macroblock.c 29.5 KB
Newer Older
Laurent Aimar's avatar
Laurent Aimar committed
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
/*****************************************************************************
 * macroblock.c: h264 encoder library
 *****************************************************************************
 * Copyright (C) 2003 Laurent Aimar
 * $Id: macroblock.c,v 1.1 2004/06/03 19:27:08 fenrir Exp $
 *
 * Authors: Laurent Aimar <fenrir@via.ecp.fr>
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 2 of the License, or
 * (at your option) any later version.
 *
 * This program 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 General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software
 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111, USA.
 *****************************************************************************/

#include <stdio.h>
#include <string.h>

27
#include "common/common.h"
Laurent Aimar's avatar
Laurent Aimar committed
28 29 30
#include "macroblock.h"


31
#define ZIG(i,y,x) level[i] = dct[x][y];
32
static inline void zigzag_scan_2x2_dc( int level[4], int16_t dct[2][2] )
Laurent Aimar's avatar
Laurent Aimar committed
33
{
34 35 36 37
    ZIG(0,0,0)
    ZIG(1,0,1)
    ZIG(2,1,0)
    ZIG(3,1,1)
Laurent Aimar's avatar
Laurent Aimar committed
38
}
39
#undef ZIG
Laurent Aimar's avatar
Laurent Aimar committed
40

41
static void quant_8x8( x264_t *h, int16_t dct[8][8], int quant_mf[6][8][8], int i_qscale, int b_intra )
42 43 44
{
    const int i_qbits = 16 + i_qscale / 6;
    const int i_mf = i_qscale % 6;
45
    const int f = h->mb.i_luma_deadzone[b_intra] << (i_qbits-6);
46
    h->quantf.quant_8x8_core( dct, quant_mf[i_mf], i_qbits, f );
47
}
48
static void quant_4x4( x264_t *h, int16_t dct[4][4], int quant_mf[6][4][4], int i_qscale, int b_intra )
49 50 51 52 53 54 55
{
    const int i_qbits = 15 + i_qscale / 6;
    const int i_mf = i_qscale % 6;
    const int f = h->mb.i_luma_deadzone[b_intra] << (i_qbits-6);
    h->quantf.quant_4x4_core( dct, quant_mf[i_mf], i_qbits, f );
}
static void quant_4x4_chroma( x264_t *h, int16_t dct[4][4], int quant_mf[6][4][4], int i_qscale, int b_intra )
Laurent Aimar's avatar
Laurent Aimar committed
56 57 58
{
    const int i_qbits = 15 + i_qscale / 6;
    const int i_mf = i_qscale % 6;
Loren Merritt's avatar
Loren Merritt committed
59
    const int f = ( 1 << (i_qbits + b_intra) ) / 6;
60
    h->quantf.quant_4x4_core( dct, quant_mf[i_mf], i_qbits, f );
Laurent Aimar's avatar
Laurent Aimar committed
61
}
62
static void quant_4x4_dc( x264_t *h, int16_t dct[4][4], int quant_mf[6][4][4], int i_qscale )
Laurent Aimar's avatar
Laurent Aimar committed
63
{
64 65
    const int i_qbits = 16 + i_qscale / 6;
    const int i_mf = i_qscale % 6;
66
    const int f = h->mb.i_luma_deadzone[1] << (i_qbits-6);
67
    h->quantf.quant_4x4_dc_core( dct, quant_mf[i_mf][0][0], i_qbits, f );
Laurent Aimar's avatar
Laurent Aimar committed
68
}
69
static void quant_2x2_dc( x264_t *h, int16_t dct[2][2], int quant_mf[6][4][4], int i_qscale, int b_intra )
Laurent Aimar's avatar
Laurent Aimar committed
70
{
71 72
    const int i_qbits = 16 + i_qscale / 6;
    const int i_mf = i_qscale % 6;
Loren Merritt's avatar
Loren Merritt committed
73
    const int f = ( 1 << (i_qbits + b_intra) ) / 6;
74
    h->quantf.quant_2x2_dc_core( dct, quant_mf[i_mf][0][0], i_qbits, f );
Laurent Aimar's avatar
Laurent Aimar committed
75
}
76

Laurent Aimar's avatar
Laurent Aimar committed
77 78 79 80 81 82 83 84 85 86
/* (ref: JVT-B118)
 * x264_mb_decimate_score: given dct coeffs it returns a score to see if we could empty this dct coeffs
 * to 0 (low score means set it to null)
 * Used in inter macroblock (luma and chroma)
 *  luma: for a 8x8 block: if score < 4 -> null
 *        for the complete mb: if score < 6 -> null
 *  chroma: for the complete mb: if score < 7 -> null
 */
static int x264_mb_decimate_score( int *dct, int i_max )
{
87 88 89 90 91 92 93 94 95
    static const int i_ds_table4[16] = {
        3,2,2,1,1,1,0,0,0,0,0,0,0,0,0,0 };
    static const int i_ds_table8[64] = {
        3,3,3,3,2,2,2,2,2,2,2,2,1,1,1,1,
        1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,
        0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
        0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 };

    const int *ds_table = (i_max == 64) ? i_ds_table8 : i_ds_table4;
Laurent Aimar's avatar
Laurent Aimar committed
96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114
    int i_score = 0;
    int idx = i_max - 1;

    while( idx >= 0 && dct[idx] == 0 )
        idx--;

    while( idx >= 0 )
    {
        int i_run;

        if( abs( dct[idx--] ) > 1 )
            return 9;

        i_run = 0;
        while( idx >= 0 && dct[idx] == 0 )
        {
            idx--;
            i_run++;
        }
115
        i_score += ds_table[i_run];
Laurent Aimar's avatar
Laurent Aimar committed
116 117 118 119 120 121 122
    }

    return i_score;
}

void x264_mb_encode_i4x4( x264_t *h, int idx, int i_qscale )
{
123 124 125 126
    int x = 4 * block_idx_x[idx];
    int y = 4 * block_idx_y[idx];
    uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
    uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
127
    DECLARE_ALIGNED( int16_t, dct4x4[4][4], 16 );
Laurent Aimar's avatar
Laurent Aimar committed
128

Loren Merritt's avatar
Loren Merritt committed
129 130
    if( h->mb.b_lossless )
    {
131
        h->zigzagf.sub_4x4( h->dct.block[idx].luma4x4, p_src, p_dst );
Loren Merritt's avatar
Loren Merritt committed
132 133 134
        return;
    }

135
    h->dctf.sub4x4_dct( dct4x4, p_src, p_dst );
Loren Merritt's avatar
Loren Merritt committed
136 137 138 139 140 141

    if( h->mb.b_trellis )
        x264_quant_4x4_trellis( h, dct4x4, CQM_4IY, i_qscale, DCT_LUMA_4x4, 1 );
    else
        quant_4x4( h, dct4x4, h->quant4_mf[CQM_4IY], i_qscale, 1 );

142
    h->zigzagf.scan_4x4( h->dct.block[idx].luma4x4, dct4x4 );
143
    h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4IY], i_qscale );
Laurent Aimar's avatar
Laurent Aimar committed
144 145

    /* output samples to fdec */
146
    h->dctf.add4x4_idct( p_dst, dct4x4 );
Laurent Aimar's avatar
Laurent Aimar committed
147 148
}

149 150
void x264_mb_encode_i8x8( x264_t *h, int idx, int i_qscale )
{
151 152 153 154
    int x = 8 * (idx&1);
    int y = 8 * (idx>>1);
    uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
    uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
155
    DECLARE_ALIGNED( int16_t, dct8x8[8][8], 16 );
156

157
    h->dctf.sub8x8_dct8( dct8x8, p_src, p_dst );
Loren Merritt's avatar
Loren Merritt committed
158 159 160 161 162 163

    if( h->mb.b_trellis )
        x264_quant_8x8_trellis( h, dct8x8, CQM_8IY, i_qscale, 1 );
    else 
        quant_8x8( h, dct8x8, h->quant8_mf[CQM_8IY], i_qscale, 1 );

164
    h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8 );
165
    h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8IY], i_qscale );
166
    h->dctf.add8x8_idct8( p_dst, dct8x8 );
167 168
}

Laurent Aimar's avatar
Laurent Aimar committed
169 170 171 172 173
static void x264_mb_encode_i16x16( x264_t *h, int i_qscale )
{
    uint8_t  *p_src = h->mb.pic.p_fenc[0];
    uint8_t  *p_dst = h->mb.pic.p_fdec[0];

174
    DECLARE_ALIGNED( int16_t, dct4x4[16+1][4][4], 16 );
Laurent Aimar's avatar
Laurent Aimar committed
175 176 177

    int i;

Loren Merritt's avatar
Loren Merritt committed
178 179 180 181
    if( h->mb.b_lossless )
    {
        for( i = 0; i < 16; i++ )
        {
182 183
            int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
            int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
184
            h->zigzagf.sub_4x4ac( h->dct.block[i].residual_ac, p_src+oe, p_dst+od );
185
            dct4x4[0][block_idx_x[i]][block_idx_y[i]] = p_src[oe] - p_dst[od];
186
            p_dst[od] = p_src[oe];
Loren Merritt's avatar
Loren Merritt committed
187
        }
188
        h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct4x4[0] );
Loren Merritt's avatar
Loren Merritt committed
189 190 191
        return;
    }

192
    h->dctf.sub16x16_dct( &dct4x4[1], p_src, p_dst );
Laurent Aimar's avatar
Laurent Aimar committed
193 194 195 196 197 198
    for( i = 0; i < 16; i++ )
    {
        /* copy dc coeff */
        dct4x4[0][block_idx_y[i]][block_idx_x[i]] = dct4x4[1+i][0][0];

        /* quant/scan/dequant */
Loren Merritt's avatar
Loren Merritt committed
199 200 201 202 203
        if( h->mb.b_trellis )
            x264_quant_4x4_trellis( h, dct4x4[1+i], CQM_4IY, i_qscale, DCT_LUMA_AC, 1 );
        else
            quant_4x4( h, dct4x4[1+i], h->quant4_mf[CQM_4IY], i_qscale, 1 );

204
        h->zigzagf.scan_4x4ac( h->dct.block[i].residual_ac, dct4x4[1+i] );
205
        h->quantf.dequant_4x4( dct4x4[1+i], h->dequant4_mf[CQM_4IY], i_qscale );
Laurent Aimar's avatar
Laurent Aimar committed
206 207 208
    }

    h->dctf.dct4x4dc( dct4x4[0] );
209
    quant_4x4_dc( h, dct4x4[0], h->quant4_mf[CQM_4IY], i_qscale );
210
    h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct4x4[0] );
Laurent Aimar's avatar
Laurent Aimar committed
211 212 213

    /* output samples to fdec */
    h->dctf.idct4x4dc( dct4x4[0] );
Loren Merritt's avatar
Loren Merritt committed
214
    x264_mb_dequant_4x4_dc( dct4x4[0], h->dequant4_mf[CQM_4IY], i_qscale );  /* XXX not inversed */
Laurent Aimar's avatar
Laurent Aimar committed
215 216 217 218 219 220 221 222

    /* calculate dct coeffs */
    for( i = 0; i < 16; i++ )
    {
        /* copy dc coeff */
        dct4x4[1+i][0][0] = dct4x4[0][block_idx_y[i]][block_idx_x[i]];
    }
    /* put pixels to fdec */
223
    h->dctf.add16x16_idct( p_dst, &dct4x4[1] );
Laurent Aimar's avatar
Laurent Aimar committed
224 225
}

226
void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qscale )
Laurent Aimar's avatar
Laurent Aimar committed
227 228
{
    int i, ch;
229
    int b_decimate = b_inter && (h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate);
Laurent Aimar's avatar
Laurent Aimar committed
230 231 232 233 234 235 236

    for( ch = 0; ch < 2; ch++ )
    {
        uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
        uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
        int i_decimate_score = 0;

237 238
        DECLARE_ALIGNED( int16_t, dct2x2[2][2] , 16 );
        DECLARE_ALIGNED( int16_t, dct4x4[4][4][4], 16 );
Laurent Aimar's avatar
Laurent Aimar committed
239

Loren Merritt's avatar
Loren Merritt committed
240 241 242 243
        if( h->mb.b_lossless )
        {
            for( i = 0; i < 4; i++ )
            {
244 245
                int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
                int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
246
                h->zigzagf.sub_4x4ac( h->dct.block[16+i+ch*4].residual_ac, p_src+oe, p_dst+od );
247 248
                h->dct.chroma_dc[ch][i] = p_src[oe] - p_dst[od];
                p_dst[od] = p_src[oe];
Loren Merritt's avatar
Loren Merritt committed
249 250 251 252
            }
            continue;
        }
            
253
        h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
Laurent Aimar's avatar
Laurent Aimar committed
254 255 256 257 258 259
        /* calculate dct coeffs */
        for( i = 0; i < 4; i++ )
        {
            /* copy dc coeff */
            dct2x2[block_idx_y[i]][block_idx_x[i]] = dct4x4[i][0][0];

Loren Merritt's avatar
Loren Merritt committed
260
            /* no trellis; it doesn't seem to help chroma noticeably */
261
            quant_4x4_chroma( h, dct4x4[i], h->quant4_mf[CQM_4IC + b_inter], i_qscale, !b_inter );
262
            h->zigzagf.scan_4x4ac( h->dct.block[16+i+ch*4].residual_ac, dct4x4[i] );
Laurent Aimar's avatar
Laurent Aimar committed
263

264
            if( b_decimate )
Laurent Aimar's avatar
Laurent Aimar committed
265 266 267 268 269 270
            {
                i_decimate_score += x264_mb_decimate_score( h->dct.block[16+i+ch*4].residual_ac, 15 );
            }
        }

        h->dctf.dct2x2dc( dct2x2 );
271
        quant_2x2_dc( h, dct2x2, h->quant4_mf[CQM_4IC + b_inter], i_qscale, !b_inter );
272
        zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
Laurent Aimar's avatar
Laurent Aimar committed
273 274 275

        /* output samples to fdec */
        h->dctf.idct2x2dc( dct2x2 );
Loren Merritt's avatar
Loren Merritt committed
276
        x264_mb_dequant_2x2_dc( dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qscale );  /* XXX not inversed */
Laurent Aimar's avatar
Laurent Aimar committed
277

278
        if( b_decimate && i_decimate_score < 7 )
Laurent Aimar's avatar
Laurent Aimar committed
279 280
        {
            /* Near null chroma 8x8 block so make it null (bits saving) */
Loren Merritt's avatar
Loren Merritt committed
281
            memset( &h->dct.block[16+ch*4], 0, 4 * sizeof( *h->dct.block ) );
282
            if( !array_non_zero( dct2x2 ) )
283 284
                continue;
            memset( dct4x4, 0, sizeof( dct4x4 ) );
Laurent Aimar's avatar
Laurent Aimar committed
285
        }
286 287 288 289 290
        else
        {
            for( i = 0; i < 4; i++ )
                h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IC + b_inter], i_qscale );
        }
Laurent Aimar's avatar
Laurent Aimar committed
291 292

        for( i = 0; i < 4; i++ )
293
            dct4x4[i][0][0] = dct2x2[0][i];
294
        h->dctf.add8x8_idct( p_dst, dct4x4 );
Laurent Aimar's avatar
Laurent Aimar committed
295
    }
296 297 298 299 300 301 302 303 304 305 306

    /* coded block pattern */
    h->mb.i_cbp_chroma = 0;
    for( i = 0; i < 8; i++ )
    {
        int nz = array_non_zero_count( h->dct.block[16+i].residual_ac, 15 );
        h->mb.cache.non_zero_count[x264_scan8[16+i]] = nz;
        h->mb.i_cbp_chroma |= nz;
    }
    if( h->mb.i_cbp_chroma )
        h->mb.i_cbp_chroma = 2;    /* dc+ac (we can't do only ac) */
307
    else if( array_non_zero( h->dct.chroma_dc ) )
308
        h->mb.i_cbp_chroma = 1;    /* dc only */
Laurent Aimar's avatar
Laurent Aimar committed
309 310
}

311 312 313 314 315 316 317 318 319 320 321 322 323 324 325
static void x264_macroblock_encode_skip( x264_t *h )
{
    int i;
    h->mb.i_cbp_luma = 0x00;
    h->mb.i_cbp_chroma = 0x00;

    for( i = 0; i < 16+8; i++ )
    {
        h->mb.cache.non_zero_count[x264_scan8[i]] = 0;
    }

    /* store cbp */
    h->mb.cbp[h->mb.i_mb_xy] = 0;
}

Laurent Aimar's avatar
Laurent Aimar committed
326 327 328 329 330 331
/*****************************************************************************
 * x264_macroblock_encode_pskip:
 *  Encode an already marked skip block
 *****************************************************************************/
void x264_macroblock_encode_pskip( x264_t *h )
{
332 333 334 335
    const int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
                                h->mb.mv_min[0], h->mb.mv_max[0] );
    const int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
                                h->mb.mv_min[1], h->mb.mv_max[1] );
Laurent Aimar's avatar
Laurent Aimar committed
336 337

    /* Motion compensation XXX probably unneeded */
338
    h->mc.mc_luma( h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
339 340
                   h->mb.pic.p_fdec[0],    FDEC_STRIDE,
                   mvx, mvy, 16, 16 );
Laurent Aimar's avatar
Laurent Aimar committed
341 342

    /* Chroma MC */
343
    h->mc.mc_chroma( h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
344 345
                     h->mb.pic.p_fdec[1],       FDEC_STRIDE,
                     mvx, mvy, 8, 8 );
Laurent Aimar's avatar
Laurent Aimar committed
346

347
    h->mc.mc_chroma( h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
348 349
                     h->mb.pic.p_fdec[2],       FDEC_STRIDE,
                     mvx, mvy, 8, 8 );
Laurent Aimar's avatar
Laurent Aimar committed
350

351
    x264_macroblock_encode_skip( h );
Laurent Aimar's avatar
Laurent Aimar committed
352 353 354 355 356 357 358 359
}

/*****************************************************************************
 * x264_macroblock_encode:
 *****************************************************************************/
void x264_macroblock_encode( x264_t *h )
{
    int i_cbp_dc = 0;
360
    int i_qp = h->mb.i_qp;
361
    int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
362
    int b_force_no_skip = 0;
Laurent Aimar's avatar
Laurent Aimar committed
363 364
    int i;

365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381
    if( h->sh.b_mbaff
        && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
        && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
    {
        /* The first skip is predicted to be a frame mb pair.
         * We don't yet support the aff part of mbaff, so force it to non-skip
         * so that we can pick the aff flag. */
        b_force_no_skip = 1;
        if( IS_SKIP(h->mb.i_type) )
        {
            if( h->mb.i_type == P_SKIP )
                h->mb.i_type = P_L0;
            else if( h->mb.i_type == B_SKIP )
                h->mb.i_type = B_DIRECT;
        }
    }

Laurent Aimar's avatar
Laurent Aimar committed
382 383 384 385 386 387
    if( h->mb.i_type == P_SKIP )
    {
        /* A bit special */
        x264_macroblock_encode_pskip( h );
        return;
    }
388 389 390 391 392 393 394
    if( h->mb.i_type == B_SKIP )
    {
        /* XXX motion compensation is probably unneeded */
        x264_mb_mc( h );
        x264_macroblock_encode_skip( h );
        return;
    }
Laurent Aimar's avatar
Laurent Aimar committed
395 396 397 398

    if( h->mb.i_type == I_16x16 )
    {
        const int i_mode = h->mb.i_intra16x16_pred_mode;
399
        h->mb.b_transform_8x8 = 0;
Laurent Aimar's avatar
Laurent Aimar committed
400
        /* do the right prediction */
401
        h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
Laurent Aimar's avatar
Laurent Aimar committed
402 403

        /* encode the 16x16 macroblock */
404
        x264_mb_encode_i16x16( h, i_qp );
Laurent Aimar's avatar
Laurent Aimar committed
405
    }
406 407
    else if( h->mb.i_type == I_8x8 )
    {
408
        DECLARE_ALIGNED( uint8_t, edge[33], 8 );
409
        h->mb.b_transform_8x8 = 1;
410 411
        for( i = 0; i < 4; i++ )
        {
412
            uint8_t  *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
413 414
            int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];

415 416
            x264_predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
            h->predict_8x8[i_mode]( p_dst, edge );
417
            x264_mb_encode_i8x8( h, i, i_qp );
418 419
        }
    }
Laurent Aimar's avatar
Laurent Aimar committed
420 421
    else if( h->mb.i_type == I_4x4 )
    {
422
        h->mb.b_transform_8x8 = 0;
Laurent Aimar's avatar
Laurent Aimar committed
423 424
        for( i = 0; i < 16; i++ )
        {
425
            uint8_t  *p_dst = &h->mb.pic.p_fdec[0][4 * block_idx_x[i] + 4 * block_idx_y[i] * FDEC_STRIDE];
Laurent Aimar's avatar
Laurent Aimar committed
426 427
            int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];

428 429
            if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
                /* emulate missing topright samples */
430
                *(uint32_t*) &p_dst[4-FDEC_STRIDE] = p_dst[3-FDEC_STRIDE] * 0x01010101U;
431

432
            h->predict_4x4[i_mode]( p_dst );
433
            x264_mb_encode_i4x4( h, i, i_qp );
Laurent Aimar's avatar
Laurent Aimar committed
434 435 436 437 438 439 440 441 442 443
        }
    }
    else    /* Inter MB */
    {
        int i8x8, i4x4, idx;
        int i_decimate_mb = 0;

        /* Motion compensation */
        x264_mb_mc( h );

Loren Merritt's avatar
Loren Merritt committed
444 445 446 447
        if( h->mb.b_lossless )
        {
            for( i4x4 = 0; i4x4 < 16; i4x4++ )
            {
448 449
                int x = 4*block_idx_x[i4x4];
                int y = 4*block_idx_y[i4x4];
450
                h->zigzagf.sub_4x4( h->dct.block[i4x4].luma4x4,
451
                                    h->mb.pic.p_fenc[0]+x+y*FENC_STRIDE,
452
                                    h->mb.pic.p_fdec[0]+x+y*FDEC_STRIDE );
Loren Merritt's avatar
Loren Merritt committed
453 454 455
            }
        }
        else if( h->mb.b_transform_8x8 )
Laurent Aimar's avatar
Laurent Aimar committed
456
        {
457
            DECLARE_ALIGNED( int16_t, dct8x8[4][8][8], 16 );
458
            int nnz8x8[4] = {1,1,1,1};
459
            b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
460
            h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
Laurent Aimar's avatar
Laurent Aimar committed
461

462
            for( idx = 0; idx < 4; idx++ )
Laurent Aimar's avatar
Laurent Aimar committed
463
            {
464 465
                if( h->mb.b_noise_reduction )
                    x264_denoise_dct( h, (int16_t*)dct8x8[idx] );
Loren Merritt's avatar
Loren Merritt committed
466 467 468 469
                if( h->mb.b_trellis )
                    x264_quant_8x8_trellis( h, dct8x8[idx], CQM_8PY, i_qp, 0 );
                else
                    quant_8x8( h, dct8x8[idx], h->quant8_mf[CQM_8PY], i_qp, 0 );
Laurent Aimar's avatar
Laurent Aimar committed
470

471
                h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
Laurent Aimar's avatar
Laurent Aimar committed
472

473
                if( b_decimate )
474
                {
Loren Merritt's avatar
Loren Merritt committed
475 476 477 478 479 480
                    int i_decimate_8x8 = x264_mb_decimate_score( h->dct.luma8x8[idx], 64 );
                    i_decimate_mb += i_decimate_8x8;
                    if( i_decimate_8x8 < 4 )
                    {
                        memset( h->dct.luma8x8[idx], 0, sizeof( h->dct.luma8x8[idx] ) );
                        memset( dct8x8[idx], 0, sizeof( dct8x8[idx] ) );
481
                        nnz8x8[idx] = 0;
Loren Merritt's avatar
Loren Merritt committed
482
                    }
483
                }
484
                else
485
                    nnz8x8[idx] = array_non_zero( dct8x8[idx] );
Laurent Aimar's avatar
Laurent Aimar committed
486 487
            }

488
            if( i_decimate_mb < 6 && b_decimate )
489 490
                memset( h->dct.luma8x8, 0, sizeof( h->dct.luma8x8 ) );
            else
491 492 493 494 495
            {
                for( idx = 0; idx < 4; idx++ )
                    if( nnz8x8[idx] )
                    {
                        h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
496
                        h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
497 498
                    }
            }
499 500 501
        }
        else
        {
502
            DECLARE_ALIGNED( int16_t, dct4x4[16][4][4], 16 );
503
            int nnz8x8[4] = {1,1,1,1};
504
            h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
505 506

            for( i8x8 = 0; i8x8 < 4; i8x8++ )
Laurent Aimar's avatar
Laurent Aimar committed
507
            {
508 509 510 511
                int i_decimate_8x8;

                /* encode one 4x4 block */
                i_decimate_8x8 = 0;
Laurent Aimar's avatar
Laurent Aimar committed
512 513 514
                for( i4x4 = 0; i4x4 < 4; i4x4++ )
                {
                    idx = i8x8 * 4 + i4x4;
515

516 517
                    if( h->mb.b_noise_reduction )
                        x264_denoise_dct( h, (int16_t*)dct4x4[idx] );
Loren Merritt's avatar
Loren Merritt committed
518 519 520 521 522
                    if( h->mb.b_trellis )
                        x264_quant_4x4_trellis( h, dct4x4[idx], CQM_4PY, i_qp, DCT_LUMA_4x4, 0 );
                    else
                        quant_4x4( h, dct4x4[idx], h->quant4_mf[CQM_4PY], i_qp, 0 );

523
                    h->zigzagf.scan_4x4( h->dct.block[idx].luma4x4, dct4x4[idx] );
524 525 526
                    
                    if( b_decimate )
                        i_decimate_8x8 += x264_mb_decimate_score( h->dct.block[idx].luma4x4, 16 );
Laurent Aimar's avatar
Laurent Aimar committed
527 528
                }

529 530
                /* decimate this 8x8 block */
                i_decimate_mb += i_decimate_8x8;
531
                if( i_decimate_8x8 < 4 && b_decimate )
Laurent Aimar's avatar
Laurent Aimar committed
532
                {
Loren Merritt's avatar
Loren Merritt committed
533 534
                    memset( &dct4x4[i8x8*4], 0, 4 * sizeof( *dct4x4 ) );
                    memset( &h->dct.block[i8x8*4], 0, 4 * sizeof( *h->dct.block ) );
535
                    nnz8x8[i8x8] = 0;
Laurent Aimar's avatar
Laurent Aimar committed
536 537
                }
            }
538

539
            if( i_decimate_mb < 6 && b_decimate )
Loren Merritt's avatar
Loren Merritt committed
540
                memset( h->dct.block, 0, 16 * sizeof( *h->dct.block ) );
541
            else
542 543 544 545 546 547
            {
                for( i8x8 = 0; i8x8 < 4; i8x8++ )
                    if( nnz8x8[i8x8] )
                    {
                        for( i = 0; i < 4; i++ )
                            h->quantf.dequant_4x4( dct4x4[i8x8*4+i], h->dequant4_mf[CQM_4PY], i_qp );
548
                        h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
549 550
                    }
            }
Laurent Aimar's avatar
Laurent Aimar committed
551 552 553 554 555 556 557
        }
    }

    /* encode chroma */
    if( IS_INTRA( h->mb.i_type ) )
    {
        const int i_mode = h->mb.i_chroma_pred_mode;
558 559
        h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
        h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
Laurent Aimar's avatar
Laurent Aimar committed
560 561 562
    }

    /* encode the 8x8 blocks */
563
    x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
Laurent Aimar's avatar
Laurent Aimar committed
564

565
    /* coded block pattern and non_zero_count */
566
    h->mb.i_cbp_luma = 0x00;
Laurent Aimar's avatar
Laurent Aimar committed
567 568 569 570 571 572 573 574
    if( h->mb.i_type == I_16x16 )
    {
        for( i = 0; i < 16; i++ )
        {
            const int nz = array_non_zero_count( h->dct.block[i].residual_ac, 15 );
            h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
            if( nz > 0 )
                h->mb.i_cbp_luma = 0x0f;
575 576 577 578
        }
    }
    else if( h->mb.b_transform_8x8 )
    {
Loren Merritt's avatar
Loren Merritt committed
579 580
        /* coded_block_flag is enough for CABAC.
         * the full non_zero_count is done only in CAVLC. */
581 582
        for( i = 0; i < 4; i++ )
        {
583
            const int nz = array_non_zero( h->dct.luma8x8[i] );
584 585 586 587 588
            int j;
            for( j = 0; j < 4; j++ )
                h->mb.cache.non_zero_count[x264_scan8[4*i+j]] = nz;
            if( nz > 0 )
                h->mb.i_cbp_luma |= 1 << i;
Laurent Aimar's avatar
Laurent Aimar committed
589 590 591 592 593 594 595 596 597 598 599 600 601 602 603
        }
    }
    else
    {
        for( i = 0; i < 16; i++ )
        {
            const int nz = array_non_zero_count( h->dct.block[i].luma4x4, 16 );
            h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
            if( nz > 0 )
                h->mb.i_cbp_luma |= 1 << (i/4);
        }
    }

    if( h->param.b_cabac )
    {
604 605 606
        i_cbp_dc = ( h->mb.i_type == I_16x16 && array_non_zero( h->dct.luma16x16_dc ) )
                 | array_non_zero( h->dct.chroma_dc[0] ) << 1
                 | array_non_zero( h->dct.chroma_dc[1] ) << 2;
Laurent Aimar's avatar
Laurent Aimar committed
607 608 609 610 611 612 613 614
    }

    /* store cbp */
    h->mb.cbp[h->mb.i_mb_xy] = (i_cbp_dc << 8) | (h->mb.i_cbp_chroma << 4) | h->mb.i_cbp_luma;

    /* Check for P_SKIP
     * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
     *      (if multiple mv give same result)*/
615
    if( !b_force_no_skip )
Laurent Aimar's avatar
Laurent Aimar committed
616
    {
617
        if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
618 619 620
            h->mb.i_cbp_luma == 0x00 && h->mb.i_cbp_chroma == 0x00 &&
            h->mb.cache.mv[0][x264_scan8[0]][0] == h->mb.cache.pskip_mv[0] &&
            h->mb.cache.mv[0][x264_scan8[0]][1] == h->mb.cache.pskip_mv[1] &&
621
            h->mb.cache.ref[0][x264_scan8[0]] == 0 )
622
        {
623
            h->mb.i_type = P_SKIP;
Laurent Aimar's avatar
Laurent Aimar committed
624
        }
625

626 627 628 629 630 631
        /* Check for B_SKIP */
        if( h->mb.i_type == B_DIRECT &&
            h->mb.i_cbp_luma == 0x00 && h->mb.i_cbp_chroma== 0x00 )
        {
            h->mb.i_type = B_SKIP;
        }
632
    }
Laurent Aimar's avatar
Laurent Aimar committed
633 634 635
}

/*****************************************************************************
636 637
 * x264_macroblock_probe_skip:
 *  Check if the current MB could be encoded as a [PB]_SKIP (it supposes you use
Laurent Aimar's avatar
Laurent Aimar committed
638 639
 *  the previous QP
 *****************************************************************************/
640
int x264_macroblock_probe_skip( x264_t *h, const int b_bidir )
Laurent Aimar's avatar
Laurent Aimar committed
641 642 643 644 645
{
    DECLARE_ALIGNED( int16_t, dct4x4[16][4][4], 16 );
    DECLARE_ALIGNED( int16_t, dct2x2[2][2], 16 );
    DECLARE_ALIGNED( int,     dctscan[16], 16 );

646
    int i_qp = h->mb.i_qp;
Laurent Aimar's avatar
Laurent Aimar committed
647 648 649 650 651 652
    int mvp[2];
    int ch;

    int i8x8, i4x4;
    int i_decimate_mb;

653 654 655
    if( !b_bidir )
    {
        /* Get the MV */
656 657
        mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
        mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
Laurent Aimar's avatar
Laurent Aimar committed
658

659
        /* Motion compensation */
660
        h->mc.mc_luma( h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
661 662
                       h->mb.pic.p_fdec[0],    FDEC_STRIDE,
                       mvp[0], mvp[1], 16, 16 );
663
    }
Laurent Aimar's avatar
Laurent Aimar committed
664 665

    /* get luma diff */
666 667
    h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0],
                                  h->mb.pic.p_fdec[0] );
Laurent Aimar's avatar
Laurent Aimar committed
668 669 670 671 672 673 674 675

    for( i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
    {
        /* encode one 4x4 block */
        for( i4x4 = 0; i4x4 < 4; i4x4++ )
        {
            const int idx = i8x8 * 4 + i4x4;

676
            quant_4x4( h, dct4x4[idx], h->quant4_mf[CQM_4PY], i_qp, 0 );
677
            h->zigzagf.scan_4x4( dctscan, dct4x4[idx] );
Laurent Aimar's avatar
Laurent Aimar committed
678 679 680 681 682 683 684 685 686 687 688 689

            i_decimate_mb += x264_mb_decimate_score( dctscan, 16 );

            if( i_decimate_mb >= 6 )
            {
                /* not as P_SKIP */
                return 0;
            }
        }
    }

    /* encode chroma */
690
    i_qp = h->mb.i_chroma_qp;
Laurent Aimar's avatar
Laurent Aimar committed
691 692 693 694 695 696

    for( ch = 0; ch < 2; ch++ )
    {
        uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
        uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];

697 698
        if( !b_bidir )
        {
699 700 701
            h->mc.mc_chroma( h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
                             h->mb.pic.p_fdec[1+ch],       FDEC_STRIDE,
                             mvp[0], mvp[1], 8, 8 );
702
        }
Laurent Aimar's avatar
Laurent Aimar committed
703

704
        h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
Laurent Aimar's avatar
Laurent Aimar committed
705 706 707 708 709 710 711

        /* calculate dct DC */
        dct2x2[0][0] = dct4x4[0][0][0];
        dct2x2[0][1] = dct4x4[1][0][0];
        dct2x2[1][0] = dct4x4[2][0][0];
        dct2x2[1][1] = dct4x4[3][0][0];
        h->dctf.dct2x2dc( dct2x2 );
712
        quant_2x2_dc( h, dct2x2, h->quant4_mf[CQM_4PC], i_qp, 0 );
Laurent Aimar's avatar
Laurent Aimar committed
713 714 715 716 717 718 719 720 721
        if( dct2x2[0][0] || dct2x2[0][1] || dct2x2[1][0] || dct2x2[1][1]  )
        {
            /* can't be */
            return 0;
        }

        /* calculate dct coeffs */
        for( i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
        {
722
            quant_4x4_chroma( h, dct4x4[i4x4], h->quant4_mf[CQM_4PC], i_qp, 0 );
723
            h->zigzagf.scan_4x4ac( dctscan, dct4x4[i4x4] );
Laurent Aimar's avatar
Laurent Aimar committed
724 725 726 727 728 729 730 731 732 733 734

            i_decimate_mb += x264_mb_decimate_score( dctscan, 15 );
            if( i_decimate_mb >= 7 )
            {
                return 0;
            }
        }
    }

    return 1;
}
735 736 737 738 739 740 741 742 743

/****************************************************************************
 * DCT-domain noise reduction / adaptive deadzone
 * from libavcodec
 ****************************************************************************/

void x264_noise_reduction_update( x264_t *h )
{
    int cat, i;
744
    for( cat = 0; cat < 2; cat++ )
745
    {
746 747
        int size = cat ? 64 : 16;
        const int *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
748

749
        if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765
        {
            for( i = 0; i < size; i++ )
                h->nr_residual_sum[cat][i] >>= 1;
            h->nr_count[cat] >>= 1;
        }

        for( i = 0; i < size; i++ )
            h->nr_offset[cat][i] =
                ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
                 + h->nr_residual_sum[cat][i]/2)
              / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
    }
}

void x264_denoise_dct( x264_t *h, int16_t *dct )
{
766
    const int cat = h->mb.b_transform_8x8;
767 768 769 770
    int i;

    h->nr_count[cat]++;

771
    for( i = (cat ? 63 : 15); i >= 1; i-- )
772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793
    {
        int level = dct[i];
        if( level )
        {
            if( level > 0 )
            {
                h->nr_residual_sum[cat][i] += level;
                level -= h->nr_offset[cat][i];
                if( level < 0 )
                    level = 0;
            }
            else
            {
                h->nr_residual_sum[cat][i] -= level;
                level += h->nr_offset[cat][i];
                if( level > 0 )
                    level = 0;
            }
            dct[i] = level;
        }
    }
}
794 795 796 797 798 799 800 801 802 803

/*****************************************************************************
 * RD only; 4 calls to this do not make up for one macroblock_encode.
 * doesn't transform chroma dc.
 *****************************************************************************/
void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
{
    int i_qp = h->mb.i_qp;
    uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
    uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
804 805
    int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
    int nnz8x8;
806 807 808 809 810 811
    int ch;

    x264_mb_mc_8x8( h, i8 );

    if( h->mb.b_transform_8x8 )
    {
812
        DECLARE_ALIGNED( int16_t, dct8x8[8][8], 16 );
813
        h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
814
        quant_8x8( h, dct8x8, h->quant8_mf[CQM_8PY], i_qp, 0 );
815
        h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
816

817 818 819
        if( b_decimate )
            nnz8x8 = 4 <= x264_mb_decimate_score( h->dct.luma8x8[i8], 64 );
        else
820
            nnz8x8 = array_non_zero( dct8x8 );
821

822 823 824
        if( nnz8x8 )
        {
            h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
825
            h->dctf.add8x8_idct8( p_fdec, dct8x8 );
826 827 828 829
        }
    }
    else
    {
830
        int i4;
831
        DECLARE_ALIGNED( int16_t, dct4x4[4][4][4], 16 );
832
        h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
833 834 835 836
        quant_4x4( h, dct4x4[0], h->quant4_mf[CQM_4PY], i_qp, 0 );
        quant_4x4( h, dct4x4[1], h->quant4_mf[CQM_4PY], i_qp, 0 );
        quant_4x4( h, dct4x4[2], h->quant4_mf[CQM_4PY], i_qp, 0 );
        quant_4x4( h, dct4x4[3], h->quant4_mf[CQM_4PY], i_qp, 0 );
837
        for( i4 = 0; i4 < 4; i4++ )
838
            h->zigzagf.scan_4x4( h->dct.block[i8*4+i4].luma4x4, dct4x4[i4] );
839

840
        if( b_decimate )
841
        {
842 843 844 845
            int i_decimate_8x8 = 0;
            for( i4 = 0; i4 < 4 && i_decimate_8x8 < 4; i4++ )
                i_decimate_8x8 += x264_mb_decimate_score( h->dct.block[i8*4+i4].luma4x4, 16 );
            nnz8x8 = 4 <= i_decimate_8x8;
846
        }
847
        else
848
            nnz8x8 = array_non_zero( dct4x4 );
849

850 851 852
        if( nnz8x8 )
        {
            for( i4 = 0; i4 < 4; i4++ )
853
                h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
854
            h->dctf.add8x8_idct( p_fdec, dct4x4 );
855 856 857
        }
    }

858
    i_qp = h->mb.i_chroma_qp;
859 860 861

    for( ch = 0; ch < 2; ch++ )
    {
862
        DECLARE_ALIGNED( int16_t, dct4x4[4][4], 16 );
863 864 865
        p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
        p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;

866
        h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
867
        quant_4x4_chroma( h, dct4x4, h->quant4_mf[CQM_4PC], i_qp, 0 );
868
        h->zigzagf.scan_4x4ac( h->dct.block[16+i8+ch*4].residual_ac, dct4x4 );
869
        if( array_non_zero( dct4x4 ) )
870 871 872 873
        {
            h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
            h->dctf.add4x4_idct( p_fdec, dct4x4 );
        }
874 875 876 877 878 879 880 881
    }

    if( nnz8x8 )
        h->mb.i_cbp_luma |= (1 << i8);
    else
        h->mb.i_cbp_luma &= ~(1 << i8);
    h->mb.i_cbp_chroma = 0x02;
}