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"
35 #include "mpegutils.h"
38 #include "svq1enc_cb.h"
43 static void svq1_write_header(SVQ1EncContext *s, int frame_type)
48 put_bits(&s->pb, 22, 0x20);
50 /* temporal reference (sure hope this is a "don't care") */
51 put_bits(&s->pb, 8, 0x00);
54 put_bits(&s->pb, 2, frame_type - 1);
56 if (frame_type == AV_PICTURE_TYPE_I) {
57 /* no checksum since frame code is 0x20 */
58 /* no embedded string either */
59 /* output 5 unknown bits (2 + 2 + 1) */
60 put_bits(&s->pb, 5, 2); /* 2 needed by quicktime decoder */
62 i = ff_match_2uint16(ff_svq1_frame_size_table,
63 FF_ARRAY_ELEMS(ff_svq1_frame_size_table),
64 s->frame_width, s->frame_height);
65 put_bits(&s->pb, 3, i);
68 put_bits(&s->pb, 12, s->frame_width);
69 put_bits(&s->pb, 12, s->frame_height);
73 /* no checksum or extra data (next 2 bits get 0) */
74 put_bits(&s->pb, 2, 0);
77 #define QUALITY_THRESHOLD 100
78 #define THRESHOLD_MULTIPLIER 0.6
80 static int ssd_int8_vs_int16_c(const int8_t *pix1, const int16_t *pix2,
85 for (i = 0; i < size; i++)
86 score += (pix1[i] - pix2[i]) * (pix1[i] - pix2[i]);
90 static int encode_block(SVQ1EncContext *s, uint8_t *src, uint8_t *ref,
91 uint8_t *decoded, int stride, int level,
92 int threshold, int lambda, int intra)
94 int count, y, x, i, j, split, best_mean, best_score, best_count;
96 int block_sum[7] = { 0, 0, 0, 0, 0, 0 };
97 int w = 2 << (level + 2 >> 1);
98 int h = 2 << (level + 1 >> 1);
100 int16_t block[7][256];
101 const int8_t *codebook_sum, *codebook;
102 const uint16_t(*mean_vlc)[2];
103 const uint8_t(*multistage_vlc)[2];
106 // FIXME: Optimize, this does not need to be done multiple times.
108 codebook_sum = svq1_intra_codebook_sum[level];
109 codebook = ff_svq1_intra_codebooks[level];
110 mean_vlc = ff_svq1_intra_mean_vlc;
111 multistage_vlc = ff_svq1_intra_multistage_vlc[level];
112 for (y = 0; y < h; y++) {
113 for (x = 0; x < w; x++) {
114 int v = src[x + y * stride];
115 block[0][x + w * y] = v;
121 codebook_sum = svq1_inter_codebook_sum[level];
122 codebook = ff_svq1_inter_codebooks[level];
123 mean_vlc = ff_svq1_inter_mean_vlc + 256;
124 multistage_vlc = ff_svq1_inter_multistage_vlc[level];
125 for (y = 0; y < h; y++) {
126 for (x = 0; x < w; x++) {
127 int v = src[x + y * stride] - ref[x + y * stride];
128 block[0][x + w * y] = v;
136 best_score -= (int)((unsigned)block_sum[0] * block_sum[0] >> (level + 3));
137 best_mean = block_sum[0] + (size >> 1) >> (level + 3);
140 for (count = 1; count < 7; count++) {
141 int best_vector_score = INT_MAX;
142 int best_vector_sum = -999, best_vector_mean = -999;
143 const int stage = count - 1;
144 const int8_t *vector;
146 for (i = 0; i < 16; i++) {
147 int sum = codebook_sum[stage * 16 + i];
148 int sqr, diff, score;
150 vector = codebook + stage * size * 16 + i * size;
151 sqr = s->ssd_int8_vs_int16(vector, block[stage], size);
152 diff = block_sum[stage] - sum;
153 score = sqr - (diff * (int64_t)diff >> (level + 3)); // FIXME: 64bit slooow
154 if (score < best_vector_score) {
155 int mean = diff + (size >> 1) >> (level + 3);
156 assert(mean > -300 && mean < 300);
157 mean = av_clip(mean, intra ? 0 : -256, 255);
158 best_vector_score = score;
159 best_vector[stage] = i;
160 best_vector_sum = sum;
161 best_vector_mean = mean;
164 assert(best_vector_mean != -999);
165 vector = codebook + stage * size * 16 + best_vector[stage] * size;
166 for (j = 0; j < size; j++)
167 block[stage + 1][j] = block[stage][j] - vector[j];
168 block_sum[stage + 1] = block_sum[stage] - best_vector_sum;
169 best_vector_score += lambda *
171 multistage_vlc[1 + count][1]
172 + mean_vlc[best_vector_mean][1]);
174 if (best_vector_score < best_score) {
175 best_score = best_vector_score;
177 best_mean = best_vector_mean;
183 if (best_score > threshold && level) {
185 int offset = level & 1 ? stride * h / 2 : w / 2;
186 PutBitContext backup[6];
188 for (i = level - 1; i >= 0; i--)
189 backup[i] = s->reorder_pb[i];
190 score += encode_block(s, src, ref, decoded, stride, level - 1,
191 threshold >> 1, lambda, intra);
192 score += encode_block(s, src + offset, ref + offset, decoded + offset,
193 stride, level - 1, threshold >> 1, lambda, intra);
196 if (score < best_score) {
200 for (i = level - 1; i >= 0; i--)
201 s->reorder_pb[i] = backup[i];
205 put_bits(&s->reorder_pb[level], 1, split);
208 assert(best_mean >= 0 && best_mean < 256 || !intra);
209 assert(best_mean >= -256 && best_mean < 256);
210 assert(best_count >= 0 && best_count < 7);
211 assert(level < 4 || best_count == 0);
213 /* output the encoding */
214 put_bits(&s->reorder_pb[level],
215 multistage_vlc[1 + best_count][1],
216 multistage_vlc[1 + best_count][0]);
217 put_bits(&s->reorder_pb[level], mean_vlc[best_mean][1],
218 mean_vlc[best_mean][0]);
220 for (i = 0; i < best_count; i++) {
221 assert(best_vector[i] >= 0 && best_vector[i] < 16);
222 put_bits(&s->reorder_pb[level], 4, best_vector[i]);
225 for (y = 0; y < h; y++)
226 for (x = 0; x < w; x++)
227 decoded[x + y * stride] = src[x + y * stride] -
228 block[best_count][x + w * y] +
235 static int svq1_encode_plane(SVQ1EncContext *s, int plane,
236 unsigned char *src_plane,
237 unsigned char *ref_plane,
238 unsigned char *decoded_plane,
239 int width, int height, int src_stride, int stride)
241 const AVFrame *f = s->avctx->coded_frame;
244 int block_width, block_height;
247 uint8_t *src = s->scratchbuf + stride * 16;
248 const int lambda = (f->quality * f->quality) >>
249 (2 * FF_LAMBDA_SHIFT);
251 /* figure out the acceptable level thresholds in advance */
252 threshold[5] = QUALITY_THRESHOLD;
253 for (level = 4; level >= 0; level--)
254 threshold[level] = threshold[level + 1] * THRESHOLD_MULTIPLIER;
256 block_width = (width + 15) / 16;
257 block_height = (height + 15) / 16;
259 if (f->pict_type == AV_PICTURE_TYPE_P) {
260 s->m.avctx = s->avctx;
261 s->m.current_picture_ptr = &s->m.current_picture;
262 s->m.last_picture_ptr = &s->m.last_picture;
263 s->m.last_picture.f->data[0] = ref_plane;
265 s->m.last_picture.f->linesize[0] =
266 s->m.new_picture.f->linesize[0] =
267 s->m.current_picture.f->linesize[0] = stride;
269 s->m.height = height;
270 s->m.mb_width = block_width;
271 s->m.mb_height = block_height;
272 s->m.mb_stride = s->m.mb_width + 1;
273 s->m.b8_stride = 2 * s->m.mb_width + 1;
275 s->m.pict_type = f->pict_type;
276 s->m.me_method = s->avctx->me_method;
277 s->m.me.scene_change_score = 0;
278 s->m.flags = s->avctx->flags;
279 // s->m.out_format = FMT_H263;
280 // s->m.unrestricted_mv = 1;
281 s->m.lambda = f->quality;
282 s->m.qscale = s->m.lambda * 139 +
283 FF_LAMBDA_SCALE * 64 >>
285 s->m.lambda2 = s->m.lambda * s->m.lambda +
286 FF_LAMBDA_SCALE / 2 >>
289 if (!s->motion_val8[plane]) {
290 s->motion_val8[plane] = av_mallocz((s->m.b8_stride *
291 block_height * 2 + 2) *
292 2 * sizeof(int16_t));
293 s->motion_val16[plane] = av_mallocz((s->m.mb_stride *
294 (block_height + 2) + 1) *
295 2 * sizeof(int16_t));
298 s->m.mb_type = s->mb_type;
300 // dummies, to avoid segfaults
301 s->m.current_picture.mb_mean = (uint8_t *)s->dummy;
302 s->m.current_picture.mb_var = (uint16_t *)s->dummy;
303 s->m.current_picture.mc_mb_var = (uint16_t *)s->dummy;
304 s->m.current_picture.mb_type = s->dummy;
306 s->m.current_picture.motion_val[0] = s->motion_val8[plane] + 2;
307 s->m.p_mv_table = s->motion_val16[plane] +
309 s->m.mecc = s->mecc; // move
312 s->m.me.dia_size = s->avctx->dia_size;
313 s->m.first_slice_line = 1;
314 for (y = 0; y < block_height; y++) {
315 s->m.new_picture.f->data[0] = src - y * 16 * stride; // ugly
318 for (i = 0; i < 16 && i + 16 * y < height; i++) {
319 memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
321 for (x = width; x < 16 * block_width; x++)
322 src[i * stride + x] = src[i * stride + x - 1];
324 for (; i < 16 && i + 16 * y < 16 * block_height; i++)
325 memcpy(&src[i * stride], &src[(i - 1) * stride],
328 for (x = 0; x < block_width; x++) {
330 ff_init_block_index(&s->m);
331 ff_update_block_index(&s->m);
333 ff_estimate_p_frame_motion(&s->m, x, y);
335 s->m.first_slice_line = 0;
338 ff_fix_long_p_mvs(&s->m);
339 ff_fix_long_mvs(&s->m, NULL, 0, s->m.p_mv_table, s->m.f_code,
340 CANDIDATE_MB_TYPE_INTER, 0);
343 s->m.first_slice_line = 1;
344 for (y = 0; y < block_height; y++) {
345 for (i = 0; i < 16 && i + 16 * y < height; i++) {
346 memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
348 for (x = width; x < 16 * block_width; x++)
349 src[i * stride + x] = src[i * stride + x - 1];
351 for (; i < 16 && i + 16 * y < 16 * block_height; i++)
352 memcpy(&src[i * stride], &src[(i - 1) * stride], 16 * block_width);
355 for (x = 0; x < block_width; x++) {
356 uint8_t reorder_buffer[3][6][7 * 32];
358 int offset = y * 16 * stride + x * 16;
359 uint8_t *decoded = decoded_plane + offset;
360 uint8_t *ref = ref_plane + offset;
361 int score[4] = { 0, 0, 0, 0 }, best;
362 uint8_t *temp = s->scratchbuf;
364 if (s->pb.buf_end - s->pb.buf -
365 (put_bits_count(&s->pb) >> 3) < 3000) { // FIXME: check size
366 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
371 ff_init_block_index(&s->m);
372 ff_update_block_index(&s->m);
374 if (f->pict_type == AV_PICTURE_TYPE_I ||
375 (s->m.mb_type[x + y * s->m.mb_stride] &
376 CANDIDATE_MB_TYPE_INTRA)) {
377 for (i = 0; i < 6; i++)
378 init_put_bits(&s->reorder_pb[i], reorder_buffer[0][i],
380 if (f->pict_type == AV_PICTURE_TYPE_P) {
381 const uint8_t *vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_INTRA];
382 put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
383 score[0] = vlc[1] * lambda;
385 score[0] += encode_block(s, src + 16 * x, NULL, temp, stride,
387 for (i = 0; i < 6; i++) {
388 count[0][i] = put_bits_count(&s->reorder_pb[i]);
389 flush_put_bits(&s->reorder_pb[i]);
396 if (f->pict_type == AV_PICTURE_TYPE_P) {
397 const uint8_t *vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_INTER];
398 int mx, my, pred_x, pred_y, dxy;
401 motion_ptr = ff_h263_pred_motion(&s->m, 0, 0, &pred_x, &pred_y);
402 if (s->m.mb_type[x + y * s->m.mb_stride] &
403 CANDIDATE_MB_TYPE_INTER) {
404 for (i = 0; i < 6; i++)
405 init_put_bits(&s->reorder_pb[i], reorder_buffer[1][i],
408 put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
410 s->m.pb = s->reorder_pb[5];
413 assert(mx >= -32 && mx <= 31);
414 assert(my >= -32 && my <= 31);
415 assert(pred_x >= -32 && pred_x <= 31);
416 assert(pred_y >= -32 && pred_y <= 31);
417 ff_h263_encode_motion(&s->m, mx - pred_x, 1);
418 ff_h263_encode_motion(&s->m, my - pred_y, 1);
419 s->reorder_pb[5] = s->m.pb;
420 score[1] += lambda * put_bits_count(&s->reorder_pb[5]);
422 dxy = (mx & 1) + 2 * (my & 1);
424 s->hdsp.put_pixels_tab[0][dxy](temp + 16,
429 score[1] += encode_block(s, src + 16 * x, temp + 16,
430 decoded, stride, 5, 64, lambda, 0);
431 best = score[1] <= score[0];
433 vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_SKIP];
434 score[2] = s->mecc.sse[0](NULL, src + 16 * x, ref,
436 score[2] += vlc[1] * lambda;
437 if (score[2] < score[best] && mx == 0 && my == 0) {
439 s->hdsp.put_pixels_tab[0][0](decoded, ref, stride, 16);
440 for (i = 0; i < 6; i++)
442 put_bits(&s->pb, vlc[1], vlc[0]);
447 for (i = 0; i < 6; i++) {
448 count[1][i] = put_bits_count(&s->reorder_pb[i]);
449 flush_put_bits(&s->reorder_pb[i]);
456 motion_ptr[0 + 2 * s->m.b8_stride] =
457 motion_ptr[1 + 2 * s->m.b8_stride] =
458 motion_ptr[2 + 2 * s->m.b8_stride] =
459 motion_ptr[3 + 2 * s->m.b8_stride] = 0;
463 s->rd_total += score[best];
465 for (i = 5; i >= 0; i--)
466 avpriv_copy_bits(&s->pb, reorder_buffer[best][i],
469 s->hdsp.put_pixels_tab[0][0](decoded, temp, stride, 16);
471 s->m.first_slice_line = 0;
476 static av_cold int svq1_encode_end(AVCodecContext *avctx)
478 SVQ1EncContext *const s = avctx->priv_data;
481 av_log(avctx, AV_LOG_DEBUG, "RD: %f\n",
482 s->rd_total / (double)(avctx->width * avctx->height *
483 avctx->frame_number));
486 ff_mpv_common_end(&s->m);
488 av_freep(&s->m.me.scratchpad);
489 av_freep(&s->m.me.map);
490 av_freep(&s->m.me.score_map);
491 av_freep(&s->mb_type);
493 av_freep(&s->scratchbuf);
495 for (i = 0; i < 3; i++) {
496 av_freep(&s->motion_val8[i]);
497 av_freep(&s->motion_val16[i]);
500 av_frame_free(&s->current_picture);
501 av_frame_free(&s->last_picture);
502 av_frame_free(&avctx->coded_frame);
507 static av_cold int svq1_encode_init(AVCodecContext *avctx)
509 SVQ1EncContext *const s = avctx->priv_data;
512 ff_hpeldsp_init(&s->hdsp, avctx->flags);
513 ff_me_cmp_init(&s->mecc, avctx);
514 ff_mpegvideoencdsp_init(&s->m.mpvencdsp, avctx);
516 avctx->coded_frame = av_frame_alloc();
517 s->current_picture = av_frame_alloc();
518 s->last_picture = av_frame_alloc();
519 if (!avctx->coded_frame || !s->current_picture || !s->last_picture) {
520 svq1_encode_end(avctx);
521 return AVERROR(ENOMEM);
524 s->frame_width = avctx->width;
525 s->frame_height = avctx->height;
527 s->y_block_width = (s->frame_width + 15) / 16;
528 s->y_block_height = (s->frame_height + 15) / 16;
530 s->c_block_width = (s->frame_width / 4 + 15) / 16;
531 s->c_block_height = (s->frame_height / 4 + 15) / 16;
536 if ((ret = ff_mpv_common_init(&s->m)) < 0) {
537 svq1_encode_end(avctx);
541 s->m.picture_structure = PICT_FRAME;
543 s->m.me.scratchpad = av_mallocz((avctx->width + 64) *
544 2 * 16 * 2 * sizeof(uint8_t));
545 s->m.me.map = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
546 s->m.me.score_map = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
547 s->mb_type = av_mallocz((s->y_block_width + 1) *
548 s->y_block_height * sizeof(int16_t));
549 s->dummy = av_mallocz((s->y_block_width + 1) *
550 s->y_block_height * sizeof(int32_t));
551 s->ssd_int8_vs_int16 = ssd_int8_vs_int16_c;
554 ff_svq1enc_init_ppc(s);
556 ff_svq1enc_init_x86(s);
558 ff_h263_encode_init(&s->m); // mv_penalty
563 static int svq1_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
564 const AVFrame *pict, int *got_packet)
566 SVQ1EncContext *const s = avctx->priv_data;
567 AVFrame *const p = avctx->coded_frame;
571 (ret = av_new_packet(pkt, s->y_block_width * s->y_block_height *
572 MAX_MB_BYTES * 3 + FF_MIN_BUFFER_SIZE)) < 0) {
573 av_log(avctx, AV_LOG_ERROR, "Error getting output packet.\n");
577 if (avctx->pix_fmt != AV_PIX_FMT_YUV410P) {
578 av_log(avctx, AV_LOG_ERROR, "unsupported pixel format\n");
582 if (!s->current_picture->data[0]) {
583 ff_get_buffer(avctx, s->current_picture, 0);
584 ff_get_buffer(avctx, s->last_picture, 0);
585 s->scratchbuf = av_malloc(s->current_picture->linesize[0] * 16 * 2);
588 FFSWAP(AVFrame*, s->current_picture, s->last_picture);
590 init_put_bits(&s->pb, pkt->data, pkt->size);
592 p->pict_type = avctx->gop_size && avctx->frame_number % avctx->gop_size ?
593 AV_PICTURE_TYPE_P : AV_PICTURE_TYPE_I;
594 p->key_frame = p->pict_type == AV_PICTURE_TYPE_I;
595 p->quality = pict->quality;
597 svq1_write_header(s, p->pict_type);
598 for (i = 0; i < 3; i++)
599 if (svq1_encode_plane(s, i,
601 s->last_picture->data[i],
602 s->current_picture->data[i],
603 s->frame_width / (i ? 4 : 1),
604 s->frame_height / (i ? 4 : 1),
606 s->current_picture->linesize[i]) < 0)
609 // avpriv_align_put_bits(&s->pb);
610 while (put_bits_count(&s->pb) & 31)
611 put_bits(&s->pb, 1, 0);
613 flush_put_bits(&s->pb);
615 pkt->size = put_bits_count(&s->pb) / 8;
616 if (p->pict_type == AV_PICTURE_TYPE_I)
617 pkt->flags |= AV_PKT_FLAG_KEY;
623 AVCodec ff_svq1_encoder = {
625 .long_name = NULL_IF_CONFIG_SMALL("Sorenson Vector Quantizer 1 / Sorenson Video 1 / SVQ1"),
626 .type = AVMEDIA_TYPE_VIDEO,
627 .id = AV_CODEC_ID_SVQ1,
628 .priv_data_size = sizeof(SVQ1EncContext),
629 .init = svq1_encode_init,
630 .encode2 = svq1_encode_frame,
631 .close = svq1_encode_end,
632 .pix_fmts = (const enum AVPixelFormat[]) { AV_PIX_FMT_YUV410P,