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/
31 #include "mpegvideo.h"
34 #include "libavutil/avassert.h"
36 #include "svq1enc_cb.h"
39 typedef struct SVQ1Context {
40 /* FIXME: Needed for motion estimation, should not be used for anything
41 * else, the idea is to make the motion estimation eventually independent
42 * of MpegEncContext, so this will be removed then. */
44 AVCodecContext *avctx;
47 AVFrame current_picture;
52 /* why ooh why this sick breadth first order,
53 * everything is slower and more complex */
54 PutBitContext reorder_pb[6];
59 /* Y plane block dimensions */
63 /* U & V plane (C planes) block dimensions */
69 int16_t (*motion_val8[3])[2];
70 int16_t (*motion_val16[3])[2];
77 static void svq1_write_header(SVQ1Context *s, int frame_type)
82 put_bits(&s->pb, 22, 0x20);
84 /* temporal reference (sure hope this is a "don't care") */
85 put_bits(&s->pb, 8, 0x00);
88 put_bits(&s->pb, 2, frame_type - 1);
90 if (frame_type == AV_PICTURE_TYPE_I) {
91 /* no checksum since frame code is 0x20 */
92 /* no embedded string either */
93 /* output 5 unknown bits (2 + 2 + 1) */
94 put_bits(&s->pb, 5, 2); /* 2 needed by quicktime decoder */
96 i = ff_match_2uint16((void*)ff_svq1_frame_size_table,
97 FF_ARRAY_ELEMS(ff_svq1_frame_size_table),
98 s->frame_width, s->frame_height);
99 put_bits(&s->pb, 3, i);
102 put_bits(&s->pb, 12, s->frame_width);
103 put_bits(&s->pb, 12, s->frame_height);
107 /* no checksum or extra data (next 2 bits get 0) */
108 put_bits(&s->pb, 2, 0);
111 #define QUALITY_THRESHOLD 100
112 #define THRESHOLD_MULTIPLIER 0.6
114 static int encode_block(SVQ1Context *s, uint8_t *src, uint8_t *ref,
115 uint8_t *decoded, int stride, int level,
116 int threshold, int lambda, int intra)
118 int count, y, x, i, j, split, best_mean, best_score, best_count;
120 int block_sum[7] = { 0, 0, 0, 0, 0, 0 };
121 int w = 2 << (level + 2 >> 1);
122 int h = 2 << (level + 1 >> 1);
124 int16_t block[7][256];
125 const int8_t *codebook_sum, *codebook;
126 const uint16_t(*mean_vlc)[2];
127 const uint8_t(*multistage_vlc)[2];
130 // FIXME: Optimize, this does not need to be done multiple times.
132 codebook_sum = svq1_intra_codebook_sum[level];
133 codebook = ff_svq1_intra_codebooks[level];
134 mean_vlc = ff_svq1_intra_mean_vlc;
135 multistage_vlc = ff_svq1_intra_multistage_vlc[level];
136 for (y = 0; y < h; y++) {
137 for (x = 0; x < w; x++) {
138 int v = src[x + y * stride];
139 block[0][x + w * y] = v;
145 codebook_sum = svq1_inter_codebook_sum[level];
146 codebook = ff_svq1_inter_codebooks[level];
147 mean_vlc = ff_svq1_inter_mean_vlc + 256;
148 multistage_vlc = ff_svq1_inter_multistage_vlc[level];
149 for (y = 0; y < h; y++) {
150 for (x = 0; x < w; x++) {
151 int v = src[x + y * stride] - ref[x + y * stride];
152 block[0][x + w * y] = v;
160 best_score -= (int)((unsigned)block_sum[0] * block_sum[0] >> (level + 3));
161 best_mean = block_sum[0] + (size >> 1) >> (level + 3);
164 for (count = 1; count < 7; count++) {
165 int best_vector_score = INT_MAX;
166 int best_vector_sum = -999, best_vector_mean = -999;
167 const int stage = count - 1;
168 const int8_t *vector;
170 for (i = 0; i < 16; i++) {
171 int sum = codebook_sum[stage * 16 + i];
172 int sqr, diff, score;
174 vector = codebook + stage * size * 16 + i * size;
175 sqr = s->dsp.ssd_int8_vs_int16(vector, block[stage], size);
176 diff = block_sum[stage] - sum;
177 score = sqr - (diff * (int64_t)diff >> (level + 3)); // FIXME: 64bit slooow
178 if (score < best_vector_score) {
179 int mean = diff + (size >> 1) >> (level + 3);
180 av_assert2(mean > -300 && mean < 300);
181 mean = av_clip(mean, intra ? 0 : -256, 255);
182 best_vector_score = score;
183 best_vector[stage] = i;
184 best_vector_sum = sum;
185 best_vector_mean = mean;
188 av_assert0(best_vector_mean != -999);
189 vector = codebook + stage * size * 16 + best_vector[stage] * size;
190 for (j = 0; j < size; j++)
191 block[stage + 1][j] = block[stage][j] - vector[j];
192 block_sum[stage + 1] = block_sum[stage] - best_vector_sum;
193 best_vector_score += lambda *
195 multistage_vlc[1 + count][1]
196 + mean_vlc[best_vector_mean][1]);
198 if (best_vector_score < best_score) {
199 best_score = best_vector_score;
201 best_mean = best_vector_mean;
207 if (best_score > threshold && level) {
209 int offset = level & 1 ? stride * h / 2 : w / 2;
210 PutBitContext backup[6];
212 for (i = level - 1; i >= 0; i--)
213 backup[i] = s->reorder_pb[i];
214 score += encode_block(s, src, ref, decoded, stride, level - 1,
215 threshold >> 1, lambda, intra);
216 score += encode_block(s, src + offset, ref + offset, decoded + offset,
217 stride, level - 1, threshold >> 1, lambda, intra);
220 if (score < best_score) {
224 for (i = level - 1; i >= 0; i--)
225 s->reorder_pb[i] = backup[i];
229 put_bits(&s->reorder_pb[level], 1, split);
232 av_assert1(best_mean >= 0 && best_mean < 256 || !intra);
233 av_assert1(best_mean >= -256 && best_mean < 256);
234 av_assert1(best_count >= 0 && best_count < 7);
235 av_assert1(level < 4 || best_count == 0);
237 /* output the encoding */
238 put_bits(&s->reorder_pb[level],
239 multistage_vlc[1 + best_count][1],
240 multistage_vlc[1 + best_count][0]);
241 put_bits(&s->reorder_pb[level], mean_vlc[best_mean][1],
242 mean_vlc[best_mean][0]);
244 for (i = 0; i < best_count; i++) {
245 av_assert2(best_vector[i] >= 0 && best_vector[i] < 16);
246 put_bits(&s->reorder_pb[level], 4, best_vector[i]);
249 for (y = 0; y < h; y++)
250 for (x = 0; x < w; x++)
251 decoded[x + y * stride] = src[x + y * stride] -
252 block[best_count][x + w * y] +
259 static int svq1_encode_plane(SVQ1Context *s, int plane,
260 unsigned char *src_plane,
261 unsigned char *ref_plane,
262 unsigned char *decoded_plane,
263 int width, int height, int src_stride, int stride)
267 int block_width, block_height;
270 uint8_t *src = s->scratchbuf + stride * 16;
271 const int lambda = (s->picture.quality * s->picture.quality) >>
272 (2 * FF_LAMBDA_SHIFT);
274 /* figure out the acceptable level thresholds in advance */
275 threshold[5] = QUALITY_THRESHOLD;
276 for (level = 4; level >= 0; level--)
277 threshold[level] = threshold[level + 1] * THRESHOLD_MULTIPLIER;
279 block_width = (width + 15) / 16;
280 block_height = (height + 15) / 16;
282 if (s->picture.pict_type == AV_PICTURE_TYPE_P) {
283 s->m.avctx = s->avctx;
284 s->m.current_picture_ptr = &s->m.current_picture;
285 s->m.last_picture_ptr = &s->m.last_picture;
286 s->m.last_picture.f.data[0] = ref_plane;
288 s->m.last_picture.f.linesize[0] =
289 s->m.new_picture.f.linesize[0] =
290 s->m.current_picture.f.linesize[0] = stride;
292 s->m.height = height;
293 s->m.mb_width = block_width;
294 s->m.mb_height = block_height;
295 s->m.mb_stride = s->m.mb_width + 1;
296 s->m.b8_stride = 2 * s->m.mb_width + 1;
298 s->m.pict_type = s->picture.pict_type;
299 s->m.me_method = s->avctx->me_method;
300 s->m.me.scene_change_score = 0;
301 s->m.flags = s->avctx->flags;
302 // s->m.out_format = FMT_H263;
303 // s->m.unrestricted_mv = 1;
304 s->m.lambda = s->picture.quality;
305 s->m.qscale = s->m.lambda * 139 +
306 FF_LAMBDA_SCALE * 64 >>
308 s->m.lambda2 = s->m.lambda * s->m.lambda +
309 FF_LAMBDA_SCALE / 2 >>
312 if (!s->motion_val8[plane]) {
313 s->motion_val8[plane] = av_mallocz((s->m.b8_stride *
314 block_height * 2 + 2) *
315 2 * sizeof(int16_t));
316 s->motion_val16[plane] = av_mallocz((s->m.mb_stride *
317 (block_height + 2) + 1) *
318 2 * sizeof(int16_t));
321 s->m.mb_type = s->mb_type;
323 // dummies, to avoid segfaults
324 s->m.current_picture.mb_mean = (uint8_t *)s->dummy;
325 s->m.current_picture.mb_var = (uint16_t *)s->dummy;
326 s->m.current_picture.mc_mb_var = (uint16_t *)s->dummy;
327 s->m.current_picture.f.mb_type = s->dummy;
329 s->m.current_picture.f.motion_val[0] = s->motion_val8[plane] + 2;
330 s->m.p_mv_table = s->motion_val16[plane] +
332 s->m.dsp = s->dsp; // move
335 s->m.me.dia_size = s->avctx->dia_size;
336 s->m.first_slice_line = 1;
337 for (y = 0; y < block_height; y++) {
338 s->m.new_picture.f.data[0] = src - y * 16 * stride; // ugly
341 for (i = 0; i < 16 && i + 16 * y < height; i++) {
342 memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
344 for (x = width; x < 16 * block_width; x++)
345 src[i * stride + x] = src[i * stride + x - 1];
347 for (; i < 16 && i + 16 * y < 16 * block_height; i++)
348 memcpy(&src[i * stride], &src[(i - 1) * stride],
351 for (x = 0; x < block_width; x++) {
353 ff_init_block_index(&s->m);
354 ff_update_block_index(&s->m);
356 ff_estimate_p_frame_motion(&s->m, x, y);
358 s->m.first_slice_line = 0;
361 ff_fix_long_p_mvs(&s->m);
362 ff_fix_long_mvs(&s->m, NULL, 0, s->m.p_mv_table, s->m.f_code,
363 CANDIDATE_MB_TYPE_INTER, 0);
366 s->m.first_slice_line = 1;
367 for (y = 0; y < block_height; y++) {
368 for (i = 0; i < 16 && i + 16 * y < height; i++) {
369 memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
371 for (x = width; x < 16 * block_width; x++)
372 src[i * stride + x] = src[i * stride + x - 1];
374 for (; i < 16 && i + 16 * y < 16 * block_height; i++)
375 memcpy(&src[i * stride], &src[(i - 1) * stride], 16 * block_width);
378 for (x = 0; x < block_width; x++) {
379 uint8_t reorder_buffer[3][6][7 * 32];
381 int offset = y * 16 * stride + x * 16;
382 uint8_t *decoded = decoded_plane + offset;
383 uint8_t *ref = ref_plane + offset;
384 int score[4] = { 0, 0, 0, 0 }, best;
385 uint8_t *temp = s->scratchbuf;
387 if (s->pb.buf_end - s->pb.buf -
388 (put_bits_count(&s->pb) >> 3) < 3000) { // FIXME: check size
389 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
394 ff_init_block_index(&s->m);
395 ff_update_block_index(&s->m);
397 if (s->picture.pict_type == AV_PICTURE_TYPE_I ||
398 (s->m.mb_type[x + y * s->m.mb_stride] &
399 CANDIDATE_MB_TYPE_INTRA)) {
400 for (i = 0; i < 6; i++)
401 init_put_bits(&s->reorder_pb[i], reorder_buffer[0][i],
403 if (s->picture.pict_type == AV_PICTURE_TYPE_P) {
404 const uint8_t *vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_INTRA];
405 put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
406 score[0] = vlc[1] * lambda;
408 score[0] += encode_block(s, src + 16 * x, NULL, temp, stride,
410 for (i = 0; i < 6; i++) {
411 count[0][i] = put_bits_count(&s->reorder_pb[i]);
412 flush_put_bits(&s->reorder_pb[i]);
419 if (s->picture.pict_type == AV_PICTURE_TYPE_P) {
420 const uint8_t *vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_INTER];
421 int mx, my, pred_x, pred_y, dxy;
424 motion_ptr = ff_h263_pred_motion(&s->m, 0, 0, &pred_x, &pred_y);
425 if (s->m.mb_type[x + y * s->m.mb_stride] &
426 CANDIDATE_MB_TYPE_INTER) {
427 for (i = 0; i < 6; i++)
428 init_put_bits(&s->reorder_pb[i], reorder_buffer[1][i],
431 put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
433 s->m.pb = s->reorder_pb[5];
436 av_assert1(mx >= -32 && mx <= 31);
437 av_assert1(my >= -32 && my <= 31);
438 av_assert1(pred_x >= -32 && pred_x <= 31);
439 av_assert1(pred_y >= -32 && pred_y <= 31);
440 ff_h263_encode_motion(&s->m, mx - pred_x, 1);
441 ff_h263_encode_motion(&s->m, my - pred_y, 1);
442 s->reorder_pb[5] = s->m.pb;
443 score[1] += lambda * put_bits_count(&s->reorder_pb[5]);
445 dxy = (mx & 1) + 2 * (my & 1);
447 s->dsp.put_pixels_tab[0][dxy](temp + 16,
452 score[1] += encode_block(s, src + 16 * x, temp + 16,
453 decoded, stride, 5, 64, lambda, 0);
454 best = score[1] <= score[0];
456 vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_SKIP];
457 score[2] = s->dsp.sse[0](NULL, src + 16 * x, ref,
459 score[2] += vlc[1] * lambda;
460 if (score[2] < score[best] && mx == 0 && my == 0) {
462 s->dsp.put_pixels_tab[0][0](decoded, ref, stride, 16);
463 for (i = 0; i < 6; i++)
465 put_bits(&s->pb, vlc[1], vlc[0]);
470 for (i = 0; i < 6; i++) {
471 count[1][i] = put_bits_count(&s->reorder_pb[i]);
472 flush_put_bits(&s->reorder_pb[i]);
479 motion_ptr[0 + 2 * s->m.b8_stride] =
480 motion_ptr[1 + 2 * s->m.b8_stride] =
481 motion_ptr[2 + 2 * s->m.b8_stride] =
482 motion_ptr[3 + 2 * s->m.b8_stride] = 0;
486 s->rd_total += score[best];
488 for (i = 5; i >= 0; i--)
489 avpriv_copy_bits(&s->pb, reorder_buffer[best][i],
492 s->dsp.put_pixels_tab[0][0](decoded, temp, stride, 16);
494 s->m.first_slice_line = 0;
499 static av_cold int svq1_encode_init(AVCodecContext *avctx)
501 SVQ1Context *const s = avctx->priv_data;
503 ff_dsputil_init(&s->dsp, avctx);
504 avctx->coded_frame = &s->picture;
506 s->frame_width = avctx->width;
507 s->frame_height = avctx->height;
509 s->y_block_width = (s->frame_width + 15) / 16;
510 s->y_block_height = (s->frame_height + 15) / 16;
512 s->c_block_width = (s->frame_width / 4 + 15) / 16;
513 s->c_block_height = (s->frame_height / 4 + 15) / 16;
517 s->m.picture_structure = PICT_FRAME;
519 s->m.me.scratchpad = av_mallocz((avctx->width + 64) *
520 2 * 16 * 2 * sizeof(uint8_t));
521 s->m.me.map = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
522 s->m.me.score_map = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
523 s->mb_type = av_mallocz((s->y_block_width + 1) *
524 s->y_block_height * sizeof(int16_t));
525 s->dummy = av_mallocz((s->y_block_width + 1) *
526 s->y_block_height * sizeof(int32_t));
527 ff_h263_encode_init(&s->m); // mv_penalty
532 static int svq1_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
533 const AVFrame *pict, int *got_packet)
535 SVQ1Context *const s = avctx->priv_data;
536 AVFrame *const p = &s->picture;
540 if ((ret = ff_alloc_packet2(avctx, pkt, s->y_block_width * s->y_block_height *
541 MAX_MB_BYTES*3 + FF_MIN_BUFFER_SIZE) < 0))
544 if (avctx->pix_fmt != AV_PIX_FMT_YUV410P) {
545 av_log(avctx, AV_LOG_ERROR, "unsupported pixel format\n");
549 if (!s->current_picture.data[0]) {
550 avctx->get_buffer(avctx, &s->current_picture);
551 avctx->get_buffer(avctx, &s->last_picture);
552 s->scratchbuf = av_malloc(s->current_picture.linesize[0] * 16 * 2);
555 temp = s->current_picture;
556 s->current_picture = s->last_picture;
557 s->last_picture = temp;
559 init_put_bits(&s->pb, pkt->data, pkt->size);
562 p->pict_type = avctx->gop_size && avctx->frame_number % avctx->gop_size ?
563 AV_PICTURE_TYPE_P : AV_PICTURE_TYPE_I;
564 p->key_frame = p->pict_type == AV_PICTURE_TYPE_I;
566 svq1_write_header(s, p->pict_type);
567 for (i = 0; i < 3; i++)
568 if (svq1_encode_plane(s, i,
570 s->last_picture.data[i],
571 s->current_picture.data[i],
572 s->frame_width / (i ? 4 : 1),
573 s->frame_height / (i ? 4 : 1),
574 s->picture.linesize[i],
575 s->current_picture.linesize[i]) < 0)
578 // avpriv_align_put_bits(&s->pb);
579 while (put_bits_count(&s->pb) & 31)
580 put_bits(&s->pb, 1, 0);
582 flush_put_bits(&s->pb);
584 pkt->size = put_bits_count(&s->pb) / 8;
585 if (p->pict_type == AV_PICTURE_TYPE_I)
586 pkt->flags |= AV_PKT_FLAG_KEY;
592 static av_cold int svq1_encode_end(AVCodecContext *avctx)
594 SVQ1Context *const s = avctx->priv_data;
597 av_log(avctx, AV_LOG_DEBUG, "RD: %f\n",
598 s->rd_total / (double)(avctx->width * avctx->height *
599 avctx->frame_number));
601 av_freep(&s->m.me.scratchpad);
602 av_freep(&s->m.me.map);
603 av_freep(&s->m.me.score_map);
604 av_freep(&s->mb_type);
606 av_freep(&s->scratchbuf);
608 for (i = 0; i < 3; i++) {
609 av_freep(&s->motion_val8[i]);
610 av_freep(&s->motion_val16[i]);
612 if(s->current_picture.data[0])
613 avctx->release_buffer(avctx, &s->current_picture);
614 if(s->last_picture.data[0])
615 avctx->release_buffer(avctx, &s->last_picture);
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"),