comparison snow.c @ 6412:1f70d64c00bc libavcodec

cosmetics: spelling/grammar

6411:5d50e8a7eb09

typedef struct SubBand{
    int level;
    int stride;
    int width;
    int height;
    int qlog;                                   ///< log(qscale)/log[2^(1/6)]
    DWTELEM *buf;
    IDWTELEM *ibuf;
    int buf_x_offset;
    int buf_y_offset;
    int stride_line; ///< Stride measured in lines, not pixels.

    int8_t last_hcoeff[HTAPS_MAX/2];
    int last_diag_mc;
}Plane;

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

    AVCodecContext *avctx;
    RangeCoder c;
    DSPContext dsp;
    AVFrame new_picture;

#define ME_CACHE_SIZE 1024
    int me_cache[ME_CACHE_SIZE];
    int me_cache_generation;
    slice_buffer sb;

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

typedef struct {
    IDWTELEM *b0;
    IDWTELEM *b1;

            }
        }
    }
}

//FIXME name clenup (b_w, block_w, b_width stuff)
static av_always_inline void add_yblock(SnowContext *s, int sliced, slice_buffer *sb, IDWTELEM *dst, uint8_t *dst8, const uint8_t *obmc, int src_x, int src_y, int b_w, int b_h, int w, int h, int dst_stride, int src_stride, int obmc_stride, int b_x, int b_y, int add, int offset_dst, int plane_index){
    const int b_width = s->b_width  << s->block_max_depth;
    const int b_height= s->b_height << s->block_max_depth;
    const int b_stride= b_width;
    BlockNode *lt= &s->block[b_x + b_y*b_stride];

    }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;
        if(!sliced && !offset_dst)
            dst -= src_x;
        src_x=0;

                int d;
                if(y<0) obmc_v += obmc[index + block_w*obmc_stride];
                if(x<0) obmc_v += obmc[index + block_w];
                if(y+block_w>h) 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

                d = -dst[index] + (1<<(FRAC_BITS-1));
                dst[index] = d;
                ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
                aa += obmc_v * obmc_v; //FIXME precalclate this
            }
        }
    }
    *b= backup;

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){
        return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
                   + av_log2(2*FFABS(left->color[1] - b->color[1]))
                   + av_log2(2*FFABS(left->color[2] - b->color[2])));
    }else{

    }

    if(block_w==16){
        /* FIXME rearrange dsputil to fit 32x32 cmp functions */
        /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
        /* FIXME cmps overlap but don't cover the wavelet's whole support,
         * so improving the score of one block is not strictly guaranteed to
         * improve the score of the whole frame, so iterative motion est
         * doesn't always converge. */
        if(s->avctx->me_cmp == FF_CMP_W97)
            distortion = w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
        else if(s->avctx->me_cmp == FF_CMP_W53)
            distortion = w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
        else{

        *block= backup;
        return 0;
    }
}

/* special case for int[2] args we discard afterward, fixes compilation prob with gcc 2.95 */
static av_always_inline int check_block_inter(SnowContext *s, int mb_x, int mb_y, int p0, int p1, const uint8_t *obmc_edged, int *best_rd){
    int p[2] = {p0, p1};
    return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
}

                if(!s->me_cache_generation)
                    memset(s->me_cache, 0, sizeof(s->me_cache));
                s->me_cache_generation += 1<<22;

                //FIXME precalc
                {
                    int x, y;
                    memcpy(obmc_edged, obmc_tab[s->block_max_depth], b_w*b_w*4);
                    if(mb_x==0)
                        for(y=0; y<b_w*2; y++)

                        check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], *obmc_edged, &best_rd);
                    if(bb)
                        check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], *obmc_edged, &best_rd);

                    /* fullpel ME */
                    //FIXME avoid subpel interpol / round to nearest integer
                    do{
                        dia_change=0;
                        for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
                            for(j=0; j<i; j++){
                                dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), *obmc_edged, &best_rd);

    width= s->avctx->width;
    height= s->avctx->height;

    s->spatial_idwt_buffer= av_mallocz(width*height*sizeof(IDWTELEM));
    s->spatial_dwt_buffer= av_mallocz(width*height*sizeof(DWTELEM)); //FIXME this doesnt belong here

    for(i=0; i<MAX_REF_FRAMES; i++)
        for(j=0; j<MAX_REF_FRAMES; j++)
            scale_mv_ref[i][j] = 256*(i+1)/(j+1);

           + 61*QROOT/8; //<64 >60
}

static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
{
    /* estimate the frame's complexity as a sum of weighted dwt coefs.
     * FIXME we know exact mv bits at this point,
     * but ratecontrol isn't set up to include them. */
    uint32_t coef_sum= 0;
    int level, orientation, delta_qlog;

    if(orientation&1)
        xo= step>>1;
    if(orientation&2)
        yo= step>>1;

    //FIXME bias for non zero ?
    //FIXME optimize
    memset(score, 0, sizeof(*score)*score_stride*((p->height + Q2_STEP-1)/Q2_STEP));
    for(y=0; y<p->height; y++){
        for(x=0; x<p->width; x++){
            int sx= (x-xo + step/2) / step / Q2_STEP;

    for(level=0; level<s->spatial_decomposition_count; level++){
        for(orientation=level ? 1 : 0; orientation<4; orientation++){
            SubBand *b= &p->band[level][orientation];
            IDWTELEM *dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
             DWTELEM *src=       buffer + (b-> buf - s->spatial_dwt_buffer);
            assert(src == b->buf); // code doesnt depen on this but its true currently

            quantize(s, b, dst, src, b->stride, s->qbias);
        }
    }
    for(pass=0; pass<1; pass++){

                        memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
                        for(y=ys; y<b->height; y+= Q2_STEP){
                            for(x=xs; x<b->width; x+= Q2_STEP){
                                if(dst[x + y*b->stride]<0) dst[x + y*b->stride]++;
                                if(dst[x + y*b->stride]>0) dst[x + y*b->stride]--;
                                //FIXME try more then just --
                            }
                        }
                        dequantize_all(s, p, idwt2_buffer, width, height);
                        ff_spatial_idwt(idwt2_buffer, width, height, stride, type, s->spatial_decomposition_count);
                        find_sse(s, p, score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);

                    }
                }
            }
        }
    }
    memcpy(s->spatial_idwt_buffer, best_dequant, height * stride * sizeof(IDWTELEM)); //FIXME work with that directly insteda of copy at the end
}

#endif /* QUANTIZE2==1 */

static int encode_init(AVCodecContext *avctx)
{
    SnowContext *s = avctx->priv_data;
    int plane_index;

    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;
    }

    if(avctx->prediction_method == DWT_97
       && (avctx->flags & CODEC_FLAG_QSCALE)
       && avctx->global_quality == 0){
        av_log(avctx, AV_LOG_ERROR, "the 9/7 wavelet is incompatible with lossless mode\n");
        return -1;
    }

    s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type

        s->lambda = pict->quality * 3/2;
    }
    if(s->qlog < 0 || (!pict->quality && (avctx->flags & CODEC_FLAG_QSCALE))){
        s->qlog= LOSSLESS_QLOG;
        s->lambda = 0;
    }//else keep previous frame's qlog until after motion est

    frame_start(s);

    s->m.current_picture_ptr= &s->m.current_picture;
    if(pict->pict_type == P_TYPE){

    }

    if(!s->block) alloc_blocks(s);

    frame_start(s);
    //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);

6412:1f70d64c00bc

typedef struct SubBand{
    int level;
    int stride;
    int width;
    int height;
    int qlog;        ///< log(qscale)/log[2^(1/6)]
    DWTELEM *buf;
    IDWTELEM *ibuf;
    int buf_x_offset;
    int buf_y_offset;
    int stride_line; ///< Stride measured in lines, not pixels.

    int8_t last_hcoeff[HTAPS_MAX/2];
    int last_diag_mc;
}Plane;

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

    AVCodecContext *avctx;
    RangeCoder c;
    DSPContext dsp;
    AVFrame new_picture;

#define ME_CACHE_SIZE 1024
    int me_cache[ME_CACHE_SIZE];
    int me_cache_generation;
    slice_buffer sb;

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

typedef struct {
    IDWTELEM *b0;
    IDWTELEM *b1;

            }
        }
    }
}

//FIXME name cleanup (b_w, block_w, b_width stuff)
static av_always_inline void add_yblock(SnowContext *s, int sliced, slice_buffer *sb, IDWTELEM *dst, uint8_t *dst8, const uint8_t *obmc, int src_x, int src_y, int b_w, int b_h, int w, int h, int dst_stride, int src_stride, int obmc_stride, int b_x, int b_y, int add, int offset_dst, int plane_index){
    const int b_width = s->b_width  << s->block_max_depth;
    const int b_height= s->b_height << s->block_max_depth;
    const int b_stride= b_width;
    BlockNode *lt= &s->block[b_x + b_y*b_stride];

    }else if(b_y + 1 >= b_height){
        lb= lt;
        rb= rt;
    }

    if(src_x<0){ //FIXME merge with prev & always round internal width up to *16
        obmc -= src_x;
        b_w += src_x;
        if(!sliced && !offset_dst)
            dst -= src_x;
        src_x=0;

                int d;
                if(y<0) obmc_v += obmc[index + block_w*obmc_stride];
                if(x<0) obmc_v += obmc[index + block_w];
                if(y+block_w>h) obmc_v += obmc[index - block_w*obmc_stride];
                if(x+block_w>w) obmc_v += obmc[index - block_w];
                //FIXME precalculate this or simplify it somehow else

                d = -dst[index] + (1<<(FRAC_BITS-1));
                dst[index] = d;
                ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
                aa += obmc_v * obmc_v; //FIXME precalculate this
            }
        }
    }
    *b= backup;

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 are not the same type
    if(b->type & BLOCK_INTRA){
        return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
                   + av_log2(2*FFABS(left->color[1] - b->color[1]))
                   + av_log2(2*FFABS(left->color[2] - b->color[2])));
    }else{

    }

    if(block_w==16){
        /* FIXME rearrange dsputil to fit 32x32 cmp functions */
        /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
        /* FIXME cmps overlap but do not cover the wavelet's whole support.
         * So improving the score of one block is not strictly guaranteed
         * to improve the score of the whole frame, thus iterative motion
         * estimation does not always converge. */
        if(s->avctx->me_cmp == FF_CMP_W97)
            distortion = w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
        else if(s->avctx->me_cmp == FF_CMP_W53)
            distortion = w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
        else{

        *block= backup;
        return 0;
    }
}

/* special case for int[2] args we discard afterwards,
 * fixes compilation problem with gcc 2.95 */
static av_always_inline int check_block_inter(SnowContext *s, int mb_x, int mb_y, int p0, int p1, const uint8_t *obmc_edged, int *best_rd){
    int p[2] = {p0, p1};
    return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
}

                if(!s->me_cache_generation)
                    memset(s->me_cache, 0, sizeof(s->me_cache));
                s->me_cache_generation += 1<<22;

                //FIXME precalculate
                {
                    int x, y;
                    memcpy(obmc_edged, obmc_tab[s->block_max_depth], b_w*b_w*4);
                    if(mb_x==0)
                        for(y=0; y<b_w*2; y++)

                        check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], *obmc_edged, &best_rd);
                    if(bb)
                        check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], *obmc_edged, &best_rd);

                    /* fullpel ME */
                    //FIXME avoid subpel interpolation / round to nearest integer
                    do{
                        dia_change=0;
                        for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
                            for(j=0; j<i; j++){
                                dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), *obmc_edged, &best_rd);

    width= s->avctx->width;
    height= s->avctx->height;

    s->spatial_idwt_buffer= av_mallocz(width*height*sizeof(IDWTELEM));
    s->spatial_dwt_buffer= av_mallocz(width*height*sizeof(DWTELEM)); //FIXME this does not belong here

    for(i=0; i<MAX_REF_FRAMES; i++)
        for(j=0; j<MAX_REF_FRAMES; j++)
            scale_mv_ref[i][j] = 256*(i+1)/(j+1);

           + 61*QROOT/8; //<64 >60
}

static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
{
    /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
     * FIXME we know exact mv bits at this point,
     * but ratecontrol isn't set up to include them. */
    uint32_t coef_sum= 0;
    int level, orientation, delta_qlog;

    if(orientation&1)
        xo= step>>1;
    if(orientation&2)
        yo= step>>1;

    //FIXME bias for nonzero ?
    //FIXME optimize
    memset(score, 0, sizeof(*score)*score_stride*((p->height + Q2_STEP-1)/Q2_STEP));
    for(y=0; y<p->height; y++){
        for(x=0; x<p->width; x++){
            int sx= (x-xo + step/2) / step / Q2_STEP;

    for(level=0; level<s->spatial_decomposition_count; level++){
        for(orientation=level ? 1 : 0; orientation<4; orientation++){
            SubBand *b= &p->band[level][orientation];
            IDWTELEM *dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
             DWTELEM *src=       buffer + (b-> buf - s->spatial_dwt_buffer);
            assert(src == b->buf); // code does not depend on this but it is true currently

            quantize(s, b, dst, src, b->stride, s->qbias);
        }
    }
    for(pass=0; pass<1; pass++){

                        memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
                        for(y=ys; y<b->height; y+= Q2_STEP){
                            for(x=xs; x<b->width; x+= Q2_STEP){
                                if(dst[x + y*b->stride]<0) dst[x + y*b->stride]++;
                                if(dst[x + y*b->stride]>0) dst[x + y*b->stride]--;
                                //FIXME try more than just --
                            }
                        }
                        dequantize_all(s, p, idwt2_buffer, width, height);
                        ff_spatial_idwt(idwt2_buffer, width, height, stride, type, s->spatial_decomposition_count);
                        find_sse(s, p, score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);

                    }
                }
            }
        }
    }
    memcpy(s->spatial_idwt_buffer, best_dequant, height * stride * sizeof(IDWTELEM)); //FIXME work with that directly instead of copy at the end
}

#endif /* QUANTIZE2==1 */

static int encode_init(AVCodecContext *avctx)
{
    SnowContext *s = avctx->priv_data;
    int plane_index;

    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;
    }

    if(avctx->prediction_method == DWT_97
       && (avctx->flags & CODEC_FLAG_QSCALE)
       && avctx->global_quality == 0){
        av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
        return -1;
    }

    s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type

        s->lambda = pict->quality * 3/2;
    }
    if(s->qlog < 0 || (!pict->quality && (avctx->flags & CODEC_FLAG_QSCALE))){
        s->qlog= LOSSLESS_QLOG;
        s->lambda = 0;
    }//else keep previous frame's qlog until after motion estimation

    frame_start(s);

    s->m.current_picture_ptr= &s->m.current_picture;
    if(pict->pict_type == P_TYPE){

    }

    if(!s->block) alloc_blocks(s);

    frame_start(s);
    //keyframe flag duplication 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);