Replace sizeof(short) by the more robust sizeof(*s->pblocks[i]) in memset and
[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 decoding
38  * coefficients in s->blocks and then copying them, copy them 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->xvmc_id == AV_XVMC_ID);
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 the macroblock 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->xvmc_id != AV_XVMC_ID ||
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->allocated_mv_blocks   < 1 ||
93         render->allocated_data_blocks <  render->allocated_mv_blocks*mb_block_count ||
94         render->start_mv_blocks_num   >= render->allocated_mv_blocks                ||
95         render->next_free_data_block_num >
96                         render->allocated_data_blocks -
97                         mb_block_count*(render->allocated_mv_blocks-render->start_mv_blocks_num)) {
98         av_log(avctx, AV_LOG_ERROR,
99                "Rendering surface doesn't provide enough block structures to work with.\n");
100         return -1;
101     }
102
103     render->picture_structure = s->picture_structure;
104     render->flags             = s->first_field ? 0 : XVMC_SECOND_FIELD;
105     render->p_future_surface  = NULL;
106     render->p_past_surface    = NULL;
107
108     switch(s->pict_type) {
109         case  FF_I_TYPE:
110             return 0; // no prediction from other frames
111         case  FF_B_TYPE:
112             next = (struct xvmc_pix_fmt*)s->next_picture.data[2];
113             if (!next)
114                 return -1;
115             if (next->xvmc_id != AV_XVMC_ID)
116                 return -1;
117             render->p_future_surface = next->p_surface;
118             // no return here, going to set forward prediction
119         case  FF_P_TYPE:
120             last = (struct xvmc_pix_fmt*)s->last_picture.data[2];
121             if (!last)
122                 last = render; // predict second field from the first
123             if (last->xvmc_id != AV_XVMC_ID)
124                 return -1;
125             render->p_past_surface = last->p_surface;
126             return 0;
127     }
128
129 return -1;
130 }
131
132 /**
133  * Complete frame/field rendering by passing any remaining blocks.
134  * Normally ff_draw_horiz_band() is called for each slice, however,
135  * some leftover blocks, for example from error_resilience(), may remain.
136  * It should be safe to call the function a few times for the same field.
137  */
138 void ff_xvmc_field_end(MpegEncContext *s)
139 {
140     struct xvmc_pix_fmt *render = (struct xvmc_pix_fmt*)s->current_picture.data[2];
141     assert(render);
142
143     if (render->filled_mv_blocks_num > 0)
144         ff_draw_horiz_band(s, 0, 0);
145 }
146
147 /**
148  * Synthesize the data needed by XvMC to render one macroblock of data.
149  * Fill all relevant fields, if necessary do IDCT.
150  */
151 void ff_xvmc_decode_mb(MpegEncContext *s)
152 {
153     XvMCMacroBlock *mv_block;
154     struct xvmc_pix_fmt *render;
155     int i, cbp, blocks_per_mb;
156
157     const int mb_xy = s->mb_y * s->mb_stride + s->mb_x;
158
159
160     if (s->encoding) {
161         av_log(s->avctx, AV_LOG_ERROR, "XVMC doesn't support encoding!!!\n");
162         return;
163     }
164
165     // from MPV_decode_mb(), update DC predictors for P macroblocks
166     if (!s->mb_intra) {
167         s->last_dc[0] =
168         s->last_dc[1] =
169         s->last_dc[2] =  128 << s->intra_dc_precision;
170     }
171
172     // MC doesn't skip blocks
173     s->mb_skipped = 0;
174
175
176     // Do I need to export quant when I could not perform postprocessing?
177     // Anyway, it doesn't hurt.
178     s->current_picture.qscale_table[mb_xy] = s->qscale;
179
180     // start of XVMC-specific code
181     render = (struct xvmc_pix_fmt*)s->current_picture.data[2];
182     assert(render);
183     assert(render->xvmc_id == AV_XVMC_ID);
184     assert(render->mv_blocks);
185
186     // take the next free macroblock
187     mv_block = &render->mv_blocks[render->start_mv_blocks_num +
188                                   render->filled_mv_blocks_num];
189
190     mv_block->x        = s->mb_x;
191     mv_block->y        = s->mb_y;
192     mv_block->dct_type = s->interlaced_dct; // XVMC_DCT_TYPE_FRAME/FIELD;
193     if (s->mb_intra) {
194         mv_block->macroblock_type = XVMC_MB_TYPE_INTRA; // no MC, all done
195     } else {
196         mv_block->macroblock_type = XVMC_MB_TYPE_PATTERN;
197
198         if (s->mv_dir & MV_DIR_FORWARD) {
199             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_FORWARD;
200             // PMV[n][dir][xy] = mv[dir][n][xy]
201             mv_block->PMV[0][0][0] = s->mv[0][0][0];
202             mv_block->PMV[0][0][1] = s->mv[0][0][1];
203             mv_block->PMV[1][0][0] = s->mv[0][1][0];
204             mv_block->PMV[1][0][1] = s->mv[0][1][1];
205         }
206         if (s->mv_dir & MV_DIR_BACKWARD) {
207             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_BACKWARD;
208             mv_block->PMV[0][1][0] = s->mv[1][0][0];
209             mv_block->PMV[0][1][1] = s->mv[1][0][1];
210             mv_block->PMV[1][1][0] = s->mv[1][1][0];
211             mv_block->PMV[1][1][1] = s->mv[1][1][1];
212         }
213
214         switch(s->mv_type) {
215             case  MV_TYPE_16X16:
216                 mv_block->motion_type = XVMC_PREDICTION_FRAME;
217                 break;
218             case  MV_TYPE_16X8:
219                 mv_block->motion_type = XVMC_PREDICTION_16x8;
220                 break;
221             case  MV_TYPE_FIELD:
222                 mv_block->motion_type = XVMC_PREDICTION_FIELD;
223                 if (s->picture_structure == PICT_FRAME) {
224                     mv_block->PMV[0][0][1] <<= 1;
225                     mv_block->PMV[1][0][1] <<= 1;
226                     mv_block->PMV[0][1][1] <<= 1;
227                     mv_block->PMV[1][1][1] <<= 1;
228                 }
229                 break;
230             case  MV_TYPE_DMV:
231                 mv_block->motion_type = XVMC_PREDICTION_DUAL_PRIME;
232                 if (s->picture_structure == PICT_FRAME) {
233
234                     mv_block->PMV[0][0][0] = s->mv[0][0][0];      // top from top
235                     mv_block->PMV[0][0][1] = s->mv[0][0][1] << 1;
236
237                     mv_block->PMV[0][1][0] = s->mv[0][0][0];      // bottom from bottom
238                     mv_block->PMV[0][1][1] = s->mv[0][0][1] << 1;
239
240                     mv_block->PMV[1][0][0] = s->mv[0][2][0];      // dmv00, top from bottom
241                     mv_block->PMV[1][0][1] = s->mv[0][2][1] << 1; // dmv01
242
243                     mv_block->PMV[1][1][0] = s->mv[0][3][0];      // dmv10, bottom from top
244                     mv_block->PMV[1][1][1] = s->mv[0][3][1] << 1; // dmv11
245
246                 } else {
247                     mv_block->PMV[0][1][0] = s->mv[0][2][0];      // dmv00
248                     mv_block->PMV[0][1][1] = s->mv[0][2][1];      // dmv01
249                 }
250                 break;
251             default:
252                 assert(0);
253         }
254
255         mv_block->motion_vertical_field_select = 0;
256
257         // set correct field references
258         if (s->mv_type == MV_TYPE_FIELD || s->mv_type == MV_TYPE_16X8) {
259             mv_block->motion_vertical_field_select |= s->field_select[0][0];
260             mv_block->motion_vertical_field_select |= s->field_select[1][0] << 1;
261             mv_block->motion_vertical_field_select |= s->field_select[0][1] << 2;
262             mv_block->motion_vertical_field_select |= s->field_select[1][1] << 3;
263         }
264     } // !intra
265     // time to handle data blocks
266     mv_block->index = render->next_free_data_block_num;
267
268     blocks_per_mb = 6;
269     if (s->chroma_format >= 2) {
270         blocks_per_mb = 4 + (1 << s->chroma_format);
271     }
272
273     // calculate cbp
274     cbp = 0;
275     for (i = 0; i < blocks_per_mb; i++) {
276         cbp += cbp;
277         if (s->block_last_index[i] >= 0)
278             cbp++;
279     }
280
281     if (s->flags & CODEC_FLAG_GRAY) {
282         if (s->mb_intra) {                                   // intra frames are always full chroma blocks
283             for (i = 4; i < blocks_per_mb; i++) {
284                 memset(s->pblocks[i], 0, sizeof(*s->pblocks[i])*64);  // so we need to clear them
285                 if (!render->unsigned_intra)
286                     s->pblocks[i][0] = 1 << 10;
287             }
288         } else {
289             cbp &= 0xf << (blocks_per_mb - 4);
290             blocks_per_mb = 4;                               // luminance blocks only
291         }
292     }
293     mv_block->coded_block_pattern = cbp;
294     if (cbp == 0)
295         mv_block->macroblock_type &= ~XVMC_MB_TYPE_PATTERN;
296
297     for (i = 0; i < blocks_per_mb; i++) {
298         if (s->block_last_index[i] >= 0) {
299             // I do not have unsigned_intra MOCO to test, hope it is OK.
300             if (s->mb_intra && (render->idct || (!render->idct && !render->unsigned_intra)))
301                 s->pblocks[i][0] -= 1 << 10;
302             if (!render->idct) {
303                 s->dsp.idct(s->pblocks[i]);
304                 /* It is unclear if MC hardware requires pixel diff values to be
305                  * in the range [-255;255]. TODO: Clipping if such hardware is
306                  * ever found. As of now it would only be an unnecessary
307                  * slowdown. */
308             }
309             // copy blocks only if the codec doesn't support pblocks reordering
310             if (s->avctx->xvmc_acceleration == 1) {
311                 memcpy(&render->data_blocks[render->next_free_data_block_num*64],
312                        s->pblocks[i], sizeof(*s->pblocks[i])*64);
313             }
314             render->next_free_data_block_num++;
315         }
316     }
317     render->filled_mv_blocks_num++;
318
319     assert(render->filled_mv_blocks_num     <= render->allocated_mv_blocks);
320     assert(render->next_free_data_block_num <= render->allocated_data_blocks);
321     /* The above conditions should not be able to fail as long as this function
322      * is used and the following 'if ()' automatically calls a callback to free
323      * blocks. */
324
325
326     if (render->filled_mv_blocks_num == render->allocated_mv_blocks)
327         ff_draw_horiz_band(s, 0, 0);
328 }