src/png/arm/arm_init.c

   1
   2 /* arm_init.c - NEON optimised filter functions
   3  *
   4  * Copyright (c) 2013 Glenn Randers-Pehrson
   5  * Written by Mans Rullgard, 2011.
   6  * Last changed in libpng 1.5.15 [March 28, 2013]
   7  *
   8  * This code is released under the libpng license.
   9  * For conditions of distribution and use, see the disclaimer
  10  * and license in png.h
  11  */
  12 /* Below, after checking __linux__, various non-C90 POSIX 1003.1 functions are
  13  * called.
  14  */
  15 #define _POSIX_SOURCE 1
  16
  17 #include "../pngpriv.h"
  18
  19 #ifdef PNG_ARM_NEON_SUPPORTED
  20 #ifdef PNG_ARM_NEON_CHECK_SUPPORTED /* Do run-time checks */
  21 #include <signal.h> /* for sig_atomic_t */
  22
  23 #ifdef __ANDROID__
  24 /* Linux provides access to information about CPU capabilites via
  25  * /proc/self/auxv, however Android blocks this while still claiming to be
  26  * Linux.  The Andoid NDK, however, provides appropriate support.
  27  *
  28  * Documentation: http://www.kandroid.org/ndk/docs/CPU-ARM-NEON.html
  29  */
  30 #include <cpu-features.h>
  31
  32 static int
  33 png_have_neon(png_structp png_ptr)
  34 {
  35    /* This is a whole lot easier than the mess below, however it is probably
  36     * implemented as below, therefore it is better to cache the result (these
  37     * function calls may be slow!)
  38     */
  39    PNG_UNUSED(png_ptr)
  40    return android_getCpuFamily() == ANDROID_CPU_FAMILY_ARM &&
  41       (android_getCpuFeatures() & ANDROID_CPU_ARM_FEATURE_NEON) != 0;
  42 }
  43 #elif defined(__linux__)
  44 /* The generic __linux__ implementation requires reading /proc/self/auxv and
  45  * looking at each element for one that records NEON capabilities.
  46  */
  47 #include <unistd.h> /* for POSIX 1003.1 */
  48 #include <errno.h>  /* for EINTR */
  49
  50 #include <sys/types.h>
  51 #include <sys/stat.h>
  52 #include <fcntl.h>
  53 #include <elf.h>
  54 #include <asm/hwcap.h>
  55
  56 /* A read call may be interrupted, in which case it returns -1 and sets errno to
  57  * EINTR if nothing was done, otherwise (if something was done) a partial read
  58  * may result.
  59  */
  60 static size_t
  61 safe_read(png_structp png_ptr, int fd, void *buffer_in, size_t nbytes)
  62 {
  63    size_t ntotal = 0;
  64    char *buffer = png_voidcast(char*, buffer_in);
  65
  66    while (nbytes > 0)
  67    {
  68       unsigned int nread;
  69       int iread;
  70
  71       /* Passing nread > INT_MAX to read is implementation defined in POSIX
  72        * 1003.1, therefore despite the unsigned argument portable code must
  73        * limit the value to INT_MAX!
  74        */
  75       if (nbytes > INT_MAX)
  76          nread = INT_MAX;
  77
  78       else
  79          nread = (unsigned int)/*SAFE*/nbytes;
  80
  81       iread = read(fd, buffer, nread);
  82
  83       if (iread == -1)
  84       {
  85          /* This is the devil in the details, a read can terminate early with 0
  86           * bytes read because of EINTR, yet it still returns -1 otherwise end
  87           * of file cannot be distinguished.
  88           */
  89          if (errno != EINTR)
  90          {
  91             png_warning(png_ptr, "/proc read failed");
  92             return 0; /* I.e. a permanent failure */
  93          }
  94       }
  95
  96       else if (iread < 0)
  97       {
  98          /* Not a valid 'read' result: */
  99          png_warning(png_ptr, "OS /proc read bug");
 100          return 0;
 101       }
 102
 103       else if (iread > 0)
 104       {
 105          /* Continue reading until a permanent failure, or EOF */
 106          buffer += iread;
 107          nbytes -= (unsigned int)/*SAFE*/iread;
 108          ntotal += (unsigned int)/*SAFE*/iread;
 109       }
 110
 111       else
 112          return ntotal;
 113    }
 114
 115    return ntotal; /* nbytes == 0 */
 116 }
 117
 118 static int
 119 png_have_neon(png_structp png_ptr)
 120 {
 121    int fd = open("/proc/self/auxv", O_RDONLY);
 122    Elf32_auxv_t aux;
 123
 124    /* Failsafe: failure to open means no NEON */
 125    if (fd == -1)
 126    {
 127       png_warning(png_ptr, "/proc/self/auxv open failed");
 128       return 0;
 129    }
 130
 131    while (safe_read(png_ptr, fd, &aux, sizeof aux) == sizeof aux)
 132    {
 133       if (aux.a_type == AT_HWCAP && (aux.a_un.a_val & HWCAP_NEON) != 0)
 134       {
 135          close(fd);
 136          return 1;
 137       }
 138    }
 139
 140    close(fd);
 141    return 0;
 142 }
 143 #else
 144    /* We don't know how to do a run-time check on this system */
 145 #  error "no support for run-time ARM NEON checks"
 146 #endif /* OS checks */
 147 #endif /* PNG_ARM_NEON_CHECK_SUPPORTED */
 148
 149 #ifndef PNG_ALIGNED_MEMORY_SUPPORTED
 150 #  error "ALIGNED_MEMORY is required; set: -DPNG_ALIGNED_MEMORY_SUPPORTED"
 151 #endif
 152
 153 void
 154 png_init_filter_functions_neon(png_structp pp, unsigned int bpp)
 155 {
 156 #ifdef PNG_ARM_NEON_API_SUPPORTED
 157    switch ((pp->options >> PNG_ARM_NEON) & 3)
 158    {
 159       case PNG_OPTION_UNSET:
 160          /* Allow the run-time check to execute if it has been enabled -
 161           * thus both API and CHECK can be turned on.  If it isn't supported
 162           * this case will fall through to the 'default' below, which just
 163           * returns.
 164           */
 165 #endif /* PNG_ARM_NEON_API_SUPPORTED */
 166 #ifdef PNG_ARM_NEON_CHECK_SUPPORTED
 167          {
 168             static volatile sig_atomic_t no_neon = -1; /* not checked */
 169
 170             if (no_neon < 0)
 171                no_neon = !png_have_neon(pp);
 172
 173             if (no_neon)
 174                return;
 175          }
 176 #ifdef PNG_ARM_NEON_API_SUPPORTED
 177          break;
 178 #endif
 179 #endif /* PNG_ARM_NEON_CHECK_SUPPORTED */
 180 #ifdef PNG_ARM_NEON_API_SUPPORTED
 181       case PNG_OPTION_ON:
 182          /* Option turned on */
 183          break;
 184
 185       default: /* OFF or INVALID */
 186          return;
 187    }
 188 #endif
 189
 190    /* IMPORTANT: any new external functions used here must be declared using
 191     * PNG_INTERNAL_FUNCTION in ../pngpriv.h.  This is required so that the
 192     * 'prefix' option to configure works:
 193     *
 194     *    ./configure --with-libpng-prefix=foobar_
 195     *
 196     * Verify you have got this right by running the above command, doing a build
 197     * and examining pngprefix.h; it must contain a #define for every external
 198     * function you add.  (Notice that this happens automatically for the
 199     * initialization function.)
 200     */
 201    pp->read_filter[PNG_FILTER_VALUE_UP-1] = png_read_filter_row_up_neon;
 202
 203    if (bpp == 3)
 204    {
 205       pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub3_neon;
 206       pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg3_neon;
 207       pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
 208          png_read_filter_row_paeth3_neon;
 209    }
 210
 211    else if (bpp == 4)
 212    {
 213       pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub4_neon;
 214       pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg4_neon;
 215       pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
 216           png_read_filter_row_paeth4_neon;
 217    }
 218 }
 219 #endif /* FILTER_OPTIMIZATIONS && __arm__ && __ARM_NEON__ */