comparison h264.c @ 4587:644687c58928 libavcodec

h264 optimization: common case hl_decode_mb patch by (Alexander Strange astrange ithinksw com) 144095->142319 dezicycles for hl_decode_mb() on duron trailing whitespace removed by me

4586:365f2fed8461

            XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+24), *(uint64_t*)(src_cr+1 +uvlinesize), temp64, 1);
        }
    }
}

static void hl_decode_mb(H264Context *h){
    MpegEncContext * const s = &h->s;
    const int mb_x= s->mb_x;
    const int mb_y= s->mb_y;
    const int mb_xy= mb_x + mb_y*s->mb_stride;
    const int mb_type= s->current_picture.mb_type[mb_xy];
    uint8_t  *dest_y, *dest_cb, *dest_cr;
    int linesize, uvlinesize /*dct_offset*/;
    int i;
    int *block_offset = &h->block_offset[0];
    const unsigned int bottom = mb_y & 1;
    const int transform_bypass = (s->qscale == 0 && h->sps.transform_bypass);
    void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
    void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);

    if(!s->decode)
        return;

    dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
    dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
    dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;

    s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
    s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);

    if (MB_FIELD) {
        linesize   = h->mb_linesize   = s->linesize * 2;
        uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
        block_offset = &h->block_offset[24];
        if(mb_y&1){ //FIXME move out of this func?
            dest_y -= s->linesize*15;

    }else{
        idct_dc_add = s->dsp.h264_idct_dc_add;
        idct_add = s->dsp.h264_idct_add;
    }

    if(FRAME_MBAFF && h->deblocking_filter && IS_INTRA(mb_type)
       && (!bottom || !IS_INTRA(s->current_picture.mb_type[mb_xy-s->mb_stride]))){
        int mbt_y = mb_y&~1;
        uint8_t *top_y  = s->current_picture.data[0] + (mbt_y * 16* s->linesize  ) + mb_x * 16;
        uint8_t *top_cb = s->current_picture.data[1] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
        uint8_t *top_cr = s->current_picture.data[2] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
        xchg_pair_border(h, top_y, top_cb, top_cr, s->linesize, s->uvlinesize, 1);
    }

    if (IS_INTRA_PCM(mb_type)) {
        unsigned int x, y;

        // The pixels are stored in h->mb array in the same order as levels,
        // copy them in output in the correct order.
        for(i=0; i<16; i++) {

                }
            }
        }
    } else {
        if(IS_INTRA(mb_type)){
            if(h->deblocking_filter && !FRAME_MBAFF)
                xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1);

            if(!(s->flags&CODEC_FLAG_GRAY)){
                h->pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
                h->pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
            }

            if(IS_INTRA4x4(mb_type)){
                if(!s->encoding){
                    if(IS_8x8DCT(mb_type)){
                        for(i=0; i<16; i+=4){
                            uint8_t * const ptr= dest_y + block_offset[i];
                            const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
                            const int nnz = h->non_zero_count_cache[ scan8[i] ];

                            topright= NULL;

                        h->pred4x4[ dir ](ptr, topright, linesize);
                        nnz = h->non_zero_count_cache[ scan8[i] ];
                        if(nnz){
                            if(s->codec_id == CODEC_ID_H264){
                                if(nnz == 1 && h->mb[i*16])
                                    idct_dc_add(ptr, h->mb + i*16, linesize);
                                else
                                    idct_add(ptr, h->mb + i*16, linesize);
                            }else

                        }
                    }
                }
            }else{
                h->pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
                if(s->codec_id == CODEC_ID_H264){
                    if(!transform_bypass)
                        h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[IS_INTRA(mb_type) ? 0:3][s->qscale][0]);
                }else
                    svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
            }
            if(h->deblocking_filter && !FRAME_MBAFF)
                xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
        }else if(s->codec_id == CODEC_ID_H264){
            hl_motion(h, dest_y, dest_cb, dest_cr,
                      s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
                      s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
                      s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
        }


        if(!IS_INTRA4x4(mb_type)){
            if(s->codec_id == CODEC_ID_H264){
                if(IS_INTRA16x16(mb_type)){
                    for(i=0; i<16; i++){
                        if(h->non_zero_count_cache[ scan8[i] ])
                            idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
                        else if(h->mb[i*16])

                    }
                }
            }
        }

        if(!(s->flags&CODEC_FLAG_GRAY)){
            uint8_t *dest[2] = {dest_cb, dest_cr};
            if(transform_bypass){
                idct_add = idct_dc_add = s->dsp.add_pixels4;
            }else{
                idct_add = s->dsp.h264_idct_add;
                idct_dc_add = s->dsp.h264_idct_dc_add;
                chroma_dc_dequant_idct_c(h->mb + 16*16, h->chroma_qp, h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp][0]);
                chroma_dc_dequant_idct_c(h->mb + 16*16+4*16, h->chroma_qp, h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp][0]);
            }
            if(s->codec_id == CODEC_ID_H264){
                for(i=16; i<16+8; i++){
                    if(h->non_zero_count_cache[ scan8[i] ])
                        idct_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
                    else if(h->mb[i*16])
                        idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);

                }
            }
        }
    }
    if(h->deblocking_filter) {
        if (FRAME_MBAFF) {
            //FIXME try deblocking one mb at a time?
            // the reduction in load/storing mvs and such might outweigh the extra backup/xchg_border
            const int mb_y = s->mb_y - 1;
            uint8_t  *pair_dest_y, *pair_dest_cb, *pair_dest_cr;
            const int mb_xy= mb_x + mb_y*s->mb_stride;

            backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
            fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
            filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
        }
    }
}

/**
 * fills the default_ref_list.
 */

4587:644687c58928

            XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+24), *(uint64_t*)(src_cr+1 +uvlinesize), temp64, 1);
        }
    }
}

static void av_always_inline hl_decode_mb_internal(H264Context *h, int simple){
    MpegEncContext * const s = &h->s;
    const int mb_x= s->mb_x;
    const int mb_y= s->mb_y;
    const int mb_xy= mb_x + mb_y*s->mb_stride;
    const int mb_type= s->current_picture.mb_type[mb_xy];
    uint8_t  *dest_y, *dest_cb, *dest_cr;
    int linesize, uvlinesize /*dct_offset*/;
    int i;
    int *block_offset = &h->block_offset[0];
    const unsigned int bottom = mb_y & 1;
    const int transform_bypass = (s->qscale == 0 && h->sps.transform_bypass), is_h264 = (simple || s->codec_id == CODEC_ID_H264);
    void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
    void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);

    dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
    dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
    dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;

    s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
    s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);

    if (!simple && MB_FIELD) {
        linesize   = h->mb_linesize   = s->linesize * 2;
        uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
        block_offset = &h->block_offset[24];
        if(mb_y&1){ //FIXME move out of this func?
            dest_y -= s->linesize*15;

    }else{
        idct_dc_add = s->dsp.h264_idct_dc_add;
        idct_add = s->dsp.h264_idct_add;
    }

    if(!simple && FRAME_MBAFF && h->deblocking_filter && IS_INTRA(mb_type)
       && (!bottom || !IS_INTRA(s->current_picture.mb_type[mb_xy-s->mb_stride]))){
        int mbt_y = mb_y&~1;
        uint8_t *top_y  = s->current_picture.data[0] + (mbt_y * 16* s->linesize  ) + mb_x * 16;
        uint8_t *top_cb = s->current_picture.data[1] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
        uint8_t *top_cr = s->current_picture.data[2] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
        xchg_pair_border(h, top_y, top_cb, top_cr, s->linesize, s->uvlinesize, 1);
    }

    if (!simple && IS_INTRA_PCM(mb_type)) {
        unsigned int x, y;

        // The pixels are stored in h->mb array in the same order as levels,
        // copy them in output in the correct order.
        for(i=0; i<16; i++) {

                }
            }
        }
    } else {
        if(IS_INTRA(mb_type)){
            if(h->deblocking_filter && (simple || !FRAME_MBAFF))
                xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1);

            if(simple || !(s->flags&CODEC_FLAG_GRAY)){
                h->pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
                h->pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
            }

            if(IS_INTRA4x4(mb_type)){
                if(simple || !s->encoding){
                    if(IS_8x8DCT(mb_type)){
                        for(i=0; i<16; i+=4){
                            uint8_t * const ptr= dest_y + block_offset[i];
                            const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
                            const int nnz = h->non_zero_count_cache[ scan8[i] ];

                            topright= NULL;

                        h->pred4x4[ dir ](ptr, topright, linesize);
                        nnz = h->non_zero_count_cache[ scan8[i] ];
                        if(nnz){
                            if(is_h264){
                                if(nnz == 1 && h->mb[i*16])
                                    idct_dc_add(ptr, h->mb + i*16, linesize);
                                else
                                    idct_add(ptr, h->mb + i*16, linesize);
                            }else

                        }
                    }
                }
            }else{
                h->pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
                if(is_h264){
                    if(!transform_bypass)
                        h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[IS_INTRA(mb_type) ? 0:3][s->qscale][0]);
                }else
                    svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
            }
            if(h->deblocking_filter && (simple || !FRAME_MBAFF))
                xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
        }else if(is_h264){
            hl_motion(h, dest_y, dest_cb, dest_cr,
                      s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
                      s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
                      s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
        }


        if(!IS_INTRA4x4(mb_type)){
            if(is_h264){
                if(IS_INTRA16x16(mb_type)){
                    for(i=0; i<16; i++){
                        if(h->non_zero_count_cache[ scan8[i] ])
                            idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
                        else if(h->mb[i*16])

                    }
                }
            }
        }

        if(simple || !(s->flags&CODEC_FLAG_GRAY)){
            uint8_t *dest[2] = {dest_cb, dest_cr};
            if(transform_bypass){
                idct_add = idct_dc_add = s->dsp.add_pixels4;
            }else{
                idct_add = s->dsp.h264_idct_add;
                idct_dc_add = s->dsp.h264_idct_dc_add;
                chroma_dc_dequant_idct_c(h->mb + 16*16, h->chroma_qp, h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp][0]);
                chroma_dc_dequant_idct_c(h->mb + 16*16+4*16, h->chroma_qp, h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp][0]);
            }
            if(is_h264){
                for(i=16; i<16+8; i++){
                    if(h->non_zero_count_cache[ scan8[i] ])
                        idct_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
                    else if(h->mb[i*16])
                        idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);

                }
            }
        }
    }
    if(h->deblocking_filter) {
        if (!simple && FRAME_MBAFF) {
            //FIXME try deblocking one mb at a time?
            // the reduction in load/storing mvs and such might outweigh the extra backup/xchg_border
            const int mb_y = s->mb_y - 1;
            uint8_t  *pair_dest_y, *pair_dest_cb, *pair_dest_cr;
            const int mb_xy= mb_x + mb_y*s->mb_stride;

            backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
            fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
            filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
        }
    }
}

/**
 * Process a macroblock; this case avoids checks for expensive uncommon cases.
 */
static void hl_decode_mb_simple(H264Context *h){
    hl_decode_mb_internal(h, 1);
}

/**
 * Process a macroblock; this handles edge cases, such as interlacing.
 */
static void av_noinline hl_decode_mb_complex(H264Context *h){
    hl_decode_mb_internal(h, 0);
}

static void hl_decode_mb(H264Context *h){
    MpegEncContext * const s = &h->s;
    const int mb_x= s->mb_x;
    const int mb_y= s->mb_y;
    const int mb_xy= mb_x + mb_y*s->mb_stride;
    const int mb_type= s->current_picture.mb_type[mb_xy];
    int is_complex = FRAME_MBAFF || MB_FIELD || IS_INTRA_PCM(mb_type) || s->codec_id != CODEC_ID_H264 || (s->flags&CODEC_FLAG_GRAY) || s->encoding;

    if(!s->decode)
        return;

    if (is_complex)
        hl_decode_mb_complex(h);
    else hl_decode_mb_simple(h);
}

/**
 * fills the default_ref_list.
 */