d9b238ac0f26740db89b7768da913733ca86169b
[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
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 #include "xvmc_internal.h"
33
34 /**
35  * Initializes 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 decodeing
38  * coefficients in s->blocks and then copy them, do that 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.data[2];
44     assert(render && render->unique_id == AV_XVMC_RENDER_MAGIC);
45
46     s->block = (DCTELEM *)(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 MB 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] = (short *)(&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.data[2];
76     const int mb_block_count = 4 + (1 << s->chroma_format);
77
78     assert(avctx);
79     if (!render || render->unique_id != AV_XVMC_RENDER_MAGIC ||
80         !render->data_blocks || !render->mv_blocks) {
81         av_log(avctx, AV_LOG_ERROR,
82                "Render token doesn't look as expected.\n");
83         return -1; // make sure that this is a render packet
84     }
85
86     if (render->filled_mv_blocks_num) {
87         av_log(avctx, AV_LOG_ERROR,
88                "Rendering surface contains %i unprocessed blocks.\n",
89                render->filled_mv_blocks_num);
90         return -1;
91     }
92     if (render->total_number_of_mv_blocks   < 1 ||
93         render->total_number_of_data_blocks < mb_block_count) {
94         av_log(avctx, AV_LOG_ERROR,
95                "Rendering surface doesn't provide enough block structures to work with.\n");
96         return -1;
97     }
98
99     render->picture_structure = s->picture_structure;
100     render->flags             = s->first_field ? 0 : XVMC_SECOND_FIELD;
101     render->p_future_surface  = NULL;
102     render->p_past_surface    = NULL;
103
104     switch(s->pict_type) {
105         case  FF_I_TYPE:
106             return 0; // no prediction from other frames
107         case  FF_B_TYPE:
108             next = (struct xvmc_pix_fmt*)s->next_picture.data[2];
109             if (!next)
110                 return -1;
111             if (next->unique_id != AV_XVMC_RENDER_MAGIC)
112                 return -1;
113             render->p_future_surface = next->p_surface;
114             // no return here, going to set forward prediction
115         case  FF_P_TYPE:
116             last = (struct xvmc_pix_fmt*)s->last_picture.data[2];
117             if (!last)
118                 last = render; // predict second field from the first
119             if (last->unique_id != AV_XVMC_RENDER_MAGIC)
120                 return -1;
121             render->p_past_surface = last->p_surface;
122             return 0;
123     }
124
125 return -1;
126 }
127
128 /**
129  * Complete frame/field rendering by passing any remaining blocks.
130  * Normally ff_draw_horiz_band() is called on each slice, however
131  * there may be some remaining blocks, for example put by error_resilience.
132  * It should be safe to call the function a few times for the same field.
133  */
134 void ff_xvmc_field_end(MpegEncContext *s)
135 {
136     struct xvmc_pix_fmt *render = (struct xvmc_pix_fmt*)s->current_picture.data[2];
137     assert(render);
138
139     if (render->filled_mv_blocks_num > 0)
140         ff_draw_horiz_band(s, 0, 0);
141 }
142
143 /**
144  * Synthesize the data needed by XvMC to render one macroblock of data.
145  * Fill all relevent fields, if necessery do IDCT.
146  */
147 void ff_xvmc_decode_mb(MpegEncContext *s)
148 {
149     XvMCMacroBlock *mv_block;
150     struct xvmc_pix_fmt *render;
151     int i, cbp, blocks_per_mb;
152
153     const int mb_xy = s->mb_y * s->mb_stride + s->mb_x;
154
155
156     if (s->encoding) {
157         av_log(s->avctx, AV_LOG_ERROR, "XVMC doesn't support encoding!!!\n");
158         return;
159     }
160
161     // from MPV_decode_mb(), update DC predictors for P macroblocks
162     if (!s->mb_intra) {
163         s->last_dc[0] =
164         s->last_dc[1] =
165         s->last_dc[2] =  128 << s->intra_dc_precision;
166     }
167
168     // MC doesn't skip blocks
169     s->mb_skipped = 0;
170
171
172     // Do I need to export quant when I could not perform postprocessing?
173     // Anyway, it doesn't hurt.
174     s->current_picture.qscale_table[mb_xy] = s->qscale;
175
176     // start of XVMC-specific code
177     render = (struct xvmc_pix_fmt*)s->current_picture.data[2];
178     assert(render);
179     assert(render->unique_id == AV_XVMC_RENDER_MAGIC);
180     assert(render->mv_blocks);
181
182     // take the next free macroblock
183     mv_block = &render->mv_blocks[render->start_mv_blocks_num +
184                                   render->filled_mv_blocks_num];
185
186     mv_block->x        = s->mb_x;
187     mv_block->y        = s->mb_y;
188     mv_block->dct_type = s->interlaced_dct; // XVMC_DCT_TYPE_FRAME/FIELD;
189     if (s->mb_intra) {
190         mv_block->macroblock_type = XVMC_MB_TYPE_INTRA; // no MC, all done
191     } else {
192         mv_block->macroblock_type = XVMC_MB_TYPE_PATTERN;
193
194         if (s->mv_dir & MV_DIR_FORWARD) {
195             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_FORWARD;
196             // PMV[n][dir][xy] = mv[dir][n][xy]
197             mv_block->PMV[0][0][0] = s->mv[0][0][0];
198             mv_block->PMV[0][0][1] = s->mv[0][0][1];
199             mv_block->PMV[1][0][0] = s->mv[0][1][0];
200             mv_block->PMV[1][0][1] = s->mv[0][1][1];
201         }
202         if (s->mv_dir & MV_DIR_BACKWARD) {
203             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_BACKWARD;
204             mv_block->PMV[0][1][0] = s->mv[1][0][0];
205             mv_block->PMV[0][1][1] = s->mv[1][0][1];
206             mv_block->PMV[1][1][0] = s->mv[1][1][0];
207             mv_block->PMV[1][1][1] = s->mv[1][1][1];
208         }
209
210         switch(s->mv_type) {
211             case  MV_TYPE_16X16:
212                 mv_block->motion_type = XVMC_PREDICTION_FRAME;
213                 break;
214             case  MV_TYPE_16X8:
215                 mv_block->motion_type = XVMC_PREDICTION_16x8;
216                 break;
217             case  MV_TYPE_FIELD:
218                 mv_block->motion_type = XVMC_PREDICTION_FIELD;
219                 if (s->picture_structure == PICT_FRAME) {
220                     mv_block->PMV[0][0][1] <<= 1;
221                     mv_block->PMV[1][0][1] <<= 1;
222                     mv_block->PMV[0][1][1] <<= 1;
223                     mv_block->PMV[1][1][1] <<= 1;
224                 }
225                 break;
226             case  MV_TYPE_DMV:
227                 mv_block->motion_type = XVMC_PREDICTION_DUAL_PRIME;
228                 if (s->picture_structure == PICT_FRAME) {
229
230                     mv_block->PMV[0][0][0] = s->mv[0][0][0];      // top from top
231                     mv_block->PMV[0][0][1] = s->mv[0][0][1] << 1;
232
233                     mv_block->PMV[0][1][0] = s->mv[0][0][0];      // bottom from bottom
234                     mv_block->PMV[0][1][1] = s->mv[0][0][1] << 1;
235
236                     mv_block->PMV[1][0][0] = s->mv[0][2][0];      // dmv00, top from bottom
237                     mv_block->PMV[1][0][1] = s->mv[0][2][1] << 1; // dmv01
238
239                     mv_block->PMV[1][1][0] = s->mv[0][3][0];      // dmv10, bottom from top
240                     mv_block->PMV[1][1][1] = s->mv[0][3][1] << 1; // dmv11
241
242                 } else {
243                     mv_block->PMV[0][1][0] = s->mv[0][2][0];      // dmv00
244                     mv_block->PMV[0][1][1] = s->mv[0][2][1];      // dmv01
245                 }
246                 break;
247             default:
248                 assert(0);
249         }
250
251         mv_block->motion_vertical_field_select = 0;
252
253         // set correct field references
254         if (s->mv_type == MV_TYPE_FIELD || s->mv_type == MV_TYPE_16X8) {
255             mv_block->motion_vertical_field_select |= s->field_select[0][0];
256             mv_block->motion_vertical_field_select |= s->field_select[1][0] << 1;
257             mv_block->motion_vertical_field_select |= s->field_select[0][1] << 2;
258             mv_block->motion_vertical_field_select |= s->field_select[1][1] << 3;
259         }
260     } // !intra
261     // time to handle data blocks
262     mv_block->index = render->next_free_data_block_num;
263
264     blocks_per_mb = 6;
265     if (s->chroma_format >= 2) {
266         blocks_per_mb = 4 + (1 << s->chroma_format);
267     }
268
269     // calculate cbp
270     cbp = 0;
271     for (i = 0; i < blocks_per_mb; i++) {
272         cbp += cbp;
273         if (s->block_last_index[i] >= 0)
274             cbp++;
275     }
276
277     if (s->flags & CODEC_FLAG_GRAY) {
278         if (s->mb_intra) {                                   // intra frames are always full chroma blocks
279             for (i = 4; i < blocks_per_mb; i++) {
280                 memset(s->pblocks[i], 0, sizeof(short)*64);  // so we need to clear them
281                 if (!render->unsigned_intra)
282                     s->pblocks[i][0] = 1 << 10;
283             }
284         } else {
285             cbp &= 0xf << (blocks_per_mb - 4);
286             blocks_per_mb = 4;                               // luminance blocks only
287         }
288     }
289     mv_block->coded_block_pattern = cbp;
290     if (cbp == 0)
291         mv_block->macroblock_type &= ~XVMC_MB_TYPE_PATTERN;
292
293     for (i = 0; i < blocks_per_mb; i++) {
294         if (s->block_last_index[i] >= 0) {
295             // I do not have unsigned_intra MOCO to test, hope it is OK.
296             if (s->mb_intra && (render->idct || (!render->idct && !render->unsigned_intra)))
297                 s->pblocks[i][0] -= 1 << 10;
298             if (!render->idct) {
299                 s->dsp.idct(s->pblocks[i]);
300                 /* It is unclear if MC hardware requires pixel diff values to be
301                  * in the range [-255;255]. TODO: Clipping if such hardware is
302                  * ever found. As of now it would only be an unnecessary
303                  * slowdown. */
304             }
305             // copy blocks only if the codec doesn't support pblocks reordering
306             if (s->avctx->xvmc_acceleration == 1) {
307                 memcpy(&render->data_blocks[render->next_free_data_block_num*64],
308                        s->pblocks[i], sizeof(short)*64);
309             }
310             render->next_free_data_block_num++;
311         }
312     }
313     render->filled_mv_blocks_num++;
314
315     assert(render->filled_mv_blocks_num     <= render->total_number_of_mv_blocks);
316     assert(render->next_free_data_block_num <= render->total_number_of_data_blocks);
317     /* The above conditions should not be able to fail as long as this function
318      * is used and the following 'if ()' automatically calls a callback to free
319      * blocks. */
320
321
322     if (render->filled_mv_blocks_num == render->total_number_of_mv_blocks)
323         ff_draw_horiz_band(s, 0, 0);
324 }