build: Generalize yasm/nasm-related variable names
[ffmpeg.git] / 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 "mpegutils.h"
27 #include "mpegvideo.h"
28
29 #undef NDEBUG
30 #include <assert.h>
31
32 #include "xvmc.h"
33 #include "xvmc_internal.h"
34 #include "version.h"
35
36 #if FF_API_XVMC
37
38 /**
39  * Initialize the block field of the MpegEncContext pointer passed as
40  * parameter after making sure that the data is not corrupted.
41  * In order to implement something like direct rendering instead of decoding
42  * coefficients in s->blocks and then copying them, copy them directly
43  * into the data_blocks array provided by xvmc.
44  */
45 void ff_xvmc_init_block(MpegEncContext *s)
46 {
47     struct xvmc_pix_fmt *render = (struct xvmc_pix_fmt*)s->current_picture.f->data[2];
48     assert(render && render->xvmc_id == AV_XVMC_ID);
49
50     s->block = (int16_t (*)[64])(render->data_blocks + render->next_free_data_block_num * 64);
51 }
52
53 /**
54  * Fill individual block pointers, so there are no gaps in the data_block array
55  * in case not all blocks in the macroblock are coded.
56  */
57 void ff_xvmc_pack_pblocks(MpegEncContext *s, int cbp)
58 {
59     int i, j = 0;
60     const int mb_block_count = 4 + (1 << s->chroma_format);
61
62     cbp <<= 12-mb_block_count;
63     for (i = 0; i < mb_block_count; i++) {
64         if (cbp & (1 << 11))
65             s->pblocks[i] = &s->block[j++];
66         else
67             s->pblocks[i] = NULL;
68         cbp += cbp;
69     }
70 }
71
72 /**
73  * Find and store the surfaces that are used as reference frames.
74  * This function should be called for every new field and/or frame.
75  * It should be safe to call the function a few times for the same field.
76  */
77 int ff_xvmc_field_start(MpegEncContext *s, AVCodecContext *avctx)
78 {
79     struct xvmc_pix_fmt *last, *next, *render = (struct xvmc_pix_fmt*)s->current_picture.f->data[2];
80     const int mb_block_count = 4 + (1 << s->chroma_format);
81
82     assert(avctx);
83     if (!render || render->xvmc_id != AV_XVMC_ID ||
84         !render->data_blocks || !render->mv_blocks ||
85         (unsigned int)render->allocated_mv_blocks   > INT_MAX/(64*6) ||
86         (unsigned int)render->allocated_data_blocks > INT_MAX/64     ||
87         !render->p_surface) {
88         av_log(avctx, AV_LOG_ERROR,
89                "Render token doesn't look as expected.\n");
90         return -1; // make sure that this is a render packet
91     }
92
93     if (render->filled_mv_blocks_num) {
94         av_log(avctx, AV_LOG_ERROR,
95                "Rendering surface contains %i unprocessed blocks.\n",
96                render->filled_mv_blocks_num);
97         return -1;
98     }
99     if (render->allocated_mv_blocks   < 1 ||
100         render->allocated_data_blocks <  render->allocated_mv_blocks*mb_block_count ||
101         render->start_mv_blocks_num   >= render->allocated_mv_blocks                ||
102         render->next_free_data_block_num >
103                         render->allocated_data_blocks -
104                         mb_block_count*(render->allocated_mv_blocks-render->start_mv_blocks_num)) {
105         av_log(avctx, AV_LOG_ERROR,
106                "Rendering surface doesn't provide enough block structures to work with.\n");
107         return -1;
108     }
109
110     render->picture_structure = s->picture_structure;
111     render->flags             = s->first_field ? 0 : XVMC_SECOND_FIELD;
112     render->p_future_surface  = NULL;
113     render->p_past_surface    = NULL;
114
115     switch(s->pict_type) {
116         case  AV_PICTURE_TYPE_I:
117             return 0; // no prediction from other frames
118         case  AV_PICTURE_TYPE_B:
119             next = (struct xvmc_pix_fmt*)s->next_picture.f->data[2];
120             if (!next)
121                 return -1;
122             if (next->xvmc_id != AV_XVMC_ID)
123                 return -1;
124             render->p_future_surface = next->p_surface;
125             // no return here, going to set forward prediction
126         case  AV_PICTURE_TYPE_P:
127             last = (struct xvmc_pix_fmt*)s->last_picture.f->data[2];
128             if (!last)
129                 last = render; // predict second field from the first
130             if (last->xvmc_id != AV_XVMC_ID)
131                 return -1;
132             render->p_past_surface = last->p_surface;
133             return 0;
134     }
135
136 return -1;
137 }
138
139 /**
140  * Complete frame/field rendering by passing any remaining blocks.
141  * Normally ff_draw_horiz_band() is called for each slice, however,
142  * some leftover blocks, for example from error_resilience(), may remain.
143  * It should be safe to call the function a few times for the same field.
144  */
145 void ff_xvmc_field_end(MpegEncContext *s)
146 {
147     struct xvmc_pix_fmt *render = (struct xvmc_pix_fmt*)s->current_picture.f->data[2];
148     assert(render);
149
150     if (render->filled_mv_blocks_num > 0)
151         ff_mpeg_draw_horiz_band(s, 0, 0);
152 }
153
154 /**
155  * Synthesize the data needed by XvMC to render one macroblock of data.
156  * Fill all relevant fields, if necessary do IDCT.
157  */
158 void ff_xvmc_decode_mb(MpegEncContext *s)
159 {
160     XvMCMacroBlock *mv_block;
161     struct xvmc_pix_fmt *render;
162     int i, cbp, blocks_per_mb;
163
164     const int mb_xy = s->mb_y * s->mb_stride + s->mb_x;
165
166
167     if (s->encoding) {
168         av_log(s->avctx, AV_LOG_ERROR, "XVMC doesn't support encoding!!!\n");
169         return;
170     }
171
172     // from ff_mpv_decode_mb(), update DC predictors for P macroblocks
173     if (!s->mb_intra) {
174         s->last_dc[0] =
175         s->last_dc[1] =
176         s->last_dc[2] =  128 << s->intra_dc_precision;
177     }
178
179     // MC doesn't skip blocks
180     s->mb_skipped = 0;
181
182
183     // Do I need to export quant when I could not perform postprocessing?
184     // Anyway, it doesn't hurt.
185     s->current_picture.qscale_table[mb_xy] = s->qscale;
186
187     // start of XVMC-specific code
188     render = (struct xvmc_pix_fmt*)s->current_picture.f->data[2];
189     assert(render);
190     assert(render->xvmc_id == AV_XVMC_ID);
191     assert(render->mv_blocks);
192
193     // take the next free macroblock
194     mv_block = &render->mv_blocks[render->start_mv_blocks_num +
195                                   render->filled_mv_blocks_num];
196
197     mv_block->x        = s->mb_x;
198     mv_block->y        = s->mb_y;
199     mv_block->dct_type = s->interlaced_dct; // XVMC_DCT_TYPE_FRAME/FIELD;
200     if (s->mb_intra) {
201         mv_block->macroblock_type = XVMC_MB_TYPE_INTRA; // no MC, all done
202     } else {
203         mv_block->macroblock_type = XVMC_MB_TYPE_PATTERN;
204
205         if (s->mv_dir & MV_DIR_FORWARD) {
206             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_FORWARD;
207             // PMV[n][dir][xy] = mv[dir][n][xy]
208             mv_block->PMV[0][0][0] = s->mv[0][0][0];
209             mv_block->PMV[0][0][1] = s->mv[0][0][1];
210             mv_block->PMV[1][0][0] = s->mv[0][1][0];
211             mv_block->PMV[1][0][1] = s->mv[0][1][1];
212         }
213         if (s->mv_dir & MV_DIR_BACKWARD) {
214             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_BACKWARD;
215             mv_block->PMV[0][1][0] = s->mv[1][0][0];
216             mv_block->PMV[0][1][1] = s->mv[1][0][1];
217             mv_block->PMV[1][1][0] = s->mv[1][1][0];
218             mv_block->PMV[1][1][1] = s->mv[1][1][1];
219         }
220
221         switch(s->mv_type) {
222             case  MV_TYPE_16X16:
223                 mv_block->motion_type = XVMC_PREDICTION_FRAME;
224                 break;
225             case  MV_TYPE_16X8:
226                 mv_block->motion_type = XVMC_PREDICTION_16x8;
227                 break;
228             case  MV_TYPE_FIELD:
229                 mv_block->motion_type = XVMC_PREDICTION_FIELD;
230                 if (s->picture_structure == PICT_FRAME) {
231                     mv_block->PMV[0][0][1] <<= 1;
232                     mv_block->PMV[1][0][1] <<= 1;
233                     mv_block->PMV[0][1][1] <<= 1;
234                     mv_block->PMV[1][1][1] <<= 1;
235                 }
236                 break;
237             case  MV_TYPE_DMV:
238                 mv_block->motion_type = XVMC_PREDICTION_DUAL_PRIME;
239                 if (s->picture_structure == PICT_FRAME) {
240
241                     mv_block->PMV[0][0][0] = s->mv[0][0][0];      // top from top
242                     mv_block->PMV[0][0][1] = s->mv[0][0][1] << 1;
243
244                     mv_block->PMV[0][1][0] = s->mv[0][0][0];      // bottom from bottom
245                     mv_block->PMV[0][1][1] = s->mv[0][0][1] << 1;
246
247                     mv_block->PMV[1][0][0] = s->mv[0][2][0];      // dmv00, top from bottom
248                     mv_block->PMV[1][0][1] = s->mv[0][2][1] << 1; // dmv01
249
250                     mv_block->PMV[1][1][0] = s->mv[0][3][0];      // dmv10, bottom from top
251                     mv_block->PMV[1][1][1] = s->mv[0][3][1] << 1; // dmv11
252
253                 } else {
254                     mv_block->PMV[0][1][0] = s->mv[0][2][0];      // dmv00
255                     mv_block->PMV[0][1][1] = s->mv[0][2][1];      // dmv01
256                 }
257                 break;
258             default:
259                 assert(0);
260         }
261
262         mv_block->motion_vertical_field_select = 0;
263
264         // set correct field references
265         if (s->mv_type == MV_TYPE_FIELD || s->mv_type == MV_TYPE_16X8) {
266             mv_block->motion_vertical_field_select |= s->field_select[0][0];
267             mv_block->motion_vertical_field_select |= s->field_select[1][0] << 1;
268             mv_block->motion_vertical_field_select |= s->field_select[0][1] << 2;
269             mv_block->motion_vertical_field_select |= s->field_select[1][1] << 3;
270         }
271     } // !intra
272     // time to handle data blocks
273     mv_block->index = render->next_free_data_block_num;
274
275     blocks_per_mb = 6;
276     if (s->chroma_format >= 2) {
277         blocks_per_mb = 4 + (1 << s->chroma_format);
278     }
279
280     // calculate cbp
281     cbp = 0;
282     for (i = 0; i < blocks_per_mb; i++) {
283         cbp += cbp;
284         if (s->block_last_index[i] >= 0)
285             cbp++;
286     }
287
288     if (s->avctx->flags & AV_CODEC_FLAG_GRAY) {
289         if (s->mb_intra) {                                   // intra frames are always full chroma blocks
290             for (i = 4; i < blocks_per_mb; i++) {
291                 memset(s->pblocks[i], 0, sizeof(*s->pblocks[i]));  // so we need to clear them
292                 if (!render->unsigned_intra)
293                     *s->pblocks[i][0] = 1 << 10;
294             }
295         } else {
296             cbp &= 0xf << (blocks_per_mb - 4);
297             blocks_per_mb = 4;                               // luminance blocks only
298         }
299     }
300     mv_block->coded_block_pattern = cbp;
301     if (cbp == 0)
302         mv_block->macroblock_type &= ~XVMC_MB_TYPE_PATTERN;
303
304     for (i = 0; i < blocks_per_mb; i++) {
305         if (s->block_last_index[i] >= 0) {
306             // I do not have unsigned_intra MOCO to test, hope it is OK.
307             if (s->mb_intra && (render->idct || !render->unsigned_intra))
308                 *s->pblocks[i][0] -= 1 << 10;
309             if (!render->idct) {
310                 s->idsp.idct(*s->pblocks[i]);
311                 /* It is unclear if MC hardware requires pixel diff values to be
312                  * in the range [-255;255]. TODO: Clipping if such hardware is
313                  * ever found. As of now it would only be an unnecessary
314                  * slowdown. */
315             }
316             // copy blocks only if the codec doesn't support pblocks reordering
317             if (s->avctx->xvmc_acceleration == 1) {
318                 memcpy(&render->data_blocks[render->next_free_data_block_num*64],
319                        s->pblocks[i], sizeof(*s->pblocks[i]));
320             }
321             render->next_free_data_block_num++;
322         }
323     }
324     render->filled_mv_blocks_num++;
325
326     assert(render->filled_mv_blocks_num     <= render->allocated_mv_blocks);
327     assert(render->next_free_data_block_num <= render->allocated_data_blocks);
328     /* The above conditions should not be able to fail as long as this function
329      * is used and the following 'if ()' automatically calls a callback to free
330      * blocks. */
331
332
333     if (render->filled_mv_blocks_num == render->allocated_mv_blocks)
334         ff_mpeg_draw_horiz_band(s, 0, 0);
335 }
336
337 #endif /* FF_API_XVMC */