git.sesse.net Git - ffmpeg/blob - libavcodec/x86/dwt.c

   1 /*
   2  * MMX optimized discrete wavelet transform
   3  * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
   4  * Copyright (c) 2010 David Conrad
   5  *
   6  * This file is part of FFmpeg.
   7  *
   8  * FFmpeg is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2.1 of the License, or (at your option) any later version.
  12  *
  13  * FFmpeg is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * Lesser General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU Lesser General Public
  19  * License along with FFmpeg; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21  */
  22
  23 #include "libavutil/x86_cpu.h"
  24 #include "dsputil_mmx.h"
  25 #include "dwt.h"
  26
  27 #define COMPOSE_VERTICAL(ext, align) \
  28 void ff_vertical_compose53iL0##ext(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width); \
  29 void ff_vertical_compose_dirac53iH0##ext(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width); \
  30 void ff_vertical_compose_dd137iL0##ext(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, int width); \
  31 void ff_vertical_compose_dd97iH0##ext(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, int width); \
  32 void ff_vertical_compose_haar##ext(IDWTELEM *b0, IDWTELEM *b1, int width); \
  33 \
  34 static void vertical_compose53iL0##ext(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width) \
  35 { \
  36     int i, width_align = width&~(align-1); \
  37 \
  38     for(i=width_align; i<width; i++) \
  39         b1[i] = COMPOSE_53iL0(b0[i], b1[i], b2[i]); \
  40 \
  41     ff_vertical_compose53iL0##ext(b0, b1, b2, width_align); \
  42 } \
  43 \
  44 static void vertical_compose_dirac53iH0##ext(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width) \
  45 { \
  46     int i, width_align = width&~(align-1); \
  47 \
  48     for(i=width_align; i<width; i++) \
  49         b1[i] = COMPOSE_DIRAC53iH0(b0[i], b1[i], b2[i]); \
  50 \
  51     ff_vertical_compose_dirac53iH0##ext(b0, b1, b2, width_align); \
  52 } \
  53 \
  54 static void vertical_compose_dd137iL0##ext(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, \
  55                                            IDWTELEM *b3, IDWTELEM *b4, int width) \
  56 { \
  57     int i, width_align = width&~(align-1); \
  58 \
  59     for(i=width_align; i<width; i++) \
  60         b2[i] = COMPOSE_DD137iL0(b0[i], b1[i], b2[i], b3[i], b4[i]); \
  61 \
  62     ff_vertical_compose_dd137iL0##ext(b0, b1, b2, b3, b4, width_align); \
  63 } \
  64 \
  65 static void vertical_compose_dd97iH0##ext(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, \
  66                                           IDWTELEM *b3, IDWTELEM *b4, int width) \
  67 { \
  68     int i, width_align = width&~(align-1); \
  69 \
  70     for(i=width_align; i<width; i++) \
  71         b2[i] = COMPOSE_DD97iH0(b0[i], b1[i], b2[i], b3[i], b4[i]); \
  72 \
  73     ff_vertical_compose_dd97iH0##ext(b0, b1, b2, b3, b4, width_align); \
  74 } \
  75 static void vertical_compose_haar##ext(IDWTELEM *b0, IDWTELEM *b1, int width) \
  76 { \
  77     int i, width_align = width&~(align-1); \
  78 \
  79     for(i=width_align; i<width; i++) { \
  80         b0[i] = COMPOSE_HAARiL0(b0[i], b1[i]); \
  81         b1[i] = COMPOSE_HAARiH0(b1[i], b0[i]); \
  82     } \
  83 \
  84     ff_vertical_compose_haar##ext(b0, b1, width_align); \
  85 } \
  86 \
  87
  88 #if HAVE_YASM
  89 #if !ARCH_X86_64
  90 COMPOSE_VERTICAL(_mmx, 4)
  91 #endif
  92 COMPOSE_VERTICAL(_sse2, 8)
  93 #endif
  94
  95
  96 void ff_horizontal_compose_dd97i_ssse3(IDWTELEM *b, IDWTELEM *tmp, int w);
  97
  98 void ff_horizontal_compose_haar0i_mmx(IDWTELEM *b, IDWTELEM *tmp, int w);
  99 void ff_horizontal_compose_haar1i_mmx(IDWTELEM *b, IDWTELEM *tmp, int w);
 100 void ff_horizontal_compose_haar0i_sse2(IDWTELEM *b, IDWTELEM *tmp, int w);
 101 void ff_horizontal_compose_haar1i_sse2(IDWTELEM *b, IDWTELEM *tmp, int w);
 102
 103 void ff_horizontal_compose_dd97i_end_c(IDWTELEM *b, IDWTELEM *tmp, int w2, int x)
 104 {
 105     for (; x < w2; x++) {
 106         b[2*x  ] = (tmp[x] + 1)>>1;
 107         b[2*x+1] = (COMPOSE_DD97iH0(tmp[x-1], tmp[x], b[x+w2], tmp[x+1], tmp[x+2]) + 1)>>1;
 108     }
 109 }
 110
 111 void ff_horizontal_compose_haar0i_end_c(IDWTELEM *b, IDWTELEM *tmp, int w2, int x)
 112 {
 113     for (; x < w2; x++) {
 114         b[2*x  ] = tmp[x];
 115         b[2*x+1] = COMPOSE_HAARiH0(b[x+w2], tmp[x]);
 116     }
 117 }
 118
 119 void ff_horizontal_compose_haar1i_end_c(IDWTELEM *b, IDWTELEM *tmp, int w2, int x)
 120 {
 121     for (; x < w2; x++) {
 122         b[2*x  ] = (tmp[x] + 1)>>1;
 123         b[2*x+1] = (COMPOSE_HAARiH0(b[x+w2], tmp[x]) + 1)>>1;
 124     }
 125 }
 126
 127 void ff_spatial_idwt_init_mmx(DWTContext *d, enum dwt_type type)
 128 {
 129 #if HAVE_YASM
 130   int mm_flags = av_get_cpu_flags();;
 131
 132 #if !ARCH_X86_64
 133     if (!(mm_flags & AV_CPU_FLAG_MMX))
 134         return;
 135
 136     switch (type) {
 137     case DWT_DIRAC_DD9_7:
 138         d->vertical_compose_l0 = vertical_compose53iL0_mmx;
 139         d->vertical_compose_h0 = vertical_compose_dd97iH0_mmx;
 140         break;
 141     case DWT_DIRAC_LEGALL5_3:
 142         d->vertical_compose_l0 = vertical_compose53iL0_mmx;
 143         d->vertical_compose_h0 = vertical_compose_dirac53iH0_mmx;
 144         break;
 145     case DWT_DIRAC_DD13_7:
 146         d->vertical_compose_l0 = vertical_compose_dd137iL0_mmx;
 147         d->vertical_compose_h0 = vertical_compose_dd97iH0_mmx;
 148         break;
 149     case DWT_DIRAC_HAAR0:
 150         d->vertical_compose   = vertical_compose_haar_mmx;
 151         d->horizontal_compose = ff_horizontal_compose_haar0i_mmx;
 152         break;
 153     case DWT_DIRAC_HAAR1:
 154         d->vertical_compose   = vertical_compose_haar_mmx;
 155         d->horizontal_compose = ff_horizontal_compose_haar1i_mmx;
 156         break;
 157     }
 158 #endif
 159
 160     if (!(mm_flags & AV_CPU_FLAG_SSE2))
 161         return;
 162
 163     switch (type) {
 164     case DWT_DIRAC_DD9_7:
 165         d->vertical_compose_l0 = vertical_compose53iL0_sse2;
 166         d->vertical_compose_h0 = vertical_compose_dd97iH0_sse2;
 167         break;
 168     case DWT_DIRAC_LEGALL5_3:
 169         d->vertical_compose_l0 = vertical_compose53iL0_sse2;
 170         d->vertical_compose_h0 = vertical_compose_dirac53iH0_sse2;
 171         break;
 172     case DWT_DIRAC_DD13_7:
 173         d->vertical_compose_l0 = vertical_compose_dd137iL0_sse2;
 174         d->vertical_compose_h0 = vertical_compose_dd97iH0_sse2;
 175         break;
 176     case DWT_DIRAC_HAAR0:
 177         d->vertical_compose   = vertical_compose_haar_sse2;
 178         d->horizontal_compose = ff_horizontal_compose_haar0i_sse2;
 179         break;
 180     case DWT_DIRAC_HAAR1:
 181         d->vertical_compose   = vertical_compose_haar_sse2;
 182         d->horizontal_compose = ff_horizontal_compose_haar1i_sse2;
 183         break;
 184     }
 185
 186     if (!(mm_flags & AV_CPU_FLAG_SSSE3))
 187         return;
 188
 189     switch (type) {
 190     case DWT_DIRAC_DD9_7:
 191 //MMXDISABLED        d->horizontal_compose = ff_horizontal_compose_dd97i_ssse3;
 192         break;
 193     }
 194 #endif // HAVE_YASM
 195 }