VP8: idct_mb optimizations
[ffmpeg.git] / libavcodec / aaccoder.c
index 763d3b9..8063fb6 100644 (file)
@@ -20,7 +20,7 @@
  */
 
 /**
- * @file libavcodec/aaccoder.c
+ * @file
  * AAC coefficients encoder
  */
 
@@ -30,6 +30,7 @@
  * add sane pulse detection
  ***********************************/
 
+#include <float.h>
 #include "avcodec.h"
 #include "put_bits.h"
 #include "aac.h"
@@ -49,7 +50,7 @@ static const uint8_t run_value_bits_short[16] = {
     3, 3, 3, 3, 3, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 9
 };
 
-static const uint8_trun_value_bits[2] = {
+static const uint8_t *run_value_bits[2] = {
     run_value_bits_long, run_value_bits_short
 };
 
@@ -61,39 +62,35 @@ static const uint8_t* run_value_bits[2] = {
  */
 static av_always_inline int quant(float coef, const float Q)
 {
-    return pow(coef * Q, 0.75) + 0.4054;
+    float a = coef * Q;
+    return sqrtf(a * sqrtf(a)) + 0.4054;
 }
 
-static void quantize_bands(int (*out)[2], const float *in, const float *scaled, int size, float Q34, int is_signed, int maxval)
+static void quantize_bands(int *out, const float *in, const float *scaled,
+                           int size, float Q34, int is_signed, int maxval)
 {
     int i;
     double qc;
     for (i = 0; i < size; i++) {
         qc = scaled[i] * Q34;
-        out[i][0] = (int)FFMIN((int)qc, maxval);
-        out[i][1] = (int)FFMIN((int)(qc + 0.4054), maxval);
+        out[i] = (int)FFMIN(qc + 0.4054, (double)maxval);
         if (is_signed && in[i] < 0.0f) {
-            out[i][0] = -out[i][0];
-            out[i][1] = -out[i][1];
+            out[i] = -out[i];
         }
     }
 }
 
-static void abs_pow34_v(float *out, const floatin, const int size)
+static void abs_pow34_v(float *out, const float *in, const int size)
 {
 #ifndef USE_REALLY_FULL_SEARCH
     int i;
     for (i = 0; i < size; i++) {
-        out[i] = pow(fabsf(in[i]), 0.75);
+        float a = fabsf(in[i]);
+        out[i] = sqrtf(a * sqrtf(a));
     }
 #endif /* USE_REALLY_FULL_SEARCH */
 }
 
-static av_always_inline int quant2(float coef, const float Q)
-{
-    return pow(coef * Q, 0.75);
-}
-
 static const uint8_t aac_cb_range [12] = {0, 3, 3, 3, 3, 9, 9, 8, 8, 13, 13, 17};
 static const uint8_t aac_cb_maxval[12] = {0, 1, 1, 2, 2, 4, 4, 7, 7, 12, 12, 16};
 
@@ -102,243 +99,198 @@ static const uint8_t aac_cb_maxval[12] = {0, 1, 1, 2, 2, 4, 4, 7, 7, 12, 12, 16}
  *
  * @return quantization distortion
  */
-static float quantize_band_cost(struct AACEncContext *s, const float *in, const float *scaled, int size, int scale_idx, int cb,
-                                 const float lambda, const float uplim, int *bits)
+static av_always_inline float quantize_and_encode_band_cost_template(
+                                struct AACEncContext *s,
+                                PutBitContext *pb, const float *in,
+                                const float *scaled, int size, int scale_idx,
+                                int cb, const float lambda, const float uplim,
+                                int *bits, int BT_ZERO, int BT_UNSIGNED,
+                                int BT_PAIR, int BT_ESC)
 {
     const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
     const float  Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
     const float CLIPPED_ESCAPE = 165140.0f*IQ;
     int i, j, k;
     float cost = 0;
-    const int dim = cb < FIRST_PAIR_BT ? 4 : 2;
+    const int dim = BT_PAIR ? 2 : 4;
     int resbits = 0;
-#ifndef USE_REALLY_FULL_SEARCH
-    const float  Q34 = pow(Q, 0.75);
-    const int range = aac_cb_range[cb];
+    const float  Q34 = sqrtf(Q * sqrtf(Q));
+    const int range  = aac_cb_range[cb];
     const int maxval = aac_cb_maxval[cb];
-    int offs[4];
-#endif /* USE_REALLY_FULL_SEARCH */
+    int off;
 
-    if(!cb){
-        for(i = 0; i < size; i++)
-            cost += in[i]*in[i]*lambda;
-        return cost;
+    if (BT_ZERO) {
+        for (i = 0; i < size; i++)
+            cost += in[i]*in[i];
+        if (bits)
+            *bits = 0;
+        return cost * lambda;
     }
-#ifndef USE_REALLY_FULL_SEARCH
-    offs[0] = 1;
-    for(i = 1; i < dim; i++)
-        offs[i] = offs[i-1]*range;
-    quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
-#endif /* USE_REALLY_FULL_SEARCH */
-    for(i = 0; i < size; i += dim){
-        float mincost;
-        int minidx = 0;
-        int minbits = 0;
+    if (!scaled) {
+        abs_pow34_v(s->scoefs, in, size);
+        scaled = s->scoefs;
+    }
+    quantize_bands(s->qcoefs, in, scaled, size, Q34, !BT_UNSIGNED, maxval);
+    if (BT_UNSIGNED) {
+        off = 0;
+    } else {
+        off = maxval;
+    }
+    for (i = 0; i < size; i += dim) {
         const float *vec;
-#ifndef USE_REALLY_FULL_SEARCH
-        int (*quants)[2] = &s->qcoefs[i];
-        mincost = 0.0f;
-        for(j = 0; j < dim; j++){
-            mincost += in[i+j]*in[i+j]*lambda;
+        int *quants = s->qcoefs + i;
+        int curidx = 0;
+        int curbits;
+        float rd = 0.0f;
+        for (j = 0; j < dim; j++) {
+            curidx *= range;
+            curidx += quants[j] + off;
         }
-        minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
-        minbits = ff_aac_spectral_bits[cb-1][minidx];
-        mincost += minbits;
-        for(j = 0; j < (1<<dim); j++){
-            float rd = 0.0f;
-            int curbits;
-            int curidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
-            int same = 0;
-            for(k = 0; k < dim; k++){
-                if((j & (1 << k)) && quants[k][0] == quants[k][1]){
-                    same = 1;
-                    break;
-                }
-            }
-            if(same)
-                continue;
-            for(k = 0; k < dim; k++)
-                curidx += quants[k][!!(j & (1 << k))] * offs[dim - 1 - k];
-            curbits = ff_aac_spectral_bits[cb-1][curidx];
-            vec = &ff_aac_codebook_vectors[cb-1][curidx*dim];
-#else
-        mincost = INFINITY;
-        vec = ff_aac_codebook_vectors[cb-1];
-        for(j = 0; j < ff_aac_spectral_sizes[cb-1]; j++, vec += dim){
-            float rd = 0.0f;
-            int curbits = ff_aac_spectral_bits[cb-1][j];
-#endif /* USE_REALLY_FULL_SEARCH */
-            if(IS_CODEBOOK_UNSIGNED(cb)){
-                for(k = 0; k < dim; k++){
+            curbits =  ff_aac_spectral_bits[cb-1][curidx];
+            vec     = &ff_aac_codebook_vectors[cb-1][curidx*dim];
+            if (BT_UNSIGNED) {
+                for (k = 0; k < dim; k++) {
                     float t = fabsf(in[i+k]);
                     float di;
-                    //do not code with escape sequence small values
-                    if(vec[k] == 64.0f && t < 39.0f*IQ){
-                        rd = INFINITY;
-                        break;
-                    }
-                    if(vec[k] == 64.0f){//FIXME: slow
+                    if (BT_ESC && vec[k] == 64.0f) { //FIXME: slow
                         if (t >= CLIPPED_ESCAPE) {
                             di = t - CLIPPED_ESCAPE;
                             curbits += 21;
-                        }else{
+                        } else {
                             int c = av_clip(quant(t, Q), 0, 8191);
-                            di = t - c*cbrt(c)*IQ;
+                            di = t - c*cbrtf(c)*IQ;
                             curbits += av_log2(c)*2 - 4 + 1;
                         }
-                    }else{
+                    } else {
                         di = t - vec[k]*IQ;
                     }
-                    if(vec[k] != 0.0f)
+                    if (vec[k] != 0.0f)
                         curbits++;
-                    rd += di*di*lambda;
+                    rd += di*di;
                 }
-            }else{
-                for(k = 0; k < dim; k++){
+            } else {
+                for (k = 0; k < dim; k++) {
                     float di = in[i+k] - vec[k]*IQ;
-                    rd += di*di*lambda;
+                    rd += di*di;
                 }
             }
-            rd += curbits;
-            if(rd < mincost){
-                mincost = rd;
-                minidx = j;
-                minbits = curbits;
+        cost    += rd * lambda + curbits;
+        resbits += curbits;
+        if (cost >= uplim)
+            return uplim;
+        if (pb) {
+            put_bits(pb, ff_aac_spectral_bits[cb-1][curidx], ff_aac_spectral_codes[cb-1][curidx]);
+            if (BT_UNSIGNED)
+                for (j = 0; j < dim; j++)
+                    if (ff_aac_codebook_vectors[cb-1][curidx*dim+j] != 0.0f)
+                        put_bits(pb, 1, in[i+j] < 0.0f);
+            if (BT_ESC) {
+                for (j = 0; j < 2; j++) {
+                    if (ff_aac_codebook_vectors[cb-1][curidx*2+j] == 64.0f) {
+                        int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
+                        int len = av_log2(coef);
+
+                        put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
+                        put_bits(pb, len, coef & ((1 << len) - 1));
+                    }
+                }
             }
         }
-        cost += mincost;
-        resbits += minbits;
-        if(cost >= uplim)
-            return uplim;
     }
 
-    if(bits)
+    if (bits)
         *bits = resbits;
     return cost;
 }
 
-static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb, const float *in, int size,
-                                     int scale_idx, int cb, const float lambda)
-{
-    const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
-    const float  Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
-    const float CLIPPED_ESCAPE = 165140.0f*IQ;
-    const int dim = (cb < FIRST_PAIR_BT) ? 4 : 2;
-    int i, j, k;
-#ifndef USE_REALLY_FULL_SEARCH
-    const float  Q34 = pow(Q, 0.75);
-    const int range = aac_cb_range[cb];
-    const int maxval = aac_cb_maxval[cb];
-    int offs[4];
-    float *scaled = s->scoefs;
-#endif /* USE_REALLY_FULL_SEARCH */
+#define QUANTIZE_AND_ENCODE_BAND_COST_FUNC(NAME, BT_ZERO, BT_UNSIGNED, BT_PAIR, BT_ESC) \
+static float quantize_and_encode_band_cost_ ## NAME(                                        \
+                                struct AACEncContext *s,                                \
+                                PutBitContext *pb, const float *in,                     \
+                                const float *scaled, int size, int scale_idx,           \
+                                int cb, const float lambda, const float uplim,          \
+                                int *bits) {                                            \
+    return quantize_and_encode_band_cost_template(                                      \
+                                s, pb, in, scaled, size, scale_idx,                     \
+                                BT_ESC ? ESC_BT : cb, lambda, uplim, bits,              \
+                                BT_ZERO, BT_UNSIGNED, BT_PAIR, BT_ESC);                 \
+}
 
-//START_TIMER
-    if(!cb)
-        return;
+QUANTIZE_AND_ENCODE_BAND_COST_FUNC(ZERO,  1, 0, 0, 0)
+QUANTIZE_AND_ENCODE_BAND_COST_FUNC(SQUAD, 0, 0, 0, 0)
+QUANTIZE_AND_ENCODE_BAND_COST_FUNC(UQUAD, 0, 1, 0, 0)
+QUANTIZE_AND_ENCODE_BAND_COST_FUNC(SPAIR, 0, 0, 1, 0)
+QUANTIZE_AND_ENCODE_BAND_COST_FUNC(UPAIR, 0, 1, 1, 0)
+QUANTIZE_AND_ENCODE_BAND_COST_FUNC(ESC,   0, 1, 1, 1)
 
-#ifndef USE_REALLY_FULL_SEARCH
-    offs[0] = 1;
-    for(i = 1; i < dim; i++)
-        offs[i] = offs[i-1]*range;
-    abs_pow34_v(scaled, in, size);
-    quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
-#endif /* USE_REALLY_FULL_SEARCH */
-    for(i = 0; i < size; i += dim){
-        float mincost;
-        int minidx = 0;
-        int minbits = 0;
-        const float *vec;
-#ifndef USE_REALLY_FULL_SEARCH
-        int (*quants)[2] = &s->qcoefs[i];
-        mincost = 0.0f;
-        for(j = 0; j < dim; j++){
-            mincost += in[i+j]*in[i+j]*lambda;
-        }
-        minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
-        minbits = ff_aac_spectral_bits[cb-1][minidx];
-        mincost += minbits;
-        for(j = 0; j < (1<<dim); j++){
-            float rd = 0.0f;
-            int curbits;
-            int curidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
-            int same = 0;
-            for(k = 0; k < dim; k++){
-                if((j & (1 << k)) && quants[k][0] == quants[k][1]){
-                    same = 1;
-                    break;
-                }
-            }
-            if(same)
-                continue;
-            for(k = 0; k < dim; k++)
-                curidx += quants[k][!!(j & (1 << k))] * offs[dim - 1 - k];
-            curbits = ff_aac_spectral_bits[cb-1][curidx];
-            vec = &ff_aac_codebook_vectors[cb-1][curidx*dim];
-#else
-        vec = ff_aac_codebook_vectors[cb-1];
-        mincost = INFINITY;
-        for(j = 0; j < ff_aac_spectral_sizes[cb-1]; j++, vec += dim){
-            float rd = 0.0f;
-            int curbits = ff_aac_spectral_bits[cb-1][j];
-            int curidx = j;
-#endif /* USE_REALLY_FULL_SEARCH */
-            if(IS_CODEBOOK_UNSIGNED(cb)){
-                for(k = 0; k < dim; k++){
-                    float t = fabsf(in[i+k]);
-                    float di;
-                    //do not code with escape sequence small values
-                    if(vec[k] == 64.0f && t < 39.0f*IQ){
-                        rd = INFINITY;
-                        break;
-                    }
-                    if(vec[k] == 64.0f){//FIXME: slow
-                        if (t >= CLIPPED_ESCAPE) {
-                            di = t - CLIPPED_ESCAPE;
-                            curbits += 21;
-                        }else{
-                            int c = av_clip(quant(t, Q), 0, 8191);
-                            di = t - c*cbrt(c)*IQ;
-                            curbits += av_log2(c)*2 - 4 + 1;
-                        }
-                    }else{
-                        di = t - vec[k]*IQ;
-                    }
-                    if(vec[k] != 0.0f)
-                        curbits++;
-                    rd += di*di*lambda;
-                }
-            }else{
-                for(k = 0; k < dim; k++){
-                    float di = in[i+k] - vec[k]*IQ;
-                    rd += di*di*lambda;
-                }
-            }
-            rd += curbits;
-            if(rd < mincost){
-                mincost = rd;
-                minidx = curidx;
-                minbits = curbits;
-            }
-        }
-        put_bits(pb, ff_aac_spectral_bits[cb-1][minidx], ff_aac_spectral_codes[cb-1][minidx]);
-        if(IS_CODEBOOK_UNSIGNED(cb))
-            for(j = 0; j < dim; j++)
-                if(ff_aac_codebook_vectors[cb-1][minidx*dim+j] != 0.0f)
-                    put_bits(pb, 1, in[i+j] < 0.0f);
-        if(cb == ESC_BT){
-            for(j = 0; j < 2; j++){
-                if(ff_aac_codebook_vectors[cb-1][minidx*2+j] == 64.0f){
-                    int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
-                    int len = av_log2(coef);
+static float (*const quantize_and_encode_band_cost_arr[])(
+                                struct AACEncContext *s,
+                                PutBitContext *pb, const float *in,
+                                const float *scaled, int size, int scale_idx,
+                                int cb, const float lambda, const float uplim,
+                                int *bits) = {
+    quantize_and_encode_band_cost_ZERO,
+    quantize_and_encode_band_cost_SQUAD,
+    quantize_and_encode_band_cost_SQUAD,
+    quantize_and_encode_band_cost_UQUAD,
+    quantize_and_encode_band_cost_UQUAD,
+    quantize_and_encode_band_cost_SPAIR,
+    quantize_and_encode_band_cost_SPAIR,
+    quantize_and_encode_band_cost_UPAIR,
+    quantize_and_encode_band_cost_UPAIR,
+    quantize_and_encode_band_cost_UPAIR,
+    quantize_and_encode_band_cost_UPAIR,
+    quantize_and_encode_band_cost_ESC,
+};
 
-                    put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
-                    put_bits(pb, len, coef & ((1 << len) - 1));
-                }
-            }
+#define quantize_and_encode_band_cost(                                  \
+                                s, pb, in, scaled, size, scale_idx, cb, \
+                                lambda, uplim, bits)                    \
+    quantize_and_encode_band_cost_arr[cb](                              \
+                                s, pb, in, scaled, size, scale_idx, cb, \
+                                lambda, uplim, bits)
+
+static float quantize_band_cost(struct AACEncContext *s, const float *in,
+                                const float *scaled, int size, int scale_idx,
+                                int cb, const float lambda, const float uplim,
+                                int *bits)
+{
+    return quantize_and_encode_band_cost(s, NULL, in, scaled, size, scale_idx,
+                                         cb, lambda, uplim, bits);
+}
+
+static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
+                                     const float *in, int size, int scale_idx,
+                                     int cb, const float lambda)
+{
+    quantize_and_encode_band_cost(s, pb, in, NULL, size, scale_idx, cb, lambda,
+                                  INFINITY, NULL);
+}
+
+static float find_max_val(int group_len, int swb_size, const float *scaled) {
+    float maxval = 0.0f;
+    int w2, i;
+    for (w2 = 0; w2 < group_len; w2++) {
+        for (i = 0; i < swb_size; i++) {
+            maxval = FFMAX(maxval, scaled[w2*128+i]);
         }
     }
-//STOP_TIMER("quantize_and_encode")
+    return maxval;
+}
+
+static int find_min_book(float maxval, int sf) {
+    float Q = ff_aac_pow2sf_tab[200 - sf + SCALE_ONE_POS - SCALE_DIV_512];
+    float Q34 = sqrtf(Q * sqrtf(Q));
+    int qmaxval, cb;
+    qmaxval = maxval * Q34 + 0.4054f;
+    if      (qmaxval ==  0) cb = 0;
+    else if (qmaxval ==  1) cb = 1;
+    else if (qmaxval ==  2) cb = 3;
+    else if (qmaxval <=  4) cb = 5;
+    else if (qmaxval <=  7) cb = 7;
+    else if (qmaxval <= 12) cb = 9;
+    else                    cb = 11;
+    return cb;
 }
 
 /**
@@ -346,7 +298,6 @@ static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
  */
 typedef struct BandCodingPath {
     int prev_idx; ///< pointer to the previous path point
-    int codebook; ///< codebook for coding band run
     float cost;   ///< path cost
     int run;
 } BandCodingPath;
@@ -360,9 +311,9 @@ static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce
     BandCodingPath path[120][12];
     int w, swb, cb, start, start2, size;
     int i, j;
-    const int max_sfb = sce->ics.max_sfb;
+    const int max_sfb  = sce->ics.max_sfb;
     const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
-    const int run_esc = (1 << run_bits) - 1;
+    const int run_esc  = (1 << run_bits) - 1;
     int idx, ppos, count;
     int stackrun[120], stackcb[120], stack_len;
     float next_minrd = INFINITY;
@@ -370,29 +321,29 @@ static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce
 
     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
     start = win*128;
-    for(cb = 0; cb < 12; cb++){
-        path[0][cb].cost = 0.0f;
+    for (cb = 0; cb < 12; cb++) {
+        path[0][cb].cost     = 0.0f;
         path[0][cb].prev_idx = -1;
-        path[0][cb].run = 0;
+        path[0][cb].run      = 0;
     }
-    for(swb = 0; swb < max_sfb; swb++){
+    for (swb = 0; swb < max_sfb; swb++) {
         start2 = start;
         size = sce->ics.swb_sizes[swb];
-        if(sce->zeroes[win*16 + swb]){
-            for(cb = 0; cb < 12; cb++){
+        if (sce->zeroes[win*16 + swb]) {
+            for (cb = 0; cb < 12; cb++) {
                 path[swb+1][cb].prev_idx = cb;
-                path[swb+1][cb].cost = path[swb][cb].cost;
-                path[swb+1][cb].run = path[swb][cb].run + 1;
+                path[swb+1][cb].cost     = path[swb][cb].cost;
+                path[swb+1][cb].run      = path[swb][cb].run + 1;
             }
-        }else{
+        } else {
             float minrd = next_minrd;
             int mincb = next_mincb;
             next_minrd = INFINITY;
             next_mincb = 0;
-            for(cb = 0; cb < 12; cb++){
+            for (cb = 0; cb < 12; cb++) {
                 float cost_stay_here, cost_get_here;
                 float rd = 0.0f;
-                for(w = 0; w < group_len; w++){
+                for (w = 0; w < group_len; w++) {
                     FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(win+w)*16+swb];
                     rd += quantize_band_cost(s, sce->coeffs + start + w*128,
                                              s->scoefs + start + w*128, size,
@@ -401,8 +352,8 @@ static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce
                 }
                 cost_stay_here = path[swb][cb].cost + rd;
                 cost_get_here  = minrd              + rd + run_bits + 4;
-                if(   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
-                   != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
+                if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
+                    != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
                     cost_stay_here += run_bits;
                 if (cost_get_here < cost_stay_here) {
                     path[swb+1][cb].prev_idx = mincb;
@@ -424,13 +375,12 @@ static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce
 
     //convert resulting path from backward-linked list
     stack_len = 0;
-    idx = 0;
-    for(cb = 1; cb < 12; cb++){
-        if(path[max_sfb][cb].cost < path[max_sfb][idx].cost)
+    idx       = 0;
+    for (cb = 1; cb < 12; cb++)
+        if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
             idx = cb;
-    }
     ppos = max_sfb;
-    while(ppos > 0){
+    while (ppos > 0) {
         cb = idx;
         stackrun[stack_len] = path[ppos][cb].run;
         stackcb [stack_len] = cb;
@@ -440,16 +390,16 @@ static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce
     }
     //perform actual band info encoding
     start = 0;
-    for(i = stack_len - 1; i >= 0; i--){
+    for (i = stack_len - 1; i >= 0; i--) {
         put_bits(&s->pb, 4, stackcb[i]);
         count = stackrun[i];
         memset(sce->zeroes + win*16 + start, !stackcb[i], count);
         //XXX: memset when band_type is also uint8_t
-        for(j = 0; j < count; j++){
+        for (j = 0; j < count; j++) {
             sce->band_type[win*16 + start] =  stackcb[i];
             start++;
         }
-        while(count >= run_esc){
+        while (count >= run_esc) {
             put_bits(&s->pb, run_bits, run_esc);
             count -= run_esc;
         }
@@ -457,185 +407,313 @@ static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce
     }
 }
 
-static void encode_window_bands_info_fixed(AACEncContext *s, SingleChannelElement *sce,
-                                           int win, int group_len, const float lambda)
+static void codebook_trellis_rate(AACEncContext *s, SingleChannelElement *sce,
+                                  int win, int group_len, const float lambda)
 {
-    encode_window_bands_info(s, sce, win, group_len, 1.0f);
+    BandCodingPath path[120][12];
+    int w, swb, cb, start, start2, size;
+    int i, j;
+    const int max_sfb  = sce->ics.max_sfb;
+    const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
+    const int run_esc  = (1 << run_bits) - 1;
+    int idx, ppos, count;
+    int stackrun[120], stackcb[120], stack_len;
+    float next_minrd = INFINITY;
+    int next_mincb = 0;
+
+    abs_pow34_v(s->scoefs, sce->coeffs, 1024);
+    start = win*128;
+    for (cb = 0; cb < 12; cb++) {
+        path[0][cb].cost     = run_bits+4;
+        path[0][cb].prev_idx = -1;
+        path[0][cb].run      = 0;
+    }
+    for (swb = 0; swb < max_sfb; swb++) {
+        start2 = start;
+        size = sce->ics.swb_sizes[swb];
+        if (sce->zeroes[win*16 + swb]) {
+            for (cb = 0; cb < 12; cb++) {
+                path[swb+1][cb].prev_idx = cb;
+                path[swb+1][cb].cost     = path[swb][cb].cost;
+                path[swb+1][cb].run      = path[swb][cb].run + 1;
+            }
+        } else {
+            float minrd = next_minrd;
+            int mincb = next_mincb;
+            int startcb = sce->band_type[win*16+swb];
+            next_minrd = INFINITY;
+            next_mincb = 0;
+            for (cb = 0; cb < startcb; cb++) {
+                path[swb+1][cb].cost = 61450;
+                path[swb+1][cb].prev_idx = -1;
+                path[swb+1][cb].run = 0;
+            }
+            for (cb = startcb; cb < 12; cb++) {
+                float cost_stay_here, cost_get_here;
+                float rd = 0.0f;
+                for (w = 0; w < group_len; w++) {
+                    rd += quantize_band_cost(s, sce->coeffs + start + w*128,
+                                             s->scoefs + start + w*128, size,
+                                             sce->sf_idx[(win+w)*16+swb], cb,
+                                             0, INFINITY, NULL);
+                }
+                cost_stay_here = path[swb][cb].cost + rd;
+                cost_get_here  = minrd              + rd + run_bits + 4;
+                if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
+                    != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
+                    cost_stay_here += run_bits;
+                if (cost_get_here < cost_stay_here) {
+                    path[swb+1][cb].prev_idx = mincb;
+                    path[swb+1][cb].cost     = cost_get_here;
+                    path[swb+1][cb].run      = 1;
+                } else {
+                    path[swb+1][cb].prev_idx = cb;
+                    path[swb+1][cb].cost     = cost_stay_here;
+                    path[swb+1][cb].run      = path[swb][cb].run + 1;
+                }
+                if (path[swb+1][cb].cost < next_minrd) {
+                    next_minrd = path[swb+1][cb].cost;
+                    next_mincb = cb;
+                }
+            }
+        }
+        start += sce->ics.swb_sizes[swb];
+    }
+
+    //convert resulting path from backward-linked list
+    stack_len = 0;
+    idx       = 0;
+    for (cb = 1; cb < 12; cb++)
+        if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
+            idx = cb;
+    ppos = max_sfb;
+    while (ppos > 0) {
+        assert(idx >= 0);
+        cb = idx;
+        stackrun[stack_len] = path[ppos][cb].run;
+        stackcb [stack_len] = cb;
+        idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
+        ppos -= path[ppos][cb].run;
+        stack_len++;
+    }
+    //perform actual band info encoding
+    start = 0;
+    for (i = stack_len - 1; i >= 0; i--) {
+        put_bits(&s->pb, 4, stackcb[i]);
+        count = stackrun[i];
+        memset(sce->zeroes + win*16 + start, !stackcb[i], count);
+        //XXX: memset when band_type is also uint8_t
+        for (j = 0; j < count; j++) {
+            sce->band_type[win*16 + start] =  stackcb[i];
+            start++;
+        }
+        while (count >= run_esc) {
+            put_bits(&s->pb, run_bits, run_esc);
+            count -= run_esc;
+        }
+        put_bits(&s->pb, run_bits, count);
+    }
+}
+
+/** Return the minimum scalefactor where the quantized coef does not clip. */
+static av_always_inline uint8_t coef2minsf(float coef) {
+    return av_clip_uint8(log2f(coef)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
 }
 
+/** Return the maximum scalefactor where the quantized coef is not zero. */
+static av_always_inline uint8_t coef2maxsf(float coef) {
+    return av_clip_uint8(log2f(coef)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
+}
 
 typedef struct TrellisPath {
     float cost;
     int prev;
-    int min_val;
-    int max_val;
 } TrellisPath;
 
+#define TRELLIS_STAGES 121
+#define TRELLIS_STATES (SCALE_MAX_DIFF+1)
+
 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
-                                       SingleChannelElement *sce, const float lambda)
+                                       SingleChannelElement *sce,
+                                       const float lambda)
 {
     int q, w, w2, g, start = 0;
-    int i;
+    int i, j;
     int idx;
-    TrellisPath paths[256*121];
-    int bandaddr[121];
+    TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
+    int bandaddr[TRELLIS_STAGES];
     int minq;
     float mincost;
+    float q0f = FLT_MAX, q1f = 0.0f, qnrgf = 0.0f;
+    int q0, q1, qcnt = 0;
+
+    for (i = 0; i < 1024; i++) {
+        float t = fabsf(sce->coeffs[i]);
+        if (t > 0.0f) {
+            q0f = FFMIN(q0f, t);
+            q1f = FFMAX(q1f, t);
+            qnrgf += t*t;
+            qcnt++;
+        }
+    }
+
+    if (!qcnt) {
+        memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
+        memset(sce->zeroes, 1, sizeof(sce->zeroes));
+        return;
+    }
+
+    //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
+    q0 = coef2minsf(q0f);
+    //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
+    q1 = coef2maxsf(q1f);
+    //av_log(NULL, AV_LOG_ERROR, "q0 %d, q1 %d\n", q0, q1);
+    if (q1 - q0 > 60) {
+        int q0low  = q0;
+        int q1high = q1;
+        //minimum scalefactor index is when maximum nonzero coefficient after quantizing is not clipped
+        int qnrg = av_clip_uint8(log2f(sqrtf(qnrgf/qcnt))*4 - 31 + SCALE_ONE_POS - SCALE_DIV_512);
+        q1 = qnrg + 30;
+        q0 = qnrg - 30;
+    //av_log(NULL, AV_LOG_ERROR, "q0 %d, q1 %d\n", q0, q1);
+        if (q0 < q0low) {
+            q1 += q0low - q0;
+            q0  = q0low;
+        } else if (q1 > q1high) {
+            q0 -= q1 - q1high;
+            q1  = q1high;
+        }
+    }
+    //av_log(NULL, AV_LOG_ERROR, "q0 %d, q1 %d\n", q0, q1);
 
-    for(i = 0; i < 256; i++){
-        paths[i].cost = 0.0f;
-        paths[i].prev = -1;
-        paths[i].min_val = i;
-        paths[i].max_val = i;
+    for (i = 0; i < TRELLIS_STATES; i++) {
+        paths[0][i].cost    = 0.0f;
+        paths[0][i].prev    = -1;
     }
-    for(i = 256; i < 256*121; i++){
-        paths[i].cost = INFINITY;
-        paths[i].prev = -2;
-        paths[i].min_val = INT_MAX;
-        paths[i].max_val = 0;
+    for (j = 1; j < TRELLIS_STAGES; j++) {
+        for (i = 0; i < TRELLIS_STATES; i++) {
+            paths[j][i].cost    = INFINITY;
+            paths[j][i].prev    = -2;
+        }
     }
-    idx = 256;
+    idx = 1;
     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
-    for(w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]){
+    for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
         start = w*128;
-        for(g = 0; g < sce->ics.num_swb; g++){
+        for (g = 0; g < sce->ics.num_swb; g++) {
             const float *coefs = sce->coeffs + start;
             float qmin, qmax;
             int nz = 0;
 
-            bandaddr[idx >> 8] = w*16+g;
+            bandaddr[idx] = w * 16 + g;
             qmin = INT_MAX;
             qmax = 0.0f;
-            for(w2 = 0; w2 < sce->ics.group_len[w]; w2++){
+            for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
-                if(band->energy <= band->threshold || band->threshold == 0.0f){
+                if (band->energy <= band->threshold || band->threshold == 0.0f) {
                     sce->zeroes[(w+w2)*16+g] = 1;
                     continue;
                 }
                 sce->zeroes[(w+w2)*16+g] = 0;
                 nz = 1;
-                for(i = 0; i < sce->ics.swb_sizes[g]; i++){
+                for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
                     float t = fabsf(coefs[w2*128+i]);
-                    if(t > 0.0f) qmin = fminf(qmin, t);
-                    qmax = fmaxf(qmax, t);
+                    if (t > 0.0f)
+                        qmin = FFMIN(qmin, t);
+                    qmax = FFMAX(qmax, t);
                 }
             }
-            if(nz){
+            if (nz) {
                 int minscale, maxscale;
                 float minrd = INFINITY;
+                float maxval;
                 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
-                minscale = av_clip_uint8(log2(qmin)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
+                minscale = coef2minsf(qmin);
                 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
-                maxscale = av_clip_uint8(log2(qmax)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
-                for(q = minscale; q < maxscale; q++){
-                    float dists[12], dist;
-                    memset(dists, 0, sizeof(dists));
-                    for(w2 = 0; w2 < sce->ics.group_len[w]; w2++){
+                maxscale = coef2maxsf(qmax);
+                minscale = av_clip(minscale - q0, 0, TRELLIS_STATES - 1);
+                maxscale = av_clip(maxscale - q0, 0, TRELLIS_STATES);
+                maxval = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], s->scoefs+start);
+                for (q = minscale; q < maxscale; q++) {
+                    float dist = 0;
+                    int cb = find_min_book(maxval, sce->sf_idx[w*16+g]);
+                    for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
                         FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
-                        int cb;
-                        for(cb = 0; cb <= ESC_BT; cb++){
-                            dists[cb] += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
-                                                            q, cb, lambda / band->threshold, INFINITY, NULL);
-                        }
+                        dist += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
+                                                   q + q0, cb, lambda / band->threshold, INFINITY, NULL);
                     }
-                    dist = dists[0];
-                    for(i = 1; i <= ESC_BT; i++)
-                        dist = fminf(dist, dists[i]);
-                    minrd = fminf(minrd, dist);
+                    minrd = FFMIN(minrd, dist);
 
-                    for(i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, 256); i++){
+                    for (i = 0; i < q1 - q0; i++) {
                         float cost;
-                        int minv, maxv;
-                        if(isinf(paths[idx - 256 + i].cost))
-                            continue;
-                        cost = paths[idx - 256 + i].cost + dist
+                        cost = paths[idx - 1][i].cost + dist
                                + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
-                        minv = FFMIN(paths[idx - 256 + i].min_val, q);
-                        maxv = FFMAX(paths[idx - 256 + i].max_val, q);
-                        if(cost < paths[idx + q].cost && maxv-minv < SCALE_MAX_DIFF){
-                            paths[idx + q].cost = cost;
-                            paths[idx + q].prev = idx - 256 + i;
-                            paths[idx + q].min_val = minv;
-                            paths[idx + q].max_val = maxv;
+                        if (cost < paths[idx][q].cost) {
+                            paths[idx][q].cost    = cost;
+                            paths[idx][q].prev    = i;
                         }
                     }
                 }
-            }else{
-                for(q = 0; q < 256; q++){
-                    if(!isinf(paths[idx - 256 + q].cost)){
-                        paths[idx + q].cost = paths[idx - 256 + q].cost + 1;
-                        paths[idx + q].prev = idx - 256 + q;
-                        paths[idx + q].min_val = FFMIN(paths[idx - 256 + q].min_val, q);
-                        paths[idx + q].max_val = FFMAX(paths[idx - 256 + q].max_val, q);
-                        continue;
-                    }
-                    for(i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, 256); i++){
-                        float cost;
-                        int minv, maxv;
-                        if(isinf(paths[idx - 256 + i].cost))
-                            continue;
-                        cost = paths[idx - 256 + i].cost + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
-                        minv = FFMIN(paths[idx - 256 + i].min_val, q);
-                        maxv = FFMAX(paths[idx - 256 + i].max_val, q);
-                        if(cost < paths[idx + q].cost && maxv-minv < SCALE_MAX_DIFF){
-                            paths[idx + q].cost = cost;
-                            paths[idx + q].prev = idx - 256 + i;
-                            paths[idx + q].min_val = minv;
-                            paths[idx + q].max_val = maxv;
-                        }
-                    }
+            } else {
+                for (q = 0; q < q1 - q0; q++) {
+                    paths[idx][q].cost = paths[idx - 1][q].cost + 1;
+                    paths[idx][q].prev = q;
                 }
             }
             sce->zeroes[w*16+g] = !nz;
             start += sce->ics.swb_sizes[g];
-            idx += 256;
+            idx++;
         }
     }
-    idx -= 256;
-    mincost = paths[idx].cost;
-    minq = idx;
-    for(i = 1; i < 256; i++){
-        if(paths[idx + i].cost < mincost){
-            mincost = paths[idx + i].cost;
-            minq = idx + i;
+    idx--;
+    mincost = paths[idx][0].cost;
+    minq    = 0;
+    for (i = 1; i < TRELLIS_STATES; i++) {
+        if (paths[idx][i].cost < mincost) {
+            mincost = paths[idx][i].cost;
+            minq = i;
         }
     }
-    while(minq >= 256){
-        sce->sf_idx[bandaddr[minq>>8]] = minq & 0xFF;
-        minq = paths[minq].prev;
+    while (idx) {
+        sce->sf_idx[bandaddr[idx]] = minq + q0;
+        minq = paths[idx][minq].prev;
+        idx--;
     }
     //set the same quantizers inside window groups
-    for(w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
-        for(g = 0;  g < sce->ics.num_swb; g++)
-            for(w2 = 1; w2 < sce->ics.group_len[w]; w2++)
+    for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
+        for (g = 0;  g < sce->ics.num_swb; g++)
+            for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
 }
 
 /**
  * two-loop quantizers search taken from ISO 13818-7 Appendix C
  */
-static void search_for_quantizers_twoloop(AVCodecContext *avctx, AACEncContext *s,
-                                          SingleChannelElement *sce, const float lambda)
+static void search_for_quantizers_twoloop(AVCodecContext *avctx,
+                                          AACEncContext *s,
+                                          SingleChannelElement *sce,
+                                          const float lambda)
 {
     int start = 0, i, w, w2, g;
     int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels;
     float dists[128], uplims[128];
+    float maxvals[128];
     int fflag, minscaler;
-    int its = 0;
+    int its  = 0;
     int allz = 0;
     float minthr = INFINITY;
 
     //XXX: some heuristic to determine initial quantizers will reduce search time
     memset(dists, 0, sizeof(dists));
     //determine zero bands and upper limits
-    for(w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]){
-        for(g = 0;  g < sce->ics.num_swb; g++){
+    for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
+        for (g = 0;  g < sce->ics.num_swb; g++) {
             int nz = 0;
             float uplim = 0.0f;
-            for(w2 = 0; w2 < sce->ics.group_len[w]; w2++){
+            for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
                 uplim += band->threshold;
-                if(band->energy <= band->threshold || band->threshold == 0.0f){
+                if (band->energy <= band->threshold || band->threshold == 0.0f) {
                     sce->zeroes[(w+w2)*16+g] = 1;
                     continue;
                 }
@@ -643,71 +721,74 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx, AACEncContext *
             }
             uplims[w*16+g] = uplim *512;
             sce->zeroes[w*16+g] = !nz;
-            if(nz)
-                minthr = fminf(minthr, uplim);
+            if (nz)
+                minthr = FFMIN(minthr, uplim);
             allz = FFMAX(allz, nz);
         }
     }
-    for(w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]){
-        for(g = 0;  g < sce->ics.num_swb; g++){
-            if(sce->zeroes[w*16+g]){
+    for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
+        for (g = 0;  g < sce->ics.num_swb; g++) {
+            if (sce->zeroes[w*16+g]) {
                 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
                 continue;
             }
-            sce->sf_idx[w*16+g] = SCALE_ONE_POS + fminf(log2(uplims[w*16+g]/minthr)*4,59);
+            sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2f(uplims[w*16+g]/minthr)*4,59);
         }
     }
 
-    if(!allz)
+    if (!allz)
         return;
     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
+
+    for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
+        start = w*128;
+        for (g = 0;  g < sce->ics.num_swb; g++) {
+            const float *scaled = s->scoefs + start;
+            maxvals[w*16+g] = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled);
+            start += sce->ics.swb_sizes[g];
+        }
+    }
+
     //perform two-loop search
     //outer loop - improve quality
-    do{
+    do {
         int tbits, qstep;
         minscaler = sce->sf_idx[0];
         //inner loop - quantize spectrum to fit into given number of bits
         qstep = its ? 1 : 32;
-        do{
+        do {
             int prev = -1;
             tbits = 0;
             fflag = 0;
-            for(w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]){
+            for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
                 start = w*128;
-                for(g = 0;  g < sce->ics.num_swb; g++){
+                for (g = 0;  g < sce->ics.num_swb; g++) {
                     const float *coefs = sce->coeffs + start;
                     const float *scaled = s->scoefs + start;
                     int bits = 0;
                     int cb;
-                    float mindist = INFINITY;
-                    int minbits = 0;
+                    float dist = 0.0f;
 
-                    if(sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218)
+                    if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
+                        start += sce->ics.swb_sizes[g];
                         continue;
+                    }
                     minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
-                    for(cb = 0; cb <= ESC_BT; cb++){
-                        float dist = 0.0f;
-                        int bb = 0;
-                        for(w2 = 0; w2 < sce->ics.group_len[w]; w2++){
-                            int b;
-                            dist += quantize_band_cost(s, coefs + w2*128,
-                                                       scaled + w2*128,
-                                                       sce->ics.swb_sizes[g],
-                                                       sce->sf_idx[w*16+g],
-                                                       ESC_BT,
-                                                       1.0,
-                                                       INFINITY,
-                                                       &b);
-                            bb += b;
-                        }
-                        if(dist < mindist){
-                            mindist = dist;
-                            minbits = bb;
-                        }
+                    cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
+                    for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
+                        int b;
+                        dist += quantize_band_cost(s, coefs + w2*128,
+                                                   scaled + w2*128,
+                                                   sce->ics.swb_sizes[g],
+                                                   sce->sf_idx[w*16+g],
+                                                   cb,
+                                                   1.0f,
+                                                   INFINITY,
+                                                   &b);
+                        bits += b;
                     }
-                    dists[w*16+g] = mindist - minbits;
-                    bits = minbits;
-                    if(prev != -1){
+                    dists[w*16+g] = dist - bits;
+                    if (prev != -1) {
                         bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
                     }
                     tbits += bits;
@@ -715,45 +796,45 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx, AACEncContext *
                     prev = sce->sf_idx[w*16+g];
                 }
             }
-            if(tbits > destbits){
-                for(i = 0; i < 128; i++){
-                    if(sce->sf_idx[i] < 218 - qstep){
+            if (tbits > destbits) {
+                for (i = 0; i < 128; i++)
+                    if (sce->sf_idx[i] < 218 - qstep)
                         sce->sf_idx[i] += qstep;
-                    }
-                }
-            }else{
-                for(i = 0; i < 128; i++){
-                    if(sce->sf_idx[i] > 60 - qstep){
+            } else {
+                for (i = 0; i < 128; i++)
+                    if (sce->sf_idx[i] > 60 - qstep)
                         sce->sf_idx[i] -= qstep;
-                    }
-                }
             }
             qstep >>= 1;
-            if(!qstep && tbits > destbits*1.02)
+            if (!qstep && tbits > destbits*1.02 && sce->sf_idx[0] < 217)
                 qstep = 1;
-            if(sce->sf_idx[0] >= 217)break;
-        }while(qstep);
+        } while (qstep);
 
         fflag = 0;
         minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
-        for(w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]){
-            start = w*128;
-            for(g = 0; g < sce->ics.num_swb; g++){
+        for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
+            for (g = 0; g < sce->ics.num_swb; g++) {
                 int prevsc = sce->sf_idx[w*16+g];
-                if(dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60)
+                if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60) {
+                    if (find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]-1))
                     sce->sf_idx[w*16+g]--;
+                    else //Try to make sure there is some energy in every band
+                        sce->sf_idx[w*16+g]-=2;
+                }
                 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
                 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
-                if(sce->sf_idx[w*16+g] != prevsc)
+                if (sce->sf_idx[w*16+g] != prevsc)
                     fflag = 1;
+                sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
             }
         }
         its++;
-    }while(fflag && its < 10);
+    } while (fflag && its < 10);
 }
 
 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
-                                       SingleChannelElement *sce, const float lambda)
+                                       SingleChannelElement *sce,
+                                       const float lambda)
 {
     int start = 0, i, w, w2, g;
     float uplim[128], maxq[128];
@@ -761,29 +842,29 @@ static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
     float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
     int last = 0, lastband = 0, curband = 0;
     float avg_energy = 0.0;
-    if(sce->ics.num_windows == 1){
+    if (sce->ics.num_windows == 1) {
         start = 0;
-        for(i = 0; i < 1024; i++){
-            if(i - start >= sce->ics.swb_sizes[curband]){
+        for (i = 0; i < 1024; i++) {
+            if (i - start >= sce->ics.swb_sizes[curband]) {
                 start += sce->ics.swb_sizes[curband];
                 curband++;
             }
-            if(sce->coeffs[i]){
+            if (sce->coeffs[i]) {
                 avg_energy += sce->coeffs[i] * sce->coeffs[i];
                 last = i;
                 lastband = curband;
             }
         }
-    }else{
-        for(w = 0; w < 8; w++){
+    } else {
+        for (w = 0; w < 8; w++) {
             const float *coeffs = sce->coeffs + w*128;
             start = 0;
-            for(i = 0; i < 128; i++){
-                if(i - start >= sce->ics.swb_sizes[curband]){
+            for (i = 0; i < 128; i++) {
+                if (i - start >= sce->ics.swb_sizes[curband]) {
                     start += sce->ics.swb_sizes[curband];
                     curband++;
                 }
-                if(coeffs[i]){
+                if (coeffs[i]) {
                     avg_energy += coeffs[i] * coeffs[i];
                     last = FFMAX(last, i);
                     lastband = FFMAX(lastband, curband);
@@ -793,183 +874,189 @@ static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
     }
     last++;
     avg_energy /= last;
-    if(avg_energy == 0.0f){
-        for(i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
+    if (avg_energy == 0.0f) {
+        for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
             sce->sf_idx[i] = SCALE_ONE_POS;
         return;
     }
-    for(w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]){
+    for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
         start = w*128;
-        for(g = 0; g < sce->ics.num_swb; g++){
-            float *coefs = sce->coeffs + start;
+        for (g = 0; g < sce->ics.num_swb; g++) {
+            float *coefs   = sce->coeffs + start;
             const int size = sce->ics.swb_sizes[g];
             int start2 = start, end2 = start + size, peakpos = start;
             float maxval = -1, thr = 0.0f, t;
             maxq[w*16+g] = 0.0f;
-            if(g > lastband){
+            if (g > lastband) {
                 maxq[w*16+g] = 0.0f;
                 start += size;
-                for(w2 = 0; w2 < sce->ics.group_len[w]; w2++)
+                for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
                     memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
                 continue;
             }
-            for(w2 = 0; w2 < sce->ics.group_len[w]; w2++){
-                for(i = 0; i < size; i++){
+            for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
+                for (i = 0; i < size; i++) {
                     float t = coefs[w2*128+i]*coefs[w2*128+i];
-                    maxq[w*16+g] = fmaxf(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
+                    maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
                     thr += t;
-                    if(sce->ics.num_windows == 1 && maxval < t){
-                        maxval = t;
+                    if (sce->ics.num_windows == 1 && maxval < t) {
+                        maxval  = t;
                         peakpos = start+i;
                     }
                 }
             }
-            if(sce->ics.num_windows == 1){
+            if (sce->ics.num_windows == 1) {
                 start2 = FFMAX(peakpos - 2, start2);
                 end2   = FFMIN(peakpos + 3, end2);
-            }else{
+            } else {
                 start2 -= start;
                 end2   -= start;
             }
             start += size;
             thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
-            t = 1.0 - (1.0 * start2 / last);
+            t   = 1.0 - (1.0 * start2 / last);
             uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
         }
     }
     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
-    for(w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]){
+    for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
         start = w*128;
-        for(g = 0;  g < sce->ics.num_swb; g++){
-            const float *coefs = sce->coeffs + start;
-            const float *scaled = s->scoefs + start;
-            const int size = sce->ics.swb_sizes[g];
+        for (g = 0;  g < sce->ics.num_swb; g++) {
+            const float *coefs  = sce->coeffs + start;
+            const float *scaled = s->scoefs   + start;
+            const int size      = sce->ics.swb_sizes[g];
             int scf, prev_scf, step;
-            int min_scf = 0, max_scf = 255;
+            int min_scf = -1, max_scf = 256;
             float curdiff;
-            if(maxq[w*16+g] < 21.544){
+            if (maxq[w*16+g] < 21.544) {
                 sce->zeroes[w*16+g] = 1;
                 start += size;
                 continue;
             }
             sce->zeroes[w*16+g] = 0;
-            scf = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2(1/maxq[w*16+g])*16/3, 60, 218);
+            scf  = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2f(1/maxq[w*16+g])*16/3, 60, 218);
             step = 16;
-            for(;;){
+            for (;;) {
                 float dist = 0.0f;
                 int quant_max;
 
-                for(w2 = 0; w2 < sce->ics.group_len[w]; w2++){
+                for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
                     int b;
                     dist += quantize_band_cost(s, coefs + w2*128,
                                                scaled + w2*128,
                                                sce->ics.swb_sizes[g],
                                                scf,
                                                ESC_BT,
-                                               1.0,
+                                               lambda,
                                                INFINITY,
                                                &b);
                     dist -= b;
                 }
-                dist *= 1.0f/512.0f;
+                dist *= 1.0f / 512.0f / lambda;
                 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[200 - scf + SCALE_ONE_POS - SCALE_DIV_512]);
-                if(quant_max >= 8191){ // too much, return to the previous quantizer
+                if (quant_max >= 8191) { // too much, return to the previous quantizer
                     sce->sf_idx[w*16+g] = prev_scf;
                     break;
                 }
                 prev_scf = scf;
                 curdiff = fabsf(dist - uplim[w*16+g]);
-                if(curdiff == 0.0f)
+                if (curdiff <= 1.0f)
                     step = 0;
                 else
-                    step = fabsf(log2(curdiff));
-                if(dist > uplim[w*16+g])
+                    step = log2f(curdiff);
+                if (dist > uplim[w*16+g])
                     step = -step;
-                if(FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)){
-                    sce->sf_idx[w*16+g] = scf;
+                scf += step;
+                scf = av_clip_uint8(scf);
+                step = scf - prev_scf;
+                if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
+                    sce->sf_idx[w*16+g] = av_clip(scf, min_scf, max_scf);
                     break;
                 }
-                scf += step;
-                if(step > 0)
-                    min_scf = scf;
+                if (step > 0)
+                    min_scf = prev_scf;
                 else
-                    max_scf = scf;
+                    max_scf = prev_scf;
             }
             start += size;
         }
     }
     minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
-    for(i = 1; i < 128; i++){
-        if(!sce->sf_idx[i])
+    for (i = 1; i < 128; i++) {
+        if (!sce->sf_idx[i])
             sce->sf_idx[i] = sce->sf_idx[i-1];
         else
             minq = FFMIN(minq, sce->sf_idx[i]);
     }
-    if(minq == INT_MAX) minq = 0;
+    if (minq == INT_MAX)
+        minq = 0;
     minq = FFMIN(minq, SCALE_MAX_POS);
     maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
-    for(i = 126; i >= 0; i--){
-        if(!sce->sf_idx[i])
+    for (i = 126; i >= 0; i--) {
+        if (!sce->sf_idx[i])
             sce->sf_idx[i] = sce->sf_idx[i+1];
         sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
     }
 }
 
 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
-                                       SingleChannelElement *sce, const float lambda)
+                                       SingleChannelElement *sce,
+                                       const float lambda)
 {
     int start = 0, i, w, w2, g;
     int minq = 255;
 
     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
-    for(w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]){
+    for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
         start = w*128;
-        for(g = 0; g < sce->ics.num_swb; g++){
-            for(w2 = 0; w2 < sce->ics.group_len[w]; w2++){
+        for (g = 0; g < sce->ics.num_swb; g++) {
+            for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
-                if(band->energy <= band->threshold){
+                if (band->energy <= band->threshold) {
                     sce->sf_idx[(w+w2)*16+g] = 218;
                     sce->zeroes[(w+w2)*16+g] = 1;
-                }else{
-                    sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2(band->threshold), 80, 218);
+                } else {
+                    sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2f(band->threshold), 80, 218);
                     sce->zeroes[(w+w2)*16+g] = 0;
                 }
                 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
             }
         }
     }
-    for(i = 0; i < 128; i++){
-        sce->sf_idx[i] = 140;//av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
+    for (i = 0; i < 128; i++) {
+        sce->sf_idx[i] = 140;
+        //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
     }
     //set the same quantizers inside window groups
-    for(w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
-        for(g = 0;  g < sce->ics.num_swb; g++)
-            for(w2 = 1; w2 < sce->ics.group_len[w]; w2++)
+    for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
+        for (g = 0;  g < sce->ics.num_swb; g++)
+            for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
 }
 
-static void search_for_ms(AACEncContext *s, ChannelElement *cpe, const float lambda)
+static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
+                          const float lambda)
 {
     int start = 0, i, w, w2, g;
     float M[128], S[128];
     float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
     SingleChannelElement *sce0 = &cpe->ch[0];
     SingleChannelElement *sce1 = &cpe->ch[1];
-    if(!cpe->common_window)
+    if (!cpe->common_window)
         return;
-    for(w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]){
-        for(g = 0;  g < sce0->ics.num_swb; g++){
-            if(!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]){
+    for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
+        for (g = 0;  g < sce0->ics.num_swb; g++) {
+            if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
                 float dist1 = 0.0f, dist2 = 0.0f;
-                for(w2 = 0; w2 < sce0->ics.group_len[w]; w2++){
+                for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
                     FFPsyBand *band0 = &s->psy.psy_bands[(s->cur_channel+0)*PSY_MAX_BANDS+(w+w2)*16+g];
                     FFPsyBand *band1 = &s->psy.psy_bands[(s->cur_channel+1)*PSY_MAX_BANDS+(w+w2)*16+g];
-                    float minthr = fminf(band0->threshold, band1->threshold);
-                    float maxthr = fmaxf(band0->threshold, band1->threshold);
-                    for(i = 0; i < sce0->ics.swb_sizes[g]; i++){
+                    float minthr = FFMIN(band0->threshold, band1->threshold);
+                    float maxthr = FFMAX(band0->threshold, band1->threshold);
+                    for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
                         M[i] = (sce0->coeffs[start+w2*128+i]
-                              + sce1->coeffs[start+w2*128+i])*0.5;
+                              + sce1->coeffs[start+w2*128+i]) * 0.5;
                         S[i] =  sce0->coeffs[start+w2*128+i]
                               - sce1->coeffs[start+w2*128+i];
                     }
@@ -1012,26 +1099,26 @@ static void search_for_ms(AACEncContext *s, ChannelElement *cpe, const float lam
 AACCoefficientsEncoder ff_aac_coders[] = {
     {
         search_for_quantizers_faac,
-        encode_window_bands_info_fixed,
+        encode_window_bands_info,
         quantize_and_encode_band,
-//        search_for_ms,
+        search_for_ms,
     },
     {
         search_for_quantizers_anmr,
         encode_window_bands_info,
         quantize_and_encode_band,
-//        search_for_ms,
+        search_for_ms,
     },
     {
         search_for_quantizers_twoloop,
-        encode_window_bands_info,
+        codebook_trellis_rate,
         quantize_and_encode_band,
-//        search_for_ms,
+        search_for_ms,
     },
     {
         search_for_quantizers_fast,
         encode_window_bands_info,
         quantize_and_encode_band,
-//        search_for_ms,
+        search_for_ms,
     },
 };