X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=common%2Fset.c;h=bbfd921d6d2e65c90bf3fb17975bac34df29272b;hb=5e0ca9aa4eab5e2cb4b124774c3ecebbc6f1ae35;hp=ecdc750deb317b91f2318a0669eb48a01775d3ec;hpb=5b5b746834be4cb5c5d6ce275aed38e90d39cbd0;p=x264 diff --git a/common/set.c b/common/set.c index ecdc750d..bbfd921d 100644 --- a/common/set.c +++ b/common/set.c @@ -1,7 +1,7 @@ /***************************************************************************** * set.c: quantization init ***************************************************************************** - * Copyright (C) 2005-2010 x264 project + * Copyright (C) 2005-2014 x264 project * * Authors: Loren Merritt * @@ -23,6 +23,7 @@ * For more information, contact us at licensing@x264.com. *****************************************************************************/ +#define _ISOC99_SOURCE #include "common.h" #define SHIFT(x,s) ((s)<=0 ? (x)<<-(s) : ((x)+(1<<((s)-1)))>>(s)) @@ -77,44 +78,56 @@ int x264_cqm_init( x264_t *h ) int def_dequant4[6][16]; int def_dequant8[6][64]; int quant4_mf[4][6][16]; - int quant8_mf[2][6][64]; + int quant8_mf[4][6][64]; int deadzone[4] = { 32 - h->param.analyse.i_luma_deadzone[1], 32 - h->param.analyse.i_luma_deadzone[0], 32 - 11, 32 - 21 }; int max_qp_err = -1; int max_chroma_qp_err = -1; int min_qp_err = QP_MAX+1; + int num_8x8_lists = h->sps->i_chroma_format_idc == CHROMA_444 ? 4 + : h->param.analyse.b_transform_8x8 ? 2 : 0; /* Checkasm may segfault if optimized out by --chroma-format */ - for( int i = 0; i < 6; i++ ) - { - int size = i<4 ? 16 : 64; - int j; - for( j = (i<4 ? 0 : 4); j < i; j++ ) - if( !memcmp( h->pps->scaling_list[i], h->pps->scaling_list[j], size*sizeof(uint8_t) ) ) - break; - if( j < i ) - { - h-> quant4_mf[i] = h-> quant4_mf[j]; - h->dequant4_mf[i] = h->dequant4_mf[j]; - h->unquant4_mf[i] = h->unquant4_mf[j]; - } - else - { - CHECKED_MALLOC( h-> quant4_mf[i], (QP_MAX+1)*size*sizeof(udctcoef) ); - CHECKED_MALLOC( h->dequant4_mf[i], 6*size*sizeof(int) ); - CHECKED_MALLOC( h->unquant4_mf[i], (QP_MAX+1)*size*sizeof(int) ); - } - - for( j = (i<4 ? 0 : 4); j < i; j++ ) - if( deadzone[j&3] == deadzone[i&3] && - !memcmp( h->pps->scaling_list[i], h->pps->scaling_list[j], size*sizeof(uint8_t) ) ) - break; - if( j < i ) - h->quant4_bias[i] = h->quant4_bias[j]; - else - CHECKED_MALLOC( h->quant4_bias[i], (QP_MAX+1)*size*sizeof(udctcoef) ); +#define CQM_ALLOC( w, count )\ + for( int i = 0; i < count; i++ )\ + {\ + int size = w*w;\ + int start = w == 8 ? 4 : 0;\ + int j;\ + for( j = 0; j < i; j++ )\ + if( !memcmp( h->pps->scaling_list[i+start], h->pps->scaling_list[j+start], size*sizeof(uint8_t) ) )\ + break;\ + if( j < i )\ + {\ + h-> quant##w##_mf[i] = h-> quant##w##_mf[j];\ + h->dequant##w##_mf[i] = h->dequant##w##_mf[j];\ + h->unquant##w##_mf[i] = h->unquant##w##_mf[j];\ + }\ + else\ + {\ + CHECKED_MALLOC( h-> quant##w##_mf[i], (QP_MAX_SPEC+1)*size*sizeof(udctcoef) );\ + CHECKED_MALLOC( h->dequant##w##_mf[i], 6*size*sizeof(int) );\ + CHECKED_MALLOC( h->unquant##w##_mf[i], (QP_MAX_SPEC+1)*size*sizeof(int) );\ + }\ + for( j = 0; j < i; j++ )\ + if( deadzone[j] == deadzone[i] &&\ + !memcmp( h->pps->scaling_list[i+start], h->pps->scaling_list[j+start], size*sizeof(uint8_t) ) )\ + break;\ + if( j < i )\ + {\ + h->quant##w##_bias[i] = h->quant##w##_bias[j];\ + h->quant##w##_bias0[i] = h->quant##w##_bias0[j];\ + }\ + else\ + {\ + CHECKED_MALLOC( h->quant##w##_bias[i], (QP_MAX_SPEC+1)*size*sizeof(udctcoef) );\ + CHECKED_MALLOC( h->quant##w##_bias0[i], (QP_MAX_SPEC+1)*size*sizeof(udctcoef) );\ + }\ } + CQM_ALLOC( 4, 4 ) + CQM_ALLOC( 8, num_8x8_lists ) + for( int q = 0; q < 6; q++ ) { for( int i = 0; i < 16; i++ ) @@ -139,14 +152,14 @@ int x264_cqm_init( x264_t *h ) h->dequant4_mf[i_list][q][i] = def_dequant4[q][i] * h->pps->scaling_list[i_list][i]; quant4_mf[i_list][q][i] = DIV(def_quant4[q][i] * 16, h->pps->scaling_list[i_list][i]); } - for( int i_list = 0; i_list < 2; i_list++ ) + for( int i_list = 0; i_list < num_8x8_lists; i_list++ ) for( int i = 0; i < 64; i++ ) { h->dequant8_mf[i_list][q][i] = def_dequant8[q][i] * h->pps->scaling_list[4+i_list][i]; quant8_mf[i_list][q][i] = DIV(def_quant8[q][i] * 16, h->pps->scaling_list[4+i_list][i]); } } - for( int q = 0; q < QP_MAX+1; q++ ) + for( int q = 0; q <= QP_MAX_SPEC; q++ ) { int j; for( int i_list = 0; i_list < 4; i_list++ ) @@ -161,13 +174,14 @@ int x264_cqm_init( x264_t *h ) } // round to nearest, unless that would cause the deadzone to be negative h->quant4_bias[i_list][q][i] = X264_MIN( DIV(deadzone[i_list]<<10, j), (1<<15)/j ); + h->quant4_bias0[i_list][q][i] = (1<<15)/j; if( j > 0xffff && q > max_qp_err && (i_list == CQM_4IY || i_list == CQM_4PY) ) max_qp_err = q; if( j > 0xffff && q > max_chroma_qp_err && (i_list == CQM_4IC || i_list == CQM_4PC) ) max_chroma_qp_err = q; } if( h->param.analyse.b_transform_8x8 ) - for( int i_list = 0; i_list < 2; i_list++ ) + for( int i_list = 0; i_list < num_8x8_lists; i_list++ ) for( int i = 0; i < 64; i++ ) { h->unquant8_mf[i_list][q][i] = (1ULL << (q/6 + 16 + 8)) / quant8_mf[i_list][q%6][i]; @@ -180,19 +194,71 @@ int x264_cqm_init( x264_t *h ) continue; } h->quant8_bias[i_list][q][i] = X264_MIN( DIV(deadzone[i_list]<<10, j), (1<<15)/j ); - if( j > 0xffff && q > max_qp_err ) + h->quant8_bias0[i_list][q][i] = (1<<15)/j; + if( j > 0xffff && q > max_qp_err && (i_list == CQM_8IY || i_list == CQM_8PY) ) max_qp_err = q; + if( j > 0xffff && q > max_chroma_qp_err && (i_list == CQM_8IC || i_list == CQM_8PC) ) + max_chroma_qp_err = q; } } + /* Emergency mode denoising. */ + x264_emms(); + CHECKED_MALLOC( h->nr_offset_emergency, sizeof(*h->nr_offset_emergency)*(QP_MAX-QP_MAX_SPEC) ); + for( int q = 0; q < QP_MAX - QP_MAX_SPEC; q++ ) + for( int cat = 0; cat < 3 + CHROMA444; cat++ ) + { + int dct8x8 = cat&1; + if( !h->param.analyse.b_transform_8x8 && dct8x8 ) + continue; + + int size = dct8x8 ? 64 : 16; + udctcoef *nr_offset = h->nr_offset_emergency[q][cat]; + /* Denoise chroma first (due to h264's chroma QP offset), then luma, then DC. */ + int dc_threshold = (QP_MAX-QP_MAX_SPEC)*2/3; + int luma_threshold = (QP_MAX-QP_MAX_SPEC)*2/3; + int chroma_threshold = 0; + + for( int i = 0; i < size; i++ ) + { + int max = (1 << (7 + BIT_DEPTH)) - 1; + /* True "emergency mode": remove all DCT coefficients */ + if( q == QP_MAX - QP_MAX_SPEC - 1 ) + { + nr_offset[i] = max; + continue; + } + + int thresh = i == 0 ? dc_threshold : cat >= 2 ? chroma_threshold : luma_threshold; + if( q < thresh ) + { + nr_offset[i] = 0; + continue; + } + double pos = (double)(q-thresh+1) / (QP_MAX - QP_MAX_SPEC - thresh); + + /* XXX: this math is largely tuned for /dev/random input. */ + double start = dct8x8 ? h->unquant8_mf[CQM_8PY][QP_MAX_SPEC][i] + : h->unquant4_mf[CQM_4PY][QP_MAX_SPEC][i]; + /* Formula chosen as an exponential scale to vaguely mimic the effects + * of a higher quantizer. */ + double bias = (pow( 2, pos*(QP_MAX - QP_MAX_SPEC)/10. )*0.003-0.003) * start; + nr_offset[i] = X264_MIN( bias + 0.5, max ); + } + } + if( !h->mb.b_lossless ) { - while( h->chroma_qp_table[h->param.rc.i_qp_min] <= max_chroma_qp_err ) + while( h->chroma_qp_table[SPEC_QP(h->param.rc.i_qp_min)] <= max_chroma_qp_err ) h->param.rc.i_qp_min++; if( min_qp_err <= h->param.rc.i_qp_max ) h->param.rc.i_qp_max = min_qp_err-1; if( max_qp_err >= h->param.rc.i_qp_min ) h->param.rc.i_qp_min = max_qp_err+1; + /* If long level-codes aren't allowed, we need to allow QP high enough to avoid them. */ + if( !h->param.b_cabac && h->sps->i_profile_idc < PROFILE_HIGH ) + while( h->chroma_qp_table[SPEC_QP(h->param.rc.i_qp_max)] <= 12 || h->param.rc.i_qp_max <= 12 ) + h->param.rc.i_qp_max++; if( h->param.rc.i_qp_min > h->param.rc.i_qp_max ) { 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 ); @@ -206,7 +272,7 @@ fail: } #define CQM_DELETE( n, max )\ - for( int i = 0; i < max; i++ )\ + for( int i = 0; i < (max); i++ )\ {\ int j;\ for( j = 0; j < i; j++ )\ @@ -222,17 +288,21 @@ fail: if( h->quant##n##_bias[i] == h->quant##n##_bias[j] )\ break;\ if( j == i )\ + {\ x264_free( h->quant##n##_bias[i] );\ + x264_free( h->quant##n##_bias0[i] );\ + }\ } void x264_cqm_delete( x264_t *h ) { CQM_DELETE( 4, 4 ); - CQM_DELETE( 8, 2 ); + CQM_DELETE( 8, CHROMA444 ? 4 : 2 ); + x264_free( h->nr_offset_emergency ); } static int x264_cqm_parse_jmlist( x264_t *h, const char *buf, const char *name, - uint8_t *cqm, const uint8_t *jvt, int length ) + uint8_t *cqm, const uint8_t *jvt, int length ) { int i; @@ -293,11 +363,16 @@ int x264_cqm_parse_file( x264_t *h, const char *filename ) memset( p, ' ', strcspn( p, "\n" ) ); b_error |= x264_cqm_parse_jmlist( h, buf, "INTRA4X4_LUMA", h->param.cqm_4iy, x264_cqm_jvt4i, 16 ); - b_error |= x264_cqm_parse_jmlist( h, buf, "INTRA4X4_CHROMA", h->param.cqm_4ic, x264_cqm_jvt4i, 16 ); b_error |= x264_cqm_parse_jmlist( h, buf, "INTER4X4_LUMA", h->param.cqm_4py, x264_cqm_jvt4p, 16 ); + b_error |= x264_cqm_parse_jmlist( h, buf, "INTRA4X4_CHROMA", h->param.cqm_4ic, x264_cqm_jvt4i, 16 ); b_error |= x264_cqm_parse_jmlist( h, buf, "INTER4X4_CHROMA", h->param.cqm_4pc, x264_cqm_jvt4p, 16 ); b_error |= x264_cqm_parse_jmlist( h, buf, "INTRA8X8_LUMA", h->param.cqm_8iy, x264_cqm_jvt8i, 64 ); b_error |= x264_cqm_parse_jmlist( h, buf, "INTER8X8_LUMA", h->param.cqm_8py, x264_cqm_jvt8p, 64 ); + if( CHROMA444 ) + { + b_error |= x264_cqm_parse_jmlist( h, buf, "INTRA8X8_CHROMA", h->param.cqm_8ic, x264_cqm_jvt8i, 64 ); + b_error |= x264_cqm_parse_jmlist( h, buf, "INTER8X8_CHROMA", h->param.cqm_8pc, x264_cqm_jvt8p, 64 ); + } x264_free( buf ); return b_error;