Rename CONFIG_FFT_FLOAT ---> FFT_FLOAT
[ffmpeg.git] / libavcodec / mpegvideo_xvmc.c
1 /*
2  * XVideo Motion Compensation
3  * Copyright (c) 2003 Ivan Kalvachev
4  *
5  * This file is part of Libav.
6  *
7  * Libav 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  * Libav 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 Libav; 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 #include <X11/extensions/XvMC.h>
24
25 #include "avcodec.h"
26 #include "mpegvideo.h"
27
28 #undef NDEBUG
29 #include <assert.h>
30
31 #include "xvmc.h"
32 #include "xvmc_internal.h"
33 #include "version.h"
34
35 #if FF_API_XVMC
36
37 /**
38  * Initialize the block field of the MpegEncContext pointer passed as
39  * parameter after making sure that the data is not corrupted.
40  * In order to implement something like direct rendering instead of decoding
41  * coefficients in s->blocks and then copying them, copy them directly
42  * into the data_blocks array provided by xvmc.
43  */
44 void ff_xvmc_init_block(MpegEncContext *s)
45 {
46     struct xvmc_pix_fmt *render = (struct xvmc_pix_fmt*)s->current_picture.f.data[2];
47     assert(render && render->xvmc_id == AV_XVMC_ID);
48
49     s->block = (int16_t (*)[64])(render->data_blocks + render->next_free_data_block_num * 64);
50 }
51
52 /**
53  * Fill individual block pointers, so there are no gaps in the data_block array
54  * in case not all blocks in the macroblock are coded.
55  */
56 void ff_xvmc_pack_pblocks(MpegEncContext *s, int cbp)
57 {
58     int i, j = 0;
59     const int mb_block_count = 4 + (1 << s->chroma_format);
60
61     cbp <<= 12-mb_block_count;
62     for (i = 0; i < mb_block_count; i++) {
63         if (cbp & (1 << 11))
64             s->pblocks[i] = &s->block[j++];
65         else
66             s->pblocks[i] = NULL;
67         cbp += cbp;
68     }
69 }
70
71 /**
72  * Find and store the surfaces that are used as reference frames.
73  * This function should be called for every new field and/or frame.
74  * It should be safe to call the function a few times for the same field.
75  */
76 int ff_xvmc_field_start(MpegEncContext *s, AVCodecContext *avctx)
77 {
78     struct xvmc_pix_fmt *last, *next, *render = (struct xvmc_pix_fmt*)s->current_picture.f.data[2];
79     const int mb_block_count = 4 + (1 << s->chroma_format);
80
81     assert(avctx);
82     if (!render || render->xvmc_id != AV_XVMC_ID ||
83         !render->data_blocks || !render->mv_blocks ||
84         (unsigned int)render->allocated_mv_blocks   > INT_MAX/(64*6) ||
85         (unsigned int)render->allocated_data_blocks > INT_MAX/64     ||
86         !render->p_surface) {
87         av_log(avctx, AV_LOG_ERROR,
88                "Render token doesn't look as expected.\n");
89         return -1; // make sure that this is a render packet
90     }
91
92     if (render->filled_mv_blocks_num) {
93         av_log(avctx, AV_LOG_ERROR,
94                "Rendering surface contains %i unprocessed blocks.\n",
95                render->filled_mv_blocks_num);
96         return -1;
97     }
98     if (render->allocated_mv_blocks   < 1 ||
99         render->allocated_data_blocks <  render->allocated_mv_blocks*mb_block_count ||
100         render->start_mv_blocks_num   >= render->allocated_mv_blocks                ||
101         render->next_free_data_block_num >
102                         render->allocated_data_blocks -
103                         mb_block_count*(render->allocated_mv_blocks-render->start_mv_blocks_num)) {
104         av_log(avctx, AV_LOG_ERROR,
105                "Rendering surface doesn't provide enough block structures to work with.\n");
106         return -1;
107     }
108
109     render->picture_structure = s->picture_structure;
110     render->flags             = s->first_field ? 0 : XVMC_SECOND_FIELD;
111     render->p_future_surface  = NULL;
112     render->p_past_surface    = NULL;
113
114     switch(s->pict_type) {
115         case  AV_PICTURE_TYPE_I:
116             return 0; // no prediction from other frames
117         case  AV_PICTURE_TYPE_B:
118             next = (struct xvmc_pix_fmt*)s->next_picture.f.data[2];
119             if (!next)
120                 return -1;
121             if (next->xvmc_id != AV_XVMC_ID)
122                 return -1;
123             render->p_future_surface = next->p_surface;
124             // no return here, going to set forward prediction
125         case  AV_PICTURE_TYPE_P:
126             last = (struct xvmc_pix_fmt*)s->last_picture.f.data[2];
127             if (!last)
128                 last = render; // predict second field from the first
129             if (last->xvmc_id != AV_XVMC_ID)
130                 return -1;
131             render->p_past_surface = last->p_surface;
132             return 0;
133     }
134
135 return -1;
136 }
137
138 /**
139  * Complete frame/field rendering by passing any remaining blocks.
140  * Normally ff_draw_horiz_band() is called for each slice, however,
141  * some leftover blocks, for example from error_resilience(), may remain.
142  * It should be safe to call the function a few times for the same field.
143  */
144 void ff_xvmc_field_end(MpegEncContext *s)
145 {
146     struct xvmc_pix_fmt *render = (struct xvmc_pix_fmt*)s->current_picture.f.data[2];
147     assert(render);
148
149     if (render->filled_mv_blocks_num > 0)
150         ff_mpeg_draw_horiz_band(s, 0, 0);
151 }
152
153 /**
154  * Synthesize the data needed by XvMC to render one macroblock of data.
155  * Fill all relevant fields, if necessary do IDCT.
156  */
157 void ff_xvmc_decode_mb(MpegEncContext *s)
158 {
159     XvMCMacroBlock *mv_block;
160     struct xvmc_pix_fmt *render;
161     int i, cbp, blocks_per_mb;
162
163     const int mb_xy = s->mb_y * s->mb_stride + s->mb_x;
164
165
166     if (s->encoding) {
167         av_log(s->avctx, AV_LOG_ERROR, "XVMC doesn't support encoding!!!\n");
168         return;
169     }
170
171     // from MPV_decode_mb(), update DC predictors for P macroblocks
172     if (!s->mb_intra) {
173         s->last_dc[0] =
174         s->last_dc[1] =
175         s->last_dc[2] =  128 << s->intra_dc_precision;
176     }
177
178     // MC doesn't skip blocks
179     s->mb_skipped = 0;
180
181
182     // Do I need to export quant when I could not perform postprocessing?
183     // Anyway, it doesn't hurt.
184     s->current_picture.qscale_table[mb_xy] = s->qscale;
185
186     // start of XVMC-specific code
187     render = (struct xvmc_pix_fmt*)s->current_picture.f.data[2];
188     assert(render);
189     assert(render->xvmc_id == AV_XVMC_ID);
190     assert(render->mv_blocks);
191
192     // take the next free macroblock
193     mv_block = &render->mv_blocks[render->start_mv_blocks_num +
194                                   render->filled_mv_blocks_num];
195
196     mv_block->x        = s->mb_x;
197     mv_block->y        = s->mb_y;
198     mv_block->dct_type = s->interlaced_dct; // XVMC_DCT_TYPE_FRAME/FIELD;
199     if (s->mb_intra) {
200         mv_block->macroblock_type = XVMC_MB_TYPE_INTRA; // no MC, all done
201     } else {
202         mv_block->macroblock_type = XVMC_MB_TYPE_PATTERN;
203
204         if (s->mv_dir & MV_DIR_FORWARD) {
205             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_FORWARD;
206             // PMV[n][dir][xy] = mv[dir][n][xy]
207             mv_block->PMV[0][0][0] = s->mv[0][0][0];
208             mv_block->PMV[0][0][1] = s->mv[0][0][1];
209             mv_block->PMV[1][0][0] = s->mv[0][1][0];
210             mv_block->PMV[1][0][1] = s->mv[0][1][1];
211         }
212         if (s->mv_dir & MV_DIR_BACKWARD) {
213             mv_block->macroblock_type |= XVMC_MB_TYPE_MOTION_BACKWARD;
214             mv_block->PMV[0][1][0] = s->mv[1][0][0];
215             mv_block->PMV[0][1][1] = s->mv[1][0][1];
216             mv_block->PMV[1][1][0] = s->mv[1][1][0];
217             mv_block->PMV[1][1][1] = s->mv[1][1][1];
218         }
219
220         switch(s->mv_type) {
221             case  MV_TYPE_16X16:
222                 mv_block->motion_type = XVMC_PREDICTION_FRAME;
223                 break;
224             case  MV_TYPE_16X8:
225                 mv_block->motion_type = XVMC_PREDICTION_16x8;
226                 break;
227             case  MV_TYPE_FIELD:
228                 mv_block->motion_type = XVMC_PREDICTION_FIELD;
229                 if (s->picture_structure == PICT_FRAME) {
230                     mv_block->PMV[0][0][1] <<= 1;
231                     mv_block->PMV[1][0][1] <<= 1;
232                     mv_block->PMV[0][1][1] <<= 1;
233                     mv_block->PMV[1][1][1] <<= 1;
234                 }
235                 break;
236             case  MV_TYPE_DMV:
237                 mv_block->motion_type = XVMC_PREDICTION_DUAL_PRIME;
238                 if (s->picture_structure == PICT_FRAME) {
239
240                     mv_block->PMV[0][0][0] = s->mv[0][0][0];      // top from top
241                     mv_block->PMV[0][0][1] = s->mv[0][0][1] << 1;
242
243                     mv_block->PMV[0][1][0] = s->mv[0][0][0];      // bottom from bottom
244                     mv_block->PMV[0][1][1] = s->mv[0][0][1] << 1;
245
246                     mv_block->PMV[1][0][0] = s->mv[0][2][0];      // dmv00, top from bottom
247                     mv_block->PMV[1][0][1] = s->mv[0][2][1] << 1; // dmv01
248
249                     mv_block->PMV[1][1][0] = s->mv[0][3][0];      // dmv10, bottom from top
250                     mv_block->PMV[1][1][1] = s->mv[0][3][1] << 1; // dmv11
251
252                 } else {
253                     mv_block->PMV[0][1][0] = s->mv[0][2][0];      // dmv00
254                     mv_block->PMV[0][1][1] = s->mv[0][2][1];      // dmv01
255                 }
256                 break;
257             default:
258                 assert(0);
259         }
260
261         mv_block->motion_vertical_field_select = 0;
262
263         // set correct field references
264         if (s->mv_type == MV_TYPE_FIELD || s->mv_type == MV_TYPE_16X8) {
265             mv_block->motion_vertical_field_select |= s->field_select[0][0];
266             mv_block->motion_vertical_field_select |= s->field_select[1][0] << 1;
267             mv_block->motion_vertical_field_select |= s->field_select[0][1] << 2;
268             mv_block->motion_vertical_field_select |= s->field_select[1][1] << 3;
269         }
270     } // !intra
271     // time to handle data blocks
272     mv_block->index = render->next_free_data_block_num;
273
274     blocks_per_mb = 6;
275     if (s->chroma_format >= 2) {
276         blocks_per_mb = 4 + (1 << s->chroma_format);
277     }
278
279     // calculate cbp
280     cbp = 0;
281     for (i = 0; i < blocks_per_mb; i++) {
282         cbp += cbp;
283         if (s->block_last_index[i] >= 0)
284             cbp++;
285     }
286
287     if (s->flags & CODEC_FLAG_GRAY) {
288         if (s->mb_intra) {                                   // intra frames are always full chroma blocks
289             for (i = 4; i < blocks_per_mb; i++) {
290                 memset(s->pblocks[i], 0, sizeof(*s->pblocks[i]));  // so we need to clear them
291                 if (!render->unsigned_intra)
292                     *s->pblocks[i][0] = 1 << 10;
293             }
294         } else {
295             cbp &= 0xf << (blocks_per_mb - 4);
296             blocks_per_mb = 4;                               // luminance blocks only
297         }
298     }
299     mv_block->coded_block_pattern = cbp;
300     if (cbp == 0)
301         mv_block->macroblock_type &= ~XVMC_MB_TYPE_PATTERN;
302
303     for (i = 0; i < blocks_per_mb; i++) {
304         if (s->block_last_index[i] >= 0) {
305             // I do not have unsigned_intra MOCO to test, hope it is OK.
306             if (s->mb_intra && (render->idct || !render->unsigned_intra))
307                 *s->pblocks[i][0] -= 1 << 10;
308             if (!render->idct) {
309                 s->dsp.idct(*s->pblocks[i]);
310                 /* It is unclear if MC hardware requires pixel diff values to be
311                  * in the range [-255;255]. TODO: Clipping if such hardware is
312                  * ever found. As of now it would only be an unnecessary
313                  * slowdown. */
314             }
315             // copy blocks only if the codec doesn't support pblocks reordering
316             if (s->avctx->xvmc_acceleration == 1) {
317                 memcpy(&render->data_blocks[render->next_free_data_block_num*64],
318                        s->pblocks[i], sizeof(*s->pblocks[i]));
319             }
320             render->next_free_data_block_num++;
321         }
322     }
323     render->filled_mv_blocks_num++;
324
325     assert(render->filled_mv_blocks_num     <= render->allocated_mv_blocks);
326     assert(render->next_free_data_block_num <= render->allocated_data_blocks);
327     /* The above conditions should not be able to fail as long as this function
328      * is used and the following 'if ()' automatically calls a callback to free
329      * blocks. */
330
331
332     if (render->filled_mv_blocks_num == render->allocated_mv_blocks)
333         ff_mpeg_draw_horiz_band(s, 0, 0);
334 }
335
336 #endif /* FF_API_XVMC */