]> git.sesse.net Git - ffmpeg/blob - libavcodec/xvmcvideo.c
fixing twice added offset bug, was triggered by 4mv + sub_cmp != mb_cmp
[ffmpeg] / libavcodec / xvmcvideo.c
1 /*
2  * XVideo Motion Compensation
3  * Copyright (c) 2003 Ivan Kalvachev
4  *
5  * This library is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU Lesser General Public
7  * License as published by the Free Software Foundation; either
8  * version 2 of the License, or (at your option) any later version.
9  *
10  * This library is distributed in the hope that it will be useful,
11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * Lesser General Public License for more details.
14  *
15  * You should have received a copy of the GNU Lesser General Public
16  * License along with this library; if not, write to the Free Software
17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
18  */
19
20 #include <limits.h>
21
22 //avcodec include
23 #include "avcodec.h"
24 #include "dsputil.h"
25 #include "mpegvideo.h"
26
27 #undef NDEBUG
28 #include <assert.h>
29
30 #ifdef USE_FASTMEMCPY
31 #include "fastmemcpy.h"
32 #endif
33
34 #ifdef HAVE_XVMC
35
36 //X11 includes are in the xvmc_render.h
37 //by replacing it with none-X one
38 //XvMC emulation could be performed
39
40 #include "xvmc_render.h"
41
42 //#include "xvmc_debug.h"
43
44 //set s->block
45 inline void XVMC_init_block(MpegEncContext *s){
46 xvmc_render_state_t * render;
47     render = (xvmc_render_state_t*)s->current_picture.data[2];
48     assert(render != NULL);
49     if( (render == NULL) || (render->magic != MP_XVMC_RENDER_MAGIC) ){
50         assert(0);
51         return;//make sure that this is render packet
52     }
53     s->block =(DCTELEM *)(render->data_blocks+(render->next_free_data_block_num)*64);
54 }
55
56 void XVMC_pack_pblocks(MpegEncContext *s, int cbp){
57 int i,j;
58 #define numblocks 6
59
60     j=0;
61     for(i=0;i<numblocks;i++){
62         if(cbp & (1<<(numblocks-1-i)) ){
63            s->pblocks[i] = (short *)(&s->block[(j++)]);
64         }else{
65            s->pblocks[i] = NULL;
66         }
67 //        printf("s->pblocks[%d]=%p ,s->block=%p cbp=%d\n",i,s->pblocks[i],s->block,cbp);
68     }
69 }
70
71 static int calc_cbp(MpegEncContext *s, int blocknum){
72 /* compute cbp */
73 // for I420 bit_offset=5
74 int  i,cbp = 0;
75     for(i=0; i<blocknum; i++) {
76         if(s->block_last_index[i] >= 0)
77             cbp |= 1 << (5 - i);
78     }
79     return cbp;
80 }
81
82
83
84 //these functions should be called on every new field or/and frame
85 //They should be safe if they are called few times for same field!
86 int XVMC_field_start(MpegEncContext*s, AVCodecContext *avctx){
87 xvmc_render_state_t * render,* last, * next;
88
89     assert(avctx != NULL);
90
91     render = (xvmc_render_state_t*)s->current_picture.data[2];
92     assert(render != NULL);
93     if( (render == NULL) || (render->magic != MP_XVMC_RENDER_MAGIC) )
94         return -1;//make sure that this is render packet
95
96     render->picture_structure = s->picture_structure;
97     render->flags = (s->first_field)? 0: XVMC_SECOND_FIELD;
98
99 //make sure that all data is drawn by XVMC_end_frame
100     assert(render->filled_mv_blocks_num==0);
101
102     render->p_future_surface = NULL;
103     render->p_past_surface = NULL;
104
105     switch(s->pict_type){
106         case  I_TYPE:
107             return 0;// no prediction from other frames
108         case  B_TYPE:
109             next = (xvmc_render_state_t*)s->next_picture.data[2];
110             assert(next!=NULL);
111             assert(next->state & MP_XVMC_STATE_PREDICTION);
112             if(next == NULL) return -1;
113             if(next->magic != MP_XVMC_RENDER_MAGIC) return -1;
114             render->p_future_surface = next->p_surface;
115             //no return here, going to set forward prediction
116         case  P_TYPE:
117             last = (xvmc_render_state_t*)s->last_picture.data[2];
118             if(last == NULL)// && !s->first_field)
119                 last = render;//predict second field from the first
120             if(last->magic != MP_XVMC_RENDER_MAGIC) return -1;
121             assert(last->state & MP_XVMC_STATE_PREDICTION);
122             render->p_past_surface = last->p_surface;
123             return 0;
124      }
125
126 return -1;
127 }
128
129 void XVMC_field_end(MpegEncContext *s){
130 xvmc_render_state_t * render;
131     render = (xvmc_render_state_t*)s->current_picture.data[2];
132     assert(render != NULL);
133
134     if(render->filled_mv_blocks_num > 0){
135 //        printf("xvmcvideo.c: rendering %d left blocks after last slice!!!\n",render->filled_mv_blocks_num );
136         ff_draw_horiz_band(s,0,0);
137     }
138 }
139
140 void XVMC_decode_mb(MpegEncContext *s){
141 XvMCMacroBlock * mv_block;
142 xvmc_render_state_t * render;
143 int i,cbp,blocks_per_mb;
144
145 const int mb_xy = s->mb_y * s->mb_stride + s->mb_x;
146
147
148     if(s->encoding){
149         fprintf(stderr,"XVMC doesn't support encoding!!!\n");
150         av_abort();
151     }
152
153    //from MPV_decode_mb(),
154     /* update DC predictors for P macroblocks */
155     if (!s->mb_intra) {
156         s->last_dc[0] =
157         s->last_dc[1] =
158         s->last_dc[2] =  128 << s->intra_dc_precision;
159     }
160
161    //MC doesn't skip blocks
162     s->mb_skiped = 0;
163
164
165    // do I need to export quant when I could not perform postprocessing?
166    // anyway, it doesn't hurrt
167     s->current_picture.qscale_table[mb_xy] = s->qscale;
168
169 //START OF XVMC specific code
170     render = (xvmc_render_state_t*)s->current_picture.data[2];
171     assert(render!=NULL);
172     assert(render->magic==MP_XVMC_RENDER_MAGIC);
173     assert(render->mv_blocks);
174
175     //take the next free macroblock
176     mv_block = &render->mv_blocks[render->start_mv_blocks_num + 
177                                    render->filled_mv_blocks_num ];
178
179 // memset(mv_block,0,sizeof(XvMCMacroBlock));
180
181     mv_block->x = s->mb_x;
182     mv_block->y = s->mb_y;
183     mv_block->dct_type = s->interlaced_dct;//XVMC_DCT_TYPE_FRAME/FIELD;
184 //    mv_block->motion_type = 0;  //zero to silense warnings
185     if(s->mb_intra){
186         mv_block->macroblock_type = XVMC_MB_TYPE_INTRA;//no MC, all done
187     }else{
188         mv_block->macroblock_type = XVMC_MB_TYPE_PATTERN;
189
190         if(s->mv_dir & MV_DIR_FORWARD){
191             mv_block->macroblock_type|= XVMC_MB_TYPE_MOTION_FORWARD;
192             //pmv[n][dir][xy]=mv[dir][n][xy]
193             mv_block->PMV[0][0][0] = s->mv[0][0][0];
194             mv_block->PMV[0][0][1] = s->mv[0][0][1];
195             mv_block->PMV[1][0][0] = s->mv[0][1][0];
196             mv_block->PMV[1][0][1] = s->mv[0][1][1];
197         }
198         if(s->mv_dir & MV_DIR_BACKWARD){
199             mv_block->macroblock_type|=XVMC_MB_TYPE_MOTION_BACKWARD;
200             mv_block->PMV[0][1][0] = s->mv[1][0][0];
201             mv_block->PMV[0][1][1] = s->mv[1][0][1];
202             mv_block->PMV[1][1][0] = s->mv[1][1][0];
203             mv_block->PMV[1][1][1] = s->mv[1][1][1];
204         }
205
206         switch(s->mv_type){
207             case  MV_TYPE_16X16:
208                 mv_block->motion_type = XVMC_PREDICTION_FRAME;
209                 break;
210             case  MV_TYPE_16X8:
211                 mv_block->motion_type = XVMC_PREDICTION_16x8;
212                 break;
213             case  MV_TYPE_FIELD:
214                 mv_block->motion_type = XVMC_PREDICTION_FIELD;
215                 if(s->picture_structure == PICT_FRAME){
216                     mv_block->PMV[0][0][1]<<=1;
217                     mv_block->PMV[1][0][1]<<=1;
218                     mv_block->PMV[0][1][1]<<=1;
219                     mv_block->PMV[1][1][1]<<=1;
220                 }
221                 break;
222             case  MV_TYPE_DMV:
223                 mv_block->motion_type = XVMC_PREDICTION_DUAL_PRIME;
224                 if(s->picture_structure == PICT_FRAME){
225
226                     mv_block->PMV[0][0][0] = s->mv[0][0][0];//top from top
227                     mv_block->PMV[0][0][1] = s->mv[0][0][1]<<1;
228
229                     mv_block->PMV[0][1][0] = s->mv[0][0][0];//bottom from bottom
230                     mv_block->PMV[0][1][1] = s->mv[0][0][1]<<1;
231
232                     mv_block->PMV[1][0][0] = s->mv[0][2][0];//dmv00, top from bottom
233                     mv_block->PMV[1][0][1] = s->mv[0][2][1]<<1;//dmv01
234
235                     mv_block->PMV[1][1][0] = s->mv[0][3][0];//dmv10, bottom from top
236                     mv_block->PMV[1][1][1] = s->mv[0][3][1]<<1;//dmv11
237
238                 }else{
239                     mv_block->PMV[0][1][0] = s->mv[0][2][0];//dmv00
240                     mv_block->PMV[0][1][1] = s->mv[0][2][1];//dmv01
241                 }
242                 break;
243             default:
244                 assert(0);
245         }
246
247         mv_block->motion_vertical_field_select = 0;
248
249 //set correct field referenses
250         if(s->mv_type == MV_TYPE_FIELD || s->mv_type == MV_TYPE_16X8){
251             if( s->field_select[0][0] ) mv_block->motion_vertical_field_select|=1;
252             if( s->field_select[1][0] ) mv_block->motion_vertical_field_select|=2;
253             if( s->field_select[0][1] ) mv_block->motion_vertical_field_select|=4;
254             if( s->field_select[1][1] ) mv_block->motion_vertical_field_select|=8;
255         }
256     }//!intra
257 //time to handle data blocks;
258     mv_block->index = render->next_free_data_block_num;
259     blocks_per_mb = 6;
260 /*
261     switch( s->chroma_format){
262         case CHROMA_422:
263             blocks_per_mb = 8;
264             break;
265         case CHROMA_444:
266             blocks_per_mb = 12;
267             break;
268     }
269 */
270     if(s->flags & CODEC_FLAG_GRAY){
271         if(s->mb_intra){//intra frames are alwasy full chroma block
272             for(i=4; i<blocks_per_mb; i++){
273                 memset(s->pblocks[i],0,sizeof(short)*8*8);//so we need to clear them
274                 if(!render->unsigned_intra)
275                     s->pblocks[i][0] = 1<<10;
276             }
277         }else
278             blocks_per_mb = 4;//Luminance blocks only
279     }
280     cbp = calc_cbp(s,blocks_per_mb);
281     mv_block->coded_block_pattern = cbp;
282     if(cbp == 0)
283         mv_block->macroblock_type &= ~XVMC_MB_TYPE_PATTERN;
284
285     for(i=0; i<blocks_per_mb; i++){
286         if(s->block_last_index[i] >= 0){
287             // i do not have unsigned_intra MOCO to test, hope it is OK
288             if( (s->mb_intra) && ( render->idct || (!render->idct && !render->unsigned_intra)) )
289                 s->pblocks[i][0]-=1<<10;
290             if(!render->idct){
291                 s->dsp.idct(s->pblocks[i]);
292                 //!!TODO!clip!!!
293             }
294 //copy blocks only if the codec doesn't support pblocks reordering
295             if(s->avctx->xvmc_acceleration == 1){
296                 memcpy(&render->data_blocks[(render->next_free_data_block_num)*64],
297                         s->pblocks[i],sizeof(short)*8*8);
298             }else{
299 /*              if(s->pblocks[i] != &render->data_blocks[
300                         (render->next_free_data_block_num)*64]){
301                    printf("ERROR mb(%d,%d) s->pblocks[i]=%p data_block[]=%p\n",
302                    s->mb_x,s->mb_y, s->pblocks[i], 
303                    &render->data_blocks[(render->next_free_data_block_num)*64]);
304                 }*/
305             }
306             render->next_free_data_block_num++;
307         }
308     }
309     render->filled_mv_blocks_num++;
310
311     assert(render->filled_mv_blocks_num <= render->total_number_of_mv_blocks);
312     assert(render->next_free_data_block_num <= render->total_number_of_data_blocks);
313
314
315     if(render->filled_mv_blocks_num >= render->total_number_of_mv_blocks)
316         ff_draw_horiz_band(s,0,0);
317
318 // DumpRenderInfo(render);
319 // DumpMBlockInfo(mv_block);
320
321 }
322
323 #endif