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