Extra options for compiler: $PNG_COPTS]]) 48 49 # Config files, substituting as above 50 AC_CONFIG_FILES([Makefile libpng.pc:libpng.pc.in]) 51 AC_CONFIG_FILES([libpng-config:libpng-config.in], 52 [chmod +x libpng-config]) 53 54 AC_OUTPUT 55 diff --git libpng-1.6.22-orig/Makefile.am libpng-1.6.22/Makefile.am 56 --- libpng-1.6.22-orig/Makefile.am 2016-05-17 18:15:12.000000000 -0400 57 +++ libpng-1.6.22/Makefile.am 2016-05-25 19:48:10.631751170 -0400 58 @@ -89,16 +89,20 @@ libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@_la_SO 59 pngset.c pngtrans.c pngwio.c pngwrite.c pngwtran.c pngwutil.c\ 60 png.h pngconf.h pngdebug.h pnginfo.h pngpriv.h pngstruct.h pngusr.dfa 61 62 if PNG_ARM_NEON 63 libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@_la_SOURCES += arm/arm_init.c\ 64 arm/filter_neon.S arm/filter_neon_intrinsics.c 65 endif 66 67 +if PNG_INTEL_SSE 68 +libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@_la_SOURCES += contrib/intel/intel_init.c\ 69 + contrib/intel/filter_sse2_intrinsics.c 70 +endif 71 nodist_libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@_la_SOURCES = pnglibconf.h 72 73 libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@_la_LDFLAGS = -no-undefined -export-dynamic \ 74 -version-number @PNGLIB_MAJOR@@PNGLIB_MINOR@:@PNGLIB_RELEASE@:0 75 76 if HAVE_LD_VERSION_SCRIPT 77 # Versioned symbols and restricted exports 78 if HAVE_SOLARIS_LD 79 diff --git libpng-1.6.22-orig/pngpriv.h libpng-1.6.22/pngpriv.h 80 --- libpng-1.6.22-orig/pngpriv.h 2016-05-25 18:59:10.000000000 -0400 81 +++ libpng-1.6.22/pngpriv.h 2016-05-25 19:48:10.635751171 -0400 82 @@ -177,16 +177,52 @@ 83 # endif /* !PNG_ARM_NEON_IMPLEMENTATION */ 84 85 # ifndef PNG_ARM_NEON_IMPLEMENTATION 86 /* Use the intrinsics code by default. */ 87 # define PNG_ARM_NEON_IMPLEMENTATION 1 88 # endif 89 #endif /* PNG_ARM_NEON_OPT > 0 */ 90 91 +#ifndef PNG_INTEL_SSE_OPT 92 +# ifdef PNG_INTEL_SSE 93 + /* Only check for SSE if the build configuration has been modified to 94 + * enable SSE optimizations. This means that these optimizations will 95 + * be off by default. See contrib/intel for more details. 96 + */ 97 +# if defined(__SSE4_1__) || defined(__AVX__) || defined(__SSSE3__) || \ 98 + defined(__SSE2__) || defined(_M_X64) || defined(_M_AMD64) || \ 99 + (defined(_M_IX86_FP) && _M_IX86_FP >= 2) 100 +# define PNG_INTEL_SSE_OPT 1 101 +# endif 102 +# endif 103 +#endif 104 + 105 +#if PNG_INTEL_SSE_OPT > 0 106 +# ifndef PNG_INTEL_SSE_IMPLEMENTATION 107 +# if defined(__SSE4_1__) || defined(__AVX__) 108 + /* We are not actually using AVX, but checking for AVX is the best 109 + way we can detect SSE4.1 and SSSE3 on MSVC. 110 + */ 111 +# define PNG_INTEL_SSE_IMPLEMENTATION 3 112 +# elif defined(__SSSE3__) 113 +# define PNG_INTEL_SSE_IMPLEMENTATION 2 114 +# elif defined(__SSE2__) || defined(_M_X64) || defined(_M_AMD64) || \ 115 + (defined(_M_IX86_FP) && _M_IX86_FP >= 2) 116 +# define PNG_INTEL_SSE_IMPLEMENTATION 1 117 +# else 118 +# define PNG_INTEL_SSE_IMPLEMENTATION 0 119 +# endif 120 +# endif 121 + 122 +# if PNG_INTEL_SSE_IMPLEMENTATION > 0 123 +# define PNG_FILTER_OPTIMIZATIONS png_init_filter_functions_sse2 124 +# endif 125 +#endif 126 + 127 /* Is this a build of a DLL where compilation of the object modules requires 128 * different preprocessor settings to those required for a simple library? If 129 * so PNG_BUILD_DLL must be set. 130 * 131 * If libpng is used inside a DLL but that DLL does not export the libpng APIs 132 * PNG_BUILD_DLL must not be set. To avoid the code below kicking in build a 133 * static library of libpng then link the DLL against that. 134 */ 135 @@ -1184,16 +1220,29 @@ PNG_INTERNAL_FUNCTION(void,png_read_filt 136 row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY); 137 PNG_INTERNAL_FUNCTION(void,png_read_filter_row_avg4_neon,(png_row_infop 138 row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY); 139 PNG_INTERNAL_FUNCTION(void,png_read_filter_row_paeth3_neon,(png_row_infop 140 row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY); 141 PNG_INTERNAL_FUNCTION(void,png_read_filter_row_paeth4_neon,(png_row_infop 142 row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY); 143 144 +PNG_INTERNAL_FUNCTION(void,png_read_filter_row_sub3_sse2,(png_row_infop 145 + row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY); 146 +PNG_INTERNAL_FUNCTION(void,png_read_filter_row_sub4_sse2,(png_row_infop 147 + row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY); 148 +PNG_INTERNAL_FUNCTION(void,png_read_filter_row_avg3_sse2,(png_row_infop 149 + row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY); 150 +PNG_INTERNAL_FUNCTION(void,png_read_filter_row_avg4_sse2,(png_row_infop 151 + row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY); 152 +PNG_INTERNAL_FUNCTION(void,png_read_filter_row_paeth3_sse2,(png_row_infop 153 + row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY); 154 +PNG_INTERNAL_FUNCTION(void,png_read_filter_row_paeth4_sse2,(png_row_infop 155 + row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY); 156 + 157 /* Choose the best filter to use and filter the row data */ 158 PNG_INTERNAL_FUNCTION(void,png_write_find_filter,(png_structrp png_ptr, 159 png_row_infop row_info),PNG_EMPTY); 160 161 #ifdef PNG_SEQUENTIAL_READ_SUPPORTED 162 PNG_INTERNAL_FUNCTION(void,png_read_IDAT_data,(png_structrp png_ptr, 163 png_bytep output, png_alloc_size_t avail_out),PNG_EMPTY); 164 /* Read 'avail_out' bytes of data from the IDAT stream. If the output buffer 165
