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

   1 /*
   2  * XVideo Motion Compensation
   3  * Copyright (c) 2003 Ivan Kalvachev
   4  *
   5  * This file is part of FFmpeg.
   6  *
   7  * FFmpeg is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * FFmpeg is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with FFmpeg; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 #include <limits.h>
  23
  24 #include "avcodec.h"
  25 #include "dsputil.h"
  26 #include "mpegvideo.h"
  27
  28 #undef NDEBUG
  29 #include <assert.h>
  30
  31 #include "xvmc.h"
  32
  33 //set s->block
  34 void ff_xvmc_init_block(MpegEncContext *s)
  35 {
  36     struct xvmc_render_state *render;
  37     render = (struct xvmc_render_state*)s->current_picture.data[2];
  38     assert(render);
  39     if (!render || render->magic != AV_XVMC_RENDER_MAGIC) {
  40         assert(0);
  41         return; // make sure that this is a render packet
  42     }
  43     s->block = (DCTELEM *)(render->data_blocks + render->next_free_data_block_num * 64);
  44 }
  45
  46 void ff_xvmc_pack_pblocks(MpegEncContext *s, int cbp)
  47 {
  48     int i, j;
  49     const int mb_block_count = 4 + (1 << s->chroma_format);
  50
  51     j = 0;
  52     cbp <<= 12-mb_block_count;
  53     for (i = 0; i < mb_block_count; i++) {
  54         if (cbp & (1 << 11))
  55             s->pblocks[i] = (short *)(&s->block[j++]);
  56         else
  57             s->pblocks[i] = NULL;
  58         cbp+=cbp;
  59     }
  60 }
  61
  62 // These functions should be called on every new field and/or frame.
  63 // They should be safe if they are called a few times for the same field!
  64 int ff_xvmc_field_start(MpegEncContext*s, AVCodecContext *avctx)
  65 {
  66     struct xvmc_render_state *render, *last, *next;
  67
  68     assert(avctx);
  69
  70     render = (struct xvmc_render_state*)s->current_picture.data[2];
  71     assert(render);
  72     if (!render || render->magic != AV_XVMC_RENDER_MAGIC)
  73         return -1; // make sure that this is a render packet
  74
  75     render->picture_structure = s->picture_structure;
  76     render->flags             = s->first_field ? 0 : XVMC_SECOND_FIELD;
  77
  78     // make sure that all data is drawn by XVMC_end_frame
  79     assert(render->filled_mv_blocks_num == 0);
  80
  81     render->p_future_surface = NULL;
  82     render->p_past_surface   = NULL;
  83
  84     switch(s->pict_type) {
  85         case  FF_I_TYPE:
  86             return 0; // no prediction from other frames
  87         case  FF_B_TYPE:
  88             next = (struct xvmc_render_state*)s->next_picture.data[2];
  89             assert(next);
  90             if (!next)
  91                 return -1;
  92             if (next->magic != AV_XVMC_RENDER_MAGIC)
  93                 return -1;
  94             render->p_future_surface = next->p_surface;
  95             // no return here, going to set forward prediction
  96         case  FF_P_TYPE:
  97             last = (struct xvmc_render_state*)s->last_picture.data[2];
  98             if (!last) // && !s->first_field)
  99                 last = render; // predict second field from the first
 100             if (last->magic != AV_XVMC_RENDER_MAGIC)
 101                 return -1;
 102             render->p_past_surface = last->p_surface;
 103             return 0;
 104     }
 105
 106 return -1;
 107 }
 108
 109 void ff_xvmc_field_end(MpegEncContext *s)
 110 {
 111     struct xvmc_render_state *render;
 112     render = (struct xvmc_render_state*)s->current_picture.data[2];
 113     assert(render);
 114
 115     if (render->filled_mv_blocks_num > 0)
 116         ff_draw_horiz_band(s,0,0);
 117 }
 118
 119 void ff_xvmc_decode_mb(MpegEncContext *s)
 120 {
 121     XvMCMacroBlock *mv_block;
 122     struct xvmc_render_state *render;
 123     int i, cbp, blocks_per_mb;
 124
 125     const int mb_xy = s->mb_y * s->mb_stride + s->mb_x;
 126
 127
 128     if (s->encoding) {
 129         av_log(s->avctx, AV_LOG_ERROR, "XVMC doesn't support encoding!!!\n");
 130         return;
 131     }
 132
 133     // from MPV_decode_mb(), update DC predictors for P macroblocks
 134     if (!s->mb_intra) {
 135         s->last_dc[0] =
 136         s->last_dc[1] =
 137         s->last_dc[2] =  128 << s->intra_dc_precision;
 138     }
 139
 140     // MC doesn't skip blocks
 141     s->mb_skipped = 0;
 142
 143
 144     // Do I need to export quant when I could not perform postprocessing?
 145     // Anyway, it doesn't hurt.
 146     s->current_picture.qscale_table[mb_xy] = s->qscale;
 147
 148     // start of XVMC-specific code
 149     render = (struct xvmc_render_state*)s->current_picture.data[2];
 150     assert(render);
 151     assert(render->magic == AV_XVMC_RENDER_MAGIC);
 152     assert(render->mv_blocks);
 153
 154     // take the next free macroblock
 155     mv_block = &render->mv_blocks[render->start_mv_blocks_num +
 156                                   render->filled_mv_blocks_num ];
 157
 158     mv_block->x        = s->mb_x;
 159     mv_block->y        = s->mb_y;
 160     mv_block->dct_type = s->interlaced_dct; // XVMC_DCT_TYPE_FRAME/FIELD;
 161     if (s->mb_intra) {
 162         mv_block->macroblock_type = XVMC_MB_TYPE_INTRA; // no MC, all done
 163     } else {
 164         mv_block->macroblock_type = XVMC_MB_TYPE_PATTERN;
 165
 166         if (s->mv_dir & MV_DIR_FORWARD) {
 167             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_FORWARD;
 168             //pmv[n][dir][xy]=mv[dir][n][xy]
 169             mv_block->PMV[0][0][0] = s->mv[0][0][0];
 170             mv_block->PMV[0][0][1] = s->mv[0][0][1];
 171             mv_block->PMV[1][0][0] = s->mv[0][1][0];
 172             mv_block->PMV[1][0][1] = s->mv[0][1][1];
 173         }
 174         if (s->mv_dir & MV_DIR_BACKWARD) {
 175             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_BACKWARD;
 176             mv_block->PMV[0][1][0] = s->mv[1][0][0];
 177             mv_block->PMV[0][1][1] = s->mv[1][0][1];
 178             mv_block->PMV[1][1][0] = s->mv[1][1][0];
 179             mv_block->PMV[1][1][1] = s->mv[1][1][1];
 180         }
 181
 182         switch(s->mv_type){
 183             case  MV_TYPE_16X16:
 184                 mv_block->motion_type = XVMC_PREDICTION_FRAME;
 185                 break;
 186             case  MV_TYPE_16X8:
 187                 mv_block->motion_type = XVMC_PREDICTION_16x8;
 188                 break;
 189             case  MV_TYPE_FIELD:
 190                 mv_block->motion_type = XVMC_PREDICTION_FIELD;
 191                 if (s->picture_structure == PICT_FRAME) {
 192                     mv_block->PMV[0][0][1] <<= 1;
 193                     mv_block->PMV[1][0][1] <<= 1;
 194                     mv_block->PMV[0][1][1] <<= 1;
 195                     mv_block->PMV[1][1][1] <<= 1;
 196                 }
 197                 break;
 198             case  MV_TYPE_DMV:
 199                 mv_block->motion_type = XVMC_PREDICTION_DUAL_PRIME;
 200                 if (s->picture_structure == PICT_FRAME) {
 201
 202                     mv_block->PMV[0][0][0] = s->mv[0][0][0]; // top from top
 203                     mv_block->PMV[0][0][1] = s->mv[0][0][1]<<1;
 204
 205                     mv_block->PMV[0][1][0] = s->mv[0][0][0]; // bottom from bottom
 206                     mv_block->PMV[0][1][1] = s->mv[0][0][1]<<1;
 207
 208                     mv_block->PMV[1][0][0] = s->mv[0][2][0]; // dmv00, top from bottom
 209                     mv_block->PMV[1][0][1] = s->mv[0][2][1]<<1; // dmv01
 210
 211                     mv_block->PMV[1][1][0] = s->mv[0][3][0]; // dmv10, bottom from top
 212                     mv_block->PMV[1][1][1] = s->mv[0][3][1]<<1; // dmv11
 213
 214                 } else {
 215                     mv_block->PMV[0][1][0] = s->mv[0][2][0]; // dmv00
 216                     mv_block->PMV[0][1][1] = s->mv[0][2][1]; // dmv01
 217                 }
 218                 break;
 219             default:
 220                 assert(0);
 221         }
 222
 223         mv_block->motion_vertical_field_select = 0;
 224
 225         // set correct field references
 226         if (s->mv_type == MV_TYPE_FIELD || s->mv_type == MV_TYPE_16X8) {
 227             mv_block->motion_vertical_field_select |= s->field_select[0][0];
 228             mv_block->motion_vertical_field_select |= s->field_select[1][0] << 1;
 229             mv_block->motion_vertical_field_select |= s->field_select[0][1] << 2;
 230             mv_block->motion_vertical_field_select |= s->field_select[1][1] << 3;
 231         }
 232     } // !intra
 233     // time to handle data blocks
 234     mv_block->index = render->next_free_data_block_num;
 235
 236     blocks_per_mb = 6;
 237     if (s->chroma_format >= 2) {
 238         blocks_per_mb = 4 + (1 << s->chroma_format);
 239     }
 240
 241     // calculate cbp
 242     cbp = 0;
 243     for (i = 0; i < blocks_per_mb; i++) {
 244         cbp += cbp;
 245         if (s->block_last_index[i] >= 0)
 246             cbp++;
 247     }
 248
 249     if (s->flags & CODEC_FLAG_GRAY) {
 250         if (s->mb_intra) { // intra frames are always full chroma blocks
 251             for (i = 4; i < blocks_per_mb; i++) {
 252                 memset(s->pblocks[i], 0, sizeof(short)*8*8); // so we need to clear them
 253                 if (!render->unsigned_intra)
 254                     s->pblocks[i][0] = 1 << 10;
 255             }
 256         } else {
 257             cbp &= 0xf << (blocks_per_mb - 4);
 258             blocks_per_mb = 4; // luminance blocks only
 259         }
 260     }
 261     mv_block->coded_block_pattern = cbp;
 262     if (cbp == 0)
 263         mv_block->macroblock_type &= ~XVMC_MB_TYPE_PATTERN;
 264
 265     for (i = 0; i < blocks_per_mb; i++) {
 266         if (s->block_last_index[i] >= 0) {
 267             // I do not have unsigned_intra MOCO to test, hope it is OK.
 268             if (s->mb_intra && (render->idct || (!render->idct && !render->unsigned_intra)))
 269                 s->pblocks[i][0] -= 1 << 10;
 270             if (!render->idct) {
 271                 s->dsp.idct(s->pblocks[i]);
 272                 /* It is unclear if MC hardware requires pixel diff values to be
 273                  * in the range [-255;255]. TODO: Clipping if such hardware is
 274                  * ever found. As of now it would only be an unnecessary
 275                  * slowdown. */
 276             }
 277             // copy blocks only if the codec doesn't support pblocks reordering
 278             if (s->avctx->xvmc_acceleration == 1) {
 279                 memcpy(&render->data_blocks[render->next_free_data_block_num*64],
 280                        s->pblocks[i],sizeof(short)*8*8);
 281             }
 282             render->next_free_data_block_num++;
 283         }
 284     }
 285     render->filled_mv_blocks_num++;
 286
 287     assert(render->filled_mv_blocks_num     <= render->total_number_of_mv_blocks);
 288     assert(render->next_free_data_block_num <= render->total_number_of_data_blocks);
 289
 290
 291     if (render->filled_mv_blocks_num >= render->total_number_of_mv_blocks)
 292         ff_draw_horiz_band(s, 0, 0);
 293 }