1 /*
2  * SVQ1 Encoder
3  * Copyright (C) 2004 Mike Melanson <melanson@pcisys.net>
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 /**
23  * @file
24  * Sorenson Vector Quantizer #1 (SVQ1) video codec.
25  * For more information of the SVQ1 algorithm, visit:
26  *   http://www.pcisys.net/~melanson/codecs/
27  */
28 
29 #include "avcodec.h"
30 #include "codec_internal.h"
31 #include "encode.h"
32 #include "hpeldsp.h"
33 #include "me_cmp.h"
34 #include "mpegvideo.h"
35 #include "h263.h"
36 #include "h263enc.h"
37 #include "internal.h"
38 #include "mpegutils.h"
39 #include "packet_internal.h"
40 #include "svq1.h"
41 #include "svq1enc.h"
42 #include "svq1enc_cb.h"
43 #include "libavutil/avassert.h"
44 
45 
svq1_write_header(SVQ1EncContext *s, int frame_type)46 static void svq1_write_header(SVQ1EncContext *s, int frame_type)
47 {
48     int i;
49 
50     /* frame code */
51     put_bits(&s->pb, 22, 0x20);
52 
53     /* temporal reference (sure hope this is a "don't care") */
54     put_bits(&s->pb, 8, 0x00);
55 
56     /* frame type */
57     put_bits(&s->pb, 2, frame_type - 1);
58 
59     if (frame_type == AV_PICTURE_TYPE_I) {
60         /* no checksum since frame code is 0x20 */
61         /* no embedded string either */
62         /* output 5 unknown bits (2 + 2 + 1) */
63         put_bits(&s->pb, 5, 2); /* 2 needed by quicktime decoder */
64 
65         i = ff_match_2uint16((void*)ff_svq1_frame_size_table,
66                              FF_ARRAY_ELEMS(ff_svq1_frame_size_table),
67                              s->frame_width, s->frame_height);
68         put_bits(&s->pb, 3, i);
69 
70         if (i == 7) {
71             put_bits(&s->pb, 12, s->frame_width);
72             put_bits(&s->pb, 12, s->frame_height);
73         }
74     }
75 
76     /* no checksum or extra data (next 2 bits get 0) */
77     put_bits(&s->pb, 2, 0);
78 }
79 
80 #define QUALITY_THRESHOLD    100
81 #define THRESHOLD_MULTIPLIER 0.6
82 
ssd_int8_vs_int16_c(const int8_t *pix1, const int16_t *pix2, intptr_t size)83 static int ssd_int8_vs_int16_c(const int8_t *pix1, const int16_t *pix2,
84                                intptr_t size)
85 {
86     int score = 0, i;
87 
88     for (i = 0; i < size; i++)
89         score += (pix1[i] - pix2[i]) * (pix1[i] - pix2[i]);
90     return score;
91 }
92 
encode_block(SVQ1EncContext *s, uint8_t *src, uint8_t *ref, uint8_t *decoded, int stride, int level, int threshold, int lambda, int intra)93 static int encode_block(SVQ1EncContext *s, uint8_t *src, uint8_t *ref,
94                         uint8_t *decoded, int stride, int level,
95                         int threshold, int lambda, int intra)
96 {
97     int count, y, x, i, j, split, best_mean, best_score, best_count;
98     int best_vector[6];
99     int block_sum[7] = { 0, 0, 0, 0, 0, 0 };
100     int w            = 2 << (level + 2 >> 1);
101     int h            = 2 << (level + 1 >> 1);
102     int size         = w * h;
103     int16_t (*block)[256] = s->encoded_block_levels[level];
104     const int8_t *codebook_sum, *codebook;
105     const uint16_t(*mean_vlc)[2];
106     const uint8_t(*multistage_vlc)[2];
107 
108     best_score = 0;
109     // FIXME: Optimize, this does not need to be done multiple times.
110     if (intra) {
111         // level is 5 when encode_block is called from svq1_encode_plane
112         // and always < 4 when called recursively from this function.
113         codebook_sum   = level < 4 ? svq1_intra_codebook_sum[level] : NULL;
114         codebook       = ff_svq1_intra_codebooks[level];
115         mean_vlc       = ff_svq1_intra_mean_vlc;
116         multistage_vlc = ff_svq1_intra_multistage_vlc[level];
117         for (y = 0; y < h; y++) {
118             for (x = 0; x < w; x++) {
119                 int v = src[x + y * stride];
120                 block[0][x + w * y] = v;
121                 best_score         += v * v;
122                 block_sum[0]       += v;
123             }
124         }
125     } else {
126         // level is 5 or < 4, see above for details.
127         codebook_sum   = level < 4 ? svq1_inter_codebook_sum[level] : NULL;
128         codebook       = ff_svq1_inter_codebooks[level];
129         mean_vlc       = ff_svq1_inter_mean_vlc + 256;
130         multistage_vlc = ff_svq1_inter_multistage_vlc[level];
131         for (y = 0; y < h; y++) {
132             for (x = 0; x < w; x++) {
133                 int v = src[x + y * stride] - ref[x + y * stride];
134                 block[0][x + w * y] = v;
135                 best_score         += v * v;
136                 block_sum[0]       += v;
137             }
138         }
139     }
140 
141     best_count  = 0;
142     best_score -= (int)((unsigned)block_sum[0] * block_sum[0] >> (level + 3));
143     best_mean   = block_sum[0] + (size >> 1) >> (level + 3);
144 
145     if (level < 4) {
146         for (count = 1; count < 7; count++) {
147             int best_vector_score = INT_MAX;
148             int best_vector_sum   = -999, best_vector_mean = -999;
149             const int stage       = count - 1;
150             const int8_t *vector;
151 
152             for (i = 0; i < 16; i++) {
153                 int sum = codebook_sum[stage * 16 + i];
154                 int sqr, diff, score;
155 
156                 vector = codebook + stage * size * 16 + i * size;
157                 sqr    = s->ssd_int8_vs_int16(vector, block[stage], size);
158                 diff   = block_sum[stage] - sum;
159                 score  = sqr - (diff * (int64_t)diff >> (level + 3)); // FIXME: 64 bits slooow
160                 if (score < best_vector_score) {
161                     int mean = diff + (size >> 1) >> (level + 3);
162                     av_assert2(mean > -300 && mean < 300);
163                     mean               = av_clip(mean, intra ? 0 : -256, 255);
164                     best_vector_score  = score;
165                     best_vector[stage] = i;
166                     best_vector_sum    = sum;
167                     best_vector_mean   = mean;
168                 }
169             }
170             av_assert0(best_vector_mean != -999);
171             vector = codebook + stage * size * 16 + best_vector[stage] * size;
172             for (j = 0; j < size; j++)
173                 block[stage + 1][j] = block[stage][j] - vector[j];
174             block_sum[stage + 1] = block_sum[stage] - best_vector_sum;
175             best_vector_score   += lambda *
176                                    (+1 + 4 * count +
177                                     multistage_vlc[1 + count][1]
178                                     + mean_vlc[best_vector_mean][1]);
179 
180             if (best_vector_score < best_score) {
181                 best_score = best_vector_score;
182                 best_count = count;
183                 best_mean  = best_vector_mean;
184             }
185         }
186     }
187 
188     split = 0;
189     if (best_score > threshold && level) {
190         int score  = 0;
191         int offset = level & 1 ? stride * h / 2 : w / 2;
192         PutBitContext backup[6];
193 
194         for (i = level - 1; i >= 0; i--)
195             backup[i] = s->reorder_pb[i];
196         score += encode_block(s, src, ref, decoded, stride, level - 1,
197                               threshold >> 1, lambda, intra);
198         score += encode_block(s, src + offset, ref + offset, decoded + offset,
199                               stride, level - 1, threshold >> 1, lambda, intra);
200         score += lambda;
201 
202         if (score < best_score) {
203             best_score = score;
204             split      = 1;
205         } else {
206             for (i = level - 1; i >= 0; i--)
207                 s->reorder_pb[i] = backup[i];
208         }
209     }
210     if (level > 0)
211         put_bits(&s->reorder_pb[level], 1, split);
212 
213     if (!split) {
214         av_assert1(best_mean >= 0 && best_mean < 256 || !intra);
215         av_assert1(best_mean >= -256 && best_mean < 256);
216         av_assert1(best_count >= 0 && best_count < 7);
217         av_assert1(level < 4 || best_count == 0);
218 
219         /* output the encoding */
220         put_bits(&s->reorder_pb[level],
221                  multistage_vlc[1 + best_count][1],
222                  multistage_vlc[1 + best_count][0]);
223         put_bits(&s->reorder_pb[level], mean_vlc[best_mean][1],
224                  mean_vlc[best_mean][0]);
225 
226         for (i = 0; i < best_count; i++) {
227             av_assert2(best_vector[i] >= 0 && best_vector[i] < 16);
228             put_bits(&s->reorder_pb[level], 4, best_vector[i]);
229         }
230 
231         for (y = 0; y < h; y++)
232             for (x = 0; x < w; x++)
233                 decoded[x + y * stride] = src[x + y * stride] -
234                                           block[best_count][x + w * y] +
235                                           best_mean;
236     }
237 
238     return best_score;
239 }
240 
init_block_index(MpegEncContext *s)241 static void init_block_index(MpegEncContext *s){
242     s->block_index[0]= s->b8_stride*(s->mb_y*2    )     + s->mb_x*2;
243     s->block_index[1]= s->b8_stride*(s->mb_y*2    ) + 1 + s->mb_x*2;
244     s->block_index[2]= s->b8_stride*(s->mb_y*2 + 1)     + s->mb_x*2;
245     s->block_index[3]= s->b8_stride*(s->mb_y*2 + 1) + 1 + s->mb_x*2;
246     s->block_index[4]= s->mb_stride*(s->mb_y + 1)                + s->b8_stride*s->mb_height*2 + s->mb_x;
247     s->block_index[5]= s->mb_stride*(s->mb_y + s->mb_height + 2) + s->b8_stride*s->mb_height*2 + s->mb_x;
248 }
249 
svq1_encode_plane(SVQ1EncContext *s, int plane, unsigned char *src_plane, unsigned char *ref_plane, unsigned char *decoded_plane, int width, int height, int src_stride, int stride)250 static int svq1_encode_plane(SVQ1EncContext *s, int plane,
251                              unsigned char *src_plane,
252                              unsigned char *ref_plane,
253                              unsigned char *decoded_plane,
254                              int width, int height, int src_stride, int stride)
255 {
256     int x, y;
257     int i;
258     int block_width, block_height;
259     int level;
260     int threshold[6];
261     uint8_t *src     = s->scratchbuf + stride * 32;
262     const int lambda = (s->quality * s->quality) >>
263                        (2 * FF_LAMBDA_SHIFT);
264 
265     /* figure out the acceptable level thresholds in advance */
266     threshold[5] = QUALITY_THRESHOLD;
267     for (level = 4; level >= 0; level--)
268         threshold[level] = threshold[level + 1] * THRESHOLD_MULTIPLIER;
269 
270     block_width  = (width  + 15) / 16;
271     block_height = (height + 15) / 16;
272 
273     if (s->pict_type == AV_PICTURE_TYPE_P) {
274         s->m.avctx                         = s->avctx;
275         s->m.current_picture_ptr           = &s->m.current_picture;
276         s->m.last_picture_ptr              = &s->m.last_picture;
277         s->m.last_picture.f->data[0]        = ref_plane;
278         s->m.linesize                      =
279         s->m.last_picture.f->linesize[0]    =
280         s->m.new_picture->linesize[0]      =
281         s->m.current_picture.f->linesize[0] = stride;
282         s->m.width                         = width;
283         s->m.height                        = height;
284         s->m.mb_width                      = block_width;
285         s->m.mb_height                     = block_height;
286         s->m.mb_stride                     = s->m.mb_width + 1;
287         s->m.b8_stride                     = 2 * s->m.mb_width + 1;
288         s->m.f_code                        = 1;
289         s->m.pict_type                     = s->pict_type;
290         s->m.motion_est                    = s->motion_est;
291         s->m.me.scene_change_score         = 0;
292         // s->m.out_format                    = FMT_H263;
293         // s->m.unrestricted_mv               = 1;
294         s->m.lambda                        = s->quality;
295         s->m.qscale                        = s->m.lambda * 139 +
296                                              FF_LAMBDA_SCALE * 64 >>
297                                              FF_LAMBDA_SHIFT + 7;
298         s->m.lambda2                       = s->m.lambda * s->m.lambda +
299                                              FF_LAMBDA_SCALE / 2 >>
300                                              FF_LAMBDA_SHIFT;
301 
302         if (!s->motion_val8[plane]) {
303             s->motion_val8[plane]  = av_mallocz((s->m.b8_stride *
304                                                  block_height * 2 + 2) *
305                                                 2 * sizeof(int16_t));
306             s->motion_val16[plane] = av_mallocz((s->m.mb_stride *
307                                                  (block_height + 2) + 1) *
308                                                 2 * sizeof(int16_t));
309             if (!s->motion_val8[plane] || !s->motion_val16[plane])
310                 return AVERROR(ENOMEM);
311         }
312 
313         s->m.mb_type = s->mb_type;
314 
315         // dummies, to avoid segfaults
316         s->m.current_picture.mb_mean   = (uint8_t *)s->dummy;
317         s->m.current_picture.mb_var    = (uint16_t *)s->dummy;
318         s->m.current_picture.mc_mb_var = (uint16_t *)s->dummy;
319         s->m.current_picture.mb_type = s->dummy;
320 
321         s->m.current_picture.motion_val[0]   = s->motion_val8[plane] + 2;
322         s->m.p_mv_table                      = s->motion_val16[plane] +
323                                                s->m.mb_stride + 1;
324         s->m.mecc                            = s->mecc; // move
325         ff_init_me(&s->m);
326 
327         s->m.me.dia_size      = s->avctx->dia_size;
328         s->m.first_slice_line = 1;
329         for (y = 0; y < block_height; y++) {
330             s->m.new_picture->data[0]  = src - y * 16 * stride; // ugly
331             s->m.mb_y                  = y;
332 
333             for (i = 0; i < 16 && i + 16 * y < height; i++) {
334                 memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
335                        width);
336                 for (x = width; x < 16 * block_width; x++)
337                     src[i * stride + x] = src[i * stride + x - 1];
338             }
339             for (; i < 16 && i + 16 * y < 16 * block_height; i++)
340                 memcpy(&src[i * stride], &src[(i - 1) * stride],
341                        16 * block_width);
342 
343             for (x = 0; x < block_width; x++) {
344                 s->m.mb_x = x;
345                 init_block_index(&s->m);
346 
347                 ff_estimate_p_frame_motion(&s->m, x, y);
348             }
349             s->m.first_slice_line = 0;
350         }
351 
352         ff_fix_long_p_mvs(&s->m, CANDIDATE_MB_TYPE_INTRA);
353         ff_fix_long_mvs(&s->m, NULL, 0, s->m.p_mv_table, s->m.f_code,
354                         CANDIDATE_MB_TYPE_INTER, 0);
355     }
356 
357     s->m.first_slice_line = 1;
358     for (y = 0; y < block_height; y++) {
359         for (i = 0; i < 16 && i + 16 * y < height; i++) {
360             memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
361                    width);
362             for (x = width; x < 16 * block_width; x++)
363                 src[i * stride + x] = src[i * stride + x - 1];
364         }
365         for (; i < 16 && i + 16 * y < 16 * block_height; i++)
366             memcpy(&src[i * stride], &src[(i - 1) * stride], 16 * block_width);
367 
368         s->m.mb_y = y;
369         for (x = 0; x < block_width; x++) {
370             uint8_t reorder_buffer[2][6][7 * 32];
371             int count[2][6];
372             int offset       = y * 16 * stride + x * 16;
373             uint8_t *decoded = decoded_plane + offset;
374             uint8_t *ref     = ref_plane + offset;
375             int score[4]     = { 0, 0, 0, 0 }, best;
376             uint8_t *temp    = s->scratchbuf;
377 
378             if (put_bytes_left(&s->pb, 0) < 3000) { // FIXME: check size
379                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
380                 return -1;
381             }
382 
383             s->m.mb_x = x;
384             init_block_index(&s->m);
385 
386             if (s->pict_type == AV_PICTURE_TYPE_I ||
387                 (s->m.mb_type[x + y * s->m.mb_stride] &
388                  CANDIDATE_MB_TYPE_INTRA)) {
389                 for (i = 0; i < 6; i++)
390                     init_put_bits(&s->reorder_pb[i], reorder_buffer[0][i],
391                                   7 * 32);
392                 if (s->pict_type == AV_PICTURE_TYPE_P) {
393                     const uint8_t *vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_INTRA];
394                     put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
395                     score[0] = vlc[1] * lambda;
396                 }
397                 score[0] += encode_block(s, src + 16 * x, NULL, temp, stride,
398                                          5, 64, lambda, 1);
399                 for (i = 0; i < 6; i++) {
400                     count[0][i] = put_bits_count(&s->reorder_pb[i]);
401                     flush_put_bits(&s->reorder_pb[i]);
402                 }
403             } else
404                 score[0] = INT_MAX;
405 
406             best = 0;
407 
408             if (s->pict_type == AV_PICTURE_TYPE_P) {
409                 const uint8_t *vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_INTER];
410                 int mx, my, pred_x, pred_y, dxy;
411                 int16_t *motion_ptr;
412 
413                 motion_ptr = ff_h263_pred_motion(&s->m, 0, 0, &pred_x, &pred_y);
414                 if (s->m.mb_type[x + y * s->m.mb_stride] &
415                     CANDIDATE_MB_TYPE_INTER) {
416                     for (i = 0; i < 6; i++)
417                         init_put_bits(&s->reorder_pb[i], reorder_buffer[1][i],
418                                       7 * 32);
419 
420                     put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
421 
422                     s->m.pb = s->reorder_pb[5];
423                     mx      = motion_ptr[0];
424                     my      = motion_ptr[1];
425                     av_assert1(mx     >= -32 && mx     <= 31);
426                     av_assert1(my     >= -32 && my     <= 31);
427                     av_assert1(pred_x >= -32 && pred_x <= 31);
428                     av_assert1(pred_y >= -32 && pred_y <= 31);
429                     ff_h263_encode_motion(&s->m.pb, mx - pred_x, 1);
430                     ff_h263_encode_motion(&s->m.pb, my - pred_y, 1);
431                     s->reorder_pb[5] = s->m.pb;
432                     score[1]        += lambda * put_bits_count(&s->reorder_pb[5]);
433 
434                     dxy = (mx & 1) + 2 * (my & 1);
435 
436                     s->hdsp.put_pixels_tab[0][dxy](temp + 16*stride,
437                                                    ref + (mx >> 1) +
438                                                    stride * (my >> 1),
439                                                    stride, 16);
440 
441                     score[1] += encode_block(s, src + 16 * x, temp + 16*stride,
442                                              decoded, stride, 5, 64, lambda, 0);
443                     best      = score[1] <= score[0];
444 
445                     vlc       = ff_svq1_block_type_vlc[SVQ1_BLOCK_SKIP];
446                     score[2]  = s->mecc.sse[0](NULL, src + 16 * x, ref,
447                                                stride, 16);
448                     score[2] += vlc[1] * lambda;
449                     if (score[2] < score[best] && mx == 0 && my == 0) {
450                         best = 2;
451                         s->hdsp.put_pixels_tab[0][0](decoded, ref, stride, 16);
452                         put_bits(&s->pb, vlc[1], vlc[0]);
453                     }
454                 }
455 
456                 if (best == 1) {
457                     for (i = 0; i < 6; i++) {
458                         count[1][i] = put_bits_count(&s->reorder_pb[i]);
459                         flush_put_bits(&s->reorder_pb[i]);
460                     }
461                 } else {
462                     motion_ptr[0]                      =
463                     motion_ptr[1]                      =
464                     motion_ptr[2]                      =
465                     motion_ptr[3]                      =
466                     motion_ptr[0 + 2 * s->m.b8_stride] =
467                     motion_ptr[1 + 2 * s->m.b8_stride] =
468                     motion_ptr[2 + 2 * s->m.b8_stride] =
469                     motion_ptr[3 + 2 * s->m.b8_stride] = 0;
470                 }
471             }
472 
473             s->rd_total += score[best];
474 
475             if (best != 2)
476             for (i = 5; i >= 0; i--)
477                 ff_copy_bits(&s->pb, reorder_buffer[best][i],
478                                  count[best][i]);
479             if (best == 0)
480                 s->hdsp.put_pixels_tab[0][0](decoded, temp, stride, 16);
481         }
482         s->m.first_slice_line = 0;
483     }
484     return 0;
485 }
486 
svq1_encode_end(AVCodecContext *avctx)487 static av_cold int svq1_encode_end(AVCodecContext *avctx)
488 {
489     SVQ1EncContext *const s = avctx->priv_data;
490     int i;
491 
492     if (avctx->frame_number)
493         av_log(avctx, AV_LOG_DEBUG, "RD: %f\n",
494                s->rd_total / (double)(avctx->width * avctx->height *
495                                       avctx->frame_number));
496 
497     s->m.mb_type = NULL;
498     ff_mpv_common_end(&s->m);
499 
500     av_freep(&s->m.me.scratchpad);
501     av_freep(&s->m.me.map);
502     av_freep(&s->m.me.score_map);
503     av_freep(&s->mb_type);
504     av_freep(&s->dummy);
505     av_freep(&s->scratchbuf);
506 
507     for (i = 0; i < 3; i++) {
508         av_freep(&s->motion_val8[i]);
509         av_freep(&s->motion_val16[i]);
510     }
511 
512     av_frame_free(&s->current_picture);
513     av_frame_free(&s->last_picture);
514 
515     return 0;
516 }
517 
svq1_encode_init(AVCodecContext *avctx)518 static av_cold int svq1_encode_init(AVCodecContext *avctx)
519 {
520     SVQ1EncContext *const s = avctx->priv_data;
521     int ret;
522 
523     if (avctx->width >= 4096 || avctx->height >= 4096) {
524         av_log(avctx, AV_LOG_ERROR, "Dimensions too large, maximum is 4095x4095\n");
525         return AVERROR(EINVAL);
526     }
527 
528     ff_hpeldsp_init(&s->hdsp, avctx->flags);
529     ff_me_cmp_init(&s->mecc, avctx);
530     ff_mpegvideoencdsp_init(&s->m.mpvencdsp, avctx);
531 
532     s->current_picture = av_frame_alloc();
533     s->last_picture    = av_frame_alloc();
534     if (!s->current_picture || !s->last_picture) {
535         return AVERROR(ENOMEM);
536     }
537 
538     s->frame_width  = avctx->width;
539     s->frame_height = avctx->height;
540 
541     s->y_block_width  = (s->frame_width  + 15) / 16;
542     s->y_block_height = (s->frame_height + 15) / 16;
543 
544     s->c_block_width  = (s->frame_width  / 4 + 15) / 16;
545     s->c_block_height = (s->frame_height / 4 + 15) / 16;
546 
547     s->avctx               = avctx;
548     s->m.avctx             = avctx;
549 
550     if ((ret = ff_mpv_common_init(&s->m)) < 0) {
551         return ret;
552     }
553 
554     s->m.picture_structure = PICT_FRAME;
555     s->m.me.temp           =
556     s->m.me.scratchpad     = av_mallocz((avctx->width + 64) *
557                                         2 * 16 * 2 * sizeof(uint8_t));
558     s->m.me.map            = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
559     s->m.me.score_map      = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
560     s->mb_type             = av_mallocz((s->y_block_width + 1) *
561                                         s->y_block_height * sizeof(int16_t));
562     s->dummy               = av_mallocz((s->y_block_width + 1) *
563                                         s->y_block_height * sizeof(int32_t));
564     s->ssd_int8_vs_int16   = ssd_int8_vs_int16_c;
565 
566     if (!s->m.me.temp || !s->m.me.scratchpad || !s->m.me.map ||
567         !s->m.me.score_map || !s->mb_type || !s->dummy) {
568         return AVERROR(ENOMEM);
569     }
570 
571 #if ARCH_PPC
572     ff_svq1enc_init_ppc(s);
573 #elif ARCH_X86
574     ff_svq1enc_init_x86(s);
575 #endif
576 
577     ff_h263_encode_init(&s->m); // mv_penalty
578 
579     return 0;
580 }
581 
svq1_encode_frame(AVCodecContext *avctx, AVPacket *pkt, const AVFrame *pict, int *got_packet)582 static int svq1_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
583                              const AVFrame *pict, int *got_packet)
584 {
585     SVQ1EncContext *const s = avctx->priv_data;
586     int i, ret;
587 
588     ret = ff_alloc_packet(avctx, pkt, s->y_block_width * s->y_block_height *
589                           MAX_MB_BYTES * 3 + AV_INPUT_BUFFER_MIN_SIZE);
590     if (ret < 0)
591         return ret;
592 
593     if (avctx->pix_fmt != AV_PIX_FMT_YUV410P) {
594         av_log(avctx, AV_LOG_ERROR, "unsupported pixel format\n");
595         return -1;
596     }
597 
598     if (!s->current_picture->data[0]) {
599         if ((ret = ff_encode_alloc_frame(avctx, s->current_picture)) < 0) {
600             return ret;
601         }
602     }
603     if (!s->last_picture->data[0]) {
604         ret = ff_encode_alloc_frame(avctx, s->last_picture);
605         if (ret < 0)
606             return ret;
607     }
608     if (!s->scratchbuf) {
609         s->scratchbuf = av_malloc_array(s->current_picture->linesize[0], 16 * 3);
610         if (!s->scratchbuf)
611             return AVERROR(ENOMEM);
612     }
613 
614     FFSWAP(AVFrame*, s->current_picture, s->last_picture);
615 
616     init_put_bits(&s->pb, pkt->data, pkt->size);
617 
618     if (avctx->gop_size && (avctx->frame_number % avctx->gop_size))
619         s->pict_type = AV_PICTURE_TYPE_P;
620     else
621         s->pict_type = AV_PICTURE_TYPE_I;
622     s->quality = pict->quality;
623 
624     ff_side_data_set_encoder_stats(pkt, pict->quality, NULL, 0, s->pict_type);
625 
626     svq1_write_header(s, s->pict_type);
627     for (i = 0; i < 3; i++) {
628         int ret = svq1_encode_plane(s, i,
629                               pict->data[i],
630                               s->last_picture->data[i],
631                               s->current_picture->data[i],
632                               s->frame_width  / (i ? 4 : 1),
633                               s->frame_height / (i ? 4 : 1),
634                               pict->linesize[i],
635                               s->current_picture->linesize[i]);
636         emms_c();
637         if (ret < 0) {
638             int j;
639             for (j = 0; j < i; j++) {
640                 av_freep(&s->motion_val8[j]);
641                 av_freep(&s->motion_val16[j]);
642             }
643             av_freep(&s->scratchbuf);
644             return -1;
645         }
646     }
647 
648     // align_put_bits(&s->pb);
649     while (put_bits_count(&s->pb) & 31)
650         put_bits(&s->pb, 1, 0);
651 
652     flush_put_bits(&s->pb);
653 
654     pkt->size = put_bytes_output(&s->pb);
655     if (s->pict_type == AV_PICTURE_TYPE_I)
656         pkt->flags |= AV_PKT_FLAG_KEY;
657     *got_packet = 1;
658 
659     return 0;
660 }
661 
662 #define OFFSET(x) offsetof(struct SVQ1EncContext, x)
663 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
664 static const AVOption options[] = {
665     { "motion-est", "Motion estimation algorithm", OFFSET(motion_est), AV_OPT_TYPE_INT, { .i64 = FF_ME_EPZS }, FF_ME_ZERO, FF_ME_XONE, VE, "motion-est"},
666         { "zero", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ZERO }, 0, 0, FF_MPV_OPT_FLAGS, "motion-est" },
667         { "epzs", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_EPZS }, 0, 0, FF_MPV_OPT_FLAGS, "motion-est" },
668         { "xone", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_XONE }, 0, 0, FF_MPV_OPT_FLAGS, "motion-est" },
669 
670     { NULL },
671 };
672 
673 static const AVClass svq1enc_class = {
674     .class_name = "svq1enc",
675     .item_name  = av_default_item_name,
676     .option     = options,
677     .version    = LIBAVUTIL_VERSION_INT,
678 };
679 
680 const FFCodec ff_svq1_encoder = {
681     .p.name         = "svq1",
682     .p.long_name    = NULL_IF_CONFIG_SMALL("Sorenson Vector Quantizer 1 / Sorenson Video 1 / SVQ1"),
683     .p.type         = AVMEDIA_TYPE_VIDEO,
684     .p.id           = AV_CODEC_ID_SVQ1,
685     .priv_data_size = sizeof(SVQ1EncContext),
686     .p.priv_class   = &svq1enc_class,
687     .init           = svq1_encode_init,
688     FF_CODEC_ENCODE_CB(svq1_encode_frame),
689     .close          = svq1_encode_end,
690     .p.pix_fmts     = (const enum AVPixelFormat[]) { AV_PIX_FMT_YUV410P,
691                                                      AV_PIX_FMT_NONE },
692     .caps_internal  = FF_CODEC_CAP_INIT_THREADSAFE | FF_CODEC_CAP_INIT_CLEANUP,
693 };
694