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/
31 #include "mpegvideo.h"
35 #include "svq1enc_cb.h"
40 typedef struct SVQ1Context {
41 /* FIXME: Needed for motion estimation, should not be used for anything
42 * else, the idea is to make the motion estimation eventually independent
43 * of MpegEncContext, so this will be removed then. */
45 AVCodecContext *avctx;
48 AVFrame current_picture;
53 /* why ooh why this sick breadth first order,
54 * everything is slower and more complex */
55 PutBitContext reorder_pb[6];
60 /* Y plane block dimensions */
64 /* U & V plane (C planes) block dimensions */
70 int16_t (*motion_val8[3])[2];
71 int16_t (*motion_val16[3])[2];
78 static void svq1_write_header(SVQ1Context *s, int frame_type)
83 put_bits(&s->pb, 22, 0x20);
85 /* temporal reference (sure hope this is a "don't care") */
86 put_bits(&s->pb, 8, 0x00);
89 put_bits(&s->pb, 2, frame_type - 1);
91 if (frame_type == AV_PICTURE_TYPE_I) {
92 /* no checksum since frame code is 0x20 */
93 /* no embedded string either */
94 /* output 5 unknown bits (2 + 2 + 1) */
95 put_bits(&s->pb, 5, 2); /* 2 needed by quicktime decoder */
97 i = ff_match_2uint16(ff_svq1_frame_size_table,
98 FF_ARRAY_ELEMS(ff_svq1_frame_size_table),
99 s->frame_width, s->frame_height);
100 put_bits(&s->pb, 3, i);
103 put_bits(&s->pb, 12, s->frame_width);
104 put_bits(&s->pb, 12, s->frame_height);
108 /* no checksum or extra data (next 2 bits get 0) */
109 put_bits(&s->pb, 2, 0);
112 #define QUALITY_THRESHOLD 100
113 #define THRESHOLD_MULTIPLIER 0.6
115 static int encode_block(SVQ1Context *s, uint8_t *src, uint8_t *ref,
116 uint8_t *decoded, int stride, int level,
117 int threshold, int lambda, int intra)
119 int count, y, x, i, j, split, best_mean, best_score, best_count;
121 int block_sum[7] = { 0, 0, 0, 0, 0, 0 };
122 int w = 2 << (level + 2 >> 1);
123 int h = 2 << (level + 1 >> 1);
125 int16_t block[7][256];
126 const int8_t *codebook_sum, *codebook;
127 const uint16_t(*mean_vlc)[2];
128 const uint8_t(*multistage_vlc)[2];
131 // FIXME: Optimize, this does not need to be done multiple times.
133 codebook_sum = svq1_intra_codebook_sum[level];
134 codebook = ff_svq1_intra_codebooks[level];
135 mean_vlc = ff_svq1_intra_mean_vlc;
136 multistage_vlc = ff_svq1_intra_multistage_vlc[level];
137 for (y = 0; y < h; y++) {
138 for (x = 0; x < w; x++) {
139 int v = src[x + y * stride];
140 block[0][x + w * y] = v;
146 codebook_sum = svq1_inter_codebook_sum[level];
147 codebook = ff_svq1_inter_codebooks[level];
148 mean_vlc = ff_svq1_inter_mean_vlc + 256;
149 multistage_vlc = ff_svq1_inter_multistage_vlc[level];
150 for (y = 0; y < h; y++) {
151 for (x = 0; x < w; x++) {
152 int v = src[x + y * stride] - ref[x + y * stride];
153 block[0][x + w * y] = v;
161 best_score -= (int)((unsigned)block_sum[0] * block_sum[0] >> (level + 3));
162 best_mean = block_sum[0] + (size >> 1) >> (level + 3);
165 for (count = 1; count < 7; count++) {
166 int best_vector_score = INT_MAX;
167 int best_vector_sum = -999, best_vector_mean = -999;
168 const int stage = count - 1;
169 const int8_t *vector;
171 for (i = 0; i < 16; i++) {
172 int sum = codebook_sum[stage * 16 + i];
173 int sqr, diff, score;
175 vector = codebook + stage * size * 16 + i * size;
176 sqr = s->dsp.ssd_int8_vs_int16(vector, block[stage], size);
177 diff = block_sum[stage] - sum;
178 score = sqr - (diff * (int64_t)diff >> (level + 3)); // FIXME: 64bit slooow
179 if (score < best_vector_score) {
180 int mean = diff + (size >> 1) >> (level + 3);
181 assert(mean > -300 && mean < 300);
182 mean = av_clip(mean, intra ? 0 : -256, 255);
183 best_vector_score = score;
184 best_vector[stage] = i;
185 best_vector_sum = sum;
186 best_vector_mean = mean;
189 assert(best_vector_mean != -999);
190 vector = codebook + stage * size * 16 + best_vector[stage] * size;
191 for (j = 0; j < size; j++)
192 block[stage + 1][j] = block[stage][j] - vector[j];
193 block_sum[stage + 1] = block_sum[stage] - best_vector_sum;
194 best_vector_score += lambda *
196 multistage_vlc[1 + count][1]
197 + mean_vlc[best_vector_mean][1]);
199 if (best_vector_score < best_score) {
200 best_score = best_vector_score;
202 best_mean = best_vector_mean;
208 if (best_score > threshold && level) {
210 int offset = level & 1 ? stride * h / 2 : w / 2;
211 PutBitContext backup[6];
213 for (i = level - 1; i >= 0; i--)
214 backup[i] = s->reorder_pb[i];
215 score += encode_block(s, src, ref, decoded, stride, level - 1,
216 threshold >> 1, lambda, intra);
217 score += encode_block(s, src + offset, ref + offset, decoded + offset,
218 stride, level - 1, threshold >> 1, lambda, intra);
221 if (score < best_score) {
225 for (i = level - 1; i >= 0; i--)
226 s->reorder_pb[i] = backup[i];
230 put_bits(&s->reorder_pb[level], 1, split);
233 assert(best_mean >= 0 && best_mean < 256 || !intra);
234 assert(best_mean >= -256 && best_mean < 256);
235 assert(best_count >= 0 && best_count < 7);
236 assert(level < 4 || best_count == 0);
238 /* output the encoding */
239 put_bits(&s->reorder_pb[level],
240 multistage_vlc[1 + best_count][1],
241 multistage_vlc[1 + best_count][0]);
242 put_bits(&s->reorder_pb[level], mean_vlc[best_mean][1],
243 mean_vlc[best_mean][0]);
245 for (i = 0; i < best_count; i++) {
246 assert(best_vector[i] >= 0 && best_vector[i] < 16);
247 put_bits(&s->reorder_pb[level], 4, best_vector[i]);
250 for (y = 0; y < h; y++)
251 for (x = 0; x < w; x++)
252 decoded[x + y * stride] = src[x + y * stride] -
253 block[best_count][x + w * y] +
260 static int svq1_encode_plane(SVQ1Context *s, int plane,
261 unsigned char *src_plane,
262 unsigned char *ref_plane,
263 unsigned char *decoded_plane,
264 int width, int height, int src_stride, int stride)
268 int block_width, block_height;
271 uint8_t *src = s->scratchbuf + stride * 16;
272 const int lambda = (s->picture.quality * s->picture.quality) >>
273 (2 * FF_LAMBDA_SHIFT);
275 /* figure out the acceptable level thresholds in advance */
276 threshold[5] = QUALITY_THRESHOLD;
277 for (level = 4; level >= 0; level--)
278 threshold[level] = threshold[level + 1] * THRESHOLD_MULTIPLIER;
280 block_width = (width + 15) / 16;
281 block_height = (height + 15) / 16;
283 if (s->picture.pict_type == AV_PICTURE_TYPE_P) {
284 s->m.avctx = s->avctx;
285 s->m.current_picture_ptr = &s->m.current_picture;
286 s->m.last_picture_ptr = &s->m.last_picture;
287 s->m.last_picture.f.data[0] = ref_plane;
289 s->m.last_picture.f.linesize[0] =
290 s->m.new_picture.f.linesize[0] =
291 s->m.current_picture.f.linesize[0] = stride;
293 s->m.height = height;
294 s->m.mb_width = block_width;
295 s->m.mb_height = block_height;
296 s->m.mb_stride = s->m.mb_width + 1;
297 s->m.b8_stride = 2 * s->m.mb_width + 1;
299 s->m.pict_type = s->picture.pict_type;
300 s->m.me_method = s->avctx->me_method;
301 s->m.me.scene_change_score = 0;
302 s->m.flags = s->avctx->flags;
303 // s->m.out_format = FMT_H263;
304 // s->m.unrestricted_mv = 1;
305 s->m.lambda = s->picture.quality;
306 s->m.qscale = s->m.lambda * 139 +
307 FF_LAMBDA_SCALE * 64 >>
309 s->m.lambda2 = s->m.lambda * s->m.lambda +
310 FF_LAMBDA_SCALE / 2 >>
313 if (!s->motion_val8[plane]) {
314 s->motion_val8[plane] = av_mallocz((s->m.b8_stride *
315 block_height * 2 + 2) *
316 2 * sizeof(int16_t));
317 s->motion_val16[plane] = av_mallocz((s->m.mb_stride *
318 (block_height + 2) + 1) *
319 2 * sizeof(int16_t));
322 s->m.mb_type = s->mb_type;
324 // dummies, to avoid segfaults
325 s->m.current_picture.mb_mean = (uint8_t *)s->dummy;
326 s->m.current_picture.mb_var = (uint16_t *)s->dummy;
327 s->m.current_picture.mc_mb_var = (uint16_t *)s->dummy;
328 s->m.current_picture.f.mb_type = s->dummy;
330 s->m.current_picture.f.motion_val[0] = s->motion_val8[plane] + 2;
331 s->m.p_mv_table = s->motion_val16[plane] +
333 s->m.dsp = s->dsp; // move
336 s->m.me.dia_size = s->avctx->dia_size;
337 s->m.first_slice_line = 1;
338 for (y = 0; y < block_height; y++) {
339 s->m.new_picture.f.data[0] = src - y * 16 * stride; // ugly
342 for (i = 0; i < 16 && i + 16 * y < height; i++) {
343 memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
345 for (x = width; x < 16 * block_width; x++)
346 src[i * stride + x] = src[i * stride + x - 1];
348 for (; i < 16 && i + 16 * y < 16 * block_height; i++)
349 memcpy(&src[i * stride], &src[(i - 1) * stride],
352 for (x = 0; x < block_width; x++) {
354 ff_init_block_index(&s->m);
355 ff_update_block_index(&s->m);
357 ff_estimate_p_frame_motion(&s->m, x, y);
359 s->m.first_slice_line = 0;
362 ff_fix_long_p_mvs(&s->m);
363 ff_fix_long_mvs(&s->m, NULL, 0, s->m.p_mv_table, s->m.f_code,
364 CANDIDATE_MB_TYPE_INTER, 0);
367 s->m.first_slice_line = 1;
368 for (y = 0; y < block_height; y++) {
369 for (i = 0; i < 16 && i + 16 * y < height; i++) {
370 memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
372 for (x = width; x < 16 * block_width; x++)
373 src[i * stride + x] = src[i * stride + x - 1];
375 for (; i < 16 && i + 16 * y < 16 * block_height; i++)
376 memcpy(&src[i * stride], &src[(i - 1) * stride], 16 * block_width);
379 for (x = 0; x < block_width; x++) {
380 uint8_t reorder_buffer[3][6][7 * 32];
382 int offset = y * 16 * stride + x * 16;
383 uint8_t *decoded = decoded_plane + offset;
384 uint8_t *ref = ref_plane + offset;
385 int score[4] = { 0, 0, 0, 0 }, best;
386 uint8_t *temp = s->scratchbuf;
388 if (s->pb.buf_end - s->pb.buf -
389 (put_bits_count(&s->pb) >> 3) < 3000) { // FIXME: check size
390 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
395 ff_init_block_index(&s->m);
396 ff_update_block_index(&s->m);
398 if (s->picture.pict_type == AV_PICTURE_TYPE_I ||
399 (s->m.mb_type[x + y * s->m.mb_stride] &
400 CANDIDATE_MB_TYPE_INTRA)) {
401 for (i = 0; i < 6; i++)
402 init_put_bits(&s->reorder_pb[i], reorder_buffer[0][i],
404 if (s->picture.pict_type == AV_PICTURE_TYPE_P) {
405 const uint8_t *vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_INTRA];
406 put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
407 score[0] = vlc[1] * lambda;
409 score[0] += encode_block(s, src + 16 * x, NULL, temp, stride,
411 for (i = 0; i < 6; i++) {
412 count[0][i] = put_bits_count(&s->reorder_pb[i]);
413 flush_put_bits(&s->reorder_pb[i]);
420 if (s->picture.pict_type == AV_PICTURE_TYPE_P) {
421 const uint8_t *vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_INTER];
422 int mx, my, pred_x, pred_y, dxy;
425 motion_ptr = ff_h263_pred_motion(&s->m, 0, 0, &pred_x, &pred_y);
426 if (s->m.mb_type[x + y * s->m.mb_stride] &
427 CANDIDATE_MB_TYPE_INTER) {
428 for (i = 0; i < 6; i++)
429 init_put_bits(&s->reorder_pb[i], reorder_buffer[1][i],
432 put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
434 s->m.pb = s->reorder_pb[5];
437 assert(mx >= -32 && mx <= 31);
438 assert(my >= -32 && my <= 31);
439 assert(pred_x >= -32 && pred_x <= 31);
440 assert(pred_y >= -32 && pred_y <= 31);
441 ff_h263_encode_motion(&s->m, mx - pred_x, 1);
442 ff_h263_encode_motion(&s->m, my - pred_y, 1);
443 s->reorder_pb[5] = s->m.pb;
444 score[1] += lambda * put_bits_count(&s->reorder_pb[5]);
446 dxy = (mx & 1) + 2 * (my & 1);
448 s->dsp.put_pixels_tab[0][dxy](temp + 16,
453 score[1] += encode_block(s, src + 16 * x, temp + 16,
454 decoded, stride, 5, 64, lambda, 0);
455 best = score[1] <= score[0];
457 vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_SKIP];
458 score[2] = s->dsp.sse[0](NULL, src + 16 * x, ref,
460 score[2] += vlc[1] * lambda;
461 if (score[2] < score[best] && mx == 0 && my == 0) {
463 s->dsp.put_pixels_tab[0][0](decoded, ref, stride, 16);
464 for (i = 0; i < 6; i++)
466 put_bits(&s->pb, vlc[1], vlc[0]);
471 for (i = 0; i < 6; i++) {
472 count[1][i] = put_bits_count(&s->reorder_pb[i]);
473 flush_put_bits(&s->reorder_pb[i]);
480 motion_ptr[0 + 2 * s->m.b8_stride] =
481 motion_ptr[1 + 2 * s->m.b8_stride] =
482 motion_ptr[2 + 2 * s->m.b8_stride] =
483 motion_ptr[3 + 2 * s->m.b8_stride] = 0;
487 s->rd_total += score[best];
489 for (i = 5; i >= 0; i--)
490 avpriv_copy_bits(&s->pb, reorder_buffer[best][i],
493 s->dsp.put_pixels_tab[0][0](decoded, temp, stride, 16);
495 s->m.first_slice_line = 0;
500 static av_cold int svq1_encode_init(AVCodecContext *avctx)
502 SVQ1Context *const s = avctx->priv_data;
504 ff_dsputil_init(&s->dsp, avctx);
505 avctx->coded_frame = &s->picture;
507 s->frame_width = avctx->width;
508 s->frame_height = avctx->height;
510 s->y_block_width = (s->frame_width + 15) / 16;
511 s->y_block_height = (s->frame_height + 15) / 16;
513 s->c_block_width = (s->frame_width / 4 + 15) / 16;
514 s->c_block_height = (s->frame_height / 4 + 15) / 16;
518 s->m.picture_structure = PICT_FRAME;
520 s->m.me.scratchpad = av_mallocz((avctx->width + 64) *
521 2 * 16 * 2 * sizeof(uint8_t));
522 s->m.me.map = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
523 s->m.me.score_map = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
524 s->mb_type = av_mallocz((s->y_block_width + 1) *
525 s->y_block_height * sizeof(int16_t));
526 s->dummy = av_mallocz((s->y_block_width + 1) *
527 s->y_block_height * sizeof(int32_t));
528 ff_h263_encode_init(&s->m); // mv_penalty
533 static int svq1_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
534 const AVFrame *pict, int *got_packet)
536 SVQ1Context *const s = avctx->priv_data;
537 AVFrame *const p = &s->picture;
542 (ret = av_new_packet(pkt, s->y_block_width * s->y_block_height *
543 MAX_MB_BYTES * 3 + FF_MIN_BUFFER_SIZE) < 0)) {
544 av_log(avctx, AV_LOG_ERROR, "Error getting output packet.\n");
548 if (avctx->pix_fmt != AV_PIX_FMT_YUV410P) {
549 av_log(avctx, AV_LOG_ERROR, "unsupported pixel format\n");
553 if (!s->current_picture.data[0]) {
554 avctx->get_buffer(avctx, &s->current_picture);
555 avctx->get_buffer(avctx, &s->last_picture);
556 s->scratchbuf = av_malloc(s->current_picture.linesize[0] * 16 * 2);
559 temp = s->current_picture;
560 s->current_picture = s->last_picture;
561 s->last_picture = temp;
563 init_put_bits(&s->pb, pkt->data, pkt->size);
566 p->pict_type = avctx->gop_size && avctx->frame_number % avctx->gop_size ?
567 AV_PICTURE_TYPE_P : AV_PICTURE_TYPE_I;
568 p->key_frame = p->pict_type == AV_PICTURE_TYPE_I;
570 svq1_write_header(s, p->pict_type);
571 for (i = 0; i < 3; i++)
572 if (svq1_encode_plane(s, i,
574 s->last_picture.data[i],
575 s->current_picture.data[i],
576 s->frame_width / (i ? 4 : 1),
577 s->frame_height / (i ? 4 : 1),
578 s->picture.linesize[i],
579 s->current_picture.linesize[i]) < 0)
582 // avpriv_align_put_bits(&s->pb);
583 while (put_bits_count(&s->pb) & 31)
584 put_bits(&s->pb, 1, 0);
586 flush_put_bits(&s->pb);
588 pkt->size = put_bits_count(&s->pb) / 8;
589 if (p->pict_type == AV_PICTURE_TYPE_I)
590 pkt->flags |= AV_PKT_FLAG_KEY;
596 static av_cold int svq1_encode_end(AVCodecContext *avctx)
598 SVQ1Context *const s = avctx->priv_data;
601 av_log(avctx, AV_LOG_DEBUG, "RD: %f\n",
602 s->rd_total / (double)(avctx->width * avctx->height *
603 avctx->frame_number));
605 av_freep(&s->m.me.scratchpad);
606 av_freep(&s->m.me.map);
607 av_freep(&s->m.me.score_map);
608 av_freep(&s->mb_type);
610 av_freep(&s->scratchbuf);
612 for (i = 0; i < 3; i++) {
613 av_freep(&s->motion_val8[i]);
614 av_freep(&s->motion_val16[i]);
620 AVCodec ff_svq1_encoder = {
622 .type = AVMEDIA_TYPE_VIDEO,
623 .id = AV_CODEC_ID_SVQ1,
624 .priv_data_size = sizeof(SVQ1Context),
625 .init = svq1_encode_init,
626 .encode2 = svq1_encode_frame,
627 .close = svq1_encode_end,
628 .pix_fmts = (const enum PixelFormat[]) { AV_PIX_FMT_YUV410P,
630 .long_name = NULL_IF_CONFIG_SMALL("Sorenson Vector Quantizer 1 / Sorenson Video 1 / SVQ1"),