svq3: use hpeldsp instead of dsputil for half-pel functions.
[ffmpeg.git] / 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
34 /**
35  * Initialize the block field of the MpegEncContext pointer passed as
36  * parameter after making sure that the data is not corrupted.
37  * In order to implement something like direct rendering instead of decoding
38  * coefficients in s->blocks and then copying them, copy them directly
39  * into the data_blocks array provided by xvmc.
40  */
41 void ff_xvmc_init_block(MpegEncContext *s)
42 {
43     struct xvmc_pix_fmt *render = (struct xvmc_pix_fmt*)s->current_picture.f.data[2];
44     assert(render && render->xvmc_id == AV_XVMC_ID);
45
46     s->block = (int16_t (*)[64])(render->data_blocks + render->next_free_data_block_num * 64);
47 }
48
49 /**
50  * Fill individual block pointers, so there are no gaps in the data_block array
51  * in case not all blocks in the macroblock are coded.
52  */
53 void ff_xvmc_pack_pblocks(MpegEncContext *s, int cbp)
54 {
55     int i, j = 0;
56     const int mb_block_count = 4 + (1 << s->chroma_format);
57
58     cbp <<= 12-mb_block_count;
59     for (i = 0; i < mb_block_count; i++) {
60         if (cbp & (1 << 11))
61             s->pblocks[i] = &s->block[j++];
62         else
63             s->pblocks[i] = NULL;
64         cbp += cbp;
65     }
66 }
67
68 /**
69  * Find and store the surfaces that are used as reference frames.
70  * This function should be called for every new field and/or frame.
71  * It should be safe to call the function a few times for the same field.
72  */
73 int ff_xvmc_field_start(MpegEncContext *s, AVCodecContext *avctx)
74 {
75     struct xvmc_pix_fmt *last, *next, *render = (struct xvmc_pix_fmt*)s->current_picture.f.data[2];
76     const int mb_block_count = 4 + (1 << s->chroma_format);
77
78     assert(avctx);
79     if (!render || render->xvmc_id != AV_XVMC_ID ||
80         !render->data_blocks || !render->mv_blocks ||
81         (unsigned int)render->allocated_mv_blocks   > INT_MAX/(64*6) ||
82         (unsigned int)render->allocated_data_blocks > INT_MAX/64     ||
83         !render->p_surface) {
84         av_log(avctx, AV_LOG_ERROR,
85                "Render token doesn't look as expected.\n");
86         return -1; // make sure that this is a render packet
87     }
88
89     if (render->filled_mv_blocks_num) {
90         av_log(avctx, AV_LOG_ERROR,
91                "Rendering surface contains %i unprocessed blocks.\n",
92                render->filled_mv_blocks_num);
93         return -1;
94     }
95     if (render->allocated_mv_blocks   < 1 ||
96         render->allocated_data_blocks <  render->allocated_mv_blocks*mb_block_count ||
97         render->start_mv_blocks_num   >= render->allocated_mv_blocks                ||
98         render->next_free_data_block_num >
99                         render->allocated_data_blocks -
100                         mb_block_count*(render->allocated_mv_blocks-render->start_mv_blocks_num)) {
101         av_log(avctx, AV_LOG_ERROR,
102                "Rendering surface doesn't provide enough block structures to work with.\n");
103         return -1;
104     }
105
106     render->picture_structure = s->picture_structure;
107     render->flags             = s->first_field ? 0 : XVMC_SECOND_FIELD;
108     render->p_future_surface  = NULL;
109     render->p_past_surface    = NULL;
110
111     switch(s->pict_type) {
112         case  AV_PICTURE_TYPE_I:
113             return 0; // no prediction from other frames
114         case  AV_PICTURE_TYPE_B:
115             next = (struct xvmc_pix_fmt*)s->next_picture.f.data[2];
116             if (!next)
117                 return -1;
118             if (next->xvmc_id != AV_XVMC_ID)
119                 return -1;
120             render->p_future_surface = next->p_surface;
121             // no return here, going to set forward prediction
122         case  AV_PICTURE_TYPE_P:
123             last = (struct xvmc_pix_fmt*)s->last_picture.f.data[2];
124             if (!last)
125                 last = render; // predict second field from the first
126             if (last->xvmc_id != AV_XVMC_ID)
127                 return -1;
128             render->p_past_surface = last->p_surface;
129             return 0;
130     }
131
132 return -1;
133 }
134
135 /**
136  * Complete frame/field rendering by passing any remaining blocks.
137  * Normally ff_draw_horiz_band() is called for each slice, however,
138  * some leftover blocks, for example from error_resilience(), may remain.
139  * It should be safe to call the function a few times for the same field.
140  */
141 void ff_xvmc_field_end(MpegEncContext *s)
142 {
143     struct xvmc_pix_fmt *render = (struct xvmc_pix_fmt*)s->current_picture.f.data[2];
144     assert(render);
145
146     if (render->filled_mv_blocks_num > 0)
147         ff_mpeg_draw_horiz_band(s, 0, 0);
148 }
149
150 /**
151  * Synthesize the data needed by XvMC to render one macroblock of data.
152  * Fill all relevant fields, if necessary do IDCT.
153  */
154 void ff_xvmc_decode_mb(MpegEncContext *s)
155 {
156     XvMCMacroBlock *mv_block;
157     struct xvmc_pix_fmt *render;
158     int i, cbp, blocks_per_mb;
159
160     const int mb_xy = s->mb_y * s->mb_stride + s->mb_x;
161
162
163     if (s->encoding) {
164         av_log(s->avctx, AV_LOG_ERROR, "XVMC doesn't support encoding!!!\n");
165         return;
166     }
167
168     // from MPV_decode_mb(), update DC predictors for P macroblocks
169     if (!s->mb_intra) {
170         s->last_dc[0] =
171         s->last_dc[1] =
172         s->last_dc[2] =  128 << s->intra_dc_precision;
173     }
174
175     // MC doesn't skip blocks
176     s->mb_skipped = 0;
177
178
179     // Do I need to export quant when I could not perform postprocessing?
180     // Anyway, it doesn't hurt.
181     s->current_picture.qscale_table[mb_xy] = s->qscale;
182
183     // start of XVMC-specific code
184     render = (struct xvmc_pix_fmt*)s->current_picture.f.data[2];
185     assert(render);
186     assert(render->xvmc_id == AV_XVMC_ID);
187     assert(render->mv_blocks);
188
189     // take the next free macroblock
190     mv_block = &render->mv_blocks[render->start_mv_blocks_num +
191                                   render->filled_mv_blocks_num];
192
193     mv_block->x        = s->mb_x;
194     mv_block->y        = s->mb_y;
195     mv_block->dct_type = s->interlaced_dct; // XVMC_DCT_TYPE_FRAME/FIELD;
196     if (s->mb_intra) {
197         mv_block->macroblock_type = XVMC_MB_TYPE_INTRA; // no MC, all done
198     } else {
199         mv_block->macroblock_type = XVMC_MB_TYPE_PATTERN;
200
201         if (s->mv_dir & MV_DIR_FORWARD) {
202             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_FORWARD;
203             // PMV[n][dir][xy] = mv[dir][n][xy]
204             mv_block->PMV[0][0][0] = s->mv[0][0][0];
205             mv_block->PMV[0][0][1] = s->mv[0][0][1];
206             mv_block->PMV[1][0][0] = s->mv[0][1][0];
207             mv_block->PMV[1][0][1] = s->mv[0][1][1];
208         }
209         if (s->mv_dir & MV_DIR_BACKWARD) {
210             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_BACKWARD;
211             mv_block->PMV[0][1][0] = s->mv[1][0][0];
212             mv_block->PMV[0][1][1] = s->mv[1][0][1];
213             mv_block->PMV[1][1][0] = s->mv[1][1][0];
214             mv_block->PMV[1][1][1] = s->mv[1][1][1];
215         }
216
217         switch(s->mv_type) {
218             case  MV_TYPE_16X16:
219                 mv_block->motion_type = XVMC_PREDICTION_FRAME;
220                 break;
221             case  MV_TYPE_16X8:
222                 mv_block->motion_type = XVMC_PREDICTION_16x8;
223                 break;
224             case  MV_TYPE_FIELD:
225                 mv_block->motion_type = XVMC_PREDICTION_FIELD;
226                 if (s->picture_structure == PICT_FRAME) {
227                     mv_block->PMV[0][0][1] <<= 1;
228                     mv_block->PMV[1][0][1] <<= 1;
229                     mv_block->PMV[0][1][1] <<= 1;
230                     mv_block->PMV[1][1][1] <<= 1;
231                 }
232                 break;
233             case  MV_TYPE_DMV:
234                 mv_block->motion_type = XVMC_PREDICTION_DUAL_PRIME;
235                 if (s->picture_structure == PICT_FRAME) {
236
237                     mv_block->PMV[0][0][0] = s->mv[0][0][0];      // top from top
238                     mv_block->PMV[0][0][1] = s->mv[0][0][1] << 1;
239
240                     mv_block->PMV[0][1][0] = s->mv[0][0][0];      // bottom from bottom
241                     mv_block->PMV[0][1][1] = s->mv[0][0][1] << 1;
242
243                     mv_block->PMV[1][0][0] = s->mv[0][2][0];      // dmv00, top from bottom
244                     mv_block->PMV[1][0][1] = s->mv[0][2][1] << 1; // dmv01
245
246                     mv_block->PMV[1][1][0] = s->mv[0][3][0];      // dmv10, bottom from top
247                     mv_block->PMV[1][1][1] = s->mv[0][3][1] << 1; // dmv11
248
249                 } else {
250                     mv_block->PMV[0][1][0] = s->mv[0][2][0];      // dmv00
251                     mv_block->PMV[0][1][1] = s->mv[0][2][1];      // dmv01
252                 }
253                 break;
254             default:
255                 assert(0);
256         }
257
258         mv_block->motion_vertical_field_select = 0;
259
260         // set correct field references
261         if (s->mv_type == MV_TYPE_FIELD || s->mv_type == MV_TYPE_16X8) {
262             mv_block->motion_vertical_field_select |= s->field_select[0][0];
263             mv_block->motion_vertical_field_select |= s->field_select[1][0] << 1;
264             mv_block->motion_vertical_field_select |= s->field_select[0][1] << 2;
265             mv_block->motion_vertical_field_select |= s->field_select[1][1] << 3;
266         }
267     } // !intra
268     // time to handle data blocks
269     mv_block->index = render->next_free_data_block_num;
270
271     blocks_per_mb = 6;
272     if (s->chroma_format >= 2) {
273         blocks_per_mb = 4 + (1 << s->chroma_format);
274     }
275
276     // calculate cbp
277     cbp = 0;
278     for (i = 0; i < blocks_per_mb; i++) {
279         cbp += cbp;
280         if (s->block_last_index[i] >= 0)
281             cbp++;
282     }
283
284     if (s->flags & CODEC_FLAG_GRAY) {
285         if (s->mb_intra) {                                   // intra frames are always full chroma blocks
286             for (i = 4; i < blocks_per_mb; i++) {
287                 memset(s->pblocks[i], 0, sizeof(*s->pblocks[i]));  // so we need to clear them
288                 if (!render->unsigned_intra)
289                     *s->pblocks[i][0] = 1 << 10;
290             }
291         } else {
292             cbp &= 0xf << (blocks_per_mb - 4);
293             blocks_per_mb = 4;                               // luminance blocks only
294         }
295     }
296     mv_block->coded_block_pattern = cbp;
297     if (cbp == 0)
298         mv_block->macroblock_type &= ~XVMC_MB_TYPE_PATTERN;
299
300     for (i = 0; i < blocks_per_mb; i++) {
301         if (s->block_last_index[i] >= 0) {
302             // I do not have unsigned_intra MOCO to test, hope it is OK.
303             if (s->mb_intra && (render->idct || !render->unsigned_intra))
304                 *s->pblocks[i][0] -= 1 << 10;
305             if (!render->idct) {
306                 s->dsp.idct(*s->pblocks[i]);
307                 /* It is unclear if MC hardware requires pixel diff values to be
308                  * in the range [-255;255]. TODO: Clipping if such hardware is
309                  * ever found. As of now it would only be an unnecessary
310                  * slowdown. */
311             }
312             // copy blocks only if the codec doesn't support pblocks reordering
313             if (s->avctx->xvmc_acceleration == 1) {
314                 memcpy(&render->data_blocks[render->next_free_data_block_num*64],
315                        s->pblocks[i], sizeof(*s->pblocks[i]));
316             }
317             render->next_free_data_block_num++;
318         }
319     }
320     render->filled_mv_blocks_num++;
321
322     assert(render->filled_mv_blocks_num     <= render->allocated_mv_blocks);
323     assert(render->next_free_data_block_num <= render->allocated_data_blocks);
324     /* The above conditions should not be able to fail as long as this function
325      * is used and the following 'if ()' automatically calls a callback to free
326      * blocks. */
327
328
329     if (render->filled_mv_blocks_num == render->allocated_mv_blocks)
330         ff_mpeg_draw_horiz_band(s, 0, 0);
331 }