]> git.sesse.net Git - ffmpeg/blobdiff - libavcodec/ffv1.c
ARM: ac3: update ff_ac3_extract_exponents_neon per 8b7b2d6
[ffmpeg] / libavcodec / ffv1.c
index 87a2050535b21a32fa12c1444c17077e075f8879..50f1062ad486a7fc5fba2ee005e75dcff1aa0527 100644 (file)
@@ -3,20 +3,20 @@
  *
  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
  *
- * This file is part of FFmpeg.
+ * This file is part of Libav.
  *
- * FFmpeg is free software; you can redistribute it and/or
+ * Libav is free software; you can redistribute it and/or
  * modify it under the terms of the GNU Lesser General Public
  * License as published by the Free Software Foundation; either
  * version 2.1 of the License, or (at your option) any later version.
  *
- * FFmpeg is distributed in the hope that it will be useful,
+ * Libav is distributed in the hope that it will be useful,
  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  * Lesser General Public License for more details.
  *
  * You should have received a copy of the GNU Lesser General Public
- * License along with FFmpeg; if not, write to the Free Software
+ * License along with Libav; if not, write to the Free Software
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
 
@@ -38,8 +38,9 @@
 #define CONTEXT_SIZE 32
 
 #define MAX_QUANT_TABLES 8
+#define MAX_CONTEXT_INPUTS 5
 
-extern const uint8_t ff_log2_run[32];
+extern const uint8_t ff_log2_run[41];
 
 static const int8_t quant3[256]={
  0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
@@ -216,6 +217,8 @@ typedef struct VlcState{
 } VlcState;
 
 typedef struct PlaneContext{
+    int16_t quant_table[MAX_CONTEXT_INPUTS][256];
+    int quant_table_index;
     int context_count;
     uint8_t (*state)[CONTEXT_SIZE];
     VlcState *vlc_state;
@@ -229,6 +232,8 @@ typedef struct FFV1Context{
     RangeCoder c;
     GetBitContext gb;
     PutBitContext pb;
+    uint64_t rc_stat[256][2];
+    uint64_t (*rc_stat2[MAX_QUANT_TABLES])[32][2];
     int version;
     int width, height;
     int chroma_h_shift, chroma_v_shift;
@@ -238,13 +243,15 @@ typedef struct FFV1Context{
     int plane_count;
     int ac;                              ///< 1=range coder <-> 0=golomb rice
     PlaneContext plane[MAX_PLANES];
-    int16_t quant_table[5][256];
-    int16_t quant_tables[MAX_QUANT_TABLES][5][256];
+    int16_t quant_table[MAX_CONTEXT_INPUTS][256];
+    int16_t quant_tables[MAX_QUANT_TABLES][MAX_CONTEXT_INPUTS][256];
     int context_count[MAX_QUANT_TABLES];
     uint8_t state_transition[256];
+    uint8_t (*initial_states[MAX_QUANT_TABLES])[32];
     int run_index;
     int colorspace;
-    int_fast16_t *sample_buffer;
+    int16_t *sample_buffer;
+    int gob_count;
 
     int quant_table_count;
 
@@ -272,7 +279,8 @@ static av_always_inline int fold(int diff, int bits){
     return diff;
 }
 
-static inline int predict(int_fast16_t *src, int_fast16_t *last){
+static inline int predict(int16_t *src, int16_t *last)
+{
     const int LT= last[-1];
     const int  T= last[ 0];
     const int L =  src[-1];
@@ -280,24 +288,77 @@ static inline int predict(int_fast16_t *src, int_fast16_t *last){
     return mid_pred(L, L + T - LT, T);
 }
 
-static inline int get_context(FFV1Context *f, int_fast16_t *src, int_fast16_t *last, int_fast16_t *last2){
+static inline int get_context(PlaneContext *p, int16_t *src,
+                              int16_t *last, int16_t *last2)
+{
     const int LT= last[-1];
     const int  T= last[ 0];
     const int RT= last[ 1];
     const int L =  src[-1];
 
-    if(f->quant_table[3][127]){
+    if(p->quant_table[3][127]){
         const int TT= last2[0];
         const int LL=  src[-2];
-        return f->quant_table[0][(L-LT) & 0xFF] + f->quant_table[1][(LT-T) & 0xFF] + f->quant_table[2][(T-RT) & 0xFF]
-              +f->quant_table[3][(LL-L) & 0xFF] + f->quant_table[4][(TT-T) & 0xFF];
+        return p->quant_table[0][(L-LT) & 0xFF] + p->quant_table[1][(LT-T) & 0xFF] + p->quant_table[2][(T-RT) & 0xFF]
+              +p->quant_table[3][(LL-L) & 0xFF] + p->quant_table[4][(TT-T) & 0xFF];
     }else
-        return f->quant_table[0][(L-LT) & 0xFF] + f->quant_table[1][(LT-T) & 0xFF] + f->quant_table[2][(T-RT) & 0xFF];
+        return p->quant_table[0][(L-LT) & 0xFF] + p->quant_table[1][(LT-T) & 0xFF] + p->quant_table[2][(T-RT) & 0xFF];
+}
+
+static void find_best_state(uint8_t best_state[256][256], const uint8_t one_state[256]){
+    int i,j,k,m;
+    double l2tab[256];
+
+    for(i=1; i<256; i++)
+        l2tab[i]= log2(i/256.0);
+
+    for(i=0; i<256; i++){
+        double best_len[256];
+        double p= i/256.0;
+
+        for(j=0; j<256; j++)
+            best_len[j]= 1<<30;
+
+        for(j=FFMAX(i-10,1); j<FFMIN(i+11,256); j++){
+            double occ[256]={0};
+            double len=0;
+            occ[j]=1.0;
+            for(k=0; k<256; k++){
+                double newocc[256]={0};
+                for(m=0; m<256; m++){
+                    if(occ[m]){
+                        len -=occ[m]*(     p *l2tab[    m]
+                                      + (1-p)*l2tab[256-m]);
+                    }
+                }
+                if(len < best_len[k]){
+                    best_len[k]= len;
+                    best_state[i][k]= j;
+                }
+                for(m=0; m<256; m++){
+                    if(occ[m]){
+                        newocc[    one_state[    m]] += occ[m]*   p ;
+                        newocc[256-one_state[256-m]] += occ[m]*(1-p);
+                    }
+                }
+                memcpy(occ, newocc, sizeof(occ));
+            }
+        }
+    }
 }
 
-static inline void put_symbol_inline(RangeCoder *c, uint8_t *state, int v, int is_signed){
+static av_always_inline av_flatten void put_symbol_inline(RangeCoder *c, uint8_t *state, int v, int is_signed, uint64_t rc_stat[256][2], uint64_t rc_stat2[32][2]){
     int i;
 
+#define put_rac(C,S,B) \
+do{\
+    if(rc_stat){\
+    rc_stat[*(S)][B]++;\
+        rc_stat2[(S)-state][B]++;\
+    }\
+    put_rac(C,S,B);\
+}while(0)
+
     if(v){
         const int a= FFABS(v);
         const int e= av_log2(a);
@@ -330,10 +391,11 @@ static inline void put_symbol_inline(RangeCoder *c, uint8_t *state, int v, int i
     }else{
         put_rac(c, state+0, 1);
     }
+#undef put_rac
 }
 
 static void av_noinline put_symbol(RangeCoder *c, uint8_t *state, int v, int is_signed){
-    put_symbol_inline(c, state, v, is_signed);
+    put_symbol_inline(c, state, v, is_signed, NULL, NULL);
 }
 
 static inline av_flatten int get_symbol_inline(RangeCoder *c, uint8_t *state, int is_signed){
@@ -447,7 +509,10 @@ static inline int get_vlc_symbol(GetBitContext *gb, VlcState * const state, int
 }
 
 #if CONFIG_FFV1_ENCODER
-static inline int encode_line(FFV1Context *s, int w, int_fast16_t *sample[2], int plane_index, int bits){
+static av_always_inline int encode_line(FFV1Context *s, int w,
+                                        int16_t *sample[2],
+                                        int plane_index, int bits)
+{
     PlaneContext * const p= &s->plane[plane_index];
     RangeCoder * const c= &s->c;
     int x;
@@ -470,7 +535,7 @@ static inline int encode_line(FFV1Context *s, int w, int_fast16_t *sample[2], in
     for(x=0; x<w; x++){
         int diff, context;
 
-        context= get_context(s, sample[0]+x, sample[1]+x, sample[2]+x);
+        context= get_context(p, sample[0]+x, sample[1]+x, sample[2]+x);
         diff= sample[0][x] - predict(sample[0]+x, sample[1]+x);
 
         if(context < 0){
@@ -481,7 +546,11 @@ static inline int encode_line(FFV1Context *s, int w, int_fast16_t *sample[2], in
         diff= fold(diff, bits);
 
         if(s->ac){
-            put_symbol_inline(c, p->state[context], diff, 1);
+            if(s->flags & CODEC_FLAG_PASS1){
+                put_symbol_inline(c, p->state[context], diff, 1, s->rc_stat, s->rc_stat2[p->quant_table_index][context]);
+            }else{
+                put_symbol_inline(c, p->state[context], diff, 1, NULL, NULL);
+            }
         }else{
             if(context == 0) run_mode=1;
 
@@ -528,7 +597,7 @@ static inline int encode_line(FFV1Context *s, int w, int_fast16_t *sample[2], in
 static void encode_plane(FFV1Context *s, uint8_t *src, int w, int h, int stride, int plane_index){
     int x,y,i;
     const int ring_size= s->avctx->context_model ? 3 : 2;
-    int_fast16_t *sample[3];
+    int16_t *sample[3];
     s->run_index=0;
 
     memset(s->sample_buffer, 0, ring_size*(w+6)*sizeof(*s->sample_buffer));
@@ -558,7 +627,7 @@ static void encode_plane(FFV1Context *s, uint8_t *src, int w, int h, int stride,
 static void encode_rgb_frame(FFV1Context *s, uint32_t *src, int w, int h, int stride){
     int x, y, p, i;
     const int ring_size= s->avctx->context_model ? 3 : 2;
-    int_fast16_t *sample[3][3];
+    int16_t *sample[3][3];
     s->run_index=0;
 
     memset(s->sample_buffer, 0, ring_size*3*(w+6)*sizeof(*s->sample_buffer));
@@ -609,7 +678,7 @@ static void write_quant_table(RangeCoder *c, int16_t *quant_table){
     put_symbol(c, state, i-last-1, 0);
 }
 
-static void write_quant_tables(RangeCoder *c, int16_t quant_table[5][256]){
+static void write_quant_tables(RangeCoder *c, int16_t quant_table[MAX_CONTEXT_INPUTS][256]){
     int i;
     for(i=0; i<5; i++)
         write_quant_table(c, quant_table[i]);
@@ -617,31 +686,41 @@ static void write_quant_tables(RangeCoder *c, int16_t quant_table[5][256]){
 
 static void write_header(FFV1Context *f){
     uint8_t state[CONTEXT_SIZE];
-    int i;
+    int i, j;
     RangeCoder * const c= &f->slice_context[0]->c;
 
     memset(state, 128, sizeof(state));
 
     if(f->version < 2){
-    put_symbol(c, state, f->version, 0);
-    put_symbol(c, state, f->ac, 0);
-    if(f->ac>1){
-        for(i=1; i<256; i++){
-            f->state_transition[i]=ver2_state[i];
-            put_symbol(c, state, ver2_state[i] - c->one_state[i], 1);
+        put_symbol(c, state, f->version, 0);
+        put_symbol(c, state, f->ac, 0);
+        if(f->ac>1){
+            for(i=1; i<256; i++){
+                put_symbol(c, state, f->state_transition[i] - c->one_state[i], 1);
+            }
         }
-    }
-    put_symbol(c, state, f->colorspace, 0); //YUV cs type
-    if(f->version>0)
-        put_symbol(c, state, f->avctx->bits_per_raw_sample, 0);
-    put_rac(c, state, 1); //chroma planes
-        put_symbol(c, state, f->chroma_h_shift, 0);
-        put_symbol(c, state, f->chroma_v_shift, 0);
-    put_rac(c, state, 0); //no transparency plane
-
-    write_quant_tables(c, f->quant_table);
+        put_symbol(c, state, f->colorspace, 0); //YUV cs type
+        if(f->version>0)
+            put_symbol(c, state, f->avctx->bits_per_raw_sample, 0);
+        put_rac(c, state, 1); //chroma planes
+            put_symbol(c, state, f->chroma_h_shift, 0);
+            put_symbol(c, state, f->chroma_v_shift, 0);
+        put_rac(c, state, 0); //no transparency plane
+
+        write_quant_tables(c, f->quant_table);
     }else{
-        put_symbol(c, state, f->avctx->context_model, 0);
+        put_symbol(c, state, f->slice_count, 0);
+        for(i=0; i<f->slice_count; i++){
+            FFV1Context *fs= f->slice_context[i];
+            put_symbol(c, state, (fs->slice_x     +1)*f->num_h_slices / f->width   , 0);
+            put_symbol(c, state, (fs->slice_y     +1)*f->num_v_slices / f->height  , 0);
+            put_symbol(c, state, (fs->slice_width +1)*f->num_h_slices / f->width -1, 0);
+            put_symbol(c, state, (fs->slice_height+1)*f->num_v_slices / f->height-1, 0);
+            for(j=0; j<f->plane_count; j++){
+                put_symbol(c, state, f->plane[j].quant_table_index, 0);
+                av_assert0(f->plane[j].quant_table_index == f->avctx->context_model);
+            }
+        }
     }
 }
 #endif /* CONFIG_FFV1_ENCODER */
@@ -666,43 +745,6 @@ static av_cold int common_init(AVCodecContext *avctx){
     return 0;
 }
 
-#if CONFIG_FFV1_ENCODER
-static int write_extra_header(FFV1Context *f){
-    RangeCoder * const c= &f->c;
-    uint8_t state[CONTEXT_SIZE];
-    int i;
-    memset(state, 128, sizeof(state));
-
-    f->avctx->extradata= av_malloc(f->avctx->extradata_size= 10000);
-    ff_init_range_encoder(c, f->avctx->extradata, f->avctx->extradata_size);
-    ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
-
-    put_symbol(c, state, f->version, 0);
-    put_symbol(c, state, f->ac, 0);
-    if(f->ac>1){
-        for(i=1; i<256; i++){
-            f->state_transition[i]=ver2_state[i];
-            put_symbol(c, state, ver2_state[i] - c->one_state[i], 1);
-        }
-    }
-    put_symbol(c, state, f->colorspace, 0); //YUV cs type
-    put_symbol(c, state, f->avctx->bits_per_raw_sample, 0);
-    put_rac(c, state, 1); //chroma planes
-        put_symbol(c, state, f->chroma_h_shift, 0);
-        put_symbol(c, state, f->chroma_v_shift, 0);
-    put_rac(c, state, 0); //no transparency plane
-    put_symbol(c, state, f->num_h_slices-1, 0);
-    put_symbol(c, state, f->num_v_slices-1, 0);
-
-    put_symbol(c, state, f->quant_table_count, 0);
-    for(i=0; i<f->quant_table_count; i++)
-        write_quant_tables(c, f->quant_tables[i]);
-
-    f->avctx->extradata_size= ff_rac_terminate(c);
-
-    return 0;
-}
-
 static int init_slice_state(FFV1Context *f){
     int i, j;
 
@@ -749,29 +791,145 @@ static av_cold int init_slice_contexts(FFV1Context *f){
         int sye= f->avctx->height*(sy+1) / f->num_v_slices;
         f->slice_context[i]= fs;
         memcpy(fs, f, sizeof(*fs));
+        memset(fs->rc_stat2, 0, sizeof(fs->rc_stat2));
 
         fs->slice_width = sxe - sxs;
         fs->slice_height= sye - sys;
         fs->slice_x     = sxs;
         fs->slice_y     = sys;
 
-        fs->sample_buffer = av_malloc(6 * (fs->slice_width+6) * sizeof(*fs->sample_buffer));
+        fs->sample_buffer = av_malloc(9 * (fs->width+6) * sizeof(*fs->sample_buffer));
         if (!fs->sample_buffer)
             return AVERROR(ENOMEM);
     }
     return 0;
 }
 
+static int allocate_initial_states(FFV1Context *f){
+    int i;
+
+    for(i=0; i<f->quant_table_count; i++){
+        f->initial_states[i]= av_malloc(f->context_count[i]*sizeof(*f->initial_states[i]));
+        if(!f->initial_states[i])
+            return AVERROR(ENOMEM);
+        memset(f->initial_states[i], 128, f->context_count[i]*sizeof(*f->initial_states[i]));
+    }
+    return 0;
+}
+
+#if CONFIG_FFV1_ENCODER
+static int write_extra_header(FFV1Context *f){
+    RangeCoder * const c= &f->c;
+    uint8_t state[CONTEXT_SIZE];
+    int i, j, k;
+    uint8_t state2[32][CONTEXT_SIZE];
+
+    memset(state2, 128, sizeof(state2));
+    memset(state, 128, sizeof(state));
+
+    f->avctx->extradata= av_malloc(f->avctx->extradata_size= 10000 + (11*11*5*5*5+11*11*11)*32);
+    ff_init_range_encoder(c, f->avctx->extradata, f->avctx->extradata_size);
+    ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
+
+    put_symbol(c, state, f->version, 0);
+    put_symbol(c, state, f->ac, 0);
+    if(f->ac>1){
+        for(i=1; i<256; i++){
+            put_symbol(c, state, f->state_transition[i] - c->one_state[i], 1);
+        }
+    }
+    put_symbol(c, state, f->colorspace, 0); //YUV cs type
+    put_symbol(c, state, f->avctx->bits_per_raw_sample, 0);
+    put_rac(c, state, 1); //chroma planes
+        put_symbol(c, state, f->chroma_h_shift, 0);
+        put_symbol(c, state, f->chroma_v_shift, 0);
+    put_rac(c, state, 0); //no transparency plane
+    put_symbol(c, state, f->num_h_slices-1, 0);
+    put_symbol(c, state, f->num_v_slices-1, 0);
+
+    put_symbol(c, state, f->quant_table_count, 0);
+    for(i=0; i<f->quant_table_count; i++)
+        write_quant_tables(c, f->quant_tables[i]);
+
+    for(i=0; i<f->quant_table_count; i++){
+        for(j=0; j<f->context_count[i]*CONTEXT_SIZE; j++)
+            if(f->initial_states[i] && f->initial_states[i][0][j] != 128)
+                break;
+        if(j<f->context_count[i]*CONTEXT_SIZE){
+            put_rac(c, state, 1);
+            for(j=0; j<f->context_count[i]; j++){
+                for(k=0; k<CONTEXT_SIZE; k++){
+                    int pred= j ? f->initial_states[i][j-1][k] : 128;
+                    put_symbol(c, state2[k], (int8_t)(f->initial_states[i][j][k]-pred), 1);
+                }
+            }
+        }else{
+            put_rac(c, state, 0);
+        }
+    }
+
+    f->avctx->extradata_size= ff_rac_terminate(c);
+
+    return 0;
+}
+
+static int sort_stt(FFV1Context *s, uint8_t stt[256]){
+    int i,i2,changed,print=0;
+
+    do{
+        changed=0;
+        for(i=12; i<244; i++){
+            for(i2=i+1; i2<245 && i2<i+4; i2++){
+#define COST(old, new) \
+    s->rc_stat[old][0]*-log2((256-(new))/256.0)\
+   +s->rc_stat[old][1]*-log2(     (new) /256.0)
+
+#define COST2(old, new) \
+    COST(old, new)\
+   +COST(256-(old), 256-(new))
+
+                double size0= COST2(i, i ) + COST2(i2, i2);
+                double sizeX= COST2(i, i2) + COST2(i2, i );
+                if(sizeX < size0 && i!=128 && i2!=128){
+                    int j;
+                    FFSWAP(int, stt[    i], stt[    i2]);
+                    FFSWAP(int, s->rc_stat[i    ][0],s->rc_stat[    i2][0]);
+                    FFSWAP(int, s->rc_stat[i    ][1],s->rc_stat[    i2][1]);
+                    if(i != 256-i2){
+                        FFSWAP(int, stt[256-i], stt[256-i2]);
+                        FFSWAP(int, s->rc_stat[256-i][0],s->rc_stat[256-i2][0]);
+                        FFSWAP(int, s->rc_stat[256-i][1],s->rc_stat[256-i2][1]);
+                    }
+                    for(j=1; j<256; j++){
+                        if     (stt[j] == i ) stt[j] = i2;
+                        else if(stt[j] == i2) stt[j] = i ;
+                        if(i != 256-i2){
+                            if     (stt[256-j] == 256-i ) stt[256-j] = 256-i2;
+                            else if(stt[256-j] == 256-i2) stt[256-j] = 256-i ;
+                        }
+                    }
+                    print=changed=1;
+                }
+            }
+        }
+    }while(changed);
+    return print;
+}
+
 static av_cold int encode_init(AVCodecContext *avctx)
 {
     FFV1Context *s = avctx->priv_data;
-    int i;
+    int i, j, k, m;
 
     common_init(avctx);
 
     s->version=0;
     s->ac= avctx->coder_type ? 2:0;
 
+    if(s->ac>1)
+        for(i=1; i<256; i++)
+            s->state_transition[i]=ver2_state[i];
+
     s->plane_count=2;
     for(i=0; i<256; i++){
         s->quant_table_count=2;
@@ -795,18 +953,21 @@ static av_cold int encode_init(AVCodecContext *avctx)
             s->quant_tables[1][4][i]= 5*5*11*11*quant5_10bit[i];
         }
     }
+    s->context_count[0]= (11*11*11+1)/2;
+    s->context_count[1]= (11*11*5*5*5+1)/2;
     memcpy(s->quant_table, s->quant_tables[avctx->context_model], sizeof(s->quant_table));
 
     for(i=0; i<s->plane_count; i++){
         PlaneContext * const p= &s->plane[i];
 
-        if(avctx->context_model==0){
-            p->context_count= (11*11*11+1)/2;
-        }else{
-            p->context_count= (11*11*5*5*5+1)/2;
-        }
+        memcpy(p->quant_table, s->quant_table, sizeof(p->quant_table));
+        p->quant_table_index= avctx->context_model;
+        p->context_count= s->context_count[p->quant_table_index];
     }
 
+    if(allocate_initial_states(s) < 0)
+        return AVERROR(ENOMEM);
+
     avctx->coded_frame= &s->picture;
     switch(avctx->pix_fmt){
     case PIX_FMT_YUV444P16:
@@ -839,6 +1000,72 @@ static av_cold int encode_init(AVCodecContext *avctx)
 
     s->picture_number=0;
 
+    if(avctx->flags & (CODEC_FLAG_PASS1|CODEC_FLAG_PASS2)){
+        for(i=0; i<s->quant_table_count; i++){
+            s->rc_stat2[i]= av_mallocz(s->context_count[i]*sizeof(*s->rc_stat2[i]));
+            if(!s->rc_stat2[i])
+                return AVERROR(ENOMEM);
+        }
+    }
+    if(avctx->stats_in){
+        char *p= avctx->stats_in;
+        uint8_t best_state[256][256];
+        int gob_count=0;
+        char *next;
+
+        av_assert0(s->version>=2);
+
+        for(;;){
+            for(j=0; j<256; j++){
+                for(i=0; i<2; i++){
+                    s->rc_stat[j][i]= strtol(p, &next, 0);
+                    if(next==p){
+                        av_log(avctx, AV_LOG_ERROR, "2Pass file invalid at %d %d [%s]\n", j,i,p);
+                        return -1;
+                    }
+                    p=next;
+                }
+            }
+            for(i=0; i<s->quant_table_count; i++){
+                for(j=0; j<s->context_count[i]; j++){
+                    for(k=0; k<32; k++){
+                        for(m=0; m<2; m++){
+                            s->rc_stat2[i][j][k][m]= strtol(p, &next, 0);
+                            if(next==p){
+                                av_log(avctx, AV_LOG_ERROR, "2Pass file invalid at %d %d %d %d [%s]\n", i,j,k,m,p);
+                                return -1;
+                            }
+                            p=next;
+                        }
+                    }
+                }
+            }
+            gob_count= strtol(p, &next, 0);
+            if(next==p || gob_count <0){
+                av_log(avctx, AV_LOG_ERROR, "2Pass file invalid\n");
+                return -1;
+            }
+            p=next;
+            while(*p=='\n' || *p==' ') p++;
+            if(p[0]==0) break;
+        }
+        sort_stt(s, s->state_transition);
+
+        find_best_state(best_state, s->state_transition);
+
+        for(i=0; i<s->quant_table_count; i++){
+            for(j=0; j<s->context_count[i]; j++){
+                for(k=0; k<32; k++){
+                    double p= 128;
+                    if(s->rc_stat2[i][j][k][0]+s->rc_stat2[i][j][k][1]){
+                        p=256.0*s->rc_stat2[i][j][k][1] / (s->rc_stat2[i][j][k][0]+s->rc_stat2[i][j][k][1]);
+                    }
+                    s->initial_states[i][j][k]= best_state[av_clip(round(p), 1, 255)][av_clip((s->rc_stat2[i][j][k][0]+s->rc_stat2[i][j][k][1])/gob_count, 0, 255)];
+                }
+            }
+        }
+    }
+
     if(s->version>1){
         s->num_h_slices=2;
         s->num_v_slices=2;
@@ -850,6 +1077,20 @@ static av_cold int encode_init(AVCodecContext *avctx)
     if(init_slice_state(s) < 0)
         return -1;
 
+#define STATS_OUT_SIZE 1024*1024*6
+    if(avctx->flags & CODEC_FLAG_PASS1){
+        avctx->stats_out= av_mallocz(STATS_OUT_SIZE);
+        for(i=0; i<s->quant_table_count; i++){
+            for(j=0; j<s->slice_count; j++){
+                FFV1Context *sf= s->slice_context[j];
+                av_assert0(!sf->rc_stat2[i]);
+                sf->rc_stat2[i]= av_mallocz(s->context_count[i]*sizeof(*sf->rc_stat2[i]));
+                if(!sf->rc_stat2[i])
+                    return AVERROR(ENOMEM);
+            }
+        }
+    }
+
     return 0;
 }
 #endif /* CONFIG_FFV1_ENCODER */
@@ -860,24 +1101,27 @@ static void clear_state(FFV1Context *f){
 
     for(si=0; si<f->slice_count; si++){
         FFV1Context *fs= f->slice_context[si];
-    for(i=0; i<f->plane_count; i++){
-        PlaneContext *p= &fs->plane[i];
+        for(i=0; i<f->plane_count; i++){
+            PlaneContext *p= &fs->plane[i];
 
-        p->interlace_bit_state[0]= 128;
-        p->interlace_bit_state[1]= 128;
+            p->interlace_bit_state[0]= 128;
+            p->interlace_bit_state[1]= 128;
 
-        for(j=0; j<p->context_count; j++){
             if(fs->ac){
-                memset(p->state[j], 128, sizeof(uint8_t)*CONTEXT_SIZE);
+                if(f->initial_states[p->quant_table_index]){
+                    memcpy(p->state, f->initial_states[p->quant_table_index], CONTEXT_SIZE*p->context_count);
+                }else
+                memset(p->state, 128, CONTEXT_SIZE*p->context_count);
             }else{
-                p->vlc_state[j].drift= 0;
-                p->vlc_state[j].error_sum= 4; //FFMAX((RANGE + 32)/64, 2);
-                p->vlc_state[j].bias= 0;
-                p->vlc_state[j].count= 1;
+            for(j=0; j<p->context_count; j++){
+                    p->vlc_state[j].drift= 0;
+                    p->vlc_state[j].error_sum= 4; //FFMAX((RANGE + 32)/64, 2);
+                    p->vlc_state[j].bias= 0;
+                    p->vlc_state[j].count= 1;
+            }
             }
         }
     }
-    }
 }
 
 #if CONFIG_FFV1_ENCODER
@@ -922,11 +1166,12 @@ static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size,
     ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
 
     *p = *pict;
-    p->pict_type= FF_I_TYPE;
+    p->pict_type= AV_PICTURE_TYPE_I;
 
     if(avctx->gop_size==0 || f->picture_number % avctx->gop_size == 0){
         put_rac(c, &keystate, 1);
         p->key_frame= 1;
+        f->gob_count++;
         write_header(f);
         clear_state(f);
     }else{
@@ -983,6 +1228,49 @@ static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size,
         buf_p += bytes;
     }
 
+    if((avctx->flags&CODEC_FLAG_PASS1) && (f->picture_number&31)==0){
+        int j, k, m;
+        char *p= avctx->stats_out;
+        char *end= p + STATS_OUT_SIZE;
+
+        memset(f->rc_stat, 0, sizeof(f->rc_stat));
+        for(i=0; i<f->quant_table_count; i++)
+            memset(f->rc_stat2[i], 0, f->context_count[i]*sizeof(*f->rc_stat2[i]));
+
+        for(j=0; j<f->slice_count; j++){
+            FFV1Context *fs= f->slice_context[j];
+            for(i=0; i<256; i++){
+                f->rc_stat[i][0] += fs->rc_stat[i][0];
+                f->rc_stat[i][1] += fs->rc_stat[i][1];
+            }
+            for(i=0; i<f->quant_table_count; i++){
+                for(k=0; k<f->context_count[i]; k++){
+                    for(m=0; m<32; m++){
+                        f->rc_stat2[i][k][m][0] += fs->rc_stat2[i][k][m][0];
+                        f->rc_stat2[i][k][m][1] += fs->rc_stat2[i][k][m][1];
+                    }
+                }
+            }
+        }
+
+        for(j=0; j<256; j++){
+            snprintf(p, end-p, "%"PRIu64" %"PRIu64" ", f->rc_stat[j][0], f->rc_stat[j][1]);
+            p+= strlen(p);
+        }
+        snprintf(p, end-p, "\n");
+
+        for(i=0; i<f->quant_table_count; i++){
+            for(j=0; j<f->context_count[i]; j++){
+                for(m=0; m<32; m++){
+                    snprintf(p, end-p, "%"PRIu64" %"PRIu64" ", f->rc_stat2[i][j][m][0], f->rc_stat2[i][j][m][1]);
+                    p+= strlen(p);
+                }
+            }
+        }
+        snprintf(p, end-p, "%d\n", f->gob_count);
+    } else if(avctx->flags&CODEC_FLAG_PASS1)
+        avctx->stats_out[0] = '\0';
+
     f->picture_number++;
     return buf_p-buf;
 }
@@ -992,21 +1280,41 @@ static av_cold int common_end(AVCodecContext *avctx){
     FFV1Context *s = avctx->priv_data;
     int i, j;
 
+    if (avctx->codec->decode && s->picture.data[0])
+        avctx->release_buffer(avctx, &s->picture);
+
     for(j=0; j<s->slice_count; j++){
         FFV1Context *fs= s->slice_context[j];
-    for(i=0; i<s->plane_count; i++){
-        PlaneContext *p= &fs->plane[i];
+        for(i=0; i<s->plane_count; i++){
+            PlaneContext *p= &fs->plane[i];
 
-        av_freep(&p->state);
-        av_freep(&p->vlc_state);
-    }
+            av_freep(&p->state);
+            av_freep(&p->vlc_state);
+        }
         av_freep(&fs->sample_buffer);
     }
 
+    av_freep(&avctx->stats_out);
+    for(j=0; j<s->quant_table_count; j++){
+        av_freep(&s->initial_states[j]);
+        for(i=0; i<s->slice_count; i++){
+            FFV1Context *sf= s->slice_context[i];
+            av_freep(&sf->rc_stat2[j]);
+        }
+        av_freep(&s->rc_stat2[j]);
+    }
+
+    for(i=0; i<s->slice_count; i++){
+        av_freep(&s->slice_context[i]);
+    }
+
     return 0;
 }
 
-static av_always_inline void decode_line(FFV1Context *s, int w, int_fast16_t *sample[2], int plane_index, int bits){
+static av_always_inline void decode_line(FFV1Context *s, int w,
+                                         int16_t *sample[2],
+                                         int plane_index, int bits)
+{
     PlaneContext * const p= &s->plane[plane_index];
     RangeCoder * const c= &s->c;
     int x;
@@ -1017,13 +1325,14 @@ static av_always_inline void decode_line(FFV1Context *s, int w, int_fast16_t *sa
     for(x=0; x<w; x++){
         int diff, context, sign;
 
-        context= get_context(s, sample[1] + x, sample[0] + x, sample[1] + x);
+        context= get_context(p, sample[1] + x, sample[0] + x, sample[1] + x);
         if(context < 0){
             context= -context;
             sign=1;
         }else
             sign=0;
 
+        av_assert2(context < p->context_count);
 
         if(s->ac){
             diff= get_symbol_inline(c, p->state[context], 1);
@@ -1065,7 +1374,7 @@ static av_always_inline void decode_line(FFV1Context *s, int w, int_fast16_t *sa
 
 static void decode_plane(FFV1Context *s, uint8_t *src, int w, int h, int stride, int plane_index){
     int x, y;
-    int_fast16_t *sample[2];
+    int16_t *sample[2];
     sample[0]=s->sample_buffer    +3;
     sample[1]=s->sample_buffer+w+6+3;
 
@@ -1074,7 +1383,7 @@ static void decode_plane(FFV1Context *s, uint8_t *src, int w, int h, int stride,
     memset(s->sample_buffer, 0, 2*(w+6)*sizeof(*s->sample_buffer));
 
     for(y=0; y<h; y++){
-        int_fast16_t *temp= sample[0]; //FIXME try a normal buffer
+        int16_t *temp = sample[0]; //FIXME try a normal buffer
 
         sample[0]= sample[1];
         sample[1]= temp;
@@ -1100,7 +1409,7 @@ static void decode_plane(FFV1Context *s, uint8_t *src, int w, int h, int stride,
 
 static void decode_rgb_frame(FFV1Context *s, uint32_t *src, int w, int h, int stride){
     int x, y, p;
-    int_fast16_t *sample[3][2];
+    int16_t *sample[3][2];
     for(x=0; x<3; x++){
         sample[x][0] = s->sample_buffer +  x*2   *(w+6) + 3;
         sample[x][1] = s->sample_buffer + (x*2+1)*(w+6) + 3;
@@ -1112,7 +1421,7 @@ static void decode_rgb_frame(FFV1Context *s, uint32_t *src, int w, int h, int st
 
     for(y=0; y<h; y++){
         for(p=0; p<3; p++){
-            int_fast16_t *temp= sample[p][0]; //FIXME try a normal buffer
+            int16_t *temp = sample[p][0]; //FIXME try a normal buffer
 
             sample[p][0]= sample[p][1];
             sample[p][1]= temp;
@@ -1196,7 +1505,7 @@ static int read_quant_table(RangeCoder *c, int16_t *quant_table, int scale){
     return 2*v - 1;
 }
 
-static int read_quant_tables(RangeCoder *c, int16_t quant_table[5][256]){
+static int read_quant_tables(RangeCoder *c, int16_t quant_table[MAX_CONTEXT_INPUTS][256]){
     int i;
     int context_count=1;
 
@@ -1212,8 +1521,10 @@ static int read_quant_tables(RangeCoder *c, int16_t quant_table[5][256]){
 static int read_extra_header(FFV1Context *f){
     RangeCoder * const c= &f->c;
     uint8_t state[CONTEXT_SIZE];
-    int i;
+    int i, j, k;
+    uint8_t state2[32][CONTEXT_SIZE];
 
+    memset(state2, 128, sizeof(state2));
     memset(state, 128, sizeof(state));
 
     ff_init_range_decoder(c, f->avctx->extradata, f->avctx->extradata_size);
@@ -1235,7 +1546,7 @@ static int read_extra_header(FFV1Context *f){
     f->plane_count= 2;
     f->num_h_slices= 1 + get_symbol(c, state, 0);
     f->num_v_slices= 1 + get_symbol(c, state, 0);
-    if(f->num_h_slices > 256U || f->num_v_slices > 256U || f->num_h_slices*f->num_v_slices > MAX_SLICES){
+    if(f->num_h_slices > (unsigned)f->width || f->num_v_slices > (unsigned)f->height){
         av_log(f->avctx, AV_LOG_ERROR, "too many slices\n");
         return -1;
     }
@@ -1250,6 +1561,20 @@ static int read_extra_header(FFV1Context *f){
         }
     }
 
+    if(allocate_initial_states(f) < 0)
+        return AVERROR(ENOMEM);
+
+    for(i=0; i<f->quant_table_count; i++){
+        if(get_rac(c, state)){
+            for(j=0; j<f->context_count[i]; j++){
+                for(k=0; k<CONTEXT_SIZE; k++){
+                    int pred= j ? f->initial_states[i][j-1][k] : 128;
+                    f->initial_states[i][j][k]= (pred+get_symbol(c, state2[k], 1))&0xFF;
+                }
+            }
+        }
+    }
+
     return 0;
 }
 
@@ -1261,21 +1586,21 @@ static int read_header(FFV1Context *f){
     memset(state, 128, sizeof(state));
 
     if(f->version < 2){
-    f->version= get_symbol(c, state, 0);
-    f->ac= f->avctx->coder_type= get_symbol(c, state, 0);
-    if(f->ac>1){
-        for(i=1; i<256; i++){
-            f->state_transition[i]= get_symbol(c, state, 1) + c->one_state[i];
+        f->version= get_symbol(c, state, 0);
+        f->ac= f->avctx->coder_type= get_symbol(c, state, 0);
+        if(f->ac>1){
+            for(i=1; i<256; i++){
+                f->state_transition[i]= get_symbol(c, state, 1) + c->one_state[i];
+            }
         }
-    }
-    f->colorspace= get_symbol(c, state, 0); //YUV cs type
-    if(f->version>0)
-        f->avctx->bits_per_raw_sample= get_symbol(c, state, 0);
-    get_rac(c, state); //no chroma = false
-    f->chroma_h_shift= get_symbol(c, state, 0);
-    f->chroma_v_shift= get_symbol(c, state, 0);
-    get_rac(c, state); //transparency plane
-    f->plane_count= 2;
+        f->colorspace= get_symbol(c, state, 0); //YUV cs type
+        if(f->version>0)
+            f->avctx->bits_per_raw_sample= get_symbol(c, state, 0);
+        get_rac(c, state); //no chroma = false
+        f->chroma_h_shift= get_symbol(c, state, 0);
+        f->chroma_v_shift= get_symbol(c, state, 0);
+        get_rac(c, state); //transparency plane
+        f->plane_count= 2;
     }
 
     if(f->colorspace==0){
@@ -1313,34 +1638,60 @@ static int read_header(FFV1Context *f){
 
 //printf("%d %d %d\n", f->chroma_h_shift, f->chroma_v_shift,f->avctx->pix_fmt);
     if(f->version < 2){
-    context_count= read_quant_tables(c, f->quant_table);
-    if(context_count < 0){
-            av_log(f->avctx, AV_LOG_ERROR, "read_quant_table error\n");
-            return -1;
+        context_count= read_quant_tables(c, f->quant_table);
+        if(context_count < 0){
+                av_log(f->avctx, AV_LOG_ERROR, "read_quant_table error\n");
+                return -1;
         }
-
     }else{
-        i=get_symbol(c, state, 0);
-        if(i > (unsigned)f->quant_table_count){
-            av_log(f->avctx, AV_LOG_ERROR, "quant_table_index out of range\n");
+        f->slice_count= get_symbol(c, state, 0);
+        if(f->slice_count > (unsigned)MAX_SLICES)
             return -1;
-        }
-        memcpy(f->quant_table, f->quant_tables[i], sizeof(f->quant_table));
-        context_count= f->context_count[i];
     }
+
     for(j=0; j<f->slice_count; j++){
         FFV1Context *fs= f->slice_context[j];
-        memcpy(fs->quant_table, f->quant_table, sizeof(fs->quant_table));
         fs->ac= f->ac;
-    for(i=0; i<f->plane_count; i++){
-        PlaneContext * const p= &fs->plane[i];
 
-        if(p->context_count < context_count){
-            av_freep(&p->state);
-            av_freep(&p->vlc_state);
+        if(f->version >= 2){
+            fs->slice_x     = get_symbol(c, state, 0)   *f->width ;
+            fs->slice_y     = get_symbol(c, state, 0)   *f->height;
+            fs->slice_width =(get_symbol(c, state, 0)+1)*f->width  + fs->slice_x;
+            fs->slice_height=(get_symbol(c, state, 0)+1)*f->height + fs->slice_y;
+
+            fs->slice_x /= f->num_h_slices;
+            fs->slice_y /= f->num_v_slices;
+            fs->slice_width  = fs->slice_width /f->num_h_slices - fs->slice_x;
+            fs->slice_height = fs->slice_height/f->num_v_slices - fs->slice_y;
+            if((unsigned)fs->slice_width > f->width || (unsigned)fs->slice_height > f->height)
+                return -1;
+            if(    (unsigned)fs->slice_x + (uint64_t)fs->slice_width  > f->width
+                || (unsigned)fs->slice_y + (uint64_t)fs->slice_height > f->height)
+                return -1;
+        }
+
+        for(i=0; i<f->plane_count; i++){
+            PlaneContext * const p= &fs->plane[i];
+
+            if(f->version >= 2){
+                int idx=get_symbol(c, state, 0);
+                if(idx > (unsigned)f->quant_table_count){
+                    av_log(f->avctx, AV_LOG_ERROR, "quant_table_index out of range\n");
+                    return -1;
+                }
+                p->quant_table_index= idx;
+                memcpy(p->quant_table, f->quant_tables[idx], sizeof(p->quant_table));
+                context_count= f->context_count[idx];
+            }else{
+                memcpy(p->quant_table, f->quant_table, sizeof(p->quant_table));
+            }
+
+            if(p->context_count < context_count){
+                av_freep(&p->state);
+                av_freep(&p->vlc_state);
+            }
+            p->context_count= context_count;
         }
-        p->context_count= context_count;
-    }
     }
 
     return 0;
@@ -1373,11 +1724,15 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, AVPac
 
     AVFrame *picture = data;
 
+    /* release previously stored data */
+    if (p->data[0])
+        avctx->release_buffer(avctx, p);
+
     ff_init_range_decoder(c, buf, buf_size);
     ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
 
 
-    p->pict_type= FF_I_TYPE; //FIXME I vs. P
+    p->pict_type= AV_PICTURE_TYPE_I; //FIXME I vs. P
     if(get_rac(c, &keystate)){
         p->key_frame= 1;
         if(read_header(f) < 0)
@@ -1435,15 +1790,12 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, AVPac
     f->picture_number++;
 
     *picture= *p;
-
-    avctx->release_buffer(avctx, p); //FIXME
-
     *data_size = sizeof(AVFrame);
 
     return buf_size;
 }
 
-AVCodec ffv1_decoder = {
+AVCodec ff_ffv1_decoder = {
     "ffv1",
     AVMEDIA_TYPE_VIDEO,
     CODEC_ID_FFV1,
@@ -1452,13 +1804,13 @@ AVCodec ffv1_decoder = {
     NULL,
     common_end,
     decode_frame,
-    CODEC_CAP_DR1 /*| CODEC_CAP_DRAW_HORIZ_BAND*/,
+    CODEC_CAP_DR1 /*| CODEC_CAP_DRAW_HORIZ_BAND*/ | CODEC_CAP_SLICE_THREADS,
     NULL,
     .long_name= NULL_IF_CONFIG_SMALL("FFmpeg video codec #1"),
 };
 
 #if CONFIG_FFV1_ENCODER
-AVCodec ffv1_encoder = {
+AVCodec ff_ffv1_encoder = {
     "ffv1",
     AVMEDIA_TYPE_VIDEO,
     CODEC_ID_FFV1,
@@ -1466,6 +1818,7 @@ AVCodec ffv1_encoder = {
     encode_init,
     encode_frame,
     common_end,
+    .capabilities = CODEC_CAP_SLICE_THREADS,
     .pix_fmts= (const enum PixelFormat[]){PIX_FMT_YUV420P, PIX_FMT_YUV444P, PIX_FMT_YUV422P, PIX_FMT_YUV411P, PIX_FMT_YUV410P, PIX_FMT_RGB32, PIX_FMT_YUV420P16, PIX_FMT_YUV422P16, PIX_FMT_YUV444P16, PIX_FMT_NONE},
     .long_name= NULL_IF_CONFIG_SMALL("FFmpeg video codec #1"),
 };