Merge commit 'a1f6a2dfdaf9beb42ca66e49d10bfaf5905a0128'
authorClément Bœsch <u@pkh.me>
Wed, 1 Feb 2017 13:47:14 +0000 (14:47 +0100)
committerClément Bœsch <u@pkh.me>
Wed, 1 Feb 2017 13:50:21 +0000 (14:50 +0100)
* commit 'a1f6a2dfdaf9beb42ca66e49d10bfaf5905a0128':
  ratecontrol: Reorder functions to avoid forward declarations

Merged, but this seems to break the clear separation of 1-pass vs
2-pass.

Merged-by: Clément Bœsch <u@pkh.me>
1  2 
libavcodec/ratecontrol.c

  #include "mpegvideo.h"
  #include "libavutil/eval.h"
  
- static int init_pass2(MpegEncContext *s);
- static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
-                          double rate_factor, int frame_num);
 -#undef NDEBUG // Always check asserts, the speed effect is far too small to disable them.
 -#include <assert.h>
 +void ff_write_pass1_stats(MpegEncContext *s)
 +{
 +    snprintf(s->avctx->stats_out, 256,
 +             "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d "
 +             "fcode:%d bcode:%d mc-var:%"PRId64" var:%"PRId64" icount:%d skipcount:%d hbits:%d;\n",
 +             s->current_picture_ptr->f->display_picture_number,
 +             s->current_picture_ptr->f->coded_picture_number,
 +             s->pict_type,
 +             s->current_picture.f->quality,
 +             s->i_tex_bits,
 +             s->p_tex_bits,
 +             s->mv_bits,
 +             s->misc_bits,
 +             s->f_code,
 +             s->b_code,
 +             s->current_picture.mc_mb_var_sum,
 +             s->current_picture.mb_var_sum,
 +             s->i_count, s->skip_count,
 +             s->header_bits);
 +}
  
- static double get_fps(AVCodecContext *avctx)
- {
-     return 1.0 / av_q2d(avctx->time_base) / FFMAX(avctx->ticks_per_frame, 1);
- }
 -#ifndef M_E
 -#define M_E 2.718281828
 -#endif
++static double get_fps(AVCodecContext *avctx)
++{
++    return 1.0 / av_q2d(avctx->time_base) / FFMAX(avctx->ticks_per_frame, 1);
++}
+ 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);
++    av_assert0(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 fps         = get_fps(s->avctx);
+     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.
+  */
+ static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
+                          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 fps             = get_fps(s->avctx);
+     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 expected_bits = 0; // init to silence gcc warning
+     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);
++    qscale         = av_malloc_array(rcc->num_entries, sizeof(double));
++    blurred_qscale = av_malloc_array(rcc->num_entries, sizeof(double));
+     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);
++        av_assert0(filter_size % 2 == 1);
+         /* fixed I/B QP relative to P mode */
++        for (i = FFMAX(0, rcc->num_entries - 300); i < rcc->num_entries; i++) {
++            RateControlEntry *rce = &rcc->entry[i];
++
++            qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
++        }
++
+         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;
+         }
  
- static inline double qp2bits(RateControlEntry *rce, double qp)
- {
-     if (qp <= 0.0) {
-         av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
+         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;
+         }
      }
-     return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / qp;
- }
+     av_free(qscale);
+     av_free(blurred_qscale);
  
- static inline double bits2qp(RateControlEntry *rce, double bits)
- {
-     if (bits < 0.9) {
-         av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
+     /* 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);
      }
-     return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / bits;
 -    assert(toobig <= 40);
++    av_assert0(toobig <= 40);
+     av_log(s->avctx, AV_LOG_DEBUG,
 -           "[lavc rc] requested bitrate: %d bps  expected bitrate: %d bps\n",
++           "[lavc rc] requested bitrate: %"PRId64" bps  expected bitrate: %"PRId64" bps\n",
+            s->bit_rate,
 -           (int)(expected_bits / ((double)all_available_bits / s->bit_rate)));
++           (int64_t)(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;
  }
  
  av_cold int ff_rate_control_init(MpegEncContext *s)
@@@ -290,282 -644,45 +678,48 @@@ av_cold void ff_rate_control_uninit(Mpe
      av_freep(&rcc->entry);
  }
  
- int ff_vbv_update(MpegEncContext *s, int frame_size)
- {
-     RateControlContext *rcc = &s->rc_context;
-     const double fps        = get_fps(s->avctx);
-     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;
-     ff_dlog(s, "%d %f %d %f %f\n",
-             buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
-     if (buffer_size) {
-         int left;
-         rcc->buffer_index -= frame_size;
-         if (rcc->buffer_index < 0) {
-             av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
-             if (frame_size > max_rate && s->qscale == s->avctx->qmax) {
-                 av_log(s->avctx, AV_LOG_ERROR, "max bitrate possibly too small or try trellis with large lmax or increase qmax\n");
-             }
-             rcc->buffer_index = 0;
-         }
-         left = buffer_size - rcc->buffer_index - 1;
-         rcc->buffer_index += av_clip(left, min_rate, max_rate);
-         if (rcc->buffer_index > buffer_size) {
-             int stuffing = ceil((rcc->buffer_index - buffer_size) / 8);
-             if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
-                 stuffing = 4;
-             rcc->buffer_index -= 8 * stuffing;
-             if (s->avctx->debug & FF_DEBUG_RC)
-                 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
-             return stuffing;
-         }
-     }
-     return 0;
- }
- /**
-  * Modify the bitrate curve from pass1 for one frame.
-  */
- static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
-                          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 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;
-     av_assert0(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         = get_fps(s->avctx);
-     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);
+ 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 double fps        = get_fps(s->avctx);
+     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;
  
-     /* modulation */
-     if (s->rc_qmod_freq &&
-         frame_num % s->rc_qmod_freq == 0 &&
-         pict_type == AV_PICTURE_TYPE_P)
-         q *= s->rc_qmod_amp;
+     ff_dlog(s, "%d %f %d %f %f\n",
+             buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
  
-     /* 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));
+         int left;
  
-             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;
+         rcc->buffer_index -= frame_size;
+         if (rcc->buffer_index < 0) {
+             av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
++            if (frame_size > max_rate && s->qscale == s->avctx->qmax) {
++                av_log(s->avctx, AV_LOG_ERROR, "max bitrate possibly too small or try trellis with large lmax or increase qmax\n");
 +            }
+             rcc->buffer_index = 0;
          }
  
-         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);
+         left = buffer_size - rcc->buffer_index - 1;
+         rcc->buffer_index += av_clip(left, min_rate, max_rate);
  
-             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);
+         if (rcc->buffer_index > buffer_size) {
+             int stuffing = ceil((rcc->buffer_index - buffer_size) / 8);
  
-         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;
+             if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
+                 stuffing = 4;
+             rcc->buffer_index -= 8 * stuffing;
  
-         q = exp(q);
-     }
+             if (s->avctx->debug & FF_DEBUG_RC)
+                 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
  
-     return q;
+             return stuffing;
+         }
+     }
+     return 0;
  }
  
- // ----------------------------------
- // 1 Pass Code
  static double predict_size(Predictor *p, double q, double var)
  {
      return p->coeff * var / (q * p->count);