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 #include <X11/extensions/XvMC.h>
  24
  25 #include "avcodec.h"
  26 #include "mpegvideo.h"
  27
  28 #undef NDEBUG
  29 #include <assert.h>
  30
  31 #include "xvmc.h"
  32 #include "xvmc_internal.h"
  33 #include "version.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 = (int16_t (*)[64])(render->data_blocks + render->next_free_data_block_num * 64);
  48 }
  49
  50 static void exchange_uv(MpegEncContext *s)
  51 {
  52     int16_t (*tmp)[64];
  53
  54     tmp           = s->pblocks[4];
  55     s->pblocks[4] = s->pblocks[5];
  56     s->pblocks[5] = tmp;
  57 }
  58
  59 /**
  60  * Fill individual block pointers, so there are no gaps in the data_block array
  61  * in case not all blocks in the macroblock are coded.
  62  */
  63 void ff_xvmc_pack_pblocks(MpegEncContext *s, int cbp)
  64 {
  65     int i, j = 0;
  66     const int mb_block_count = 4 + (1 << s->chroma_format);
  67
  68     cbp <<= 12-mb_block_count;
  69     for (i = 0; i < mb_block_count; i++) {
  70         if (cbp & (1 << 11))
  71             s->pblocks[i] = &s->block[j++];
  72         else
  73             s->pblocks[i] = NULL;
  74         cbp += cbp;
  75     }
  76     if (s->swap_uv) {
  77         exchange_uv(s);
  78     }
  79 }
  80
  81 /**
  82  * Find and store the surfaces that are used as reference frames.
  83  * This function should be called for every new field and/or frame.
  84  * It should be safe to call the function a few times for the same field.
  85  */
  86 static int ff_xvmc_field_start(AVCodecContext *avctx, const uint8_t *buf, uint32_t buf_size)
  87 {
  88     struct MpegEncContext *s = avctx->priv_data;
  89     struct xvmc_pix_fmt *last, *next, *render = (struct xvmc_pix_fmt*)s->current_picture.f.data[2];
  90     const int mb_block_count = 4 + (1 << s->chroma_format);
  91
  92     assert(avctx);
  93     if (!render || render->xvmc_id != AV_XVMC_ID ||
  94         !render->data_blocks || !render->mv_blocks ||
  95         (unsigned int)render->allocated_mv_blocks   > INT_MAX/(64*6) ||
  96         (unsigned int)render->allocated_data_blocks > INT_MAX/64     ||
  97         !render->p_surface) {
  98         av_log(avctx, AV_LOG_ERROR,
  99                "Render token doesn't look as expected.\n");
 100         return -1; // make sure that this is a render packet
 101     }
 102
 103     if (render->filled_mv_blocks_num) {
 104         av_log(avctx, AV_LOG_ERROR,
 105                "Rendering surface contains %i unprocessed blocks.\n",
 106                render->filled_mv_blocks_num);
 107         return -1;
 108     }
 109     if (render->allocated_mv_blocks   < 1 ||
 110         render->allocated_data_blocks <  render->allocated_mv_blocks*mb_block_count ||
 111         render->start_mv_blocks_num   >= render->allocated_mv_blocks                ||
 112         render->next_free_data_block_num >
 113                         render->allocated_data_blocks -
 114                         mb_block_count*(render->allocated_mv_blocks-render->start_mv_blocks_num)) {
 115         av_log(avctx, AV_LOG_ERROR,
 116                "Rendering surface doesn't provide enough block structures to work with.\n");
 117         return -1;
 118     }
 119
 120     render->picture_structure = s->picture_structure;
 121     render->flags             = s->first_field ? 0 : XVMC_SECOND_FIELD;
 122     render->p_future_surface  = NULL;
 123     render->p_past_surface    = NULL;
 124
 125     switch(s->pict_type) {
 126         case  AV_PICTURE_TYPE_I:
 127             return 0; // no prediction from other frames
 128         case  AV_PICTURE_TYPE_B:
 129             next = (struct xvmc_pix_fmt*)s->next_picture.f.data[2];
 130             if (!next)
 131                 return -1;
 132             if (next->xvmc_id != AV_XVMC_ID)
 133                 return -1;
 134             render->p_future_surface = next->p_surface;
 135             // no return here, going to set forward prediction
 136         case  AV_PICTURE_TYPE_P:
 137             last = (struct xvmc_pix_fmt*)s->last_picture.f.data[2];
 138             if (!last)
 139                 last = render; // predict second field from the first
 140             if (last->xvmc_id != AV_XVMC_ID)
 141                 return -1;
 142             render->p_past_surface = last->p_surface;
 143             return 0;
 144     }
 145
 146 return -1;
 147 }
 148
 149 /**
 150  * Complete frame/field rendering by passing any remaining blocks.
 151  * Normally ff_draw_horiz_band() is called for each slice, however,
 152  * some leftover blocks, for example from error_resilience(), may remain.
 153  * It should be safe to call the function a few times for the same field.
 154  */
 155 static int ff_xvmc_field_end(AVCodecContext *avctx)
 156 {
 157     struct MpegEncContext *s = avctx->priv_data;
 158     struct xvmc_pix_fmt *render = (struct xvmc_pix_fmt*)s->current_picture.f.data[2];
 159     assert(render);
 160
 161     if (render->filled_mv_blocks_num > 0)
 162         ff_mpeg_draw_horiz_band(s, 0, 0);
 163     return 0;
 164 }
 165
 166 /**
 167  * Synthesize the data needed by XvMC to render one macroblock of data.
 168  * Fill all relevant fields, if necessary do IDCT.
 169  */
 170 static void ff_xvmc_decode_mb(struct MpegEncContext *s)
 171 {
 172     XvMCMacroBlock *mv_block;
 173     struct xvmc_pix_fmt *render;
 174     int i, cbp, blocks_per_mb;
 175
 176     const int mb_xy = s->mb_y * s->mb_stride + s->mb_x;
 177
 178
 179     if (s->encoding) {
 180         av_log(s->avctx, AV_LOG_ERROR, "XVMC doesn't support encoding!!!\n");
 181         return;
 182     }
 183
 184     // from MPV_decode_mb(), update DC predictors for P macroblocks
 185     if (!s->mb_intra) {
 186         s->last_dc[0] =
 187         s->last_dc[1] =
 188         s->last_dc[2] =  128 << s->intra_dc_precision;
 189     }
 190
 191     // MC doesn't skip blocks
 192     s->mb_skipped = 0;
 193
 194
 195     // Do I need to export quant when I could not perform postprocessing?
 196     // Anyway, it doesn't hurt.
 197     s->current_picture.qscale_table[mb_xy] = s->qscale;
 198
 199     // start of XVMC-specific code
 200     render = (struct xvmc_pix_fmt*)s->current_picture.f.data[2];
 201     assert(render);
 202     assert(render->xvmc_id == AV_XVMC_ID);
 203     assert(render->mv_blocks);
 204
 205     // take the next free macroblock
 206     mv_block = &render->mv_blocks[render->start_mv_blocks_num +
 207                                   render->filled_mv_blocks_num];
 208
 209     mv_block->x        = s->mb_x;
 210     mv_block->y        = s->mb_y;
 211     mv_block->dct_type = s->interlaced_dct; // XVMC_DCT_TYPE_FRAME/FIELD;
 212     if (s->mb_intra) {
 213         mv_block->macroblock_type = XVMC_MB_TYPE_INTRA; // no MC, all done
 214     } else {
 215         mv_block->macroblock_type = XVMC_MB_TYPE_PATTERN;
 216
 217         if (s->mv_dir & MV_DIR_FORWARD) {
 218             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_FORWARD;
 219             // PMV[n][dir][xy] = mv[dir][n][xy]
 220             mv_block->PMV[0][0][0] = s->mv[0][0][0];
 221             mv_block->PMV[0][0][1] = s->mv[0][0][1];
 222             mv_block->PMV[1][0][0] = s->mv[0][1][0];
 223             mv_block->PMV[1][0][1] = s->mv[0][1][1];
 224         }
 225         if (s->mv_dir & MV_DIR_BACKWARD) {
 226             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_BACKWARD;
 227             mv_block->PMV[0][1][0] = s->mv[1][0][0];
 228             mv_block->PMV[0][1][1] = s->mv[1][0][1];
 229             mv_block->PMV[1][1][0] = s->mv[1][1][0];
 230             mv_block->PMV[1][1][1] = s->mv[1][1][1];
 231         }
 232
 233         switch(s->mv_type) {
 234             case  MV_TYPE_16X16:
 235                 mv_block->motion_type = XVMC_PREDICTION_FRAME;
 236                 break;
 237             case  MV_TYPE_16X8:
 238                 mv_block->motion_type = XVMC_PREDICTION_16x8;
 239                 break;
 240             case  MV_TYPE_FIELD:
 241                 mv_block->motion_type = XVMC_PREDICTION_FIELD;
 242                 if (s->picture_structure == PICT_FRAME) {
 243                     mv_block->PMV[0][0][1] <<= 1;
 244                     mv_block->PMV[1][0][1] <<= 1;
 245                     mv_block->PMV[0][1][1] <<= 1;
 246                     mv_block->PMV[1][1][1] <<= 1;
 247                 }
 248                 break;
 249             case  MV_TYPE_DMV:
 250                 mv_block->motion_type = XVMC_PREDICTION_DUAL_PRIME;
 251                 if (s->picture_structure == PICT_FRAME) {
 252
 253                     mv_block->PMV[0][0][0] = s->mv[0][0][0];      // top from top
 254                     mv_block->PMV[0][0][1] = s->mv[0][0][1] << 1;
 255
 256                     mv_block->PMV[0][1][0] = s->mv[0][0][0];      // bottom from bottom
 257                     mv_block->PMV[0][1][1] = s->mv[0][0][1] << 1;
 258
 259                     mv_block->PMV[1][0][0] = s->mv[0][2][0];      // dmv00, top from bottom
 260                     mv_block->PMV[1][0][1] = s->mv[0][2][1] << 1; // dmv01
 261
 262                     mv_block->PMV[1][1][0] = s->mv[0][3][0];      // dmv10, bottom from top
 263                     mv_block->PMV[1][1][1] = s->mv[0][3][1] << 1; // dmv11
 264
 265                 } else {
 266                     mv_block->PMV[0][1][0] = s->mv[0][2][0];      // dmv00
 267                     mv_block->PMV[0][1][1] = s->mv[0][2][1];      // dmv01
 268                 }
 269                 break;
 270             default:
 271                 assert(0);
 272         }
 273
 274         mv_block->motion_vertical_field_select = 0;
 275
 276         // set correct field references
 277         if (s->mv_type == MV_TYPE_FIELD || s->mv_type == MV_TYPE_16X8) {
 278             mv_block->motion_vertical_field_select |= s->field_select[0][0];
 279             mv_block->motion_vertical_field_select |= s->field_select[1][0] << 1;
 280             mv_block->motion_vertical_field_select |= s->field_select[0][1] << 2;
 281             mv_block->motion_vertical_field_select |= s->field_select[1][1] << 3;
 282         }
 283     } // !intra
 284     // time to handle data blocks
 285     mv_block->index = render->next_free_data_block_num;
 286
 287     blocks_per_mb = 6;
 288     if (s->chroma_format >= 2) {
 289         blocks_per_mb = 4 + (1 << s->chroma_format);
 290     }
 291
 292     // calculate cbp
 293     cbp = 0;
 294     for (i = 0; i < blocks_per_mb; i++) {
 295         cbp += cbp;
 296         if (s->block_last_index[i] >= 0)
 297             cbp++;
 298     }
 299
 300     if (s->flags & CODEC_FLAG_GRAY) {
 301         if (s->mb_intra) {                                   // intra frames are always full chroma blocks
 302             for (i = 4; i < blocks_per_mb; i++) {
 303                 memset(s->pblocks[i], 0, sizeof(*s->pblocks[i]));  // so we need to clear them
 304                 if (!render->unsigned_intra)
 305                     *s->pblocks[i][0] = 1 << 10;
 306             }
 307         } else {
 308             cbp &= 0xf << (blocks_per_mb - 4);
 309             blocks_per_mb = 4;                               // luminance blocks only
 310         }
 311     }
 312     mv_block->coded_block_pattern = cbp;
 313     if (cbp == 0)
 314         mv_block->macroblock_type &= ~XVMC_MB_TYPE_PATTERN;
 315
 316     for (i = 0; i < blocks_per_mb; i++) {
 317         if (s->block_last_index[i] >= 0) {
 318             // I do not have unsigned_intra MOCO to test, hope it is OK.
 319             if (s->mb_intra && (render->idct || !render->unsigned_intra))
 320                 *s->pblocks[i][0] -= 1 << 10;
 321             if (!render->idct) {
 322                 s->dsp.idct(*s->pblocks[i]);
 323                 /* It is unclear if MC hardware requires pixel diff values to be
 324                  * in the range [-255;255]. TODO: Clipping if such hardware is
 325                  * ever found. As of now it would only be an unnecessary
 326                  * slowdown. */
 327             }
 328             // copy blocks only if the codec doesn't support pblocks reordering
 329             if (!s->pack_pblocks) {
 330                 memcpy(&render->data_blocks[render->next_free_data_block_num*64],
 331                        s->pblocks[i], sizeof(*s->pblocks[i]));
 332             }
 333             render->next_free_data_block_num++;
 334         }
 335     }
 336     render->filled_mv_blocks_num++;
 337
 338     assert(render->filled_mv_blocks_num     <= render->allocated_mv_blocks);
 339     assert(render->next_free_data_block_num <= render->allocated_data_blocks);
 340     /* The above conditions should not be able to fail as long as this function
 341      * is used and the following 'if ()' automatically calls a callback to free
 342      * blocks. */
 343
 344
 345     if (render->filled_mv_blocks_num == render->allocated_mv_blocks)
 346         ff_mpeg_draw_horiz_band(s, 0, 0);
 347 }
 348
 349 #if CONFIG_MPEG1_XVMC_HWACCEL
 350 AVHWAccel ff_mpeg1_xvmc_hwaccel = {
 351     .name           = "mpeg1_xvmc",
 352     .type           = AVMEDIA_TYPE_VIDEO,
 353     .id             = AV_CODEC_ID_MPEG1VIDEO,
 354     .pix_fmt        = AV_PIX_FMT_XVMC,
 355     .start_frame    = ff_xvmc_field_start,
 356     .end_frame      = ff_xvmc_field_end,
 357     .decode_slice   = NULL,
 358     .decode_mb      = ff_xvmc_decode_mb,
 359     .priv_data_size = 0,
 360 };
 361 #endif
 362
 363 #if CONFIG_MPEG2_XVMC_HWACCEL
 364 AVHWAccel ff_mpeg2_xvmc_hwaccel = {
 365     .name           = "mpeg2_xvmc",
 366     .type           = AVMEDIA_TYPE_VIDEO,
 367     .id             = AV_CODEC_ID_MPEG2VIDEO,
 368     .pix_fmt        = AV_PIX_FMT_XVMC,
 369     .start_frame    = ff_xvmc_field_start,
 370     .end_frame      = ff_xvmc_field_end,
 371     .decode_slice   = NULL,
 372     .decode_mb      = ff_xvmc_decode_mb,
 373     .priv_data_size = 0,
 374 };
 375 #endif