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/
30 #include "mpegvideo.h"
33 #include "libavutil/avassert.h"
35 #include "svq1enc_cb.h"
38 typedef struct SVQ1Context {
39 /* FIXME: Needed for motion estimation, should not be used for anything
40 * else, the idea is to make the motion estimation eventually independent
41 * of MpegEncContext, so this will be removed then. */
43 AVCodecContext *avctx;
46 AVFrame current_picture;
51 /* why ooh why this sick breadth first order,
52 * everything is slower and more complex */
53 PutBitContext reorder_pb[6];
58 /* Y plane block dimensions */
62 /* U & V plane (C planes) block dimensions */
68 int16_t (*motion_val8[3])[2];
69 int16_t (*motion_val16[3])[2];
76 static void svq1_write_header(SVQ1Context *s, int frame_type)
81 put_bits(&s->pb, 22, 0x20);
83 /* temporal reference (sure hope this is a "don't care") */
84 put_bits(&s->pb, 8, 0x00);
87 put_bits(&s->pb, 2, frame_type - 1);
89 if (frame_type == AV_PICTURE_TYPE_I) {
90 /* no checksum since frame code is 0x20 */
91 /* no embedded string either */
92 /* output 5 unknown bits (2 + 2 + 1) */
93 put_bits(&s->pb, 5, 2); /* 2 needed by quicktime decoder */
95 i = ff_match_2uint16((void*)ff_svq1_frame_size_table,
96 FF_ARRAY_ELEMS(ff_svq1_frame_size_table),
97 s->frame_width, s->frame_height);
98 put_bits(&s->pb, 3, i);
101 put_bits(&s->pb, 12, s->frame_width);
102 put_bits(&s->pb, 12, s->frame_height);
106 /* no checksum or extra data (next 2 bits get 0) */
107 put_bits(&s->pb, 2, 0);
110 #define QUALITY_THRESHOLD 100
111 #define THRESHOLD_MULTIPLIER 0.6
113 static int encode_block(SVQ1Context *s, uint8_t *src, uint8_t *ref,
114 uint8_t *decoded, int stride, int level,
115 int threshold, int lambda, int intra)
117 int count, y, x, i, j, split, best_mean, best_score, best_count;
119 int block_sum[7] = { 0, 0, 0, 0, 0, 0 };
120 int w = 2 << (level + 2 >> 1);
121 int h = 2 << (level + 1 >> 1);
123 int16_t block[7][256];
124 const int8_t *codebook_sum, *codebook;
125 const uint16_t(*mean_vlc)[2];
126 const uint8_t(*multistage_vlc)[2];
129 // FIXME: Optimize, this does not need to be done multiple times.
131 codebook_sum = svq1_intra_codebook_sum[level];
132 codebook = ff_svq1_intra_codebooks[level];
133 mean_vlc = ff_svq1_intra_mean_vlc;
134 multistage_vlc = ff_svq1_intra_multistage_vlc[level];
135 for (y = 0; y < h; y++) {
136 for (x = 0; x < w; x++) {
137 int v = src[x + y * stride];
138 block[0][x + w * y] = v;
144 codebook_sum = svq1_inter_codebook_sum[level];
145 codebook = ff_svq1_inter_codebooks[level];
146 mean_vlc = ff_svq1_inter_mean_vlc + 256;
147 multistage_vlc = ff_svq1_inter_multistage_vlc[level];
148 for (y = 0; y < h; y++) {
149 for (x = 0; x < w; x++) {
150 int v = src[x + y * stride] - ref[x + y * stride];
151 block[0][x + w * y] = v;
159 best_score -= (int)((unsigned)block_sum[0] * block_sum[0] >> (level + 3));
160 best_mean = block_sum[0] + (size >> 1) >> (level + 3);
163 for (count = 1; count < 7; count++) {
164 int best_vector_score = INT_MAX;
165 int best_vector_sum = -999, best_vector_mean = -999;
166 const int stage = count - 1;
167 const int8_t *vector;
169 for (i = 0; i < 16; i++) {
170 int sum = codebook_sum[stage * 16 + i];
171 int sqr, diff, score;
173 vector = codebook + stage * size * 16 + i * size;
174 sqr = s->dsp.ssd_int8_vs_int16(vector, block[stage], size);
175 diff = block_sum[stage] - sum;
176 score = sqr - (diff * (int64_t)diff >> (level + 3)); // FIXME: 64bit slooow
177 if (score < best_vector_score) {
178 int mean = diff + (size >> 1) >> (level + 3);
179 av_assert2(mean > -300 && mean < 300);
180 mean = av_clip(mean, intra ? 0 : -256, 255);
181 best_vector_score = score;
182 best_vector[stage] = i;
183 best_vector_sum = sum;
184 best_vector_mean = mean;
187 av_assert0(best_vector_mean != -999);
188 vector = codebook + stage * size * 16 + best_vector[stage] * size;
189 for (j = 0; j < size; j++)
190 block[stage + 1][j] = block[stage][j] - vector[j];
191 block_sum[stage + 1] = block_sum[stage] - best_vector_sum;
192 best_vector_score += lambda *
194 multistage_vlc[1 + count][1]
195 + mean_vlc[best_vector_mean][1]);
197 if (best_vector_score < best_score) {
198 best_score = best_vector_score;
200 best_mean = best_vector_mean;
206 if (best_score > threshold && level) {
208 int offset = level & 1 ? stride * h / 2 : w / 2;
209 PutBitContext backup[6];
211 for (i = level - 1; i >= 0; i--)
212 backup[i] = s->reorder_pb[i];
213 score += encode_block(s, src, ref, decoded, stride, level - 1,
214 threshold >> 1, lambda, intra);
215 score += encode_block(s, src + offset, ref + offset, decoded + offset,
216 stride, level - 1, threshold >> 1, lambda, intra);
219 if (score < best_score) {
223 for (i = level - 1; i >= 0; i--)
224 s->reorder_pb[i] = backup[i];
228 put_bits(&s->reorder_pb[level], 1, split);
231 av_assert1(best_mean >= 0 && best_mean < 256 || !intra);
232 av_assert1(best_mean >= -256 && best_mean < 256);
233 av_assert1(best_count >= 0 && best_count < 7);
234 av_assert1(level < 4 || best_count == 0);
236 /* output the encoding */
237 put_bits(&s->reorder_pb[level],
238 multistage_vlc[1 + best_count][1],
239 multistage_vlc[1 + best_count][0]);
240 put_bits(&s->reorder_pb[level], mean_vlc[best_mean][1],
241 mean_vlc[best_mean][0]);
243 for (i = 0; i < best_count; i++) {
244 av_assert2(best_vector[i] >= 0 && best_vector[i] < 16);
245 put_bits(&s->reorder_pb[level], 4, best_vector[i]);
248 for (y = 0; y < h; y++)
249 for (x = 0; x < w; x++)
250 decoded[x + y * stride] = src[x + y * stride] -
251 block[best_count][x + w * y] +
258 static int svq1_encode_plane(SVQ1Context *s, int plane,
259 unsigned char *src_plane,
260 unsigned char *ref_plane,
261 unsigned char *decoded_plane,
262 int width, int height, int src_stride, int stride)
266 int block_width, block_height;
269 uint8_t *src = s->scratchbuf + stride * 16;
270 const int lambda = (s->picture.quality * s->picture.quality) >>
271 (2 * FF_LAMBDA_SHIFT);
273 /* figure out the acceptable level thresholds in advance */
274 threshold[5] = QUALITY_THRESHOLD;
275 for (level = 4; level >= 0; level--)
276 threshold[level] = threshold[level + 1] * THRESHOLD_MULTIPLIER;
278 block_width = (width + 15) / 16;
279 block_height = (height + 15) / 16;
281 if (s->picture.pict_type == AV_PICTURE_TYPE_P) {
282 s->m.avctx = s->avctx;
283 s->m.current_picture_ptr = &s->m.current_picture;
284 s->m.last_picture_ptr = &s->m.last_picture;
285 s->m.last_picture.f.data[0] = ref_plane;
287 s->m.last_picture.f.linesize[0] =
288 s->m.new_picture.f.linesize[0] =
289 s->m.current_picture.f.linesize[0] = stride;
291 s->m.height = height;
292 s->m.mb_width = block_width;
293 s->m.mb_height = block_height;
294 s->m.mb_stride = s->m.mb_width + 1;
295 s->m.b8_stride = 2 * s->m.mb_width + 1;
297 s->m.pict_type = s->picture.pict_type;
298 s->m.me_method = s->avctx->me_method;
299 s->m.me.scene_change_score = 0;
300 s->m.flags = s->avctx->flags;
301 // s->m.out_format = FMT_H263;
302 // s->m.unrestricted_mv = 1;
303 s->m.lambda = s->picture.quality;
304 s->m.qscale = s->m.lambda * 139 +
305 FF_LAMBDA_SCALE * 64 >>
307 s->m.lambda2 = s->m.lambda * s->m.lambda +
308 FF_LAMBDA_SCALE / 2 >>
311 if (!s->motion_val8[plane]) {
312 s->motion_val8[plane] = av_mallocz((s->m.b8_stride *
313 block_height * 2 + 2) *
314 2 * sizeof(int16_t));
315 s->motion_val16[plane] = av_mallocz((s->m.mb_stride *
316 (block_height + 2) + 1) *
317 2 * sizeof(int16_t));
320 s->m.mb_type = s->mb_type;
322 // dummies, to avoid segfaults
323 s->m.current_picture.mb_mean = (uint8_t *)s->dummy;
324 s->m.current_picture.mb_var = (uint16_t *)s->dummy;
325 s->m.current_picture.mc_mb_var = (uint16_t *)s->dummy;
326 s->m.current_picture.f.mb_type = s->dummy;
328 s->m.current_picture.f.motion_val[0] = s->motion_val8[plane] + 2;
329 s->m.p_mv_table = s->motion_val16[plane] +
331 s->m.dsp = s->dsp; // move
334 s->m.me.dia_size = s->avctx->dia_size;
335 s->m.first_slice_line = 1;
336 for (y = 0; y < block_height; y++) {
337 s->m.new_picture.f.data[0] = src - y * 16 * stride; // ugly
340 for (i = 0; i < 16 && i + 16 * y < height; i++) {
341 memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
343 for (x = width; x < 16 * block_width; x++)
344 src[i * stride + x] = src[i * stride + x - 1];
346 for (; i < 16 && i + 16 * y < 16 * block_height; i++)
347 memcpy(&src[i * stride], &src[(i - 1) * stride],
350 for (x = 0; x < block_width; x++) {
352 ff_init_block_index(&s->m);
353 ff_update_block_index(&s->m);
355 ff_estimate_p_frame_motion(&s->m, x, y);
357 s->m.first_slice_line = 0;
360 ff_fix_long_p_mvs(&s->m);
361 ff_fix_long_mvs(&s->m, NULL, 0, s->m.p_mv_table, s->m.f_code,
362 CANDIDATE_MB_TYPE_INTER, 0);
365 s->m.first_slice_line = 1;
366 for (y = 0; y < block_height; y++) {
367 for (i = 0; i < 16 && i + 16 * y < height; i++) {
368 memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
370 for (x = width; x < 16 * block_width; x++)
371 src[i * stride + x] = src[i * stride + x - 1];
373 for (; i < 16 && i + 16 * y < 16 * block_height; i++)
374 memcpy(&src[i * stride], &src[(i - 1) * stride], 16 * block_width);
377 for (x = 0; x < block_width; x++) {
378 uint8_t reorder_buffer[3][6][7 * 32];
380 int offset = y * 16 * stride + x * 16;
381 uint8_t *decoded = decoded_plane + offset;
382 uint8_t *ref = ref_plane + offset;
383 int score[4] = { 0, 0, 0, 0 }, best;
384 uint8_t *temp = s->scratchbuf;
386 if (s->pb.buf_end - s->pb.buf -
387 (put_bits_count(&s->pb) >> 3) < 3000) { // FIXME: check size
388 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
393 ff_init_block_index(&s->m);
394 ff_update_block_index(&s->m);
396 if (s->picture.pict_type == AV_PICTURE_TYPE_I ||
397 (s->m.mb_type[x + y * s->m.mb_stride] &
398 CANDIDATE_MB_TYPE_INTRA)) {
399 for (i = 0; i < 6; i++)
400 init_put_bits(&s->reorder_pb[i], reorder_buffer[0][i],
402 if (s->picture.pict_type == AV_PICTURE_TYPE_P) {
403 const uint8_t *vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_INTRA];
404 put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
405 score[0] = vlc[1] * lambda;
407 score[0] += encode_block(s, src + 16 * x, NULL, temp, stride,
409 for (i = 0; i < 6; i++) {
410 count[0][i] = put_bits_count(&s->reorder_pb[i]);
411 flush_put_bits(&s->reorder_pb[i]);
418 if (s->picture.pict_type == AV_PICTURE_TYPE_P) {
419 const uint8_t *vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_INTER];
420 int mx, my, pred_x, pred_y, dxy;
423 motion_ptr = ff_h263_pred_motion(&s->m, 0, 0, &pred_x, &pred_y);
424 if (s->m.mb_type[x + y * s->m.mb_stride] &
425 CANDIDATE_MB_TYPE_INTER) {
426 for (i = 0; i < 6; i++)
427 init_put_bits(&s->reorder_pb[i], reorder_buffer[1][i],
430 put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
432 s->m.pb = s->reorder_pb[5];
435 av_assert1(mx >= -32 && mx <= 31);
436 av_assert1(my >= -32 && my <= 31);
437 av_assert1(pred_x >= -32 && pred_x <= 31);
438 av_assert1(pred_y >= -32 && pred_y <= 31);
439 ff_h263_encode_motion(&s->m, mx - pred_x, 1);
440 ff_h263_encode_motion(&s->m, my - pred_y, 1);
441 s->reorder_pb[5] = s->m.pb;
442 score[1] += lambda * put_bits_count(&s->reorder_pb[5]);
444 dxy = (mx & 1) + 2 * (my & 1);
446 s->dsp.put_pixels_tab[0][dxy](temp + 16,
451 score[1] += encode_block(s, src + 16 * x, temp + 16,
452 decoded, stride, 5, 64, lambda, 0);
453 best = score[1] <= score[0];
455 vlc = ff_svq1_block_type_vlc[SVQ1_BLOCK_SKIP];
456 score[2] = s->dsp.sse[0](NULL, src + 16 * x, ref,
458 score[2] += vlc[1] * lambda;
459 if (score[2] < score[best] && mx == 0 && my == 0) {
461 s->dsp.put_pixels_tab[0][0](decoded, ref, stride, 16);
462 for (i = 0; i < 6; i++)
464 put_bits(&s->pb, vlc[1], vlc[0]);
469 for (i = 0; i < 6; i++) {
470 count[1][i] = put_bits_count(&s->reorder_pb[i]);
471 flush_put_bits(&s->reorder_pb[i]);
478 motion_ptr[0 + 2 * s->m.b8_stride] =
479 motion_ptr[1 + 2 * s->m.b8_stride] =
480 motion_ptr[2 + 2 * s->m.b8_stride] =
481 motion_ptr[3 + 2 * s->m.b8_stride] = 0;
485 s->rd_total += score[best];
487 for (i = 5; i >= 0; i--)
488 avpriv_copy_bits(&s->pb, reorder_buffer[best][i],
491 s->dsp.put_pixels_tab[0][0](decoded, temp, stride, 16);
493 s->m.first_slice_line = 0;
498 static av_cold int svq1_encode_init(AVCodecContext *avctx)
500 SVQ1Context *const s = avctx->priv_data;
502 ff_dsputil_init(&s->dsp, avctx);
503 avctx->coded_frame = &s->picture;
505 s->frame_width = avctx->width;
506 s->frame_height = avctx->height;
508 s->y_block_width = (s->frame_width + 15) / 16;
509 s->y_block_height = (s->frame_height + 15) / 16;
511 s->c_block_width = (s->frame_width / 4 + 15) / 16;
512 s->c_block_height = (s->frame_height / 4 + 15) / 16;
516 s->m.picture_structure = PICT_FRAME;
518 s->m.me.scratchpad = av_mallocz((avctx->width + 64) *
519 2 * 16 * 2 * sizeof(uint8_t));
520 s->m.me.map = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
521 s->m.me.score_map = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
522 s->mb_type = av_mallocz((s->y_block_width + 1) *
523 s->y_block_height * sizeof(int16_t));
524 s->dummy = av_mallocz((s->y_block_width + 1) *
525 s->y_block_height * sizeof(int32_t));
526 ff_h263_encode_init(&s->m); // mv_penalty
531 static int svq1_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
532 const AVFrame *pict, int *got_packet)
534 SVQ1Context *const s = avctx->priv_data;
535 AVFrame *const p = &s->picture;
539 if ((ret = ff_alloc_packet2(avctx, pkt, s->y_block_width * s->y_block_height *
540 MAX_MB_BYTES*3 + FF_MIN_BUFFER_SIZE)) < 0)
543 if (avctx->pix_fmt != AV_PIX_FMT_YUV410P) {
544 av_log(avctx, AV_LOG_ERROR, "unsupported pixel format\n");
548 if (!s->current_picture.data[0]) {
549 if ((ret = ff_get_buffer(avctx, &s->current_picture))< 0 ||
550 (ret = ff_get_buffer(avctx, &s->last_picture)) < 0) {
553 s->scratchbuf = av_malloc(s->current_picture.linesize[0] * 16 * 2);
556 temp = s->current_picture;
557 s->current_picture = s->last_picture;
558 s->last_picture = temp;
560 init_put_bits(&s->pb, pkt->data, pkt->size);
563 p->pict_type = avctx->gop_size && avctx->frame_number % avctx->gop_size ?
564 AV_PICTURE_TYPE_P : AV_PICTURE_TYPE_I;
565 p->key_frame = p->pict_type == AV_PICTURE_TYPE_I;
567 svq1_write_header(s, p->pict_type);
568 for (i = 0; i < 3; i++)
569 if (svq1_encode_plane(s, i,
571 s->last_picture.data[i],
572 s->current_picture.data[i],
573 s->frame_width / (i ? 4 : 1),
574 s->frame_height / (i ? 4 : 1),
575 s->picture.linesize[i],
576 s->current_picture.linesize[i]) < 0)
579 // avpriv_align_put_bits(&s->pb);
580 while (put_bits_count(&s->pb) & 31)
581 put_bits(&s->pb, 1, 0);
583 flush_put_bits(&s->pb);
585 pkt->size = put_bits_count(&s->pb) / 8;
586 if (p->pict_type == AV_PICTURE_TYPE_I)
587 pkt->flags |= AV_PKT_FLAG_KEY;
593 static av_cold int svq1_encode_end(AVCodecContext *avctx)
595 SVQ1Context *const s = avctx->priv_data;
598 av_log(avctx, AV_LOG_DEBUG, "RD: %f\n",
599 s->rd_total / (double)(avctx->width * avctx->height *
600 avctx->frame_number));
602 av_freep(&s->m.me.scratchpad);
603 av_freep(&s->m.me.map);
604 av_freep(&s->m.me.score_map);
605 av_freep(&s->mb_type);
607 av_freep(&s->scratchbuf);
609 for (i = 0; i < 3; i++) {
610 av_freep(&s->motion_val8[i]);
611 av_freep(&s->motion_val16[i]);
613 if(s->current_picture.data[0])
614 avctx->release_buffer(avctx, &s->current_picture);
615 if(s->last_picture.data[0])
616 avctx->release_buffer(avctx, &s->last_picture);
621 AVCodec ff_svq1_encoder = {
623 .type = AVMEDIA_TYPE_VIDEO,
624 .id = AV_CODEC_ID_SVQ1,
625 .priv_data_size = sizeof(SVQ1Context),
626 .init = svq1_encode_init,
627 .encode2 = svq1_encode_frame,
628 .close = svq1_encode_end,
629 .pix_fmts = (const enum PixelFormat[]) { AV_PIX_FMT_YUV410P,
631 .long_name = NULL_IF_CONFIG_SMALL("Sorenson Vector Quantizer 1 / Sorenson Video 1 / SVQ1"),