comparison mpeg12.c @ 1580:628bf341e099 libavcodec

XvMC speedup by removing one memcpy and doing MB packing

1579:039cb2296de2

static int mpeg_decode_motion(MpegEncContext *s, int fcode, int pred);

#ifdef HAVE_XVMC
extern int XVMC_field_start(MpegEncContext *s, AVCodecContext *avctx);
extern int XVMC_field_end(MpegEncContext *s);
#endif

#ifdef CONFIG_ENCODERS
static uint8_t (*mv_penalty)[MAX_MV*2+1]= NULL;
static uint8_t fcode_tab[MAX_MV*2+1];

            skip_bits1(&s->gb); /* marker */
        }else
            memset(s->last_mv, 0, sizeof(s->last_mv)); /* reset mv prediction */
        s->mb_intra = 1;

        if (s->codec_id == CODEC_ID_MPEG2VIDEO) {
            for(i=0;i<6;i++) {
                if (mpeg2_decode_block_intra(s, block[i], i) < 0)
                    return -1;
            }
        } else {
            for(i=0;i<6;i++) {
                if (mpeg1_decode_block_intra(s, block[i], i) < 0)
                    return -1;
            }
        }
    } else {
        if (mb_type & MB_TYPE_ZERO_MV){

                fprintf(stderr, "invalid cbp at %d %d\n", s->mb_x, s->mb_y);
                return -1;
            }
            cbp++;

            if (s->codec_id == CODEC_ID_MPEG2VIDEO) {
                for(i=0;i<6;i++) {
                    if (cbp & 32) {
                        if (mpeg2_decode_block_non_intra(s, block[i], i) < 0)
                            return -1;
                    } else {
                        s->block_last_index[i] = -1;
                    }
                    cbp+=cbp;
                }
            } else {
                for(i=0;i<6;i++) {
                    if (cbp & 32) {
                        if (mpeg1_decode_block_inter(s, block[i], i) < 0)
                            return -1;
                    } else {
                        s->block_last_index[i] = -1;
                    }
                    cbp+=cbp;

        mpeg_decode_picture_coding_extension(s);
        break;
    }
}

static void exchange_uv(AVFrame *f){
    uint8_t *t= f->data[1];
    f->data[1]= f->data[2];
    f->data[2]= t;
}

#define DECODE_SLICE_FATAL_ERROR -2
#define DECODE_SLICE_ERROR -1
#define DECODE_SLICE_OK 0

    s->resync_mb_y= s->mb_y = start_code;
    s->mb_skip_run= 0;
    ff_init_block_index(s);

    for(;;) {
	s->dsp.clear_blocks(s->block[0]);

        ret = mpeg_decode_mb(s, s->block);

        dprintf("ret=%d\n", ret);

        s->dest[2] += 8;

        MPV_decode_mb(s, s->block);
        
        if (++s->mb_x >= s->mb_width) {
            if(s->avctx->codec_tag == ff_get_fourcc("VCR2"))
                exchange_uv((AVFrame*)s->current_picture_ptr);

            ff_draw_horiz_band(s, 16*s->mb_y, 16);

            if(s->avctx->codec_tag == ff_get_fourcc("VCR2"))
                exchange_uv((AVFrame*)s->current_picture_ptr);

            s->mb_x = 0;
            s->mb_y++;

            if(s->mb_y<<field_pic >= s->mb_height){

            if (s->last_picture_ptr != NULL) {
                *pict= *(AVFrame*)s->last_picture_ptr;
                 ff_print_debug_info(s, s->last_picture_ptr);
            }
        }
        if(s->avctx->codec_tag == ff_get_fourcc("VCR2"))
            exchange_uv(pict);

        return 1;
    } else {
        return 0;
    }

        avctx->bit_rate = s->bit_rate;
        
        //get_format() or set_video(width,height,aspect,pix_fmt);
        //until then pix_fmt may be changed right after codec init
        if( avctx->pix_fmt == PIX_FMT_XVMC_MPEG2_IDCT )
            avctx->idct_algo = FF_IDCT_SIMPLE;

        if (MPV_common_init(s) < 0)
            return -1;
        s1->mpeg_enc_ctx_allocated = 1;
    }

    skip_bits(&s->gb, 10); /* vbv_buffer_size */
    skip_bits(&s->gb, 1);

    s->avctx = avctx;

    //get_format() or set_video(width,height,aspect,pix_fmt);
    //until then pix_fmt may be changed right after codec init
    if( avctx->pix_fmt == PIX_FMT_XVMC_MPEG2_IDCT )
        avctx->idct_algo = FF_IDCT_SIMPLE;
    
    if (MPV_common_init(s) < 0)
        return -1;
    s1->mpeg_enc_ctx_allocated = 1;

    for(i=0;i<64;i++) {
        int j= s->dsp.idct_permutation[i];
        v = ff_mpeg1_default_intra_matrix[i];

static int mpeg_mc_decode_init(AVCodecContext *avctx){
    Mpeg1Context *s;

    if( !(avctx->slice_flags & SLICE_FLAG_CODED_ORDER) )
        return -1;
    if( !(avctx->slice_flags & SLICE_FLAG_ALLOW_FIELD) )
        dprintf("mpeg12.c: XvMC decoder will work better if SLICE_FLAG_ALLOW_FIELD is set\n");

    mpeg_decode_init(avctx);
    s = avctx->priv_data;

    avctx->pix_fmt = PIX_FMT_XVMC_MPEG2_IDCT;
    avctx->xvmc_acceleration = 1;

    return 0;
}

AVCodec mpeg_xvmc_decoder = {

1580:628bf341e099

static int mpeg_decode_motion(MpegEncContext *s, int fcode, int pred);

#ifdef HAVE_XVMC
extern int XVMC_field_start(MpegEncContext *s, AVCodecContext *avctx);
extern int XVMC_field_end(MpegEncContext *s);
extern void XVMC_pack_pblocks(MpegEncContext *s,int cbp);
extern void XVMC_init_block(s);//set s->block
#endif

#ifdef CONFIG_ENCODERS
static uint8_t (*mv_penalty)[MAX_MV*2+1]= NULL;
static uint8_t fcode_tab[MAX_MV*2+1];

            skip_bits1(&s->gb); /* marker */
        }else
            memset(s->last_mv, 0, sizeof(s->last_mv)); /* reset mv prediction */
        s->mb_intra = 1;
#ifdef HAVE_XVMC
        //one 1 we memcpy blocks in xvmcvideo
        if(s->avctx->xvmc_acceleration > 1){
            XVMC_pack_pblocks(s,-1);//inter are always full blocks
            if(s->swap_uv){
                exchange_uv(s);
            }
        }
#endif

        if (s->codec_id == CODEC_ID_MPEG2VIDEO) {
            for(i=0;i<6;i++) {
                if (mpeg2_decode_block_intra(s, s->pblocks[i], i) < 0)
                    return -1;
            }
        } else {
            for(i=0;i<6;i++) {
                if (mpeg1_decode_block_intra(s, s->pblocks[i], i) < 0)
                    return -1;
            }
        }
    } else {
        if (mb_type & MB_TYPE_ZERO_MV){

                fprintf(stderr, "invalid cbp at %d %d\n", s->mb_x, s->mb_y);
                return -1;
            }
            cbp++;

#ifdef HAVE_XVMC
            //on 1 we memcpy blocks in xvmcvideo
            if(s->avctx->xvmc_acceleration > 1){
                XVMC_pack_pblocks(s,cbp);
                if(s->swap_uv){
                    exchange_uv(s);
                }
            }    
#endif

            if (s->codec_id == CODEC_ID_MPEG2VIDEO) {
                for(i=0;i<6;i++) {
                    if (cbp & 32) {
                        if (mpeg2_decode_block_non_intra(s, s->pblocks[i], i) < 0)
                            return -1;
                    } else {
                        s->block_last_index[i] = -1;
                    }
                    cbp+=cbp;
                }
            } else {
                for(i=0;i<6;i++) {
                    if (cbp & 32) {
                        if (mpeg1_decode_block_inter(s, s->pblocks[i], i) < 0)
                            return -1;
                    } else {
                        s->block_last_index[i] = -1;
                    }
                    cbp+=cbp;

        mpeg_decode_picture_coding_extension(s);
        break;
    }
}

static void exchange_uv(MpegEncContext *s){
short * tmp;

    tmp = s->pblocks[4];
    s->pblocks[4] = s->pblocks[5];
    s->pblocks[5] = tmp;
}

#define DECODE_SLICE_FATAL_ERROR -2
#define DECODE_SLICE_ERROR -1
#define DECODE_SLICE_OK 0

    s->resync_mb_y= s->mb_y = start_code;
    s->mb_skip_run= 0;
    ff_init_block_index(s);

    for(;;) {
#ifdef HAVE_XVMC
        //one 1 we memcpy blocks in xvmcvideo
        if(s->avctx->xvmc_acceleration > 1)
            XVMC_init_block(s);//set s->block
#endif

	s->dsp.clear_blocks(s->block[0]);

        ret = mpeg_decode_mb(s, s->block);

        dprintf("ret=%d\n", ret);

        s->dest[2] += 8;

        MPV_decode_mb(s, s->block);
        
        if (++s->mb_x >= s->mb_width) {

            ff_draw_horiz_band(s, 16*s->mb_y, 16);

            s->mb_x = 0;
            s->mb_y++;

            if(s->mb_y<<field_pic >= s->mb_height){

            if (s->last_picture_ptr != NULL) {
                *pict= *(AVFrame*)s->last_picture_ptr;
                 ff_print_debug_info(s, s->last_picture_ptr);
            }
        }

        return 1;
    } else {
        return 0;
    }

        avctx->bit_rate = s->bit_rate;
        
        //get_format() or set_video(width,height,aspect,pix_fmt);
        //until then pix_fmt may be changed right after codec init
        if( avctx->pix_fmt == PIX_FMT_XVMC_MPEG2_IDCT )
            if( avctx->idct_algo == FF_IDCT_AUTO )
                avctx->idct_algo = FF_IDCT_SIMPLE;

        if (MPV_common_init(s) < 0)
            return -1;
        s1->mpeg_enc_ctx_allocated = 1;
        s->swap_uv = 0;//just in case vcr2 and mpeg2 stream have been concatinated
    }

    skip_bits(&s->gb, 10); /* vbv_buffer_size */
    skip_bits(&s->gb, 1);

    s->avctx = avctx;

    //get_format() or set_video(width,height,aspect,pix_fmt);
    //until then pix_fmt may be changed right after codec init
    if( avctx->pix_fmt == PIX_FMT_XVMC_MPEG2_IDCT )
        if( avctx->idct_algo == FF_IDCT_AUTO )
            avctx->idct_algo = FF_IDCT_SIMPLE;
    
    if (MPV_common_init(s) < 0)
        return -1;
    exchange_uv(s);//common init reset pblocks, so we swap them here
    s->swap_uv = 1;// in case of xvmc we need to swap uv for each MB 
    s1->mpeg_enc_ctx_allocated = 1;

    for(i=0;i<64;i++) {
        int j= s->dsp.idct_permutation[i];
        v = ff_mpeg1_default_intra_matrix[i];

static int mpeg_mc_decode_init(AVCodecContext *avctx){
    Mpeg1Context *s;

    if( !(avctx->slice_flags & SLICE_FLAG_CODED_ORDER) )
        return -1;
    if( !(avctx->slice_flags & SLICE_FLAG_ALLOW_FIELD) ){
        dprintf("mpeg12.c: XvMC decoder will work better if SLICE_FLAG_ALLOW_FIELD is set\n");
    }
    mpeg_decode_init(avctx);
    s = avctx->priv_data;

    avctx->pix_fmt = PIX_FMT_XVMC_MPEG2_IDCT;
    avctx->xvmc_acceleration = 2;//2 - the blocks are packed!

    return 0;
}

AVCodec mpeg_xvmc_decoder = {