comparison snow.c @ 2562:d0a58dca5ad2 libavcodec

slice-based idwt (10% faster decoding)

2561:eb72c01df6ed

    BlockNode *block;

    MpegEncContext m; // needed for motion estimation, should not be used for anything else, the idea is to make the motion estimation eventually independant of MpegEncContext, so this will be removed then (FIXME/XXX)
}SnowContext;

#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

    for(i=0; i<width; i++){
        b1[i] -= (b0[i] + b2[i] + 2)>>2;
    }
}

static void spatial_compose53i(DWTELEM *buffer, int width, int height, int stride){
    int 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;

{START_TIMER
        if(b1 <= b3) vertical_compose53iL0(b1, b2, b3, width);
        if(b0 <= b2) 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);
STOP_TIMER("horizontal_compose53i")}

        b0=b2;
        b1=b3;
    }
}   

 
static void horizontal_compose97i(DWTELEM *b, int width){
    DWTELEM temp[width];

    for(i=0; i<width; i++){
        b1[i] -= (W_DM*(b0[i] + b2[i])+W_DO)>>W_DS;
    }
}

static void spatial_compose97i(DWTELEM *buffer, int width, int height, int stride){
    int 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;

    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;

        if(stride == width && y+4 < height && 0){ 
            int x;
            for(x=0; x<width/2; x++)
                b5[x] += 64*2;

{START_TIMER
        if(y-1>=  0) horizontal_compose97i(b0, width);
        if(b0 <= b2) horizontal_compose97i(b1, width);
if(width>400 && b0 <= b2){
STOP_TIMER("horizontal_compose97i")}}
        
        b0=b2;
        b1=b3;
        b2=b4;
        b3=b5;
    }
}

void ff_spatial_idwt(DWTELEM *buffer, int width, int height, int stride, 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); break;
        case 1: spatial_compose53i(buffer, width>>level, height>>level, stride<<level); break;
        case 2: spatial_composeX  (buffer, width>>level, height>>level, stride<<level); break;
        }
    }
}

static int encode_subband_c0run(SnowContext *s, SubBand *b, DWTELEM *src, DWTELEM *parent, int stride, int orientation){
    const int w= b->width;

        }
    }
#endif
}

static always_inline 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;
    int x, y, mb_x, mb_y;
    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];

    int w= p->width;
    int h= p->height;
    START_TIMER
    
    if(s->keyframe || (s->avctx->debug&512)){
        if(add){
            for(y=0; y<h; y++){
                for(x=0; x<w; x++){
                    int v= buf[x + y*w] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
                    v >>= FRAC_BITS;
                    if(v&(~255)) v= ~(v>>31);
                    dst8[x + y*ref_stride]= v;
                }
            }
        }else{
            for(y=0; y<h; y++){
                for(x=0; x<w; x++){
                    buf[x + y*w]-= 128<<FRAC_BITS;
                }
            }
        }

        return;
    }
    
    for(mb_y=0; mb_y<=mb_h; mb_y++){
        for(mb_x=0; mb_x<=mb_w; mb_x++){
            START_TIMER

            add_yblock(s, buf, dst8, ref, obmc, 
                       block_w*mb_x - block_w/2,

                       mb_x - 1, mb_y - 1,
                       add, plane_index);
            
            STOP_TIMER("add_yblock")
        }
    }
    
    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;

                    assert(b->buf == s->spatial_dwt_buffer);
                }
            }
        }

        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; y<h; y++){
                for(x=0; x<w; x++){
                    s->spatial_dwt_buffer[y*w + x]<<=FRAC_BITS;
                }
            }
        }
{START_TIMER
        predict_plane(s, s->spatial_dwt_buffer, plane_index, 1);
STOP_TIMER("predict_plane conv2")}
    }
            
    emms_c();

    if(s->last_picture.data[0])

2562:d0a58dca5ad2

    BlockNode *block;

    MpegEncContext m; // needed for motion estimation, should not be used for anything else, the idea is to make the motion estimation eventually independant of MpegEncContext, so this will be removed then (FIXME/XXX)
}SnowContext;

typedef struct {
    DWTELEM *b0;
    DWTELEM *b1;
    DWTELEM *b2;
    DWTELEM *b3;
    int y;
} dwt_compose_t;

#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

    for(i=0; i<width; i++){
        b1[i] -= (b0[i] + b2[i] + 2)>>2;
    }
}

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(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(b1 <= b3) vertical_compose53iL0(b1, b2, b3, width);
        if(b0 <= b2) 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);
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(DWTELEM *b, int width){
    DWTELEM temp[width];

    for(i=0; i<width; i++){
        b1[i] -= (W_DM*(b0[i] + b2[i])+W_DO)>>W_DS;
    }
}

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(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; x<width/2; x++)
                b5[x] += 64*2;

{START_TIMER
        if(y-1>=  0) horizontal_compose97i(b0, width);
        if(b0 <= b2) horizontal_compose97i(b1, width);
if(width>400 && b0 <= b2){
STOP_TIMER("horizontal_compose97i")}}

    cs->b0=b2;
    cs->b1=b3;
    cs->b2=b4;
    cs->b3=b5;
    cs->y += 2;
}

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_init(dwt_compose_t *cs, DWTELEM *buffer, int width, int height, int stride, int type, int decomposition_count){
    int level;
    for(level=decomposition_count-1; level>=0; level--){
        switch(type){
        case 0: spatial_compose97i_init(cs+level, buffer, height>>level, stride<<level); break;
        case 1: spatial_compose53i_init(cs+level, buffer, height>>level, stride<<level); break;
        /* not slicified yet */
        case 2: spatial_composeX(buffer, width>>level, height>>level, stride<<level); break;
        }
    }
}

void ff_spatial_idwt_slice(dwt_compose_t *cs, DWTELEM *buffer, int width, int height, int stride, int type, int decomposition_count, int y){
    const int support = type==1 ? 3 : 5;
    int level;
    if(type==2) return;

    for(level=decomposition_count-1; level>=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);
                    break;
            case 1: spatial_compose53i_dy(cs+level, buffer, width>>level, height>>level, stride<<level);
                    break;
            case 2: break;
            }
        }
    }
}

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);
    }else{
        dwt_compose_t cs[MAX_DECOMPOSITIONS];
        int y;
        ff_spatial_idwt_init(cs, buffer, width, height, stride, type, decomposition_count);
        for(y=0; y<height; y+=4)
            ff_spatial_idwt_slice(cs, buffer, width, height, stride, type, decomposition_count, y);
    }
}

static int encode_subband_c0run(SnowContext *s, SubBand *b, DWTELEM *src, DWTELEM *parent, int stride, int orientation){
    const int w= b->width;

        }
    }
#endif
}

static always_inline void predict_slice(SnowContext *s, DWTELEM *buf, int plane_index, int add, int mb_y){
    Plane *p= &s->plane[plane_index];
    const int mb_w= s->b_width  << s->block_max_depth;
    const int mb_h= s->b_height << s->block_max_depth;
    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];

    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<block_w*(mb_y+1); y++){
                for(x=0; x<w; x++){
                    int v= buf[x + y*w] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
                    v >>= FRAC_BITS;
                    if(v&(~255)) v= ~(v>>31);
                    dst8[x + y*ref_stride]= v;
                }
            }
        }else{
            for(y=block_w*mb_y; y<block_w*(mb_y+1); y++){
                for(x=0; x<w; x++){
                    buf[x + y*w]-= 128<<FRAC_BITS;
                }
            }
        }

        return;
    }
    
        for(mb_x=0; mb_x<=mb_w; mb_x++){
            START_TIMER

            add_yblock(s, buf, dst8, ref, obmc, 
                       block_w*mb_x - block_w/2,

                       mb_x - 1, mb_y - 1,
                       add, plane_index);
            
            STOP_TIMER("add_yblock")
        }
    
    STOP_TIMER("predict_slice")
}

static always_inline void predict_plane(SnowContext *s, DWTELEM *buf, int plane_index, int add){
    const int mb_h= s->b_height << s->block_max_depth;
    int mb_y;
    for(mb_y=0; mb_y<=mb_h; mb_y++)
        predict_slice(s, buf, plane_index, add, mb_y);
}

static void quantize(SnowContext *s, SubBand *b, DWTELEM *src, int stride, int bias){
    const int level= b->level;
    const int w= b->width;

                    assert(b->buf == s->spatial_dwt_buffer);
                }
            }
        }

{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;

    ff_spatial_idwt_init(cs, s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
    for(mb_y=0; mb_y<=mb_h; mb_y++){
        const int slice_h = block_w*(mb_y+1);
        for(; yd<slice_h; yd+=4)
            ff_spatial_idwt_slice(cs, s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count, yd);

        if(s->qlog == LOSSLESS_QLOG){
            for(; yq<slice_h && yq<h; yq++){
                for(x=0; x<w; x++){
                    s->spatial_dwt_buffer[yq*w + x]<<=FRAC_BITS;
                }
            }
        }

        predict_slice(s, s->spatial_dwt_buffer, plane_index, 1, mb_y);
    }
STOP_TIMER("idwt + predict_slices")}
    }
            
    emms_c();

    if(s->last_picture.data[0])