3 * Copyright (C) 2004 Mike Melanson <melanson@pcisys.net>
5 * This file is part of Libav.
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.
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.
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
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"
36 #include "svq1enc_cb.h"
41 typedef struct SVQ1Context {
42 /* FIXME: Needed for motion estimation, should not be used for anything
43 * else, the idea is to make the motion estimation eventually independent
44 * of MpegEncContext, so this will be removed then. */
46 AVCodecContext *avctx;
50 AVFrame current_picture;
55 /* why ooh why this sick breadth first order,
56 * everything is slower and more complex */
57 PutBitContext reorder_pb[6];
62 /* Y plane block dimensions */
66 /* U & V plane (C planes) block dimensions */
72 int16_t (*motion_val8[3])[2];
73 int16_t (*motion_val16[3])[2];
80 static void svq1_write_header(SVQ1Context *s, int frame_type)
85 put_bits(&s->pb, 22, 0x20);
87 /* temporal reference (sure hope this is a "don't care") */
88 put_bits(&s->pb, 8, 0x00);
91 put_bits(&s->pb, 2, frame_type - 1);
93 if (frame_type == AV_PICTURE_TYPE_I) {
94 /* no checksum since frame code is 0x20 */
95 /* no embedded string either */
96 /* output 5 unknown bits (2 + 2 + 1) */
97 put_bits(&s->pb, 5, 2); /* 2 needed by quicktime decoder */
99 i = ff_match_2uint16(ff_svq1_frame_size_table,
100 FF_ARRAY_ELEMS(ff_svq1_frame_size_table),
101 s->frame_width, s->frame_height);
102 put_bits(&s->pb, 3, i);
105 put_bits(&s->pb, 12, s->frame_width);
106 put_bits(&s->pb, 12, s->frame_height);
110 /* no checksum or extra data (next 2 bits get 0) */
111 put_bits(&s->pb, 2, 0);
114 #define QUALITY_THRESHOLD 100
115 #define THRESHOLD_MULTIPLIER 0.6
117 static int encode_block(SVQ1Context *s, uint8_t *src, uint8_t *ref,
118 uint8_t *decoded, int stride, int level,
119 int threshold, int lambda, int intra)
121 int count, y, x, i, j, split, best_mean, best_score, best_count;
123 int block_sum[7] = { 0, 0, 0, 0, 0, 0 };
124 int w = 2 << (level + 2 >> 1);
125 int h = 2 << (level + 1 >> 1);
127 int16_t block[7][256];
128 const int8_t *codebook_sum, *codebook;
129 const uint16_t(*mean_vlc)[2];
130 const uint8_t(*multistage_vlc)[2];
133 // FIXME: Optimize, this does not need to be done multiple times.
135 codebook_sum = svq1_intra_codebook_sum[level];
136 codebook = ff_svq1_intra_codebooks[level];
137 mean_vlc = ff_svq1_intra_mean_vlc;
138 multistage_vlc = ff_svq1_intra_multistage_vlc[level];
139 for (y = 0; y < h; y++) {
140 for (x = 0; x < w; x++) {
141 int v = src[x + y * stride];
142 block[0][x + w * y] = v;
148 codebook_sum = svq1_inter_codebook_sum[level];
149 codebook = ff_svq1_inter_codebooks[level];
150 mean_vlc = ff_svq1_inter_mean_vlc + 256;
151 multistage_vlc = ff_svq1_inter_multistage_vlc[level];
152 for (y = 0; y < h; y++) {
153 for (x = 0; x < w; x++) {
154 int v = src[x + y * stride] - ref[x + y * stride];
155 block[0][x + w * y] = v;
163 best_score -= (int)((unsigned)block_sum[0] * block_sum[0] >> (level + 3));
164 best_mean = block_sum[0] + (size >> 1) >> (level + 3);
167 for (count = 1; count < 7; count++) {
168 int best_vector_score = INT_MAX;
169 int best_vector_sum = -999, best_vector_mean = -999;
170 const int stage = count - 1;
171 const int8_t *vector;
173 for (i = 0; i < 16; i++) {
174 int sum = codebook_sum[stage * 16 + i];
175 int sqr, diff, score;
177 vector = codebook + stage * size * 16 + i * size;
178 sqr = s->dsp.ssd_int8_vs_int16(vector, block[stage], size);
179 diff = block_sum[stage] - sum;
180 score = sqr - (diff * (int64_t)diff >> (level + 3)); // FIXME: 64bit slooow
181 if (score < best_vector_score) {
182 int mean = diff + (size >> 1) >> (level + 3);
183 assert(mean > -300 && mean < 300);
184 mean = av_clip(mean, intra ? 0 : -256, 255);
185 best_vector_score = score;
186 best_vector[stage] = i;
187 best_vector_sum = sum;
188 best_vector_mean = mean;
191 assert(best_vector_mean != -999);
192 vector = codebook + stage * size * 16 + best_vector[stage] * size;
193 for (j = 0; j < size; j++)
194 block[stage + 1][j] = block[stage][j] - vector[j];
195 block_sum[stage + 1] = block_sum[stage] - best_vector_sum;
196 best_vector_score += lambda *
198 multistage_vlc[1 + count][1]
199 + mean_vlc[best_vector_mean][1]);
201 if (best_vector_score < best_score) {
202 best_score = best_vector_score;
204 best_mean = best_vector_mean;
210 if (best_score > threshold && level) {
212 int offset = level & 1 ? stride * h / 2 : w / 2;
213 PutBitContext backup[6];
215 for (i = level - 1; i >= 0; i--)
216 backup[i] = s->reorder_pb[i];
217 score += encode_block(s, src, ref, decoded, stride, level - 1,
218 threshold >> 1, lambda, intra);
219 score += encode_block(s, src + offset, ref + offset, decoded + offset,
220 stride, level - 1, threshold >> 1, lambda, intra);
223 if (score < best_score) {
227 for (i = level - 1; i >= 0; i--)
228 s->reorder_pb[i] = backup[i];
232 put_bits(&s->reorder_pb[level], 1, split);
235 assert(best_mean >= 0 && best_mean < 256 || !intra);
236 assert(best_mean >= -256 && best_mean < 256);
237 assert(best_count >= 0 && best_count < 7);
238 assert(level < 4 || best_count == 0);
240 /* output the encoding */
241 put_bits(&s->reorder_pb[level],
242 multistage_vlc[1 + best_count][1],
243 multistage_vlc[1 + best_count][0]);
244 put_bits(&s->reorder_pb[level], mean_vlc[best_mean][1],
245 mean_vlc[best_mean][0]);
247 for (i = 0; i < best_count; i++) {
248 assert(best_vector[i] >= 0 && best_vector[i] < 16);
249 put_bits(&s->reorder_pb[level], 4, best_vector[i]);
252 for (y = 0; y < h; y++)
253 for (x = 0; x < w; x++)
254 decoded[x + y * stride] = src[x + y * stride] -
255 block[best_count][x + w * y] +
262 static int svq1_encode_plane(SVQ1Context *s, int plane,
263 unsigned char *src_plane,
264 unsigned char *ref_plane,
265 unsigned char *decoded_plane,
266 int width, int height, int src_stride, int stride)
270 int block_width, block_height;
273 uint8_t *src = s->scratchbuf + stride * 16;
274 const int lambda = (s->picture.quality * s->picture.quality) >>
275 (2 * FF_LAMBDA_SHIFT);
277 /* figure out the acceptable level thresholds in advance */
278 threshold[5] = QUALITY_THRESHOLD;
279 for (level = 4; level >= 0; level--)
280 threshold[level] = threshold[level + 1] * THRESHOLD_MULTIPLIER;
282 block_width = (width + 15) / 16;
283 block_height = (height + 15) / 16;
285 if (s->picture.pict_type == AV_PICTURE_TYPE_P) {
286 s->m.avctx = s->avctx;
287 s->m.current_picture_ptr = &s->m.current_picture;
288 s->m.last_picture_ptr = &s->m.last_picture;
289 s->m.last_picture.f.data[0] = ref_plane;
291 s->m.last_picture.f.linesize[0] =
292 s->m.new_picture.f.linesize[0] =
293 s->m.current_picture.f.linesize[0] = stride;
295 s->m.height = height;
296 s->m.mb_width = block_width;
297 s->m.mb_height = block_height;
298 s->m.mb_stride = s->m.mb_width + 1;
299 s->m.b8_stride = 2 * s->m.mb_width + 1;
301 s->m.pict_type = s->picture.pict_type;
302 s->m.me_method = s->avctx->me_method;
303 s->m.me.scene_change_score = 0;
304 s->m.flags = s->avctx->flags;
305 // s->m.out_format = FMT_H263;
306 // s->m.unrestricted_mv = 1;
307 s->m.lambda = s->picture.quality;
308 s->m.qscale = s->m.lambda * 139 +
309 FF_LAMBDA_SCALE * 64 >>
311 s->m.lambda2 = s->m.lambda * s->m.lambda +
312 FF_LAMBDA_SCALE / 2 >>
315 if (!s->motion_val8[plane]) {
316 s->motion_val8[plane] = av_mallocz((s->m.b8_stride *
317 block_height * 2 + 2) *
318 2 * sizeof(int16_t));
319 s->motion_val16[plane] = av_mallocz((s->m.mb_stride *
320 (block_height + 2) + 1) *
321 2 * sizeof(int16_t));
324 s->m.mb_type = s->mb_type;
326 // dummies, to avoid segfaults
327 s->m.current_picture.mb_mean = (uint8_t *)s->dummy;
328 s->m.current_picture.mb_var = (uint16_t *)s->dummy;
329 s->m.current_picture.mc_mb_var = (uint16_t *)s->dummy;
330 s->m.current_picture.mb_type = s->dummy;
332 s->m.current_picture.motion_val[0] = s->motion_val8[plane] + 2;
333 s->m.p_mv_table = s->motion_val16[plane] +
335 s->m.dsp = s->dsp; // move
338 s->m.me.dia_size = s->avctx->dia_size;
339 s->m.first_slice_line = 1;
340 for (y = 0; y < block_height; y++) {
341 s->m.new_picture.f.data[0] = src - y * 16 * stride; // ugly
344 for (i = 0; i < 16 && i + 16 * y < height; i++) {
345 memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
347 for (x = width; x < 16 * block_width; x++)
348 src[i * stride + x] = src[i * stride + x - 1];
350 for (; i < 16 && i + 16 * y < 16 * block_height; i++)
351 memcpy(&src[i * stride], &src[(i - 1) * stride],
354 for (x = 0; x < block_width; x++) {
356 ff_init_block_index(&s->m);
357 ff_update_block_index(&s->m);
359 ff_estimate_p_frame_motion(&s->m, x, y);
361 s->m.first_slice_line = 0;
364 ff_fix_long_p_mvs(&s->m);
365 ff_fix_long_mvs(&s->m, NULL, 0, s->m.p_mv_table, s->m.f_code,
366 CANDIDATE_MB_TYPE_INTER, 0);
369 s->m.first_slice_line = 1;
370 for (y = 0; y < block_height; y++) {
371 for (i = 0; i < 16 && i + 16 * y < height; i++) {
372 memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
374 for (x = width; x < 16 * block_width; x++)
375 src[i * stride + x] = src[i * stride + x - 1];
377 for (; i < 16 && i + 16 * y < 16 * block_height; i++)
378 memcpy(&src[i * stride], &src[(i - 1) * stride], 16 * block_width);
381 for (x = 0; x < block_width; x++) {
382 uint8_t reorder_buffer[3][6][7 * 32];
384 int offset = y * 16 * stride + x * 16;
385 uint8_t *decoded = decoded_plane + offset;
386 uint8_t *ref = ref_plane + offset;
387 int score[4] = { 0, 0, 0, 0 }, best;
388 uint8_t *temp = s->scratchbuf;
390 if (s->pb.buf_end - s->pb.buf -
391 (put_bits_count(&s->pb) >> 3) < 3000) { // FIXME: check size
392 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
397 ff_init_block_index(&s->m);
398 ff_update_block_index(&s->m);
400 if (s->picture.pict_type == AV_PICTURE_TYPE_I ||
401 (s->m.mb_type[x + y * s->m.mb_stride] &
402 CANDIDATE_MB_TYPE_INTRA)) {
403 for (i = 0; i < 6; i++)
404 init_put_bits(&s->reorder_pb[i], reorder_buffer[0][i],
406 if (s->picture.pict_type == AV_PICTURE_TYPE_P) {
407 const uint8_t *vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_INTRA];
408 put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
409 score[0] = vlc[1] * lambda;
411 score[0] += encode_block(s, src + 16 * x, NULL, temp, stride,
413 for (i = 0; i < 6; i++) {
414 count[0][i] = put_bits_count(&s->reorder_pb[i]);
415 flush_put_bits(&s->reorder_pb[i]);
422 if (s->picture.pict_type == AV_PICTURE_TYPE_P) {
423 const uint8_t *vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_INTER];
424 int mx, my, pred_x, pred_y, dxy;
427 motion_ptr = ff_h263_pred_motion(&s->m, 0, 0, &pred_x, &pred_y);
428 if (s->m.mb_type[x + y * s->m.mb_stride] &
429 CANDIDATE_MB_TYPE_INTER) {
430 for (i = 0; i < 6; i++)
431 init_put_bits(&s->reorder_pb[i], reorder_buffer[1][i],
434 put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
436 s->m.pb = s->reorder_pb[5];
439 assert(mx >= -32 && mx <= 31);
440 assert(my >= -32 && my <= 31);
441 assert(pred_x >= -32 && pred_x <= 31);
442 assert(pred_y >= -32 && pred_y <= 31);
443 ff_h263_encode_motion(&s->m, mx - pred_x, 1);
444 ff_h263_encode_motion(&s->m, my - pred_y, 1);
445 s->reorder_pb[5] = s->m.pb;
446 score[1] += lambda * put_bits_count(&s->reorder_pb[5]);
448 dxy = (mx & 1) + 2 * (my & 1);
450 s->hdsp.put_pixels_tab[0][dxy](temp + 16,
455 score[1] += encode_block(s, src + 16 * x, temp + 16,
456 decoded, stride, 5, 64, lambda, 0);
457 best = score[1] <= score[0];
459 vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_SKIP];
460 score[2] = s->dsp.sse[0](NULL, src + 16 * x, ref,
462 score[2] += vlc[1] * lambda;
463 if (score[2] < score[best] && mx == 0 && my == 0) {
465 s->hdsp.put_pixels_tab[0][0](decoded, ref, stride, 16);
466 for (i = 0; i < 6; i++)
468 put_bits(&s->pb, vlc[1], vlc[0]);
473 for (i = 0; i < 6; i++) {
474 count[1][i] = put_bits_count(&s->reorder_pb[i]);
475 flush_put_bits(&s->reorder_pb[i]);
482 motion_ptr[0 + 2 * s->m.b8_stride] =
483 motion_ptr[1 + 2 * s->m.b8_stride] =
484 motion_ptr[2 + 2 * s->m.b8_stride] =
485 motion_ptr[3 + 2 * s->m.b8_stride] = 0;
489 s->rd_total += score[best];
491 for (i = 5; i >= 0; i--)
492 avpriv_copy_bits(&s->pb, reorder_buffer[best][i],
495 s->hdsp.put_pixels_tab[0][0](decoded, temp, stride, 16);
497 s->m.first_slice_line = 0;
502 static av_cold int svq1_encode_init(AVCodecContext *avctx)
504 SVQ1Context *const s = avctx->priv_data;
506 ff_dsputil_init(&s->dsp, avctx);
507 ff_hpeldsp_init(&s->hdsp, avctx->flags);
508 avctx->coded_frame = &s->picture;
510 s->frame_width = avctx->width;
511 s->frame_height = avctx->height;
513 s->y_block_width = (s->frame_width + 15) / 16;
514 s->y_block_height = (s->frame_height + 15) / 16;
516 s->c_block_width = (s->frame_width / 4 + 15) / 16;
517 s->c_block_height = (s->frame_height / 4 + 15) / 16;
521 s->m.picture_structure = PICT_FRAME;
523 s->m.me.scratchpad = av_mallocz((avctx->width + 64) *
524 2 * 16 * 2 * sizeof(uint8_t));
525 s->m.me.map = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
526 s->m.me.score_map = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
527 s->mb_type = av_mallocz((s->y_block_width + 1) *
528 s->y_block_height * sizeof(int16_t));
529 s->dummy = av_mallocz((s->y_block_width + 1) *
530 s->y_block_height * sizeof(int32_t));
531 ff_h263_encode_init(&s->m); // mv_penalty
536 static int svq1_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
537 const AVFrame *pict, int *got_packet)
539 SVQ1Context *const s = avctx->priv_data;
540 AVFrame *const p = &s->picture;
545 (ret = av_new_packet(pkt, s->y_block_width * s->y_block_height *
546 MAX_MB_BYTES * 3 + FF_MIN_BUFFER_SIZE)) < 0) {
547 av_log(avctx, AV_LOG_ERROR, "Error getting output packet.\n");
551 if (avctx->pix_fmt != AV_PIX_FMT_YUV410P) {
552 av_log(avctx, AV_LOG_ERROR, "unsupported pixel format\n");
556 if (!s->current_picture.data[0]) {
557 ff_get_buffer(avctx, &s->current_picture, 0);
558 ff_get_buffer(avctx, &s->last_picture, 0);
559 s->scratchbuf = av_malloc(s->current_picture.linesize[0] * 16 * 2);
562 temp = s->current_picture;
563 s->current_picture = s->last_picture;
564 s->last_picture = temp;
566 init_put_bits(&s->pb, pkt->data, pkt->size);
569 p->pict_type = avctx->gop_size && avctx->frame_number % avctx->gop_size ?
570 AV_PICTURE_TYPE_P : AV_PICTURE_TYPE_I;
571 p->key_frame = p->pict_type == AV_PICTURE_TYPE_I;
573 svq1_write_header(s, p->pict_type);
574 for (i = 0; i < 3; i++)
575 if (svq1_encode_plane(s, i,
577 s->last_picture.data[i],
578 s->current_picture.data[i],
579 s->frame_width / (i ? 4 : 1),
580 s->frame_height / (i ? 4 : 1),
581 s->picture.linesize[i],
582 s->current_picture.linesize[i]) < 0)
585 // avpriv_align_put_bits(&s->pb);
586 while (put_bits_count(&s->pb) & 31)
587 put_bits(&s->pb, 1, 0);
589 flush_put_bits(&s->pb);
591 pkt->size = put_bits_count(&s->pb) / 8;
592 if (p->pict_type == AV_PICTURE_TYPE_I)
593 pkt->flags |= AV_PKT_FLAG_KEY;
599 static av_cold int svq1_encode_end(AVCodecContext *avctx)
601 SVQ1Context *const s = avctx->priv_data;
604 av_log(avctx, AV_LOG_DEBUG, "RD: %f\n",
605 s->rd_total / (double)(avctx->width * avctx->height *
606 avctx->frame_number));
608 av_freep(&s->m.me.scratchpad);
609 av_freep(&s->m.me.map);
610 av_freep(&s->m.me.score_map);
611 av_freep(&s->mb_type);
613 av_freep(&s->scratchbuf);
615 for (i = 0; i < 3; i++) {
616 av_freep(&s->motion_val8[i]);
617 av_freep(&s->motion_val16[i]);
620 av_frame_unref(&s->current_picture);
621 av_frame_unref(&s->last_picture);
626 AVCodec ff_svq1_encoder = {
628 .type = AVMEDIA_TYPE_VIDEO,
629 .id = AV_CODEC_ID_SVQ1,
630 .priv_data_size = sizeof(SVQ1Context),
631 .init = svq1_encode_init,
632 .encode2 = svq1_encode_frame,
633 .close = svq1_encode_end,
634 .pix_fmts = (const enum AVPixelFormat[]) { AV_PIX_FMT_YUV410P,
636 .long_name = NULL_IF_CONFIG_SMALL("Sorenson Vector Quantizer 1 / Sorenson Video 1 / SVQ1"),