package/libs/zlib/patches/001-neon-implementation-of-adler32.patch

   1 From d2f06cd65d7ac39c6dd6761eef162abc946b155b Mon Sep 17 00:00:00 2001
   2 From: Adenilson Cavalcanti <adenilson.cavalcanti@arm.com>
   3 Date: Tue, 11 Apr 2017 17:13:02 -0700
   4 Subject: [PATCH] NEON implementation for Adler32
   5
   6 The checksum is calculated in the uncompressed PNG data
   7 and can be made much faster by using SIMD.
   8
   9 Tests in ARMv8 yielded an improvement of about 3x
  10 (e.g. walltime was 350ms x 125ms for a 4096x4096 bytes
  11 executed 30 times). That results in at least 18% improvement
  12 in image decoding in Chromium.
  13
  14 Further details at:
  15 https://bugs.chromium.org/p/chromium/issues/detail?id=688601
  16 ---
  17  CMakeLists.txt             |  29 +++++++---
  18  adler32.c                  |   5 ++
  19  contrib/README.contrib     |   3 +
  20  contrib/arm/neon_adler32.c | 137 +++++++++++++++++++++++++++++++++++++++++++++
  21  4 files changed, 166 insertions(+), 8 deletions(-)
  22  create mode 100644 contrib/arm/neon_adler32.c
  23
  24 --- a/CMakeLists.txt
  25 +++ b/CMakeLists.txt
  26 @@ -7,6 +7,7 @@ set(VERSION "1.2.12")
  27
  28  option(ASM686 "Enable building i686 assembly implementation")
  29  option(AMD64 "Enable building amd64 assembly implementation")
  30 +option(ARMv8 "Enable building ARM NEON intrinsics implementation")
  31
  32  set(INSTALL_BIN_DIR "${CMAKE_INSTALL_PREFIX}/bin" CACHE PATH "Installation directory for executables")
  33  set(INSTALL_LIB_DIR "${CMAKE_INSTALL_PREFIX}/lib" CACHE PATH "Installation directory for libraries")
  34 @@ -132,14 +133,26 @@ endif()
  35  if(CMAKE_COMPILER_IS_GNUCC)
  36      if(ASM686)
  37          set(ZLIB_ASMS contrib/asm686/match.S)
  38 -    elseif (AMD64)
  39 +    elseif(AMD64)
  40          set(ZLIB_ASMS contrib/amd64/amd64-match.S)
  41 -    endif ()
  42 +    elseif(ARMv8)
  43 +        set(ZLIB_ARMv8 contrib/arm/neon_adler32.c)
  44 +    endif()
  45
  46 -       if(ZLIB_ASMS)
  47 -               add_definitions(-DASMV)
  48 -               set_source_files_properties(${ZLIB_ASMS} PROPERTIES LANGUAGE C COMPILE_FLAGS -DNO_UNDERLINE)
  49 -       endif()
  50 +    if(ZLIB_ASMS)
  51 +        add_definitions(-DASMV)
  52 +           set_source_files_properties(${ZLIB_ASMS} PROPERTIES LANGUAGE C COMPILE_FLAGS -DNO_UNDERLINE)
  53 +    elseif(ZLIB_ARMv8)
  54 +        add_definitions(-DARMv8)
  55 +        set(COMPILER ${CMAKE_C_COMPILER})
  56 +        # NEON is mandatory in ARMv8.
  57 +        if(${COMPILER} MATCHES "aarch64")
  58 +            set_source_files_properties(${ZLIB_ARMv8} PROPERTIES LANGUAGE C COMPILE_FLAGS -march=armv8-a)
  59 +        # But it was optional for ARMv7.
  60 +        elseif(${COMPILER} MATCHES "arm")
  61 +            set_source_files_properties(${ZLIB_ARMv8} PROPERTIES LANGUAGE C COMPILE_FLAGS -mfpu=neon)
  62 +        endif()
  63 +    endif()
  64  endif()
  65
  66  if(MSVC)
  67 @@ -183,8 +196,8 @@ if(MINGW)
  68      set(ZLIB_DLL_SRCS ${CMAKE_CURRENT_BINARY_DIR}/zlib1rc.obj)
  69  endif(MINGW)
  70
  71 -add_library(zlib SHARED ${ZLIB_SRCS} ${ZLIB_ASMS} ${ZLIB_DLL_SRCS} ${ZLIB_PUBLIC_HDRS} ${ZLIB_PRIVATE_HDRS})
  72 -add_library(zlibstatic STATIC ${ZLIB_SRCS} ${ZLIB_ASMS} ${ZLIB_PUBLIC_HDRS} ${ZLIB_PRIVATE_HDRS})
  73 +add_library(zlib SHARED ${ZLIB_SRCS} ${ZLIB_ASMS} ${ZLIB_ARMv8} ${ZLIB_DLL_SRCS} ${ZLIB_PUBLIC_HDRS} ${ZLIB_PRIVATE_HDRS})
  74 +add_library(zlibstatic STATIC ${ZLIB_SRCS} ${ZLIB_ASMS} ${ZLIB_ARMv8} ${ZLIB_PUBLIC_HDRS} ${ZLIB_PRIVATE_HDRS})
  75  set_target_properties(zlib PROPERTIES DEFINE_SYMBOL ZLIB_DLL)
  76  set_target_properties(zlib PROPERTIES SOVERSION 1)
  77
  78 diff --git a/adler32.c b/adler32.c
  79 index d0be4380..45ebaa4b 100644
  80 --- a/adler32.c
  81 +++ b/adler32.c
  82 @@ -136,7 +136,12 @@ uLong ZEXPORT adler32(adler, buf, len)
  83      const Bytef *buf;
  84      uInt len;
  85  {
  86 +#ifdef ARMv8
  87 +#  pragma message("Using NEON-ized Adler32.")
  88 +    return NEON_adler32(adler, buf, len);
  89 +#else
  90      return adler32_z(adler, buf, len);
  91 +#endif
  92  }
  93
  94  /* ========================================================================= */
  95 --- a/contrib/README.contrib
  96 +++ b/contrib/README.contrib
  97 @@ -8,6 +8,9 @@ ada/        by Dmitriy Anisimkov <anisim
  98          Support for Ada
  99          See http://zlib-ada.sourceforge.net/
 100
 101 +arm/        by Adenilson Cavalcanti <cavalcantii@chromium.org>
 102 +        ARM optimizations (NEON and ARMv8 code).
 103 +
 104  blast/      by Mark Adler <madler@alumni.caltech.edu>
 105          Decompressor for output of PKWare Data Compression Library (DCL)
 106
 107 --- /dev/null
 108 +++ b/contrib/arm/neon_adler32.c
 109 @@ -0,0 +1,137 @@
 110 +/* Copyright (C) 1995-2011, 2016 Mark Adler
 111 + * Copyright (C) 2017 ARM Holdings Inc.
 112 + * Authors: Adenilson Cavalcanti <adenilson.cavalcanti@arm.com>
 113 + *          Simon Hosie <simon.hosie@arm.com>
 114 + * This software is provided 'as-is', without any express or implied
 115 + * warranty.  In no event will the authors be held liable for any damages
 116 + * arising from the use of this software.
 117 + * Permission is granted to anyone to use this software for any purpose,
 118 + * including commercial applications, and to alter it and redistribute it
 119 + * freely, subject to the following restrictions:
 120 + * 1. The origin of this software must not be misrepresented; you must not
 121 + *  claim that you wrote the original software. If you use this software
 122 + *    in a product, an acknowledgment in the product documentation would be
 123 + *    appreciated but is not required.
 124 + * 2. Altered source versions must be plainly marked as such, and must not be
 125 + *    misrepresented as being the original software.
 126 + * 3. This notice may not be removed or altered from any source distribution.
 127 + */
 128 +
 129 +#if (defined(__ARM_NEON__) || defined(__ARM_NEON))
 130 +#include <arm_neon.h>
 131 +
 132 +static void NEON_accum32(uint32_t *s, const unsigned char *buf,
 133 +                         unsigned int len)
 134 +{
 135 +    static const uint8_t taps[32] = {
 136 +        32, 31, 30, 29, 28, 27, 26, 25,
 137 +        24, 23, 22, 21, 20, 19, 18, 17,
 138 +        16, 15, 14, 13, 12, 11, 10, 9,
 139 +        8, 7, 6, 5, 4, 3, 2, 1 };
 140 +
 141 +    uint32x2_t adacc2, s2acc2, as;
 142 +    uint8x16_t t0 = vld1q_u8(taps), t1 = vld1q_u8(taps + 16);
 143 +
 144 +    uint32x4_t adacc = vdupq_n_u32(0), s2acc = vdupq_n_u32(0);
 145 +    adacc = vsetq_lane_u32(s[0], adacc, 0);
 146 +    s2acc = vsetq_lane_u32(s[1], s2acc, 0);
 147 +
 148 +    while (len >= 2) {
 149 +        uint8x16_t d0 = vld1q_u8(buf), d1 = vld1q_u8(buf + 16);
 150 +        uint16x8_t adler, sum2;
 151 +        s2acc = vaddq_u32(s2acc, vshlq_n_u32(adacc, 5));
 152 +        adler = vpaddlq_u8(       d0);
 153 +        adler = vpadalq_u8(adler, d1);
 154 +        sum2 = vmull_u8(      vget_low_u8(t0), vget_low_u8(d0));
 155 +        sum2 = vmlal_u8(sum2, vget_high_u8(t0), vget_high_u8(d0));
 156 +        sum2 = vmlal_u8(sum2, vget_low_u8(t1), vget_low_u8(d1));
 157 +        sum2 = vmlal_u8(sum2, vget_high_u8(t1), vget_high_u8(d1));
 158 +        adacc = vpadalq_u16(adacc, adler);
 159 +        s2acc = vpadalq_u16(s2acc, sum2);
 160 +        len -= 2;
 161 +        buf += 32;
 162 +    }
 163 +
 164 +    while (len > 0) {
 165 +        uint8x16_t d0 = vld1q_u8(buf);
 166 +        uint16x8_t adler, sum2;
 167 +        s2acc = vaddq_u32(s2acc, vshlq_n_u32(adacc, 4));
 168 +        adler = vpaddlq_u8(d0);
 169 +        sum2 = vmull_u8(      vget_low_u8(t1), vget_low_u8(d0));
 170 +        sum2 = vmlal_u8(sum2, vget_high_u8(t1), vget_high_u8(d0));
 171 +        adacc = vpadalq_u16(adacc, adler);
 172 +        s2acc = vpadalq_u16(s2acc, sum2);
 173 +        buf += 16;
 174 +        len--;
 175 +    }
 176 +
 177 +    adacc2 = vpadd_u32(vget_low_u32(adacc), vget_high_u32(adacc));
 178 +    s2acc2 = vpadd_u32(vget_low_u32(s2acc), vget_high_u32(s2acc));
 179 +    as = vpadd_u32(adacc2, s2acc2);
 180 +    s[0] = vget_lane_u32(as, 0);
 181 +    s[1] = vget_lane_u32(as, 1);
 182 +}
 183 +
 184 +static void NEON_handle_tail(uint32_t *pair, const unsigned char *buf,
 185 +                             unsigned int len)
 186 +{
 187 +    /* Oldie K&R code integration. */
 188 +    unsigned int i;
 189 +    for (i = 0; i < len; ++i) {
 190 +        pair[0] += buf[i];
 191 +        pair[1] += pair[0];
 192 +    }
 193 +}
 194 +
 195 +extern unsigned long NEON_adler32(unsigned long adler, const unsigned char *buf,
 196 +                                  const unsigned int len)
 197 +{
 198 +    /* initial Adler-32 value (deferred check for len == 1 speed) */
 199 +    if (!buf)
 200 +        return 1L;
 201 +
 202 +    /* The largest prime smaller than 65536. */
 203 +    const uint32_t M_BASE = 65521;
 204 +    /* This is the threshold where doing accumulation may overflow. */
 205 +    const int M_NMAX = 5552;
 206 +
 207 +    unsigned long sum2;
 208 +    uint32_t pair[2];
 209 +    int n = M_NMAX;
 210 +    unsigned int done = 0;
 211 +    /* Oldie K&R code integration. */
 212 +    unsigned int i;
 213 +
 214 +    /* Split Adler-32 into component sums, it can be supplied by
 215 +     * the caller sites (e.g. in a PNG file).
 216 +     */
 217 +    sum2 = (adler >> 16) & 0xffff;
 218 +    adler &= 0xffff;
 219 +    pair[0] = adler;
 220 +    pair[1] = sum2;
 221 +
 222 +    for (i = 0; i < len; i += n) {
 223 +        if ((i + n) > len)
 224 +            n = len - i;
 225 +
 226 +        if (n < 16)
 227 +            break;
 228 +
 229 +        NEON_accum32(pair, buf + i, n / 16);
 230 +        pair[0] %= M_BASE;
 231 +        pair[1] %= M_BASE;
 232 +
 233 +        done += (n / 16) * 16;
 234 +    }
 235 +
 236 +    /* Handle the tail elements. */
 237 +    if (done < len) {
 238 +        NEON_handle_tail(pair, (buf + done), len - done);
 239 +        pair[0] %= M_BASE;
 240 +        pair[1] %= M_BASE;
 241 +    }
 242 +
 243 +    /* D = B * 65536 + A, see: https://en.wikipedia.org/wiki/Adler-32. */
 244 +    return (pair[1] << 16) | pair[0];
 245 +}
 246 +#endif