X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libavcodec%2Fsnow.c;h=dc50941fac5655113451e0f7db2e74a2c72d0b60;hb=226d6947bc6e6592eabbcb7b401bf164be5bfe6e;hp=d9f58e2d028b5076ce5a488924aca5780b7b72dc;hpb=4f59b684a6470d4dc337cc7abca662347039c8c6;p=ffmpeg diff --git a/libavcodec/snow.c b/libavcodec/snow.c index d9f58e2d028..dc50941fac5 100644 --- a/libavcodec/snow.c +++ b/libavcodec/snow.c @@ -1,41 +1,35 @@ /* * Copyright (C) 2004 Michael Niedermayer * - * This library is free software; you can redistribute it and/or + * This file is part of FFmpeg. + * + * FFmpeg 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 of the License, or (at your option) any later version. + * version 2.1 of the License, or (at your option) any later version. * - * This library is distributed in the hope that it will be useful, + * FFmpeg 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 this library; if not, write to the Free Software + * License along with FFmpeg; if not, write to the Free Software * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA */ #include "avcodec.h" #include "common.h" #include "dsputil.h" +#include "snow.h" #include "rangecoder.h" -#define MID_STATE 128 #include "mpegvideo.h" #undef NDEBUG #include -#define MAX_DECOMPOSITIONS 8 -#define MAX_PLANES 4 -#define DWTELEM int -#define QSHIFT 5 -#define QROOT (1<data_count - 1; i >= 0; i--) { assert(buf->data_stack[i]); - av_free(buf->data_stack[i]); + av_freep(&buf->data_stack[i]); } assert(buf->data_stack); - av_free(buf->data_stack); + av_freep(&buf->data_stack); assert(buf->line); - av_free(buf->line); + av_freep(&buf->line); } #ifdef __sgi @@ -600,7 +595,7 @@ static inline void put_symbol(RangeCoder *c, uint8_t *state, int v, int is_signe int i; if(v){ - const int a= ABS(v); + const int a= FFABS(v); const int e= av_log2(a); #if 1 const int el= FFMIN(e, 10); @@ -719,7 +714,7 @@ static inline int get_symbol2(RangeCoder *c, uint8_t *state, int log2){ return v; } -static always_inline void lift(DWTELEM *dst, DWTELEM *src, DWTELEM *ref, int dst_step, int src_step, int ref_step, int width, int mul, int add, int shift, int highpass, int inverse){ +static av_always_inline void lift(DWTELEM *dst, DWTELEM *src, DWTELEM *ref, int dst_step, int src_step, int ref_step, int width, int mul, int add, int shift, int highpass, int inverse){ const int mirror_left= !highpass; const int mirror_right= (width&1) ^ highpass; const int w= (width>>1) - 1 + (highpass & width); @@ -741,7 +736,8 @@ static always_inline void lift(DWTELEM *dst, DWTELEM *src, DWTELEM *ref, int dst } } -static always_inline void lift5(DWTELEM *dst, DWTELEM *src, DWTELEM *ref, int dst_step, int src_step, int ref_step, int width, int mul, int add, int shift, int highpass, int inverse){ +#ifndef lift5 +static av_always_inline void lift5(DWTELEM *dst, DWTELEM *src, DWTELEM *ref, int dst_step, int src_step, int ref_step, int width, int mul, int add, int shift, int highpass, int inverse){ const int mirror_left= !highpass; const int mirror_right= (width&1) ^ highpass; const int w= (width>>1) - 1 + (highpass & width); @@ -770,8 +766,10 @@ static always_inline void lift5(DWTELEM *dst, DWTELEM *src, DWTELEM *ref, int ds dst[w*dst_step] = LIFT(src[w*src_step], ((r+add)>>shift), inverse); } } +#endif -static always_inline void liftS(DWTELEM *dst, DWTELEM *src, DWTELEM *ref, int dst_step, int src_step, int ref_step, int width, int mul, int add, int shift, int highpass, int inverse){ +#ifndef liftS +static av_always_inline void liftS(DWTELEM *dst, DWTELEM *src, DWTELEM *ref, int dst_step, int src_step, int ref_step, int width, int mul, int add, int shift, int highpass, int inverse){ const int mirror_left= !highpass; const int mirror_right= (width&1) ^ highpass; const int w= (width>>1) - 1 + (highpass & width); @@ -793,6 +791,7 @@ static always_inline void liftS(DWTELEM *dst, DWTELEM *src, DWTELEM *ref, int ds dst[w*dst_step] = LIFTS(src[w*src_step], mul*2*ref[w*ref_step]+add, inverse); } } +#endif static void inplace_lift(DWTELEM *dst, int width, int *coeffs, int n, int shift, int start, int inverse){ @@ -1111,76 +1110,6 @@ STOP_TIMER("vertical_decompose53i*")} } } -#define liftS lift -#define lift5 lift -#if 1 -#define W_AM 3 -#define W_AO 0 -#define W_AS 1 - -#undef liftS -#define W_BM 1 -#define W_BO 8 -#define W_BS 4 - -#define W_CM 1 -#define W_CO 0 -#define W_CS 0 - -#define W_DM 3 -#define W_DO 4 -#define W_DS 3 -#elif 0 -#define W_AM 55 -#define W_AO 16 -#define W_AS 5 - -#define W_BM 3 -#define W_BO 32 -#define W_BS 6 - -#define W_CM 127 -#define W_CO 64 -#define W_CS 7 - -#define W_DM 7 -#define W_DO 8 -#define W_DS 4 -#elif 0 -#define W_AM 97 -#define W_AO 32 -#define W_AS 6 - -#define W_BM 63 -#define W_BO 512 -#define W_BS 10 - -#define W_CM 13 -#define W_CO 8 -#define W_CS 4 - -#define W_DM 15 -#define W_DO 16 -#define W_DS 5 - -#else - -#define W_AM 203 -#define W_AO 64 -#define W_AS 7 - -#define W_BM 217 -#define W_BO 2048 -#define W_BS 12 - -#define W_CM 113 -#define W_CO 64 -#define W_CS 7 - -#define W_DM 227 -#define W_DO 128 -#define W_DS 9 -#endif static void horizontal_decompose97i(DWTELEM *b, int width){ DWTELEM temp[width]; const int w2= (width+1)>>1; @@ -1275,9 +1204,9 @@ void ff_spatial_dwt(DWTELEM *buffer, int width, int height, int stride, int type for(level=0; level>level, height>>level, stride<>level, height>>level, stride<>level, height>>level, stride<>level, height>>level, stride<>level, height>>level, stride<>level, height>>level, stride<>1; @@ -1463,7 +1392,7 @@ static void vertical_compose97iL1(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int wid } } -static void vertical_compose97i(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width){ +void ff_snow_vertical_compose97i(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width){ int i; for(i=0; iy = -3; } -static void spatial_compose97i_dy_buffered(dwt_compose_t *cs, slice_buffer * sb, int width, int height, int stride_line){ +static void spatial_compose97i_dy_buffered(DSPContext *dsp, dwt_compose_t *cs, slice_buffer * sb, int width, int height, int stride_line){ int y = cs->y; DWTELEM *b0= cs->b0; @@ -1516,7 +1445,7 @@ static void spatial_compose97i_dy_buffered(dwt_compose_t *cs, slice_buffer * sb, {START_TIMER if(y>0 && y+4vertical_compose97i(b0, b1, b2, b3, b4, b5, width); }else{ if(y+3<(unsigned)height) vertical_compose97iL1(b3, b4, b5, width); if(y+2<(unsigned)height) vertical_compose97iH1(b2, b3, b4, width); @@ -1527,8 +1456,8 @@ if(width>400){ STOP_TIMER("vertical_compose97i")}} {START_TIMER - if(y-1<(unsigned)height) horizontal_compose97i(b0, width); - if(y+0<(unsigned)height) horizontal_compose97i(b1, width); + if(y-1<(unsigned)height) dsp->horizontal_compose97i(b0, width); + if(y+0<(unsigned)height) dsp->horizontal_compose97i(b1, width); if(width>400 && y+0<(unsigned)height){ STOP_TIMER("horizontal_compose97i")}} @@ -1557,8 +1486,8 @@ if(width>400){ STOP_TIMER("vertical_compose97i")}} {START_TIMER - if(y-1<(unsigned)height) horizontal_compose97i(b0, width); - if(y+0<(unsigned)height) horizontal_compose97i(b1, width); + if(y-1<(unsigned)height) ff_snow_horizontal_compose97i(b0, width); + if(y+0<(unsigned)height) ff_snow_horizontal_compose97i(b1, width); if(width>400 && b0 <= b2){ STOP_TIMER("horizontal_compose97i")}} @@ -1576,32 +1505,32 @@ static void spatial_compose97i(DWTELEM *buffer, int width, int height, int strid spatial_compose97i_dy(&cs, buffer, width, height, stride); } -void ff_spatial_idwt_buffered_init(dwt_compose_t *cs, slice_buffer * sb, int width, int height, int stride_line, int type, int decomposition_count){ +static void ff_spatial_idwt_buffered_init(dwt_compose_t *cs, slice_buffer * sb, int width, int height, int stride_line, int type, int decomposition_count){ int level; for(level=decomposition_count-1; level>=0; level--){ switch(type){ - case 0: spatial_compose97i_buffered_init(cs+level, sb, height>>level, stride_line<>level, stride_line<>level, stride_line<>level, stride_line<>level, height>>level, stride<>level, height>>level, stride<=0; level--){ switch(type){ - case 0: spatial_compose97i_init(cs+level, buffer, height>>level, stride<>level, stride<>level, stride<>level, stride<>level, height>>level, stride<>level, height>>level, stride<=0; level--){ while(cs[level].y <= FFMIN((y>>level)+support, height>>level)){ switch(type){ - case 0: spatial_compose97i_dy(cs+level, buffer, width>>level, height>>level, stride<>level, height>>level, stride<>level, height>>level, stride<>level, height>>level, stride<=0; level--){ while(cs[level].y <= FFMIN((y>>level)+support, height>>level)){ switch(type){ - case 0: spatial_compose97i_dy_buffered(cs+level, slice_buf, width>>level, height>>level, stride_line<>level, height>>level, stride_line<>level, height>>level, stride_line<>level, height>>level, stride_line<=0; level--) @@ -1742,7 +1671,7 @@ static int encode_subband_c0run(SnowContext *s, SubBand *b, DWTELEM *src, DWTELE p= parent[px + py*2*stride]; } if(/*ll|*/l|lt|t|rt|p){ - int context= av_log2(/*ABS(ll) + */3*ABS(l) + ABS(lt) + 2*ABS(t) + ABS(rt) + ABS(p)); + int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p)); put_rac(&s->c, &b->state[0][context], !!v); }else{ @@ -1758,11 +1687,11 @@ static int encode_subband_c0run(SnowContext *s, SubBand *b, DWTELEM *src, DWTELE } } if(v){ - int context= av_log2(/*ABS(ll) + */3*ABS(l) + ABS(lt) + 2*ABS(t) + ABS(rt) + ABS(p)); - int l2= 2*ABS(l) + (l<0); - int t2= 2*ABS(t) + (t<0); + int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p)); + int l2= 2*FFABS(l) + (l<0); + int t2= 2*FFABS(t) + (t<0); - put_symbol2(&s->c, b->state[context + 2], ABS(v)-1, context-4); + put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4); put_rac(&s->c, &b->state[0][16 + 1 + 3 + quant3bA[l2&0xFF] + 3*quant3bA[t2&0xFF]], v<0); } } @@ -1825,7 +1754,7 @@ static inline void unpack_coeffs(SnowContext *s, SubBand *b, SubBand * parent, i } } if(/*ll|*/l|lt|t|rt|p){ - int context= av_log2(/*ABS(ll) + */3*(l>>1) + (lt>>1) + (t&~1) + (rt>>1) + (p>>1)); + int context= av_log2(/*FFABS(ll) + */3*(l>>1) + (lt>>1) + (t&~1) + (rt>>1) + (p>>1)); v=get_rac(&s->c, &b->state[0][context]); if(v){ @@ -1881,7 +1810,7 @@ static inline void unpack_coeffs(SnowContext *s, SubBand *b, SubBand * parent, i static inline void decode_subband_slice_buffered(SnowContext *s, SubBand *b, slice_buffer * sb, int start_y, int h, int save_state[1]){ const int w= b->width; int y; - const int qlog= clip(s->qlog + b->qlog, 0, QROOT*16); + const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16); int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT); int qadd= (s->qbias*qmul)>>QBIAS_SHIFT; int new_index = 0; @@ -1925,7 +1854,7 @@ static inline void decode_subband_slice_buffered(SnowContext *s, SubBand *b, sli return; } -static void reset_contexts(SnowContext *s){ +static void reset_contexts(SnowContext *s){ //FIXME better initial contexts int plane_index, level, orientation; for(plane_index=0; plane_index<3; plane_index++){ @@ -1978,7 +1907,7 @@ static int pix_sum(uint8_t * pix, int line_size, int w) static int pix_norm1(uint8_t * pix, int line_size, int w) { int s, i, j; - uint32_t *sq = squareTbl + 256; + uint32_t *sq = ff_squareTbl + 256; s = 0; for (i = 0; i < w; i++) { @@ -1991,7 +1920,7 @@ static int pix_norm1(uint8_t * pix, int line_size, int w) return s; } -static inline void set_blocks(SnowContext *s, int level, int x, int y, int l, int cb, int cr, int mx, int my, int type){ +static inline void set_blocks(SnowContext *s, int level, int x, int y, int l, int cb, int cr, int mx, int my, int ref, int type){ const int w= s->b_width << s->block_max_depth; const int rem_depth= s->block_max_depth - level; const int index= (x + y*w) << rem_depth; @@ -2004,6 +1933,7 @@ static inline void set_blocks(SnowContext *s, int level, int x, int y, int l, in block.color[2]= cr; block.mx= mx; block.my= my; + block.ref= ref; block.type= type; block.level= level; @@ -2028,6 +1958,22 @@ static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3 assert(!ref_index); } +static inline void pred_mv(SnowContext *s, int *mx, int *my, int ref, + const BlockNode *left, const BlockNode *top, const BlockNode *tr){ + if(s->ref_frames == 1){ + *mx = mid_pred(left->mx, top->mx, tr->mx); + *my = mid_pred(left->my, top->my, tr->my); + }else{ + const int *scale = scale_mv_ref[ref]; + *mx = mid_pred((left->mx * scale[left->ref] + 128) >>8, + (top ->mx * scale[top ->ref] + 128) >>8, + (tr ->mx * scale[tr ->ref] + 128) >>8); + *my = mid_pred((left->my * scale[left->ref] + 128) >>8, + (top ->my * scale[top ->ref] + 128) >>8, + (tr ->my * scale[tr ->ref] + 128) >>8); + } +} + //FIXME copy&paste #define P_LEFT P[1] #define P_TOP P[2] @@ -2052,17 +1998,16 @@ static int encode_q_branch(SnowContext *s, int level, int x, int y){ const int block_w= 1<<(LOG2_MB_SIZE - level); int trx= (x+1)<block[index-1] : &null_block; - BlockNode *top = y ? &s->block[index-w] : &null_block; - BlockNode *right = trxblock[index+1] : &null_block; - BlockNode *bottom= tryblock[index+w] : &null_block; - BlockNode *tl = y && x ? &s->block[index-w-1] : left; - BlockNode *tr = y && trxblock[index-w+(1<block[index-1] : &null_block; + const BlockNode *top = y ? &s->block[index-w] : &null_block; + const BlockNode *right = trxblock[index+1] : &null_block; + const BlockNode *bottom= tryblock[index+w] : &null_block; + const BlockNode *tl = y && x ? &s->block[index-w-1] : left; + const BlockNode *tr = y && trxblock[index-w+(1<color[0]; int pcb= left->color[1]; int pcr= left->color[2]; - int pmx= mid_pred(left->mx, top->mx, tr->mx); - int pmy= mid_pred(left->my, top->my, tr->my); + int pmx, pmy; int mx=0, my=0; int l,cr,cb; const int stride= s->current_picture.linesize[0]; @@ -2075,13 +2020,15 @@ static int encode_q_branch(SnowContext *s, int level, int x, int y){ int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused const int shift= 1+qpel; MotionEstContext *c= &s->m.me; - int mx_context= av_log2(2*ABS(left->mx - top->mx)); - int my_context= av_log2(2*ABS(left->my - top->my)); + int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref); + int mx_context= av_log2(2*FFABS(left->mx - top->mx)); + int my_context= av_log2(2*FFABS(left->my - top->my)); int s_context= 2*left->level + 2*top->level + tl->level + tr->level; + int ref, best_ref, ref_score, ref_mx, ref_my; assert(sizeof(s->block_state) >= 256); if(s->keyframe){ - set_blocks(s, level, x, y, pl, pcb, pcr, pmx, pmy, BLOCK_INTRA); + set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA); return 0; } @@ -2104,12 +2051,10 @@ static int encode_q_branch(SnowContext *s, int level, int x, int y){ s->m.mb_stride=2; s->m.mb_x= s->m.mb_y= 0; - s->m.me.skip= 0; - - init_ref(c, current_data, s->last_picture.data, NULL, block_w*x, block_w*y, 0); + c->skip= 0; - assert(s->m.me. stride == stride); - assert(s->m.me.uvstride == uvstride); + assert(c-> stride == stride); + assert(c->uvstride == uvstride); c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp); c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp); @@ -2140,16 +2085,34 @@ static int encode_q_branch(SnowContext *s, int level, int x, int y){ c->pred_y = P_MEDIAN[1]; } - score= ff_epzs_motion_search(&s->m, &mx, &my, P, 0, /*ref_index*/ 0, last_mv, - (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w); + score= INT_MAX; + best_ref= 0; + for(ref=0; refref_frames; ref++){ + init_ref(c, current_data, s->last_picture[ref].data, NULL, block_w*x, block_w*y, 0); - assert(mx >= c->xmin); - assert(mx <= c->xmax); - assert(my >= c->ymin); - assert(my <= c->ymax); + ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv, + (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w); - score= s->m.me.sub_motion_search(&s->m, &mx, &my, score, 0, 0, level-LOG2_MB_SIZE+4, block_w); - score= ff_get_mb_score(&s->m, mx, my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0); + assert(ref_mx >= c->xmin); + assert(ref_mx <= c->xmax); + assert(ref_my >= c->ymin); + assert(ref_my <= c->ymax); + + ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w); + ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0); + ref_score+= 2*av_log2(2*ref)*c->penalty_factor; + if(s->ref_mvs[ref]){ + s->ref_mvs[ref][index][0]= ref_mx; + s->ref_mvs[ref][index][1]= ref_my; + s->ref_scores[ref][index]= ref_score; + } + if(score > ref_score){ + score= ref_score; + best_ref= ref; + mx= ref_mx; + my= ref_my; + } + } //FIXME if mb_cmp != SSE then intra cant be compared currently and mb_penalty vs. lambda2 // subpel search @@ -2161,8 +2124,11 @@ static int encode_q_branch(SnowContext *s, int level, int x, int y){ if(level!=s->block_max_depth) put_rac(&pc, &p_state[4 + s_context], 1); put_rac(&pc, &p_state[1 + left->type + top->type], 0); - put_symbol(&pc, &p_state[128 + 32*mx_context], mx - pmx, 1); - put_symbol(&pc, &p_state[128 + 32*my_context], my - pmy, 1); + if(s->ref_frames > 1) + put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0); + pred_mv(s, &pmx, &pmy, best_ref, left, top, tr); + put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1); + put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1); p_len= pc.bytestream - pc.bytestream_start; score += (s->lambda2*(p_len*8 + (pc.outstanding_count - s->c.outstanding_count)*8 @@ -2226,11 +2192,12 @@ static int encode_q_branch(SnowContext *s, int level, int x, int y){ } if(iscore < score){ + pred_mv(s, &pmx, &pmy, 0, left, top, tr); memcpy(pbbak, i_buffer, i_len); s->c= ic; s->c.bytestream_start= pbbak_start; s->c.bytestream= pbbak + i_len; - set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, BLOCK_INTRA); + set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA); memcpy(s->block_state, i_state, sizeof(s->block_state)); return iscore; }else{ @@ -2238,17 +2205,17 @@ static int encode_q_branch(SnowContext *s, int level, int x, int y){ s->c= pc; s->c.bytestream_start= pbbak_start; s->c.bytestream= pbbak + p_len; - set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, 0); + set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0); memcpy(s->block_state, p_state, sizeof(s->block_state)); return score; } } -static always_inline int same_block(BlockNode *a, BlockNode *b){ +static av_always_inline int same_block(BlockNode *a, BlockNode *b){ if((a->type&BLOCK_INTRA) && (b->type&BLOCK_INTRA)){ return !((a->color[0] - b->color[0]) | (a->color[1] - b->color[1]) | (a->color[2] - b->color[2])); }else{ - return !((a->mx - b->mx) | (a->my - b->my) | ((a->type ^ b->type)&BLOCK_INTRA)); + return !((a->mx - b->mx) | (a->my - b->my) | (a->ref - b->ref) | ((a->type ^ b->type)&BLOCK_INTRA)); } } @@ -2258,21 +2225,21 @@ static void encode_q_branch2(SnowContext *s, int level, int x, int y){ const int index= (x + y*w) << rem_depth; int trx= (x+1)<block[index]; - BlockNode *left = x ? &s->block[index-1] : &null_block; - BlockNode *top = y ? &s->block[index-w] : &null_block; - BlockNode *tl = y && x ? &s->block[index-w-1] : left; - BlockNode *tr = y && trxblock[index-w+(1<block[index-1] : &null_block; + const BlockNode *top = y ? &s->block[index-w] : &null_block; + const BlockNode *tl = y && x ? &s->block[index-w-1] : left; + const BlockNode *tr = y && trxblock[index-w+(1<color[0]; int pcb= left->color[1]; int pcr= left->color[2]; - int pmx= mid_pred(left->mx, top->mx, tr->mx); - int pmy= mid_pred(left->my, top->my, tr->my); - int mx_context= av_log2(2*ABS(left->mx - top->mx)); - int my_context= av_log2(2*ABS(left->my - top->my)); + int pmx, pmy; + int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref); + int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref; + int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref; int s_context= 2*left->level + 2*top->level + tl->level + tr->level; if(s->keyframe){ - set_blocks(s, level, x, y, pl, pcb, pcr, pmx, pmy, BLOCK_INTRA); + set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA); return; } @@ -2289,16 +2256,20 @@ static void encode_q_branch2(SnowContext *s, int level, int x, int y){ } } if(b->type & BLOCK_INTRA){ + pred_mv(s, &pmx, &pmy, 0, left, top, tr); put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1); put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1); put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1); put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1); - set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, BLOCK_INTRA); + set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA); }else{ + pred_mv(s, &pmx, &pmy, b->ref, left, top, tr); put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0); + if(s->ref_frames > 1) + put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0); put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1); put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1); - set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, 0); + set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0); } } @@ -2307,38 +2278,42 @@ static void decode_q_branch(SnowContext *s, int level, int x, int y){ const int rem_depth= s->block_max_depth - level; const int index= (x + y*w) << rem_depth; int trx= (x+1)<block[index-1] : &null_block; - BlockNode *top = y ? &s->block[index-w] : &null_block; - BlockNode *tl = y && x ? &s->block[index-w-1] : left; - BlockNode *tr = y && trxblock[index-w+(1<block[index-1] : &null_block; + const BlockNode *top = y ? &s->block[index-w] : &null_block; + const BlockNode *tl = y && x ? &s->block[index-w-1] : left; + const BlockNode *tr = y && trxblock[index-w+(1<level + 2*top->level + tl->level + tr->level; if(s->keyframe){ - set_blocks(s, level, x, y, null_block.color[0], null_block.color[1], null_block.color[2], null_block.mx, null_block.my, BLOCK_INTRA); + set_blocks(s, level, x, y, null_block.color[0], null_block.color[1], null_block.color[2], null_block.mx, null_block.my, null_block.ref, BLOCK_INTRA); return; } if(level==s->block_max_depth || get_rac(&s->c, &s->block_state[4 + s_context])){ - int type; + int type, mx, my; int l = left->color[0]; int cb= left->color[1]; int cr= left->color[2]; - int mx= mid_pred(left->mx, top->mx, tr->mx); - int my= mid_pred(left->my, top->my, tr->my); - int mx_context= av_log2(2*ABS(left->mx - top->mx)) + 0*av_log2(2*ABS(tr->mx - top->mx)); - int my_context= av_log2(2*ABS(left->my - top->my)) + 0*av_log2(2*ABS(tr->my - top->my)); + int ref = 0; + int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref); + int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 0*av_log2(2*FFABS(tr->mx - top->mx)); + int my_context= av_log2(2*FFABS(left->my - top->my)) + 0*av_log2(2*FFABS(tr->my - top->my)); type= get_rac(&s->c, &s->block_state[1 + left->type + top->type]) ? BLOCK_INTRA : 0; if(type){ + pred_mv(s, &mx, &my, 0, left, top, tr); l += get_symbol(&s->c, &s->block_state[32], 1); cb+= get_symbol(&s->c, &s->block_state[64], 1); cr+= get_symbol(&s->c, &s->block_state[96], 1); }else{ - mx+= get_symbol(&s->c, &s->block_state[128 + 32*mx_context], 1); - my+= get_symbol(&s->c, &s->block_state[128 + 32*my_context], 1); + if(s->ref_frames > 1) + ref= get_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], 0); + pred_mv(s, &mx, &my, ref, left, top, tr); + mx+= get_symbol(&s->c, &s->block_state[128 + 32*(mx_context + 16*!!ref)], 1); + my+= get_symbol(&s->c, &s->block_state[128 + 32*(my_context + 16*!!ref)], 1); } - set_blocks(s, level, x, y, l, cb, cr, mx, my, type); + set_blocks(s, level, x, y, l, cb, cr, mx, my, ref, type); }else{ decode_q_branch(s, level+1, 2*x+0, 2*y+0); decode_q_branch(s, level+1, 2*x+1, 2*y+0); @@ -2347,12 +2322,12 @@ static void decode_q_branch(SnowContext *s, int level, int x, int y){ } } -static void encode_blocks(SnowContext *s){ +static void encode_blocks(SnowContext *s, int search){ int x, y; int w= s->b_width; int h= s->b_height; - if(s->avctx->me_method == ME_ITER && !s->keyframe) + if(s->avctx->me_method == ME_ITER && !s->keyframe && search) iterative_me(s); for(y=0; yavctx->me_method == ME_ITER) + if(s->avctx->me_method == ME_ITER || !search) encode_q_branch2(s, 0, x, y); else encode_q_branch (s, 0, x, y); @@ -2466,7 +2441,7 @@ mca( 8, 0,8) mca( 0, 8,8) mca( 8, 8,8) -static void pred_block(SnowContext *s, uint8_t *dst, uint8_t *src, uint8_t *tmp, int stride, int sx, int sy, int b_w, int b_h, BlockNode *block, int plane_index, int w, int h){ +static void pred_block(SnowContext *s, uint8_t *dst, uint8_t *tmp, int stride, int sx, int sy, int b_w, int b_h, BlockNode *block, int plane_index, int w, int h){ if(block->type & BLOCK_INTRA){ int x, y; const int color = block->color[plane_index]; @@ -2506,6 +2481,7 @@ static void pred_block(SnowContext *s, uint8_t *dst, uint8_t *src, uint8_t *tmp, } } }else{ + uint8_t *src= s->last_picture[block->ref].data[plane_index]; const int scale= plane_index ? s->mv_scale : 2*s->mv_scale; int mx= block->mx*scale; int my= block->my*scale; @@ -2520,11 +2496,11 @@ static void pred_block(SnowContext *s, uint8_t *dst, uint8_t *src, uint8_t *tmp, ff_emulated_edge_mc(tmp + MB_SIZE, src, stride, b_w+5, b_h+5, sx, sy, w, h); src= tmp + MB_SIZE; } - assert(b_w == b_h || 2*b_w == b_h || b_w == 2*b_h); - assert(!(b_w&(b_w-1))); +// assert(b_w == b_h || 2*b_w == b_h || b_w == 2*b_h); +// assert(!(b_w&(b_w-1))); assert(b_w>1 && b_h>1); assert(tab_index>=0 && tab_index<4 || b_w==32); - if((dx&3) || (dy&3)) + if((dx&3) || (dy&3) || !(b_w == b_h || 2*b_w == b_h || b_w == 2*b_h) || (b_w&(b_w-1))) mc_block(dst, src, tmp, stride, b_w, b_h, dx, dy); else if(b_w==32){ int y; @@ -2545,136 +2521,16 @@ static void pred_block(SnowContext *s, uint8_t *dst, uint8_t *src, uint8_t *tmp, } } -//FIXME name clenup (b_w, block_w, b_width stuff) -static always_inline void add_yblock_buffered(SnowContext *s, slice_buffer * sb, DWTELEM *old_dst, uint8_t *dst8, uint8_t *src, uint8_t *obmc, int src_x, int src_y, int b_w, int b_h, int w, int h, int dst_stride, int src_stride, int obmc_stride, int b_x, int b_y, int add, int plane_index){ - DWTELEM * dst = NULL; - const int b_width = s->b_width << s->block_max_depth; - const int b_height= s->b_height << s->block_max_depth; - const int b_stride= b_width; - BlockNode *lt= &s->block[b_x + b_y*b_stride]; - BlockNode *rt= lt+1; - BlockNode *lb= lt+b_stride; - BlockNode *rb= lb+1; - uint8_t *block[4]; - int tmp_step= src_stride >= 7*MB_SIZE ? MB_SIZE : MB_SIZE*src_stride; - uint8_t tmp[src_stride*7*MB_SIZE]; //FIXME align - uint8_t *ptmp; - int x,y; - - if(b_x<0){ - lt= rt; - lb= rb; - }else if(b_x + 1 >= b_width){ - rt= lt; - rb= lb; - } - if(b_y<0){ - lt= lb; - rt= rb; - }else if(b_y + 1 >= b_height){ - lb= lt; - rb= rt; - } - - if(src_x<0){ //FIXME merge with prev & always round internal width upto *16 - obmc -= src_x; - b_w += src_x; - src_x=0; - }else if(src_x + b_w > w){ - b_w = w - src_x; - } - if(src_y<0){ - obmc -= src_y*obmc_stride; - b_h += src_y; - src_y=0; - }else if(src_y + b_h> h){ - b_h = h - src_y; - } - - if(b_w<=0 || b_h<=0) return; - -assert(src_stride > 2*MB_SIZE + 5); -// old_dst += src_x + src_y*dst_stride; - dst8+= src_x + src_y*src_stride; -// src += src_x + src_y*src_stride; - - ptmp= tmp + 3*tmp_step; - block[0]= ptmp; - ptmp+=tmp_step; - pred_block(s, block[0], src, tmp, src_stride, src_x, src_y, b_w, b_h, lt, plane_index, w, h); - - if(same_block(lt, rt)){ - block[1]= block[0]; - }else{ - block[1]= ptmp; - ptmp+=tmp_step; - pred_block(s, block[1], src, tmp, src_stride, src_x, src_y, b_w, b_h, rt, plane_index, w, h); - } - - if(same_block(lt, lb)){ - block[2]= block[0]; - }else if(same_block(rt, lb)){ - block[2]= block[1]; - }else{ - block[2]= ptmp; - ptmp+=tmp_step; - pred_block(s, block[2], src, tmp, src_stride, src_x, src_y, b_w, b_h, lb, plane_index, w, h); - } - - if(same_block(lt, rb) ){ - block[3]= block[0]; - }else if(same_block(rt, rb)){ - block[3]= block[1]; - }else if(same_block(lb, rb)){ - block[3]= block[2]; - }else{ - block[3]= ptmp; - pred_block(s, block[3], src, tmp, src_stride, src_x, src_y, b_w, b_h, rb, plane_index, w, h); - } -#if 0 - for(y=0; y>1); - for(x=0; x>1); - for(x=0; x>1); - uint8_t *obmc4= obmc3+ (obmc_stride>>1); - for(x=0; x>1); - uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1); - uint8_t *obmc4= obmc3+ (obmc_stride>>1); + const uint8_t *obmc1= obmc + y*obmc_stride; + const uint8_t *obmc2= obmc1+ (obmc_stride>>1); + const uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1); + const uint8_t *obmc4= obmc3+ (obmc_stride>>1); dst = slice_buffer_get_line(sb, src_y + y); for(x=0; x 2*MB_SIZE + 5); v >>= 8 - FRAC_BITS; } if(add){ -// v += old_dst[x + y*dst_stride]; v += dst[x + src_x]; v = (v + (1<<(FRAC_BITS-1))) >> FRAC_BITS; if(v&(~255)) v= ~(v>>31); dst8[x + y*src_stride] = v; }else{ -// old_dst[x + y*dst_stride] -= v; dst[x + src_x] -= v; } } } - STOP_TIMER("Inner add y block") -} -#endif } //FIXME name clenup (b_w, block_w, b_width stuff) -static always_inline void add_yblock(SnowContext *s, DWTELEM *dst, uint8_t *dst8, uint8_t *src, uint8_t *obmc, int src_x, int src_y, int b_w, int b_h, int w, int h, int dst_stride, int src_stride, int obmc_stride, int b_x, int b_y, int add, int offset_dst, int plane_index){ +static av_always_inline void add_yblock(SnowContext *s, int sliced, slice_buffer *sb, DWTELEM *dst, uint8_t *dst8, const uint8_t *obmc, int src_x, int src_y, int b_w, int b_h, int w, int h, int dst_stride, int src_stride, int obmc_stride, int b_x, int b_y, int add, int offset_dst, int plane_index){ const int b_width = s->b_width << s->block_max_depth; const int b_height= s->b_height << s->block_max_depth; const int b_stride= b_width; @@ -2737,7 +2588,7 @@ static always_inline void add_yblock(SnowContext *s, DWTELEM *dst, uint8_t *dst8 if(src_x<0){ //FIXME merge with prev & always round internal width upto *16 obmc -= src_x; b_w += src_x; - if(!offset_dst) + if(!sliced && !offset_dst) dst -= src_x; src_x=0; }else if(src_x + b_w > w){ @@ -2746,7 +2597,7 @@ static always_inline void add_yblock(SnowContext *s, DWTELEM *dst, uint8_t *dst8 if(src_y<0){ obmc -= src_y*obmc_stride; b_h += src_y; - if(!offset_dst) + if(!sliced && !offset_dst) dst -= src_y*dst_stride; src_y=0; }else if(src_y + b_h> h){ @@ -2756,7 +2607,7 @@ static always_inline void add_yblock(SnowContext *s, DWTELEM *dst, uint8_t *dst8 if(b_w<=0 || b_h<=0) return; assert(src_stride > 2*MB_SIZE + 5); - if(offset_dst) + if(!sliced && offset_dst) dst += src_x + src_y*dst_stride; dst8+= src_x + src_y*src_stride; // src += src_x + src_y*src_stride; @@ -2764,14 +2615,14 @@ assert(src_stride > 2*MB_SIZE + 5); ptmp= tmp + 3*tmp_step; block[0]= ptmp; ptmp+=tmp_step; - pred_block(s, block[0], src, tmp, src_stride, src_x, src_y, b_w, b_h, lt, plane_index, w, h); + pred_block(s, block[0], tmp, src_stride, src_x, src_y, b_w, b_h, lt, plane_index, w, h); if(same_block(lt, rt)){ block[1]= block[0]; }else{ block[1]= ptmp; ptmp+=tmp_step; - pred_block(s, block[1], src, tmp, src_stride, src_x, src_y, b_w, b_h, rt, plane_index, w, h); + pred_block(s, block[1], tmp, src_stride, src_x, src_y, b_w, b_h, rt, plane_index, w, h); } if(same_block(lt, lb)){ @@ -2781,7 +2632,7 @@ assert(src_stride > 2*MB_SIZE + 5); }else{ block[2]= ptmp; ptmp+=tmp_step; - pred_block(s, block[2], src, tmp, src_stride, src_x, src_y, b_w, b_h, lb, plane_index, w, h); + pred_block(s, block[2], tmp, src_stride, src_x, src_y, b_w, b_h, lb, plane_index, w, h); } if(same_block(lt, rb) ){ @@ -2792,7 +2643,7 @@ assert(src_stride > 2*MB_SIZE + 5); block[3]= block[2]; }else{ block[3]= ptmp; - pred_block(s, block[3], src, tmp, src_stride, src_x, src_y, b_w, b_h, rb, plane_index, w, h); + pred_block(s, block[3], tmp, src_stride, src_x, src_y, b_w, b_h, rb, plane_index, w, h); } #if 0 for(y=0; y 2*MB_SIZE + 5); } } #else + if(sliced){ + START_TIMER + + s->dsp.inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8); + STOP_TIMER("inner_add_yblock") + }else for(y=0; y>1); - uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1); - uint8_t *obmc4= obmc3+ (obmc_stride>>1); + const uint8_t *obmc1= obmc + y*obmc_stride; + const uint8_t *obmc2= obmc1+ (obmc_stride>>1); + const uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1); + const uint8_t *obmc4= obmc3+ (obmc_stride>>1); for(x=0; x 2*MB_SIZE + 5); #endif } -static always_inline void predict_slice_buffered(SnowContext *s, slice_buffer * sb, DWTELEM * old_buffer, int plane_index, int add, int mb_y){ +static av_always_inline void predict_slice_buffered(SnowContext *s, slice_buffer * sb, DWTELEM * old_buffer, int plane_index, int add, int mb_y){ Plane *p= &s->plane[plane_index]; const int mb_w= s->b_width << s->block_max_depth; const int mb_h= s->b_height << s->block_max_depth; @@ -2868,7 +2725,6 @@ static always_inline void predict_slice_buffered(SnowContext *s, slice_buffer * const uint8_t *obmc = plane_index ? obmc_tab[s->block_max_depth+1] : obmc_tab[s->block_max_depth]; int obmc_stride= plane_index ? block_size : 2*block_size; int ref_stride= s->current_picture.linesize[plane_index]; - uint8_t *ref = s->last_picture.data[plane_index]; uint8_t *dst8= s->current_picture.data[plane_index]; int w= p->width; int h= p->height; @@ -2911,14 +2767,14 @@ static always_inline void predict_slice_buffered(SnowContext *s, slice_buffer * for(mb_x=0; mb_x<=mb_w; mb_x++){ START_TIMER - add_yblock_buffered(s, sb, old_buffer, dst8, ref, obmc, + add_yblock(s, 1, sb, old_buffer, dst8, obmc, block_w*mb_x - block_w/2, block_w*mb_y - block_w/2, block_w, block_w, w, h, w, ref_stride, obmc_stride, mb_x - 1, mb_y - 1, - add, plane_index); + add, 0, plane_index); STOP_TIMER("add_yblock") } @@ -2926,7 +2782,7 @@ static always_inline void predict_slice_buffered(SnowContext *s, slice_buffer * STOP_TIMER("predict_slice") } -static always_inline void predict_slice(SnowContext *s, DWTELEM *buf, int plane_index, int add, int mb_y){ +static av_always_inline void predict_slice(SnowContext *s, DWTELEM *buf, int plane_index, int add, int mb_y){ Plane *p= &s->plane[plane_index]; const int mb_w= s->b_width << s->block_max_depth; const int mb_h= s->b_height << s->block_max_depth; @@ -2936,7 +2792,6 @@ static always_inline void predict_slice(SnowContext *s, DWTELEM *buf, int plane_ const uint8_t *obmc = plane_index ? obmc_tab[s->block_max_depth+1] : obmc_tab[s->block_max_depth]; const int obmc_stride= plane_index ? block_size : 2*block_size; int ref_stride= s->current_picture.linesize[plane_index]; - uint8_t *ref = s->last_picture.data[plane_index]; uint8_t *dst8= s->current_picture.data[plane_index]; int w= p->width; int h= p->height; @@ -2969,7 +2824,7 @@ static always_inline void predict_slice(SnowContext *s, DWTELEM *buf, int plane_ for(mb_x=0; mb_x<=mb_w; mb_x++){ START_TIMER - add_yblock(s, buf, dst8, ref, obmc, + add_yblock(s, 0, NULL, buf, dst8, obmc, block_w*mb_x - block_w/2, block_w*mb_y - block_w/2, block_w, block_w, @@ -2984,7 +2839,7 @@ static always_inline void predict_slice(SnowContext *s, DWTELEM *buf, int plane_ STOP_TIMER("predict_slice") } -static always_inline void predict_plane(SnowContext *s, DWTELEM *buf, int plane_index, int add){ +static av_always_inline void predict_plane(SnowContext *s, DWTELEM *buf, int plane_index, int add){ const int mb_h= s->b_height << s->block_max_depth; int mb_y; for(mb_y=0; mb_y<=mb_h; mb_y++) @@ -2999,7 +2854,6 @@ static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){ const uint8_t *obmc = plane_index ? obmc_tab[s->block_max_depth+1] : obmc_tab[s->block_max_depth]; const int obmc_stride= plane_index ? block_size : 2*block_size; const int ref_stride= s->current_picture.linesize[plane_index]; - uint8_t *ref= s-> last_picture.data[plane_index]; uint8_t *src= s-> input_picture.data[plane_index]; DWTELEM *dst= (DWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; const int b_stride = s->b_width << s->block_max_depth; @@ -3021,7 +2875,7 @@ static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){ int x= block_w*mb_x2 + block_w/2; int y= block_w*mb_y2 + block_w/2; - add_yblock(s, dst + ((i&1)+(i>>1)*obmc_stride)*block_w, NULL, ref, obmc, + add_yblock(s, 0, NULL, dst + ((i&1)+(i>>1)*obmc_stride)*block_w, NULL, obmc, x, y, block_w, block_w, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index); for(y2= FFMAX(y, 0); y2b_width << s->block_max_depth; const int b_height = s->b_height<< s->block_max_depth; int index= x + y*b_stride; - BlockNode *b = &s->block[index]; - BlockNode *left = x ? &s->block[index-1] : &null_block; - BlockNode *top = y ? &s->block[index-b_stride] : &null_block; - BlockNode *tl = y && x ? &s->block[index-b_stride-1] : left; - BlockNode *tr = y && x+wblock[index-b_stride+w] : tl; + const BlockNode *b = &s->block[index]; + const BlockNode *left = x ? &s->block[index-1] : &null_block; + const BlockNode *top = y ? &s->block[index-b_stride] : &null_block; + const BlockNode *tl = y && x ? &s->block[index-b_stride-1] : left; + const BlockNode *tr = y && x+wblock[index-b_stride+w] : tl; int dmx, dmy; -// int mx_context= av_log2(2*ABS(left->mx - top->mx)); -// int my_context= av_log2(2*ABS(left->my - top->my)); +// int mx_context= av_log2(2*FFABS(left->mx - top->mx)); +// int my_context= av_log2(2*FFABS(left->my - top->my)); if(x<0 || x>=b_stride || y>=b_height) return 0; - dmx= b->mx - mid_pred(left->mx, top->mx, tr->mx); - dmy= b->my - mid_pred(left->my, top->my, tr->my); /* 1 0 0 01X 1-2 1 @@ -3074,22 +2926,25 @@ static inline int get_block_bits(SnowContext *s, int x, int y, int w){ //FIXME try accurate rate //FIXME intra and inter predictors if surrounding blocks arent the same type if(b->type & BLOCK_INTRA){ - return 3+2*( av_log2(2*ABS(left->color[0] - b->color[0])) - + av_log2(2*ABS(left->color[1] - b->color[1])) - + av_log2(2*ABS(left->color[2] - b->color[2]))); - }else - return 2*(1 + av_log2(2*ABS(dmx)) - + av_log2(2*ABS(dmy))); //FIXME kill the 2* can be merged in lambda + return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0])) + + av_log2(2*FFABS(left->color[1] - b->color[1])) + + av_log2(2*FFABS(left->color[2] - b->color[2]))); + }else{ + pred_mv(s, &dmx, &dmy, b->ref, left, top, tr); + dmx-= b->mx; + dmy-= b->my; + return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda + + av_log2(2*FFABS(dmy)) + + av_log2(2*b->ref)); + } } static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, const uint8_t *obmc_edged){ Plane *p= &s->plane[plane_index]; const int block_size = MB_SIZE >> s->block_max_depth; const int block_w = plane_index ? block_size/2 : block_size; - const uint8_t *obmc = plane_index ? obmc_tab[s->block_max_depth+1] : obmc_tab[s->block_max_depth]; const int obmc_stride= plane_index ? block_size : 2*block_size; const int ref_stride= s->current_picture.linesize[plane_index]; - uint8_t *ref= s-> last_picture.data[plane_index]; uint8_t *dst= s->current_picture.data[plane_index]; uint8_t *src= s-> input_picture.data[plane_index]; DWTELEM *pred= (DWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; @@ -3104,13 +2959,13 @@ static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, con const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp); int sx= block_w*mb_x - block_w/2; int sy= block_w*mb_y - block_w/2; - const int x0= FFMAX(0,-sx); - const int y0= FFMAX(0,-sy); - const int x1= FFMIN(block_w*2, w-sx); - const int y1= FFMIN(block_w*2, h-sy); + int x0= FFMAX(0,-sx); + int y0= FFMAX(0,-sy); + int x1= FFMIN(block_w*2, w-sx); + int y1= FFMIN(block_w*2, h-sy); int i,x,y; - pred_block(s, cur, ref, tmp, ref_stride, sx, sy, block_w*2, block_w*2, &s->block[mb_x + mb_y*b_stride], plane_index, w, h); + pred_block(s, cur, tmp, ref_stride, sx, sy, block_w*2, block_w*2, &s->block[mb_x + mb_y*b_stride], plane_index, w, h); for(y=y0; y>1))*ref_stride; - distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16); + /* FIXME rearrange dsputil to fit 32x32 cmp functions */ + /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */ + /* FIXME cmps overlap but don't cover the wavelet's whole support, + * so improving the score of one block is not strictly guaranteed to + * improve the score of the whole frame, so iterative motion est + * doesn't always converge. */ + if(s->avctx->me_cmp == FF_CMP_W97) + distortion = w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32); + else if(s->avctx->me_cmp == FF_CMP_W53) + distortion = w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32); + else{ + distortion = 0; + for(i=0; i<4; i++){ + int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride; + distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16); + } } }else{ assert(block_w==8); @@ -3159,12 +3041,10 @@ static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){ const uint8_t *obmc = plane_index ? obmc_tab[s->block_max_depth+1] : obmc_tab[s->block_max_depth]; const int obmc_stride= plane_index ? block_size : 2*block_size; const int ref_stride= s->current_picture.linesize[plane_index]; - uint8_t *ref= s-> last_picture.data[plane_index]; uint8_t *dst= s->current_picture.data[plane_index]; uint8_t *src= s-> input_picture.data[plane_index]; - const static DWTELEM zero_dst[4096]; //FIXME + static const DWTELEM zero_dst[4096]; //FIXME const int b_stride = s->b_width << s->block_max_depth; - const int b_height = s->b_height<< s->block_max_depth; const int w= p->width; const int h= p->height; int distortion= 0; @@ -3177,7 +3057,7 @@ static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){ int x= block_w*mb_x2 + block_w/2; int y= block_w*mb_y2 + block_w/2; - add_yblock(s, zero_dst, dst, ref, obmc, + add_yblock(s, 0, NULL, zero_dst, dst, obmc, x, y, block_w, block_w, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index); //FIXME find a cleaner/simpler way to skip the outside stuff @@ -3217,7 +3097,7 @@ static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){ return distortion + rate*penalty_factor; } -static always_inline int check_block(SnowContext *s, int mb_x, int mb_y, int p[3], int intra, const uint8_t *obmc_edged, int *best_rd){ +static av_always_inline int check_block(SnowContext *s, int mb_x, int mb_y, int p[3], int intra, const uint8_t *obmc_edged, int *best_rd){ const int b_stride= s->b_width << s->block_max_depth; BlockNode *block= &s->block[mb_x + mb_y * b_stride]; BlockNode backup= *block; @@ -3233,7 +3113,7 @@ static always_inline int check_block(SnowContext *s, int mb_x, int mb_y, int p[3 block->type |= BLOCK_INTRA; }else{ index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1); - value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6); + value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12); if(s->me_cache[index] == value) return 0; s->me_cache[index]= value; @@ -3256,12 +3136,12 @@ static always_inline int check_block(SnowContext *s, int mb_x, int mb_y, int p[3 } /* special case for int[2] args we discard afterward, fixes compilation prob with gcc 2.95 */ -static always_inline int check_block_inter(SnowContext *s, int mb_x, int mb_y, int p0, int p1, int intra, const uint8_t *obmc_edged, int *best_rd){ +static av_always_inline int check_block_inter(SnowContext *s, int mb_x, int mb_y, int p0, int p1, const uint8_t *obmc_edged, int *best_rd){ int p[2] = {p0, p1}; - return check_block(s, mb_x, mb_y, p, intra, obmc_edged, best_rd); + return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd); } -static always_inline int check_4block_inter(SnowContext *s, int mb_x, int mb_y, int p0, int p1, int *best_rd){ +static av_always_inline int check_4block_inter(SnowContext *s, int mb_x, int mb_y, int p0, int p1, int ref, int *best_rd){ const int b_stride= s->b_width << s->block_max_depth; BlockNode *block= &s->block[mb_x + mb_y * b_stride]; BlockNode backup[4]= {block[0], block[1], block[b_stride], block[b_stride+1]}; @@ -3272,13 +3152,14 @@ static always_inline int check_4block_inter(SnowContext *s, int mb_x, int mb_y, assert(((mb_x|mb_y)&1) == 0); index= (p0 + 31*p1) & (ME_CACHE_SIZE-1); - value= s->me_cache_generation + (p0>>10) + (p1<<6); + value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12); if(s->me_cache[index] == value) return 0; s->me_cache[index]= value; block->mx= p0; block->my= p1; + block->ref= ref; block->type &= ~BLOCK_INTRA; block[1]= block[b_stride]= block[b_stride+1]= *block; @@ -3304,24 +3185,35 @@ static void iterative_me(SnowContext *s){ const int b_stride= b_width; int color[3]; - for(pass=0; pass<50; pass++){ + { + RangeCoder r = s->c; + uint8_t state[sizeof(s->block_state)]; + memcpy(state, s->block_state, sizeof(s->block_state)); + for(mb_y= 0; mb_yb_height; mb_y++) + for(mb_x= 0; mb_xb_width; mb_x++) + encode_q_branch(s, 0, mb_x, mb_y); + s->c = r; + memcpy(s->block_state, state, sizeof(s->block_state)); + } + + for(pass=0; pass<25; pass++){ int change= 0; for(mb_y= 0; mb_yblock[index]; - BlockNode *tb = mb_y ? &s->block[index-b_stride ] : &null_block; - BlockNode *lb = mb_x ? &s->block[index -1] : &null_block; - BlockNode *rb = mb_x+1block[index +1] : &null_block; - BlockNode *bb = mb_y+1block[index+b_stride ] : &null_block; - BlockNode *tlb= mb_x && mb_y ? &s->block[index-b_stride-1] : &null_block; - BlockNode *trb= mb_x+1block[index-b_stride+1] : &null_block; - BlockNode *blb= mb_x && mb_y+1block[index+b_stride-1] : &null_block; - BlockNode *brb= mb_x+1block[index+b_stride+1] : &null_block; + BlockNode *tb = mb_y ? &s->block[index-b_stride ] : NULL; + BlockNode *lb = mb_x ? &s->block[index -1] : NULL; + BlockNode *rb = mb_x+1block[index +1] : NULL; + BlockNode *bb = mb_y+1block[index+b_stride ] : NULL; + BlockNode *tlb= mb_x && mb_y ? &s->block[index-b_stride-1] : NULL; + BlockNode *trb= mb_x+1block[index-b_stride+1] : NULL; + BlockNode *blb= mb_x && mb_y+1block[index+b_stride-1] : NULL; + BlockNode *brb= mb_x+1block[index+b_stride+1] : NULL; const int b_w= (MB_SIZE >> s->block_max_depth); uint8_t obmc_edged[b_w*2][b_w*2]; @@ -3395,48 +3287,72 @@ static void iterative_me(SnowContext *s){ int color0[3]= {block->color[0], block->color[1], block->color[2]}; check_block(s, mb_x, mb_y, color0, 1, *obmc_edged, &best_rd); }else - check_block_inter(s, mb_x, mb_y, block->mx, block->my, 0, *obmc_edged, &best_rd); - - check_block_inter(s, mb_x, mb_y, 0, 0, 0, *obmc_edged, &best_rd); - check_block_inter(s, mb_x, mb_y, tb->mx, tb->my, 0, *obmc_edged, &best_rd); - check_block_inter(s, mb_x, mb_y, lb->mx, lb->my, 0, *obmc_edged, &best_rd); - check_block_inter(s, mb_x, mb_y, rb->mx, rb->my, 0, *obmc_edged, &best_rd); - check_block_inter(s, mb_x, mb_y, bb->mx, bb->my, 0, *obmc_edged, &best_rd); - - /* fullpel ME */ - //FIXME avoid subpel interpol / round to nearest integer - do{ - dia_change=0; - for(i=0; iavctx->dia_size, 1); i++){ - for(j=0; jmx+4*(i-j), block->my+(4*j), 0, *obmc_edged, &best_rd); - dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), 0, *obmc_edged, &best_rd); - dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), 0, *obmc_edged, &best_rd); - dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), 0, *obmc_edged, &best_rd); + check_block_inter(s, mb_x, mb_y, block->mx, block->my, *obmc_edged, &best_rd); + + ref_b= *block; + ref_rd= best_rd; + for(ref=0; ref < s->ref_frames; ref++){ + int16_t (*mvr)[2]= &s->ref_mvs[ref][index]; + if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold + continue; + block->ref= ref; + best_rd= INT_MAX; + + check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], *obmc_edged, &best_rd); + check_block_inter(s, mb_x, mb_y, 0, 0, *obmc_edged, &best_rd); + if(tb) + check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], *obmc_edged, &best_rd); + if(lb) + check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], *obmc_edged, &best_rd); + if(rb) + check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], *obmc_edged, &best_rd); + if(bb) + check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], *obmc_edged, &best_rd); + + /* fullpel ME */ + //FIXME avoid subpel interpol / round to nearest integer + do{ + dia_change=0; + for(i=0; iavctx->dia_size, 1); i++){ + for(j=0; jmx+4*(i-j), block->my+(4*j), *obmc_edged, &best_rd); + dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), *obmc_edged, &best_rd); + dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), *obmc_edged, &best_rd); + dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), *obmc_edged, &best_rd); + } } + }while(dia_change); + /* subpel ME */ + do{ + static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},}; + dia_change=0; + for(i=0; i<8; i++) + dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], *obmc_edged, &best_rd); + }while(dia_change); + //FIXME or try the standard 2 pass qpel or similar + + mvr[0][0]= block->mx; + mvr[0][1]= block->my; + if(ref_rd > best_rd){ + ref_rd= best_rd; + ref_b= *block; } - }while(dia_change); - /* subpel ME */ - do{ - static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},}; - dia_change=0; - for(i=0; i<8; i++) - dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], 0, *obmc_edged, &best_rd); - }while(dia_change); - //FIXME or try the standard 2 pass qpel or similar + } + best_rd= ref_rd; + *block= ref_b; #if 1 check_block(s, mb_x, mb_y, color, 1, *obmc_edged, &best_rd); //FIXME RD style color selection #endif if(!same_block(block, &backup)){ - if(tb != &null_block) tb ->type &= ~BLOCK_OPT; - if(lb != &null_block) lb ->type &= ~BLOCK_OPT; - if(rb != &null_block) rb ->type &= ~BLOCK_OPT; - if(bb != &null_block) bb ->type &= ~BLOCK_OPT; - if(tlb!= &null_block) tlb->type &= ~BLOCK_OPT; - if(trb!= &null_block) trb->type &= ~BLOCK_OPT; - if(blb!= &null_block) blb->type &= ~BLOCK_OPT; - if(brb!= &null_block) brb->type &= ~BLOCK_OPT; + if(tb ) tb ->type &= ~BLOCK_OPT; + if(lb ) lb ->type &= ~BLOCK_OPT; + if(rb ) rb ->type &= ~BLOCK_OPT; + if(bb ) bb ->type &= ~BLOCK_OPT; + if(tlb) tlb->type &= ~BLOCK_OPT; + if(trb) trb->type &= ~BLOCK_OPT; + if(blb) blb->type &= ~BLOCK_OPT; + if(brb) brb->type &= ~BLOCK_OPT; change ++; } } @@ -3450,7 +3366,7 @@ static void iterative_me(SnowContext *s){ int change= 0; for(mb_y= 0; mb_ymx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2, - (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, &best_rd); + (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd); for(i=0; i<4; i++) if(!(b[i]->type&BLOCK_INTRA)) - check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, &best_rd); + check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd); if(init_rd != best_rd) change++; @@ -3490,7 +3407,7 @@ static void quantize(SnowContext *s, SubBand *b, DWTELEM *src, int stride, int b const int level= b->level; const int w= b->width; const int h= b->height; - const int qlog= clip(s->qlog + b->qlog, 0, QROOT*16); + const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16); const int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT); int x,y, thres1, thres2; // START_TIMER @@ -3549,7 +3466,7 @@ static void quantize(SnowContext *s, SubBand *b, DWTELEM *src, int stride, int b static void dequantize_slice_buffered(SnowContext *s, slice_buffer * sb, SubBand *b, DWTELEM *src, int stride, int start_y, int end_y){ const int w= b->width; - const int qlog= clip(s->qlog + b->qlog, 0, QROOT*16); + const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16); const int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT); const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT; int x,y; @@ -3577,7 +3494,7 @@ static void dequantize_slice_buffered(SnowContext *s, slice_buffer * sb, SubBand static void dequantize(SnowContext *s, SubBand *b, DWTELEM *src, int stride){ const int w= b->width; const int h= b->height; - const int qlog= clip(s->qlog + b->qlog, 0, QROOT*16); + const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16); const int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT); const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT; int x,y; @@ -3630,7 +3547,7 @@ static void correlate_slice_buffered(SnowContext *s, slice_buffer * sb, SubBand // START_TIMER - DWTELEM * line; + DWTELEM * line=0; // silence silly "could be used without having been initialized" warning DWTELEM * prev; if (start_y != 0) @@ -3689,8 +3606,14 @@ static void encode_header(SnowContext *s){ memset(kstate, MID_STATE, sizeof(kstate)); put_rac(&s->c, kstate, s->keyframe); - if(s->keyframe || s->always_reset) + if(s->keyframe || s->always_reset){ reset_contexts(s); + s->last_spatial_decomposition_type= + s->last_qlog= + s->last_qbias= + s->last_mv_scale= + s->last_block_max_depth= 0; + } if(s->keyframe){ put_symbol(&s->c, s->header_state, s->version, 0); put_rac(&s->c, s->header_state, s->always_reset); @@ -3702,6 +3625,7 @@ static void encode_header(SnowContext *s){ put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0); put_rac(&s->c, s->header_state, s->spatial_scalability); // put_rac(&s->c, s->header_state, s->rate_scalability); + put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0); for(plane_index=0; plane_index<2; plane_index++){ for(level=0; levelspatial_decomposition_count; level++){ @@ -3712,11 +3636,17 @@ static void encode_header(SnowContext *s){ } } } - put_symbol(&s->c, s->header_state, s->spatial_decomposition_type, 0); - put_symbol(&s->c, s->header_state, s->qlog, 1); - put_symbol(&s->c, s->header_state, s->mv_scale, 0); - put_symbol(&s->c, s->header_state, s->qbias, 1); - put_symbol(&s->c, s->header_state, s->block_max_depth, 0); + put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1); + put_symbol(&s->c, s->header_state, s->qlog - s->last_qlog , 1); + put_symbol(&s->c, s->header_state, s->mv_scale - s->last_mv_scale, 1); + put_symbol(&s->c, s->header_state, s->qbias - s->last_qbias , 1); + put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1); + + s->last_spatial_decomposition_type= s->spatial_decomposition_type; + s->last_qlog = s->qlog; + s->last_qbias = s->qbias; + s->last_mv_scale = s->mv_scale; + s->last_block_max_depth = s->block_max_depth; } static int decode_header(SnowContext *s){ @@ -3726,8 +3656,14 @@ static int decode_header(SnowContext *s){ memset(kstate, MID_STATE, sizeof(kstate)); s->keyframe= get_rac(&s->c, kstate); - if(s->keyframe || s->always_reset) + if(s->keyframe || s->always_reset){ reset_contexts(s); + s->spatial_decomposition_type= + s->qlog= + s->qbias= + s->mv_scale= + s->block_max_depth= 0; + } if(s->keyframe){ s->version= get_symbol(&s->c, s->header_state, 0); if(s->version>0){ @@ -3743,6 +3679,7 @@ static int decode_header(SnowContext *s){ s->chroma_v_shift= get_symbol(&s->c, s->header_state, 0); s->spatial_scalability= get_rac(&s->c, s->header_state); // s->rate_scalability= get_rac(&s->c, s->header_state); + s->max_ref_frames= get_symbol(&s->c, s->header_state, 0)+1; for(plane_index=0; plane_index<3; plane_index++){ for(level=0; levelspatial_decomposition_count; level++){ @@ -3757,17 +3694,17 @@ static int decode_header(SnowContext *s){ } } - s->spatial_decomposition_type= get_symbol(&s->c, s->header_state, 0); + s->spatial_decomposition_type+= get_symbol(&s->c, s->header_state, 1); if(s->spatial_decomposition_type > 2){ av_log(s->avctx, AV_LOG_ERROR, "spatial_decomposition_type %d not supported", s->spatial_decomposition_type); return -1; } - s->qlog= get_symbol(&s->c, s->header_state, 1); - s->mv_scale= get_symbol(&s->c, s->header_state, 0); - s->qbias= get_symbol(&s->c, s->header_state, 1); - s->block_max_depth= get_symbol(&s->c, s->header_state, 0); - if(s->block_max_depth > 1){ + s->qlog += get_symbol(&s->c, s->header_state, 1); + s->mv_scale += get_symbol(&s->c, s->header_state, 1); + s->qbias += get_symbol(&s->c, s->header_state, 1); + s->block_max_depth+= get_symbol(&s->c, s->header_state, 1); + if(s->block_max_depth > 1 || s->block_max_depth < 0){ av_log(s->avctx, AV_LOG_ERROR, "block_max_depth= %d is too large", s->block_max_depth); s->block_max_depth= 0; return -1; @@ -3776,7 +3713,7 @@ static int decode_header(SnowContext *s){ return 0; } -static void init_qexp(){ +static void init_qexp(void){ int i; double v=128; @@ -3790,6 +3727,7 @@ static int common_init(AVCodecContext *avctx){ SnowContext *s = avctx->priv_data; int width, height; int level, orientation, plane_index, dec; + int i, j; s->avctx= avctx; @@ -3895,6 +3833,10 @@ static int common_init(AVCodecContext *avctx){ } } + for(i=0; iwidth= avctx->width; @@ -3907,6 +3849,60 @@ static int common_init(AVCodecContext *avctx){ return 0; } +static int qscale2qlog(int qscale){ + return rint(QROOT*log(qscale / (float)FF_QP2LAMBDA)/log(2)) + + 61*QROOT/8; //<64 >60 +} + +static int ratecontrol_1pass(SnowContext *s, AVFrame *pict) +{ + /* estimate the frame's complexity as a sum of weighted dwt coefs. + * FIXME we know exact mv bits at this point, + * but ratecontrol isn't set up to include them. */ + uint32_t coef_sum= 0; + int level, orientation, delta_qlog; + + for(level=0; levelspatial_decomposition_count; level++){ + for(orientation=level ? 1 : 0; orientation<4; orientation++){ + SubBand *b= &s->plane[0].band[level][orientation]; + DWTELEM *buf= b->buf; + const int w= b->width; + const int h= b->height; + const int stride= b->stride; + const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16); + const int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT); + const int qdiv= (1<<16)/qmul; + int x, y; + if(orientation==0) + decorrelate(s, b, buf, stride, 1, 0); + for(y=0; y> 16; + if(orientation==0) + correlate(s, b, buf, stride, 1, 0); + } + } + + /* ugly, ratecontrol just takes a sqrt again */ + coef_sum = (uint64_t)coef_sum * coef_sum >> 16; + assert(coef_sum < INT_MAX); + + if(pict->pict_type == I_TYPE){ + s->m.current_picture.mb_var_sum= coef_sum; + s->m.current_picture.mc_mb_var_sum= 0; + }else{ + s->m.current_picture.mc_mb_var_sum= coef_sum; + s->m.current_picture.mb_var_sum= 0; + } + + pict->quality= ff_rate_estimate_qscale(&s->m, 1); + if (pict->quality < 0) + return INT_MIN; + s->lambda= pict->quality * 3/2; + delta_qlog= qscale2qlog(pict->quality) - s->qlog; + s->qlog+= delta_qlog; + return delta_qlog; +} static void calculate_vissual_weight(SnowContext *s, Plane *p){ int width = p->width; @@ -3946,6 +3942,13 @@ static int encode_init(AVCodecContext *avctx) return -1; } + if(avctx->prediction_method == DWT_97 + && (avctx->flags & CODEC_FLAG_QSCALE) + && avctx->global_quality == 0){ + av_log(avctx, AV_LOG_ERROR, "the 9/7 wavelet is incompatible with lossless mode\n"); + return -1; + } + common_init(avctx); alloc_blocks(s); @@ -3961,14 +3964,17 @@ static int encode_init(AVCodecContext *avctx) s->m.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t)); h263_encode_init(&s->m); //mv_penalty + s->max_ref_frames = FFMAX(FFMIN(avctx->refs, MAX_REF_FRAMES), 1); + if(avctx->flags&CODEC_FLAG_PASS1){ if(!avctx->stats_out) avctx->stats_out = av_mallocz(256); } - if(avctx->flags&CODEC_FLAG_PASS2){ + if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){ if(ff_rate_control_init(&s->m) < 0) return -1; } + s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2)); for(plane_index=0; plane_index<3; plane_index++){ calculate_vissual_weight(s, &s->plane[plane_index]); @@ -3985,7 +3991,7 @@ static int encode_init(AVCodecContext *avctx) // case PIX_FMT_YUV410P: s->colorspace_type= 0; break; -/* case PIX_FMT_RGBA32: +/* case PIX_FMT_RGB32: s->colorspace= 1; break;*/ default: @@ -4001,6 +4007,15 @@ static int encode_init(AVCodecContext *avctx) s->avctx->get_buffer(s->avctx, &s->input_picture); + if(s->avctx->me_method == ME_ITER){ + int i; + int size= s->b_width * s->b_height << 2*s->block_max_depth; + for(i=0; imax_ref_frames; i++){ + s->ref_mvs[i]= av_mallocz(size*sizeof(int16_t[2])); + s->ref_scores[i]= av_mallocz(size*sizeof(uint32_t)); + } + } + return 0; } @@ -4015,16 +4030,29 @@ static int frame_start(SnowContext *s){ draw_edges(s->current_picture.data[2], s->current_picture.linesize[2], w>>1, h>>1, EDGE_WIDTH/2); } - tmp= s->last_picture; - s->last_picture= s->current_picture; + tmp= s->last_picture[s->max_ref_frames-1]; + memmove(s->last_picture+1, s->last_picture, (s->max_ref_frames-1)*sizeof(AVFrame)); + s->last_picture[0]= s->current_picture; s->current_picture= tmp; + if(s->keyframe){ + s->ref_frames= 0; + }else{ + int i; + for(i=0; imax_ref_frames && s->last_picture[i].data[0]; i++) + if(i && s->last_picture[i-1].key_frame) + break; + s->ref_frames= i; + } + s->current_picture.reference= 1; if(s->avctx->get_buffer(s->avctx, &s->current_picture) < 0){ av_log(s->avctx, AV_LOG_ERROR, "get_buffer() failed\n"); return -1; } + s->current_picture.key_frame= s->keyframe; + return 0; } @@ -4035,6 +4063,8 @@ static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size, const int width= s->avctx->width; const int height= s->avctx->height; int level, orientation, plane_index, i, y; + uint8_t rc_header_bak[sizeof(s->header_state)]; + uint8_t rc_block_bak[sizeof(s->block_state)]; ff_init_range_encoder(c, buf, buf_size); ff_build_rac_states(c, 0.05*(1LL<<32), 256-8); @@ -4048,27 +4078,34 @@ static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size, } s->new_picture = *pict; + s->m.picture_number= avctx->frame_number; if(avctx->flags&CODEC_FLAG_PASS2){ s->m.pict_type = pict->pict_type= s->m.rc_context.entry[avctx->frame_number].new_pict_type; s->keyframe= pict->pict_type==FF_I_TYPE; - s->m.picture_number= avctx->frame_number; - pict->quality= ff_rate_estimate_qscale(&s->m, 0); + if(!(avctx->flags&CODEC_FLAG_QSCALE)) { + pict->quality= ff_rate_estimate_qscale(&s->m, 0); + if (pict->quality < 0) + return -1; + } }else{ s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0; + s->m.pict_type= pict->pict_type= s->keyframe ? FF_I_TYPE : FF_P_TYPE; } + if(s->pass1_rc && avctx->frame_number == 0) + pict->quality= 2*FF_QP2LAMBDA; if(pict->quality){ - s->qlog= rint(QROOT*log(pict->quality / (float)FF_QP2LAMBDA)/log(2)); - //<64 >60 - s->qlog += 61*QROOT/8; - }else{ - s->qlog= LOSSLESS_QLOG; + s->qlog= qscale2qlog(pict->quality); + s->lambda = pict->quality * 3/2; } + if(s->qlog < 0 || (!pict->quality && (avctx->flags & CODEC_FLAG_QSCALE))){ + s->qlog= LOSSLESS_QLOG; + s->lambda = 0; + }//else keep previous frame's qlog until after motion est frame_start(s); - s->current_picture.key_frame= s->keyframe; s->m.current_picture_ptr= &s->m.current_picture; if(pict->pict_type == P_TYPE){ @@ -4077,11 +4114,11 @@ static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size, int stride= s->current_picture.linesize[0]; assert(s->current_picture.data[0]); - assert(s->last_picture.data[0]); + assert(s->last_picture[0].data[0]); s->m.avctx= s->avctx; s->m.current_picture.data[0]= s->current_picture.data[0]; - s->m. last_picture.data[0]= s-> last_picture.data[0]; + s->m. last_picture.data[0]= s->last_picture[0].data[0]; s->m. new_picture.data[0]= s-> input_picture.data[0]; s->m. last_picture_ptr= &s->m. last_picture; s->m.linesize= @@ -4104,7 +4141,7 @@ static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size, s->m.out_format= FMT_H263; s->m.unrestricted_mv= 1; - s->lambda = s->m.lambda= pict->quality * 3/2; //FIXME bug somewhere else + s->m.lambda = s->lambda; s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7); s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT; @@ -4113,13 +4150,19 @@ static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size, s->dsp= s->m.dsp; } + if(s->pass1_rc){ + memcpy(rc_header_bak, s->header_state, sizeof(s->header_state)); + memcpy(rc_block_bak, s->block_state, sizeof(s->block_state)); + } + redo_frame: + s->m.pict_type = pict->pict_type; s->qbias= pict->pict_type == P_TYPE ? 2 : 0; encode_header(s); s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start); - encode_blocks(s); + encode_blocks(s, 1); s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits; for(plane_index=0; plane_index<3; plane_index++){ @@ -4129,6 +4172,7 @@ redo_frame: int x, y; // int bits= put_bits_count(&s->c.pb); + if(!(avctx->flags2 & CODEC_FLAG2_MEMC_ONLY)){ //FIXME optimize if(pict->data[plane_index]) //FIXME gray hack for(y=0; ypict_type == P_TYPE + && !(avctx->flags&CODEC_FLAG_PASS2) && s->m.me.scene_change_score > s->avctx->scenechange_threshold){ ff_init_range_encoder(c, buf, buf_size); ff_build_rac_states(c, 0.05*(1LL<<32), 256-8); pict->pict_type= FF_I_TYPE; s->keyframe=1; - reset_contexts(s); + s->current_picture.key_frame=1; goto redo_frame; } @@ -4159,6 +4204,20 @@ redo_frame: ff_spatial_dwt(s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count); + if(s->pass1_rc && plane_index==0){ + int delta_qlog = ratecontrol_1pass(s, pict); + if (delta_qlog <= INT_MIN) + return -1; + if(delta_qlog){ + //reordering qlog in the bitstream would eliminate this reset + ff_init_range_encoder(c, buf, buf_size); + memcpy(s->header_state, rc_header_bak, sizeof(s->header_state)); + memcpy(s->block_state, rc_block_bak, sizeof(s->block_state)); + encode_header(s); + encode_blocks(s, 0); + } + } + for(level=0; levelspatial_decomposition_count; level++){ for(orientation=level ? 1 : 0; orientation<4; orientation++){ SubBand *b= &p->band[level][orientation]; @@ -4193,6 +4252,20 @@ redo_frame: {START_TIMER predict_plane(s, s->spatial_dwt_buffer, plane_index, 1); STOP_TIMER("pred-conv")} + }else{ + //ME/MC only + if(pict->pict_type == I_TYPE){ + for(y=0; ycurrent_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x]= + pict->data[plane_index][y*pict->linesize[plane_index] + x]; + } + } + }else{ + memset(s->spatial_dwt_buffer, 0, sizeof(DWTELEM)*w*h); + predict_plane(s, s->spatial_dwt_buffer, plane_index, 1); + } + } if(s->avctx->flags&CODEC_FLAG_PSNR){ int64_t error= 0; @@ -4208,23 +4281,28 @@ STOP_TIMER("pred-conv")} } } - if(s->last_picture.data[0]) - avctx->release_buffer(avctx, &s->last_picture); + if(s->last_picture[s->max_ref_frames-1].data[0]) + avctx->release_buffer(avctx, &s->last_picture[s->max_ref_frames-1]); s->current_picture.coded_picture_number = avctx->frame_number; s->current_picture.pict_type = pict->pict_type; s->current_picture.quality = pict->quality; - if(avctx->flags&CODEC_FLAG_PASS1){ - s->m.p_tex_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits - s->m.mv_bits; - s->m.current_picture.display_picture_number = - s->m.current_picture.coded_picture_number = avctx->frame_number; - s->m.pict_type = pict->pict_type; - s->m.current_picture.quality = pict->quality; + s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start); + s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits; + s->m.current_picture.display_picture_number = + s->m.current_picture.coded_picture_number = avctx->frame_number; + s->m.current_picture.quality = pict->quality; + s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start); + if(s->pass1_rc) + if (ff_rate_estimate_qscale(&s->m, 0) < 0) + return -1; + if(avctx->flags&CODEC_FLAG_PASS1) ff_write_pass1_stats(&s->m); - } - if(avctx->flags&CODEC_FLAG_PASS2){ - s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start); - } + s->m.last_pict_type = s->m.pict_type; + avctx->frame_bits = s->m.frame_bits; + avctx->mv_bits = s->m.mv_bits; + avctx->misc_bits = s->m.misc_bits; + avctx->p_tex_bits = s->m.p_tex_bits; emms_c(); @@ -4232,7 +4310,7 @@ STOP_TIMER("pred-conv")} } static void common_end(SnowContext *s){ - int plane_index, level, orientation; + int plane_index, level, orientation, i; av_freep(&s->spatial_dwt_buffer); @@ -4243,6 +4321,13 @@ static void common_end(SnowContext *s){ av_freep(&s->block); + for(i=0; iref_mvs[i]); + av_freep(&s->ref_scores[i]); + if(s->last_picture[i].data[0]) + s->avctx->release_buffer(s->avctx, &s->last_picture[i]); + } + for(plane_index=0; plane_index<3; plane_index++){ for(level=s->spatial_decomposition_count-1; level>=0; level--){ for(orientation=level ? 1 : 0; orientation<4; orientation++){ @@ -4387,7 +4472,7 @@ if(s->avctx->debug&2048){ { START_TIMER for(; ydsb, w, h, 1, s->spatial_decomposition_type, s->spatial_decomposition_count, yd); + ff_spatial_idwt_buffered_slice(&s->dsp, cs, &s->sb, w, h, 1, s->spatial_decomposition_type, s->spatial_decomposition_count, yd); } STOP_TIMER("idwt slice");} @@ -4416,8 +4501,8 @@ STOP_TIMER("idwt + predict_slices")} emms_c(); - if(s->last_picture.data[0]) - avctx->release_buffer(avctx, &s->last_picture); + if(s->last_picture[s->max_ref_frames-1].data[0]) + avctx->release_buffer(avctx, &s->last_picture[s->max_ref_frames-1]); if(!(s->avctx->debug&2048)) *picture= s->current_picture; @@ -4502,7 +4587,7 @@ int main(){ ff_spatial_idwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count); for(i=0; i20) printf("fsck: %d %d %d\n",i, buffer[0][i], buffer[1][i]); + if(FFABS(buffer[0][i] - buffer[1][i])>20) printf("fsck: %d %d %d\n",i, buffer[0][i], buffer[1][i]); #if 0 printf("testing AC coder\n"); @@ -4512,7 +4597,7 @@ int main(){ for(i=-256; i<256; i++){ START_TIMER - put_symbol(&s.c, s.header_state, i*i*i/3*ABS(i), 1); + put_symbol(&s.c, s.header_state, i*i*i/3*FFABS(i), 1); STOP_TIMER("put_symbol") } ff_rac_terminate(&s.c); @@ -4526,7 +4611,7 @@ STOP_TIMER("put_symbol") START_TIMER j= get_symbol(&s.c, s.header_state, 1); STOP_TIMER("get_symbol") - if(j!=i*i*i/3*ABS(i)) printf("fsck: %d != %d\n", i, j); + if(j!=i*i*i/3*FFABS(i)) printf("fsck: %d != %d\n", i, j); } #endif { @@ -4555,9 +4640,9 @@ int64_t g=0; for(x=0; x