comparison dv.c @ 1567:e08df4d22d27 libavcodec

* introducing dct248 into the DSP context. * simple/accurate implementation of dct248 * DV encoding now supports 2-4-8 DCT * DV encoding gets a bit faster (but still miles away from what I think it could do) * misc. DV codec cleanups

1566:396e8526e82c

#include "simple_idct.h"
#include "dvdata.h"

typedef struct DVVideoDecodeContext {
    const DVprofile* sys;
    GetBitContext gb;
    AVFrame picture;
    DCTELEM block[5*6][64] __align8;
    
    /* FIXME: the following is extracted from DSP */
    uint8_t dv_zigzag[2][64];
    uint8_t idct_permutation[64];
    void (*get_pixels)(DCTELEM *block, const uint8_t *pixels, int line_size);
    void (*fdct)(DCTELEM *block);
    
    /* XXX: move it to static storage ? */
    uint8_t dv_shift[2][22][64];
    void (*idct_put[2])(uint8_t *dest, int line_size, DCTELEM *block);
} DVVideoDecodeContext;

#define TEX_VLC_BITS 9
/* XXX: also include quantization */
static RL_VLC_ELEM *dv_rl_vlc[1];
static VLC_TYPE dv_vlc_codes[15][23];

static void dv_build_unquantize_tables(DVVideoDecodeContext *s)
{
    int i, q, j;

    /* NOTE: max left shift is 6 */
    for(q = 0; q < 22; q++) {
        /* 88 unquant */
        for(i = 1; i < 64; i++) {
            /* 88 table */
            j = s->idct_permutation[i];
            s->dv_shift[0][q][j] =
                dv_quant_shifts[q][dv_88_areas[i]] + 1;
        }
        
        /* 248 unquant */
        for(i = 1; i < 64; i++) {
            /* 248 table */
            s->dv_shift[1][q][i] =  
                    dv_quant_shifts[q][dv_248_areas[i]] + 1;
        }
    }
}

static int dvvideo_init(AVCodecContext *avctx)
{
    DVVideoDecodeContext *s = avctx->priv_data;
    MpegEncContext s2;
    static int done=0;

    if (!done) {
        int i;
        VLC dv_vlc;

	   if (dv_vlc_run[i] < 15 && dv_vlc_level[i] < 23 && dv_vlc_len[i] < 15)
	       dv_vlc_codes[dv_vlc_run[i]][dv_vlc_level[i]] = i;
	}
    }

    /* ugly way to get the idct & scantable */
    /* XXX: fix it */
    memset(&s2, 0, sizeof(MpegEncContext));
    s2.avctx = avctx;
    dsputil_init(&s2.dsp, avctx);
    if (DCT_common_init(&s2) < 0)
       return -1;

    s->get_pixels = s2.dsp.get_pixels;
    s->fdct = s2.dsp.fdct;
    
    s->idct_put[0] = s2.dsp.idct_put;
    memcpy(s->idct_permutation, s2.dsp.idct_permutation, 64);
    memcpy(s->dv_zigzag[0], s2.intra_scantable.permutated, 64);

    /* XXX: use MMX also for idct248 */
    s->idct_put[1] = simple_idct248_put;
    memcpy(s->dv_zigzag[1], dv_248_zigzag, 64);

    /* XXX: do it only for constant case */
    dv_build_unquantize_tables(s);

    /* FIXME: I really don't think this should be here */
    if (dv_codec_profile(avctx))
	avctx->pix_fmt = dv_codec_profile(avctx)->pix_fmt; 
    avctx->coded_frame = &s->picture;

            dct_mode = get_bits1(&s->gb);
            mb->dct_mode = dct_mode;
            mb->scan_table = s->dv_zigzag[dct_mode];
            class1 = get_bits(&s->gb, 2);
            mb->shift_offset = (class1 == 3);
            mb->shift_table = s->dv_shift[dct_mode]
                [quant + dv_quant_offset[class1]];
            dc = dc << 2;
            /* convert to unsigned because 128 is not added in the
               standard IDCT */
            dc += 1024;

    int cno;
    int dct_mode;
    int block_size;
    DCTELEM *mb;
    PutBitContext pb;
} EncBlockInfo;

static inline int dv_bits_left(EncBlockInfo* bi)
{
    return (bi->block_size - get_bit_count(&bi->pb));

static inline void dv_encode_ac(EncBlockInfo* bi, PutBitContext* heap)
{
    int i, level, size, run = 0;
    uint32_t vlc;
    PutBitContext* cpb = &bi->pb;
    
    for (i=1; i<64; i++) {
       level = bi->mb[ff_zigzag_direct[i]] / 
               (1<<(dv_quant_shifts[bi->qno + dv_quant_offset[bi->cno]]
			       [dv_88_areas[ff_zigzag_direct[i]]] + 4 + (bi->cno == 3)));
       if (level != 0) {
	   size = dv_rl2vlc(run, level, &vlc);
put_vlc:

#ifdef VLC_DEBUG

    
    if (bi->cno > 3)
        bi->cno = 3;
}

/*
 * This is a very rough initial implementaion. The performance is
 * horrible and some features are missing, mainly 2-4-8 DCT encoding.
 * The weighting is missing as well, but it's missing from the decoding
 * step also -- so at least we're on the same page with decoder ;-)
 */
static inline void dv_encode_video_segment(DVVideoDecodeContext *s, 
                                           uint8_t *dif, 
                                           const uint16_t *mb_pos_ptr)
{

    int       QNO = 15;
   
    /* Stage 1 -- doing DCT on 5 MBs */
    block = &s->block[0][0];
    for(mb_index = 0; mb_index < 5; mb_index++) {
        v = *mb_pos_ptr++;
        mb_x = v & 0xff;
        mb_y = v >> 8;
        y_ptr = s->picture.data[0] + (mb_y * s->picture.linesize[0] * 8) + (mb_x * 8);

		   b += 8;
		}
	    } else {             /* Simple copy: 8x8 -> 8x8 */
	        s->get_pixels(block, data, linesize);
	    }
            
	    s->fdct(block);
	    
	    block += 64;
        }
    }

    /* Stage 2 -- setup for encoding phase */
    enc_blk = &enc_blks[0];
    block = &s->block[0][0];
    for (i=0; i<5; i++) {
       for (j=0; j<6; j++) {
	  enc_blk->mb = block;
	  enc_blk->dct_mode = 0;
	  enc_blk->block_size = block_sizes[j];
	  
	  dv_set_class_number(enc_blk, j/4*(j%2));
	  
	  block += 64;
	  enc_blk++;
       }
    }
   
    /* Stage 3 -- encoding by trial-and-error */
encode_vs:
    enc_blk = &enc_blks[0];
    for (i=0; i<5; i++) {
       uint8_t* p = dif + i*80 + 4;
       for (j=0; j<6; j++) {
          enc_blk->qno = QNO;
	  init_put_bits(&enc_blk->pb, p, block_sizes[j]/8);
	  enc_blk++;
	  p += block_sizes[j]/8;
       }
    }

1567:e08df4d22d27

#include "simple_idct.h"
#include "dvdata.h"

typedef struct DVVideoDecodeContext {
    const DVprofile* sys;
    AVFrame picture;
    
    uint8_t dv_zigzag[2][64];
    uint8_t dv_idct_shift[2][22][64];
    uint8_t dv_dct_shift[2][22][64];
  
    void (*get_pixels)(DCTELEM *block, const uint8_t *pixels, int line_size);
    void (*fdct[2])(DCTELEM *block);
    void (*idct_put[2])(uint8_t *dest, int line_size, DCTELEM *block);
    
    GetBitContext gb;
    DCTELEM block[5*6][64] __align8;
} DVVideoDecodeContext;

#define TEX_VLC_BITS 9
/* XXX: also include quantization */
static RL_VLC_ELEM *dv_rl_vlc[1];
static VLC_TYPE dv_vlc_codes[15][23];

static void dv_build_unquantize_tables(DVVideoDecodeContext *s, uint8_t* perm)
{
    int i, q, j;

    /* NOTE: max left shift is 6 */
    for(q = 0; q < 22; q++) {
        /* 88DCT */
        for(i = 1; i < 64; i++) {
            /* 88 table */
            j = perm[i];
            s->dv_idct_shift[0][q][j] =
                dv_quant_shifts[q][dv_88_areas[i]] + 1;
            s->dv_dct_shift[0][q][i] =
                dv_quant_shifts[q][dv_88_areas[ff_zigzag_direct[i]]] + 4;
        }
        
        /* 248DCT */
        for(i = 1; i < 64; i++) {
            /* 248 table */
            s->dv_idct_shift[1][q][i] =  
                dv_quant_shifts[q][dv_248_areas[i]] + 1;
	    s->dv_dct_shift[1][q][i] =  
                dv_quant_shifts[q][dv_248_areas[ff_zigzag248_direct[i]]] + 4;
        }
    }
}

static int dvvideo_init(AVCodecContext *avctx)
{
    DVVideoDecodeContext *s = avctx->priv_data;
    DSPContext dsp;
    static int done=0;
    int i;

    if (!done) {
        int i;
        VLC dv_vlc;

	   if (dv_vlc_run[i] < 15 && dv_vlc_level[i] < 23 && dv_vlc_len[i] < 15)
	       dv_vlc_codes[dv_vlc_run[i]][dv_vlc_level[i]] = i;
	}
    }

    /* Generic DSP setup */
    dsputil_init(&dsp, avctx);
    s->get_pixels = dsp.get_pixels;

    /* 88DCT setup */
    s->fdct[0] = dsp.fdct;
    s->idct_put[0] = dsp.idct_put;
    for (i=0; i<64; i++)
       s->dv_zigzag[0][i] = dsp.idct_permutation[ff_zigzag_direct[i]];

    /* 248DCT setup */
    s->fdct[1] = dsp.fdct248;
    s->idct_put[1] = simple_idct248_put;  // FIXME: need to add it to DSP
    memcpy(s->dv_zigzag[1], ff_zigzag248_direct, 64);

    /* XXX: do it only for constant case */
    dv_build_unquantize_tables(s, dsp.idct_permutation);

    /* FIXME: I really don't think this should be here */
    if (dv_codec_profile(avctx))
	avctx->pix_fmt = dv_codec_profile(avctx)->pix_fmt; 
    avctx->coded_frame = &s->picture;

            dct_mode = get_bits1(&s->gb);
            mb->dct_mode = dct_mode;
            mb->scan_table = s->dv_zigzag[dct_mode];
            class1 = get_bits(&s->gb, 2);
            mb->shift_offset = (class1 == 3);
            mb->shift_table = s->dv_idct_shift[dct_mode]
                [quant + dv_quant_offset[class1]];
            dc = dc << 2;
            /* convert to unsigned because 128 is not added in the
               standard IDCT */
            dc += 1024;

    int cno;
    int dct_mode;
    int block_size;
    DCTELEM *mb;
    PutBitContext pb;
    const uint8_t* zigzag_scan;
    uint8_t *dv_shift;
} EncBlockInfo;

static inline int dv_bits_left(EncBlockInfo* bi)
{
    return (bi->block_size - get_bit_count(&bi->pb));

static inline void dv_encode_ac(EncBlockInfo* bi, PutBitContext* heap)
{
    int i, level, size, run = 0;
    uint32_t vlc;
    PutBitContext* cpb = &bi->pb;
    int bias = (bi->cno == 3);
    
    for (i=1; i<64; i++) {
       level = bi->mb[bi->zigzag_scan[i]] / (1<<(bi->dv_shift[i] + bias));
       if (level != 0) {
	   size = dv_rl2vlc(run, level, &vlc);
put_vlc:

#ifdef VLC_DEBUG

    
    if (bi->cno > 3)
        bi->cno = 3;
}

#define SQ(a) ((a)*(a))
static int dv_score_lines(DCTELEM *s, int stride) {
    int score=0;
    int x, y;
    
    for(y=0; y<4; y++) {
        for(x=0; x<8; x+=4){
            score+= SQ(s[x  ] - s[x  +stride]) + SQ(s[x+1] - s[x+1+stride]) 
                   +SQ(s[x+2] - s[x+2+stride]) + SQ(s[x+3] - s[x+3+stride]);
        }
        s+= stride;
    }
    
    return score;
}

/*
 * This is a very rough initial implementaion. The performance is
 * horrible and the weighting is missing. But it's missing from the 
 * decoding step also -- so at least we're on the same page with decoder ;-)
 */
static inline void dv_encode_video_segment(DVVideoDecodeContext *s, 
                                           uint8_t *dif, 
                                           const uint16_t *mb_pos_ptr)
{

    int       QNO = 15;
   
    /* Stage 1 -- doing DCT on 5 MBs */
    block = &s->block[0][0];
    enc_blk = &enc_blks[0];
    for(mb_index = 0; mb_index < 5; mb_index++) {
        v = *mb_pos_ptr++;
        mb_x = v & 0xff;
        mb_y = v >> 8;
        y_ptr = s->picture.data[0] + (mb_y * s->picture.linesize[0] * 8) + (mb_x * 8);

		   b += 8;
		}
	    } else {             /* Simple copy: 8x8 -> 8x8 */
	        s->get_pixels(block, data, linesize);
	    }
	  
	    if (dv_score_lines(block, 8) + dv_score_lines(block+8*4, 8) - 100 >
	        dv_score_lines(block, 16) + dv_score_lines(block+8, 16)) {
               enc_blk->dct_mode = 1;
	       enc_blk->zigzag_scan = ff_zigzag248_direct; 
	    } else {
	       enc_blk->dct_mode = 0;
	       enc_blk->zigzag_scan = ff_zigzag_direct;
	    }
	    enc_blk->mb = block;
            enc_blk->block_size = block_sizes[j];
	    
	    s->fdct[enc_blk->dct_mode](block);
	    
	    dv_set_class_number(enc_blk, j/4*(j%2));

	    block += 64;
	    enc_blk++;
        }
    }

    /* Stage 2 -- encoding by trial-and-error */
encode_vs:
    enc_blk = &enc_blks[0];
    for (i=0; i<5; i++) {
       uint8_t* p = dif + i*80 + 4;
       for (j=0; j<6; j++) {
          enc_blk->qno = QNO;
	  enc_blk->dv_shift = &(s->dv_dct_shift[0]
	                           [QNO + dv_quant_offset[enc_blk->cno]][0]);
	  init_put_bits(&enc_blk->pb, p, block_sizes[j]/8);
	  enc_blk++;
	  p += block_sizes[j]/8;
       }
    }