3 * Copyright (C) 2004 Mike Melanson <melanson@pcisys.net>
5 * This file is part of FFmpeg.
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.
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.
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
24 * Sorenson Vector Quantizer #1 (SVQ1) video codec.
25 * For more information of the SVQ1 algorithm, visit:
26 * http://www.pcisys.net/~melanson/codecs/
32 #include "mpegvideo.h"
35 #include "mpegutils.h"
38 #include "svq1enc_cb.h"
39 #include "libavutil/avassert.h"
42 static void svq1_write_header(SVQ1EncContext *s, int frame_type)
47 put_bits(&s->pb, 22, 0x20);
49 /* temporal reference (sure hope this is a "don't care") */
50 put_bits(&s->pb, 8, 0x00);
53 put_bits(&s->pb, 2, frame_type - 1);
55 if (frame_type == AV_PICTURE_TYPE_I) {
56 /* no checksum since frame code is 0x20 */
57 /* no embedded string either */
58 /* output 5 unknown bits (2 + 2 + 1) */
59 put_bits(&s->pb, 5, 2); /* 2 needed by quicktime decoder */
61 i = ff_match_2uint16((void*)ff_svq1_frame_size_table,
62 FF_ARRAY_ELEMS(ff_svq1_frame_size_table),
63 s->frame_width, s->frame_height);
64 put_bits(&s->pb, 3, i);
67 put_bits(&s->pb, 12, s->frame_width);
68 put_bits(&s->pb, 12, s->frame_height);
72 /* no checksum or extra data (next 2 bits get 0) */
73 put_bits(&s->pb, 2, 0);
76 #define QUALITY_THRESHOLD 100
77 #define THRESHOLD_MULTIPLIER 0.6
79 static int ssd_int8_vs_int16_c(const int8_t *pix1, const int16_t *pix2,
84 for (i = 0; i < size; i++)
85 score += (pix1[i] - pix2[i]) * (pix1[i] - pix2[i]);
89 static int encode_block(SVQ1EncContext *s, uint8_t *src, uint8_t *ref,
90 uint8_t *decoded, int stride, int level,
91 int threshold, int lambda, int intra)
93 int count, y, x, i, j, split, best_mean, best_score, best_count;
95 int block_sum[7] = { 0, 0, 0, 0, 0, 0 };
96 int w = 2 << (level + 2 >> 1);
97 int h = 2 << (level + 1 >> 1);
99 int16_t (*block)[256] = s->encoded_block_levels[level];
100 const int8_t *codebook_sum, *codebook;
101 const uint16_t(*mean_vlc)[2];
102 const uint8_t(*multistage_vlc)[2];
105 // FIXME: Optimize, this does not need to be done multiple times.
107 codebook_sum = svq1_intra_codebook_sum[level];
108 codebook = ff_svq1_intra_codebooks[level];
109 mean_vlc = ff_svq1_intra_mean_vlc;
110 multistage_vlc = ff_svq1_intra_multistage_vlc[level];
111 for (y = 0; y < h; y++) {
112 for (x = 0; x < w; x++) {
113 int v = src[x + y * stride];
114 block[0][x + w * y] = v;
120 codebook_sum = svq1_inter_codebook_sum[level];
121 codebook = ff_svq1_inter_codebooks[level];
122 mean_vlc = ff_svq1_inter_mean_vlc + 256;
123 multistage_vlc = ff_svq1_inter_multistage_vlc[level];
124 for (y = 0; y < h; y++) {
125 for (x = 0; x < w; x++) {
126 int v = src[x + y * stride] - ref[x + y * stride];
127 block[0][x + w * y] = v;
135 best_score -= (int)((unsigned)block_sum[0] * block_sum[0] >> (level + 3));
136 best_mean = block_sum[0] + (size >> 1) >> (level + 3);
139 for (count = 1; count < 7; count++) {
140 int best_vector_score = INT_MAX;
141 int best_vector_sum = -999, best_vector_mean = -999;
142 const int stage = count - 1;
143 const int8_t *vector;
145 for (i = 0; i < 16; i++) {
146 int sum = codebook_sum[stage * 16 + i];
147 int sqr, diff, score;
149 vector = codebook + stage * size * 16 + i * size;
150 sqr = s->ssd_int8_vs_int16(vector, block[stage], size);
151 diff = block_sum[stage] - sum;
152 score = sqr - (diff * (int64_t)diff >> (level + 3)); // FIXME: 64bit slooow
153 if (score < best_vector_score) {
154 int mean = diff + (size >> 1) >> (level + 3);
155 av_assert2(mean > -300 && mean < 300);
156 mean = av_clip(mean, intra ? 0 : -256, 255);
157 best_vector_score = score;
158 best_vector[stage] = i;
159 best_vector_sum = sum;
160 best_vector_mean = mean;
163 av_assert0(best_vector_mean != -999);
164 vector = codebook + stage * size * 16 + best_vector[stage] * size;
165 for (j = 0; j < size; j++)
166 block[stage + 1][j] = block[stage][j] - vector[j];
167 block_sum[stage + 1] = block_sum[stage] - best_vector_sum;
168 best_vector_score += lambda *
170 multistage_vlc[1 + count][1]
171 + mean_vlc[best_vector_mean][1]);
173 if (best_vector_score < best_score) {
174 best_score = best_vector_score;
176 best_mean = best_vector_mean;
182 if (best_score > threshold && level) {
184 int offset = level & 1 ? stride * h / 2 : w / 2;
185 PutBitContext backup[6];
187 for (i = level - 1; i >= 0; i--)
188 backup[i] = s->reorder_pb[i];
189 score += encode_block(s, src, ref, decoded, stride, level - 1,
190 threshold >> 1, lambda, intra);
191 score += encode_block(s, src + offset, ref + offset, decoded + offset,
192 stride, level - 1, threshold >> 1, lambda, intra);
195 if (score < best_score) {
199 for (i = level - 1; i >= 0; i--)
200 s->reorder_pb[i] = backup[i];
204 put_bits(&s->reorder_pb[level], 1, split);
207 av_assert1(best_mean >= 0 && best_mean < 256 || !intra);
208 av_assert1(best_mean >= -256 && best_mean < 256);
209 av_assert1(best_count >= 0 && best_count < 7);
210 av_assert1(level < 4 || best_count == 0);
212 /* output the encoding */
213 put_bits(&s->reorder_pb[level],
214 multistage_vlc[1 + best_count][1],
215 multistage_vlc[1 + best_count][0]);
216 put_bits(&s->reorder_pb[level], mean_vlc[best_mean][1],
217 mean_vlc[best_mean][0]);
219 for (i = 0; i < best_count; i++) {
220 av_assert2(best_vector[i] >= 0 && best_vector[i] < 16);
221 put_bits(&s->reorder_pb[level], 4, best_vector[i]);
224 for (y = 0; y < h; y++)
225 for (x = 0; x < w; x++)
226 decoded[x + y * stride] = src[x + y * stride] -
227 block[best_count][x + w * y] +
234 static void init_block_index(MpegEncContext *s){
235 s->block_index[0]= s->b8_stride*(s->mb_y*2 ) + s->mb_x*2;
236 s->block_index[1]= s->b8_stride*(s->mb_y*2 ) + 1 + s->mb_x*2;
237 s->block_index[2]= s->b8_stride*(s->mb_y*2 + 1) + s->mb_x*2;
238 s->block_index[3]= s->b8_stride*(s->mb_y*2 + 1) + 1 + s->mb_x*2;
239 s->block_index[4]= s->mb_stride*(s->mb_y + 1) + s->b8_stride*s->mb_height*2 + s->mb_x;
240 s->block_index[5]= s->mb_stride*(s->mb_y + s->mb_height + 2) + s->b8_stride*s->mb_height*2 + s->mb_x;
243 static int svq1_encode_plane(SVQ1EncContext *s, int plane,
244 unsigned char *src_plane,
245 unsigned char *ref_plane,
246 unsigned char *decoded_plane,
247 int width, int height, int src_stride, int stride)
249 const AVFrame *f = s->avctx->coded_frame;
252 int block_width, block_height;
255 uint8_t *src = s->scratchbuf + stride * 32;
256 const int lambda = (f->quality * f->quality) >>
257 (2 * FF_LAMBDA_SHIFT);
259 /* figure out the acceptable level thresholds in advance */
260 threshold[5] = QUALITY_THRESHOLD;
261 for (level = 4; level >= 0; level--)
262 threshold[level] = threshold[level + 1] * THRESHOLD_MULTIPLIER;
264 block_width = (width + 15) / 16;
265 block_height = (height + 15) / 16;
267 if (f->pict_type == AV_PICTURE_TYPE_P) {
268 s->m.avctx = s->avctx;
269 s->m.current_picture_ptr = &s->m.current_picture;
270 s->m.last_picture_ptr = &s->m.last_picture;
271 s->m.last_picture.f->data[0] = ref_plane;
273 s->m.last_picture.f->linesize[0] =
274 s->m.new_picture.f->linesize[0] =
275 s->m.current_picture.f->linesize[0] = stride;
277 s->m.height = height;
278 s->m.mb_width = block_width;
279 s->m.mb_height = block_height;
280 s->m.mb_stride = s->m.mb_width + 1;
281 s->m.b8_stride = 2 * s->m.mb_width + 1;
283 s->m.pict_type = f->pict_type;
284 s->m.me_method = s->avctx->me_method;
285 s->m.me.scene_change_score = 0;
286 s->m.flags = s->avctx->flags;
287 // s->m.out_format = FMT_H263;
288 // s->m.unrestricted_mv = 1;
289 s->m.lambda = f->quality;
290 s->m.qscale = s->m.lambda * 139 +
291 FF_LAMBDA_SCALE * 64 >>
293 s->m.lambda2 = s->m.lambda * s->m.lambda +
294 FF_LAMBDA_SCALE / 2 >>
297 if (!s->motion_val8[plane]) {
298 s->motion_val8[plane] = av_mallocz((s->m.b8_stride *
299 block_height * 2 + 2) *
300 2 * sizeof(int16_t));
301 s->motion_val16[plane] = av_mallocz((s->m.mb_stride *
302 (block_height + 2) + 1) *
303 2 * sizeof(int16_t));
304 if (!s->motion_val8[plane] || !s->motion_val16[plane])
305 return AVERROR(ENOMEM);
308 s->m.mb_type = s->mb_type;
310 // dummies, to avoid segfaults
311 s->m.current_picture.mb_mean = (uint8_t *)s->dummy;
312 s->m.current_picture.mb_var = (uint16_t *)s->dummy;
313 s->m.current_picture.mc_mb_var = (uint16_t *)s->dummy;
314 s->m.current_picture.mb_type = s->dummy;
316 s->m.current_picture.motion_val[0] = s->motion_val8[plane] + 2;
317 s->m.p_mv_table = s->motion_val16[plane] +
319 s->m.mecc = s->mecc; // move
322 s->m.me.dia_size = s->avctx->dia_size;
323 s->m.first_slice_line = 1;
324 for (y = 0; y < block_height; y++) {
325 s->m.new_picture.f->data[0] = src - y * 16 * stride; // ugly
328 for (i = 0; i < 16 && i + 16 * y < height; i++) {
329 memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
331 for (x = width; x < 16 * block_width; x++)
332 src[i * stride + x] = src[i * stride + x - 1];
334 for (; i < 16 && i + 16 * y < 16 * block_height; i++)
335 memcpy(&src[i * stride], &src[(i - 1) * stride],
338 for (x = 0; x < block_width; x++) {
340 init_block_index(&s->m);
342 ff_estimate_p_frame_motion(&s->m, x, y);
344 s->m.first_slice_line = 0;
347 ff_fix_long_p_mvs(&s->m);
348 ff_fix_long_mvs(&s->m, NULL, 0, s->m.p_mv_table, s->m.f_code,
349 CANDIDATE_MB_TYPE_INTER, 0);
352 s->m.first_slice_line = 1;
353 for (y = 0; y < block_height; y++) {
354 for (i = 0; i < 16 && i + 16 * y < height; i++) {
355 memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
357 for (x = width; x < 16 * block_width; x++)
358 src[i * stride + x] = src[i * stride + x - 1];
360 for (; i < 16 && i + 16 * y < 16 * block_height; i++)
361 memcpy(&src[i * stride], &src[(i - 1) * stride], 16 * block_width);
364 for (x = 0; x < block_width; x++) {
365 uint8_t reorder_buffer[2][6][7 * 32];
367 int offset = y * 16 * stride + x * 16;
368 uint8_t *decoded = decoded_plane + offset;
369 uint8_t *ref = ref_plane + offset;
370 int score[4] = { 0, 0, 0, 0 }, best;
371 uint8_t *temp = s->scratchbuf;
373 if (s->pb.buf_end - s->pb.buf -
374 (put_bits_count(&s->pb) >> 3) < 3000) { // FIXME: check size
375 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
380 init_block_index(&s->m);
382 if (f->pict_type == AV_PICTURE_TYPE_I ||
383 (s->m.mb_type[x + y * s->m.mb_stride] &
384 CANDIDATE_MB_TYPE_INTRA)) {
385 for (i = 0; i < 6; i++)
386 init_put_bits(&s->reorder_pb[i], reorder_buffer[0][i],
388 if (f->pict_type == AV_PICTURE_TYPE_P) {
389 const uint8_t *vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_INTRA];
390 put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
391 score[0] = vlc[1] * lambda;
393 score[0] += encode_block(s, src + 16 * x, NULL, temp, stride,
395 for (i = 0; i < 6; i++) {
396 count[0][i] = put_bits_count(&s->reorder_pb[i]);
397 flush_put_bits(&s->reorder_pb[i]);
404 if (f->pict_type == AV_PICTURE_TYPE_P) {
405 const uint8_t *vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_INTER];
406 int mx, my, pred_x, pred_y, dxy;
409 motion_ptr = ff_h263_pred_motion(&s->m, 0, 0, &pred_x, &pred_y);
410 if (s->m.mb_type[x + y * s->m.mb_stride] &
411 CANDIDATE_MB_TYPE_INTER) {
412 for (i = 0; i < 6; i++)
413 init_put_bits(&s->reorder_pb[i], reorder_buffer[1][i],
416 put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
418 s->m.pb = s->reorder_pb[5];
421 av_assert1(mx >= -32 && mx <= 31);
422 av_assert1(my >= -32 && my <= 31);
423 av_assert1(pred_x >= -32 && pred_x <= 31);
424 av_assert1(pred_y >= -32 && pred_y <= 31);
425 ff_h263_encode_motion(&s->m, mx - pred_x, 1);
426 ff_h263_encode_motion(&s->m, my - pred_y, 1);
427 s->reorder_pb[5] = s->m.pb;
428 score[1] += lambda * put_bits_count(&s->reorder_pb[5]);
430 dxy = (mx & 1) + 2 * (my & 1);
432 s->hdsp.put_pixels_tab[0][dxy](temp + 16*stride,
437 score[1] += encode_block(s, src + 16 * x, temp + 16*stride,
438 decoded, stride, 5, 64, lambda, 0);
439 best = score[1] <= score[0];
441 vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_SKIP];
442 score[2] = s->mecc.sse[0](NULL, src + 16 * x, ref,
444 score[2] += vlc[1] * lambda;
445 if (score[2] < score[best] && mx == 0 && my == 0) {
447 s->hdsp.put_pixels_tab[0][0](decoded, ref, stride, 16);
448 put_bits(&s->pb, vlc[1], vlc[0]);
453 for (i = 0; i < 6; i++) {
454 count[1][i] = put_bits_count(&s->reorder_pb[i]);
455 flush_put_bits(&s->reorder_pb[i]);
462 motion_ptr[0 + 2 * s->m.b8_stride] =
463 motion_ptr[1 + 2 * s->m.b8_stride] =
464 motion_ptr[2 + 2 * s->m.b8_stride] =
465 motion_ptr[3 + 2 * s->m.b8_stride] = 0;
469 s->rd_total += score[best];
472 for (i = 5; i >= 0; i--)
473 avpriv_copy_bits(&s->pb, reorder_buffer[best][i],
476 s->hdsp.put_pixels_tab[0][0](decoded, temp, stride, 16);
478 s->m.first_slice_line = 0;
483 static av_cold int svq1_encode_end(AVCodecContext *avctx)
485 SVQ1EncContext *const s = avctx->priv_data;
488 av_log(avctx, AV_LOG_DEBUG, "RD: %f\n",
489 s->rd_total / (double)(avctx->width * avctx->height *
490 avctx->frame_number));
493 ff_mpv_common_end(&s->m);
495 av_freep(&s->m.me.scratchpad);
496 av_freep(&s->m.me.map);
497 av_freep(&s->m.me.score_map);
498 av_freep(&s->mb_type);
500 av_freep(&s->scratchbuf);
502 for (i = 0; i < 3; i++) {
503 av_freep(&s->motion_val8[i]);
504 av_freep(&s->motion_val16[i]);
507 av_frame_free(&s->current_picture);
508 av_frame_free(&s->last_picture);
509 av_frame_free(&avctx->coded_frame);
514 static av_cold int svq1_encode_init(AVCodecContext *avctx)
516 SVQ1EncContext *const s = avctx->priv_data;
519 ff_hpeldsp_init(&s->hdsp, avctx->flags);
520 ff_me_cmp_init(&s->mecc, avctx);
521 ff_mpegvideoencdsp_init(&s->m.mpvencdsp, avctx);
523 avctx->coded_frame = av_frame_alloc();
524 s->current_picture = av_frame_alloc();
525 s->last_picture = av_frame_alloc();
526 if (!avctx->coded_frame || !s->current_picture || !s->last_picture) {
527 svq1_encode_end(avctx);
528 return AVERROR(ENOMEM);
531 s->frame_width = avctx->width;
532 s->frame_height = avctx->height;
534 s->y_block_width = (s->frame_width + 15) / 16;
535 s->y_block_height = (s->frame_height + 15) / 16;
537 s->c_block_width = (s->frame_width / 4 + 15) / 16;
538 s->c_block_height = (s->frame_height / 4 + 15) / 16;
543 if ((ret = ff_mpv_common_init(&s->m)) < 0) {
544 svq1_encode_end(avctx);
548 s->m.picture_structure = PICT_FRAME;
550 s->m.me.scratchpad = av_mallocz((avctx->width + 64) *
551 2 * 16 * 2 * sizeof(uint8_t));
552 s->m.me.map = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
553 s->m.me.score_map = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
554 s->mb_type = av_mallocz((s->y_block_width + 1) *
555 s->y_block_height * sizeof(int16_t));
556 s->dummy = av_mallocz((s->y_block_width + 1) *
557 s->y_block_height * sizeof(int32_t));
558 s->ssd_int8_vs_int16 = ssd_int8_vs_int16_c;
560 if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map ||
561 !s->mb_type || !s->dummy) {
562 svq1_encode_end(avctx);
563 return AVERROR(ENOMEM);
567 ff_svq1enc_init_ppc(s);
569 ff_svq1enc_init_x86(s);
571 ff_h263_encode_init(&s->m); // mv_penalty
576 static int svq1_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
577 const AVFrame *pict, int *got_packet)
579 SVQ1EncContext *const s = avctx->priv_data;
580 AVFrame *const p = avctx->coded_frame;
583 if ((ret = ff_alloc_packet2(avctx, pkt, s->y_block_width * s->y_block_height *
584 MAX_MB_BYTES*3 + FF_MIN_BUFFER_SIZE)) < 0)
587 if (avctx->pix_fmt != AV_PIX_FMT_YUV410P) {
588 av_log(avctx, AV_LOG_ERROR, "unsupported pixel format\n");
592 if (!s->current_picture->data[0]) {
593 if ((ret = ff_get_buffer(avctx, s->current_picture, 0))< 0 ||
594 (ret = ff_get_buffer(avctx, s->last_picture, 0)) < 0) {
597 s->scratchbuf = av_malloc(s->current_picture->linesize[0] * 16 * 3);
600 return AVERROR(ENOMEM);
602 FFSWAP(AVFrame*, s->current_picture, s->last_picture);
604 init_put_bits(&s->pb, pkt->data, pkt->size);
606 p->pict_type = avctx->gop_size && avctx->frame_number % avctx->gop_size ?
607 AV_PICTURE_TYPE_P : AV_PICTURE_TYPE_I;
608 p->key_frame = p->pict_type == AV_PICTURE_TYPE_I;
609 p->quality = pict->quality;
611 svq1_write_header(s, p->pict_type);
612 for (i = 0; i < 3; i++)
613 if (svq1_encode_plane(s, i,
615 s->last_picture->data[i],
616 s->current_picture->data[i],
617 s->frame_width / (i ? 4 : 1),
618 s->frame_height / (i ? 4 : 1),
620 s->current_picture->linesize[i]) < 0)
623 // avpriv_align_put_bits(&s->pb);
624 while (put_bits_count(&s->pb) & 31)
625 put_bits(&s->pb, 1, 0);
627 flush_put_bits(&s->pb);
629 pkt->size = put_bits_count(&s->pb) / 8;
630 if (p->pict_type == AV_PICTURE_TYPE_I)
631 pkt->flags |= AV_PKT_FLAG_KEY;
637 AVCodec ff_svq1_encoder = {
639 .long_name = NULL_IF_CONFIG_SMALL("Sorenson Vector Quantizer 1 / Sorenson Video 1 / SVQ1"),
640 .type = AVMEDIA_TYPE_VIDEO,
641 .id = AV_CODEC_ID_SVQ1,
642 .priv_data_size = sizeof(SVQ1EncContext),
643 .init = svq1_encode_init,
644 .encode2 = svq1_encode_frame,
645 .close = svq1_encode_end,
646 .pix_fmts = (const enum AVPixelFormat[]) { AV_PIX_FMT_YUV410P,