X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libavcodec%2Fsnow.c;h=0300e58c7ce5c2af898e2e0579dea78cd692f2f9;hb=a27d6e714f850ead1505a805cacc5d6479b2de76;hp=58560e9570f78a81f5a24b63e55f96cabf2a961e;hpb=87f20c2f314c9dd426d1f64b300ebc7a656202f8;p=ffmpeg diff --git a/libavcodec/snow.c b/libavcodec/snow.c index 58560e9570f..0300e58c7ce 100644 --- a/libavcodec/snow.c +++ b/libavcodec/snow.c @@ -1,41 +1,34 @@ /* * 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<line_count = line_count; buf->line_width = line_width; buf->data_count = max_allocated_lines; - buf->line = (DWTELEM * *) av_mallocz (sizeof(DWTELEM *) * line_count); - buf->data_stack = (DWTELEM * *) av_malloc (sizeof(DWTELEM *) * max_allocated_lines); + buf->line = av_mallocz (sizeof(IDWTELEM *) * line_count); + buf->data_stack = av_malloc (sizeof(IDWTELEM *) * max_allocated_lines); for (i = 0; i < max_allocated_lines; i++) { - buf->data_stack[i] = (DWTELEM *) av_malloc (sizeof(DWTELEM) * line_width); + buf->data_stack[i] = av_malloc (sizeof(IDWTELEM) * line_width); } buf->data_stack_top = max_allocated_lines - 1; } -static DWTELEM * slice_buffer_load_line(slice_buffer * buf, int line) +static IDWTELEM * slice_buffer_load_line(slice_buffer * buf, int line) { int offset; - DWTELEM * buffer; + IDWTELEM * buffer; // av_log(NULL, AV_LOG_DEBUG, "Cache hit: %d\n", line); @@ -538,7 +547,7 @@ static DWTELEM * slice_buffer_load_line(slice_buffer * buf, int line) static void slice_buffer_release(slice_buffer * buf, int line) { int offset; - DWTELEM * buffer; + IDWTELEM * buffer; assert(line >= 0 && line < buf->line_count); assert(buf->line[line]); @@ -572,13 +581,10 @@ static void slice_buffer_destroy(slice_buffer * buf) for (i = buf->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); - assert(buf->line); - av_free(buf->line); + av_freep(&buf->data_stack); + av_freep(&buf->line); } #ifdef __sgi @@ -600,7 +606,7 @@ 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 +725,11 @@ 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); @@ -733,52 +743,68 @@ static always_inline void lift(DWTELEM *dst, DWTELEM *src, DWTELEM *ref, int dst } for(i=0; i>shift), inverse); + dst[i*dst_step] = + LIFT(src[i*src_step], + ((mul*(ref[i*ref_step] + ref[(i+1)*ref_step])+add)>>shift), + inverse); } if(mirror_right){ - dst[w*dst_step] = LIFT(src[w*src_step], ((mul*2*ref[w*ref_step]+add)>>shift), inverse); + dst[w*dst_step] = + LIFT(src[w*src_step], + ((mul*2*ref[w*ref_step]+add)>>shift), + inverse); } } -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){ +static av_always_inline void +inv_lift(IDWTELEM *dst, IDWTELEM *src, IDWTELEM *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); int i; +#define LIFT(src, ref, inv) ((src) + ((inv) ? - (ref) : + (ref))) if(mirror_left){ - int r= 3*2*ref[0]; - r += r>>4; - r += r>>8; - dst[0] = LIFT(src[0], ((r+add)>>shift), inverse); + dst[0] = LIFT(src[0], ((mul*2*ref[0]+add)>>shift), inverse); dst += dst_step; src += src_step; } for(i=0; i>4; - r += r>>8; - dst[i*dst_step] = LIFT(src[i*src_step], ((r+add)>>shift), inverse); + dst[i*dst_step] = + LIFT(src[i*src_step], + ((mul*(ref[i*ref_step] + ref[(i+1)*ref_step])+add)>>shift), + inverse); } if(mirror_right){ - int r= 3*2*ref[w*ref_step]; - r += r>>4; - r += r>>8; - dst[w*dst_step] = LIFT(src[w*src_step], ((r+add)>>shift), inverse); + dst[w*dst_step] = + LIFT(src[w*src_step], + ((mul*2*ref[w*ref_step]+add)>>shift), + inverse); } } -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); int i; assert(shift == 4); -#define LIFTS(src, ref, inv) ((inv) ? (src) - (((ref) - 4*(src))>>shift): (16*4*(src) + 4*(ref) + 8 + (5<<27))/(5*16) - (1<<23)) +#define LIFTS(src, ref, inv) \ + ((inv) ? \ + (src) + (((ref) + 4*(src))>>shift): \ + -((-16*(src) + (ref) + add/4 + 1 + (5<<25))/(5*4) - (1<<23))) if(mirror_left){ dst[0] = LIFTS(src[0], mul*2*ref[0]+add, inverse); dst += dst_step; @@ -786,241 +812,51 @@ static always_inline void liftS(DWTELEM *dst, DWTELEM *src, DWTELEM *ref, int ds } for(i=0; i=width) x2= 2*width-x2-2; - sum += coeffs[i]*(int64_t)dst[x2]; - } - if(inverse) dst[x] -= (sum + (1<>shift; - else dst[x] += (sum + (1<>shift; - } -} - -static void inplace_liftV(DWTELEM *dst, int width, int height, int stride, int *coeffs, int n, int shift, int start, int inverse){ - int x, y, i; - for(y=start; y=height) y2= 2*height-y2-2; - sum += coeffs[i]*(int64_t)dst[x + y2*stride]; - } - if(inverse) dst[x + y*stride] -= (sum + (1<>shift; - else dst[x + y*stride] += (sum + (1<>shift; - } + dst[w*dst_step] = + LIFTS(src[w*src_step], mul*2*ref[w*ref_step]+add, inverse); } } +static av_always_inline void +inv_liftS(IDWTELEM *dst, IDWTELEM *src, IDWTELEM *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); + int i; -#define SCALEX 1 -#define LX0 0 -#define LX1 1 - -#if 0 // more accurate 9/7 -#define N1 2 -#define SHIFT1 14 -#define COEFFS1 (int[]){-25987,-25987} -#define N2 2 -#define SHIFT2 19 -#define COEFFS2 (int[]){-27777,-27777} -#define N3 2 -#define SHIFT3 15 -#define COEFFS3 (int[]){28931,28931} -#define N4 2 -#define SHIFT4 15 -#define COEFFS4 (int[]){14533,14533} -#elif 1 // 13/7 CRF -#define N1 4 -#define SHIFT1 4 -#define COEFFS1 (int[]){1,-9,-9,1} -#define N2 4 -#define SHIFT2 4 -#define COEFFS2 (int[]){-1,5,5,-1} -#define N3 0 -#define SHIFT3 1 -#define COEFFS3 NULL -#define N4 0 -#define SHIFT4 1 -#define COEFFS4 NULL -#elif 1 // 3/5 -#define LX0 1 -#define LX1 0 -#define SCALEX 0.5 -#define N1 2 -#define SHIFT1 1 -#define COEFFS1 (int[]){1,1} -#define N2 2 -#define SHIFT2 2 -#define COEFFS2 (int[]){-1,-1} -#define N3 0 -#define SHIFT3 0 -#define COEFFS3 NULL -#define N4 0 -#define SHIFT4 0 -#define COEFFS4 NULL -#elif 1 // 11/5 -#define N1 0 -#define SHIFT1 1 -#define COEFFS1 NULL -#define N2 2 -#define SHIFT2 2 -#define COEFFS2 (int[]){-1,-1} -#define N3 2 -#define SHIFT3 0 -#define COEFFS3 (int[]){-1,-1} -#define N4 4 -#define SHIFT4 7 -#define COEFFS4 (int[]){-5,29,29,-5} -#define SCALEX 4 -#elif 1 // 9/7 CDF -#define N1 2 -#define SHIFT1 7 -#define COEFFS1 (int[]){-203,-203} -#define N2 2 -#define SHIFT2 12 -#define COEFFS2 (int[]){-217,-217} -#define N3 2 -#define SHIFT3 7 -#define COEFFS3 (int[]){113,113} -#define N4 2 -#define SHIFT4 9 -#define COEFFS4 (int[]){227,227} -#define SCALEX 1 -#elif 1 // 7/5 CDF -#define N1 0 -#define SHIFT1 1 -#define COEFFS1 NULL -#define N2 2 -#define SHIFT2 2 -#define COEFFS2 (int[]){-1,-1} -#define N3 2 -#define SHIFT3 0 -#define COEFFS3 (int[]){-1,-1} -#define N4 2 -#define SHIFT4 4 -#define COEFFS4 (int[]){3,3} -#elif 1 // 9/7 MN -#define N1 4 -#define SHIFT1 4 -#define COEFFS1 (int[]){1,-9,-9,1} -#define N2 2 -#define SHIFT2 2 -#define COEFFS2 (int[]){1,1} -#define N3 0 -#define SHIFT3 1 -#define COEFFS3 NULL -#define N4 0 -#define SHIFT4 1 -#define COEFFS4 NULL -#else // 13/7 CRF -#define N1 4 -#define SHIFT1 4 -#define COEFFS1 (int[]){1,-9,-9,1} -#define N2 4 -#define SHIFT2 4 -#define COEFFS2 (int[]){-1,5,5,-1} -#define N3 0 -#define SHIFT3 1 -#define COEFFS3 NULL -#define N4 0 -#define SHIFT4 1 -#define COEFFS4 NULL -#endif -static void horizontal_decomposeX(DWTELEM *b, int width){ - DWTELEM temp[width]; - const int width2= width>>1; - const int w2= (width+1)>>1; - int x; - - inplace_lift(b, width, COEFFS1, N1, SHIFT1, LX1, 0); - inplace_lift(b, width, COEFFS2, N2, SHIFT2, LX0, 0); - inplace_lift(b, width, COEFFS3, N3, SHIFT3, LX1, 0); - inplace_lift(b, width, COEFFS4, N4, SHIFT4, LX0, 0); - - for(x=0; x>1; - int x; - const int w2= (width+1)>>1; - - memcpy(temp, b, width*sizeof(int)); - for(x=0; x>shift): \ + -((-16*(src) + (ref) + add/4 + 1 + (5<<25))/(5*4) - (1<<23))) + if(mirror_left){ + dst[0] = LIFTS(src[0], mul*2*ref[0]+add, inverse); + dst += dst_step; + src += src_step; } - inplace_liftV(buffer, width, height, stride, COEFFS1, N1, SHIFT1, LX1, 0); - inplace_liftV(buffer, width, height, stride, COEFFS2, N2, SHIFT2, LX0, 0); - inplace_liftV(buffer, width, height, stride, COEFFS3, N3, SHIFT3, LX1, 0); - inplace_liftV(buffer, width, height, stride, COEFFS4, N4, SHIFT4, LX0, 0); -} - -static void spatial_composeX(DWTELEM *buffer, int width, int height, int stride){ - int x, y; - - inplace_liftV(buffer, width, height, stride, COEFFS4, N4, SHIFT4, LX0, 1); - inplace_liftV(buffer, width, height, stride, COEFFS3, N3, SHIFT3, LX1, 1); - inplace_liftV(buffer, width, height, stride, COEFFS2, N2, SHIFT2, LX0, 1); - inplace_liftV(buffer, width, height, stride, COEFFS1, N1, SHIFT1, LX1, 1); - - for(y=0; y>1; - lift (temp+w2, b +1, b , 1, 2, 2, width, -W_AM, W_AO, W_AS, 1, 0); - liftS(temp , b , temp+w2, 1, 2, 1, width, -W_BM, W_BO, W_BS, 0, 0); - lift5(b +w2, temp+w2, temp , 1, 1, 1, width, W_CM, W_CO, W_CS, 1, 0); + lift (temp+w2, b +1, b , 1, 2, 2, width, W_AM, W_AO, W_AS, 1, 1); + liftS(temp , b , temp+w2, 1, 2, 1, width, W_BM, W_BO, W_BS, 0, 0); + lift (b +w2, temp+w2, temp , 1, 1, 1, width, W_CM, W_CO, W_CS, 1, 0); lift (b , temp , b +w2, 1, 1, 1, width, W_DM, W_DO, W_DS, 0, 0); } @@ -1204,14 +970,7 @@ static void vertical_decompose97iH1(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int w int i; for(i=0; i>W_CS; -#else - int r= 3*(b0[i] + b2[i]); - r+= r>>4; - r+= r>>8; - b1[i] += (r+W_CO)>>W_CS; -#endif } } @@ -1222,7 +981,7 @@ static void vertical_decompose97iL0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int w #ifdef liftS b1[i] -= (W_BM*(b0[i] + b2[i])+W_BO)>>W_BS; #else - b1[i] = (16*4*b1[i] - 4*(b0[i] + b2[i]) + 8*5 + (5<<27)) / (5*16) - (1<<23); + b1[i] = (16*4*b1[i] - 4*(b0[i] + b2[i]) + W_BO*5 + (5<<27)) / (5*16) - (1<<23); #endif } } @@ -1275,15 +1034,14 @@ 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<>1; const int w2= (width+1)>>1; int x; @@ -1318,8 +1076,8 @@ static void horizontal_compose53i(DWTELEM *b, int width){ b[width -1] = A3; b[width2-1] = A2; #else - lift(temp , b , b+w2, 1, 1, 1, width, 1, 2, 2, 0, 1); - lift(temp+w2, b+w2, temp, 1, 1, 1, width, -1, 0, 1, 1, 1); + inv_lift(temp , b , b+w2, 1, 1, 1, width, 1, 2, 2, 0, 1); + inv_lift(temp+w2, b+w2, temp, 1, 1, 1, width, -1, 0, 1, 1, 1); #endif for(x=0; xy = -1; } -static void spatial_compose53i_init(dwt_compose_t *cs, DWTELEM *buffer, int height, int stride){ +static void spatial_compose53i_init(dwt_compose_t *cs, IDWTELEM *buffer, int height, int stride){ cs->b0 = buffer + mirror(-1-1, height-1)*stride; cs->b1 = buffer + mirror(-1 , height-1)*stride; cs->y = -1; @@ -1360,10 +1118,10 @@ static void spatial_compose53i_init(dwt_compose_t *cs, DWTELEM *buffer, int heig static void spatial_compose53i_dy_buffered(dwt_compose_t *cs, slice_buffer * sb, int width, int height, int stride_line){ int y= cs->y; - DWTELEM *b0= cs->b0; - DWTELEM *b1= cs->b1; - DWTELEM *b2= slice_buffer_get_line(sb, mirror(y+1, height-1) * stride_line); - DWTELEM *b3= slice_buffer_get_line(sb, mirror(y+2, height-1) * stride_line); + IDWTELEM *b0= cs->b0; + IDWTELEM *b1= cs->b1; + IDWTELEM *b2= slice_buffer_get_line(sb, mirror(y+1, height-1) * stride_line); + IDWTELEM *b3= slice_buffer_get_line(sb, mirror(y+2, height-1) * stride_line); {START_TIMER if(y+1<(unsigned)height) vertical_compose53iL0(b1, b2, b3, width); @@ -1380,12 +1138,12 @@ STOP_TIMER("horizontal_compose53i")} cs->y += 2; } -static void spatial_compose53i_dy(dwt_compose_t *cs, DWTELEM *buffer, int width, int height, int stride){ +static void spatial_compose53i_dy(dwt_compose_t *cs, IDWTELEM *buffer, int width, int height, int stride){ int y= cs->y; - DWTELEM *b0= cs->b0; - DWTELEM *b1= cs->b1; - DWTELEM *b2= buffer + mirror(y+1, height-1)*stride; - DWTELEM *b3= buffer + mirror(y+2, height-1)*stride; + IDWTELEM *b0= cs->b0; + IDWTELEM *b1= cs->b1; + IDWTELEM *b2= buffer + mirror(y+1, height-1)*stride; + IDWTELEM *b3= buffer + mirror(y+2, height-1)*stride; {START_TIMER if(y+1<(unsigned)height) vertical_compose53iL0(b1, b2, b3, width); @@ -1402,7 +1160,7 @@ STOP_TIMER("horizontal_compose53i")} cs->y += 2; } -static void spatial_compose53i(DWTELEM *buffer, int width, int height, int stride){ +static void spatial_compose53i(IDWTELEM *buffer, int width, int height, int stride){ dwt_compose_t cs; spatial_compose53i_init(&cs, buffer, height, stride); while(cs.y <= height) @@ -1410,17 +1168,17 @@ static void spatial_compose53i(DWTELEM *buffer, int width, int height, int strid } -static void horizontal_compose97i(DWTELEM *b, int width){ - DWTELEM temp[width]; +void ff_snow_horizontal_compose97i(IDWTELEM *b, int width){ + IDWTELEM temp[width]; const int w2= (width+1)>>1; - lift (temp , b , b +w2, 1, 1, 1, width, W_DM, W_DO, W_DS, 0, 1); - lift5(temp+w2, b +w2, temp , 1, 1, 1, width, W_CM, W_CO, W_CS, 1, 1); - liftS(b , temp , temp+w2, 2, 1, 1, width, -W_BM, W_BO, W_BS, 0, 1); - lift (b+1 , temp+w2, b , 2, 1, 2, width, -W_AM, W_AO, W_AS, 1, 1); + inv_lift (temp , b , b +w2, 1, 1, 1, width, W_DM, W_DO, W_DS, 0, 1); + inv_lift (temp+w2, b +w2, temp , 1, 1, 1, width, W_CM, W_CO, W_CS, 1, 1); + inv_liftS(b , temp , temp+w2, 2, 1, 1, width, W_BM, W_BO, W_BS, 0, 1); + inv_lift (b+1 , temp+w2, b , 2, 1, 2, width, W_AM, W_AO, W_AS, 1, 0); } -static void vertical_compose97iH0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){ +static void vertical_compose97iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width){ int i; for(i=0; i>W_CS; -#else - int r= 3*(b0[i] + b2[i]); - r+= r>>4; - r+= r>>8; - b1[i] -= (r+W_CO)>>W_CS; -#endif } } -static void vertical_compose97iL0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){ +static void vertical_compose97iL0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width){ int i; for(i=0; i>W_DS; -#ifdef lift5 b3[i] -= (W_CM*(b2[i] + b4[i])+W_CO)>>W_CS; -#else - r= 3*(b2[i] + b4[i]); - r+= r>>4; - r+= r>>8; - b3[i] -= (r+W_CO)>>W_CS; -#endif #ifdef liftS b2[i] += (W_BM*(b1[i] + b3[i])+W_BO)>>W_BS; #else @@ -1496,7 +1237,7 @@ static void spatial_compose97i_buffered_init(dwt_compose_t *cs, slice_buffer * s cs->y = -3; } -static void spatial_compose97i_init(dwt_compose_t *cs, DWTELEM *buffer, int height, int stride){ +static void spatial_compose97i_init(dwt_compose_t *cs, IDWTELEM *buffer, int height, int stride){ cs->b0 = buffer + mirror(-3-1, height-1)*stride; cs->b1 = buffer + mirror(-3 , height-1)*stride; cs->b2 = buffer + mirror(-3+1, height-1)*stride; @@ -1504,19 +1245,19 @@ static void spatial_compose97i_init(dwt_compose_t *cs, DWTELEM *buffer, int heig cs->y = -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; - DWTELEM *b1= cs->b1; - DWTELEM *b2= cs->b2; - DWTELEM *b3= cs->b3; - DWTELEM *b4= slice_buffer_get_line(sb, mirror(y + 3, height - 1) * stride_line); - DWTELEM *b5= slice_buffer_get_line(sb, mirror(y + 4, height - 1) * stride_line); + IDWTELEM *b0= cs->b0; + IDWTELEM *b1= cs->b1; + IDWTELEM *b2= cs->b2; + IDWTELEM *b3= cs->b3; + IDWTELEM *b4= slice_buffer_get_line(sb, mirror(y + 3, height - 1) * stride_line); + IDWTELEM *b5= slice_buffer_get_line(sb, mirror(y + 4, height - 1) * stride_line); {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 +1268,8 @@ 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")}} @@ -1539,14 +1280,14 @@ STOP_TIMER("horizontal_compose97i")}} cs->y += 2; } -static void spatial_compose97i_dy(dwt_compose_t *cs, DWTELEM *buffer, int width, int height, int stride){ +static void spatial_compose97i_dy(dwt_compose_t *cs, IDWTELEM *buffer, int width, int height, int stride){ int y = cs->y; - DWTELEM *b0= cs->b0; - DWTELEM *b1= cs->b1; - DWTELEM *b2= cs->b2; - DWTELEM *b3= cs->b3; - DWTELEM *b4= buffer + mirror(y+3, height-1)*stride; - DWTELEM *b5= buffer + mirror(y+4, height-1)*stride; + IDWTELEM *b0= cs->b0; + IDWTELEM *b1= cs->b1; + IDWTELEM *b2= cs->b2; + IDWTELEM *b3= cs->b3; + IDWTELEM *b4= buffer + mirror(y+3, height-1)*stride; + IDWTELEM *b5= buffer + mirror(y+4, height-1)*stride; {START_TIMER if(y+3<(unsigned)height) vertical_compose97iL1(b3, b4, b5, width); @@ -1557,8 +1298,8 @@ 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")}} @@ -1569,7 +1310,7 @@ STOP_TIMER("horizontal_compose97i")}} cs->y += 2; } -static void spatial_compose97i(DWTELEM *buffer, int width, int height, int stride){ +static void spatial_compose97i(IDWTELEM *buffer, int width, int height, int stride){ dwt_compose_t cs; spatial_compose97i_init(&cs, buffer, height, stride); while(cs.y <= height) @@ -1580,28 +1321,23 @@ static void ff_spatial_idwt_buffered_init(dwt_compose_t *cs, slice_buffer * sb, 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, height>>level, stride<>level, stride_line<>level, stride_line<=0; level--){ switch(type){ - case 0: spatial_compose97i_init(cs+level, buffer, height>>level, stride<>level, stride<>level, height>>level, stride<>level, stride<>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--) - spatial_composeX (buffer, width>>level, height>>level, stride<width; const int h= b->height; int x, y; @@ -1742,7 +1470,7 @@ 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 +1486,11 @@ 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); } } @@ -1771,7 +1499,7 @@ static int encode_subband_c0run(SnowContext *s, SubBand *b, DWTELEM *src, DWTELE return 0; } -static int encode_subband(SnowContext *s, SubBand *b, DWTELEM *src, DWTELEM *parent, int stride, int orientation){ +static int encode_subband(SnowContext *s, SubBand *b, IDWTELEM *src, IDWTELEM *parent, int stride, int orientation){ // encode_subband_qtree(s, b, src, parent, stride, orientation); // encode_subband_z0run(s, b, src, parent, stride, orientation); return encode_subband_c0run(s, b, src, parent, stride, orientation); @@ -1825,7 +1553,7 @@ 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){ @@ -1881,14 +1609,14 @@ 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; START_TIMER - if(b->buf == s->spatial_dwt_buffer || s->qlog == LOSSLESS_QLOG){ + if(b->ibuf == s->spatial_idwt_buffer || s->qlog == LOSSLESS_QLOG){ qadd= 0; qmul= 1<stride_line + b->buf_y_offset) + b->buf_x_offset; - memset(line, 0, b->width*sizeof(DWTELEM)); + IDWTELEM * line = slice_buffer_get_line(sb, y * b->stride_line + b->buf_y_offset) + b->buf_x_offset; + memset(line, 0, b->width*sizeof(IDWTELEM)); v = b->x_coeff[new_index].coeff; x = b->x_coeff[new_index++].x; while(x < w) @@ -1925,11 +1653,11 @@ 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++){ - for(level=0; levelspatial_decomposition_count; level++){ + for(level=0; levelplane[plane_index].band[level][orientation].state, MID_STATE, sizeof(s->plane[plane_index].band[level][orientation].state)); } @@ -1978,7 +1706,7 @@ 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 +1719,7 @@ 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 +1732,7 @@ 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 +1757,22 @@ 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] @@ -2044,7 +1789,7 @@ static int encode_q_branch(SnowContext *s, int level, int x, int y){ RangeCoder pc, ic; uint8_t *pbbak= s->c.bytestream; uint8_t *pbbak_start= s->c.bytestream_start; - int score, score2, iscore, i_len, p_len, block_s, sum; + int score, score2, iscore, i_len, p_len, block_s, sum, base_bits; const int w= s->b_width << s->block_max_depth; const int h= s->b_height << s->block_max_depth; const int rem_depth= s->block_max_depth - level; @@ -2052,17 +1797,16 @@ 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 +1819,15 @@ 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 +1850,10 @@ 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); @@ -2140,19 +1884,38 @@ 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); - //FIXME if mb_cmp != SSE then intra cant be compared currently and mb_penalty vs. lambda2 + 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 cannot be compared currently and mb_penalty vs. lambda2 // subpel search + base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start); pc= s->c; pc.bytestream_start= pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo @@ -2161,13 +1924,13 @@ 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 - + (-av_log2(pc.range) + av_log2(s->c.range)) - ))>>FF_LAMBDA_SHIFT; + score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT; block_s= block_w*block_w; sum = pix_sum(current_data[0], stride, block_w); @@ -2193,10 +1956,7 @@ static int encode_q_branch(SnowContext *s, int level, int x, int y){ put_symbol(&ic, &i_state[64], cb-pcb, 1); put_symbol(&ic, &i_state[96], cr-pcr, 1); i_len= ic.bytestream - ic.bytestream_start; - iscore += (s->lambda2*(i_len*8 - + (ic.outstanding_count - s->c.outstanding_count)*8 - + (-av_log2(ic.range) + av_log2(s->c.range)) - ))>>FF_LAMBDA_SHIFT; + iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT; // assert(score==256*256*256*64-1); assert(iscore < 255*255*256 + s->lambda2*10); @@ -2226,11 +1986,12 @@ 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 +1999,17 @@ 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 +2019,21 @@ 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 +2050,20 @@ 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 +2072,42 @@ 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 +2116,12 @@ 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); @@ -2381,80 +2150,197 @@ static void decode_blocks(SnowContext *s){ } } -static void mc_block(uint8_t *dst, uint8_t *src, uint8_t *tmp, int stride, int b_w, int b_h, int dx, int dy){ - int x, y; +static void mc_block(Plane *p, uint8_t *dst, const uint8_t *src, uint8_t *tmp, int stride, int b_w, int b_h, int dx, int dy){ + const static uint8_t weight[64]={ + 8,7,6,5,4,3,2,1, + 7,7,0,0,0,0,0,1, + 6,0,6,0,0,0,2,0, + 5,0,0,5,0,3,0,0, + 4,0,0,0,4,0,0,0, + 3,0,0,5,0,3,0,0, + 2,0,6,0,0,0,2,0, + 1,7,0,0,0,0,0,1, + }; + + const static uint8_t brane[256]={ + 0x00,0x01,0x01,0x01,0x01,0x01,0x01,0x01,0x11,0x12,0x12,0x12,0x12,0x12,0x12,0x12, + 0x04,0x05,0xcc,0xcc,0xcc,0xcc,0xcc,0x41,0x15,0x16,0xcc,0xcc,0xcc,0xcc,0xcc,0x52, + 0x04,0xcc,0x05,0xcc,0xcc,0xcc,0x41,0xcc,0x15,0xcc,0x16,0xcc,0xcc,0xcc,0x52,0xcc, + 0x04,0xcc,0xcc,0x05,0xcc,0x41,0xcc,0xcc,0x15,0xcc,0xcc,0x16,0xcc,0x52,0xcc,0xcc, + 0x04,0xcc,0xcc,0xcc,0x41,0xcc,0xcc,0xcc,0x15,0xcc,0xcc,0xcc,0x16,0xcc,0xcc,0xcc, + 0x04,0xcc,0xcc,0x41,0xcc,0x05,0xcc,0xcc,0x15,0xcc,0xcc,0x52,0xcc,0x16,0xcc,0xcc, + 0x04,0xcc,0x41,0xcc,0xcc,0xcc,0x05,0xcc,0x15,0xcc,0x52,0xcc,0xcc,0xcc,0x16,0xcc, + 0x04,0x41,0xcc,0xcc,0xcc,0xcc,0xcc,0x05,0x15,0x52,0xcc,0xcc,0xcc,0xcc,0xcc,0x16, + 0x44,0x45,0x45,0x45,0x45,0x45,0x45,0x45,0x55,0x56,0x56,0x56,0x56,0x56,0x56,0x56, + 0x48,0x49,0xcc,0xcc,0xcc,0xcc,0xcc,0x85,0x59,0x5A,0xcc,0xcc,0xcc,0xcc,0xcc,0x96, + 0x48,0xcc,0x49,0xcc,0xcc,0xcc,0x85,0xcc,0x59,0xcc,0x5A,0xcc,0xcc,0xcc,0x96,0xcc, + 0x48,0xcc,0xcc,0x49,0xcc,0x85,0xcc,0xcc,0x59,0xcc,0xcc,0x5A,0xcc,0x96,0xcc,0xcc, + 0x48,0xcc,0xcc,0xcc,0x49,0xcc,0xcc,0xcc,0x59,0xcc,0xcc,0xcc,0x96,0xcc,0xcc,0xcc, + 0x48,0xcc,0xcc,0x85,0xcc,0x49,0xcc,0xcc,0x59,0xcc,0xcc,0x96,0xcc,0x5A,0xcc,0xcc, + 0x48,0xcc,0x85,0xcc,0xcc,0xcc,0x49,0xcc,0x59,0xcc,0x96,0xcc,0xcc,0xcc,0x5A,0xcc, + 0x48,0x85,0xcc,0xcc,0xcc,0xcc,0xcc,0x49,0x59,0x96,0xcc,0xcc,0xcc,0xcc,0xcc,0x5A, + }; + + const static uint8_t needs[16]={ + 0,1,0,0, + 2,4,2,0, + 0,1,0,0, + 15 + }; + + int x, y, b, r, l; + int16_t tmpIt [64*(32+HTAPS_MAX)]; + uint8_t tmp2t[3][stride*(32+HTAPS_MAX)]; + int16_t *tmpI= tmpIt; + uint8_t *tmp2= tmp2t[0]; + uint8_t *hpel[11]; START_TIMER - for(y=0; y < b_h+5; y++){ - for(x=0; x < b_w; x++){ - int a0= src[x ]; - int a1= src[x + 1]; - int a2= src[x + 2]; - int a3= src[x + 3]; - int a4= src[x + 4]; - int a5= src[x + 5]; -// int am= 9*(a1+a2) - (a0+a3); - int am= 20*(a2+a3) - 5*(a1+a4) + (a0+a5); -// int am= 18*(a2+a3) - 2*(a1+a4); -// int aL= (-7*a0 + 105*a1 + 35*a2 - 5*a3)>>3; -// int aR= (-7*a3 + 105*a2 + 35*a1 - 5*a0)>>3; - -// if(b_w==16) am= 8*(a1+a2); - - if(dx<8) am = (32*a2*( 8-dx) + am* dx + 128)>>8; - else am = ( am*(16-dx) + 32*a3*(dx-8) + 128)>>8; - - /* FIXME Try increasing tmp buffer to 16 bits and not clipping here. Should give marginally better results. - Robert*/ - if(am&(~255)) am= ~(am>>31); - - tmp[x] = am; - -/* if (dx< 4) tmp[x + y*stride]= (16*a1*( 4-dx) + aL* dx + 32)>>6; - else if(dx< 8) tmp[x + y*stride]= ( aL*( 8-dx) + am*(dx- 4) + 32)>>6; - else if(dx<12) tmp[x + y*stride]= ( am*(12-dx) + aR*(dx- 8) + 32)>>6; - else tmp[x + y*stride]= ( aR*(16-dx) + 16*a2*(dx-12) + 32)>>6;*/ + assert(dx<16 && dy<16); + r= brane[dx + 16*dy]&15; + l= brane[dx + 16*dy]>>4; + + b= needs[l] | needs[r]; + if(p && !p->diag_mc) + b= 15; + + if(b&5){ + for(y=0; y < b_h+HTAPS_MAX-1; y++){ + for(x=0; x < b_w; x++){ + int a_1=src[x + HTAPS_MAX/2-4]; + int a0= src[x + HTAPS_MAX/2-3]; + int a1= src[x + HTAPS_MAX/2-2]; + int a2= src[x + HTAPS_MAX/2-1]; + int a3= src[x + HTAPS_MAX/2+0]; + int a4= src[x + HTAPS_MAX/2+1]; + int a5= src[x + HTAPS_MAX/2+2]; + int a6= src[x + HTAPS_MAX/2+3]; + int am=0; + if(!p || p->fast_mc){ + am= 20*(a2+a3) - 5*(a1+a4) + (a0+a5); + tmpI[x]= am; + am= (am+16)>>5; + }else{ + am= p->hcoeff[0]*(a2+a3) + p->hcoeff[1]*(a1+a4) + p->hcoeff[2]*(a0+a5) + p->hcoeff[3]*(a_1+a6); + tmpI[x]= am; + am= (am+32)>>6; + } + + if(am&(~255)) am= ~(am>>31); + tmp2[x]= am; + } + tmpI+= 64; + tmp2+= stride; + src += stride; + } + src -= stride*y; + } + src += HTAPS_MAX/2 - 1; + tmp2= tmp2t[1]; + + if(b&2){ + for(y=0; y < b_h; y++){ + for(x=0; x < b_w+1; x++){ + int a_1=src[x + (HTAPS_MAX/2-4)*stride]; + int a0= src[x + (HTAPS_MAX/2-3)*stride]; + int a1= src[x + (HTAPS_MAX/2-2)*stride]; + int a2= src[x + (HTAPS_MAX/2-1)*stride]; + int a3= src[x + (HTAPS_MAX/2+0)*stride]; + int a4= src[x + (HTAPS_MAX/2+1)*stride]; + int a5= src[x + (HTAPS_MAX/2+2)*stride]; + int a6= src[x + (HTAPS_MAX/2+3)*stride]; + int am=0; + if(!p || p->fast_mc) + am= (20*(a2+a3) - 5*(a1+a4) + (a0+a5) + 16)>>5; + else + am= (p->hcoeff[0]*(a2+a3) + p->hcoeff[1]*(a1+a4) + p->hcoeff[2]*(a0+a5) + p->hcoeff[3]*(a_1+a6) + 32)>>6; + + if(am&(~255)) am= ~(am>>31); + tmp2[x]= am; + } + src += stride; + tmp2+= stride; + } + src -= stride*y; + } + src += stride*(HTAPS_MAX/2 - 1); + tmp2= tmp2t[2]; + tmpI= tmpIt; + if(b&4){ + for(y=0; y < b_h; y++){ + for(x=0; x < b_w; x++){ + int a_1=tmpI[x + (HTAPS_MAX/2-4)*64]; + int a0= tmpI[x + (HTAPS_MAX/2-3)*64]; + int a1= tmpI[x + (HTAPS_MAX/2-2)*64]; + int a2= tmpI[x + (HTAPS_MAX/2-1)*64]; + int a3= tmpI[x + (HTAPS_MAX/2+0)*64]; + int a4= tmpI[x + (HTAPS_MAX/2+1)*64]; + int a5= tmpI[x + (HTAPS_MAX/2+2)*64]; + int a6= tmpI[x + (HTAPS_MAX/2+3)*64]; + int am=0; + if(!p || p->fast_mc) + am= (20*(a2+a3) - 5*(a1+a4) + (a0+a5) + 512)>>10; + else + am= (p->hcoeff[0]*(a2+a3) + p->hcoeff[1]*(a1+a4) + p->hcoeff[2]*(a0+a5) + p->hcoeff[3]*(a_1+a6) + 2048)>>12; + if(am&(~255)) am= ~(am>>31); + tmp2[x]= am; + } + tmpI+= 64; + tmp2+= stride; } - tmp += stride; - src += stride; - } - tmp -= (b_h+5)*stride; - - for(y=0; y < b_h; y++){ - for(x=0; x < b_w; x++){ - int a0= tmp[x + 0*stride]; - int a1= tmp[x + 1*stride]; - int a2= tmp[x + 2*stride]; - int a3= tmp[x + 3*stride]; - int a4= tmp[x + 4*stride]; - int a5= tmp[x + 5*stride]; - int am= 20*(a2+a3) - 5*(a1+a4) + (a0+a5); -// int am= 18*(a2+a3) - 2*(a1+a4); -/* int aL= (-7*a0 + 105*a1 + 35*a2 - 5*a3)>>3; - int aR= (-7*a3 + 105*a2 + 35*a1 - 5*a0)>>3;*/ - -// if(b_w==16) am= 8*(a1+a2); - - if(dy<8) am = (32*a2*( 8-dy) + am* dy + 128)>>8; - else am = ( am*(16-dy) + 32*a3*(dy-8) + 128)>>8; - - if(am&(~255)) am= ~(am>>31); - - dst[x] = am; -/* if (dy< 4) tmp[x + y*stride]= (16*a1*( 4-dy) + aL* dy + 32)>>6; - else if(dy< 8) tmp[x + y*stride]= ( aL*( 8-dy) + am*(dy- 4) + 32)>>6; - else if(dy<12) tmp[x + y*stride]= ( am*(12-dy) + aR*(dy- 8) + 32)>>6; - else tmp[x + y*stride]= ( aR*(16-dy) + 16*a2*(dy-12) + 32)>>6;*/ + } + + hpel[ 0]= src; + hpel[ 1]= tmp2t[0] + stride*(HTAPS_MAX/2-1); + hpel[ 2]= src + 1; + + hpel[ 4]= tmp2t[1]; + hpel[ 5]= tmp2t[2]; + hpel[ 6]= tmp2t[1] + 1; + + hpel[ 8]= src + stride; + hpel[ 9]= hpel[1] + stride; + hpel[10]= hpel[8] + 1; + + if(b==15){ + uint8_t *src1= hpel[dx/8 + dy/8*4 ]; + uint8_t *src2= hpel[dx/8 + dy/8*4+1]; + uint8_t *src3= hpel[dx/8 + dy/8*4+4]; + uint8_t *src4= hpel[dx/8 + dy/8*4+5]; + dx&=7; + dy&=7; + for(y=0; y < b_h; y++){ + for(x=0; x < b_w; x++){ + dst[x]= ((8-dx)*(8-dy)*src1[x] + dx*(8-dy)*src2[x]+ + (8-dx)* dy *src3[x] + dx* dy *src4[x]+32)>>6; + } + src1+=stride; + src2+=stride; + src3+=stride; + src4+=stride; + dst +=stride; + } + }else{ + uint8_t *src1= hpel[l]; + uint8_t *src2= hpel[r]; + int a= weight[((dx&7) + (8*(dy&7)))]; + int b= 8-a; + for(y=0; y < b_h; y++){ + for(x=0; x < b_w; x++){ + dst[x]= (a*src1[x] + b*src2[x] + 4)>>3; + } + src1+=stride; + src2+=stride; + dst +=stride; } - dst += stride; - tmp += stride; } STOP_TIMER("mc_block") } #define mca(dx,dy,b_w)\ -static void mc_block_hpel ## dx ## dy ## b_w(uint8_t *dst, uint8_t *src, int stride, int h){\ - uint8_t tmp[stride*(b_w+5)];\ +static void mc_block_hpel ## dx ## dy ## b_w(uint8_t *dst, const uint8_t *src, int stride, int h){\ + uint8_t tmp[stride*(b_w+HTAPS_MAX-1)];\ assert(h==b_w);\ - mc_block(dst, src-2-2*stride, tmp, stride, b_w, b_w, dx, dy);\ + mc_block(NULL, dst, src-(HTAPS_MAX/2-1)-(HTAPS_MAX/2-1)*stride, tmp, stride, b_w, b_w, dx, dy);\ } mca( 0, 0,16) @@ -2466,7 +2352,7 @@ 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]; @@ -2506,175 +2392,56 @@ 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; const int dx= mx&15; const int dy= my&15; const int tab_index= 3 - (b_w>>2) + (b_w>>4); - sx += (mx>>4) - 2; - sy += (my>>4) - 2; + sx += (mx>>4) - (HTAPS_MAX/2-1); + sy += (my>>4) - (HTAPS_MAX/2-1); src += sx + sy*stride; - if( (unsigned)sx >= w - b_w - 4 - || (unsigned)sy >= h - b_h - 4){ - ff_emulated_edge_mc(tmp + MB_SIZE, src, stride, b_w+5, b_h+5, sx, sy, w, h); + if( (unsigned)sx >= w - b_w - (HTAPS_MAX-2) + || (unsigned)sy >= h - b_h - (HTAPS_MAX-2)){ + ff_emulated_edge_mc(tmp + MB_SIZE, src, stride, b_w+HTAPS_MAX-1, b_h+HTAPS_MAX-1, 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>1 && b_h>1); assert(tab_index>=0 && tab_index<4 || b_w==32); - 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); + if((dx&3) || (dy&3) || !(b_w == b_h || 2*b_w == b_h || b_w == 2*b_h) || (b_w&(b_w-1)) || !s->plane[plane_index].fast_mc ) + mc_block(&s->plane[plane_index], dst, src, tmp, stride, b_w, b_h, dx, dy); else if(b_w==32){ int y; for(y=0; ydsp.put_h264_qpel_pixels_tab[0][dy+(dx>>2)](dst + y*stride, src + 2 + (y+2)*stride,stride); - s->dsp.put_h264_qpel_pixels_tab[0][dy+(dx>>2)](dst + 16 + y*stride, src + 18 + (y+2)*stride,stride); + s->dsp.put_h264_qpel_pixels_tab[0][dy+(dx>>2)](dst + y*stride, src + 3 + (y+3)*stride,stride); + s->dsp.put_h264_qpel_pixels_tab[0][dy+(dx>>2)](dst + 16 + y*stride, src + 19 + (y+3)*stride,stride); } }else if(b_w==b_h) - s->dsp.put_h264_qpel_pixels_tab[tab_index ][dy+(dx>>2)](dst,src + 2 + 2*stride,stride); + s->dsp.put_h264_qpel_pixels_tab[tab_index ][dy+(dx>>2)](dst,src + 3 + 3*stride,stride); else if(b_w==2*b_h){ - s->dsp.put_h264_qpel_pixels_tab[tab_index+1][dy+(dx>>2)](dst ,src + 2 + 2*stride,stride); - s->dsp.put_h264_qpel_pixels_tab[tab_index+1][dy+(dx>>2)](dst+b_h,src + 2 + b_h + 2*stride,stride); + s->dsp.put_h264_qpel_pixels_tab[tab_index+1][dy+(dx>>2)](dst ,src + 3 + 3*stride,stride); + s->dsp.put_h264_qpel_pixels_tab[tab_index+1][dy+(dx>>2)](dst+b_h,src + 3 + b_h + 3*stride,stride); }else{ assert(2*b_w==b_h); - s->dsp.put_h264_qpel_pixels_tab[tab_index ][dy+(dx>>2)](dst ,src + 2 + 2*stride ,stride); - s->dsp.put_h264_qpel_pixels_tab[tab_index ][dy+(dx>>2)](dst+b_w*stride,src + 2 + 2*stride+b_w*stride,stride); + s->dsp.put_h264_qpel_pixels_tab[tab_index ][dy+(dx>>2)](dst ,src + 3 + 3*stride ,stride); + s->dsp.put_h264_qpel_pixels_tab[tab_index ][dy+(dx>>2)](dst+b_w*stride,src + 3 + 3*stride+b_w*stride,stride); } } } -//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); + //FIXME ugly misuse of obmc_stride + 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 - LOG2_OBMC_MAX; if(FRAC_BITS != 8){ - v += 1<<(7 - FRAC_BITS); 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, IDWTELEM *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 +2498,7 @@ 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 +2507,7 @@ 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 +2517,7 @@ 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 +2525,14 @@ 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 +2542,7 @@ 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 +2553,7 @@ 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); + //FIXME ugly misuse of obmc_stride + 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); v <<= 8 - LOG2_OBMC_MAX; if(FRAC_BITS != 8){ - v += 1<<(7 - FRAC_BITS); v >>= 8 - FRAC_BITS; } if(add){ @@ -2858,7 +2624,7 @@ assert(src_stride > 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, IDWTELEM * 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 +2634,6 @@ 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; @@ -2882,7 +2647,7 @@ static always_inline void predict_slice_buffered(SnowContext *s, slice_buffer * for(y=block_w*mb_y; yline[y]; + IDWTELEM * line = sb->line[y]; for(x=0; xline[y]; + IDWTELEM * line = sb->line[y]; for(x=0; xplane[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 +2701,6 @@ 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 +2733,7 @@ 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 +2748,7 @@ 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, IDWTELEM *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++) @@ -2999,9 +2763,8 @@ 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; + IDWTELEM *dst= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned const int b_stride = s->b_width << s->block_max_depth; const int w= p->width; const int h= p->height; @@ -3013,7 +2776,7 @@ static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){ b->type|= BLOCK_INTRA; b->color[plane_index]= 0; - memset(dst, 0, obmc_stride*obmc_stride*sizeof(DWTELEM)); + memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM)); for(i=0; i<4; i++){ int mb_x2= mb_x + (i &1) - 1; @@ -3021,7 +2784,7 @@ 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 @@ -3074,27 +2835,30 @@ 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; + IDWTELEM *pred= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; uint8_t cur[ref_stride*2*MB_SIZE]; //FIXME alignment - uint8_t tmp[ref_stride*(2*MB_SIZE+5)]; + uint8_t tmp[ref_stride*(2*MB_SIZE+HTAPS_MAX-1)]; 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; @@ -3104,33 +2868,64 @@ 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= LOG2_OBMC_MAX int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX); +#else + int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS); +#endif v = (v + pred1[x]) >> FRAC_BITS; if(v&(~255)) v= ~(v>>31); dst1[x] = v; } } - //FIXME sad/ssd can be broken up, but wavelet cmp should be one 32x32 block + /* copy the regions where obmc[] = (uint8_t)256 */ + if(LOG2_OBMC_MAX == 8 + && (mb_x == 0 || mb_x == b_stride-1) + && (mb_y == 0 || mb_y == b_height-1)){ + if(mb_x == 0) + x1 = block_w; + else + x0 = block_w; + if(mb_y == 0) + y1 = block_w; + else + y0 = block_w; + 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 +2954,10 @@ 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 IDWTELEM 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 +2970,7 @@ 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 +3010,7 @@ 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 +3026,7 @@ 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 +3049,12 @@ 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 +3065,14 @@ 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 +3098,35 @@ 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]; @@ -3390,53 +3195,77 @@ static void iterative_me(SnowContext *s){ for(i=0; i<3; i++) color[i]= get_dc(s, mb_x, mb_y, i); - // get previous score (cant be cached due to OBMC) + // get previous score (cannot be cached due to OBMC) if(pass > 0 && (block->type&BLOCK_INTRA)){ 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 +3279,7 @@ 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++; @@ -3486,16 +3316,21 @@ static void iterative_me(SnowContext *s){ } } -static void quantize(SnowContext *s, SubBand *b, DWTELEM *src, int stride, int bias){ +static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){ 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 qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT); + const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16); + const int qmul= qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS); int x,y, thres1, thres2; // START_TIMER - if(s->qlog == LOSSLESS_QLOG) return; + if(s->qlog == LOSSLESS_QLOG){ + for(y=0; y>3; thres1= ((qmul - bias)>>QEXPSHIFT) - 1; @@ -3510,15 +3345,15 @@ static void quantize(SnowContext *s, SubBand *b, DWTELEM *src, int stride, int b if(i>=0){ i<<= QEXPSHIFT; i/= qmul; //FIXME optimize - src[x + y*stride]= i; + dst[x + y*stride]= i; }else{ i= -i; i<<= QEXPSHIFT; i/= qmul; //FIXME optimize - src[x + y*stride]= -i; + dst[x + y*stride]= -i; } }else - src[x + y*stride]= 0; + dst[x + y*stride]= 0; } } }else{ @@ -3530,15 +3365,15 @@ static void quantize(SnowContext *s, SubBand *b, DWTELEM *src, int stride, int b if(i>=0){ i<<= QEXPSHIFT; i= (i + bias) / qmul; //FIXME optimize - src[x + y*stride]= i; + dst[x + y*stride]= i; }else{ i= -i; i<<= QEXPSHIFT; i= (i + bias) / qmul; //FIXME optimize - src[x + y*stride]= -i; + dst[x + y*stride]= -i; } }else - src[x + y*stride]= 0; + dst[x + y*stride]= 0; } } } @@ -3547,9 +3382,9 @@ 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){ +static void dequantize_slice_buffered(SnowContext *s, slice_buffer * sb, SubBand *b, IDWTELEM *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; @@ -3559,7 +3394,7 @@ static void dequantize_slice_buffered(SnowContext *s, slice_buffer * sb, SubBand for(y=start_y; ystride_line) + b->buf_y_offset) + b->buf_x_offset; + IDWTELEM * line = slice_buffer_get_line(sb, (y * b->stride_line) + b->buf_y_offset) + b->buf_x_offset; for(x=0; xwidth; 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; @@ -3600,7 +3435,7 @@ static void dequantize(SnowContext *s, SubBand *b, DWTELEM *src, int stride){ } } -static void decorrelate(SnowContext *s, SubBand *b, DWTELEM *src, int stride, int inverse, int use_median){ +static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){ const int w= b->width; const int h= b->height; int x,y; @@ -3624,14 +3459,14 @@ static void decorrelate(SnowContext *s, SubBand *b, DWTELEM *src, int stride, in } } -static void correlate_slice_buffered(SnowContext *s, slice_buffer * sb, SubBand *b, DWTELEM *src, int stride, int inverse, int use_median, int start_y, int end_y){ +static void correlate_slice_buffered(SnowContext *s, slice_buffer * sb, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median, int start_y, int end_y){ const int w= b->width; int x,y; // START_TIMER - DWTELEM * line; - DWTELEM * prev; + IDWTELEM * line=0; // silence silly "could be used without having been initialized" warning + IDWTELEM * prev; if (start_y != 0) line = slice_buffer_get_line(sb, ((start_y - 1) * b->stride_line) + b->buf_y_offset) + b->buf_x_offset; @@ -3658,7 +3493,7 @@ static void correlate_slice_buffered(SnowContext *s, slice_buffer * sb, SubBand // STOP_TIMER("correlate") } -static void correlate(SnowContext *s, SubBand *b, DWTELEM *src, int stride, int inverse, int use_median){ +static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){ const int w= b->width; const int h= b->height; int x,y; @@ -3682,15 +3517,40 @@ static void correlate(SnowContext *s, SubBand *b, DWTELEM *src, int stride, int } } -static void encode_header(SnowContext *s){ +static void encode_qlogs(SnowContext *s){ int plane_index, level, orientation; + + for(plane_index=0; plane_index<2; plane_index++){ + for(level=0; levelspatial_decomposition_count; level++){ + for(orientation=level ? 1:0; orientation<4; orientation++){ + if(orientation==2) continue; + put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1); + } + } + } +} + +static void encode_header(SnowContext *s){ + int plane_index, i; uint8_t kstate[32]; 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; + for(plane_index=0; plane_index<2; plane_index++){ + Plane *p= &s->plane[plane_index]; + p->last_htaps=0; + p->last_diag_mc=0; + memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff)); + } + } if(s->keyframe){ put_symbol(&s->c, s->header_state, s->version, 0); put_rac(&s->c, s->header_state, s->always_reset); @@ -3702,32 +3562,86 @@ 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); + + encode_qlogs(s); + } + if(!s->keyframe){ + int update_mc=0; for(plane_index=0; plane_index<2; plane_index++){ - for(level=0; levelspatial_decomposition_count; level++){ - for(orientation=level ? 1:0; orientation<4; orientation++){ - if(orientation==2) continue; - put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1); - } + Plane *p= &s->plane[plane_index]; + update_mc |= p->last_htaps != p->htaps; + update_mc |= p->last_diag_mc != p->diag_mc; + update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff)); + } + put_rac(&s->c, s->header_state, update_mc); + if(update_mc){ + for(plane_index=0; plane_index<2; plane_index++){ + Plane *p= &s->plane[plane_index]; + put_rac(&s->c, s->header_state, p->diag_mc); + put_symbol(&s->c, s->header_state, p->htaps/2-1, 0); + for(i= p->htaps/2; i; i--) + put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0); + + p->last_diag_mc= p->diag_mc; + p->last_htaps= p->htaps; + memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff)); + } + } + if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){ + put_rac(&s->c, s->header_state, 1); + put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0); + encode_qlogs(s); + }else + put_rac(&s->c, s->header_state, 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; + s->last_spatial_decomposition_count= s->spatial_decomposition_count; +} + +static void decode_qlogs(SnowContext *s){ + int plane_index, level, orientation; + + for(plane_index=0; plane_index<3; plane_index++){ + for(level=0; levelspatial_decomposition_count; level++){ + for(orientation=level ? 1:0; orientation<4; orientation++){ + int q; + if (plane_index==2) q= s->plane[1].band[level][orientation].qlog; + else if(orientation==2) q= s->plane[plane_index].band[level][1].qlog; + else q= get_symbol(&s->c, s->header_state, 1); + s->plane[plane_index].band[level][orientation].qlog= q; } } } - 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); } static int decode_header(SnowContext *s){ - int plane_index, level, orientation; + int plane_index; uint8_t kstate[32]; 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){ @@ -3743,31 +3657,48 @@ 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); - - for(plane_index=0; plane_index<3; plane_index++){ - for(level=0; levelspatial_decomposition_count; level++){ - for(orientation=level ? 1:0; orientation<4; orientation++){ - int q; - if (plane_index==2) q= s->plane[1].band[level][orientation].qlog; - else if(orientation==2) q= s->plane[plane_index].band[level][1].qlog; - else q= get_symbol(&s->c, s->header_state, 1); - s->plane[plane_index].band[level][orientation].qlog= q; + s->max_ref_frames= get_symbol(&s->c, s->header_state, 0)+1; + + decode_qlogs(s); + } + + if(!s->keyframe){ + if(get_rac(&s->c, s->header_state)){ + for(plane_index=0; plane_index<2; plane_index++){ + int htaps, i, sum=0, absum=0; + Plane *p= &s->plane[plane_index]; + p->diag_mc= get_rac(&s->c, s->header_state); + htaps= get_symbol(&s->c, s->header_state, 0)*2 + 2; + if((unsigned)htaps > HTAPS_MAX || htaps==0) + return -1; + p->htaps= htaps; + for(i= htaps/2; i; i--){ + p->hcoeff[i]= get_symbol(&s->c, s->header_state, 0) * (1-2*(i&1)); + sum += p->hcoeff[i]; } + p->hcoeff[0]= 32-sum; } + s->plane[2].diag_mc= s->plane[1].diag_mc; + s->plane[2].htaps = s->plane[1].htaps; + memcpy(s->plane[2].hcoeff, s->plane[1].hcoeff, sizeof(s->plane[1].hcoeff)); + } + if(get_rac(&s->c, s->header_state)){ + s->spatial_decomposition_count= get_symbol(&s->c, s->header_state, 0); + decode_qlogs(s); } } - s->spatial_decomposition_type= get_symbol(&s->c, s->header_state, 0); - if(s->spatial_decomposition_type > 2){ + s->spatial_decomposition_type+= get_symbol(&s->c, s->header_state, 1); + if(s->spatial_decomposition_type > 1){ 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; @@ -3789,7 +3720,7 @@ static void init_qexp(void){ 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; @@ -3836,21 +3767,26 @@ static int common_init(AVCodecContext *avctx){ if(!qexp[0]) init_qexp(); - dec= s->spatial_decomposition_count= 5; - s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type - - s->chroma_h_shift= 1; //FIXME XXX - s->chroma_v_shift= 1; - // dec += FFMAX(s->chroma_h_shift, s->chroma_v_shift); width= s->avctx->width; height= s->avctx->height; - s->spatial_dwt_buffer= av_mallocz(width*height*sizeof(DWTELEM)); + s->spatial_idwt_buffer= av_mallocz(width*height*sizeof(IDWTELEM)); + s->spatial_dwt_buffer= av_mallocz(width*height*sizeof(DWTELEM)); //FIXME this doesnt belong here - s->mv_scale= (s->avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4; - s->block_max_depth= (s->avctx->flags & CODEC_FLAG_4MV) ? 1 : 0; + for(i=0; iavctx->get_buffer(s->avctx, &s->mconly_picture); + + return 0; +} + +static int common_init_after_header(AVCodecContext *avctx){ + SnowContext *s = avctx->priv_data; + int plane_index, level, orientation; for(plane_index=0; plane_index<3; plane_index++){ int w= s->avctx->width; @@ -3862,6 +3798,7 @@ static int common_init(AVCodecContext *avctx){ } s->plane[plane_index].width = w; s->plane[plane_index].height= h; + //av_log(NULL, AV_LOG_DEBUG, "%d %d\n", w, h); for(level=s->spatial_decomposition_count-1; level>=0; level--){ for(orientation=level ? 1 : 0; orientation<4; orientation++){ @@ -3885,9 +3822,12 @@ static int common_init(AVCodecContext *avctx){ b->buf += b->stride>>1; b->buf_y_offset = b->stride_line >> 1; } + b->ibuf= s->spatial_idwt_buffer + (b->buf - s->spatial_dwt_buffer); if(level) b->parent= &s->plane[plane_index].band[level-1][orientation]; + //FIXME avoid this realloc + av_freep(&b->x_coeff); b->x_coeff=av_mallocz(((b->width+1) * b->height+1)*sizeof(x_and_coeff)); } w= (w+1)>>1; @@ -3895,18 +3835,65 @@ static int common_init(AVCodecContext *avctx){ } } - reset_contexts(s); -/* - width= s->width= avctx->width; - height= s->height= avctx->height; - - assert(width && height); -*/ - s->avctx->get_buffer(s->avctx, &s->mconly_picture); - 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]; + IDWTELEM *buf= b->ibuf; + 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; + //FIXME this is ugly + for(y=0; ybuf[x+y*stride]; + if(orientation==0) + decorrelate(s, b, buf, stride, 1, 0); + for(y=0; y> 16; + } + } + + /* 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; @@ -3916,15 +3903,15 @@ static void calculate_vissual_weight(SnowContext *s, Plane *p){ for(level=0; levelspatial_decomposition_count; level++){ for(orientation=level ? 1 : 0; orientation<4; orientation++){ SubBand *b= &p->band[level][orientation]; - DWTELEM *buf= b->buf; + IDWTELEM *ibuf= b->ibuf; int64_t error=0; - memset(s->spatial_dwt_buffer, 0, sizeof(int)*width*height); - buf[b->width/2 + b->height/2*b->stride]= 256*256; - ff_spatial_idwt(s->spatial_dwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count); + memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height); + ibuf[b->width/2 + b->height/2*b->stride]= 256*16; + ff_spatial_idwt(s->spatial_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count); for(y=0; yspatial_dwt_buffer[x + y*width]; + int64_t d= s->spatial_idwt_buffer[x + y*width]*16; error += d*d; } } @@ -3946,6 +3933,30 @@ 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; + } + + s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type + + s->chroma_h_shift= 1; //FIXME XXX + s->chroma_v_shift= 1; + + s->mv_scale = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4; + s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0; + + for(plane_index=0; plane_index<3; plane_index++){ + s->plane[plane_index].diag_mc= 1; + s->plane[plane_index].htaps= 6; + s->plane[plane_index].hcoeff[0]= 40; + s->plane[plane_index].hcoeff[1]= -10; + s->plane[plane_index].hcoeff[2]= 2; + s->plane[plane_index].fast_mc= 1; + } + common_init(avctx); alloc_blocks(s); @@ -3961,19 +3972,17 @@ 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; } - - for(plane_index=0; plane_index<3; plane_index++){ - calculate_vissual_weight(s, &s->plane[plane_index]); - } - + s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2)); avctx->coded_frame= &s->current_picture; switch(avctx->pix_fmt){ @@ -3985,7 +3994,7 @@ 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: @@ -4001,9 +4010,62 @@ 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; } +static void halfpel_interpol(SnowContext *s, uint8_t *halfpel[4][4], AVFrame *frame){ + int p,x,y; + + assert(!(s->avctx->flags & CODEC_FLAG_EMU_EDGE)); + + for(p=0; p<3; p++){ + int is_chroma= !!p; + int w= s->avctx->width >>is_chroma; + int h= s->avctx->height >>is_chroma; + int ls= frame->linesize[p]; + uint8_t *src= frame->data[p]; + + halfpel[1][p]= (uint8_t*)av_malloc(ls * (h+2*EDGE_WIDTH)) + EDGE_WIDTH*(1+ls); + halfpel[2][p]= (uint8_t*)av_malloc(ls * (h+2*EDGE_WIDTH)) + EDGE_WIDTH*(1+ls); + halfpel[3][p]= (uint8_t*)av_malloc(ls * (h+2*EDGE_WIDTH)) + EDGE_WIDTH*(1+ls); + + halfpel[0][p]= src; + for(y=0; y>5; + } + } + for(y=0; y>5; + } + } + src= halfpel[1][p]; + for(y=0; y>5; + } + } + +//FIXME border! + } +} + static int frame_start(SnowContext *s){ AVFrame tmp; int w= s->avctx->width; //FIXME round up to x16 ? @@ -4015,16 +4077,34 @@ 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)); + memmove(s->halfpel_plane+1, s->halfpel_plane, (s->max_ref_frames-1)*sizeof(void*)*4*4); +#ifdef USE_HALFPEL_PLANE + if(s->current_picture.data[0]) + halfpel_interpol(s, s->halfpel_plane[0], &s->current_picture); +#endif + 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 +4115,8 @@ 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 +4130,34 @@ 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 +4166,11 @@ 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 +4193,7 @@ 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 +4202,32 @@ 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: + if(pict->pict_type == I_TYPE) + s->spatial_decomposition_count= 5; + else + s->spatial_decomposition_count= 5; + + s->m.pict_type = pict->pict_type; s->qbias= pict->pict_type == P_TYPE ? 2 : 0; + common_init_after_header(avctx); + + if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){ + for(plane_index=0; plane_index<3; plane_index++){ + calculate_vissual_weight(s, &s->plane[plane_index]); + } + } + 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++){ @@ -4129,47 +4237,69 @@ 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; yspatial_dwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<spatial_dwt_buffer, plane_index, 0); + predict_plane(s, s->spatial_idwt_buffer, plane_index, 0); if( plane_index==0 && pict->pict_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; } if(s->qlog == LOSSLESS_QLOG){ for(y=0; yspatial_dwt_buffer[y*w + x]= (s->spatial_dwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS; + s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS; + } + } + }else{ + for(y=0; yspatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<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]; - quantize(s, b, b->buf, b->stride, s->qbias); + quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias); if(orientation==0) - decorrelate(s, b, b->buf, b->stride, pict->pict_type == P_TYPE, 0); - encode_subband(s, b, b->buf, b->parent ? b->parent->buf : NULL, b->stride, orientation); + decorrelate(s, b, b->ibuf, b->stride, pict->pict_type == P_TYPE, 0); + encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation); assert(b->parent==NULL || b->parent->stride == b->stride*2); if(orientation==0) - correlate(s, b, b->buf, b->stride, 1, 0); + correlate(s, b, b->ibuf, b->stride, 1, 0); } } // av_log(NULL, AV_LOG_DEBUG, "plane:%d bits:%d\n", plane_index, put_bits_count(&s->c.pb) - bits); @@ -4178,21 +4308,35 @@ redo_frame: for(orientation=level ? 1 : 0; orientation<4; orientation++){ SubBand *b= &p->band[level][orientation]; - dequantize(s, b, b->buf, b->stride); + dequantize(s, b, b->ibuf, b->stride); } } - ff_spatial_idwt(s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count); + ff_spatial_idwt(s->spatial_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count); if(s->qlog == LOSSLESS_QLOG){ for(y=0; yspatial_dwt_buffer[y*w + x]<<=FRAC_BITS; + s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS; } } } {START_TIMER - predict_plane(s, s->spatial_dwt_buffer, plane_index, 1); + predict_plane(s, s->spatial_idwt_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_idwt_buffer, 0, sizeof(IDWTELEM)*w*h); + predict_plane(s, s->spatial_idwt_buffer, plane_index, 1); + } + } if(s->avctx->flags&CODEC_FLAG_PSNR){ int64_t error= 0; @@ -4208,23 +4352,32 @@ 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]); + for(i=0; i<9; i++) + if(s->halfpel_plane[s->max_ref_frames-1][1+i/3][i%3]) + av_free(s->halfpel_plane[s->max_ref_frames-1][1+i/3][i%3] - EDGE_WIDTH*(1+s->current_picture.linesize[i%3])); + } 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(); @@ -4232,9 +4385,10 @@ 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); + av_freep(&s->spatial_idwt_buffer); av_freep(&s->m.me.scratchpad); av_freep(&s->m.me.map); @@ -4243,6 +4397,13 @@ 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++){ @@ -4267,15 +4428,11 @@ static int encode_end(AVCodecContext *avctx) static int decode_init(AVCodecContext *avctx) { SnowContext *s = avctx->priv_data; - int block_size; avctx->pix_fmt= PIX_FMT_YUV420P; common_init(avctx); - block_size = MB_SIZE >> s->block_max_depth; - slice_buffer_init(&s->sb, s->plane[0].height, (block_size) + (s->spatial_decomposition_count * (s->spatial_decomposition_count + 3)) + 1, s->plane[0].width, s->spatial_dwt_buffer); - return 0; } @@ -4284,13 +4441,27 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, uint8 RangeCoder * const c= &s->c; int bytes_read; AVFrame *picture = data; - int level, orientation, plane_index; + int level, orientation, plane_index, i; ff_init_range_decoder(c, buf, buf_size); ff_build_rac_states(c, 0.05*(1LL<<32), 256-8); s->current_picture.pict_type= FF_I_TYPE; //FIXME I vs. P - decode_header(s); + if(decode_header(s)<0) + return -1; + common_init_after_header(avctx); + + // realloc slice buffer for the case that spatial_decomposition_count changed + slice_buffer_destroy(&s->sb); + slice_buffer_init(&s->sb, s->plane[0].height, (MB_SIZE >> s->block_max_depth) + s->spatial_decomposition_count * 8 + 1, s->plane[0].width, s->spatial_idwt_buffer); + + for(plane_index=0; plane_index<3; plane_index++){ + Plane *p= &s->plane[plane_index]; + p->fast_mc= p->diag_mc && p->htaps==6 && p->hcoeff[0]==40 + && p->hcoeff[1]==-10 + && p->hcoeff[2]==2; + } + if(!s->block) alloc_blocks(s); frame_start(s); @@ -4309,7 +4480,7 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, uint8 if(s->avctx->debug&2048){ memset(s->spatial_dwt_buffer, 0, sizeof(DWTELEM)*w*h); - predict_plane(s, s->spatial_dwt_buffer, plane_index, 1); + predict_plane(s, s->spatial_idwt_buffer, plane_index, 1); for(y=0; yavctx->debug&2048){ int correlate_end_y = FFMIN(b->height, end_y + 1); int correlate_start_y = FFMIN(b->height, (start_y ? start_y + 1 : 0)); decode_subband_slice_buffered(s, correlate_band, &s->sb, correlate_start_y, correlate_end_y, decode_state[0][0]); - correlate_slice_buffered(s, &s->sb, correlate_band, correlate_band->buf, correlate_band->stride, 1, 0, correlate_start_y, correlate_end_y); - dequantize_slice_buffered(s, &s->sb, correlate_band, correlate_band->buf, correlate_band->stride, start_y, end_y); + correlate_slice_buffered(s, &s->sb, correlate_band, correlate_band->ibuf, correlate_band->stride, 1, 0, correlate_start_y, correlate_end_y); + dequantize_slice_buffered(s, &s->sb, correlate_band, correlate_band->ibuf, correlate_band->stride, start_y, end_y); } else decode_subband_slice_buffered(s, b, &s->sb, start_y, end_y, decode_state[level][orientation]); @@ -4387,21 +4558,21 @@ 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");} if(s->qlog == LOSSLESS_QLOG){ for(; yqsb, yq); + IDWTELEM * line = slice_buffer_get_line(&s->sb, yq); for(x=0; xsb, s->spatial_dwt_buffer, plane_index, 1, mb_y); + predict_slice_buffered(s, &s->sb, s->spatial_idwt_buffer, plane_index, 1, mb_y); y = FFMIN(p->height, slice_starty); end_y = FFMIN(p->height, slice_h); @@ -4416,8 +4587,12 @@ 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]); + for(i=0; i<9; i++) + if(s->halfpel_plane[s->max_ref_frames-1][1+i/3][i%3]) + av_free(s->halfpel_plane[s->max_ref_frames-1][1+i/3][i%3] - EDGE_WIDTH*(1+s->current_picture.linesize[i%3])); + } if(!(s->avctx->debug&2048)) *picture= s->current_picture; @@ -4456,7 +4631,7 @@ AVCodec snow_decoder = { NULL }; -#ifdef CONFIG_ENCODERS +#ifdef CONFIG_SNOW_ENCODER AVCodec snow_encoder = { "snow", CODEC_TYPE_VIDEO, @@ -4473,6 +4648,7 @@ AVCodec snow_encoder = { #undef malloc #undef free #undef printf +#undef random int main(){ int width=256; @@ -4502,7 +4678,7 @@ 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 +4688,7 @@ 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 +4702,7 @@ 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 +4731,9 @@ int64_t g=0; for(x=0; x