X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libavcodec%2Fsnow.c;h=0aa56dba61f4799674c2f3dec635b0aa5bbbfd73;hb=2692ceab8bcc0e7787ce707e9040b4d3ca79723e;hp=18be35218e22bfbd0660f110449a9c986436899c;hpb=fff6d4ea4e66f04eb2fb4172ca5ae16c8ddb3a73;p=ffmpeg diff --git a/libavcodec/snow.c b/libavcodec/snow.c index 18be35218e2..0aa56dba61f 100644 --- a/libavcodec/snow.c +++ b/libavcodec/snow.c @@ -19,7 +19,9 @@ #include "avcodec.h" #include "common.h" #include "dsputil.h" -#include "cabac.h" + +#include "rangecoder.h" +#define MID_STATE 128 #include "mpegvideo.h" @@ -29,8 +31,10 @@ #define MAX_DECOMPOSITIONS 8 #define MAX_PLANES 4 #define DWTELEM int -#define QROOT 8 +#define QSHIFT 5 +#define QROOT (1<line[line_num] ? (slice_buf)->line[line_num] : slice_buffer_load_line((slice_buf), (line_num))) +//#define slice_buffer_get_line(slice_buf, line_num) (slice_buffer_load_line((slice_buf), (line_num))) + +static void iterative_me(SnowContext *s); + +static void slice_buffer_init(slice_buffer * buf, int line_count, int max_allocated_lines, int line_width, DWTELEM * base_buffer) +{ + int i; + + buf->base_buffer = base_buffer; + buf->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); + + for (i = 0; i < max_allocated_lines; i++) + { + buf->data_stack[i] = (DWTELEM *) av_malloc (sizeof(DWTELEM) * line_width); + } + + buf->data_stack_top = max_allocated_lines - 1; +} + +static DWTELEM * slice_buffer_load_line(slice_buffer * buf, int line) +{ + int offset; + DWTELEM * buffer; + +// av_log(NULL, AV_LOG_DEBUG, "Cache hit: %d\n", line); + + assert(buf->data_stack_top >= 0); +// assert(!buf->line[line]); + if (buf->line[line]) + return buf->line[line]; + + offset = buf->line_width * line; + buffer = buf->data_stack[buf->data_stack_top]; + buf->data_stack_top--; + buf->line[line] = buffer; + +// av_log(NULL, AV_LOG_DEBUG, "slice_buffer_load_line: line: %d remaining: %d\n", line, buf->data_stack_top + 1); + + return buffer; +} + +static void slice_buffer_release(slice_buffer * buf, int line) +{ + int offset; + DWTELEM * buffer; + + assert(line >= 0 && line < buf->line_count); + assert(buf->line[line]); + + offset = buf->line_width * line; + buffer = buf->line[line]; + buf->data_stack_top++; + buf->data_stack[buf->data_stack_top] = buffer; + buf->line[line] = NULL; + +// av_log(NULL, AV_LOG_DEBUG, "slice_buffer_release: line: %d remaining: %d\n", line, buf->data_stack_top + 1); +} + +static void slice_buffer_flush(slice_buffer * buf) +{ + int i; + for (i = 0; i < buf->line_count; i++) + { + if (buf->line[i]) + { +// av_log(NULL, AV_LOG_DEBUG, "slice_buffer_flush: line: %d \n", i); + slice_buffer_release(buf, i); + } + } +} + +static void slice_buffer_destroy(slice_buffer * buf) +{ + int i; + slice_buffer_flush(buf); + + for (i = buf->data_count - 1; i >= 0; i--) + { + assert(buf->data_stack[i]); + av_free(buf->data_stack[i]); + } + assert(buf->data_stack); + av_free(buf->data_stack); + assert(buf->line); + av_free(buf->line); +} + +#ifdef __sgi +// Avoid a name clash on SGI IRIX +#undef qexp +#endif +#define QEXPSHIFT (7-FRAC_BITS+8) //FIXME try to change this to 0 +static uint8_t qexp[QROOT]; static inline int mirror(int v, int m){ - if (v<0) return -v; - else if(v>m) return 2*m-v; - else return v; + while((unsigned)v > (unsigned)m){ + v=-v; + if(v<0) v+= 2*m; + } + return v; } -static inline void put_symbol(CABACContext *c, uint8_t *state, int v, int is_signed){ +static inline void put_symbol(RangeCoder *c, uint8_t *state, int v, int is_signed){ int i; if(v){ const int a= ABS(v); const int e= av_log2(a); #if 1 - const int el= FFMIN(e, 10); - put_cabac(c, state+0, 0); + const int el= FFMIN(e, 10); + put_rac(c, state+0, 0); for(i=0; i=el; i--){ - put_cabac(c, state+22+9, (a>>i)&1); //22..31 + put_rac(c, state+22+9, (a>>i)&1); //22..31 } for(; i>=0; i--){ - put_cabac(c, state+22+i, (a>>i)&1); //22..31 + put_rac(c, state+22+i, (a>>i)&1); //22..31 } if(is_signed) - put_cabac(c, state+11 + el, v < 0); //11..21 + put_rac(c, state+11 + el, v < 0); //11..21 #else - - put_cabac(c, state+0, 0); + + put_rac(c, state+0, 0); if(e<=9){ for(i=0; i=0; i--){ - put_cabac(c, state+22+i, (a>>i)&1); //22..31 + put_rac(c, state+22+i, (a>>i)&1); //22..31 } if(is_signed) - put_cabac(c, state+11 + e, v < 0); //11..21 + put_rac(c, state+11 + e, v < 0); //11..21 }else{ for(i=0; i=0; i--){ - put_cabac(c, state+22+FFMIN(i,9), (a>>i)&1); //22..31 + put_rac(c, state+22+FFMIN(i,9), (a>>i)&1); //22..31 } if(is_signed) - put_cabac(c, state+11 + FFMIN(e,10), v < 0); //11..21 + put_rac(c, state+11 + FFMIN(e,10), v < 0); //11..21 } #endif }else{ - put_cabac(c, state+0, 1); + put_rac(c, state+0, 1); } } -static inline int get_symbol(CABACContext *c, uint8_t *state, int is_signed){ - if(get_cabac(c, state+0)) +static inline int get_symbol(RangeCoder *c, uint8_t *state, int is_signed){ + if(get_rac(c, state+0)) return 0; else{ - int i, e, a, el; - //FIXME try to merge loops with FFMIN() maybe they are equally fast and they are surly cuter - for(e=0; e<10; e++){ - if(get_cabac(c, state + 1 + e)==0) // 1..10 - break; - } - el= e; - - if(e==10){ - while(get_cabac(c, state + 1 + 9)) //10 - e++; + int i, e, a; + e= 0; + while(get_rac(c, state+1 + FFMIN(e,9))){ //1..10 + e++; } + a= 1; - for(i=e-1; i>=el; i--){ - a += a + get_cabac(c, state+22+9); //31 - } - for(; i>=0; i--){ - a += a + get_cabac(c, state+22+i); //22..31 + for(i=e-1; i>=0; i--){ + a += a + get_rac(c, state+22 + FFMIN(i,9)); //22..31 } - if(is_signed && get_cabac(c, state+11 + el)) //11..21 + if(is_signed && get_rac(c, state+11 + FFMIN(e,10))) //11..21 return -a; else return a; } } -static inline void put_symbol2(CABACContext *c, uint8_t *state, int v, int log2){ +static inline void put_symbol2(RangeCoder *c, uint8_t *state, int v, int log2){ int i; int r= log2>=0 ? 1<=-4); while(v >= r){ - put_cabac(c, state+4+log2, 1); + put_rac(c, state+4+log2, 1); v -= r; log2++; if(log2>0) r+=r; } - put_cabac(c, state+4+log2, 0); - + put_rac(c, state+4+log2, 0); + for(i=log2-1; i>=0; i--){ - put_cabac(c, state+31-i, (v>>i)&1); + put_rac(c, state+31-i, (v>>i)&1); } } -static inline int get_symbol2(CABACContext *c, uint8_t *state, int log2){ +static inline int get_symbol2(RangeCoder *c, uint8_t *state, int log2){ int i; int r= log2>=0 ? 1<=-4); - while(get_cabac(c, state+4+log2)){ + while(get_rac(c, state+4+log2)){ v+= r; log2++; if(log2>0) r+=r; } - + for(i=log2-1; i>=0; i--){ - v+= get_cabac(c, state+31-i)<>shift), inverse); } - + if(mirror_right){ dst[w*dst_step] = LIFT(src[w*src_step], ((mul*2*ref[w*ref_step]+add)>>shift), inverse); } @@ -609,14 +755,14 @@ static always_inline void lift5(DWTELEM *dst, DWTELEM *src, DWTELEM *ref, int ds 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); } - + if(mirror_right){ int r= 3*2*ref[w*ref_step]; r += r>>4; @@ -625,10 +771,33 @@ static always_inline void lift5(DWTELEM *dst, DWTELEM *src, DWTELEM *ref, int ds } } +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){ + 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)) + if(mirror_left){ + dst[0] = LIFTS(src[0], mul*2*ref[0]+add, inverse); + dst += dst_step; + src += src_step; + } + + for(i=0; i>1; const int w2= (width+1)>>1; - int A1,A2,A3,A4, x; + 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 A1,A2,A3,A4, x; + int x; const int w2= (width+1)>>1; memcpy(temp, b, width*sizeof(int)); @@ -815,9 +984,9 @@ static void horizontal_composeX(int *b, int width){ inplace_lift(b, width, COEFFS1, N1, SHIFT1, LX1, 1); } -static void spatial_decomposeX(int *buffer, int width, int height, int stride){ +static void spatial_decomposeX(DWTELEM *buffer, int width, int height, int stride){ int x, y; - + for(y=0; y>1; - int A1,A2,A3,A4, x; + int x; const int w2= (width+1)>>1; for(x=0; x>2; b[width -1] = A3; b[width2-1] = A2; -#else + } +#else lift(b+w2, temp+w2, temp, 1, 1, 1, width, -1, 0, 1, 1, 0); lift(b , temp , b+w2, 1, 1, 1, width, 1, 2, 2, 0, 0); #endif } -static void vertical_decompose53iH0(int *b0, int *b1, int *b2, int width){ +static void vertical_decompose53iH0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){ int i; - + for(i=0; i>1; } } -static void vertical_decompose53iL0(int *b0, int *b1, int *b2, int width){ +static void vertical_decompose53iL0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){ int i; - + for(i=0; i>2; } } -static void spatial_decompose53i(int *buffer, int width, int height, int stride){ - int x, y; +static void spatial_decompose53i(DWTELEM *buffer, int width, int height, int stride){ + int y; DWTELEM *b0= buffer + mirror(-2-1, height-1)*stride; DWTELEM *b1= buffer + mirror(-2 , height-1)*stride; - + for(y=-2; y>1; lift (temp+w2, b +1, b , 1, 2, 2, width, -W_AM, W_AO, W_AS, 1, 0); - lift (temp , b , temp+w2, 1, 2, 1, width, -W_BM, W_BO, W_BS, 0, 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 (b , temp , b +w2, 1, 1, 1, width, W_DM, W_DO, W_DS, 0, 0); } -static void vertical_decompose97iH0(int *b0, int *b1, int *b2, int width){ +static void vertical_decompose97iH0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){ int i; - + for(i=0; i>W_AS; } } -static void vertical_decompose97iH1(int *b0, int *b1, int *b2, int width){ +static void vertical_decompose97iH1(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){ int i; - + for(i=0; i>W_CS; @@ -1042,50 +1215,54 @@ static void vertical_decompose97iH1(int *b0, int *b1, int *b2, int width){ } } -static void vertical_decompose97iL0(int *b0, int *b1, int *b2, int width){ +static void vertical_decompose97iL0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){ int i; - + for(i=0; i>W_BS; +#else + b1[i] = (16*4*b1[i] - 4*(b0[i] + b2[i]) + 8*5 + (5<<27)) / (5*16) - (1<<23); +#endif } } -static void vertical_decompose97iL1(int *b0, int *b1, int *b2, int width){ +static void vertical_decompose97iL1(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){ int i; - + for(i=0; i>W_DS; } } -static void spatial_decompose97i(int *buffer, int width, int height, int stride){ - int x, y; +static void spatial_decompose97i(DWTELEM *buffer, int width, int height, int stride){ + int y; DWTELEM *b0= buffer + mirror(-4-1, height-1)*stride; DWTELEM *b1= buffer + mirror(-4 , height-1)*stride; DWTELEM *b2= buffer + mirror(-4+1, height-1)*stride; DWTELEM *b3= buffer + mirror(-4+2, height-1)*stride; - + for(y=-4; y400){ STOP_TIMER("horizontal_decompose97i") }} - + {START_TIMER - if(b3 <= b5) vertical_decompose97iH0(b3, b4, b5, width); - if(b2 <= b4) vertical_decompose97iL0(b2, b3, b4, width); - if(b1 <= b3) vertical_decompose97iH1(b1, b2, b3, width); - if(b0 <= b2) vertical_decompose97iL1(b0, b1, b2, width); + if(y+3<(unsigned)height) vertical_decompose97iH0(b3, b4, b5, width); + if(y+2<(unsigned)height) vertical_decompose97iL0(b2, b3, b4, width); + if(y+1<(unsigned)height) vertical_decompose97iH1(b1, b2, b3, width); + if(y+0<(unsigned)height) vertical_decompose97iL1(b0, b1, b2, width); if(width>400){ STOP_TIMER("vertical_decompose97i") }} - + b0=b2; b1=b3; b2=b4; @@ -1093,9 +1270,9 @@ STOP_TIMER("vertical_decompose97i") } } -void ff_spatial_dwt(int *buffer, int width, int height, int stride, int type, int decomposition_count){ +void ff_spatial_dwt(DWTELEM *buffer, int width, int height, int stride, int type, int decomposition_count){ int level; - + for(level=0; level>level, height>>level, stride<>1; const int w2= (width+1)>>1; - int A1,A2,A3,A4, x; + int x; #if 0 + int A1,A2,A3,A4; A2= temp[1 ]; A4= temp[0 ]; A1= temp[0+width2]; @@ -1139,7 +1317,7 @@ static void horizontal_compose53i(int *b, int width){ A2 += (A1 + A3 + 2)>>2; b[width -1] = A3; b[width2-1] = A2; -#else +#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); #endif @@ -1151,68 +1329,108 @@ static void horizontal_compose53i(int *b, int width){ b[2*x ]= temp[x ]; } -static void vertical_compose53iH0(int *b0, int *b1, int *b2, int width){ +static void vertical_compose53iH0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){ int i; - + for(i=0; i>1; } } -static void vertical_compose53iL0(int *b0, int *b1, int *b2, int width){ +static void vertical_compose53iL0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){ int i; - + for(i=0; i>2; } } -static void spatial_compose53i(int *buffer, int width, int height, int stride){ - int x, y; - DWTELEM *b0= buffer + mirror(-1-1, height-1)*stride; - DWTELEM *b1= buffer + mirror(-1 , height-1)*stride; - - for(y=-1; y<=height; y+=2){ - DWTELEM *b2= buffer + mirror(y+1, height-1)*stride; - DWTELEM *b3= buffer + mirror(y+2, height-1)*stride; +static void spatial_compose53i_buffered_init(dwt_compose_t *cs, slice_buffer * sb, int height, int stride_line){ + cs->b0 = slice_buffer_get_line(sb, mirror(-1-1, height-1) * stride_line); + cs->b1 = slice_buffer_get_line(sb, mirror(-1 , height-1) * stride_line); + cs->y = -1; +} + +static void spatial_compose53i_init(dwt_compose_t *cs, DWTELEM *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; +} + +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); {START_TIMER - if(b1 <= b3) vertical_compose53iL0(b1, b2, b3, width); - if(b0 <= b2) vertical_compose53iH0(b0, b1, b2, width); + if(y+1<(unsigned)height) vertical_compose53iL0(b1, b2, b3, width); + if(y+0<(unsigned)height) vertical_compose53iH0(b0, b1, b2, width); STOP_TIMER("vertical_compose53i*")} {START_TIMER - if(y-1 >= 0) horizontal_compose53i(b0, width); - if(b0 <= b2) horizontal_compose53i(b1, width); + if(y-1<(unsigned)height) horizontal_compose53i(b0, width); + if(y+0<(unsigned)height) horizontal_compose53i(b1, width); STOP_TIMER("horizontal_compose53i")} - b0=b2; - b1=b3; - } -} + cs->b0 = b2; + cs->b1 = b3; + cs->y += 2; +} + +static void spatial_compose53i_dy(dwt_compose_t *cs, DWTELEM *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; + +{START_TIMER + if(y+1<(unsigned)height) vertical_compose53iL0(b1, b2, b3, width); + if(y+0<(unsigned)height) vertical_compose53iH0(b0, b1, b2, width); +STOP_TIMER("vertical_compose53i*")} + +{START_TIMER + if(y-1<(unsigned)height) horizontal_compose53i(b0, width); + if(y+0<(unsigned)height) horizontal_compose53i(b1, width); +STOP_TIMER("horizontal_compose53i")} + + cs->b0 = b2; + cs->b1 = b3; + cs->y += 2; +} + +static void spatial_compose53i(DWTELEM *buffer, int width, int height, int stride){ + dwt_compose_t cs; + spatial_compose53i_init(&cs, buffer, height, stride); + while(cs.y <= height) + spatial_compose53i_dy(&cs, buffer, width, height, stride); +} + - -static void horizontal_compose97i(int *b, int width){ - int temp[width]; +static void horizontal_compose97i(DWTELEM *b, int width){ + DWTELEM 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); - lift (b , temp , temp+w2, 2, 1, 1, width, -W_BM, W_BO, W_BS, 0, 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); } -static void vertical_compose97iH0(int *b0, int *b1, int *b2, int width){ +static void vertical_compose97iH0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){ int i; - + for(i=0; i>W_AS; } } -static void vertical_compose97iH1(int *b0, int *b1, int *b2, int width){ +static void vertical_compose97iH1(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){ int i; - + for(i=0; i>W_CS; @@ -1225,618 +1443,277 @@ static void vertical_compose97iH1(int *b0, int *b1, int *b2, int width){ } } -static void vertical_compose97iL0(int *b0, int *b1, int *b2, int width){ +static void vertical_compose97iL0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){ int i; - + for(i=0; i>W_BS; +#else + b1[i] += (W_BM*(b0[i] + b2[i])+4*b1[i]+W_BO)>>W_BS; +#endif } } -static void vertical_compose97iL1(int *b0, int *b1, int *b2, int width){ +static void vertical_compose97iL1(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, int width){ int i; - + for(i=0; i>W_DS; } } -static void spatial_compose97i(int *buffer, int width, int height, int stride){ - int x, y; - DWTELEM *b0= buffer + mirror(-3-1, height-1)*stride; - DWTELEM *b1= buffer + mirror(-3 , height-1)*stride; - DWTELEM *b2= buffer + mirror(-3+1, height-1)*stride; - DWTELEM *b3= buffer + mirror(-3+2, height-1)*stride; +static void vertical_compose97i(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width){ + int i; - for(y=-3; y<=height; y+=2){ - DWTELEM *b4= buffer + mirror(y+3, height-1)*stride; - DWTELEM *b5= buffer + mirror(y+4, height-1)*stride; + 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 + b2[i] += (W_BM*(b1[i] + b3[i])+4*b2[i]+W_BO)>>W_BS; +#endif + b1[i] += (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS; + } +} + +static void spatial_compose97i_buffered_init(dwt_compose_t *cs, slice_buffer * sb, int height, int stride_line){ + cs->b0 = slice_buffer_get_line(sb, mirror(-3-1, height-1) * stride_line); + cs->b1 = slice_buffer_get_line(sb, mirror(-3 , height-1) * stride_line); + cs->b2 = slice_buffer_get_line(sb, mirror(-3+1, height-1) * stride_line); + cs->b3 = slice_buffer_get_line(sb, mirror(-3+2, height-1) * stride_line); + cs->y = -3; +} + +static void spatial_compose97i_init(dwt_compose_t *cs, DWTELEM *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; + cs->b3 = buffer + mirror(-3+2, height-1)*stride; + cs->y = -3; +} + +static void spatial_compose97i_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= 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); + +{START_TIMER + if(y>0 && y+4400){ +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(width>400 && y+0<(unsigned)height){ +STOP_TIMER("horizontal_compose97i")}} + + cs->b0=b2; + cs->b1=b3; + cs->b2=b4; + cs->b3=b5; + cs->y += 2; +} + +static void spatial_compose97i_dy(dwt_compose_t *cs, DWTELEM *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; - if(stride == width && y+4 < height && 0){ - int x; - for(x=0; x400){ STOP_TIMER("vertical_compose97i")}} {START_TIMER - if(y-1>= 0) horizontal_compose97i(b0, width); - if(b0 <= b2) horizontal_compose97i(b1, width); + if(y-1<(unsigned)height) horizontal_compose97i(b0, width); + if(y+0<(unsigned)height) horizontal_compose97i(b1, width); if(width>400 && b0 <= b2){ STOP_TIMER("horizontal_compose97i")}} - - b0=b2; - b1=b3; - b2=b4; - b3=b5; - } + + cs->b0=b2; + cs->b1=b3; + cs->b2=b4; + cs->b3=b5; + cs->y += 2; } -void ff_spatial_idwt(int *buffer, int width, int height, int stride, int type, int decomposition_count){ - int level; +static void spatial_compose97i(DWTELEM *buffer, int width, int height, int stride){ + dwt_compose_t cs; + spatial_compose97i_init(&cs, buffer, height, stride); + while(cs.y <= height) + spatial_compose97i_dy(&cs, buffer, width, height, stride); +} +void ff_spatial_idwt_buffered_init(dwt_compose_t *cs, slice_buffer * sb, int width, int height, int stride_line, int type, int decomposition_count){ + int level; for(level=decomposition_count-1; level>=0; level--){ switch(type){ - case 0: spatial_compose97i(buffer, width>>level, height>>level, stride<>level, height>>level, stride<>level, height>>level, stride<>level, stride_line<>level, stride_line<>level, height>>level, stride<=0; level--){ + switch(type){ + case 0: spatial_compose97i_init(cs+level, buffer, height>>level, stride<>level, stride<>level, height>>level, stride<=0; level--){ + while(cs[level].y <= FFMIN((y>>level)+support, height>>level)){ + switch(type){ + case 0: spatial_compose97i_dy(cs+level, buffer, width>>level, height>>level, stride<>level, height>>level, stride<=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<>15; +void ff_spatial_idwt(DWTELEM *buffer, int width, int height, int stride, int type, int decomposition_count){ + if(type==2){ + int level; + for(level=decomposition_count-1; level>=0; level--) + spatial_composeX (buffer, width>>level, height>>level, stride<level; +static int encode_subband_c0run(SnowContext *s, SubBand *b, DWTELEM *src, DWTELEM *parent, int stride, int orientation){ const int w= b->width; const int h= b->height; - int x, y, pos; + int x, y; if(1){ int run=0; int runs[w*h]; int run_index=0; - int count=0; - - for(pos=0; ; pos++){ - int x= deint(pos ); - int y= deint(pos>>1); - int v, p=0, pr=0, pd=0; - int /*ll=0, */l=0, lt=0, t=0/*, rt=0*/; - - if(x>=w || y>=h){ - if(x>=w && y>=h) - break; - continue; - } - count++; - - v= src[x + y*stride]; + int max_index; - if(y){ - t= src[x + (y-1)*stride]; - if(x){ - lt= src[x - 1 + (y-1)*stride]; + for(y=0; y 1){ + if(orientation==1) ll= src[y + (x-2)*stride]; + else ll= src[x - 2 + y*stride]; + }*/ } - } - if(x){ - l= src[x - 1 + y*stride]; - /*if(x > 1){ - if(orientation==1) ll= src[y + (x-2)*stride]; - else ll= src[x - 2 + y*stride]; - }*/ - } - if(parent){ - int px= x>>1; - int py= y>>1; - if(pxparent->width && pyparent->height){ - p= parent[px + py*2*stride]; - /*if(px+1parent->width) - pr= parent[px + 1 + py*2*stride]; - if(py+1parent->height) - pd= parent[px + (py+1)*2*stride];*/ + if(parent){ + int px= x>>1; + int py= y>>1; + if(pxparent->width && pyparent->height) + p= parent[px + py*2*stride]; } - } - if(!(/*ll|*/l|lt|t|/*rt|*/p)){ - if(v){ - runs[run_index++]= run; - run=0; - }else{ - run++; + if(!(/*ll|*/l|lt|t|rt|p)){ + if(v){ + runs[run_index++]= run; + run=0; + }else{ + run++; + } } } } - assert(count==w*h); + max_index= run_index; runs[run_index++]= run; run_index=0; run= runs[run_index++]; - put_symbol(&s->c, b->state[1], run, 0); - - for(pos=0; ; pos++){ - int x= deint(pos ); - int y= deint(pos>>1); - int v, p=0, pr=0, pd=0; - int /*ll=0, */l=0, lt=0, t=0/*, rt=0*/; + put_symbol2(&s->c, b->state[30], max_index, 0); + if(run_index <= max_index) + put_symbol2(&s->c, b->state[1], run, 3); - if(x>=w || y>=h){ - if(x>=w && y>=h) - break; - continue; + for(y=0; yc.bytestream_end - s->c.bytestream < w*40){ + av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n"); + return -1; } - v= src[x + y*stride]; - - if(y){ - t= src[x + (y-1)*stride]; - if(x){ - lt= src[x - 1 + (y-1)*stride]; - } - if(x + 1 < w){ -// rt= src[x + 1 + (y-1)*stride]; - } - } - if(x){ - l= src[x - 1 + y*stride]; - /*if(x > 1){ - if(orientation==1) ll= src[y + (x-2)*stride]; - else ll= src[x - 2 + y*stride]; - }*/ - } - - if(parent){ - int px= x>>1; - int py= y>>1; - if(pxparent->width && pyparent->height){ - p= parent[px + py*2*stride]; -/* if(px+1parent->width) - pr= parent[px + 1 + py*2*stride]; - if(py+1parent->height) - pd= parent[px + (py+1)*2*stride];*/ - } - } - if(/*ll|*/l|lt|t|/*rt|*/p){ - int context= av_log2(/*ABS(ll) + */2*(3*ABS(l) + ABS(lt) + 2*ABS(t) + /*ABS(rt) +*/ ABS(p))); - - put_cabac(&s->c, &b->state[0][context], !!v); - }else{ - if(!run){ - run= runs[run_index++]; - put_symbol(&s->c, b->state[1], run, 0); - assert(v); - }else{ - run--; - assert(!v); - } - } - if(v){ - int context= av_log2(/*ABS(ll) + */3*ABS(l) + ABS(lt) + 2*ABS(t) + /*ABS(rt) +*/ ABS(p)); - - put_symbol(&s->c, b->state[context + 2], ABS(v)-1, 0); - put_cabac(&s->c, &b->state[0][16 + 1 + 3 + quant3b[l&0xFF] + 3*quant3b[t&0xFF]], v<0); - } - } - } -} - -static void encode_subband_bp(SnowContext *s, SubBand *b, DWTELEM *src, DWTELEM *parent, int stride, int orientation){ - const int level= b->level; - const int w= b->width; - const int h= b->height; - int x, y; - -#if 0 - int plane; - for(plane=24; plane>=0; plane--){ - int run=0; - int runs[w*h]; - int run_index=0; - - for(y=0; y 1){ - if(orientation==1) ll= src[y + (x-2)*stride]; - else ll= src[x - 2 + y*stride]; - }*/ - } - if(y+1>1; - int py= y>>1; - if(pxparent->width && pyparent->height) - p= parent[px + py*2*stride]; - } -#define HIDE(c, plane) c= c>=0 ? c&((-1)<<(plane)) : -((-c)&((-1)<<(plane))); - lv=v; - HIDE( v, plane) - HIDE(lv, plane+1) - HIDE( p, plane) - HIDE( l, plane) - HIDE(lt, plane) - HIDE( t, plane) - HIDE(rt, plane) - HIDE( r, plane+1) - HIDE(rd, plane+1) - HIDE( d, plane+1) - HIDE(ld, plane+1) - if(!(/*ll|*/l|lt|t|rt|r|rd|ld|d|p|lv)){ - if(v){ - runs[run_index++]= run; - run=0; - }else{ - run++; - } - } - } - } - runs[run_index++]= run; - run_index=0; - run= runs[run_index++]; - - put_symbol(&s->c, b->state[1], run, 0); - - for(y=0; y 1){ - if(orientation==1) ll= src[y + (x-2)*stride]; - else ll= src[x - 2 + y*stride]; - }*/ - } - if(y+1>1; - int py= y>>1; - if(pxparent->width && pyparent->height) - p= parent[px + py*2*stride]; - } - lv=v; - HIDE( v, plane) - HIDE(lv, plane+1) - HIDE( p, plane) - HIDE( l, plane) - HIDE(lt, plane) - HIDE( t, plane) - HIDE(rt, plane) - HIDE( r, plane+1) - HIDE(rd, plane+1) - HIDE( d, plane+1) - HIDE(ld, plane+1) - if(/*ll|*/l|lt|t|rt|r|rd|ld|d|p|lv){ - int context= av_log2(/*ABS(ll) + */3*ABS(l) + ABS(lt) + 2*ABS(t) + ABS(rt) + ABS(p) - +3*ABS(r) + ABS(rd) + 2*ABS(d) + ABS(ld)); - - if(lv) put_cabac(&s->c, &b->state[99][context + 8*(av_log2(ABS(lv))-plane)], !!(v-lv)); - else put_cabac(&s->c, &b->state[ 0][context], !!v); - }else{ - assert(!lv); - if(!run){ - run= runs[run_index++]; - put_symbol(&s->c, b->state[1], run, 0); - assert(v); - }else{ - run--; - assert(!v); - } - } - if(v && !lv){ - int context= clip(quant3b[l&0xFF] + quant3b[r&0xFF], -1,1) - + 3*clip(quant3b[t&0xFF] + quant3b[d&0xFF], -1,1); - put_cabac(&s->c, &b->state[0][16 + 1 + 3 + context], v<0); - } - } - } - } - return; -#endif -} - -static void encode_subband_X(SnowContext *s, SubBand *b, DWTELEM *src, DWTELEM *parent, int stride, int orientation){ - const int level= b->level; - const int w= b->width; - const int h= b->height; - int x, y; - -#if 0 - if(orientation==3 && parent && 0){ - int16_t candidate[w*h][2]; - uint8_t state[w*h]; - int16_t boarder[3][w*h*4][2]; - int16_t significant[w*h][2]; - int candidate_count=0; - int boarder_count[3]={0,0,0}; - int significant_count=0; - int rle_pos=0; - int v, last_v; - int primary= orientation==1; - - memset(candidate, 0, sizeof(candidate)); - memset(state, 0, sizeof(state)); - memset(boarder, 0, sizeof(boarder)); - - for(y=0; y>1) + (y>>1)*2*stride]) - SVI(candidate, candidate_count, x, y) - } - } - - for(;;){ - while(candidate_count && !boarder_count[0] && !boarder_count[1] && !boarder_count[2]){ - candidate_count--; - x= candidate[ candidate_count][0]; - y= candidate[ candidate_count][1]; - if(state[x + y*w]) - continue; - state[x + y*w]= 1; - v= !!src[x + y*stride]; - put_cabac(&s->c, &b->state[0][0], v); - if(v){ - SVI(significant, significant_count, x,y) - if(x && !state[x - 1 + y *w]) SVI(boarder[0],boarder_count[0],x-1,y ) - if(y && !state[x + (y-1)*w]) SVI(boarder[1],boarder_count[1],x ,y-1) - if(x+1c, b->state[1], run, 0); - SVI(significant, significant_count, x,y) - if(x && !state[x - 1 + y *w]) SVI(boarder[0],boarder_count[0],x-1,y ) - if(y && !state[x + (y-1)*w]) SVI(boarder[1],boarder_count[1],x ,y-1) - if(x+1c, &b->state[0][index+1], v); - if(v){ - SVI(significant, significant_count, x,y) - if(x && !state[x - 1 + y *w]) SVI(boarder[0],boarder_count[0],x-1,y ) - if(y && !state[x + (y-1)*w]) SVI(boarder[1],boarder_count[1],x ,y-1) - if(x+1c, b->state[context + 2], v, 1); //FIXME try to avoid first bit, try this with the old code too!! - last_v= v; - } - } -#endif -} - -static void encode_subband_c0run(SnowContext *s, SubBand *b, DWTELEM *src, DWTELEM *parent, int stride, int orientation){ - const int level= b->level; - const int w= b->width; - const int h= b->height; - int x, y; - - if(1){ - int run=0; - int runs[w*h]; - int run_index=0; - - for(y=0; y 1){ - if(orientation==1) ll= src[y + (x-2)*stride]; - else ll= src[x - 2 + y*stride]; - }*/ - } - if(parent){ - int px= x>>1; - int py= y>>1; - if(pxparent->width && pyparent->height) - p= parent[px + py*2*stride]; - } - if(!(/*ll|*/l|lt|t|rt|p)){ - if(v){ - runs[run_index++]= run; - run=0; - }else{ - run++; - } - } - } - } - runs[run_index++]= run; - run_index=0; - run= runs[run_index++]; - - put_symbol2(&s->c, b->state[1], run, 3); - - for(y=0; y>1; int py= y>>1; - if(pxparent->width && pyparent->height) + if(pxparent->width && pyparent->height) 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)); - put_cabac(&s->c, &b->state[0][context], !!v); + put_rac(&s->c, &b->state[0][context], !!v); }else{ if(!run){ run= runs[run_index++]; - put_symbol2(&s->c, b->state[1], run, 3); + if(run_index <= max_index) + put_symbol2(&s->c, b->state[1], run, 3); assert(v); }else{ run--; @@ -1881,167 +1759,203 @@ static void encode_subband_c0run(SnowContext *s, SubBand *b, DWTELEM *src, DWTEL } 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); put_symbol2(&s->c, b->state[context + 2], ABS(v)-1, context-4); - put_cabac(&s->c, &b->state[0][16 + 1 + 3 + quant3b[l&0xFF] + 3*quant3b[t&0xFF]], v<0); + put_rac(&s->c, &b->state[0][16 + 1 + 3 + quant3bA[l2&0xFF] + 3*quant3bA[t2&0xFF]], v<0); } } } } + return 0; } -static void encode_subband(SnowContext *s, SubBand *b, DWTELEM *src, DWTELEM *parent, int stride, int orientation){ +static int encode_subband(SnowContext *s, SubBand *b, DWTELEM *src, DWTELEM *parent, int stride, int orientation){ // encode_subband_qtree(s, b, src, parent, stride, orientation); // encode_subband_z0run(s, b, src, parent, stride, orientation); - encode_subband_c0run(s, b, src, parent, stride, orientation); + return encode_subband_c0run(s, b, src, parent, stride, orientation); // encode_subband_dzr(s, b, src, parent, stride, orientation); } -static inline void decode_subband(SnowContext *s, SubBand *b, DWTELEM *src, DWTELEM *parent, int stride, int orientation){ - const int level= b->level; +static inline void unpack_coeffs(SnowContext *s, SubBand *b, SubBand * parent, int orientation){ const int w= b->width; const int h= b->height; int x,y; - const int qlog= clip(s->qlog + b->qlog, 0, 128); - int qmul= qexp[qlog&7]<<(qlog>>3); - int qadd= (s->qbias*qmul)>>QBIAS_SHIFT; - - START_TIMER - - if(b->buf == s->spatial_dwt_buffer || s->qlog == LOSSLESS_QLOG){ - qadd= 0; - qmul= 1<height; y++) - memset(&src[y*stride], 0, b->width*sizeof(DWTELEM)); - - run= get_symbol2(&s->c, b->state[1], 3); + int run, runs; + x_and_coeff *xc= b->x_coeff; + x_and_coeff *prev_xc= NULL; + x_and_coeff *prev2_xc= xc; + x_and_coeff *parent_xc= parent ? parent->x_coeff : NULL; + x_and_coeff *prev_parent_xc= parent_xc; + + runs= get_symbol2(&s->c, b->state[30], 0); + if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3); + else run= INT_MAX; + for(y=0; yx[prev_index] == 0){ - rt= b->coeff[prev_index]; + if(y && prev_xc->x == 0){ + rt= prev_xc->coeff; } for(x=0; xx[prev_index] <= x) - prev_index++; - if(b->x[prev_index] == x + 1) - rt= b->coeff[prev_index]; + if(prev_xc->x <= x) + prev_xc++; + if(prev_xc->x == x + 1) + rt= prev_xc->coeff; else rt=0; } - if(parent){ - if(x>>1 > b->parent->x[parent_index]){ - parent_index++; + if(parent_xc){ + if(x>>1 > parent_xc->x){ + parent_xc++; } - if(x>>1 == b->parent->x[parent_index]){ - p= b->parent->coeff[parent_index]; + if(x>>1 == parent_xc->x){ + p= parent_xc->coeff; } } 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(/*ABS(ll) + */3*(l>>1) + (lt>>1) + (t&~1) + (rt>>1) + (p>>1)); + + v=get_rac(&s->c, &b->state[0][context]); + if(v){ + v= 2*(get_symbol2(&s->c, b->state[context + 2], context-4) + 1); + v+=get_rac(&s->c, &b->state[0][16 + 1 + 3 + quant3bA[l&0xFF] + 3*quant3bA[t&0xFF]]); - v=get_cabac(&s->c, &b->state[0][context]); + xc->x=x; + (xc++)->coeff= v; + } }else{ if(!run){ - run= get_symbol2(&s->c, b->state[1], 3); - v=1; + if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3); + else run= INT_MAX; + v= 2*(get_symbol2(&s->c, b->state[0 + 2], 0-4) + 1); + v+=get_rac(&s->c, &b->state[0][16 + 1 + 3]); + + xc->x=x; + (xc++)->coeff= v; }else{ + int max_run; run--; v=0; - if(y && parent){ - int max_run; - - max_run= FFMIN(run, b->x[prev_index] - x - 2); - max_run= FFMIN(max_run, 2*b->parent->x[parent_index] - x - 1); - x+= max_run; - run-= max_run; - } + if(y) max_run= FFMIN(run, prev_xc->x - x - 2); + else max_run= FFMIN(run, w-x-1); + if(parent_xc) + max_run= FFMIN(max_run, 2*parent_xc->x - x - 1); + x+= max_run; + run-= max_run; } } - if(v){ - int context= av_log2(/*ABS(ll) + */3*ABS(l) + ABS(lt) + 2*ABS(t) + ABS(rt) + ABS(p)); - v= get_symbol2(&s->c, b->state[context + 2], context-4) + 1; - if(get_cabac(&s->c, &b->state[0][16 + 1 + 3 + quant3b[l&0xFF] + 3*quant3b[t&0xFF]])){ - src[x + y*stride]=-(( v*qmul + qadd)>>(QEXPSHIFT)); - v= -v; - }else{ - src[x + y*stride]= (( v*qmul + qadd)>>(QEXPSHIFT)); - } - b->x[index]=x; //FIXME interleave x/coeff - b->coeff[index++]= v; - } } - b->x[index++]= w+1; //end marker - prev_index= prev2_index; - prev2_index= index; - - if(parent){ - while(b->parent->x[parent_index] != b->parent->width+1) - parent_index++; - parent_index++; + (xc++)->x= w+1; //end marker + prev_xc= prev2_xc; + prev2_xc= xc; + + if(parent_xc){ if(y&1){ - prev_parent_index= parent_index; + while(parent_xc->x != parent->width+1) + parent_xc++; + parent_xc++; + prev_parent_xc= parent_xc; }else{ - parent_index= prev_parent_index; + parent_xc= prev_parent_xc; } } } - b->x[index++]= w+1; //end marker - if(w > 200 /*level+1 == s->spatial_decomposition_count*/){ - STOP_TIMER("decode_subband") + + (xc++)->x= w+1; //end marker + } +} + +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); + 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){ + qadd= 0; + qmul= 1<stride_line + b->buf_y_offset) + b->buf_x_offset; + memset(line, 0, b->width*sizeof(DWTELEM)); + v = b->x_coeff[new_index].coeff; + x = b->x_coeff[new_index++].x; + while(x < w) + { + register int t= ( (v>>1)*qmul + qadd)>>QEXPSHIFT; + register int u= -(v&1); + line[x] = (t^u) - u; + + v = b->x_coeff[new_index].coeff; + x = b->x_coeff[new_index++].x; } - - return; } + if(w > 200 && start_y != 0/*level+1 == s->spatial_decomposition_count*/){ + STOP_TIMER("decode_subband") + } + + /* Save our variables for the next slice. */ + save_state[0] = new_index; + + return; } static void reset_contexts(SnowContext *s){ int plane_index, level, orientation; - for(plane_index=0; plane_index<2; plane_index++){ + for(plane_index=0; plane_index<3; plane_index++){ for(level=0; levelspatial_decomposition_count; level++){ for(orientation=level ? 1:0; orientation<4; orientation++){ - memset(s->plane[plane_index].band[level][orientation].state, 0, sizeof(s->plane[plane_index].band[level][orientation].state)); + memset(s->plane[plane_index].band[level][orientation].state, MID_STATE, sizeof(s->plane[plane_index].band[level][orientation].state)); } } } - memset(s->header_state, 0, sizeof(s->header_state)); - memset(s->block_state, 0, sizeof(s->block_state)); + memset(s->header_state, MID_STATE, sizeof(s->header_state)); + memset(s->block_state, MID_STATE, sizeof(s->block_state)); } static int alloc_blocks(SnowContext *s){ int w= -((-s->avctx->width )>>LOG2_MB_SIZE); int h= -((-s->avctx->height)>>LOG2_MB_SIZE); - + s->b_width = w; s->b_height= h; - + s->block= av_mallocz(w * h * sizeof(BlockNode) << (s->block_max_depth*2)); return 0; } -static inline void copy_cabac_state(CABACContext *d, CABACContext *s){ - PutBitContext bak= d->pb; +static inline void copy_rac_state(RangeCoder *d, RangeCoder *s){ + uint8_t *bytestream= d->bytestream; + uint8_t *bytestream_start= d->bytestream_start; *d= *s; - d->pb= bak; + d->bytestream= bytestream; + d->bytestream_start= bytestream_start; } //near copy & paste from dsputil, FIXME @@ -2084,7 +1998,7 @@ static inline void set_blocks(SnowContext *s, int level, int x, int y, int l, in const int block_w= 1<block_state)]; uint8_t i_state[sizeof(s->block_state)]; - CABACContext pc, ic; - PutBitContext pbbak= s->c.pb; + 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; 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; const int index= (x + y*w) << rem_depth; const int block_w= 1<<(LOG2_MB_SIZE - level); - static BlockNode null_block= { //FIXME add border maybe - .color= {128,128,128}, - .mx= 0, - .my= 0, - .type= 0, - .level= 0, - }; int trx= (x+1)<block[index-1] : &null_block; @@ -2156,16 +2064,12 @@ static int encode_q_branch(SnowContext *s, int level, int x, int y){ int pmx= mid_pred(left->mx, top->mx, tr->mx); int pmy= mid_pred(left->my, top->my, tr->my); int mx=0, my=0; - int l,cr,cb, i; + int l,cr,cb; const int stride= s->current_picture.linesize[0]; const int uvstride= s->current_picture.linesize[1]; - const int instride= s->input_picture.linesize[0]; - const int uvinstride= s->input_picture.linesize[1]; - uint8_t *new_l = s->input_picture.data[0] + (x + y* instride)*block_w; - uint8_t *new_cb= s->input_picture.data[1] + (x + y*uvinstride)*block_w/2; - uint8_t *new_cr= s->input_picture.data[2] + (x + y*uvinstride)*block_w/2; - uint8_t current_mb[3][stride*block_w]; - uint8_t *current_data[3]= {¤t_mb[0][0], ¤t_mb[1][0], ¤t_mb[2][0]}; + uint8_t *current_data[3]= { s->input_picture.data[0] + (x + y* stride)*block_w, + s->input_picture.data[1] + (x + y*uvstride)*block_w/2, + s->input_picture.data[2] + (x + y*uvstride)*block_w/2}; int P[10][2]; int16_t last_mv[3][2]; int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused @@ -2181,14 +2085,6 @@ static int encode_q_branch(SnowContext *s, int level, int x, int y){ return 0; } - //FIXME optimize - for(i=0; i>1; i++) - memcpy(¤t_mb[1][0] + uvstride*i, new_cb + uvinstride*i, block_w>>1); - for(i=0; i>1; i++) - memcpy(¤t_mb[2][0] + uvstride*i, new_cr + uvinstride*i, block_w>>1); - // clip predictors / edge ? P_LEFT[0]= left->mx; @@ -2197,36 +2093,36 @@ static int encode_q_branch(SnowContext *s, int level, int x, int y){ P_TOP [1]= top->my; P_TOPRIGHT[0]= tr->mx; P_TOPRIGHT[1]= tr->my; - + last_mv[0][0]= s->block[index].mx; last_mv[0][1]= s->block[index].my; last_mv[1][0]= right->mx; last_mv[1][1]= right->my; last_mv[2][0]= bottom->mx; last_mv[2][1]= bottom->my; - + s->m.mb_stride=2; - s->m.mb_x= + 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); - + assert(s->m.me. stride == stride); assert(s->m.me.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); c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp); c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV; - - c->xmin = - x*block_w - 16; - c->ymin = - y*block_w - 16; - c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16; - c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16; + + c->xmin = - x*block_w - 16+2; + c->ymin = - y*block_w - 16+2; + c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-2; + c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-2; if(P_LEFT[0] > (c->xmax<xmax< (c->ymax<ymax< (c->ymax<ymax< (c->xmax<xmax< (c->ymax<ymax<xmin<xmin<pred_y = P_MEDIAN[1]; } - score= ff_epzs_motion_search(&s->m, &mx, &my, P, 0, /*ref_index*/ 0, last_mv, + 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); assert(mx >= c->xmin); assert(mx <= c->xmax); assert(my >= c->ymin); assert(my <= c->ymax); - + 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 - + // subpel search pc= s->c; - init_put_bits(&pc.pb, p_buffer, sizeof(p_buffer)); + pc.bytestream_start= + pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo memcpy(p_state, s->block_state, sizeof(s->block_state)); if(level!=s->block_max_depth) - put_cabac(&pc, &p_state[4 + s_context], 1); - put_cabac(&pc, &p_state[1 + left->type + top->type], 0); + 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); - p_len= put_bits_count(&pc.pb); - score += (s->lambda2*(p_len + pc.outstanding_count - s->c.outstanding_count))>>FF_LAMBDA_SHIFT; + 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; block_s= block_w*block_w; - sum = pix_sum(¤t_mb[0][0], stride, block_w); + sum = pix_sum(current_data[0], stride, block_w); l= (sum + block_s/2)/block_s; - iscore = pix_norm1(¤t_mb[0][0], stride, block_w) - 2*l*sum + l*l*block_s; - + iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s; + block_s= block_w*block_w>>2; - sum = pix_sum(¤t_mb[1][0], uvstride, block_w>>1); + sum = pix_sum(current_data[1], uvstride, block_w>>1); cb= (sum + block_s/2)/block_s; // iscore += pix_norm1(¤t_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s; - sum = pix_sum(¤t_mb[2][0], uvstride, block_w>>1); + sum = pix_sum(current_data[2], uvstride, block_w>>1); cr= (sum + block_s/2)/block_s; // iscore += pix_norm1(¤t_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s; ic= s->c; - init_put_bits(&ic.pb, i_buffer, sizeof(i_buffer)); + ic.bytestream_start= + ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo memcpy(i_state, s->block_state, sizeof(s->block_state)); if(level!=s->block_max_depth) - put_cabac(&ic, &i_state[4 + s_context], 1); - put_cabac(&ic, &i_state[1 + left->type + top->type], 1); + put_rac(&ic, &i_state[4 + s_context], 1); + put_rac(&ic, &i_state[1 + left->type + top->type], 1); put_symbol(&ic, &i_state[32], l-pl , 1); put_symbol(&ic, &i_state[64], cb-pcb, 1); put_symbol(&ic, &i_state[96], cr-pcr, 1); - i_len= put_bits_count(&ic.pb); - iscore += (s->lambda2*(i_len + ic.outstanding_count - s->c.outstanding_count))>>FF_LAMBDA_SHIFT; + 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; // assert(score==256*256*256*64-1); assert(iscore < 255*255*256 + s->lambda2*10); @@ -2308,64 +2212,113 @@ static int encode_q_branch(SnowContext *s, int level, int x, int y){ else c->scene_change_score+= s->m.qscale; } - + if(level!=s->block_max_depth){ - put_cabac(&s->c, &s->block_state[4 + s_context], 0); + put_rac(&s->c, &s->block_state[4 + s_context], 0); score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0); score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0); score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1); score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1); score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead - + if(score2 < score && score2 < iscore) return score2; } - + if(iscore < score){ - flush_put_bits(&ic.pb); - ff_copy_bits(&pbbak, i_buffer, i_len); + memcpy(pbbak, i_buffer, i_len); s->c= ic; - s->c.pb= pbbak; + 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); memcpy(s->block_state, i_state, sizeof(s->block_state)); return iscore; }else{ - flush_put_bits(&pc.pb); - ff_copy_bits(&pbbak, p_buffer, p_len); + memcpy(pbbak, p_buffer, p_len); s->c= pc; - s->c.pb= pbbak; + s->c.bytestream_start= pbbak_start; + s->c.bytestream= pbbak + p_len; set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, 0); memcpy(s->block_state, p_state, sizeof(s->block_state)); return score; } } +static 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)); + } +} + +static void encode_q_branch2(SnowContext *s, int level, int x, int y){ + 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; + 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<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 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); + return; + } + + if(level!=s->block_max_depth){ + if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){ + put_rac(&s->c, &s->block_state[4 + s_context], 1); + }else{ + put_rac(&s->c, &s->block_state[4 + s_context], 0); + encode_q_branch2(s, level+1, 2*x+0, 2*y+0); + encode_q_branch2(s, level+1, 2*x+1, 2*y+0); + encode_q_branch2(s, level+1, 2*x+0, 2*y+1); + encode_q_branch2(s, level+1, 2*x+1, 2*y+1); + return; + } + } + if(b->type & BLOCK_INTRA){ + 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); + }else{ + put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 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); + } +} + static void decode_q_branch(SnowContext *s, int level, int x, int y){ 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; const int index= (x + y*w) << rem_depth; - static BlockNode null_block= { //FIXME add border maybe - .color= {128,128,128}, - .mx= 0, - .my= 0, - .type= 0, - .level= 0, - }; 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<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); return; } - if(level==s->block_max_depth || get_cabac(&s->c, &s->block_state[4 + s_context])){ + if(level==s->block_max_depth || get_rac(&s->c, &s->block_state[4 + s_context])){ int type; int l = left->color[0]; int cb= left->color[1]; @@ -2374,8 +2327,8 @@ static void decode_q_branch(SnowContext *s, int level, int x, int y){ 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)); - - type= get_cabac(&s->c, &s->block_state[1 + left->type + top->type]) ? BLOCK_INTRA : 0; + + type= get_rac(&s->c, &s->block_state[1 + left->type + top->type]) ? BLOCK_INTRA : 0; if(type){ l += get_symbol(&s->c, &s->block_state[32], 1); @@ -2399,9 +2352,19 @@ static void encode_blocks(SnowContext *s){ int w= s->b_width; int h= s->b_height; + if(s->avctx->me_method == ME_ITER && !s->keyframe) + iterative_me(s); + for(y=0; yc.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit + av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n"); + return; + } for(x=0; xavctx->me_method == ME_ITER) + encode_q_branch2(s, 0, x, y); + else + encode_q_branch (s, 0, x, y); } } } @@ -2420,15 +2383,15 @@ 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; - +START_TIMER for(y=0; y < b_h+5; y++){ for(x=0; x < b_w; x++){ - int a0= src[x + y*stride]; - int a1= src[x + 1 + y*stride]; - int a2= src[x + 2 + y*stride]; - int a3= src[x + 3 + y*stride]; - int a4= src[x + 4 + y*stride]; - int a5= src[x + 5 + y*stride]; + 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); @@ -2437,203 +2400,871 @@ static void mc_block(uint8_t *dst, uint8_t *src, uint8_t *tmp, int stride, int b // if(b_w==16) am= 8*(a1+a2); - if(dx<8) tmp[x + y*stride]= (32*a2*( 8-dx) + am* dx + 128)>>8; - else tmp[x + y*stride]= ( am*(16-dx) + 32*a3*(dx-8) + 128)>>8; + 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;*/ } + 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 + y *stride]; - int a1= tmp[x + (y + 1)*stride]; - int a2= tmp[x + (y + 2)*stride]; - int a3= tmp[x + (y + 3)*stride]; - int a4= tmp[x + (y + 4)*stride]; - int a5= tmp[x + (y + 5)*stride]; + 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) dst[x + y*stride]= (32*a2*( 8-dy) + am* dy + 128)>>8; - else dst[x + y*stride]= ( am*(16-dy) + 32*a3*(dy-8) + 128)>>8; + 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;*/ } + 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)];\ + assert(h==b_w);\ + mc_block(dst, src-2-2*stride, tmp, stride, b_w, b_w, dx, dy);\ +} + +mca( 0, 0,16) +mca( 8, 0,16) +mca( 0, 8,16) +mca( 8, 8,16) +mca( 0, 0,8) +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){ + if(block->type & BLOCK_INTRA){ + int x, y; + const int color = block->color[plane_index]; + const int color4= color*0x01010101; + if(b_w==16){ + for(y=0; y < b_h; y++){ + *(uint32_t*)&dst[0 + y*stride]= color4; + *(uint32_t*)&dst[4 + y*stride]= color4; + *(uint32_t*)&dst[8 + y*stride]= color4; + *(uint32_t*)&dst[12+ y*stride]= color4; + } + }else if(b_w==8){ + for(y=0; y < b_h; y++){ + *(uint32_t*)&dst[0 + y*stride]= color4; + *(uint32_t*)&dst[4 + y*stride]= color4; + } + }else if(b_w==4){ + for(y=0; y < b_h; y++){ + *(uint32_t*)&dst[0 + y*stride]= color4; + } + }else{ + for(y=0; y < b_h; y++){ + for(x=0; x < b_w; x++){ + dst[x + y*stride]= color; + } + } + } + }else{ + 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; + sx += (mx>>4) - 2; + sy += (my>>4) - 2; + 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); + 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); + if((dx&3) || (dy&3) || b_w==2 || b_h==2) + mc_block(dst, src, tmp, stride, b_w, b_h, dx, dy); + else if(b_w==b_h) + s->dsp.put_h264_qpel_pixels_tab[2-(b_w>>3)][dy+(dx>>2)](dst,src + 2 + 2*stride,stride); + else if(b_w==2*b_h){ + s->dsp.put_h264_qpel_pixels_tab[2-(b_h>>3)][dy+(dx>>2)](dst ,src + 2 + 2*stride,stride); + s->dsp.put_h264_qpel_pixels_tab[2-(b_h>>3)][dy+(dx>>2)](dst+b_h,src + 2 + b_h + 2*stride,stride); + }else{ + assert(2*b_w==b_h); + s->dsp.put_h264_qpel_pixels_tab[2-(b_w>>3)][dy+(dx>>2)](dst ,src + 2 + 2*stride ,stride); + s->dsp.put_h264_qpel_pixels_tab[2-(b_w>>3)][dy+(dx>>2)](dst+b_w*stride,src + 2 + 2*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); + dst = slice_buffer_get_line(sb, src_y + y); + for(x=0; x>= 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 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; + 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); + 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); + for(x=0; x>= 8 - FRAC_BITS; + } + if(add){ + v += dst[x + y*dst_stride]; + v = (v + (1<<(FRAC_BITS-1))) >> FRAC_BITS; + if(v&(~255)) v= ~(v>>31); + dst8[x + y*src_stride] = v; + }else{ + dst[x + y*dst_stride] -= v; + } + } + } +#endif +} + +static always_inline void predict_slice_buffered(SnowContext *s, slice_buffer * sb, DWTELEM * old_buffer, int plane_index, int add, int mb_y){ + Plane *p= &s->plane[plane_index]; + const int mb_w= s->b_width << s->block_max_depth; + const int mb_h= s->b_height << s->block_max_depth; + int x, y, mb_x; + int block_size = MB_SIZE >> s->block_max_depth; + 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]; + 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; + START_TIMER + + if(s->keyframe || (s->avctx->debug&512)){ + if(mb_y==mb_h) + return; + + if(add){ + for(y=block_w*mb_y; yline[y]; + for(x=0; x>= FRAC_BITS; + if(v&(~255)) v= ~(v>>31); + dst8[x + y*ref_stride]= v; + } + } + }else{ + for(y=block_w*mb_y; yline[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; + int x, y, mb_x; + int block_size = MB_SIZE >> s->block_max_depth; + 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; + 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; + START_TIMER + + if(s->keyframe || (s->avctx->debug&512)){ + if(mb_y==mb_h) + return; + + if(add){ + for(y=block_w*mb_y; y>= FRAC_BITS; + if(v&(~255)) v= ~(v>>31); + dst8[x + y*ref_stride]= v; + } + } + }else{ + for(y=block_w*mb_y; yb_height << s->block_max_depth; + int mb_y; + for(mb_y=0; mb_y<=mb_h; mb_y++) + predict_slice(s, buf, plane_index, add, mb_y); +} + +static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){ + int i, x2, y2; + 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]; + const static DWTELEM zero_dst[4096]; //FIXME + const int b_stride = s->b_width << s->block_max_depth; + const int w= p->width; + const int h= p->height; + int index= mb_x + mb_y*b_stride; + BlockNode *b= &s->block[index]; + BlockNode backup= *b; + int ab=0; + int aa=0; + + b->type|= BLOCK_INTRA; + b->color[plane_index]= 0; + + for(i=0; i<4; i++){ + int mb_x2= mb_x + (i &1) - 1; + int mb_y2= mb_y + (i>>1) - 1; + 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, + x, y, block_w, block_w, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, plane_index); + + for(y2= FFMAX(y, 0); y2h) obmc_v += obmc[index - block_w*obmc_stride]; + if(x+block_w>w) obmc_v += obmc[index - block_w]; + //FIXME precalc this or simplify it somehow else + + ab += (src[x2 + y2*ref_stride] - dst[x2 + y2*ref_stride]) * obmc_v; + aa += obmc_v * obmc_v; //FIXME precalclate this + } + } } -} + *b= backup; -#define mcb(dx,dy,b_w)\ -static void mc_block ## dx ## dy(uint8_t *dst, uint8_t *src, int stride){\ - uint8_t tmp[stride*(b_w+5)];\ - mc_block(dst, src-2-2*stride, tmp, stride, b_w, b_w, dx, dy);\ + return clip(((ab<<6) + aa/2)/aa, 0, 255); //FIXME we shouldnt need cliping } -mcb( 0, 0,16) -mcb( 4, 0,16) -mcb( 8, 0,16) -mcb(12, 0,16) -mcb( 0, 4,16) -mcb( 4, 4,16) -mcb( 8, 4,16) -mcb(12, 4,16) -mcb( 0, 8,16) -mcb( 4, 8,16) -mcb( 8, 8,16) -mcb(12, 8,16) -mcb( 0,12,16) -mcb( 4,12,16) -mcb( 8,12,16) -mcb(12,12,16) +static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){ + int i, y2; + 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]; + const static DWTELEM zero_dst[4096]; //FIXME + const int b_stride = s->b_width << s->block_max_depth; + const int b_height = s->b_height<< s->block_max_depth; + const int w= p->width; + const int h= p->height; + int distortion= 0; + int rate= 0; + const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp); + + for(i=0; i<4; i++){ + int mb_x2= mb_x + (i &1) - 1; + int mb_y2= mb_y + (i>>1) - 1; + 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, + x, y, block_w, block_w, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, plane_index); + + //FIXME find a cleaner/simpler way to skip the outside stuff + for(y2= y; y2<0; y2++) + memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w); + for(y2= h; y2 w){ + for(y2= y; y2dsp.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_w); + } -#define mca(dx,dy,b_w)\ -static void mc_block_hpel ## dx ## dy(uint8_t *dst, uint8_t *src, int stride, int h){\ - uint8_t tmp[stride*(b_w+5)];\ - assert(h==b_w);\ - mc_block(dst, src-2-2*stride, tmp, stride, b_w, b_w, dx, dy);\ + if(plane_index==0){ + for(i=0; i<4; i++){ +/* ..RRr + * .RXx. + * rxx.. + */ + int x= mb_x + (i&1) - (i>>1); + int y= mb_y + (i>>1); + 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+1block[index-b_stride+1] : tl; + int dmx= b->mx - mid_pred(left->mx, top->mx, tr->mx); + int dmy= b->my - 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)); + + if(x<0 || x>=b_stride || y>=b_height) + continue; +/* +1 0 0 +01X 1-2 1 +001XX 3-6 2-3 +0001XXX 7-14 4-7 +00001XXXX 15-30 8-15 +*/ +//FIXME try accurate rate +//FIXME intra and inter predictors if surrounding blocks arent the same type + if(b->type & BLOCK_INTRA){ + rate += 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 + rate += 2*(1 + av_log2(2*ABS(dmx)) + + av_log2(2*ABS(dmy))); //FIXME kill the 2* can be merged in lambda + } + } + + return distortion + rate*penalty_factor; } -mca( 0, 0,16) -mca( 8, 0,16) -mca( 0, 8,16) -mca( 8, 8,16) +static always_inline int check_block(SnowContext *s, int mb_x, int mb_y, int p[3], int intra, 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; + int rd, index, value; -static always_inline void add_xblock(SnowContext *s, DWTELEM *dst, uint8_t *src, uint8_t *obmc, int s_x, int s_y, int b_w, int b_h, int mv_x, int mv_y, int w, int h, int dst_stride, int src_stride, int obmc_stride, int mb_type, int add, int color){ - uint8_t tmp[src_stride*(b_h+5)]; //FIXME move to context to gurantee alignment - int x,y; + assert(mb_x>=0 && mb_y>=0); + assert(mb_x w){ - b_w = w - s_x; - } - if(s_y<0){ - obmc -= s_y*obmc_stride; - b_h += s_y; - s_y=0; - }else if(s_y + b_h> h){ - b_h = h - s_y; + if(intra){ + block->color[0] = p[0]; + block->color[1] = p[1]; + block->color[2] = p[2]; + 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); + if(s->me_cache[index] == value) + return 0; + s->me_cache[index]= value; + + block->mx= p[0]; + block->my= p[1]; + block->type &= ~BLOCK_INTRA; } - if(b_w<=0 || b_h<=0) return; - - dst += s_x + s_y*dst_stride; - - if(mb_type==BLOCK_INTRA){ - for(y=0; y < b_h; y++){ - for(x=0; x < b_w; x++){ - if(add) dst[x + y*dst_stride] += obmc[x + y*obmc_stride] * color * (256/OBMC_MAX); - else dst[x + y*dst_stride] -= obmc[x + y*obmc_stride] * color * (256/OBMC_MAX); - } - } + rd= get_block_rd(s, mb_x, mb_y, 0); + +//FIXME chroma + if(rd < *best_rd){ + *best_rd= rd; + return 1; }else{ - int dx= mv_x&15; - int dy= mv_y&15; -// int dxy= (mv_x&1) + 2*(mv_y&1); - - s_x += (mv_x>>4) - 2; - s_y += (mv_y>>4) - 2; - src += s_x + s_y*src_stride; - //use dsputil - - if( (unsigned)s_x >= w - b_w - 4 - || (unsigned)s_y >= h - b_h - 4){ - ff_emulated_edge_mc(tmp + 32, src, src_stride, b_w+5, b_h+5, s_x, s_y, w, h); - src= tmp + 32; - } - - assert(mb_type==0); - mc_block(tmp, src, tmp + 64+8, src_stride, b_w, b_h, dx, dy); - for(y=0; y < b_h; y++){ - for(x=0; x < b_w; x++){ - if(add) dst[x + y*dst_stride] += obmc[x + y*obmc_stride] * tmp[x + y*src_stride] * (256/OBMC_MAX); - else dst[x + y*dst_stride] -= obmc[x + y*obmc_stride] * tmp[x + y*src_stride] * (256/OBMC_MAX); - } - } + *block= backup; + return 0; } } -static void predict_plane(SnowContext *s, DWTELEM *buf, int plane_index, int add){ - 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; - const int mb_stride= mb_w; - int x, y, mb_x, mb_y; - int scale = plane_index ? s->mv_scale : 2*s->mv_scale; - int block_size = MB_SIZE >> s->block_max_depth; - int block_w = plane_index ? block_size/2 : block_size; - 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->last_picture.linesize[plane_index]; - uint8_t *ref = s->last_picture.data[plane_index]; - int w= p->width; - int h= p->height; - START_TIMER - -if(s->avctx->debug&512){ - for(y=0; yblock[index].mx*scale, s->block[index].my*scale, - w, h, - w, ref_stride, obmc_stride, - s->block[index].type, add, s->block[index].color[plane_index]); - STOP_TIMER("add_xblock") +static void iterative_me(SnowContext *s){ + int pass, mb_x, mb_y; + 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; + int color[3]; + + for(pass=0; pass<50; 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_xblock[index +1] : &null_block; + BlockNode *bb = mb_yblock[index+b_stride ] : &null_block; + BlockNode *tlb= mb_x && mb_y ? &s->block[index-b_stride-1] : &null_block; + BlockNode *trb= mb_xblock[index-b_stride+1] : &null_block; + BlockNode *blb= mb_x && mb_yblock[index+b_stride-1] : &null_block; + BlockNode *brb= mb_xblock[index+b_stride+1] : &null_block; + + if(pass && (block->type & BLOCK_OPT)) + continue; + block->type |= BLOCK_OPT; + + backup= *block; + + if(!s->me_cache_generation) + memset(s->me_cache, 0, sizeof(s->me_cache)); + s->me_cache_generation += 1<<22; + + // get previous score (cant be cached due to OBMC) + check_block_inter(s, mb_x, mb_y, block->mx, block->my, 0, &best_rd); + check_block_inter(s, mb_x, mb_y, 0, 0, 0, &best_rd); + check_block_inter(s, mb_x, mb_y, tb->mx, tb->my, 0, &best_rd); + check_block_inter(s, mb_x, mb_y, lb->mx, lb->my, 0, &best_rd); + check_block_inter(s, mb_x, mb_y, rb->mx, rb->my, 0, &best_rd); + check_block_inter(s, mb_x, mb_y, bb->mx, bb->my, 0, &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, &best_rd); + dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), 0, &best_rd); + dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), 0, &best_rd); + dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), 0, &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], 0, &best_rd); + }while(dia_change); + //FIXME or try the standard 2 pass qpel or similar +#if 1 + for(i=0; i<3; i++){ + color[i]= get_dc(s, mb_x, mb_y, i); + } + check_block(s, mb_x, mb_y, color, 1, &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; + change ++; + } + } } + av_log(NULL, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change); + if(!change) + break; } - - STOP_TIMER("predict_plane") } static void quantize(SnowContext *s, SubBand *b, 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, 128); - const int qmul= qexp[qlog&7]<<(qlog>>3); + const int qlog= clip(s->qlog + b->qlog, 0, QROOT*16); + const int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT); int x,y, thres1, thres2; - START_TIMER - - assert(QROOT==8); +// START_TIMER if(s->qlog == LOSSLESS_QLOG) return; - + bias= bias ? 0 : (3*qmul)>>3; thres1= ((qmul - bias)>>QEXPSHIFT) - 1; thres2= 2*thres1; - + if(!bias){ for(y=0; y thres2){ if(i>=0){ i<<= QEXPSHIFT; @@ -2652,8 +3283,8 @@ static void quantize(SnowContext *s, SubBand *b, DWTELEM *src, int stride, int b }else{ for(y=0; y thres2){ if(i>=0){ i<<= QEXPSHIFT; @@ -2675,19 +3306,43 @@ static void quantize(SnowContext *s, SubBand *b, DWTELEM *src, int stride, int b } } +static void dequantize_slice_buffered(SnowContext *s, slice_buffer * sb, SubBand *b, DWTELEM *src, int stride, int start_y, int end_y){ + const int w= b->width; + const int qlog= clip(s->qlog + b->qlog, 0, QROOT*16); + const int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT); + const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT; + int x,y; + START_TIMER + + if(s->qlog == LOSSLESS_QLOG) return; + + for(y=start_y; ystride_line) + b->buf_y_offset) + b->buf_x_offset; + for(x=0; x>(QEXPSHIFT)); //FIXME try different bias + }else if(i>0){ + line[x]= (( i*qmul + qadd)>>(QEXPSHIFT)); + } + } + } + if(w > 200 /*level+1 == s->spatial_decomposition_count*/){ + STOP_TIMER("dquant") + } +} + static void dequantize(SnowContext *s, SubBand *b, DWTELEM *src, int stride){ - const int level= b->level; const int w= b->width; const int h= b->height; - const int qlog= clip(s->qlog + b->qlog, 0, 128); - const int qmul= qexp[qlog&7]<<(qlog>>3); + const int qlog= 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; START_TIMER - + if(s->qlog == LOSSLESS_QLOG) return; - - assert(QROOT==8); for(y=0; ywidth; const int h= b->height; int x,y; - + for(y=h-1; y>=0; y--){ for(x=w-1; x>=0; x--){ int i= x + y*stride; - + if(x){ if(use_median){ if(y && x+1width; + int x,y; + +// START_TIMER + + DWTELEM * line; + DWTELEM * 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; + + for(y=start_y; ystride_line) + b->buf_y_offset) + b->buf_x_offset; + for(x=0; xwidth; const int h= b->height; int x,y; - + for(y=0; yc, s->header_state, s->keyframe); // state clearing stuff? + memset(kstate, MID_STATE, sizeof(kstate)); + + put_rac(&s->c, kstate, s->keyframe); + if(s->keyframe || s->always_reset) + reset_contexts(s); if(s->keyframe){ put_symbol(&s->c, s->header_state, s->version, 0); + put_rac(&s->c, s->header_state, s->always_reset); put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0); put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0); put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0); put_symbol(&s->c, s->header_state, s->colorspace_type, 0); put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0); put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0); - put_cabac(&s->c, s->header_state, s->spatial_scalability); -// put_cabac(&s->c, s->header_state, s->rate_scalability); + put_rac(&s->c, s->header_state, s->spatial_scalability); +// put_rac(&s->c, s->header_state, s->rate_scalability); for(plane_index=0; plane_index<2; plane_index++){ for(level=0; levelspatial_decomposition_count; level++){ @@ -2777,30 +3472,36 @@ static void encode_header(SnowContext *s){ } } put_symbol(&s->c, s->header_state, s->spatial_decomposition_type, 0); - put_symbol(&s->c, s->header_state, s->qlog, 1); - put_symbol(&s->c, s->header_state, s->mv_scale, 0); + put_symbol(&s->c, s->header_state, s->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; + uint8_t kstate[32]; - s->keyframe= get_cabac(&s->c, s->header_state); + memset(kstate, MID_STATE, sizeof(kstate)); + + s->keyframe= get_rac(&s->c, kstate); + if(s->keyframe || s->always_reset) + reset_contexts(s); if(s->keyframe){ s->version= get_symbol(&s->c, s->header_state, 0); if(s->version>0){ av_log(s->avctx, AV_LOG_ERROR, "version %d not supported", s->version); return -1; } + s->always_reset= get_rac(&s->c, s->header_state); s->temporal_decomposition_type= get_symbol(&s->c, s->header_state, 0); s->temporal_decomposition_count= get_symbol(&s->c, s->header_state, 0); s->spatial_decomposition_count= get_symbol(&s->c, s->header_state, 0); s->colorspace_type= get_symbol(&s->c, s->header_state, 0); s->chroma_h_shift= get_symbol(&s->c, s->header_state, 0); s->chroma_v_shift= get_symbol(&s->c, s->header_state, 0); - s->spatial_scalability= get_cabac(&s->c, s->header_state); -// s->rate_scalability= get_cabac(&s->c, s->header_state); + 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++){ @@ -2814,34 +3515,52 @@ static int decode_header(SnowContext *s){ } } } - + s->spatial_decomposition_type= get_symbol(&s->c, s->header_state, 0); if(s->spatial_decomposition_type > 2){ av_log(s->avctx, AV_LOG_ERROR, "spatial_decomposition_type %d not supported", s->spatial_decomposition_type); return -1; } - + s->qlog= get_symbol(&s->c, s->header_state, 1); s->mv_scale= get_symbol(&s->c, s->header_state, 0); s->qbias= get_symbol(&s->c, s->header_state, 1); s->block_max_depth= get_symbol(&s->c, s->header_state, 0); + if(s->block_max_depth > 1){ + 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; + } return 0; } +static void init_qexp(){ + int i; + double v=128; + + for(i=0; ipriv_data; int width, height; int level, orientation, plane_index, dec; s->avctx= avctx; - + dsputil_init(&s->dsp, avctx); #define mcf(dx,dy)\ s->dsp.put_qpel_pixels_tab [0][dy+dx/4]=\ s->dsp.put_no_rnd_qpel_pixels_tab[0][dy+dx/4]=\ - mc_block ## dx ## dy; + s->dsp.put_h264_qpel_pixels_tab[0][dy+dx/4];\ + s->dsp.put_qpel_pixels_tab [1][dy+dx/4]=\ + s->dsp.put_no_rnd_qpel_pixels_tab[1][dy+dx/4]=\ + s->dsp.put_h264_qpel_pixels_tab[1][dy+dx/4]; mcf( 0, 0) mcf( 4, 0) @@ -2863,31 +3582,36 @@ static int common_init(AVCodecContext *avctx){ #define mcfh(dx,dy)\ s->dsp.put_pixels_tab [0][dy/4+dx/8]=\ s->dsp.put_no_rnd_pixels_tab[0][dy/4+dx/8]=\ - mc_block_hpel ## dx ## dy; + mc_block_hpel ## dx ## dy ## 16;\ + s->dsp.put_pixels_tab [1][dy/4+dx/8]=\ + s->dsp.put_no_rnd_pixels_tab[1][dy/4+dx/8]=\ + mc_block_hpel ## dx ## dy ## 8; mcfh(0, 0) mcfh(8, 0) mcfh(0, 8) mcfh(8, 8) - + + 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->pred_buffer= av_mallocz(width*height*sizeof(DWTELEM)); - + s->mv_scale= (s->avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4; s->block_max_depth= (s->avctx->flags & CODEC_FLAG_4MV) ? 1 : 0; - - for(plane_index=0; plane_index<3; plane_index++){ + + for(plane_index=0; plane_index<3; plane_index++){ int w= s->avctx->width; int h= s->avctx->height; @@ -2901,35 +3625,44 @@ static int common_init(AVCodecContext *avctx){ for(level=s->spatial_decomposition_count-1; level>=0; level--){ for(orientation=level ? 1 : 0; orientation<4; orientation++){ SubBand *b= &s->plane[plane_index].band[level][orientation]; - + b->buf= s->spatial_dwt_buffer; b->level= level; b->stride= s->plane[plane_index].width << (s->spatial_decomposition_count - level); b->width = (w + !(orientation&1))>>1; b->height= (h + !(orientation>1))>>1; - - if(orientation&1) b->buf += (w+1)>>1; - if(orientation>1) b->buf += b->stride>>1; - + + b->stride_line = 1 << (s->spatial_decomposition_count - level); + b->buf_x_offset = 0; + b->buf_y_offset = 0; + + if(orientation&1){ + b->buf += (w+1)>>1; + b->buf_x_offset = (w+1)>>1; + } + if(orientation>1){ + b->buf += b->stride>>1; + b->buf_y_offset = b->stride_line >> 1; + } + if(level) b->parent= &s->plane[plane_index].band[level-1][orientation]; - b->x = av_mallocz(((b->width+1) * b->height+1)*sizeof(int16_t)); - b->coeff= av_mallocz(((b->width+1) * b->height+1)*sizeof(DWTELEM)); + b->x_coeff=av_mallocz(((b->width+1) * b->height+1)*sizeof(x_and_coeff)); } w= (w+1)>>1; h= (h+1)>>1; } } - + 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; } @@ -2937,14 +3670,14 @@ static int common_init(AVCodecContext *avctx){ static void calculate_vissual_weight(SnowContext *s, Plane *p){ int width = p->width; int height= p->height; - int i, level, orientation, x, y; + int level, orientation, x, y; 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; 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); @@ -2964,30 +3697,42 @@ static void calculate_vissual_weight(SnowContext *s, Plane *p){ static int encode_init(AVCodecContext *avctx) { SnowContext *s = avctx->priv_data; - int i; - int level, orientation, plane_index; + int plane_index; - if(avctx->strict_std_compliance >= 0){ - av_log(avctx, AV_LOG_ERROR, "this codec is under development, files encoded with it wont be decodeable with future versions!!!\n" - "use vstrict=-1 to use it anyway\n"); + if(avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL){ + av_log(avctx, AV_LOG_ERROR, "this codec is under development, files encoded with it may not be decodable with future versions!!!\n" + "use vstrict=-2 / -strict -2 to use it anyway\n"); return -1; } - + common_init(avctx); alloc_blocks(s); - + s->version=0; - + + s->m.avctx = avctx; + s->m.flags = avctx->flags; + s->m.bit_rate= avctx->bit_rate; + s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t)); s->m.me.map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t)); s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t)); h263_encode_init(&s->m); //mv_penalty + if(avctx->flags&CODEC_FLAG_PASS1){ + if(!avctx->stats_out) + avctx->stats_out = av_mallocz(256); + } + if(avctx->flags&CODEC_FLAG_PASS2){ + 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]); } - - + + avctx->coded_frame= &s->current_picture; switch(avctx->pix_fmt){ // case PIX_FMT_YUV444P: @@ -3008,6 +3753,12 @@ static int encode_init(AVCodecContext *avctx) // avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift); s->chroma_h_shift= 1; s->chroma_v_shift= 1; + + ff_set_cmp(&s->dsp, s->dsp.me_cmp, s->avctx->me_cmp); + ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp); + + s->avctx->get_buffer(s->avctx, &s->input_picture); + return 0; } @@ -3016,9 +3767,6 @@ static int frame_start(SnowContext *s){ int w= s->avctx->width; //FIXME round up to x16 ? int h= s->avctx->height; - if(s->keyframe) - reset_contexts(s); - if(s->current_picture.data[0]){ draw_edges(s->current_picture.data[0], s->current_picture.linesize[0], w , h , EDGE_WIDTH ); draw_edges(s->current_picture.data[1], s->current_picture.linesize[1], w>>1, h>>1, EDGE_WIDTH/2); @@ -3028,61 +3776,71 @@ static int frame_start(SnowContext *s){ tmp= s->last_picture; s->last_picture= s->current_picture; s->current_picture= tmp; - + 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; } - + return 0; } static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data){ SnowContext *s = avctx->priv_data; - CABACContext * const c= &s->c; + RangeCoder * const c= &s->c; AVFrame *pict = data; const int width= s->avctx->width; const int height= s->avctx->height; - int used_count= 0; - int log2_threshold, level, orientation, plane_index, i; + int level, orientation, plane_index, i, y; - ff_init_cabac_encoder(c, buf, buf_size); - ff_init_cabac_states(c, ff_h264_lps_range, ff_h264_mps_state, ff_h264_lps_state, 64); - - s->input_picture = *pict; + ff_init_range_encoder(c, buf, buf_size); + ff_build_rac_states(c, 0.05*(1LL<<32), 256-8); - memset(s->header_state, 0, sizeof(s->header_state)); + for(i=0; i<3; i++){ + int shift= !!i; + for(y=0; y<(height>>shift); y++) + memcpy(&s->input_picture.data[i][y * s->input_picture.linesize[i]], + &pict->data[i][y * pict->linesize[i]], + width>>shift); + } + s->new_picture = *pict; + + 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); + }else{ + s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0; + pict->pict_type= s->keyframe ? FF_I_TYPE : FF_P_TYPE; + } - s->keyframe=avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0; - pict->pict_type= s->keyframe ? FF_I_TYPE : FF_P_TYPE; - if(pict->quality){ s->qlog= rint(QROOT*log(pict->quality / (float)FF_QP2LAMBDA)/log(2)); //<64 >60 - s->qlog += 61; + s->qlog += 61*QROOT/8; }else{ s->qlog= LOSSLESS_QLOG; } 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){ int block_width = (width +15)>>4; int block_height= (height+15)>>4; int stride= s->current_picture.linesize[0]; - uint8_t *src_plane= s->input_picture.data[0]; - int src_stride= s->input_picture.linesize[0]; - int x,y; - + assert(s->current_picture.data[0]); assert(s->last_picture.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. new_picture.data[0]= s-> input_picture.data[0]; - s->m.current_picture_ptr= &s->m.current_picture; s->m. last_picture_ptr= &s->m. last_picture; s->m.linesize= s->m. last_picture.linesize[0]= @@ -3110,56 +3868,59 @@ static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size, s->m.dsp= s->dsp; //move ff_init_me(&s->m); + s->dsp= s->m.dsp; } - + redo_frame: - + s->qbias= pict->pict_type == P_TYPE ? 2 : 0; encode_header(s); + s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start); encode_blocks(s); - + 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++){ Plane *p= &s->plane[plane_index]; int w= p->width; int h= p->height; int x, y; - int bits= put_bits_count(&s->c.pb); +// int bits= put_bits_count(&s->c.pb); //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]<<8; + s->spatial_dwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<spatial_dwt_buffer, plane_index, 0); - - if( plane_index==0 - && pict->pict_type == P_TYPE + + if( plane_index==0 + && pict->pict_type == P_TYPE && s->m.me.scene_change_score > s->avctx->scenechange_threshold){ - ff_init_cabac_encoder(c, buf, buf_size); - ff_init_cabac_states(c, ff_h264_lps_range, ff_h264_mps_state, ff_h264_lps_state, 64); + 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); goto redo_frame; } - + if(s->qlog == LOSSLESS_QLOG){ for(y=0; yspatial_dwt_buffer[y*w + x]= (s->spatial_dwt_buffer[y*w + x] + 127)>>8; + s->spatial_dwt_buffer[y*w + x]= (s->spatial_dwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS; } } } - + ff_spatial_dwt(s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count); 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); if(orientation==0) decorrelate(s, b, b->buf, b->stride, pict->pict_type == P_TYPE, 0); @@ -3183,22 +3944,16 @@ redo_frame: if(s->qlog == LOSSLESS_QLOG){ for(y=0; yspatial_dwt_buffer[y*w + x]<<=8; + s->spatial_dwt_buffer[y*w + x]<<=FRAC_BITS; } } } +{START_TIMER predict_plane(s, s->spatial_dwt_buffer, plane_index, 1); - //FIXME optimize - for(y=0; yspatial_dwt_buffer[y*w + x]+128)>>8; - if(v&(~255)) v= ~(v>>31); - s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x]= v; - } - } +STOP_TIMER("pred-conv")} if(s->avctx->flags&CODEC_FLAG_PSNR){ int64_t error= 0; - + if(pict->data[plane_index]) //FIXME gray hack for(y=0; yavctx->error[plane_index] += error; - s->avctx->error[3] += error; + s->current_picture.error[plane_index] = error; } } if(s->last_picture.data[0]) avctx->release_buffer(avctx, &s->last_picture); + 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; + ff_write_pass1_stats(&s->m); + } + if(avctx->flags&CODEC_FLAG_PASS2){ + s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start); + } + emms_c(); - - return put_cabac_terminate(c, 1); + + return ff_rac_terminate(c); } static void common_end(SnowContext *s){ @@ -3224,19 +3994,18 @@ static void common_end(SnowContext *s){ av_freep(&s->spatial_dwt_buffer); - av_freep(&s->m.me.scratchpad); + av_freep(&s->m.me.scratchpad); av_freep(&s->m.me.map); av_freep(&s->m.me.score_map); - + av_freep(&s->block); - for(plane_index=0; plane_index<3; plane_index++){ + 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++){ SubBand *b= &s->plane[plane_index].band[level][orientation]; - - av_freep(&b->x); - av_freep(&b->coeff); + + av_freep(&b->x_coeff); } } } @@ -3247,37 +4016,35 @@ static int encode_end(AVCodecContext *avctx) SnowContext *s = avctx->priv_data; common_end(s); + av_free(avctx->stats_out); return 0; } static int decode_init(AVCodecContext *avctx) { -// SnowContext *s = avctx->priv_data; + 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; } static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, uint8_t *buf, int buf_size){ SnowContext *s = avctx->priv_data; - CABACContext * const c= &s->c; - const int width= s->avctx->width; - const int height= s->avctx->height; + RangeCoder * const c= &s->c; int bytes_read; AVFrame *picture = data; - int log2_threshold, level, orientation, plane_index; - - - /* no supplementary picture */ - if (buf_size == 0) - return 0; - - ff_init_cabac_decoder(c, buf, buf_size); - ff_init_cabac_states(c, ff_h264_lps_range, ff_h264_mps_state, ff_h264_lps_state, 64); + int level, orientation, plane_index; - memset(s->header_state, 0, sizeof(s->header_state)); + 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); @@ -3287,7 +4054,7 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, uint8 //keyframe flag dupliaction mess FIXME if(avctx->debug&FF_DEBUG_PICT_INFO) av_log(avctx, AV_LOG_ERROR, "keyframe:%d qlog:%d\n", s->keyframe, s->qlog); - + decode_blocks(s); for(plane_index=0; plane_index<3; plane_index++){ @@ -3295,66 +4062,129 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, uint8 int w= p->width; int h= p->height; int x, y; - + int decode_state[MAX_DECOMPOSITIONS][4][1]; /* Stored state info for unpack_coeffs. 1 variable per instance. */ + 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); for(y=0; yspatial_dwt_buffer[y*w + x]+128)>>8; - if(v&(~255)) v= ~(v>>31); + int v= s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x]; s->mconly_picture.data[plane_index][y*s->mconly_picture.linesize[plane_index] + x]= v; } } } + +{ START_TIMER + for(level=0; levelspatial_decomposition_count; level++){ + for(orientation=level ? 1 : 0; orientation<4; orientation++){ + SubBand *b= &p->band[level][orientation]; + unpack_coeffs(s, b, b->parent, orientation); + } + } + STOP_TIMER("unpack coeffs"); +} + +{START_TIMER + const int mb_h= s->b_height << s->block_max_depth; + const int block_size = MB_SIZE >> s->block_max_depth; + const int block_w = plane_index ? block_size/2 : block_size; + int mb_y; + dwt_compose_t cs[MAX_DECOMPOSITIONS]; + int yd=0, yq=0; + int y; + int end_y; + + ff_spatial_idwt_buffered_init(cs, &s->sb, w, h, 1, s->spatial_decomposition_type, s->spatial_decomposition_count); + for(mb_y=0; mb_y<=mb_h; mb_y++){ + + int slice_starty = block_w*mb_y; + int slice_h = block_w*(mb_y+1); + if (!(s->keyframe || s->avctx->debug&512)){ + slice_starty = FFMAX(0, slice_starty - (block_w >> 1)); + slice_h -= (block_w >> 1); + } + + { + START_TIMER for(level=0; levelspatial_decomposition_count; level++){ for(orientation=level ? 1 : 0; orientation<4; orientation++){ SubBand *b= &p->band[level][orientation]; - - decode_subband(s, b, b->buf, b->parent ? b->parent->buf : NULL, b->stride, orientation); - if(orientation==0){ - correlate(s, b, b->buf, b->stride, 1, 0); - dequantize(s, b, b->buf, b->stride); - assert(b->buf == s->spatial_dwt_buffer); + int start_y; + int end_y; + int our_mb_start = mb_y; + int our_mb_end = (mb_y + 1); + const int extra= 3; + start_y = (mb_y ? ((block_w * our_mb_start) >> (s->spatial_decomposition_count - level)) + s->spatial_decomposition_count - level + extra: 0); + end_y = (((block_w * our_mb_end) >> (s->spatial_decomposition_count - level)) + s->spatial_decomposition_count - level + extra); + if (!(s->keyframe || s->avctx->debug&512)){ + start_y = FFMAX(0, start_y - (block_w >> (1+s->spatial_decomposition_count - level))); + end_y = FFMAX(0, end_y - (block_w >> (1+s->spatial_decomposition_count - level))); + } + start_y = FFMIN(b->height, start_y); + end_y = FFMIN(b->height, end_y); + + if (start_y != end_y){ + if (orientation == 0){ + SubBand * correlate_band = &p->band[0][0]; + 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); + } + else + decode_subband_slice_buffered(s, b, &s->sb, start_y, end_y, decode_state[level][orientation]); } } } + STOP_TIMER("decode_subband_slice"); + } + +{ START_TIMER + for(; ydsb, w, h, 1, s->spatial_decomposition_type, s->spatial_decomposition_count, yd); + } + STOP_TIMER("idwt slice");} + - ff_spatial_idwt(s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count); if(s->qlog == LOSSLESS_QLOG){ - for(y=0; ysb, yq); for(x=0; xspatial_dwt_buffer[y*w + x]<<=8; + line[x] <<= FRAC_BITS; } } } - predict_plane(s, s->spatial_dwt_buffer, plane_index, 1); - //FIXME optimize - for(y=0; yspatial_dwt_buffer[y*w + x]+128)>>8; - if(v&(~255)) v= ~(v>>31); - s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x]= v; - } - } + predict_slice_buffered(s, &s->sb, s->spatial_dwt_buffer, plane_index, 1, mb_y); + + y = FFMIN(p->height, slice_starty); + end_y = FFMIN(p->height, slice_h); + while(y < end_y) + slice_buffer_release(&s->sb, y++); + } + + slice_buffer_flush(&s->sb); + +STOP_TIMER("idwt + predict_slices")} } - + emms_c(); if(s->last_picture.data[0]) avctx->release_buffer(avctx, &s->last_picture); -if(!(s->avctx->debug&2048)) +if(!(s->avctx->debug&2048)) *picture= s->current_picture; else *picture= s->mconly_picture; - + *data_size = sizeof(AVFrame); - - bytes_read= get_cabac_terminate(c); - if(bytes_read ==0) av_log(s->avctx, AV_LOG_ERROR, "error at end of frame\n"); + + bytes_read= c->bytestream - c->bytestream_start; + if(bytes_read ==0) av_log(s->avctx, AV_LOG_ERROR, "error at end of frame\n"); //FIXME return bytes_read; } @@ -3363,6 +4193,8 @@ static int decode_end(AVCodecContext *avctx) { SnowContext *s = avctx->priv_data; + slice_buffer_destroy(&s->sb); + common_end(s); return 0; @@ -3381,6 +4213,7 @@ AVCodec snow_decoder = { NULL }; +#ifdef CONFIG_ENCODERS AVCodec snow_encoder = { "snow", CODEC_TYPE_VIDEO, @@ -3390,6 +4223,7 @@ AVCodec snow_encoder = { encode_frame, encode_end, }; +#endif #if 0 @@ -3405,14 +4239,14 @@ int main(){ int i; s.spatial_decomposition_count=6; s.spatial_decomposition_type=1; - + printf("testing 5/3 DWT\n"); for(i=0; ispatial_decomposition_type, s->spatial_decomposition_count); - ff_spatial_idwt(buffer[0], width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count); - + + ff_spatial_dwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count); + ff_spatial_idwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count); + for(i=0; ispatial_decomposition_type, s->spatial_decomposition_count); - ff_spatial_idwt(buffer[0], width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count); - + + ff_spatial_dwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count); + 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 0 printf("testing AC coder\n"); memset(s.header_state, 0, sizeof(s.header_state)); - ff_init_cabac_encoder(&s.c, buffer[0], 256*256); + ff_init_range_encoder(&s.c, buffer[0], 256*256); ff_init_cabac_states(&s.c, ff_h264_lps_range, ff_h264_mps_state, ff_h264_lps_state, 64); - + for(i=-256; i<256; i++){ START_TIMER put_symbol(&s.c, s.header_state, i*i*i/3*ABS(i), 1); STOP_TIMER("put_symbol") } - put_cabac_terminate(&s.c, 1); + ff_rac_terminate(&s.c); memset(s.header_state, 0, sizeof(s.header_state)); - ff_init_cabac_decoder(&s.c, buffer[0], 256*256); + ff_init_range_decoder(&s.c, buffer[0], 256*256); ff_init_cabac_states(&s.c, ff_h264_lps_range, ff_h264_mps_state, ff_h264_lps_state, 64); - + for(i=-256; i<256; i++){ int j; START_TIMER @@ -3450,6 +4285,7 @@ START_TIMER STOP_TIMER("get_symbol") if(j!=i*i*i/3*ABS(i)) printf("fsck: %d != %d\n", i, j); } +#endif { int level, orientation, x, y; int64_t errors[8][4]; @@ -3468,10 +4304,10 @@ int64_t g=0; if(orientation&1) buf+=w; if(orientation>1) buf+=stride>>1; - + memset(buffer[0], 0, sizeof(int)*width*height); buf[w/2 + h/2*stride]= 256*256; - ff_spatial_idwt(buffer[0], width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count); + ff_spatial_idwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count); for(y=0; y>1; - + memset(buffer[0], 0, sizeof(int)*width*height); #if 1 for(y=0; yspatial_decomposition_type, s->spatial_decomposition_count); + ff_spatial_dwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count); #else for(y=0; yspatial_decomposition_type, s->spatial_decomposition_count); + ff_spatial_idwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count); #endif for(y=0; y