1 /*****************************************************************************
2 * set.c: quantization init
3 *****************************************************************************
4 * Copyright (C) 2005-2016 x264 project
6 * Authors: Loren Merritt <lorenm@u.washington.edu>
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2 of the License, or
11 * (at your option) any later version.
13 * This program is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with this program; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
22 * This program is also available under a commercial proprietary license.
23 * For more information, contact us at licensing@x264.com.
24 *****************************************************************************/
28 #define SHIFT(x,s) ((s)<=0 ? (x)<<-(s) : ((x)+(1<<((s)-1)))>>(s))
29 #define DIV(n,d) (((n) + ((d)>>1)) / (d))
31 static const uint8_t dequant4_scale[6][3] =
40 static const uint16_t quant4_scale[6][3] =
42 { 13107, 8066, 5243 },
43 { 11916, 7490, 4660 },
44 { 10082, 6554, 4194 },
50 static const uint8_t quant8_scan[16] =
52 0,3,4,3, 3,1,5,1, 4,5,2,5, 3,1,5,1
54 static const uint8_t dequant8_scale[6][6] =
56 { 20, 18, 32, 19, 25, 24 },
57 { 22, 19, 35, 21, 28, 26 },
58 { 26, 23, 42, 24, 33, 31 },
59 { 28, 25, 45, 26, 35, 33 },
60 { 32, 28, 51, 30, 40, 38 },
61 { 36, 32, 58, 34, 46, 43 },
63 static const uint16_t quant8_scale[6][6] =
65 { 13107, 11428, 20972, 12222, 16777, 15481 },
66 { 11916, 10826, 19174, 11058, 14980, 14290 },
67 { 10082, 8943, 15978, 9675, 12710, 11985 },
68 { 9362, 8228, 14913, 8931, 11984, 11259 },
69 { 8192, 7346, 13159, 7740, 10486, 9777 },
70 { 7282, 6428, 11570, 6830, 9118, 8640 }
73 int x264_cqm_init( x264_t *h )
75 int def_quant4[6][16];
76 int def_quant8[6][64];
77 int def_dequant4[6][16];
78 int def_dequant8[6][64];
79 int quant4_mf[4][6][16];
80 int quant8_mf[4][6][64];
81 int deadzone[4] = { 32 - h->param.analyse.i_luma_deadzone[1],
82 32 - h->param.analyse.i_luma_deadzone[0],
85 int max_chroma_qp_err = -1;
86 int min_qp_err = QP_MAX+1;
87 int num_8x8_lists = h->sps->i_chroma_format_idc == CHROMA_444 ? 4
88 : h->param.analyse.b_transform_8x8 ? 2 : 0; /* Checkasm may segfault if optimized out by --chroma-format */
90 #define CQM_ALLOC( w, count )\
91 for( int i = 0; i < count; i++ )\
94 int start = w == 8 ? 4 : 0;\
96 for( j = 0; j < i; j++ )\
97 if( !memcmp( h->pps->scaling_list[i+start], h->pps->scaling_list[j+start], size*sizeof(uint8_t) ) )\
101 h-> quant##w##_mf[i] = h-> quant##w##_mf[j];\
102 h->dequant##w##_mf[i] = h->dequant##w##_mf[j];\
103 h->unquant##w##_mf[i] = h->unquant##w##_mf[j];\
107 CHECKED_MALLOC( h-> quant##w##_mf[i], (QP_MAX_SPEC+1)*size*sizeof(udctcoef) );\
108 CHECKED_MALLOC( h->dequant##w##_mf[i], 6*size*sizeof(int) );\
109 CHECKED_MALLOC( h->unquant##w##_mf[i], (QP_MAX_SPEC+1)*size*sizeof(int) );\
111 for( j = 0; j < i; j++ )\
112 if( deadzone[j] == deadzone[i] &&\
113 !memcmp( h->pps->scaling_list[i+start], h->pps->scaling_list[j+start], size*sizeof(uint8_t) ) )\
117 h->quant##w##_bias[i] = h->quant##w##_bias[j];\
118 h->quant##w##_bias0[i] = h->quant##w##_bias0[j];\
122 CHECKED_MALLOC( h->quant##w##_bias[i], (QP_MAX_SPEC+1)*size*sizeof(udctcoef) );\
123 CHECKED_MALLOC( h->quant##w##_bias0[i], (QP_MAX_SPEC+1)*size*sizeof(udctcoef) );\
128 CQM_ALLOC( 8, num_8x8_lists )
130 for( int q = 0; q < 6; q++ )
132 for( int i = 0; i < 16; i++ )
134 int j = (i&1) + ((i>>2)&1);
135 def_dequant4[q][i] = dequant4_scale[q][j];
136 def_quant4[q][i] = quant4_scale[q][j];
138 for( int i = 0; i < 64; i++ )
140 int j = quant8_scan[((i>>1)&12) | (i&3)];
141 def_dequant8[q][i] = dequant8_scale[q][j];
142 def_quant8[q][i] = quant8_scale[q][j];
146 for( int q = 0; q < 6; q++ )
148 for( int i_list = 0; i_list < 4; i_list++ )
149 for( int i = 0; i < 16; i++ )
151 h->dequant4_mf[i_list][q][i] = def_dequant4[q][i] * h->pps->scaling_list[i_list][i];
152 quant4_mf[i_list][q][i] = DIV(def_quant4[q][i] * 16, h->pps->scaling_list[i_list][i]);
154 for( int i_list = 0; i_list < num_8x8_lists; i_list++ )
155 for( int i = 0; i < 64; i++ )
157 h->dequant8_mf[i_list][q][i] = def_dequant8[q][i] * h->pps->scaling_list[4+i_list][i];
158 quant8_mf[i_list][q][i] = DIV(def_quant8[q][i] * 16, h->pps->scaling_list[4+i_list][i]);
161 for( int q = 0; q <= QP_MAX_SPEC; q++ )
164 for( int i_list = 0; i_list < 4; i_list++ )
165 for( int i = 0; i < 16; i++ )
167 h->unquant4_mf[i_list][q][i] = (1ULL << (q/6 + 15 + 8)) / quant4_mf[i_list][q%6][i];
168 h->quant4_mf[i_list][q][i] = j = SHIFT(quant4_mf[i_list][q%6][i], q/6 - 1);
171 min_qp_err = X264_MIN( min_qp_err, q );
174 // round to nearest, unless that would cause the deadzone to be negative
175 h->quant4_bias[i_list][q][i] = X264_MIN( DIV(deadzone[i_list]<<10, j), (1<<15)/j );
176 h->quant4_bias0[i_list][q][i] = (1<<15)/j;
177 if( j > 0xffff && q > max_qp_err && (i_list == CQM_4IY || i_list == CQM_4PY) )
179 if( j > 0xffff && q > max_chroma_qp_err && (i_list == CQM_4IC || i_list == CQM_4PC) )
180 max_chroma_qp_err = q;
182 if( h->param.analyse.b_transform_8x8 )
183 for( int i_list = 0; i_list < num_8x8_lists; i_list++ )
184 for( int i = 0; i < 64; i++ )
186 h->unquant8_mf[i_list][q][i] = (1ULL << (q/6 + 16 + 8)) / quant8_mf[i_list][q%6][i];
187 j = SHIFT(quant8_mf[i_list][q%6][i], q/6);
188 h->quant8_mf[i_list][q][i] = (uint16_t)j;
192 min_qp_err = X264_MIN( min_qp_err, q );
195 h->quant8_bias[i_list][q][i] = X264_MIN( DIV(deadzone[i_list]<<10, j), (1<<15)/j );
196 h->quant8_bias0[i_list][q][i] = (1<<15)/j;
197 if( j > 0xffff && q > max_qp_err && (i_list == CQM_8IY || i_list == CQM_8PY) )
199 if( j > 0xffff && q > max_chroma_qp_err && (i_list == CQM_8IC || i_list == CQM_8PC) )
200 max_chroma_qp_err = q;
204 /* Emergency mode denoising. */
206 CHECKED_MALLOC( h->nr_offset_emergency, sizeof(*h->nr_offset_emergency)*(QP_MAX-QP_MAX_SPEC) );
207 for( int q = 0; q < QP_MAX - QP_MAX_SPEC; q++ )
208 for( int cat = 0; cat < 3 + CHROMA444; cat++ )
211 if( !h->param.analyse.b_transform_8x8 && dct8x8 )
214 int size = dct8x8 ? 64 : 16;
215 udctcoef *nr_offset = h->nr_offset_emergency[q][cat];
216 /* Denoise chroma first (due to h264's chroma QP offset), then luma, then DC. */
217 int dc_threshold = (QP_MAX-QP_MAX_SPEC)*2/3;
218 int luma_threshold = (QP_MAX-QP_MAX_SPEC)*2/3;
219 int chroma_threshold = 0;
221 for( int i = 0; i < size; i++ )
223 int max = (1 << (7 + BIT_DEPTH)) - 1;
224 /* True "emergency mode": remove all DCT coefficients */
225 if( q == QP_MAX - QP_MAX_SPEC - 1 )
231 int thresh = i == 0 ? dc_threshold : cat >= 2 ? chroma_threshold : luma_threshold;
237 double pos = (double)(q-thresh+1) / (QP_MAX - QP_MAX_SPEC - thresh);
239 /* XXX: this math is largely tuned for /dev/random input. */
240 double start = dct8x8 ? h->unquant8_mf[CQM_8PY][QP_MAX_SPEC][i]
241 : h->unquant4_mf[CQM_4PY][QP_MAX_SPEC][i];
242 /* Formula chosen as an exponential scale to vaguely mimic the effects
243 * of a higher quantizer. */
244 double bias = (pow( 2, pos*(QP_MAX - QP_MAX_SPEC)/10. )*0.003-0.003) * start;
245 nr_offset[i] = X264_MIN( bias + 0.5, max );
249 if( !h->mb.b_lossless )
251 while( h->chroma_qp_table[SPEC_QP(h->param.rc.i_qp_min)] <= max_chroma_qp_err )
252 h->param.rc.i_qp_min++;
253 if( min_qp_err <= h->param.rc.i_qp_max )
254 h->param.rc.i_qp_max = min_qp_err-1;
255 if( max_qp_err >= h->param.rc.i_qp_min )
256 h->param.rc.i_qp_min = max_qp_err+1;
257 /* If long level-codes aren't allowed, we need to allow QP high enough to avoid them. */
258 if( !h->param.b_cabac && h->sps->i_profile_idc < PROFILE_HIGH )
259 while( h->chroma_qp_table[SPEC_QP(h->param.rc.i_qp_max)] <= 12 || h->param.rc.i_qp_max <= 12 )
260 h->param.rc.i_qp_max++;
261 if( h->param.rc.i_qp_min > h->param.rc.i_qp_max )
263 x264_log( h, X264_LOG_ERROR, "Impossible QP constraints for CQM (min=%d, max=%d)\n", h->param.rc.i_qp_min, h->param.rc.i_qp_max );
269 x264_cqm_delete( h );
273 #define CQM_DELETE( n, max )\
274 for( int i = 0; i < (max); i++ )\
277 for( j = 0; j < i; j++ )\
278 if( h->quant##n##_mf[i] == h->quant##n##_mf[j] )\
282 x264_free( h-> quant##n##_mf[i] );\
283 x264_free( h->dequant##n##_mf[i] );\
284 x264_free( h->unquant##n##_mf[i] );\
286 for( j = 0; j < i; j++ )\
287 if( h->quant##n##_bias[i] == h->quant##n##_bias[j] )\
291 x264_free( h->quant##n##_bias[i] );\
292 x264_free( h->quant##n##_bias0[i] );\
296 void x264_cqm_delete( x264_t *h )
299 CQM_DELETE( 8, CHROMA444 ? 4 : 2 );
300 x264_free( h->nr_offset_emergency );
303 static int x264_cqm_parse_jmlist( x264_t *h, const char *buf, const char *name,
304 uint8_t *cqm, const uint8_t *jvt, int length )
308 char *p = strstr( buf, name );
311 memset( cqm, 16, length );
316 if( *p == 'U' || *p == 'V' )
319 char *nextvar = strstr( p, "INT" );
321 for( i = 0; i < length && (p = strpbrk( p, " \t\n," )) && (p = strpbrk( p, "0123456789" )); i++ )
324 sscanf( p, "%d", &coef );
325 if( i == 0 && coef == 0 )
327 memcpy( cqm, jvt, length );
330 if( coef < 1 || coef > 255 )
332 x264_log( h, X264_LOG_ERROR, "bad coefficient in list '%s'\n", name );
338 if( (nextvar && p > nextvar) || i != length )
340 x264_log( h, X264_LOG_ERROR, "not enough coefficients in list '%s'\n", name );
347 int x264_cqm_parse_file( x264_t *h, const char *filename )
352 h->param.i_cqm_preset = X264_CQM_CUSTOM;
354 char *buf = x264_slurp_file( filename );
357 x264_log( h, X264_LOG_ERROR, "can't open file '%s'\n", filename );
361 while( (p = strchr( buf, '#' )) != NULL )
362 memset( p, ' ', strcspn( p, "\n" ) );
364 b_error |= x264_cqm_parse_jmlist( h, buf, "INTRA4X4_LUMA", h->param.cqm_4iy, x264_cqm_jvt4i, 16 );
365 b_error |= x264_cqm_parse_jmlist( h, buf, "INTER4X4_LUMA", h->param.cqm_4py, x264_cqm_jvt4p, 16 );
366 b_error |= x264_cqm_parse_jmlist( h, buf, "INTRA4X4_CHROMA", h->param.cqm_4ic, x264_cqm_jvt4i, 16 );
367 b_error |= x264_cqm_parse_jmlist( h, buf, "INTER4X4_CHROMA", h->param.cqm_4pc, x264_cqm_jvt4p, 16 );
368 b_error |= x264_cqm_parse_jmlist( h, buf, "INTRA8X8_LUMA", h->param.cqm_8iy, x264_cqm_jvt8i, 64 );
369 b_error |= x264_cqm_parse_jmlist( h, buf, "INTER8X8_LUMA", h->param.cqm_8py, x264_cqm_jvt8p, 64 );
372 b_error |= x264_cqm_parse_jmlist( h, buf, "INTRA8X8_CHROMA", h->param.cqm_8ic, x264_cqm_jvt8i, 64 );
373 b_error |= x264_cqm_parse_jmlist( h, buf, "INTER8X8_CHROMA", h->param.cqm_8pc, x264_cqm_jvt8p, 64 );