git.sesse.net Git - ffmpeg/blob - libavcodec/idctdsp.c

   1 /*
   2  * This file is part of Libav.
   3  *
   4  * Libav 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  * Libav 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 Libav; if not, write to the Free Software
  16  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  17  */
  18
  19 #include "config.h"
  20 #include "libavutil/attributes.h"
  21 #include "libavutil/common.h"
  22 #include "avcodec.h"
  23 #include "dct.h"
  24 #include "faanidct.h"
  25 #include "idctdsp.h"
  26 #include "simple_idct.h"
  27
  28 av_cold void ff_init_scantable(uint8_t *permutation, ScanTable *st,
  29                                const uint8_t *src_scantable)
  30 {
  31     int i, end;
  32
  33     st->scantable = src_scantable;
  34
  35     for (i = 0; i < 64; i++) {
  36         int j = src_scantable[i];
  37         st->permutated[i] = permutation[j];
  38     }
  39
  40     end = -1;
  41     for (i = 0; i < 64; i++) {
  42         int j = st->permutated[i];
  43         if (j > end)
  44             end = j;
  45         st->raster_end[i] = end;
  46     }
  47 }
  48
  49 av_cold void ff_init_scantable_permutation(uint8_t *idct_permutation,
  50                                            enum idct_permutation_type perm_type)
  51 {
  52     int i;
  53
  54     if (ARCH_X86)
  55         if (ff_init_scantable_permutation_x86(idct_permutation,
  56                                               perm_type))
  57             return;
  58
  59     switch (perm_type) {
  60     case FF_IDCT_PERM_NONE:
  61         for (i = 0; i < 64; i++)
  62             idct_permutation[i] = i;
  63         break;
  64     case FF_IDCT_PERM_LIBMPEG2:
  65         for (i = 0; i < 64; i++)
  66             idct_permutation[i] = (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2);
  67         break;
  68     case FF_IDCT_PERM_TRANSPOSE:
  69         for (i = 0; i < 64; i++)
  70             idct_permutation[i] = ((i & 7) << 3) | (i >> 3);
  71         break;
  72     case FF_IDCT_PERM_PARTTRANS:
  73         for (i = 0; i < 64; i++)
  74             idct_permutation[i] = (i & 0x24) | ((i & 3) << 3) | ((i >> 3) & 3);
  75         break;
  76     default:
  77         av_log(NULL, AV_LOG_ERROR,
  78                "Internal error, IDCT permutation not set\n");
  79     }
  80 }
  81
  82 static void put_pixels_clamped_c(const int16_t *block, uint8_t *restrict pixels,
  83                                  int line_size)
  84 {
  85     int i;
  86
  87     /* read the pixels */
  88     for (i = 0; i < 8; i++) {
  89         pixels[0] = av_clip_uint8(block[0]);
  90         pixels[1] = av_clip_uint8(block[1]);
  91         pixels[2] = av_clip_uint8(block[2]);
  92         pixels[3] = av_clip_uint8(block[3]);
  93         pixels[4] = av_clip_uint8(block[4]);
  94         pixels[5] = av_clip_uint8(block[5]);
  95         pixels[6] = av_clip_uint8(block[6]);
  96         pixels[7] = av_clip_uint8(block[7]);
  97
  98         pixels += line_size;
  99         block  += 8;
 100     }
 101 }
 102
 103 static void put_signed_pixels_clamped_c(const int16_t *block,
 104                                         uint8_t *restrict pixels,
 105                                         int line_size)
 106 {
 107     int i, j;
 108
 109     for (i = 0; i < 8; i++) {
 110         for (j = 0; j < 8; j++) {
 111             if (*block < -128)
 112                 *pixels = 0;
 113             else if (*block > 127)
 114                 *pixels = 255;
 115             else
 116                 *pixels = (uint8_t) (*block + 128);
 117             block++;
 118             pixels++;
 119         }
 120         pixels += (line_size - 8);
 121     }
 122 }
 123
 124 static void add_pixels_clamped_c(const int16_t *block, uint8_t *restrict pixels,
 125                                  int line_size)
 126 {
 127     int i;
 128
 129     /* read the pixels */
 130     for (i = 0; i < 8; i++) {
 131         pixels[0] = av_clip_uint8(pixels[0] + block[0]);
 132         pixels[1] = av_clip_uint8(pixels[1] + block[1]);
 133         pixels[2] = av_clip_uint8(pixels[2] + block[2]);
 134         pixels[3] = av_clip_uint8(pixels[3] + block[3]);
 135         pixels[4] = av_clip_uint8(pixels[4] + block[4]);
 136         pixels[5] = av_clip_uint8(pixels[5] + block[5]);
 137         pixels[6] = av_clip_uint8(pixels[6] + block[6]);
 138         pixels[7] = av_clip_uint8(pixels[7] + block[7]);
 139         pixels   += line_size;
 140         block    += 8;
 141     }
 142 }
 143
 144 static void jref_idct_put(uint8_t *dest, int line_size, int16_t *block)
 145 {
 146     ff_j_rev_dct(block);
 147     put_pixels_clamped_c(block, dest, line_size);
 148 }
 149
 150 static void jref_idct_add(uint8_t *dest, int line_size, int16_t *block)
 151 {
 152     ff_j_rev_dct(block);
 153     add_pixels_clamped_c(block, dest, line_size);
 154 }
 155
 156 av_cold void ff_idctdsp_init(IDCTDSPContext *c, AVCodecContext *avctx)
 157 {
 158     const unsigned high_bit_depth = avctx->bits_per_raw_sample > 8;
 159
 160     if (avctx->bits_per_raw_sample == 10) {
 161         c->idct_put              = ff_simple_idct_put_10;
 162         c->idct_add              = ff_simple_idct_add_10;
 163         c->idct                  = ff_simple_idct_10;
 164         c->perm_type             = FF_IDCT_PERM_NONE;
 165     } else {
 166         if (avctx->idct_algo == FF_IDCT_INT) {
 167             c->idct_put              = jref_idct_put;
 168             c->idct_add              = jref_idct_add;
 169             c->idct                  = ff_j_rev_dct;
 170             c->perm_type             = FF_IDCT_PERM_LIBMPEG2;
 171         } else if (avctx->idct_algo == FF_IDCT_FAAN) {
 172             c->idct_put              = ff_faanidct_put;
 173             c->idct_add              = ff_faanidct_add;
 174             c->idct                  = ff_faanidct;
 175             c->perm_type             = FF_IDCT_PERM_NONE;
 176         } else { // accurate/default
 177             c->idct_put              = ff_simple_idct_put_8;
 178             c->idct_add              = ff_simple_idct_add_8;
 179             c->idct                  = ff_simple_idct_8;
 180             c->perm_type             = FF_IDCT_PERM_NONE;
 181         }
 182     }
 183
 184     c->put_pixels_clamped        = put_pixels_clamped_c;
 185     c->put_signed_pixels_clamped = put_signed_pixels_clamped_c;
 186     c->add_pixels_clamped        = add_pixels_clamped_c;
 187
 188     if (ARCH_ARM)
 189         ff_idctdsp_init_arm(c, avctx, high_bit_depth);
 190     if (ARCH_PPC)
 191         ff_idctdsp_init_ppc(c, avctx, high_bit_depth);
 192     if (ARCH_X86)
 193         ff_idctdsp_init_x86(c, avctx, high_bit_depth);
 194
 195     ff_init_scantable_permutation(c->idct_permutation,
 196                                   c->perm_type);
 197 }