git.sesse.net Git - ffmpeg/blob - libavfilter/x86/scene_sad_init.c

   1 /*
   2  * This file is part of FFmpeg.
   3  *
   4  * FFmpeg is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU Lesser General Public
   6  * License as published by the Free Software Foundation; either
   7  * version 2.1 of the License, or (at your option) any later version.
   8  *
   9  * FFmpeg is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  12  * Lesser General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU Lesser General Public
  15  * License along with FFmpeg; if not, write to the Free Software
  16  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  17  */
  18
  19 #include "libavutil/cpu.h"
  20 #include "libavutil/x86/cpu.h"
  21 #include "libavfilter/scene_sad.h"
  22
  23 #define SCENE_SAD_FUNC(FUNC_NAME, ASM_FUNC_NAME, MMSIZE)                      \
  24 void ASM_FUNC_NAME(SCENE_SAD_PARAMS);                                         \
  25                                                                               \
  26 static void FUNC_NAME(SCENE_SAD_PARAMS) {                                     \
  27     uint64_t sad[MMSIZE / 8] = {0};                                           \
  28     ptrdiff_t awidth = width & ~(MMSIZE - 1);                                 \
  29     *sum = 0;                                                                 \
  30     ASM_FUNC_NAME(src1, stride1, src2, stride2, awidth, height, sad);         \
  31     for (int i = 0; i < MMSIZE / 8; i++)                                      \
  32         *sum += sad[i];                                                       \
  33     ff_scene_sad_c(src1 + awidth, stride1,                                    \
  34                    src2 + awidth, stride2,                                    \
  35                    width - awidth, height, sad);                              \
  36     *sum += sad[0];                                                           \
  37 }
  38
  39 #if HAVE_X86ASM
  40 SCENE_SAD_FUNC(scene_sad_sse2, ff_scene_sad_sse2, 16);
  41 #if HAVE_AVX2_EXTERNAL
  42 SCENE_SAD_FUNC(scene_sad_avx2, ff_scene_sad_avx2, 32);
  43 #endif
  44 #endif
  45
  46 ff_scene_sad_fn ff_scene_sad_get_fn_x86(int depth)
  47 {
  48 #if HAVE_X86ASM
  49     int cpu_flags = av_get_cpu_flags();
  50     if (depth == 8) {
  51 #if HAVE_AVX2_EXTERNAL
  52         if (EXTERNAL_AVX2_FAST(cpu_flags))
  53             return scene_sad_avx2;
  54 #endif
  55         if (EXTERNAL_SSE2(cpu_flags))
  56             return scene_sad_sse2;
  57     }
  58 #endif
  59     return NULL;
  60 }