Almost cosmetics.
[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_pixfmt_render *render = (struct xvmc_pixfmt_render*)s->current_picture.data[2];
41     assert(render && render->magic_id == AV_XVMC_RENDER_MAGIC);
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_pixfmt_render *last, *next, *render = (struct xvmc_pixfmt_render*)s->current_picture.data[2];
68     const int mb_block_count = 4 + (1 << s->chroma_format);
69
70     assert(avctx);
71     if (!render || render->magic_id != AV_XVMC_RENDER_MAGIC ||
72         !render->data_blocks || !render->mv_blocks){
73         av_log(avctx, AV_LOG_ERROR,
74                "Render token doesn't look as expected.\n");
75         return -1; // make sure that this is a render packet
76     }
77
78     if (render->filled_mv_blocks_num) {
79         av_log(avctx, AV_LOG_ERROR,
80                "Rendering surface contains %i unprocessed blocks.\n",
81                render->filled_mv_blocks_num);
82         return -1;
83     }
84     if (render->total_number_of_mv_blocks   < 1 ||
85         render->total_number_of_data_blocks < mb_block_count) {
86         av_log(avctx, AV_LOG_ERROR,
87                "Rendering surface doesn't provide enough block structures to work with.\n");
88         return -1;
89     }
90     if (render->total_number_of_mv_blocks   < 1 ||
91         render->total_number_of_data_blocks < mb_block_count) {
92         av_log(avctx, AV_LOG_ERROR,
93                "Rendering surface doesn't provide enough block structures to work with.\n");
94         return -1;
95     }
96
97     render->picture_structure = s->picture_structure;
98     render->flags             = s->first_field ? 0 : XVMC_SECOND_FIELD;
99     render->p_future_surface = NULL;
100     render->p_past_surface   = NULL;
101
102     switch(s->pict_type) {
103         case  FF_I_TYPE:
104             return 0; // no prediction from other frames
105         case  FF_B_TYPE:
106             next = (struct xvmc_pixfmt_render*)s->next_picture.data[2];
107             if (!next)
108                 return -1;
109             if (next->magic_id != AV_XVMC_RENDER_MAGIC)
110                 return -1;
111             render->p_future_surface = next->p_surface;
112             // no return here, going to set forward prediction
113         case  FF_P_TYPE:
114             last = (struct xvmc_pixfmt_render*)s->last_picture.data[2];
115             if (!last)
116                 last = render; // predict second field from the first
117             if (last->magic_id != AV_XVMC_RENDER_MAGIC)
118                 return -1;
119             render->p_past_surface = last->p_surface;
120             return 0;
121     }
122
123 return -1;
124 }
125
126 /**
127  * This function should be called for every new field and/or frame.
128  * It should be safe to call the function a few times for the same field.
129  */
130 void ff_xvmc_field_end(MpegEncContext *s)
131 {
132     struct xvmc_pixfmt_render *render = (struct xvmc_pixfmt_render*)s->current_picture.data[2];
133     assert(render);
134
135     if (render->filled_mv_blocks_num > 0)
136         ff_draw_horiz_band(s, 0, 0);
137 }
138
139 void ff_xvmc_decode_mb(MpegEncContext *s)
140 {
141     XvMCMacroBlock *mv_block;
142     struct xvmc_pixfmt_render *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         av_log(s->avctx, AV_LOG_ERROR, "XVMC doesn't support encoding!!!\n");
150         return;
151     }
152
153     // from MPV_decode_mb(), update DC predictors for P macroblocks
154     if (!s->mb_intra) {
155         s->last_dc[0] =
156         s->last_dc[1] =
157         s->last_dc[2] =  128 << s->intra_dc_precision;
158     }
159
160     // MC doesn't skip blocks
161     s->mb_skipped = 0;
162
163
164     // Do I need to export quant when I could not perform postprocessing?
165     // Anyway, it doesn't hurt.
166     s->current_picture.qscale_table[mb_xy] = s->qscale;
167
168     // start of XVMC-specific code
169     render = (struct xvmc_pixfmt_render*)s->current_picture.data[2];
170     assert(render);
171     assert(render->magic_id == AV_XVMC_RENDER_MAGIC);
172     assert(render->mv_blocks);
173
174     // take the next free macroblock
175     mv_block = &render->mv_blocks[render->start_mv_blocks_num +
176                                   render->filled_mv_blocks_num];
177
178     mv_block->x        = s->mb_x;
179     mv_block->y        = s->mb_y;
180     mv_block->dct_type = s->interlaced_dct; // XVMC_DCT_TYPE_FRAME/FIELD;
181     if (s->mb_intra) {
182         mv_block->macroblock_type = XVMC_MB_TYPE_INTRA; // no MC, all done
183     } else {
184         mv_block->macroblock_type = XVMC_MB_TYPE_PATTERN;
185
186         if (s->mv_dir & MV_DIR_FORWARD) {
187             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_FORWARD;
188             // PMV[n][dir][xy] = mv[dir][n][xy]
189             mv_block->PMV[0][0][0] = s->mv[0][0][0];
190             mv_block->PMV[0][0][1] = s->mv[0][0][1];
191             mv_block->PMV[1][0][0] = s->mv[0][1][0];
192             mv_block->PMV[1][0][1] = s->mv[0][1][1];
193         }
194         if (s->mv_dir & MV_DIR_BACKWARD) {
195             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_BACKWARD;
196             mv_block->PMV[0][1][0] = s->mv[1][0][0];
197             mv_block->PMV[0][1][1] = s->mv[1][0][1];
198             mv_block->PMV[1][1][0] = s->mv[1][1][0];
199             mv_block->PMV[1][1][1] = s->mv[1][1][1];
200         }
201
202         switch(s->mv_type) {
203             case  MV_TYPE_16X16:
204                 mv_block->motion_type = XVMC_PREDICTION_FRAME;
205                 break;
206             case  MV_TYPE_16X8:
207                 mv_block->motion_type = XVMC_PREDICTION_16x8;
208                 break;
209             case  MV_TYPE_FIELD:
210                 mv_block->motion_type = XVMC_PREDICTION_FIELD;
211                 if (s->picture_structure == PICT_FRAME) {
212                     mv_block->PMV[0][0][1] <<= 1;
213                     mv_block->PMV[1][0][1] <<= 1;
214                     mv_block->PMV[0][1][1] <<= 1;
215                     mv_block->PMV[1][1][1] <<= 1;
216                 }
217                 break;
218             case  MV_TYPE_DMV:
219                 mv_block->motion_type = XVMC_PREDICTION_DUAL_PRIME;
220                 if (s->picture_structure == PICT_FRAME) {
221
222                     mv_block->PMV[0][0][0] = s->mv[0][0][0];      // top from top
223                     mv_block->PMV[0][0][1] = s->mv[0][0][1] << 1;
224
225                     mv_block->PMV[0][1][0] = s->mv[0][0][0];      // bottom from bottom
226                     mv_block->PMV[0][1][1] = s->mv[0][0][1] << 1;
227
228                     mv_block->PMV[1][0][0] = s->mv[0][2][0];      // dmv00, top from bottom
229                     mv_block->PMV[1][0][1] = s->mv[0][2][1] << 1; // dmv01
230
231                     mv_block->PMV[1][1][0] = s->mv[0][3][0];      // dmv10, bottom from top
232                     mv_block->PMV[1][1][1] = s->mv[0][3][1] << 1; // dmv11
233
234                 } else {
235                     mv_block->PMV[0][1][0] = s->mv[0][2][0];      // dmv00
236                     mv_block->PMV[0][1][1] = s->mv[0][2][1];      // dmv01
237                 }
238                 break;
239             default:
240                 assert(0);
241         }
242
243         mv_block->motion_vertical_field_select = 0;
244
245         // set correct field references
246         if (s->mv_type == MV_TYPE_FIELD || s->mv_type == MV_TYPE_16X8) {
247             mv_block->motion_vertical_field_select |= s->field_select[0][0];
248             mv_block->motion_vertical_field_select |= s->field_select[1][0] << 1;
249             mv_block->motion_vertical_field_select |= s->field_select[0][1] << 2;
250             mv_block->motion_vertical_field_select |= s->field_select[1][1] << 3;
251         }
252     } // !intra
253     // time to handle data blocks
254     mv_block->index = render->next_free_data_block_num;
255
256     blocks_per_mb = 6;
257     if (s->chroma_format >= 2) {
258         blocks_per_mb = 4 + (1 << s->chroma_format);
259     }
260
261     // calculate cbp
262     cbp = 0;
263     for (i = 0; i < blocks_per_mb; i++) {
264         cbp += cbp;
265         if (s->block_last_index[i] >= 0)
266             cbp++;
267     }
268
269     if (s->flags & CODEC_FLAG_GRAY) {
270         if (s->mb_intra) {                                   // intra frames are always full chroma blocks
271             for (i = 4; i < blocks_per_mb; i++) {
272                 memset(s->pblocks[i], 0, sizeof(short)*64);  // so we need to clear them
273                 if (!render->unsigned_intra)
274                     s->pblocks[i][0] = 1 << 10;
275             }
276         } else {
277             cbp &= 0xf << (blocks_per_mb - 4);
278             blocks_per_mb = 4;                               // luminance blocks only
279         }
280     }
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                 /* It is unclear if MC hardware requires pixel diff values to be
293                  * in the range [-255;255]. TODO: Clipping if such hardware is
294                  * ever found. As of now it would only be an unnecessary
295                  * slowdown. */
296             }
297             // copy blocks only if the codec doesn't support pblocks reordering
298             if (s->avctx->xvmc_acceleration == 1) {
299                 memcpy(&render->data_blocks[render->next_free_data_block_num*64],
300                        s->pblocks[i], sizeof(short)*64);
301             }
302             render->next_free_data_block_num++;
303         }
304     }
305     render->filled_mv_blocks_num++;
306
307     assert(render->filled_mv_blocks_num     <= render->total_number_of_mv_blocks);
308     assert(render->next_free_data_block_num <= render->total_number_of_data_blocks);
309     /* The above conditions should not be able to fail as long as this function
310      * is used and the following 'if ()' automatically calls a callback to free
311      * blocks. */
312
313
314     if (render->filled_mv_blocks_num == render->total_number_of_mv_blocks)
315         ff_draw_horiz_band(s, 0, 0);
316 }