whitespace cosmetics: remove stray space
[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 //set s->block
35 void ff_xvmc_init_block(MpegEncContext *s)
36 {
37     struct xvmc_render_state *render = (struct xvmc_render_state*)s->current_picture.data[2];
38     assert(render);
39     if (!render || render->magic != AV_XVMC_RENDER_MAGIC) {
40         assert(0);
41         return; // make sure that this is a render packet
42     }
43     s->block = (DCTELEM *)(render->data_blocks + render->next_free_data_block_num * 64);
44 }
45
46 void ff_xvmc_pack_pblocks(MpegEncContext *s, int cbp)
47 {
48     int i, j = 0;
49     const int mb_block_count = 4 + (1 << s->chroma_format);
50
51     cbp <<= 12-mb_block_count;
52     for (i = 0; i < mb_block_count; i++) {
53         if (cbp & (1 << 11))
54             s->pblocks[i] = (short *)(&s->block[j++]);
55         else
56             s->pblocks[i] = NULL;
57         cbp+=cbp;
58     }
59 }
60
61 /**
62  * This function should be called for every new field and/or frame.
63  * It should be safe to call the function a few times for the same field.
64  */
65 int ff_xvmc_field_start(MpegEncContext*s, AVCodecContext *avctx)
66 {
67     struct xvmc_render_state *last, *next, *render = (struct xvmc_render_state*)s->current_picture.data[2];
68
69     assert(avctx);
70     assert(render);
71     if (!render || render->magic != AV_XVMC_RENDER_MAGIC)
72         return -1; // make sure that this is a render packet
73
74     render->picture_structure = s->picture_structure;
75     render->flags             = s->first_field ? 0 : XVMC_SECOND_FIELD;
76
77     if (render->filled_mv_blocks_num) {
78         av_log(avctx, AV_LOG_ERROR,
79                "Rendering surface contains %i unprocessed blocks\n",
80                render->filled_mv_blocks_num);
81         return -1;
82     }
83
84     render->p_future_surface = NULL;
85     render->p_past_surface   = NULL;
86
87     switch(s->pict_type) {
88         case  FF_I_TYPE:
89             return 0; // no prediction from other frames
90         case  FF_B_TYPE:
91             next = (struct xvmc_render_state*)s->next_picture.data[2];
92             assert(next);
93             if (!next)
94                 return -1;
95             if (next->magic != AV_XVMC_RENDER_MAGIC)
96                 return -1;
97             render->p_future_surface = next->p_surface;
98             // no return here, going to set forward prediction
99         case  FF_P_TYPE:
100             last = (struct xvmc_render_state*)s->last_picture.data[2];
101             if (!last)
102                 last = render; // predict second field from the first
103             if (last->magic != AV_XVMC_RENDER_MAGIC)
104                 return -1;
105             render->p_past_surface = last->p_surface;
106             return 0;
107     }
108
109 return -1;
110 }
111
112 void ff_xvmc_field_end(MpegEncContext *s)
113 {
114     struct xvmc_render_state *render = (struct xvmc_render_state*)s->current_picture.data[2];
115     assert(render);
116
117     if (render->filled_mv_blocks_num > 0)
118         ff_draw_horiz_band(s, 0, 0);
119 }
120
121 void ff_xvmc_decode_mb(MpegEncContext *s)
122 {
123     XvMCMacroBlock *mv_block;
124     struct xvmc_render_state *render;
125     int i, cbp, blocks_per_mb;
126
127     const int mb_xy = s->mb_y * s->mb_stride + s->mb_x;
128
129
130     if (s->encoding) {
131         av_log(s->avctx, AV_LOG_ERROR, "XVMC doesn't support encoding!!!\n");
132         return;
133     }
134
135     // from MPV_decode_mb(), update DC predictors for P macroblocks
136     if (!s->mb_intra) {
137         s->last_dc[0] =
138         s->last_dc[1] =
139         s->last_dc[2] =  128 << s->intra_dc_precision;
140     }
141
142     // MC doesn't skip blocks
143     s->mb_skipped = 0;
144
145
146     // Do I need to export quant when I could not perform postprocessing?
147     // Anyway, it doesn't hurt.
148     s->current_picture.qscale_table[mb_xy] = s->qscale;
149
150     // start of XVMC-specific code
151     render = (struct xvmc_render_state*)s->current_picture.data[2];
152     assert(render);
153     assert(render->magic == AV_XVMC_RENDER_MAGIC);
154     assert(render->mv_blocks);
155
156     // take the next free macroblock
157     mv_block = &render->mv_blocks[render->start_mv_blocks_num +
158                                   render->filled_mv_blocks_num];
159
160     mv_block->x        = s->mb_x;
161     mv_block->y        = s->mb_y;
162     mv_block->dct_type = s->interlaced_dct; // XVMC_DCT_TYPE_FRAME/FIELD;
163     if (s->mb_intra) {
164         mv_block->macroblock_type = XVMC_MB_TYPE_INTRA; // no MC, all done
165     } else {
166         mv_block->macroblock_type = XVMC_MB_TYPE_PATTERN;
167
168         if (s->mv_dir & MV_DIR_FORWARD) {
169             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_FORWARD;
170             // PMV[n][dir][xy] = mv[dir][n][xy]
171             mv_block->PMV[0][0][0] = s->mv[0][0][0];
172             mv_block->PMV[0][0][1] = s->mv[0][0][1];
173             mv_block->PMV[1][0][0] = s->mv[0][1][0];
174             mv_block->PMV[1][0][1] = s->mv[0][1][1];
175         }
176         if (s->mv_dir & MV_DIR_BACKWARD) {
177             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_BACKWARD;
178             mv_block->PMV[0][1][0] = s->mv[1][0][0];
179             mv_block->PMV[0][1][1] = s->mv[1][0][1];
180             mv_block->PMV[1][1][0] = s->mv[1][1][0];
181             mv_block->PMV[1][1][1] = s->mv[1][1][1];
182         }
183
184         switch(s->mv_type) {
185             case  MV_TYPE_16X16:
186                 mv_block->motion_type = XVMC_PREDICTION_FRAME;
187                 break;
188             case  MV_TYPE_16X8:
189                 mv_block->motion_type = XVMC_PREDICTION_16x8;
190                 break;
191             case  MV_TYPE_FIELD:
192                 mv_block->motion_type = XVMC_PREDICTION_FIELD;
193                 if (s->picture_structure == PICT_FRAME) {
194                     mv_block->PMV[0][0][1] <<= 1;
195                     mv_block->PMV[1][0][1] <<= 1;
196                     mv_block->PMV[0][1][1] <<= 1;
197                     mv_block->PMV[1][1][1] <<= 1;
198                 }
199                 break;
200             case  MV_TYPE_DMV:
201                 mv_block->motion_type = XVMC_PREDICTION_DUAL_PRIME;
202                 if (s->picture_structure == PICT_FRAME) {
203
204                     mv_block->PMV[0][0][0] = s->mv[0][0][0];      // top from top
205                     mv_block->PMV[0][0][1] = s->mv[0][0][1] << 1;
206
207                     mv_block->PMV[0][1][0] = s->mv[0][0][0];      // bottom from bottom
208                     mv_block->PMV[0][1][1] = s->mv[0][0][1] << 1;
209
210                     mv_block->PMV[1][0][0] = s->mv[0][2][0];      // dmv00, top from bottom
211                     mv_block->PMV[1][0][1] = s->mv[0][2][1] << 1; // dmv01
212
213                     mv_block->PMV[1][1][0] = s->mv[0][3][0];      // dmv10, bottom from top
214                     mv_block->PMV[1][1][1] = s->mv[0][3][1] << 1; // dmv11
215
216                 } else {
217                     mv_block->PMV[0][1][0] = s->mv[0][2][0];      // dmv00
218                     mv_block->PMV[0][1][1] = s->mv[0][2][1];      // dmv01
219                 }
220                 break;
221             default:
222                 assert(0);
223         }
224
225         mv_block->motion_vertical_field_select = 0;
226
227         // set correct field references
228         if (s->mv_type == MV_TYPE_FIELD || s->mv_type == MV_TYPE_16X8) {
229             mv_block->motion_vertical_field_select |= s->field_select[0][0];
230             mv_block->motion_vertical_field_select |= s->field_select[1][0] << 1;
231             mv_block->motion_vertical_field_select |= s->field_select[0][1] << 2;
232             mv_block->motion_vertical_field_select |= s->field_select[1][1] << 3;
233         }
234     } // !intra
235     // time to handle data blocks
236     mv_block->index = render->next_free_data_block_num;
237
238     blocks_per_mb = 6;
239     if (s->chroma_format >= 2) {
240         blocks_per_mb = 4 + (1 << s->chroma_format);
241     }
242
243     // calculate cbp
244     cbp = 0;
245     for (i = 0; i < blocks_per_mb; i++) {
246         cbp += cbp;
247         if (s->block_last_index[i] >= 0)
248             cbp++;
249     }
250
251     if (s->flags & CODEC_FLAG_GRAY) {
252         if (s->mb_intra) {                                   // intra frames are always full chroma blocks
253             for (i = 4; i < blocks_per_mb; i++) {
254                 memset(s->pblocks[i], 0, sizeof(short)*64);  // so we need to clear them
255                 if (!render->unsigned_intra)
256                     s->pblocks[i][0] = 1 << 10;
257             }
258         } else {
259             cbp &= 0xf << (blocks_per_mb - 4);
260             blocks_per_mb = 4;                               // luminance blocks only
261         }
262     }
263     mv_block->coded_block_pattern = cbp;
264     if (cbp == 0)
265         mv_block->macroblock_type &= ~XVMC_MB_TYPE_PATTERN;
266
267     for (i = 0; i < blocks_per_mb; i++) {
268         if (s->block_last_index[i] >= 0) {
269             // I do not have unsigned_intra MOCO to test, hope it is OK.
270             if (s->mb_intra && (render->idct || (!render->idct && !render->unsigned_intra)))
271                 s->pblocks[i][0] -= 1 << 10;
272             if (!render->idct) {
273                 s->dsp.idct(s->pblocks[i]);
274                 /* It is unclear if MC hardware requires pixel diff values to be
275                  * in the range [-255;255]. TODO: Clipping if such hardware is
276                  * ever found. As of now it would only be an unnecessary
277                  * slowdown. */
278             }
279             // copy blocks only if the codec doesn't support pblocks reordering
280             if (s->avctx->xvmc_acceleration == 1) {
281                 memcpy(&render->data_blocks[render->next_free_data_block_num*64],
282                        s->pblocks[i], sizeof(short)*64);
283             }
284             render->next_free_data_block_num++;
285         }
286     }
287     render->filled_mv_blocks_num++;
288
289     assert(render->filled_mv_blocks_num     <= render->total_number_of_mv_blocks);
290     assert(render->next_free_data_block_num <= render->total_number_of_data_blocks);
291     /* The above conditions should not be able to fail as long as this function
292      * is used and the following 'if ()' automatically calls a callback to free
293      * blocks. */
294
295
296     if (render->filled_mv_blocks_num >= render->total_number_of_mv_blocks)
297         ff_draw_horiz_band(s, 0, 0);
298 }