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 Libav.
   6  *
   7  * Libav 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  * Libav 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 Libav; 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 #include <X11/extensions/XvMC.h>
  24
  25 #include "avcodec.h"
  26 #include "dsputil.h"
  27 #include "mpegvideo.h"
  28
  29 #undef NDEBUG
  30 #include <assert.h>
  31
  32 #include "xvmc.h"
  33 #include "xvmc_internal.h"
  34
  35 /**
  36  * Initialize the block field of the MpegEncContext pointer passed as
  37  * parameter after making sure that the data is not corrupted.
  38  * In order to implement something like direct rendering instead of decoding
  39  * coefficients in s->blocks and then copying them, copy them directly
  40  * into the data_blocks array provided by xvmc.
  41  */
  42 void ff_xvmc_init_block(MpegEncContext *s)
  43 {
  44     struct xvmc_pix_fmt *render = (struct xvmc_pix_fmt*)s->current_picture.f.data[2];
  45     assert(render && render->xvmc_id == AV_XVMC_ID);
  46
  47     s->block = (DCTELEM (*)[64])(render->data_blocks + render->next_free_data_block_num * 64);
  48 }
  49
  50 /**
  51  * Fill individual block pointers, so there are no gaps in the data_block array
  52  * in case not all blocks in the macroblock are coded.
  53  */
  54 void ff_xvmc_pack_pblocks(MpegEncContext *s, int cbp)
  55 {
  56     int i, j = 0;
  57     const int mb_block_count = 4 + (1 << s->chroma_format);
  58
  59     cbp <<= 12-mb_block_count;
  60     for (i = 0; i < mb_block_count; i++) {
  61         if (cbp & (1 << 11))
  62             s->pblocks[i] = &s->block[j++];
  63         else
  64             s->pblocks[i] = NULL;
  65         cbp += cbp;
  66     }
  67 }
  68
  69 /**
  70  * Find and store the surfaces that are used as reference frames.
  71  * This function should be called for every new field and/or frame.
  72  * It should be safe to call the function a few times for the same field.
  73  */
  74 int ff_xvmc_field_start(MpegEncContext *s, AVCodecContext *avctx)
  75 {
  76     struct xvmc_pix_fmt *last, *next, *render = (struct xvmc_pix_fmt*)s->current_picture.f.data[2];
  77     const int mb_block_count = 4 + (1 << s->chroma_format);
  78
  79     assert(avctx);
  80     if (!render || render->xvmc_id != AV_XVMC_ID ||
  81         !render->data_blocks || !render->mv_blocks ||
  82         (unsigned int)render->allocated_mv_blocks   > INT_MAX/(64*6) ||
  83         (unsigned int)render->allocated_data_blocks > INT_MAX/64     ||
  84         !render->p_surface) {
  85         av_log(avctx, AV_LOG_ERROR,
  86                "Render token doesn't look as expected.\n");
  87         return -1; // make sure that this is a render packet
  88     }
  89
  90     if (render->filled_mv_blocks_num) {
  91         av_log(avctx, AV_LOG_ERROR,
  92                "Rendering surface contains %i unprocessed blocks.\n",
  93                render->filled_mv_blocks_num);
  94         return -1;
  95     }
  96     if (render->allocated_mv_blocks   < 1 ||
  97         render->allocated_data_blocks <  render->allocated_mv_blocks*mb_block_count ||
  98         render->start_mv_blocks_num   >= render->allocated_mv_blocks                ||
  99         render->next_free_data_block_num >
 100                         render->allocated_data_blocks -
 101                         mb_block_count*(render->allocated_mv_blocks-render->start_mv_blocks_num)) {
 102         av_log(avctx, AV_LOG_ERROR,
 103                "Rendering surface doesn't provide enough block structures to work with.\n");
 104         return -1;
 105     }
 106
 107     render->picture_structure = s->picture_structure;
 108     render->flags             = s->first_field ? 0 : XVMC_SECOND_FIELD;
 109     render->p_future_surface  = NULL;
 110     render->p_past_surface    = NULL;
 111
 112     switch(s->pict_type) {
 113         case  AV_PICTURE_TYPE_I:
 114             return 0; // no prediction from other frames
 115         case  AV_PICTURE_TYPE_B:
 116             next = (struct xvmc_pix_fmt*)s->next_picture.f.data[2];
 117             if (!next)
 118                 return -1;
 119             if (next->xvmc_id != AV_XVMC_ID)
 120                 return -1;
 121             render->p_future_surface = next->p_surface;
 122             // no return here, going to set forward prediction
 123         case  AV_PICTURE_TYPE_P:
 124             last = (struct xvmc_pix_fmt*)s->last_picture.f.data[2];
 125             if (!last)
 126                 last = render; // predict second field from the first
 127             if (last->xvmc_id != AV_XVMC_ID)
 128                 return -1;
 129             render->p_past_surface = last->p_surface;
 130             return 0;
 131     }
 132
 133 return -1;
 134 }
 135
 136 /**
 137  * Complete frame/field rendering by passing any remaining blocks.
 138  * Normally ff_draw_horiz_band() is called for each slice, however,
 139  * some leftover blocks, for example from error_resilience(), may remain.
 140  * It should be safe to call the function a few times for the same field.
 141  */
 142 void ff_xvmc_field_end(MpegEncContext *s)
 143 {
 144     struct xvmc_pix_fmt *render = (struct xvmc_pix_fmt*)s->current_picture.f.data[2];
 145     assert(render);
 146
 147     if (render->filled_mv_blocks_num > 0)
 148         ff_draw_horiz_band(s, 0, 0);
 149 }
 150
 151 /**
 152  * Synthesize the data needed by XvMC to render one macroblock of data.
 153  * Fill all relevant fields, if necessary do IDCT.
 154  */
 155 void ff_xvmc_decode_mb(MpegEncContext *s)
 156 {
 157     XvMCMacroBlock *mv_block;
 158     struct xvmc_pix_fmt *render;
 159     int i, cbp, blocks_per_mb;
 160
 161     const int mb_xy = s->mb_y * s->mb_stride + s->mb_x;
 162
 163
 164     if (s->encoding) {
 165         av_log(s->avctx, AV_LOG_ERROR, "XVMC doesn't support encoding!!!\n");
 166         return;
 167     }
 168
 169     // from MPV_decode_mb(), update DC predictors for P macroblocks
 170     if (!s->mb_intra) {
 171         s->last_dc[0] =
 172         s->last_dc[1] =
 173         s->last_dc[2] =  128 << s->intra_dc_precision;
 174     }
 175
 176     // MC doesn't skip blocks
 177     s->mb_skipped = 0;
 178
 179
 180     // Do I need to export quant when I could not perform postprocessing?
 181     // Anyway, it doesn't hurt.
 182     s->current_picture.f.qscale_table[mb_xy] = s->qscale;
 183
 184     // start of XVMC-specific code
 185     render = (struct xvmc_pix_fmt*)s->current_picture.f.data[2];
 186     assert(render);
 187     assert(render->xvmc_id == AV_XVMC_ID);
 188     assert(render->mv_blocks);
 189
 190     // take the next free macroblock
 191     mv_block = &render->mv_blocks[render->start_mv_blocks_num +
 192                                   render->filled_mv_blocks_num];
 193
 194     mv_block->x        = s->mb_x;
 195     mv_block->y        = s->mb_y;
 196     mv_block->dct_type = s->interlaced_dct; // XVMC_DCT_TYPE_FRAME/FIELD;
 197     if (s->mb_intra) {
 198         mv_block->macroblock_type = XVMC_MB_TYPE_INTRA; // no MC, all done
 199     } else {
 200         mv_block->macroblock_type = XVMC_MB_TYPE_PATTERN;
 201
 202         if (s->mv_dir & MV_DIR_FORWARD) {
 203             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_FORWARD;
 204             // PMV[n][dir][xy] = mv[dir][n][xy]
 205             mv_block->PMV[0][0][0] = s->mv[0][0][0];
 206             mv_block->PMV[0][0][1] = s->mv[0][0][1];
 207             mv_block->PMV[1][0][0] = s->mv[0][1][0];
 208             mv_block->PMV[1][0][1] = s->mv[0][1][1];
 209         }
 210         if (s->mv_dir & MV_DIR_BACKWARD) {
 211             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_BACKWARD;
 212             mv_block->PMV[0][1][0] = s->mv[1][0][0];
 213             mv_block->PMV[0][1][1] = s->mv[1][0][1];
 214             mv_block->PMV[1][1][0] = s->mv[1][1][0];
 215             mv_block->PMV[1][1][1] = s->mv[1][1][1];
 216         }
 217
 218         switch(s->mv_type) {
 219             case  MV_TYPE_16X16:
 220                 mv_block->motion_type = XVMC_PREDICTION_FRAME;
 221                 break;
 222             case  MV_TYPE_16X8:
 223                 mv_block->motion_type = XVMC_PREDICTION_16x8;
 224                 break;
 225             case  MV_TYPE_FIELD:
 226                 mv_block->motion_type = XVMC_PREDICTION_FIELD;
 227                 if (s->picture_structure == PICT_FRAME) {
 228                     mv_block->PMV[0][0][1] <<= 1;
 229                     mv_block->PMV[1][0][1] <<= 1;
 230                     mv_block->PMV[0][1][1] <<= 1;
 231                     mv_block->PMV[1][1][1] <<= 1;
 232                 }
 233                 break;
 234             case  MV_TYPE_DMV:
 235                 mv_block->motion_type = XVMC_PREDICTION_DUAL_PRIME;
 236                 if (s->picture_structure == PICT_FRAME) {
 237
 238                     mv_block->PMV[0][0][0] = s->mv[0][0][0];      // top from top
 239                     mv_block->PMV[0][0][1] = s->mv[0][0][1] << 1;
 240
 241                     mv_block->PMV[0][1][0] = s->mv[0][0][0];      // bottom from bottom
 242                     mv_block->PMV[0][1][1] = s->mv[0][0][1] << 1;
 243
 244                     mv_block->PMV[1][0][0] = s->mv[0][2][0];      // dmv00, top from bottom
 245                     mv_block->PMV[1][0][1] = s->mv[0][2][1] << 1; // dmv01
 246
 247                     mv_block->PMV[1][1][0] = s->mv[0][3][0];      // dmv10, bottom from top
 248                     mv_block->PMV[1][1][1] = s->mv[0][3][1] << 1; // dmv11
 249
 250                 } else {
 251                     mv_block->PMV[0][1][0] = s->mv[0][2][0];      // dmv00
 252                     mv_block->PMV[0][1][1] = s->mv[0][2][1];      // dmv01
 253                 }
 254                 break;
 255             default:
 256                 assert(0);
 257         }
 258
 259         mv_block->motion_vertical_field_select = 0;
 260
 261         // set correct field references
 262         if (s->mv_type == MV_TYPE_FIELD || s->mv_type == MV_TYPE_16X8) {
 263             mv_block->motion_vertical_field_select |= s->field_select[0][0];
 264             mv_block->motion_vertical_field_select |= s->field_select[1][0] << 1;
 265             mv_block->motion_vertical_field_select |= s->field_select[0][1] << 2;
 266             mv_block->motion_vertical_field_select |= s->field_select[1][1] << 3;
 267         }
 268     } // !intra
 269     // time to handle data blocks
 270     mv_block->index = render->next_free_data_block_num;
 271
 272     blocks_per_mb = 6;
 273     if (s->chroma_format >= 2) {
 274         blocks_per_mb = 4 + (1 << s->chroma_format);
 275     }
 276
 277     // calculate cbp
 278     cbp = 0;
 279     for (i = 0; i < blocks_per_mb; i++) {
 280         cbp += cbp;
 281         if (s->block_last_index[i] >= 0)
 282             cbp++;
 283     }
 284
 285     if (s->flags & CODEC_FLAG_GRAY) {
 286         if (s->mb_intra) {                                   // intra frames are always full chroma blocks
 287             for (i = 4; i < blocks_per_mb; i++) {
 288                 memset(s->pblocks[i], 0, sizeof(*s->pblocks[i]));  // so we need to clear them
 289                 if (!render->unsigned_intra)
 290                     *s->pblocks[i][0] = 1 << 10;
 291             }
 292         } else {
 293             cbp &= 0xf << (blocks_per_mb - 4);
 294             blocks_per_mb = 4;                               // luminance blocks only
 295         }
 296     }
 297     mv_block->coded_block_pattern = cbp;
 298     if (cbp == 0)
 299         mv_block->macroblock_type &= ~XVMC_MB_TYPE_PATTERN;
 300
 301     for (i = 0; i < blocks_per_mb; i++) {
 302         if (s->block_last_index[i] >= 0) {
 303             // I do not have unsigned_intra MOCO to test, hope it is OK.
 304             if (s->mb_intra && (render->idct || !render->unsigned_intra))
 305                 *s->pblocks[i][0] -= 1 << 10;
 306             if (!render->idct) {
 307                 s->dsp.idct(*s->pblocks[i]);
 308                 /* It is unclear if MC hardware requires pixel diff values to be
 309                  * in the range [-255;255]. TODO: Clipping if such hardware is
 310                  * ever found. As of now it would only be an unnecessary
 311                  * slowdown. */
 312             }
 313             // copy blocks only if the codec doesn't support pblocks reordering
 314             if (s->avctx->xvmc_acceleration == 1) {
 315                 memcpy(&render->data_blocks[render->next_free_data_block_num*64],
 316                        s->pblocks[i], sizeof(*s->pblocks[i]));
 317             }
 318             render->next_free_data_block_num++;
 319         }
 320     }
 321     render->filled_mv_blocks_num++;
 322
 323     assert(render->filled_mv_blocks_num     <= render->allocated_mv_blocks);
 324     assert(render->next_free_data_block_num <= render->allocated_data_blocks);
 325     /* The above conditions should not be able to fail as long as this function
 326      * is used and the following 'if ()' automatically calls a callback to free
 327      * blocks. */
 328
 329
 330     if (render->filled_mv_blocks_num == render->allocated_mv_blocks)
 331         ff_draw_horiz_band(s, 0, 0);
 332 }