yading@10: /* yading@10: * Rate control for video encoders yading@10: * yading@10: * Copyright (c) 2002-2004 Michael Niedermayer yading@10: * yading@10: * This file is part of FFmpeg. yading@10: * yading@10: * FFmpeg is free software; you can redistribute it and/or yading@10: * modify it under the terms of the GNU Lesser General Public yading@10: * License as published by the Free Software Foundation; either yading@10: * version 2.1 of the License, or (at your option) any later version. yading@10: * yading@10: * FFmpeg is distributed in the hope that it will be useful, yading@10: * but WITHOUT ANY WARRANTY; without even the implied warranty of yading@10: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@10: * Lesser General Public License for more details. yading@10: * yading@10: * You should have received a copy of the GNU Lesser General Public yading@10: * License along with FFmpeg; if not, write to the Free Software yading@10: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@10: */ yading@10: yading@10: /** yading@10: * @file yading@10: * Rate control for video encoders. yading@10: */ yading@10: yading@10: #include "avcodec.h" yading@10: #include "ratecontrol.h" yading@10: #include "mpegvideo.h" yading@10: #include "libavutil/eval.h" yading@10: yading@10: #undef NDEBUG // Always check asserts, the speed effect is far too small to disable them. yading@10: #include yading@10: yading@10: #ifndef M_E yading@10: #define M_E 2.718281828 yading@10: #endif yading@10: yading@10: static int init_pass2(MpegEncContext *s); yading@10: static double get_qscale(MpegEncContext *s, RateControlEntry *rce, yading@10: double rate_factor, int frame_num); yading@10: yading@10: void ff_write_pass1_stats(MpegEncContext *s) yading@10: { yading@10: snprintf(s->avctx->stats_out, 256, yading@10: "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d " yading@10: "fcode:%d bcode:%d mc-var:%d var:%d icount:%d skipcount:%d hbits:%d;\n", yading@10: s->current_picture_ptr->f.display_picture_number, yading@10: s->current_picture_ptr->f.coded_picture_number, yading@10: s->pict_type, yading@10: s->current_picture.f.quality, yading@10: s->i_tex_bits, yading@10: s->p_tex_bits, yading@10: s->mv_bits, yading@10: s->misc_bits, yading@10: s->f_code, yading@10: s->b_code, yading@10: s->current_picture.mc_mb_var_sum, yading@10: s->current_picture.mb_var_sum, yading@10: s->i_count, s->skip_count, yading@10: s->header_bits); yading@10: } yading@10: yading@10: static double get_fps(AVCodecContext *avctx) yading@10: { yading@10: return 1.0 / av_q2d(avctx->time_base) / FFMAX(avctx->ticks_per_frame, 1); yading@10: } yading@10: yading@10: static inline double qp2bits(RateControlEntry *rce, double qp) yading@10: { yading@10: if (qp <= 0.0) { yading@10: av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n"); yading@10: } yading@10: return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / qp; yading@10: } yading@10: yading@10: static inline double bits2qp(RateControlEntry *rce, double bits) yading@10: { yading@10: if (bits < 0.9) { yading@10: av_log(NULL, AV_LOG_ERROR, "bits<0.9\n"); yading@10: } yading@10: return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / bits; yading@10: } yading@10: yading@10: int ff_rate_control_init(MpegEncContext *s) yading@10: { yading@10: RateControlContext *rcc = &s->rc_context; yading@10: int i, res; yading@10: static const char * const const_names[] = { yading@10: "PI", yading@10: "E", yading@10: "iTex", yading@10: "pTex", yading@10: "tex", yading@10: "mv", yading@10: "fCode", yading@10: "iCount", yading@10: "mcVar", yading@10: "var", yading@10: "isI", yading@10: "isP", yading@10: "isB", yading@10: "avgQP", yading@10: "qComp", yading@10: #if 0 yading@10: "lastIQP", yading@10: "lastPQP", yading@10: "lastBQP", yading@10: "nextNonBQP", yading@10: #endif yading@10: "avgIITex", yading@10: "avgPITex", yading@10: "avgPPTex", yading@10: "avgBPTex", yading@10: "avgTex", yading@10: NULL yading@10: }; yading@10: static double (* const func1[])(void *, double) = { yading@10: (void *)bits2qp, yading@10: (void *)qp2bits, yading@10: NULL yading@10: }; yading@10: static const char * const func1_names[] = { yading@10: "bits2qp", yading@10: "qp2bits", yading@10: NULL yading@10: }; yading@10: emms_c(); yading@10: yading@10: if (!s->avctx->rc_max_available_vbv_use && s->avctx->rc_buffer_size) { yading@10: if (s->avctx->rc_max_rate) { yading@10: s->avctx->rc_max_available_vbv_use = av_clipf(s->avctx->rc_max_rate/(s->avctx->rc_buffer_size*get_fps(s->avctx)), 1.0/3, 1.0); yading@10: } else yading@10: s->avctx->rc_max_available_vbv_use = 1.0; yading@10: } yading@10: yading@10: res = av_expr_parse(&rcc->rc_eq_eval, yading@10: s->avctx->rc_eq ? s->avctx->rc_eq : "tex^qComp", yading@10: const_names, func1_names, func1, yading@10: NULL, NULL, 0, s->avctx); yading@10: if (res < 0) { yading@10: av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->avctx->rc_eq); yading@10: return res; yading@10: } yading@10: yading@10: for (i = 0; i < 5; i++) { yading@10: rcc->pred[i].coeff = FF_QP2LAMBDA * 7.0; yading@10: rcc->pred[i].count = 1.0; yading@10: rcc->pred[i].decay = 0.4; yading@10: yading@10: rcc->i_cplx_sum [i] = yading@10: rcc->p_cplx_sum [i] = yading@10: rcc->mv_bits_sum[i] = yading@10: rcc->qscale_sum [i] = yading@10: rcc->frame_count[i] = 1; // 1 is better because of 1/0 and such yading@10: yading@10: rcc->last_qscale_for[i] = FF_QP2LAMBDA * 5; yading@10: } yading@10: rcc->buffer_index = s->avctx->rc_initial_buffer_occupancy; yading@10: if (!rcc->buffer_index) yading@10: rcc->buffer_index = s->avctx->rc_buffer_size * 3 / 4; yading@10: yading@10: if (s->flags & CODEC_FLAG_PASS2) { yading@10: int i; yading@10: char *p; yading@10: yading@10: /* find number of pics */ yading@10: p = s->avctx->stats_in; yading@10: for (i = -1; p; i++) yading@10: p = strchr(p + 1, ';'); yading@10: i += s->max_b_frames; yading@10: if (i <= 0 || i >= INT_MAX / sizeof(RateControlEntry)) yading@10: return -1; yading@10: rcc->entry = av_mallocz(i * sizeof(RateControlEntry)); yading@10: rcc->num_entries = i; yading@10: yading@10: /* init all to skipped p frames yading@10: * (with b frames we might have a not encoded frame at the end FIXME) */ yading@10: for (i = 0; i < rcc->num_entries; i++) { yading@10: RateControlEntry *rce = &rcc->entry[i]; yading@10: yading@10: rce->pict_type = rce->new_pict_type = AV_PICTURE_TYPE_P; yading@10: rce->qscale = rce->new_qscale = FF_QP2LAMBDA * 2; yading@10: rce->misc_bits = s->mb_num + 10; yading@10: rce->mb_var_sum = s->mb_num * 100; yading@10: } yading@10: yading@10: /* read stats */ yading@10: p = s->avctx->stats_in; yading@10: for (i = 0; i < rcc->num_entries - s->max_b_frames; i++) { yading@10: RateControlEntry *rce; yading@10: int picture_number; yading@10: int e; yading@10: char *next; yading@10: yading@10: next = strchr(p, ';'); yading@10: if (next) { yading@10: (*next) = 0; // sscanf in unbelievably slow on looong strings // FIXME copy / do not write yading@10: next++; yading@10: } yading@10: e = sscanf(p, " in:%d ", &picture_number); yading@10: yading@10: assert(picture_number >= 0); yading@10: assert(picture_number < rcc->num_entries); yading@10: rce = &rcc->entry[picture_number]; yading@10: yading@10: 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", yading@10: &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits, yading@10: &rce->mv_bits, &rce->misc_bits, yading@10: &rce->f_code, &rce->b_code, yading@10: &rce->mc_mb_var_sum, &rce->mb_var_sum, yading@10: &rce->i_count, &rce->skip_count, &rce->header_bits); yading@10: if (e != 14) { yading@10: av_log(s->avctx, AV_LOG_ERROR, yading@10: "statistics are damaged at line %d, parser out=%d\n", yading@10: i, e); yading@10: return -1; yading@10: } yading@10: yading@10: p = next; yading@10: } yading@10: yading@10: if (init_pass2(s) < 0) yading@10: return -1; yading@10: yading@10: // FIXME maybe move to end yading@10: if ((s->flags & CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID) { yading@10: #if CONFIG_LIBXVID yading@10: return ff_xvid_rate_control_init(s); yading@10: #else yading@10: av_log(s->avctx, AV_LOG_ERROR, yading@10: "Xvid ratecontrol requires libavcodec compiled with Xvid support.\n"); yading@10: return -1; yading@10: #endif yading@10: } yading@10: } yading@10: yading@10: if (!(s->flags & CODEC_FLAG_PASS2)) { yading@10: rcc->short_term_qsum = 0.001; yading@10: rcc->short_term_qcount = 0.001; yading@10: yading@10: rcc->pass1_rc_eq_output_sum = 0.001; yading@10: rcc->pass1_wanted_bits = 0.001; yading@10: yading@10: if (s->avctx->qblur > 1.0) { yading@10: av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n"); yading@10: return -1; yading@10: } yading@10: /* init stuff with the user specified complexity */ yading@10: if (s->avctx->rc_initial_cplx) { yading@10: for (i = 0; i < 60 * 30; i++) { yading@10: double bits = s->avctx->rc_initial_cplx * (i / 10000.0 + 1.0) * s->mb_num; yading@10: RateControlEntry rce; yading@10: yading@10: if (i % ((s->gop_size + 3) / 4) == 0) yading@10: rce.pict_type = AV_PICTURE_TYPE_I; yading@10: else if (i % (s->max_b_frames + 1)) yading@10: rce.pict_type = AV_PICTURE_TYPE_B; yading@10: else yading@10: rce.pict_type = AV_PICTURE_TYPE_P; yading@10: yading@10: rce.new_pict_type = rce.pict_type; yading@10: rce.mc_mb_var_sum = bits * s->mb_num / 100000; yading@10: rce.mb_var_sum = s->mb_num; yading@10: yading@10: rce.qscale = FF_QP2LAMBDA * 2; yading@10: rce.f_code = 2; yading@10: rce.b_code = 1; yading@10: rce.misc_bits = 1; yading@10: yading@10: if (s->pict_type == AV_PICTURE_TYPE_I) { yading@10: rce.i_count = s->mb_num; yading@10: rce.i_tex_bits = bits; yading@10: rce.p_tex_bits = 0; yading@10: rce.mv_bits = 0; yading@10: } else { yading@10: rce.i_count = 0; // FIXME we do know this approx yading@10: rce.i_tex_bits = 0; yading@10: rce.p_tex_bits = bits * 0.9; yading@10: rce.mv_bits = bits * 0.1; yading@10: } yading@10: rcc->i_cplx_sum[rce.pict_type] += rce.i_tex_bits * rce.qscale; yading@10: rcc->p_cplx_sum[rce.pict_type] += rce.p_tex_bits * rce.qscale; yading@10: rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits; yading@10: rcc->frame_count[rce.pict_type]++; yading@10: yading@10: get_qscale(s, &rce, rcc->pass1_wanted_bits / rcc->pass1_rc_eq_output_sum, i); yading@10: yading@10: // FIXME misbehaves a little for variable fps yading@10: rcc->pass1_wanted_bits += s->bit_rate / get_fps(s->avctx); yading@10: } yading@10: } yading@10: } yading@10: yading@10: return 0; yading@10: } yading@10: yading@10: void ff_rate_control_uninit(MpegEncContext *s) yading@10: { yading@10: RateControlContext *rcc = &s->rc_context; yading@10: emms_c(); yading@10: yading@10: av_expr_free(rcc->rc_eq_eval); yading@10: av_freep(&rcc->entry); yading@10: yading@10: #if CONFIG_LIBXVID yading@10: if ((s->flags & CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID) yading@10: ff_xvid_rate_control_uninit(s); yading@10: #endif yading@10: } yading@10: yading@10: int ff_vbv_update(MpegEncContext *s, int frame_size) yading@10: { yading@10: RateControlContext *rcc = &s->rc_context; yading@10: const double fps = get_fps(s->avctx); yading@10: const int buffer_size = s->avctx->rc_buffer_size; yading@10: const double min_rate = s->avctx->rc_min_rate / fps; yading@10: const double max_rate = s->avctx->rc_max_rate / fps; yading@10: yading@10: av_dlog(s, "%d %f %d %f %f\n", yading@10: buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate); yading@10: yading@10: if (buffer_size) { yading@10: int left; yading@10: yading@10: rcc->buffer_index -= frame_size; yading@10: if (rcc->buffer_index < 0) { yading@10: av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n"); yading@10: rcc->buffer_index = 0; yading@10: } yading@10: yading@10: left = buffer_size - rcc->buffer_index - 1; yading@10: rcc->buffer_index += av_clip(left, min_rate, max_rate); yading@10: yading@10: if (rcc->buffer_index > buffer_size) { yading@10: int stuffing = ceil((rcc->buffer_index - buffer_size) / 8); yading@10: yading@10: if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4) yading@10: stuffing = 4; yading@10: rcc->buffer_index -= 8 * stuffing; yading@10: yading@10: if (s->avctx->debug & FF_DEBUG_RC) yading@10: av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing); yading@10: yading@10: return stuffing; yading@10: } yading@10: } yading@10: return 0; yading@10: } yading@10: yading@10: /** yading@10: * Modify the bitrate curve from pass1 for one frame. yading@10: */ yading@10: static double get_qscale(MpegEncContext *s, RateControlEntry *rce, yading@10: double rate_factor, int frame_num) yading@10: { yading@10: RateControlContext *rcc = &s->rc_context; yading@10: AVCodecContext *a = s->avctx; yading@10: const int pict_type = rce->new_pict_type; yading@10: const double mb_num = s->mb_num; yading@10: double q, bits; yading@10: int i; yading@10: yading@10: double const_values[] = { yading@10: M_PI, yading@10: M_E, yading@10: rce->i_tex_bits * rce->qscale, yading@10: rce->p_tex_bits * rce->qscale, yading@10: (rce->i_tex_bits + rce->p_tex_bits) * (double)rce->qscale, yading@10: rce->mv_bits / mb_num, yading@10: rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code) * 0.5 : rce->f_code, yading@10: rce->i_count / mb_num, yading@10: rce->mc_mb_var_sum / mb_num, yading@10: rce->mb_var_sum / mb_num, yading@10: rce->pict_type == AV_PICTURE_TYPE_I, yading@10: rce->pict_type == AV_PICTURE_TYPE_P, yading@10: rce->pict_type == AV_PICTURE_TYPE_B, yading@10: rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type], yading@10: a->qcompress, yading@10: #if 0 yading@10: rcc->last_qscale_for[AV_PICTURE_TYPE_I], yading@10: rcc->last_qscale_for[AV_PICTURE_TYPE_P], yading@10: rcc->last_qscale_for[AV_PICTURE_TYPE_B], yading@10: rcc->next_non_b_qscale, yading@10: #endif yading@10: rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I], yading@10: rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P], yading@10: rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P], yading@10: rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B], yading@10: (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type], yading@10: 0 yading@10: }; yading@10: yading@10: bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce); yading@10: if (isnan(bits)) { yading@10: av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->avctx->rc_eq); yading@10: return -1; yading@10: } yading@10: yading@10: rcc->pass1_rc_eq_output_sum += bits; yading@10: bits *= rate_factor; yading@10: if (bits < 0.0) yading@10: bits = 0.0; yading@10: bits += 1.0; // avoid 1/0 issues yading@10: yading@10: /* user override */ yading@10: for (i = 0; i < s->avctx->rc_override_count; i++) { yading@10: RcOverride *rco = s->avctx->rc_override; yading@10: if (rco[i].start_frame > frame_num) yading@10: continue; yading@10: if (rco[i].end_frame < frame_num) yading@10: continue; yading@10: yading@10: if (rco[i].qscale) yading@10: bits = qp2bits(rce, rco[i].qscale); // FIXME move at end to really force it? yading@10: else yading@10: bits *= rco[i].quality_factor; yading@10: } yading@10: yading@10: q = bits2qp(rce, bits); yading@10: yading@10: /* I/B difference */ yading@10: if (pict_type == AV_PICTURE_TYPE_I && s->avctx->i_quant_factor < 0.0) yading@10: q = -q * s->avctx->i_quant_factor + s->avctx->i_quant_offset; yading@10: else if (pict_type == AV_PICTURE_TYPE_B && s->avctx->b_quant_factor < 0.0) yading@10: q = -q * s->avctx->b_quant_factor + s->avctx->b_quant_offset; yading@10: if (q < 1) yading@10: q = 1; yading@10: yading@10: return q; yading@10: } yading@10: yading@10: static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q) yading@10: { yading@10: RateControlContext *rcc = &s->rc_context; yading@10: AVCodecContext *a = s->avctx; yading@10: const int pict_type = rce->new_pict_type; yading@10: const double last_p_q = rcc->last_qscale_for[AV_PICTURE_TYPE_P]; yading@10: const double last_non_b_q = rcc->last_qscale_for[rcc->last_non_b_pict_type]; yading@10: yading@10: if (pict_type == AV_PICTURE_TYPE_I && yading@10: (a->i_quant_factor > 0.0 || rcc->last_non_b_pict_type == AV_PICTURE_TYPE_P)) yading@10: q = last_p_q * FFABS(a->i_quant_factor) + a->i_quant_offset; yading@10: else if (pict_type == AV_PICTURE_TYPE_B && yading@10: a->b_quant_factor > 0.0) yading@10: q = last_non_b_q * a->b_quant_factor + a->b_quant_offset; yading@10: if (q < 1) yading@10: q = 1; yading@10: yading@10: /* last qscale / qdiff stuff */ yading@10: if (rcc->last_non_b_pict_type == pict_type || pict_type != AV_PICTURE_TYPE_I) { yading@10: double last_q = rcc->last_qscale_for[pict_type]; yading@10: const int maxdiff = FF_QP2LAMBDA * a->max_qdiff; yading@10: yading@10: if (q > last_q + maxdiff) yading@10: q = last_q + maxdiff; yading@10: else if (q < last_q - maxdiff) yading@10: q = last_q - maxdiff; yading@10: } yading@10: yading@10: rcc->last_qscale_for[pict_type] = q; // Note we cannot do that after blurring yading@10: yading@10: if (pict_type != AV_PICTURE_TYPE_B) yading@10: rcc->last_non_b_pict_type = pict_type; yading@10: yading@10: return q; yading@10: } yading@10: yading@10: /** yading@10: * Get the qmin & qmax for pict_type. yading@10: */ yading@10: static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type) yading@10: { yading@10: int qmin = s->avctx->lmin; yading@10: int qmax = s->avctx->lmax; yading@10: yading@10: assert(qmin <= qmax); yading@10: yading@10: switch (pict_type) { yading@10: case AV_PICTURE_TYPE_B: yading@10: qmin = (int)(qmin * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5); yading@10: qmax = (int)(qmax * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5); yading@10: break; yading@10: case AV_PICTURE_TYPE_I: yading@10: qmin = (int)(qmin * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5); yading@10: qmax = (int)(qmax * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5); yading@10: break; yading@10: } yading@10: yading@10: qmin = av_clip(qmin, 1, FF_LAMBDA_MAX); yading@10: qmax = av_clip(qmax, 1, FF_LAMBDA_MAX); yading@10: yading@10: if (qmax < qmin) yading@10: qmax = qmin; yading@10: yading@10: *qmin_ret = qmin; yading@10: *qmax_ret = qmax; yading@10: } yading@10: yading@10: static double modify_qscale(MpegEncContext *s, RateControlEntry *rce, yading@10: double q, int frame_num) yading@10: { yading@10: RateControlContext *rcc = &s->rc_context; yading@10: const double buffer_size = s->avctx->rc_buffer_size; yading@10: const double fps = get_fps(s->avctx); yading@10: const double min_rate = s->avctx->rc_min_rate / fps; yading@10: const double max_rate = s->avctx->rc_max_rate / fps; yading@10: const int pict_type = rce->new_pict_type; yading@10: int qmin, qmax; yading@10: yading@10: get_qminmax(&qmin, &qmax, s, pict_type); yading@10: yading@10: /* modulation */ yading@10: if (s->avctx->rc_qmod_freq && yading@10: frame_num % s->avctx->rc_qmod_freq == 0 && yading@10: pict_type == AV_PICTURE_TYPE_P) yading@10: q *= s->avctx->rc_qmod_amp; yading@10: yading@10: /* buffer overflow/underflow protection */ yading@10: if (buffer_size) { yading@10: double expected_size = rcc->buffer_index; yading@10: double q_limit; yading@10: yading@10: if (min_rate) { yading@10: double d = 2 * (buffer_size - expected_size) / buffer_size; yading@10: if (d > 1.0) yading@10: d = 1.0; yading@10: else if (d < 0.0001) yading@10: d = 0.0001; yading@10: q *= pow(d, 1.0 / s->avctx->rc_buffer_aggressivity); yading@10: yading@10: q_limit = bits2qp(rce, yading@10: FFMAX((min_rate - buffer_size + rcc->buffer_index) * yading@10: s->avctx->rc_min_vbv_overflow_use, 1)); yading@10: yading@10: if (q > q_limit) { yading@10: if (s->avctx->debug & FF_DEBUG_RC) yading@10: av_log(s->avctx, AV_LOG_DEBUG, yading@10: "limiting QP %f -> %f\n", q, q_limit); yading@10: q = q_limit; yading@10: } yading@10: } yading@10: yading@10: if (max_rate) { yading@10: double d = 2 * expected_size / buffer_size; yading@10: if (d > 1.0) yading@10: d = 1.0; yading@10: else if (d < 0.0001) yading@10: d = 0.0001; yading@10: q /= pow(d, 1.0 / s->avctx->rc_buffer_aggressivity); yading@10: yading@10: q_limit = bits2qp(rce, yading@10: FFMAX(rcc->buffer_index * yading@10: s->avctx->rc_max_available_vbv_use, yading@10: 1)); yading@10: if (q < q_limit) { yading@10: if (s->avctx->debug & FF_DEBUG_RC) yading@10: av_log(s->avctx, AV_LOG_DEBUG, yading@10: "limiting QP %f -> %f\n", q, q_limit); yading@10: q = q_limit; yading@10: } yading@10: } yading@10: } yading@10: av_dlog(s, "q:%f max:%f min:%f size:%f index:%f agr:%f\n", yading@10: q, max_rate, min_rate, buffer_size, rcc->buffer_index, yading@10: s->avctx->rc_buffer_aggressivity); yading@10: if (s->avctx->rc_qsquish == 0.0 || qmin == qmax) { yading@10: if (q < qmin) yading@10: q = qmin; yading@10: else if (q > qmax) yading@10: q = qmax; yading@10: } else { yading@10: double min2 = log(qmin); yading@10: double max2 = log(qmax); yading@10: yading@10: q = log(q); yading@10: q = (q - min2) / (max2 - min2) - 0.5; yading@10: q *= -4.0; yading@10: q = 1.0 / (1.0 + exp(q)); yading@10: q = q * (max2 - min2) + min2; yading@10: yading@10: q = exp(q); yading@10: } yading@10: yading@10: return q; yading@10: } yading@10: yading@10: // ---------------------------------- yading@10: // 1 Pass Code yading@10: yading@10: static double predict_size(Predictor *p, double q, double var) yading@10: { yading@10: return p->coeff * var / (q * p->count); yading@10: } yading@10: yading@10: static void update_predictor(Predictor *p, double q, double var, double size) yading@10: { yading@10: double new_coeff = size * q / (var + 1); yading@10: if (var < 10) yading@10: return; yading@10: yading@10: p->count *= p->decay; yading@10: p->coeff *= p->decay; yading@10: p->count++; yading@10: p->coeff += new_coeff; yading@10: } yading@10: yading@10: static void adaptive_quantization(MpegEncContext *s, double q) yading@10: { yading@10: int i; yading@10: const float lumi_masking = s->avctx->lumi_masking / (128.0 * 128.0); yading@10: const float dark_masking = s->avctx->dark_masking / (128.0 * 128.0); yading@10: const float temp_cplx_masking = s->avctx->temporal_cplx_masking; yading@10: const float spatial_cplx_masking = s->avctx->spatial_cplx_masking; yading@10: const float p_masking = s->avctx->p_masking; yading@10: const float border_masking = s->avctx->border_masking; yading@10: float bits_sum = 0.0; yading@10: float cplx_sum = 0.0; yading@10: float *cplx_tab = s->cplx_tab; yading@10: float *bits_tab = s->bits_tab; yading@10: const int qmin = s->avctx->mb_lmin; yading@10: const int qmax = s->avctx->mb_lmax; yading@10: Picture *const pic = &s->current_picture; yading@10: const int mb_width = s->mb_width; yading@10: const int mb_height = s->mb_height; yading@10: yading@10: for (i = 0; i < s->mb_num; i++) { yading@10: const int mb_xy = s->mb_index2xy[i]; yading@10: float temp_cplx = sqrt(pic->mc_mb_var[mb_xy]); // FIXME merge in pow() yading@10: float spat_cplx = sqrt(pic->mb_var[mb_xy]); yading@10: const int lumi = pic->mb_mean[mb_xy]; yading@10: float bits, cplx, factor; yading@10: int mb_x = mb_xy % s->mb_stride; yading@10: int mb_y = mb_xy / s->mb_stride; yading@10: int mb_distance; yading@10: float mb_factor = 0.0; yading@10: if (spat_cplx < 4) yading@10: spat_cplx = 4; // FIXME finetune yading@10: if (temp_cplx < 4) yading@10: temp_cplx = 4; // FIXME finetune yading@10: yading@10: if ((s->mb_type[mb_xy] & CANDIDATE_MB_TYPE_INTRA)) { // FIXME hq mode yading@10: cplx = spat_cplx; yading@10: factor = 1.0 + p_masking; yading@10: } else { yading@10: cplx = temp_cplx; yading@10: factor = pow(temp_cplx, -temp_cplx_masking); yading@10: } yading@10: factor *= pow(spat_cplx, -spatial_cplx_masking); yading@10: yading@10: if (lumi > 127) yading@10: factor *= (1.0 - (lumi - 128) * (lumi - 128) * lumi_masking); yading@10: else yading@10: factor *= (1.0 - (lumi - 128) * (lumi - 128) * dark_masking); yading@10: yading@10: if (mb_x < mb_width / 5) { yading@10: mb_distance = mb_width / 5 - mb_x; yading@10: mb_factor = (float)mb_distance / (float)(mb_width / 5); yading@10: } else if (mb_x > 4 * mb_width / 5) { yading@10: mb_distance = mb_x - 4 * mb_width / 5; yading@10: mb_factor = (float)mb_distance / (float)(mb_width / 5); yading@10: } yading@10: if (mb_y < mb_height / 5) { yading@10: mb_distance = mb_height / 5 - mb_y; yading@10: mb_factor = FFMAX(mb_factor, yading@10: (float)mb_distance / (float)(mb_height / 5)); yading@10: } else if (mb_y > 4 * mb_height / 5) { yading@10: mb_distance = mb_y - 4 * mb_height / 5; yading@10: mb_factor = FFMAX(mb_factor, yading@10: (float)mb_distance / (float)(mb_height / 5)); yading@10: } yading@10: yading@10: factor *= 1.0 - border_masking * mb_factor; yading@10: yading@10: if (factor < 0.00001) yading@10: factor = 0.00001; yading@10: yading@10: bits = cplx * factor; yading@10: cplx_sum += cplx; yading@10: bits_sum += bits; yading@10: cplx_tab[i] = cplx; yading@10: bits_tab[i] = bits; yading@10: } yading@10: yading@10: /* handle qmin/qmax clipping */ yading@10: if (s->flags & CODEC_FLAG_NORMALIZE_AQP) { yading@10: float factor = bits_sum / cplx_sum; yading@10: for (i = 0; i < s->mb_num; i++) { yading@10: float newq = q * cplx_tab[i] / bits_tab[i]; yading@10: newq *= factor; yading@10: yading@10: if (newq > qmax) { yading@10: bits_sum -= bits_tab[i]; yading@10: cplx_sum -= cplx_tab[i] * q / qmax; yading@10: } else if (newq < qmin) { yading@10: bits_sum -= bits_tab[i]; yading@10: cplx_sum -= cplx_tab[i] * q / qmin; yading@10: } yading@10: } yading@10: if (bits_sum < 0.001) yading@10: bits_sum = 0.001; yading@10: if (cplx_sum < 0.001) yading@10: cplx_sum = 0.001; yading@10: } yading@10: yading@10: for (i = 0; i < s->mb_num; i++) { yading@10: const int mb_xy = s->mb_index2xy[i]; yading@10: float newq = q * cplx_tab[i] / bits_tab[i]; yading@10: int intq; yading@10: yading@10: if (s->flags & CODEC_FLAG_NORMALIZE_AQP) { yading@10: newq *= bits_sum / cplx_sum; yading@10: } yading@10: yading@10: intq = (int)(newq + 0.5); yading@10: yading@10: if (intq > qmax) yading@10: intq = qmax; yading@10: else if (intq < qmin) yading@10: intq = qmin; yading@10: s->lambda_table[mb_xy] = intq; yading@10: } yading@10: } yading@10: yading@10: void ff_get_2pass_fcode(MpegEncContext *s) yading@10: { yading@10: RateControlContext *rcc = &s->rc_context; yading@10: RateControlEntry *rce = &rcc->entry[s->picture_number]; yading@10: yading@10: s->f_code = rce->f_code; yading@10: s->b_code = rce->b_code; yading@10: } yading@10: yading@10: // FIXME rd or at least approx for dquant yading@10: yading@10: float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run) yading@10: { yading@10: float q; yading@10: int qmin, qmax; yading@10: float br_compensation; yading@10: double diff; yading@10: double short_term_q; yading@10: double fps; yading@10: int picture_number = s->picture_number; yading@10: int64_t wanted_bits; yading@10: RateControlContext *rcc = &s->rc_context; yading@10: AVCodecContext *a = s->avctx; yading@10: RateControlEntry local_rce, *rce; yading@10: double bits; yading@10: double rate_factor; yading@10: int var; yading@10: const int pict_type = s->pict_type; yading@10: Picture * const pic = &s->current_picture; yading@10: emms_c(); yading@10: yading@10: #if CONFIG_LIBXVID yading@10: if ((s->flags & CODEC_FLAG_PASS2) && yading@10: s->avctx->rc_strategy == FF_RC_STRATEGY_XVID) yading@10: return ff_xvid_rate_estimate_qscale(s, dry_run); yading@10: #endif yading@10: yading@10: get_qminmax(&qmin, &qmax, s, pict_type); yading@10: yading@10: fps = get_fps(s->avctx); yading@10: /* update predictors */ yading@10: if (picture_number > 2 && !dry_run) { yading@10: const int last_var = s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum yading@10: : rcc->last_mc_mb_var_sum; yading@10: av_assert1(s->frame_bits >= s->stuffing_bits); yading@10: update_predictor(&rcc->pred[s->last_pict_type], yading@10: rcc->last_qscale, yading@10: sqrt(last_var), yading@10: s->frame_bits - s->stuffing_bits); yading@10: } yading@10: yading@10: if (s->flags & CODEC_FLAG_PASS2) { yading@10: assert(picture_number >= 0); yading@10: if (picture_number >= rcc->num_entries) { yading@10: av_log(s, AV_LOG_ERROR, "Input is longer than 2-pass log file\n"); yading@10: return -1; yading@10: } yading@10: rce = &rcc->entry[picture_number]; yading@10: wanted_bits = rce->expected_bits; yading@10: } else { yading@10: Picture *dts_pic; yading@10: rce = &local_rce; yading@10: yading@10: /* FIXME add a dts field to AVFrame and ensure it is set and use it yading@10: * here instead of reordering but the reordering is simpler for now yading@10: * until H.264 B-pyramid must be handled. */ yading@10: if (s->pict_type == AV_PICTURE_TYPE_B || s->low_delay) yading@10: dts_pic = s->current_picture_ptr; yading@10: else yading@10: dts_pic = s->last_picture_ptr; yading@10: yading@10: if (!dts_pic || dts_pic->f.pts == AV_NOPTS_VALUE) yading@10: wanted_bits = (uint64_t)(s->bit_rate * (double)picture_number / fps); yading@10: else yading@10: wanted_bits = (uint64_t)(s->bit_rate * (double)dts_pic->f.pts / fps); yading@10: } yading@10: yading@10: diff = s->total_bits - wanted_bits; yading@10: br_compensation = (a->bit_rate_tolerance - diff) / a->bit_rate_tolerance; yading@10: if (br_compensation <= 0.0) yading@10: br_compensation = 0.001; yading@10: yading@10: var = pict_type == AV_PICTURE_TYPE_I ? pic->mb_var_sum : pic->mc_mb_var_sum; yading@10: yading@10: short_term_q = 0; /* avoid warning */ yading@10: if (s->flags & CODEC_FLAG_PASS2) { yading@10: if (pict_type != AV_PICTURE_TYPE_I) yading@10: assert(pict_type == rce->new_pict_type); yading@10: yading@10: q = rce->new_qscale / br_compensation; yading@10: av_dlog(s, "%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale, yading@10: br_compensation, s->frame_bits, var, pict_type); yading@10: } else { yading@10: rce->pict_type = yading@10: rce->new_pict_type = pict_type; yading@10: rce->mc_mb_var_sum = pic->mc_mb_var_sum; yading@10: rce->mb_var_sum = pic->mb_var_sum; yading@10: rce->qscale = FF_QP2LAMBDA * 2; yading@10: rce->f_code = s->f_code; yading@10: rce->b_code = s->b_code; yading@10: rce->misc_bits = 1; yading@10: yading@10: bits = predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var)); yading@10: if (pict_type == AV_PICTURE_TYPE_I) { yading@10: rce->i_count = s->mb_num; yading@10: rce->i_tex_bits = bits; yading@10: rce->p_tex_bits = 0; yading@10: rce->mv_bits = 0; yading@10: } else { yading@10: rce->i_count = 0; // FIXME we do know this approx yading@10: rce->i_tex_bits = 0; yading@10: rce->p_tex_bits = bits * 0.9; yading@10: rce->mv_bits = bits * 0.1; yading@10: } yading@10: rcc->i_cplx_sum[pict_type] += rce->i_tex_bits * rce->qscale; yading@10: rcc->p_cplx_sum[pict_type] += rce->p_tex_bits * rce->qscale; yading@10: rcc->mv_bits_sum[pict_type] += rce->mv_bits; yading@10: rcc->frame_count[pict_type]++; yading@10: yading@10: bits = rce->i_tex_bits + rce->p_tex_bits; yading@10: rate_factor = rcc->pass1_wanted_bits / yading@10: rcc->pass1_rc_eq_output_sum * br_compensation; yading@10: yading@10: q = get_qscale(s, rce, rate_factor, picture_number); yading@10: if (q < 0) yading@10: return -1; yading@10: yading@10: assert(q > 0.0); yading@10: q = get_diff_limited_q(s, rce, q); yading@10: assert(q > 0.0); yading@10: yading@10: // FIXME type dependent blur like in 2-pass yading@10: if (pict_type == AV_PICTURE_TYPE_P || s->intra_only) { yading@10: rcc->short_term_qsum *= a->qblur; yading@10: rcc->short_term_qcount *= a->qblur; yading@10: yading@10: rcc->short_term_qsum += q; yading@10: rcc->short_term_qcount++; yading@10: q = short_term_q = rcc->short_term_qsum / rcc->short_term_qcount; yading@10: } yading@10: assert(q > 0.0); yading@10: yading@10: q = modify_qscale(s, rce, q, picture_number); yading@10: yading@10: rcc->pass1_wanted_bits += s->bit_rate / fps; yading@10: yading@10: assert(q > 0.0); yading@10: } yading@10: yading@10: if (s->avctx->debug & FF_DEBUG_RC) { yading@10: av_log(s->avctx, AV_LOG_DEBUG, yading@10: "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f " yading@10: "size:%d var:%d/%d br:%d fps:%d\n", yading@10: av_get_picture_type_char(pict_type), yading@10: qmin, q, qmax, picture_number, yading@10: (int)wanted_bits / 1000, (int)s->total_bits / 1000, yading@10: br_compensation, short_term_q, s->frame_bits, yading@10: pic->mb_var_sum, pic->mc_mb_var_sum, yading@10: s->bit_rate / 1000, (int)fps); yading@10: } yading@10: yading@10: if (q < qmin) yading@10: q = qmin; yading@10: else if (q > qmax) yading@10: q = qmax; yading@10: yading@10: if (s->adaptive_quant) yading@10: adaptive_quantization(s, q); yading@10: else yading@10: q = (int)(q + 0.5); yading@10: yading@10: if (!dry_run) { yading@10: rcc->last_qscale = q; yading@10: rcc->last_mc_mb_var_sum = pic->mc_mb_var_sum; yading@10: rcc->last_mb_var_sum = pic->mb_var_sum; yading@10: } yading@10: return q; yading@10: } yading@10: yading@10: // ---------------------------------------------- yading@10: // 2-Pass code yading@10: yading@10: static int init_pass2(MpegEncContext *s) yading@10: { yading@10: RateControlContext *rcc = &s->rc_context; yading@10: AVCodecContext *a = s->avctx; yading@10: int i, toobig; yading@10: double fps = get_fps(s->avctx); yading@10: double complexity[5] = { 0 }; // approximate bits at quant=1 yading@10: uint64_t const_bits[5] = { 0 }; // quantizer independent bits yading@10: uint64_t all_const_bits; yading@10: uint64_t all_available_bits = (uint64_t)(s->bit_rate * yading@10: (double)rcc->num_entries / fps); yading@10: double rate_factor = 0; yading@10: double step; yading@10: const int filter_size = (int)(a->qblur * 4) | 1; yading@10: double expected_bits; yading@10: double *qscale, *blurred_qscale, qscale_sum; yading@10: yading@10: /* find complexity & const_bits & decide the pict_types */ yading@10: for (i = 0; i < rcc->num_entries; i++) { yading@10: RateControlEntry *rce = &rcc->entry[i]; yading@10: yading@10: rce->new_pict_type = rce->pict_type; yading@10: rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale; yading@10: rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale; yading@10: rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits; yading@10: rcc->frame_count[rce->pict_type]++; yading@10: yading@10: complexity[rce->new_pict_type] += (rce->i_tex_bits + rce->p_tex_bits) * yading@10: (double)rce->qscale; yading@10: const_bits[rce->new_pict_type] += rce->mv_bits + rce->misc_bits; yading@10: } yading@10: yading@10: all_const_bits = const_bits[AV_PICTURE_TYPE_I] + yading@10: const_bits[AV_PICTURE_TYPE_P] + yading@10: const_bits[AV_PICTURE_TYPE_B]; yading@10: yading@10: if (all_available_bits < all_const_bits) { yading@10: av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n"); yading@10: return -1; yading@10: } yading@10: yading@10: qscale = av_malloc(sizeof(double) * rcc->num_entries); yading@10: blurred_qscale = av_malloc(sizeof(double) * rcc->num_entries); yading@10: toobig = 0; yading@10: yading@10: for (step = 256 * 256; step > 0.0000001; step *= 0.5) { yading@10: expected_bits = 0; yading@10: rate_factor += step; yading@10: yading@10: rcc->buffer_index = s->avctx->rc_buffer_size / 2; yading@10: yading@10: /* find qscale */ yading@10: for (i = 0; i < rcc->num_entries; i++) { yading@10: RateControlEntry *rce = &rcc->entry[i]; yading@10: yading@10: qscale[i] = get_qscale(s, &rcc->entry[i], rate_factor, i); yading@10: rcc->last_qscale_for[rce->pict_type] = qscale[i]; yading@10: } yading@10: assert(filter_size % 2 == 1); yading@10: yading@10: /* fixed I/B QP relative to P mode */ yading@10: for (i = FFMAX(0, rcc->num_entries - 300); i < rcc->num_entries; i++) { yading@10: RateControlEntry *rce = &rcc->entry[i]; yading@10: yading@10: qscale[i] = get_diff_limited_q(s, rce, qscale[i]); yading@10: } yading@10: yading@10: for (i = rcc->num_entries - 1; i >= 0; i--) { yading@10: RateControlEntry *rce = &rcc->entry[i]; yading@10: yading@10: qscale[i] = get_diff_limited_q(s, rce, qscale[i]); yading@10: } yading@10: yading@10: /* smooth curve */ yading@10: for (i = 0; i < rcc->num_entries; i++) { yading@10: RateControlEntry *rce = &rcc->entry[i]; yading@10: const int pict_type = rce->new_pict_type; yading@10: int j; yading@10: double q = 0.0, sum = 0.0; yading@10: yading@10: for (j = 0; j < filter_size; j++) { yading@10: int index = i + j - filter_size / 2; yading@10: double d = index - i; yading@10: double coeff = a->qblur == 0 ? 1.0 : exp(-d * d / (a->qblur * a->qblur)); yading@10: yading@10: if (index < 0 || index >= rcc->num_entries) yading@10: continue; yading@10: if (pict_type != rcc->entry[index].new_pict_type) yading@10: continue; yading@10: q += qscale[index] * coeff; yading@10: sum += coeff; yading@10: } yading@10: blurred_qscale[i] = q / sum; yading@10: } yading@10: yading@10: /* find expected bits */ yading@10: for (i = 0; i < rcc->num_entries; i++) { yading@10: RateControlEntry *rce = &rcc->entry[i]; yading@10: double bits; yading@10: yading@10: rce->new_qscale = modify_qscale(s, rce, blurred_qscale[i], i); yading@10: yading@10: bits = qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits; yading@10: bits += 8 * ff_vbv_update(s, bits); yading@10: yading@10: rce->expected_bits = expected_bits; yading@10: expected_bits += bits; yading@10: } yading@10: yading@10: av_dlog(s->avctx, yading@10: "expected_bits: %f all_available_bits: %d rate_factor: %f\n", yading@10: expected_bits, (int)all_available_bits, rate_factor); yading@10: if (expected_bits > all_available_bits) { yading@10: rate_factor -= step; yading@10: ++toobig; yading@10: } yading@10: } yading@10: av_free(qscale); yading@10: av_free(blurred_qscale); yading@10: yading@10: /* check bitrate calculations and print info */ yading@10: qscale_sum = 0.0; yading@10: for (i = 0; i < rcc->num_entries; i++) { yading@10: av_dlog(s, "[lavc rc] entry[%d].new_qscale = %.3f qp = %.3f\n", yading@10: i, yading@10: rcc->entry[i].new_qscale, yading@10: rcc->entry[i].new_qscale / FF_QP2LAMBDA); yading@10: qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA, yading@10: s->avctx->qmin, s->avctx->qmax); yading@10: } yading@10: assert(toobig <= 40); yading@10: av_log(s->avctx, AV_LOG_DEBUG, yading@10: "[lavc rc] requested bitrate: %d bps expected bitrate: %d bps\n", yading@10: s->bit_rate, yading@10: (int)(expected_bits / ((double)all_available_bits / s->bit_rate))); yading@10: av_log(s->avctx, AV_LOG_DEBUG, yading@10: "[lavc rc] estimated target average qp: %.3f\n", yading@10: (float)qscale_sum / rcc->num_entries); yading@10: if (toobig == 0) { yading@10: av_log(s->avctx, AV_LOG_INFO, yading@10: "[lavc rc] Using all of requested bitrate is not " yading@10: "necessary for this video with these parameters.\n"); yading@10: } else if (toobig == 40) { yading@10: av_log(s->avctx, AV_LOG_ERROR, yading@10: "[lavc rc] Error: bitrate too low for this video " yading@10: "with these parameters.\n"); yading@10: return -1; yading@10: } else if (fabs(expected_bits / all_available_bits - 1.0) > 0.01) { yading@10: av_log(s->avctx, AV_LOG_ERROR, yading@10: "[lavc rc] Error: 2pass curve failed to converge\n"); yading@10: return -1; yading@10: } yading@10: yading@10: return 0; yading@10: }