ratecontrol.c 32.6 KB
Newer Older
1
/*
2 3
 * Rate control for video encoders
 *
4
 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
5
 *
6
 * This file is part of Libav.
7
 *
8
 * Libav is free software; you can redistribute it and/or
9 10
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
11
 * version 2.1 of the License, or (at your option) any later version.
12
 *
13
 * Libav is distributed in the hope that it will be useful,
14 15 16 17 18
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
19
 * License along with Libav; if not, write to the Free Software
20
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21
 */
Michael Niedermayer's avatar
Michael Niedermayer committed
22 23

/**
24
 * @file
Michael Niedermayer's avatar
Michael Niedermayer committed
25
 * Rate control for video encoders.
26
 */
Michael Niedermayer's avatar
Michael Niedermayer committed
27

28
#include "libavutil/attributes.h"
29 30
#include "libavutil/internal.h"

31
#include "avcodec.h"
32
#include "internal.h"
33
#include "ratecontrol.h"
34
#include "mpegutils.h"
35
#include "mpegvideo.h"
36
#include "libavutil/eval.h"
37

Diego Biurrun's avatar
Diego Biurrun committed
38
#undef NDEBUG // Always check asserts, the speed effect is far too small to disable them.
39
#include <assert.h>
40

41 42 43 44
#ifndef M_E
#define M_E 2.718281828
#endif

45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218
static inline double qp2bits(RateControlEntry *rce, double qp)
{
    if (qp <= 0.0) {
        av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
    }
    return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / qp;
}

static inline double bits2qp(RateControlEntry *rce, double bits)
{
    if (bits < 0.9) {
        av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
    }
    return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / bits;
}

static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q)
{
    RateControlContext *rcc   = &s->rc_context;
    AVCodecContext *a         = s->avctx;
    const int pict_type       = rce->new_pict_type;
    const double last_p_q     = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
    const double last_non_b_q = rcc->last_qscale_for[rcc->last_non_b_pict_type];

    if (pict_type == AV_PICTURE_TYPE_I &&
        (a->i_quant_factor > 0.0 || rcc->last_non_b_pict_type == AV_PICTURE_TYPE_P))
        q = last_p_q * FFABS(a->i_quant_factor) + a->i_quant_offset;
    else if (pict_type == AV_PICTURE_TYPE_B &&
             a->b_quant_factor > 0.0)
        q = last_non_b_q * a->b_quant_factor + a->b_quant_offset;
    if (q < 1)
        q = 1;

    /* last qscale / qdiff stuff */
    if (rcc->last_non_b_pict_type == pict_type || pict_type != AV_PICTURE_TYPE_I) {
        double last_q     = rcc->last_qscale_for[pict_type];
        const int maxdiff = FF_QP2LAMBDA * a->max_qdiff;

        if (q > last_q + maxdiff)
            q = last_q + maxdiff;
        else if (q < last_q - maxdiff)
            q = last_q - maxdiff;
    }

    rcc->last_qscale_for[pict_type] = q; // Note we cannot do that after blurring

    if (pict_type != AV_PICTURE_TYPE_B)
        rcc->last_non_b_pict_type = pict_type;

    return q;
}

/**
 * Get the qmin & qmax for pict_type.
 */
static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type)
{
    int qmin = s->lmin;
    int qmax = s->lmax;

    assert(qmin <= qmax);

    switch (pict_type) {
    case AV_PICTURE_TYPE_B:
        qmin = (int)(qmin * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
        qmax = (int)(qmax * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
        break;
    case AV_PICTURE_TYPE_I:
        qmin = (int)(qmin * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
        qmax = (int)(qmax * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
        break;
    }

    qmin = av_clip(qmin, 1, FF_LAMBDA_MAX);
    qmax = av_clip(qmax, 1, FF_LAMBDA_MAX);

    if (qmax < qmin)
        qmax = qmin;

    *qmin_ret = qmin;
    *qmax_ret = qmax;
}

static double modify_qscale(MpegEncContext *s, RateControlEntry *rce,
                            double q, int frame_num)
{
    RateControlContext *rcc  = &s->rc_context;
    const double buffer_size = s->avctx->rc_buffer_size;
    const double fps         = 1 / av_q2d(s->avctx->time_base);
    const double min_rate    = s->avctx->rc_min_rate / fps;
    const double max_rate    = s->avctx->rc_max_rate / fps;
    const int pict_type      = rce->new_pict_type;
    int qmin, qmax;

    get_qminmax(&qmin, &qmax, s, pict_type);

    /* modulation */
    if (s->rc_qmod_freq &&
        frame_num % s->rc_qmod_freq == 0 &&
        pict_type == AV_PICTURE_TYPE_P)
        q *= s->rc_qmod_amp;

    /* buffer overflow/underflow protection */
    if (buffer_size) {
        double expected_size = rcc->buffer_index;
        double q_limit;

        if (min_rate) {
            double d = 2 * (buffer_size - expected_size) / buffer_size;
            if (d > 1.0)
                d = 1.0;
            else if (d < 0.0001)
                d = 0.0001;
            q *= pow(d, 1.0 / s->rc_buffer_aggressivity);

            q_limit = bits2qp(rce,
                              FFMAX((min_rate - buffer_size + rcc->buffer_index) *
                                    s->avctx->rc_min_vbv_overflow_use, 1));

            if (q > q_limit) {
                if (s->avctx->debug & FF_DEBUG_RC)
                    av_log(s->avctx, AV_LOG_DEBUG,
                           "limiting QP %f -> %f\n", q, q_limit);
                q = q_limit;
            }
        }

        if (max_rate) {
            double d = 2 * expected_size / buffer_size;
            if (d > 1.0)
                d = 1.0;
            else if (d < 0.0001)
                d = 0.0001;
            q /= pow(d, 1.0 / s->rc_buffer_aggressivity);

            q_limit = bits2qp(rce,
                              FFMAX(rcc->buffer_index *
                                    s->avctx->rc_max_available_vbv_use,
                                    1));
            if (q < q_limit) {
                if (s->avctx->debug & FF_DEBUG_RC)
                    av_log(s->avctx, AV_LOG_DEBUG,
                           "limiting QP %f -> %f\n", q, q_limit);
                q = q_limit;
            }
        }
    }
    ff_dlog(s, "q:%f max:%f min:%f size:%f index:%f agr:%f\n",
            q, max_rate, min_rate, buffer_size, rcc->buffer_index,
            s->rc_buffer_aggressivity);
    if (s->rc_qsquish == 0.0 || qmin == qmax) {
        if (q < qmin)
            q = qmin;
        else if (q > qmax)
            q = qmax;
    } else {
        double min2 = log(qmin);
        double max2 = log(qmax);

        q  = log(q);
        q  = (q - min2) / (max2 - min2) - 0.5;
        q *= -4.0;
        q  = 1.0 / (1.0 + exp(q));
        q  = q * (max2 - min2) + min2;

        q = exp(q);
    }

    return q;
}

/**
 * Modify the bitrate curve from pass1 for one frame.
 */
219
static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398
                         double rate_factor, int frame_num)
{
    RateControlContext *rcc = &s->rc_context;
    AVCodecContext *a       = s->avctx;
    const int pict_type     = rce->new_pict_type;
    const double mb_num     = s->mb_num;
    double q, bits;
    int i;

    double const_values[] = {
        M_PI,
        M_E,
        rce->i_tex_bits * rce->qscale,
        rce->p_tex_bits * rce->qscale,
        (rce->i_tex_bits + rce->p_tex_bits) * (double)rce->qscale,
        rce->mv_bits / mb_num,
        rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code) * 0.5 : rce->f_code,
        rce->i_count / mb_num,
        rce->mc_mb_var_sum / mb_num,
        rce->mb_var_sum / mb_num,
        rce->pict_type == AV_PICTURE_TYPE_I,
        rce->pict_type == AV_PICTURE_TYPE_P,
        rce->pict_type == AV_PICTURE_TYPE_B,
        rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
        a->qcompress,
        rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I],
        rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
        rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
        rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B],
        (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
        0
    };

    bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
    if (isnan(bits)) {
        av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->rc_eq);
        return -1;
    }

    rcc->pass1_rc_eq_output_sum += bits;
    bits *= rate_factor;
    if (bits < 0.0)
        bits = 0.0;
    bits += 1.0; // avoid 1/0 issues

    /* user override */
    for (i = 0; i < s->avctx->rc_override_count; i++) {
        RcOverride *rco = s->avctx->rc_override;
        if (rco[i].start_frame > frame_num)
            continue;
        if (rco[i].end_frame < frame_num)
            continue;

        if (rco[i].qscale)
            bits = qp2bits(rce, rco[i].qscale);  // FIXME move at end to really force it?
        else
            bits *= rco[i].quality_factor;
    }

    q = bits2qp(rce, bits);

    /* I/B difference */
    if (pict_type == AV_PICTURE_TYPE_I && s->avctx->i_quant_factor < 0.0)
        q = -q * s->avctx->i_quant_factor + s->avctx->i_quant_offset;
    else if (pict_type == AV_PICTURE_TYPE_B && s->avctx->b_quant_factor < 0.0)
        q = -q * s->avctx->b_quant_factor + s->avctx->b_quant_offset;
    if (q < 1)
        q = 1;

    return q;
}

static int init_pass2(MpegEncContext *s)
{
    RateControlContext *rcc = &s->rc_context;
    AVCodecContext *a       = s->avctx;
    int i, toobig;
    double fps             = 1 / av_q2d(s->avctx->time_base);
    double complexity[5]   = { 0 }; // approximate bits at quant=1
    uint64_t const_bits[5] = { 0 }; // quantizer independent bits
    uint64_t all_const_bits;
    uint64_t all_available_bits = (uint64_t)(s->bit_rate *
                                             (double)rcc->num_entries / fps);
    double rate_factor          = 0;
    double step;
    const int filter_size = (int)(a->qblur * 4) | 1;
    double expected_bits;
    double *qscale, *blurred_qscale, qscale_sum;

    /* find complexity & const_bits & decide the pict_types */
    for (i = 0; i < rcc->num_entries; i++) {
        RateControlEntry *rce = &rcc->entry[i];

        rce->new_pict_type                = rce->pict_type;
        rcc->i_cplx_sum[rce->pict_type]  += rce->i_tex_bits * rce->qscale;
        rcc->p_cplx_sum[rce->pict_type]  += rce->p_tex_bits * rce->qscale;
        rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
        rcc->frame_count[rce->pict_type]++;

        complexity[rce->new_pict_type] += (rce->i_tex_bits + rce->p_tex_bits) *
                                          (double)rce->qscale;
        const_bits[rce->new_pict_type] += rce->mv_bits + rce->misc_bits;
    }

    all_const_bits = const_bits[AV_PICTURE_TYPE_I] +
                     const_bits[AV_PICTURE_TYPE_P] +
                     const_bits[AV_PICTURE_TYPE_B];

    if (all_available_bits < all_const_bits) {
        av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
        return -1;
    }

    qscale         = av_malloc(sizeof(double) * rcc->num_entries);
    blurred_qscale = av_malloc(sizeof(double) * rcc->num_entries);
    if (!qscale || !blurred_qscale) {
        av_free(qscale);
        av_free(blurred_qscale);
        return AVERROR(ENOMEM);
    }
    toobig = 0;

    for (step = 256 * 256; step > 0.0000001; step *= 0.5) {
        expected_bits = 0;
        rate_factor  += step;

        rcc->buffer_index = s->avctx->rc_buffer_size / 2;

        /* find qscale */
        for (i = 0; i < rcc->num_entries; i++) {
            RateControlEntry *rce = &rcc->entry[i];

            qscale[i] = get_qscale(s, &rcc->entry[i], rate_factor, i);
            rcc->last_qscale_for[rce->pict_type] = qscale[i];
        }
        assert(filter_size % 2 == 1);

        /* fixed I/B QP relative to P mode */
        for (i = rcc->num_entries - 1; i >= 0; i--) {
            RateControlEntry *rce = &rcc->entry[i];

            qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
        }

        /* smooth curve */
        for (i = 0; i < rcc->num_entries; i++) {
            RateControlEntry *rce = &rcc->entry[i];
            const int pict_type   = rce->new_pict_type;
            int j;
            double q = 0.0, sum = 0.0;

            for (j = 0; j < filter_size; j++) {
                int index    = i + j - filter_size / 2;
                double d     = index - i;
                double coeff = a->qblur == 0 ? 1.0 : exp(-d * d / (a->qblur * a->qblur));

                if (index < 0 || index >= rcc->num_entries)
                    continue;
                if (pict_type != rcc->entry[index].new_pict_type)
                    continue;
                q   += qscale[index] * coeff;
                sum += coeff;
            }
            blurred_qscale[i] = q / sum;
        }

        /* find expected bits */
        for (i = 0; i < rcc->num_entries; i++) {
            RateControlEntry *rce = &rcc->entry[i];
            double bits;

            rce->new_qscale = modify_qscale(s, rce, blurred_qscale[i], i);

            bits  = qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
            bits += 8 * ff_vbv_update(s, bits);

            rce->expected_bits = expected_bits;
            expected_bits     += bits;
        }
399

400 401 402 403 404 405 406
        ff_dlog(s->avctx,
                "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
                expected_bits, (int)all_available_bits, rate_factor);
        if (expected_bits > all_available_bits) {
            rate_factor -= step;
            ++toobig;
        }
407
    }
408 409
    av_free(qscale);
    av_free(blurred_qscale);
410

411 412 413 414 415 416 417 418 419
    /* check bitrate calculations and print info */
    qscale_sum = 0.0;
    for (i = 0; i < rcc->num_entries; i++) {
        ff_dlog(s, "[lavc rc] entry[%d].new_qscale = %.3f  qp = %.3f\n",
                i,
                rcc->entry[i].new_qscale,
                rcc->entry[i].new_qscale / FF_QP2LAMBDA);
        qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA,
                              s->avctx->qmin, s->avctx->qmax);
420
    }
421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444
    assert(toobig <= 40);
    av_log(s->avctx, AV_LOG_DEBUG,
           "[lavc rc] requested bitrate: %d bps  expected bitrate: %d bps\n",
           s->bit_rate,
           (int)(expected_bits / ((double)all_available_bits / s->bit_rate)));
    av_log(s->avctx, AV_LOG_DEBUG,
           "[lavc rc] estimated target average qp: %.3f\n",
           (float)qscale_sum / rcc->num_entries);
    if (toobig == 0) {
        av_log(s->avctx, AV_LOG_INFO,
               "[lavc rc] Using all of requested bitrate is not "
               "necessary for this video with these parameters.\n");
    } else if (toobig == 40) {
        av_log(s->avctx, AV_LOG_ERROR,
               "[lavc rc] Error: bitrate too low for this video "
               "with these parameters.\n");
        return -1;
    } else if (fabs(expected_bits / all_available_bits - 1.0) > 0.01) {
        av_log(s->avctx, AV_LOG_ERROR,
               "[lavc rc] Error: 2pass curve failed to converge\n");
        return -1;
    }

    return 0;
445 446
}

447
av_cold int ff_rate_control_init(MpegEncContext *s)
448
{
449
    RateControlContext *rcc = &s->rc_context;
450
    int i, res;
451
    static const char * const const_names[] = {
452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473
        "PI",
        "E",
        "iTex",
        "pTex",
        "tex",
        "mv",
        "fCode",
        "iCount",
        "mcVar",
        "var",
        "isI",
        "isP",
        "isB",
        "avgQP",
        "qComp",
        "avgIITex",
        "avgPITex",
        "avgPPTex",
        "avgBPTex",
        "avgTex",
        NULL
    };
474
    static double (* const func1[])(void *, double) = {
475 476
        (double (*)(void *, double)) bits2qp,
        (double (*)(void *, double)) qp2bits,
477 478
        NULL
    };
479
    static const char * const func1_names[] = {
480 481 482 483
        "bits2qp",
        "qp2bits",
        NULL
    };
484 485
    emms_c();

486
    res = av_expr_parse(&rcc->rc_eq_eval,
487
                        s->rc_eq ? s->rc_eq : "tex^qComp",
488 489
                        const_names, func1_names, func1,
                        NULL, NULL, 0, s->avctx);
490
    if (res < 0) {
491
        av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->rc_eq);
492
        return res;
493 494
    }

495 496 497 498 499 500 501 502 503 504 505 506
    for (i = 0; i < 5; i++) {
        rcc->pred[i].coeff = FF_QP2LAMBDA * 7.0;
        rcc->pred[i].count = 1.0;
        rcc->pred[i].decay = 0.4;

        rcc->i_cplx_sum [i] =
        rcc->p_cplx_sum [i] =
        rcc->mv_bits_sum[i] =
        rcc->qscale_sum [i] =
        rcc->frame_count[i] = 1; // 1 is better because of 1/0 and such

        rcc->last_qscale_for[i] = FF_QP2LAMBDA * 5;
507
    }
508
    rcc->buffer_index = s->avctx->rc_initial_buffer_occupancy;
509

510
    if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
511
        int i;
512
        char *p;
513

514
        /* find number of pics */
515 516 517 518 519
        p = s->avctx->stats_in;
        for (i = -1; p; i++)
            p = strchr(p + 1, ';');
        i += s->max_b_frames;
        if (i <= 0 || i >= INT_MAX / sizeof(RateControlEntry))
520
            return -1;
521 522
        rcc->entry       = av_mallocz(i * sizeof(RateControlEntry));
        rcc->num_entries = i;
523 524
        if (!rcc->entry)
            return AVERROR(ENOMEM);
525

526 527
        /* init all to skipped P-frames
         * (with B-frames we might have a not encoded frame at the end FIXME) */
528 529 530 531 532 533 534
        for (i = 0; i < rcc->num_entries; i++) {
            RateControlEntry *rce = &rcc->entry[i];

            rce->pict_type  = rce->new_pict_type = AV_PICTURE_TYPE_P;
            rce->qscale     = rce->new_qscale    = FF_QP2LAMBDA * 2;
            rce->misc_bits  = s->mb_num + 10;
            rce->mb_var_sum = s->mb_num * 100;
535 536
        }

537
        /* read stats */
538 539
        p = s->avctx->stats_in;
        for (i = 0; i < rcc->num_entries - s->max_b_frames; i++) {
540 541 542
            RateControlEntry *rce;
            int picture_number;
            int e;
543 544
            char *next;

545 546
            next = strchr(p, ';');
            if (next) {
547
                (*next) = 0; // sscanf is unbelievably slow on looong strings // FIXME copy / do not write
548 549
                next++;
            }
550
            e = sscanf(p, " in:%d ", &picture_number);
551 552 553

            assert(picture_number >= 0);
            assert(picture_number < rcc->num_entries);
554 555 556 557 558 559 560 561 562 563 564 565
            rce = &rcc->entry[picture_number];

            e += sscanf(p, " in:%*d out:%*d type:%d q:%f itex:%d ptex:%d mv:%d misc:%d fcode:%d bcode:%d mc-var:%d var:%d icount:%d skipcount:%d hbits:%d",
                        &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits,
                        &rce->mv_bits, &rce->misc_bits,
                        &rce->f_code, &rce->b_code,
                        &rce->mc_mb_var_sum, &rce->mb_var_sum,
                        &rce->i_count, &rce->skip_count, &rce->header_bits);
            if (e != 14) {
                av_log(s->avctx, AV_LOG_ERROR,
                       "statistics are damaged at line %d, parser out=%d\n",
                       i, e);
566 567
                return -1;
            }
568

569
            p = next;
570
        }
571

572 573
        if (init_pass2(s) < 0) {
            ff_rate_control_uninit(s);
574
            return -1;
575
        }
576
    }
577

578
    if (!(s->avctx->flags & AV_CODEC_FLAG_PASS2)) {
579 580
        rcc->short_term_qsum   = 0.001;
        rcc->short_term_qcount = 0.001;
581

582 583
        rcc->pass1_rc_eq_output_sum = 0.001;
        rcc->pass1_wanted_bits      = 0.001;
584

585
        if (s->avctx->qblur > 1.0) {
586 587 588
            av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
            return -1;
        }
589
        /* init stuff with the user specified complexity */
590
        if (s->rc_initial_cplx) {
591
            for (i = 0; i < 60 * 30; i++) {
592
                double bits = s->rc_initial_cplx * (i / 10000.0 + 1.0) * s->mb_num;
593
                RateControlEntry rce;
594

595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620
                if (i % ((s->gop_size + 3) / 4) == 0)
                    rce.pict_type = AV_PICTURE_TYPE_I;
                else if (i % (s->max_b_frames + 1))
                    rce.pict_type = AV_PICTURE_TYPE_B;
                else
                    rce.pict_type = AV_PICTURE_TYPE_P;

                rce.new_pict_type = rce.pict_type;
                rce.mc_mb_var_sum = bits * s->mb_num / 100000;
                rce.mb_var_sum    = s->mb_num;

                rce.qscale    = FF_QP2LAMBDA * 2;
                rce.f_code    = 2;
                rce.b_code    = 1;
                rce.misc_bits = 1;

                if (s->pict_type == AV_PICTURE_TYPE_I) {
                    rce.i_count    = s->mb_num;
                    rce.i_tex_bits = bits;
                    rce.p_tex_bits = 0;
                    rce.mv_bits    = 0;
                } else {
                    rce.i_count    = 0; // FIXME we do know this approx
                    rce.i_tex_bits = 0;
                    rce.p_tex_bits = bits * 0.9;
                    rce.mv_bits    = bits * 0.1;
621
                }
622 623
                rcc->i_cplx_sum[rce.pict_type]  += rce.i_tex_bits * rce.qscale;
                rcc->p_cplx_sum[rce.pict_type]  += rce.p_tex_bits * rce.qscale;
624
                rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
625 626 627
                rcc->frame_count[rce.pict_type]++;

                get_qscale(s, &rce, rcc->pass1_wanted_bits / rcc->pass1_rc_eq_output_sum, i);
628

629 630
                // FIXME misbehaves a little for variable fps
                rcc->pass1_wanted_bits += s->bit_rate / (1 / av_q2d(s->avctx->time_base));
631 632 633
            }
        }
    }
634

635 636 637
    return 0;
}

638 639 640 641
av_cold void ff_rate_control_uninit(MpegEncContext *s)
{
    RateControlContext *rcc = &s->rc_context;
    emms_c();
642

643 644 645
    av_expr_free(rcc->rc_eq_eval);
    av_freep(&rcc->entry);
}
646

647 648 649 650 651 652 653
int ff_vbv_update(MpegEncContext *s, int frame_size)
{
    RateControlContext *rcc = &s->rc_context;
    const double fps        = 1 / av_q2d(s->avctx->time_base);
    const int buffer_size   = s->avctx->rc_buffer_size;
    const double min_rate   = s->avctx->rc_min_rate / fps;
    const double max_rate   = s->avctx->rc_max_rate / fps;
654

655 656
    ff_dlog(s, "%d %f %d %f %f\n",
            buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
657

658 659
    if (buffer_size) {
        int left;
660

661 662 663 664
        rcc->buffer_index -= frame_size;
        if (rcc->buffer_index < 0) {
            av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
            rcc->buffer_index = 0;
665 666
        }

667 668
        left = buffer_size - rcc->buffer_index - 1;
        rcc->buffer_index += av_clip(left, min_rate, max_rate);
669

670 671
        if (rcc->buffer_index > buffer_size) {
            int stuffing = ceil((rcc->buffer_index - buffer_size) / 8);
672

673 674 675
            if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
                stuffing = 4;
            rcc->buffer_index -= 8 * stuffing;
676

677 678
            if (s->avctx->debug & FF_DEBUG_RC)
                av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
679

680 681 682 683
            return stuffing;
        }
    }
    return 0;
684 685 686
}

static double predict_size(Predictor *p, double q, double var)
687
{
688
    return p->coeff * var / (q * p->count);
689 690 691 692
}

static void update_predictor(Predictor *p, double q, double var, double size)
{
693 694 695
    double new_coeff = size * q / (var + 1);
    if (var < 10)
        return;
696

697 698
    p->count *= p->decay;
    p->coeff *= p->decay;
699
    p->count++;
700
    p->coeff += new_coeff;
701 702
}

703 704
static void adaptive_quantization(MpegEncContext *s, double q)
{
705
    int i;
706 707 708
    const float lumi_masking         = s->avctx->lumi_masking / (128.0 * 128.0);
    const float dark_masking         = s->avctx->dark_masking / (128.0 * 128.0);
    const float temp_cplx_masking    = s->avctx->temporal_cplx_masking;
709
    const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
710
    const float p_masking            = s->avctx->p_masking;
711
    const float border_masking       = s->border_masking;
712 713 714 715 716 717 718 719 720 721 722 723 724 725 726
    float bits_sum                   = 0.0;
    float cplx_sum                   = 0.0;
    float *cplx_tab                  = s->cplx_tab;
    float *bits_tab                  = s->bits_tab;
    const int qmin                   = s->avctx->mb_lmin;
    const int qmax                   = s->avctx->mb_lmax;
    Picture *const pic               = &s->current_picture;
    const int mb_width               = s->mb_width;
    const int mb_height              = s->mb_height;

    for (i = 0; i < s->mb_num; i++) {
        const int mb_xy = s->mb_index2xy[i];
        float temp_cplx = sqrt(pic->mc_mb_var[mb_xy]); // FIXME merge in pow()
        float spat_cplx = sqrt(pic->mb_var[mb_xy]);
        const int lumi  = pic->mb_mean[mb_xy];
727
        float bits, cplx, factor;
728 729 730 731
        int mb_x = mb_xy % s->mb_stride;
        int mb_y = mb_xy / s->mb_stride;
        int mb_distance;
        float mb_factor = 0.0;
732
        if (spat_cplx < 4)
733
            spat_cplx = 4;              // FIXME fine-tune
734
        if (temp_cplx < 4)
735
            temp_cplx = 4;              // FIXME fine-tune
736 737 738 739 740 741 742

        if ((s->mb_type[mb_xy] & CANDIDATE_MB_TYPE_INTRA)) { // FIXME hq mode
            cplx   = spat_cplx;
            factor = 1.0 + p_masking;
        } else {
            cplx   = temp_cplx;
            factor = pow(temp_cplx, -temp_cplx_masking);
743
        }
744
        factor *= pow(spat_cplx, -spatial_cplx_masking);
745

746 747
        if (lumi > 127)
            factor *= (1.0 - (lumi - 128) * (lumi - 128) * lumi_masking);
748
        else
749 750 751 752 753 754 755 756
            factor *= (1.0 - (lumi - 128) * (lumi - 128) * dark_masking);

        if (mb_x < mb_width / 5) {
            mb_distance = mb_width / 5 - mb_x;
            mb_factor   = (float)mb_distance / (float)(mb_width / 5);
        } else if (mb_x > 4 * mb_width / 5) {
            mb_distance = mb_x - 4 * mb_width / 5;
            mb_factor   = (float)mb_distance / (float)(mb_width / 5);
757
        }
758 759 760 761 762 763 764 765
        if (mb_y < mb_height / 5) {
            mb_distance = mb_height / 5 - mb_y;
            mb_factor   = FFMAX(mb_factor,
                                (float)mb_distance / (float)(mb_height / 5));
        } else if (mb_y > 4 * mb_height / 5) {
            mb_distance = mb_y - 4 * mb_height / 5;
            mb_factor   = FFMAX(mb_factor,
                                (float)mb_distance / (float)(mb_height / 5));
766 767
        }

768
        factor *= 1.0 - border_masking * mb_factor;
769

770 771
        if (factor < 0.00001)
            factor = 0.00001;
772

773 774 775 776 777
        bits        = cplx * factor;
        cplx_sum   += cplx;
        bits_sum   += bits;
        cplx_tab[i] = cplx;
        bits_tab[i] = bits;
778 779
    }

Diego Biurrun's avatar
Diego Biurrun committed
780
    /* handle qmin/qmax clipping */
781
    if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
782 783 784 785
        float factor = bits_sum / cplx_sum;
        for (i = 0; i < s->mb_num; i++) {
            float newq = q * cplx_tab[i] / bits_tab[i];
            newq *= factor;
786

787
            if (newq > qmax) {
788
                bits_sum -= bits_tab[i];
789 790
                cplx_sum -= cplx_tab[i] * q / qmax;
            } else if (newq < qmin) {
791
                bits_sum -= bits_tab[i];
792
                cplx_sum -= cplx_tab[i] * q / qmin;
793 794
            }
        }
795 796 797 798
        if (bits_sum < 0.001)
            bits_sum = 0.001;
        if (cplx_sum < 0.001)
            cplx_sum = 0.001;
799
    }
800

801 802 803
    for (i = 0; i < s->mb_num; i++) {
        const int mb_xy = s->mb_index2xy[i];
        float newq      = q * cplx_tab[i] / bits_tab[i];
804 805
        int intq;

806
        if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
807
            newq *= bits_sum / cplx_sum;
808 809
        }

810
        intq = (int)(newq + 0.5);
811

812 813 814 815 816
        if (intq > qmax)
            intq = qmax;
        else if (intq < qmin)
            intq = qmin;
        s->lambda_table[mb_xy] = intq;
817 818
    }
}
819

820 821 822 823
void ff_get_2pass_fcode(MpegEncContext *s)
{
    RateControlContext *rcc = &s->rc_context;
    RateControlEntry *rce   = &rcc->entry[s->picture_number];
824

825 826
    s->f_code = rce->f_code;
    s->b_code = rce->b_code;
827 828
}

829
// FIXME rd or at least approx for dquant
830

831
float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
832 833
{
    float q;
834
    int qmin, qmax;
835 836 837 838
    float br_compensation;
    double diff;
    double short_term_q;
    double fps;
839
    int picture_number = s->picture_number;
840
    int64_t wanted_bits;
841 842
    RateControlContext *rcc = &s->rc_context;
    AVCodecContext *a       = s->avctx;
843 844 845 846
    RateControlEntry local_rce, *rce;
    double bits;
    double rate_factor;
    int var;
847 848
    const int pict_type = s->pict_type;
    Picture * const pic = &s->current_picture;
849 850
    emms_c();

851
    get_qminmax(&qmin, &qmax, s, pict_type);
852

853 854 855 856 857 858 859 860
    fps = 1 / av_q2d(s->avctx->time_base);
    /* update predictors */
    if (picture_number > 2 && !dry_run) {
        const int last_var = s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum
                                                                    : rcc->last_mc_mb_var_sum;
        update_predictor(&rcc->pred[s->last_pict_type],
                         rcc->last_qscale,
                         sqrt(last_var), s->frame_bits);
861 862
    }

863
    if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
864 865 866 867 868
        assert(picture_number >= 0);
        assert(picture_number < rcc->num_entries);
        rce         = &rcc->entry[picture_number];
        wanted_bits = rce->expected_bits;
    } else {
869
        Picture *dts_pic;
870
        rce = &local_rce;
871

872 873 874 875 876
        /* FIXME add a dts field to AVFrame and ensure it is set and use it
         * here instead of reordering but the reordering is simpler for now
         * until H.264 B-pyramid must be handled. */
        if (s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
            dts_pic = s->current_picture_ptr;
877
        else
878
            dts_pic = s->last_picture_ptr;
879

880
        if (!dts_pic || dts_pic->f->pts == AV_NOPTS_VALUE)
881
            wanted_bits = (uint64_t)(s->bit_rate * (double)picture_number / fps);
882
        else
883
            wanted_bits = (uint64_t)(s->bit_rate * (double)dts_pic->f->pts / fps);
884
    }
885

886 887 888 889
    diff = s->total_bits - wanted_bits;
    br_compensation = (a->bit_rate_tolerance - diff) / a->bit_rate_tolerance;
    if (br_compensation <= 0.0)
        br_compensation = 0.001;
890

891
    var = pict_type == AV_PICTURE_TYPE_I ? pic->mb_var_sum : pic->mc_mb_var_sum;
892

Fabrice Bellard's avatar
Fabrice Bellard committed
893
    short_term_q = 0; /* avoid warning */
894
    if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
895
        if (pict_type != AV_PICTURE_TYPE_I)
896 897
            assert(pict_type == rce->new_pict_type);

898
        q = rce->new_qscale / br_compensation;
899
        ff_dlog(s, "%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale,
900
                br_compensation, s->frame_bits, var, pict_type);
901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921
    } else {
        rce->pict_type     =
        rce->new_pict_type = pict_type;
        rce->mc_mb_var_sum = pic->mc_mb_var_sum;
        rce->mb_var_sum    = pic->mb_var_sum;
        rce->qscale        = FF_QP2LAMBDA * 2;
        rce->f_code        = s->f_code;
        rce->b_code        = s->b_code;
        rce->misc_bits     = 1;

        bits = predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
        if (pict_type == AV_PICTURE_TYPE_I) {
            rce->i_count    = s->mb_num;
            rce->i_tex_bits = bits;
            rce->p_tex_bits = 0;
            rce->mv_bits    = 0;
        } else {
            rce->i_count    = 0;    // FIXME we do know this approx
            rce->i_tex_bits = 0;
            rce->p_tex_bits = bits * 0.9;
            rce->mv_bits    = bits * 0.1;
922
        }
923 924
        rcc->i_cplx_sum[pict_type]  += rce->i_tex_bits * rce->qscale;
        rcc->p_cplx_sum[pict_type]  += rce->p_tex_bits * rce->qscale;
925
        rcc->mv_bits_sum[pict_type] += rce->mv_bits;
926
        rcc->frame_count[pict_type]++;
927

928 929 930
        bits        = rce->i_tex_bits + rce->p_tex_bits;
        rate_factor = rcc->pass1_wanted_bits /
                      rcc->pass1_rc_eq_output_sum * br_compensation;
931

932
        q = get_qscale(s, rce, rate_factor, picture_number);
933 934
        if (q < 0)
            return -1;
935

936 937 938
        assert(q > 0.0);
        q = get_diff_limited_q(s, rce, q);
        assert(q > 0.0);
939

940 941 942 943
        // FIXME type dependent blur like in 2-pass
        if (pict_type == AV_PICTURE_TYPE_P || s->intra_only) {
            rcc->short_term_qsum   *= a->qblur;
            rcc->short_term_qcount *= a->qblur;
944

945
            rcc->short_term_qsum += q;
946
            rcc->short_term_qcount++;
947
            q = short_term_q = rcc->short_term_qsum / rcc->short_term_qcount;
948
        }
949
        assert(q > 0.0);
950

951
        q = modify_qscale(s, rce, q, picture_number);
952

953
        rcc->pass1_wanted_bits += s->bit_rate / fps;
954

955
        assert(q > 0.0);
956
    }
Michael Niedermayer's avatar
Michael Niedermayer committed
957

958 959 960 961 962 963 964 965 966 967
    if (s->avctx->debug & FF_DEBUG_RC) {
        av_log(s->avctx, AV_LOG_DEBUG,
               "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f "
               "size:%d var:%d/%d br:%d fps:%d\n",
               av_get_picture_type_char(pict_type),
               qmin, q, qmax, picture_number,
               (int)wanted_bits / 1000, (int)s->total_bits / 1000,
               br_compensation, short_term_q, s->frame_bits,
               pic->mb_var_sum, pic->mc_mb_var_sum,
               s->bit_rate / 1000, (int)fps);
Michael Niedermayer's avatar
Michael Niedermayer committed
968
    }
969

970 971 972 973
    if (q < qmin)
        q = qmin;
    else if (q > qmax)
        q = qmax;
974

975
    if (s->adaptive_quant)
976 977
        adaptive_quantization(s, q);
    else
978
        q = (int)(q + 0.5);
979

980 981 982 983
    if (!dry_run) {
        rcc->last_qscale        = q;
        rcc->last_mc_mb_var_sum = pic->mc_mb_var_sum;
        rcc->last_mb_var_sum    = pic->mb_var_sum;
984
    }
985
    return q;
986
}