svq1enc.c

mediastreamer2是开源的网络传输媒体流的库

/*
 * SVQ1 Encoder
 * Copyright (C) 2004 Mike Melanson <[e-mail]>
 *
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2.1 of the License, or (at your option) any later version.
 *
 * FFmpeg is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with FFmpeg; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

/**
 * @file svq1enc.c
 * Sorenson Vector Quantizer #1 (SVQ1) video codec.
 * For more information of the SVQ1 algorithm, visit:
 *   http://www.pcisys.net/~melanson/codecs/
 */


#include "avcodec.h"
#include "dsputil.h"
#include "mpegvideo.h"

#include "svq1.h"
#include "svq1enc_cb.h"

#undef NDEBUG
#include <assert.h>


typedef struct SVQ1Context {
    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;
    DSPContext dsp;
    AVFrame picture;
    AVFrame current_picture;
    AVFrame last_picture;
    PutBitContext pb;
    GetBitContext gb;

    PutBitContext reorder_pb[6]; //why ooh why this sick breadth first order, everything is slower and more complex

    int frame_width;
    int frame_height;

    /* Y plane block dimensions */
    int y_block_width;
    int y_block_height;

    /* U & V plane (C planes) block dimensions */
    int c_block_width;
    int c_block_height;

    uint16_t *mb_type;
    uint32_t *dummy;
    int16_t (*motion_val8[3])[2];
    int16_t (*motion_val16[3])[2];

    int64_t rd_total;
} SVQ1Context;

static void svq1_write_header(SVQ1Context *s, int frame_type)
{
    int i;

    /* frame code */
    put_bits(&s->pb, 22, 0x20);

    /* temporal reference (sure hope this is a "don't care") */
    put_bits(&s->pb, 8, 0x00);

    /* frame type */
    put_bits(&s->pb, 2, frame_type - 1);

    if (frame_type == I_TYPE) {

        /* no checksum since frame code is 0x20 */

        /* no embedded string either */

        /* output 5 unknown bits (2 + 2 + 1) */
        put_bits(&s->pb, 5, 2); /* 2 needed by quicktime decoder */

        for (i = 0; i < 7; i++)
        {
            if ((ff_svq1_frame_size_table[i].width == s->frame_width) &&
                (ff_svq1_frame_size_table[i].height == s->frame_height))
            {
                put_bits(&s->pb, 3, i);
                break;
            }
        }

        if (i == 7)
        {
            put_bits(&s->pb, 3, 7);
                put_bits(&s->pb, 12, s->frame_width);
                put_bits(&s->pb, 12, s->frame_height);
        }
    }

    /* no checksum or extra data (next 2 bits get 0) */
    put_bits(&s->pb, 2, 0);
}


#define QUALITY_THRESHOLD 100
#define THRESHOLD_MULTIPLIER 0.6

#if defined(HAVE_ALTIVEC)
#undef vector
#endif

static int encode_block(SVQ1Context *s, uint8_t *src, uint8_t *ref, uint8_t *decoded, int stride, int level, int threshold, int lambda, int intra){
    int count, y, x, i, j, split, best_mean, best_score, best_count;
    int best_vector[6];
    int block_sum[7]= {0, 0, 0, 0, 0, 0};
    int w= 2<<((level+2)>>1);
    int h= 2<<((level+1)>>1);
    int size=w*h;
    int16_t block[7][256];
    const int8_t *codebook_sum, *codebook;
    const uint16_t (*mean_vlc)[2];
    const uint8_t (*multistage_vlc)[2];

    best_score=0;
    //FIXME optimize, this doenst need to be done multiple times
    if(intra){
        codebook_sum= svq1_intra_codebook_sum[level];
        codebook= ff_svq1_intra_codebooks[level];
        mean_vlc= ff_svq1_intra_mean_vlc;
        multistage_vlc= ff_svq1_intra_multistage_vlc[level];
        for(y=0; y<h; y++){
            for(x=0; x<w; x++){
                int v= src[x + y*stride];
                block[0][x + w*y]= v;
                best_score += v*v;
                block_sum[0] += v;
            }
        }
    }else{
        codebook_sum= svq1_inter_codebook_sum[level];
        codebook= ff_svq1_inter_codebooks[level];
        mean_vlc= ff_svq1_inter_mean_vlc + 256;
        multistage_vlc= ff_svq1_inter_multistage_vlc[level];
        for(y=0; y<h; y++){
            for(x=0; x<w; x++){
                int v= src[x + y*stride] - ref[x + y*stride];
                block[0][x + w*y]= v;
                best_score += v*v;
                block_sum[0] += v;
            }
        }
    }

    best_count=0;
    best_score -= ((block_sum[0]*block_sum[0])>>(level+3));
    best_mean= (block_sum[0] + (size>>1)) >> (level+3);

    if(level<4){
        for(count=1; count<7; count++){
            int best_vector_score= INT_MAX;
            int best_vector_sum=-999, best_vector_mean=-999;
            const int stage= count-1;
            const int8_t *vector;

            for(i=0; i<16; i++){
                int sum= codebook_sum[stage*16 + i];
                int sqr, diff, score;

                vector = codebook + stage*size*16 + i*size;
                sqr = s->dsp.ssd_int8_vs_int16(vector, block[stage], size);
                diff= block_sum[stage] - sum;
                score= sqr - ((diff*(int64_t)diff)>>(level+3)); //FIXME 64bit slooow
                if(score < best_vector_score){
                    int mean= (diff + (size>>1)) >> (level+3);
                    assert(mean >-300 && mean<300);
                    mean= av_clip(mean, intra?0:-256, 255);
                    best_vector_score= score;
                    best_vector[stage]= i;
                    best_vector_sum= sum;
                    best_vector_mean= mean;
                }
            }
            assert(best_vector_mean != -999);
            vector= codebook + stage*size*16 + best_vector[stage]*size;
            for(j=0; j<size; j++){
                block[stage+1][j] = block[stage][j] - vector[j];
            }
            block_sum[stage+1]= block_sum[stage] - best_vector_sum;
            best_vector_score +=
                lambda*(+ 1 + 4*count
                        + multistage_vlc[1+count][1]
                        + mean_vlc[best_vector_mean][1]);

            if(best_vector_score < best_score){
                best_score= best_vector_score;
                best_count= count;
                best_mean= best_vector_mean;
            }
        }
    }

    split=0;
    if(best_score > threshold && level){
        int score=0;
        int offset= (level&1) ? stride*h/2 : w/2;
        PutBitContext backup[6];

        for(i=level-1; i>=0; i--){
            backup[i]= s->reorder_pb[i];
        }
        score += encode_block(s, src         , ref         , decoded         , stride, level-1, threshold>>1, lambda, intra);
        score += encode_block(s, src + offset, ref + offset, decoded + offset, stride, level-1, threshold>>1, lambda, intra);
        score += lambda;

        if(score < best_score){
            best_score= score;
            split=1;
        }else{
            for(i=level-1; i>=0; i--){
                s->reorder_pb[i]= backup[i];
            }
        }
    }
    if (level > 0)
        put_bits(&s->reorder_pb[level], 1, split);

    if(!split){
        assert((best_mean >= 0 && best_mean<256) || !intra);
        assert(best_mean >= -256 && best_mean<256);
        assert(best_count >=0 && best_count<7);
        assert(level<4 || best_count==0);

        /* output the encoding */
        put_bits(&s->reorder_pb[level],
            multistage_vlc[1 + best_count][1],
            multistage_vlc[1 + best_count][0]);
        put_bits(&s->reorder_pb[level], mean_vlc[best_mean][1],
            mean_vlc[best_mean][0]);

        for (i = 0; i < best_count; i++){
            assert(best_vector[i]>=0 && best_vector[i]<16);
            put_bits(&s->reorder_pb[level], 4, best_vector[i]);
        }

        for(y=0; y<h; y++){
            for(x=0; x<w; x++){
                decoded[x + y*stride]= src[x + y*stride] - block[best_count][x + w*y] + best_mean;
            }
        }
    }

    return best_score;
}


static int svq1_encode_plane(SVQ1Context *s, int plane, unsigned char *src_plane, unsigned char *ref_plane, unsigned char *decoded_plane,
    int width, int height, int src_stride, int stride)
{
    int x, y;
    int i;
    int block_width, block_height;
    int level;
    int threshold[6];
    const int lambda= (s->picture.quality*s->picture.quality) >> (2*FF_LAMBDA_SHIFT);

    /* figure out the acceptable level thresholds in advance */
    threshold[5] = QUALITY_THRESHOLD;
    for (level = 4; level >= 0; level--)
        threshold[level] = threshold[level + 1] * THRESHOLD_MULTIPLIER;

    block_width = (width + 15) / 16;
    block_height = (height + 15) / 16;

    if(s->picture.pict_type == P_TYPE){
        s->m.avctx= s->avctx;
        s->m.current_picture_ptr= &s->m.current_picture;
        s->m.last_picture_ptr   = &s->m.last_picture;
        s->m.last_picture.data[0]= ref_plane;
        s->m.linesize=
        s->m.last_picture.linesize[0]=
        s->m.new_picture.linesize[0]=
        s->m.current_picture.linesize[0]= stride;
        s->m.width= width;