X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=encoder%2Fanalyse.c;h=c1c9314a0a38d2958179e1a63e6aa5c19026a52d;hb=9dd6842dc649734219b1207481c6746bbc6e2198;hp=c052fd8135f79b0d73b68847377c6a7a7fdfaa85;hpb=9ea7b69df504b8990f339e2c8578a516f9df00c7;p=x264 diff --git a/encoder/analyse.c b/encoder/analyse.c index c052fd81..c1c9314a 100644 --- a/encoder/analyse.c +++ b/encoder/analyse.c @@ -24,7 +24,6 @@ #define _ISOC99_SOURCE #include -#include #ifndef _MSC_VER #include #endif @@ -47,7 +46,7 @@ typedef struct /* 8x8 */ int i_cost8x8; /* [ref][0] is 16x16 mv, [ref][1..4] are 8x8 mv from partition [0..3] */ - DECLARE_ALIGNED_4( int16_t mvc[32][5][2] ); + ALIGNED_4( int16_t mvc[32][5][2] ); x264_me_t me8x8[4]; /* Sub 4x4 */ @@ -78,7 +77,7 @@ typedef struct int i_lambda; int i_lambda2; int i_qp; - int16_t *p_cost_mv; + uint16_t *p_cost_mv; uint16_t *p_cost_ref0; uint16_t *p_cost_ref1; int i_mbrd; @@ -156,6 +155,37 @@ const int x264_lambda2_tab[52] = { 943718, 1189010, 1498059, 1887436 /* 48 - 51 */ }; +const uint8_t x264_exp2_lut[64] = { + 1, 4, 7, 10, 13, 16, 19, 22, 25, 28, 31, 34, 37, 40, 44, 47, + 50, 53, 57, 60, 64, 67, 71, 74, 78, 81, 85, 89, 93, 96, 100, 104, + 108, 112, 116, 120, 124, 128, 132, 137, 141, 145, 150, 154, 159, 163, 168, 172, + 177, 182, 186, 191, 196, 201, 206, 211, 216, 221, 226, 232, 237, 242, 248, 253, +}; + +const float x264_log2_lut[128] = { + 0.00000, 0.01123, 0.02237, 0.03342, 0.04439, 0.05528, 0.06609, 0.07682, + 0.08746, 0.09803, 0.10852, 0.11894, 0.12928, 0.13955, 0.14975, 0.15987, + 0.16993, 0.17991, 0.18982, 0.19967, 0.20945, 0.21917, 0.22882, 0.23840, + 0.24793, 0.25739, 0.26679, 0.27612, 0.28540, 0.29462, 0.30378, 0.31288, + 0.32193, 0.33092, 0.33985, 0.34873, 0.35755, 0.36632, 0.37504, 0.38370, + 0.39232, 0.40088, 0.40939, 0.41785, 0.42626, 0.43463, 0.44294, 0.45121, + 0.45943, 0.46761, 0.47573, 0.48382, 0.49185, 0.49985, 0.50779, 0.51570, + 0.52356, 0.53138, 0.53916, 0.54689, 0.55459, 0.56224, 0.56986, 0.57743, + 0.58496, 0.59246, 0.59991, 0.60733, 0.61471, 0.62205, 0.62936, 0.63662, + 0.64386, 0.65105, 0.65821, 0.66534, 0.67243, 0.67948, 0.68650, 0.69349, + 0.70044, 0.70736, 0.71425, 0.72110, 0.72792, 0.73471, 0.74147, 0.74819, + 0.75489, 0.76155, 0.76818, 0.77479, 0.78136, 0.78790, 0.79442, 0.80090, + 0.80735, 0.81378, 0.82018, 0.82655, 0.83289, 0.83920, 0.84549, 0.85175, + 0.85798, 0.86419, 0.87036, 0.87652, 0.88264, 0.88874, 0.89482, 0.90087, + 0.90689, 0.91289, 0.91886, 0.92481, 0.93074, 0.93664, 0.94251, 0.94837, + 0.95420, 0.96000, 0.96578, 0.97154, 0.97728, 0.98299, 0.98868, 0.99435, +}; + +/* Avoid an int/float conversion. */ +const float x264_log2_lz_lut[32] = { + 31,30,29,28,27,26,25,24,23,22,21,20,19,18,17,16,15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0 +}; + // should the intra and inter lambdas be different? // I'm just matching the behaviour of deadzone quant. static const int x264_trellis_lambda2_tab[2][52] = { @@ -207,50 +237,64 @@ static const int i_sub_mb_p_cost_table[4] = { static void x264_analyse_update_cache( x264_t *h, x264_mb_analysis_t *a ); -/* Indexed by lambda instead of qp because, due to rounding, - * some quantizers share lambdas. This saves memory. */ -uint16_t *x264_cost_mv_fpel[92][4]; -uint16_t x264_cost_ref[92][3][33]; +static uint16_t x264_cost_ref[92][3][33]; +static x264_pthread_mutex_t cost_ref_mutex = X264_PTHREAD_MUTEX_INITIALIZER; -/* initialize an array of lambda*nbits for all possible mvs */ -static void x264_mb_analyse_load_costs( x264_t *h, x264_mb_analysis_t *a ) +int x264_analyse_init_costs( x264_t *h, int qp ) { - static int16_t *p_cost_mv[92]; int i, j; - - if( !p_cost_mv[a->i_lambda] ) + int lambda = x264_lambda_tab[qp]; + if( h->cost_mv[lambda] ) + return 0; + /* factor of 4 from qpel, 2 from sign, and 2 because mv can be opposite from mvp */ + CHECKED_MALLOC( h->cost_mv[lambda], (4*4*2048 + 1) * sizeof(uint16_t) ); + h->cost_mv[lambda] += 2*4*2048; + for( i = 0; i <= 2*4*2048; i++ ) + { + h->cost_mv[lambda][-i] = + h->cost_mv[lambda][i] = lambda * (log2f(i+1)*2 + 0.718f + !!i) + .5f; + } + x264_pthread_mutex_lock( &cost_ref_mutex ); + for( i = 0; i < 3; i++ ) + for( j = 0; j < 33; j++ ) + x264_cost_ref[lambda][i][j] = i ? lambda * bs_size_te( i, j ) : 0; + x264_pthread_mutex_unlock( &cost_ref_mutex ); + if( h->param.analyse.i_me_method >= X264_ME_ESA && !h->cost_mv_fpel[lambda][0] ) { - x264_emms(); - /* could be faster, but isn't called many times */ - /* factor of 4 from qpel, 2 from sign, and 2 because mv can be opposite from mvp */ - p_cost_mv[a->i_lambda] = x264_malloc( (4*4*2048 + 1) * sizeof(int16_t) ); - p_cost_mv[a->i_lambda] += 2*4*2048; - for( i = 0; i <= 2*4*2048; i++ ) + for( j=0; j<4; j++ ) { - p_cost_mv[a->i_lambda][-i] = - p_cost_mv[a->i_lambda][i] = a->i_lambda * (log2f(i+1)*2 + 0.718f + !!i) + .5f; + CHECKED_MALLOC( h->cost_mv_fpel[lambda][j], (4*2048 + 1) * sizeof(uint16_t) ); + h->cost_mv_fpel[lambda][j] += 2*2048; + for( i = -2*2048; i < 2*2048; i++ ) + h->cost_mv_fpel[lambda][j][i] = h->cost_mv[lambda][i*4+j]; } - for( i = 0; i < 3; i++ ) - for( j = 0; j < 33; j++ ) - x264_cost_ref[a->i_lambda][i][j] = i ? a->i_lambda * bs_size_te( i, j ) : 0; } - a->p_cost_mv = p_cost_mv[a->i_lambda]; - a->p_cost_ref0 = x264_cost_ref[a->i_lambda][x264_clip3(h->sh.i_num_ref_idx_l0_active-1,0,2)]; - a->p_cost_ref1 = x264_cost_ref[a->i_lambda][x264_clip3(h->sh.i_num_ref_idx_l1_active-1,0,2)]; + return 0; +fail: + return -1; +} - /* FIXME is this useful for all me methods? */ - if( h->param.analyse.i_me_method >= X264_ME_ESA && !x264_cost_mv_fpel[a->i_lambda][0] ) +void x264_analyse_free_costs( x264_t *h ) +{ + int i, j; + for( i = 0; i < 92; i++ ) { - for( j=0; j<4; j++ ) - { - x264_cost_mv_fpel[a->i_lambda][j] = x264_malloc( (4*2048 + 1) * sizeof(int16_t) ); - x264_cost_mv_fpel[a->i_lambda][j] += 2*2048; - for( i = -2*2048; i < 2*2048; i++ ) - x264_cost_mv_fpel[a->i_lambda][j][i] = p_cost_mv[a->i_lambda][i*4+j]; - } + if( h->cost_mv[i] ) + x264_free( h->cost_mv[i] - 2*4*2048 ); + if( h->cost_mv_fpel[i][0] ) + for( j = 0; j < 4; j++ ) + x264_free( h->cost_mv_fpel[i][j] - 2*2048 ); } } +/* initialize an array of lambda*nbits for all possible mvs */ +static void x264_mb_analyse_load_costs( x264_t *h, x264_mb_analysis_t *a ) +{ + a->p_cost_mv = h->cost_mv[a->i_lambda]; + a->p_cost_ref0 = x264_cost_ref[a->i_lambda][x264_clip3(h->sh.i_num_ref_idx_l0_active-1,0,2)]; + a->p_cost_ref1 = x264_cost_ref[a->i_lambda][x264_clip3(h->sh.i_num_ref_idx_l1_active-1,0,2)]; +} + static void x264_mb_analyse_init( x264_t *h, x264_mb_analysis_t *a, int i_qp ) { int i = h->param.analyse.i_subpel_refine - (h->sh.i_type == SLICE_TYPE_B); @@ -276,8 +320,8 @@ static void x264_mb_analyse_init( x264_t *h, x264_mb_analysis_t *a, int i_qp ) h->mb.i_trellis_lambda2[1][1] = x264_trellis_lambda2_tab[1][h->mb.i_chroma_qp]; } h->mb.i_psy_rd_lambda = a->i_lambda; - /* Adjusting chroma lambda based on QP offset hurts PSNR, so we'll leave it as part of psy-RD. */ - h->mb.i_chroma_lambda2_offset = h->mb.i_psy_rd ? x264_chroma_lambda2_offset_tab[h->mb.i_qp-h->mb.i_chroma_qp+12] : 256; + /* Adjusting chroma lambda based on QP offset hurts PSNR but improves visual quality. */ + h->mb.i_chroma_lambda2_offset = h->param.analyse.b_psy ? x264_chroma_lambda2_offset_tab[h->mb.i_qp-h->mb.i_chroma_qp+12] : 256; h->mb.i_me_method = h->param.analyse.i_me_method; h->mb.i_subpel_refine = h->param.analyse.i_subpel_refine; @@ -423,23 +467,30 @@ static void x264_mb_analyse_init( x264_t *h, x264_mb_analysis_t *a, int i_qp ) /* Max = 4 */ static void predict_16x16_mode_available( unsigned int i_neighbour, int *mode, int *pi_count ) { - if( i_neighbour & MB_TOPLEFT ) + int b_top = i_neighbour & MB_TOP; + int b_left = i_neighbour & MB_LEFT; + if( b_top && b_left ) { /* top and left available */ *mode++ = I_PRED_16x16_V; *mode++ = I_PRED_16x16_H; *mode++ = I_PRED_16x16_DC; - *mode++ = I_PRED_16x16_P; - *pi_count = 4; + *pi_count = 3; + if( i_neighbour & MB_TOPLEFT ) + { + /* top left available*/ + *mode++ = I_PRED_16x16_P; + *pi_count = 4; + } } - else if( i_neighbour & MB_LEFT ) + else if( b_left ) { /* left available*/ *mode++ = I_PRED_16x16_DC_LEFT; *mode++ = I_PRED_16x16_H; *pi_count = 2; } - else if( i_neighbour & MB_TOP ) + else if( b_top ) { /* top available*/ *mode++ = I_PRED_16x16_DC_TOP; @@ -457,23 +508,30 @@ static void predict_16x16_mode_available( unsigned int i_neighbour, int *mode, i /* Max = 4 */ static void predict_8x8chroma_mode_available( unsigned int i_neighbour, int *mode, int *pi_count ) { - if( i_neighbour & MB_TOPLEFT ) + int b_top = i_neighbour & MB_TOP; + int b_left = i_neighbour & MB_LEFT; + if( b_top && b_left ) { /* top and left available */ *mode++ = I_PRED_CHROMA_V; *mode++ = I_PRED_CHROMA_H; *mode++ = I_PRED_CHROMA_DC; - *mode++ = I_PRED_CHROMA_P; - *pi_count = 4; + *pi_count = 3; + if( i_neighbour & MB_TOPLEFT ) + { + /* top left available */ + *mode++ = I_PRED_CHROMA_P; + *pi_count = 4; + } } - else if( i_neighbour & MB_LEFT ) + else if( b_left ) { /* left available*/ *mode++ = I_PRED_CHROMA_DC_LEFT; *mode++ = I_PRED_CHROMA_H; *pi_count = 2; } - else if( i_neighbour & MB_TOP ) + else if( b_top ) { /* top available*/ *mode++ = I_PRED_CHROMA_DC_TOP; @@ -492,10 +550,9 @@ static void predict_8x8chroma_mode_available( unsigned int i_neighbour, int *mod static void predict_4x4_mode_available( unsigned int i_neighbour, int *mode, int *pi_count ) { - int b_l = i_neighbour & MB_LEFT; - int b_t = i_neighbour & MB_TOP; - - if( b_l && b_t ) + int b_top = i_neighbour & MB_TOP; + int b_left = i_neighbour & MB_LEFT; + if( b_top && b_left ) { *pi_count = 6; *mode++ = I_PRED_4x4_DC; @@ -512,14 +569,14 @@ static void predict_4x4_mode_available( unsigned int i_neighbour, *mode++ = I_PRED_4x4_VL; *mode++ = I_PRED_4x4_HU; } - else if( b_l ) + else if( b_left ) { *mode++ = I_PRED_4x4_DC_LEFT; *mode++ = I_PRED_4x4_H; *mode++ = I_PRED_4x4_HU; *pi_count = 3; } - else if( b_t ) + else if( b_top ) { *mode++ = I_PRED_4x4_DC_TOP; *mode++ = I_PRED_4x4_V; @@ -537,9 +594,9 @@ static void predict_4x4_mode_available( unsigned int i_neighbour, /* For trellis=2, we need to do this for both sizes of DCT, for trellis=1 we only need to use it on the chosen mode. */ static void inline x264_psy_trellis_init( x264_t *h, int do_both_dct ) { - DECLARE_ALIGNED_16( int16_t dct8x8[4][8][8] ); - DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] ); - DECLARE_ALIGNED_16( static uint8_t zero[16*FDEC_STRIDE] ) = {0}; + ALIGNED_ARRAY_16( int16_t, dct8x8,[4],[8][8] ); + ALIGNED_ARRAY_16( int16_t, dct4x4,[16],[4][4] ); + ALIGNED_16( static uint8_t zero[16*FDEC_STRIDE] ) = {0}; int i; if( do_both_dct || h->mb.b_transform_8x8 ) @@ -559,7 +616,7 @@ static void inline x264_psy_trellis_init( x264_t *h, int do_both_dct ) /* Pre-calculate fenc satd scores for psy RD, minus DC coefficients */ static inline void x264_mb_cache_fenc_satd( x264_t *h ) { - DECLARE_ALIGNED_16( static uint8_t zero[16] ) = {0}; + ALIGNED_16( static uint8_t zero[16] ) = {0}; uint8_t *fenc; int x, y, satd_sum = 0, sa8d_sum = 0; if( h->param.analyse.i_trellis == 2 && h->mb.i_psy_trellis ) @@ -716,7 +773,7 @@ static void x264_mb_analyse_intra( x264_t *h, x264_mb_analysis_t *a, int i_satd_ /* 8x8 prediction selection */ if( flags & X264_ANALYSE_I8x8 ) { - DECLARE_ALIGNED_16( uint8_t edge[33] ); + ALIGNED_ARRAY_16( uint8_t, edge,[33] ); x264_pixel_cmp_t sa8d = (h->pixf.mbcmp[0] == h->pixf.satd[0]) ? h->pixf.sa8d[PIXEL_8x8] : h->pixf.mbcmp[PIXEL_8x8]; int i_satd_thresh = a->i_mbrd ? COST_MAX : X264_MIN( i_satd_inter, a->i_satd_i16x16 ); int i_cost = 0; @@ -1041,7 +1098,7 @@ static void x264_intra_rd_refine( x264_t *h, x264_mb_analysis_t *a ) } else if( h->mb.i_type == I_8x8 ) { - DECLARE_ALIGNED_16( uint8_t edge[33] ); + ALIGNED_ARRAY_16( uint8_t, edge,[33] ); for( idx = 0; idx < 4; idx++ ) { uint64_t pels_h = 0; @@ -1122,7 +1179,7 @@ static void x264_mb_analyse_inter_p16x16( x264_t *h, x264_mb_analysis_t *a ) { x264_me_t m; int i_ref, i_mvc; - DECLARE_ALIGNED_4( int16_t mvc[8][2] ); + ALIGNED_4( int16_t mvc[8][2] ); int i_halfpel_thresh = INT_MAX; int *p_halfpel_thresh = h->mb.pic.i_fref[0]>1 ? &i_halfpel_thresh : NULL; @@ -1183,6 +1240,8 @@ static void x264_mb_analyse_inter_p16x16( x264_t *h, x264_mb_analysis_t *a ) h->mb.i_partition = D_16x16; x264_macroblock_cache_mv_ptr( h, 0, 0, 4, 4, 0, a->l0.me16x16.mv ); a->l0.i_rd16x16 = x264_rd_cost_mb( h, a->i_lambda2 ); + if( !(h->mb.i_cbp_luma|h->mb.i_cbp_chroma) ) + h->mb.i_type = P_SKIP; } } } @@ -1319,7 +1378,7 @@ static void x264_mb_analyse_inter_p16x8( x264_t *h, x264_mb_analysis_t *a ) { x264_me_t m; uint8_t **p_fenc = h->mb.pic.p_fenc; - DECLARE_ALIGNED_4( int16_t mvc[3][2] ); + ALIGNED_4( int16_t mvc[3][2] ); int i, j; /* XXX Needed for x264_mb_predict_mv */ @@ -1369,7 +1428,7 @@ static void x264_mb_analyse_inter_p8x16( x264_t *h, x264_mb_analysis_t *a ) { x264_me_t m; uint8_t **p_fenc = h->mb.pic.p_fenc; - DECLARE_ALIGNED_4( int16_t mvc[3][2] ); + ALIGNED_4( int16_t mvc[3][2] ); int i, j; /* XXX Needed for x264_mb_predict_mv */ @@ -1416,7 +1475,7 @@ static void x264_mb_analyse_inter_p8x16( x264_t *h, x264_mb_analysis_t *a ) static int x264_mb_analyse_inter_p4x4_chroma( x264_t *h, x264_mb_analysis_t *a, uint8_t **p_fref, int i8x8, int pixel ) { - DECLARE_ALIGNED_8( uint8_t pix1[16*8] ); + ALIGNED_8( uint8_t pix1[16*8] ); uint8_t *pix2 = pix1+8; const int i_stride = h->mb.pic.i_stride[1]; const int or = 4*(i8x8&1) + 2*(i8x8&2)*i_stride; @@ -1592,14 +1651,14 @@ static void x264_mb_analyse_inter_direct( x264_t *h, x264_mb_analysis_t *a ) static void x264_mb_analyse_inter_b16x16( x264_t *h, x264_mb_analysis_t *a ) { - DECLARE_ALIGNED_16( uint8_t pix0[16*16] ); - DECLARE_ALIGNED_16( uint8_t pix1[16*16] ); + ALIGNED_ARRAY_16( uint8_t, pix0,[16*16] ); + ALIGNED_ARRAY_16( uint8_t, pix1,[16*16] ); uint8_t *src0, *src1; int stride0 = 16, stride1 = 16; x264_me_t m; int i_ref, i_mvc; - DECLARE_ALIGNED_4( int16_t mvc[9][2] ); + ALIGNED_4( int16_t mvc[9][2] ); int i_halfpel_thresh = INT_MAX; int *p_halfpel_thresh = h->mb.pic.i_fref[0]>1 ? &i_halfpel_thresh : NULL; @@ -1776,7 +1835,7 @@ static void x264_mb_analyse_inter_b8x8( x264_t *h, x264_mb_analysis_t *a ) uint8_t **p_fref[2] = { h->mb.pic.p_fref[0][a->l0.i_ref], h->mb.pic.p_fref[1][a->l1.i_ref] }; - DECLARE_ALIGNED_8( uint8_t pix[2][8*8] ); + ALIGNED_8( uint8_t pix[2][8*8] ); int i, l; /* XXX Needed for x264_mb_predict_mv */ @@ -1841,8 +1900,8 @@ static void x264_mb_analyse_inter_b16x8( x264_t *h, x264_mb_analysis_t *a ) uint8_t **p_fref[2] = { h->mb.pic.p_fref[0][a->l0.i_ref], h->mb.pic.p_fref[1][a->l1.i_ref] }; - DECLARE_ALIGNED_16( uint8_t pix[2][16*8] ); - DECLARE_ALIGNED_4( int16_t mvc[2][2] ); + ALIGNED_ARRAY_16( uint8_t, pix,[2],[16*8] ); + ALIGNED_4( int16_t mvc[2][2] ); int i, l; h->mb.i_partition = D_16x8; @@ -1911,8 +1970,8 @@ static void x264_mb_analyse_inter_b8x16( x264_t *h, x264_mb_analysis_t *a ) uint8_t **p_fref[2] = { h->mb.pic.p_fref[0][a->l0.i_ref], h->mb.pic.p_fref[1][a->l1.i_ref] }; - DECLARE_ALIGNED_8( uint8_t pix[2][8*16] ); - DECLARE_ALIGNED_4( int16_t mvc[2][2] ); + ALIGNED_8( uint8_t pix[2][8*16] ); + ALIGNED_4( int16_t mvc[2][2] ); int i, l; h->mb.i_partition = D_8x16; @@ -1987,7 +2046,6 @@ static void x264_mb_analyse_p_rd( x264_t *h, x264_mb_analysis_t *a, int i_satd ) x264_analyse_update_cache( h, a ); a->l0.i_rd16x16 = x264_rd_cost_mb( h, a->i_lambda2 ); } - a->l0.me16x16.cost = a->l0.i_rd16x16; if( a->l0.i_cost16x8 <= thresh ) { @@ -2128,25 +2186,25 @@ static void x264_refine_bidir( x264_t *h, x264_mb_analysis_t *a ) switch( h->mb.i_partition ) { - case D_16x16: - if( h->mb.i_type == B_BI_BI ) - x264_me_refine_bidir_satd( h, &a->l0.me16x16, &a->l1.me16x16, i_biweight ); - break; - case D_16x8: - for( i=0; i<2; i++ ) - if( a->i_mb_partition16x8[i] == D_BI_8x8 ) - x264_me_refine_bidir_satd( h, &a->l0.me16x8[i], &a->l1.me16x8[i], i_biweight ); - break; - case D_8x16: - for( i=0; i<2; i++ ) - if( a->i_mb_partition8x16[i] == D_BI_8x8 ) - x264_me_refine_bidir_satd( h, &a->l0.me8x16[i], &a->l1.me8x16[i], i_biweight ); - break; - case D_8x8: - for( i=0; i<4; i++ ) - if( h->mb.i_sub_partition[i] == D_BI_8x8 ) - x264_me_refine_bidir_satd( h, &a->l0.me8x8[i], &a->l1.me8x8[i], i_biweight ); - break; + case D_16x16: + if( h->mb.i_type == B_BI_BI ) + x264_me_refine_bidir_satd( h, &a->l0.me16x16, &a->l1.me16x16, i_biweight ); + break; + case D_16x8: + for( i=0; i<2; i++ ) + if( a->i_mb_partition16x8[i] == D_BI_8x8 ) + x264_me_refine_bidir_satd( h, &a->l0.me16x8[i], &a->l1.me16x8[i], i_biweight ); + break; + case D_8x16: + for( i=0; i<2; i++ ) + if( a->i_mb_partition8x16[i] == D_BI_8x8 ) + x264_me_refine_bidir_satd( h, &a->l0.me8x16[i], &a->l1.me8x16[i], i_biweight ); + break; + case D_8x8: + for( i=0; i<4; i++ ) + if( h->mb.i_sub_partition[i] == D_BI_8x8 ) + x264_me_refine_bidir_satd( h, &a->l0.me8x8[i], &a->l1.me8x8[i], i_biweight ); + break; } } @@ -2200,17 +2258,30 @@ static inline void x264_mb_analyse_qp_rd( x264_t *h, x264_mb_analysis_t *a ) { int bcost, cost, direction, failures, prevcost, origcost; int orig_qp = h->mb.i_qp, bqp = h->mb.i_qp; + int last_qp_tried = 0; origcost = bcost = x264_rd_cost_mb( h, a->i_lambda2 ); /* If CBP is already zero, don't raise the quantizer any higher. */ for( direction = h->mb.cbp[h->mb.i_mb_xy] ? 1 : -1; direction >= -1; direction-=2 ) { + /* Without psy-RD, require monotonicity when moving quant away from previous + * macroblock's quant; allow 1 failure when moving quant towards previous quant. + * With psy-RD, allow 1 failure when moving quant away from previous quant, + * allow 2 failures when moving quant towards previous quant. + * Psy-RD generally seems to result in more chaotic RD score-vs-quantizer curves. */ + int threshold = (!!h->mb.i_psy_rd); + /* Raise the threshold for failures if we're moving towards the last QP. */ + if( ( h->mb.i_last_qp < orig_qp && direction == -1 ) || + ( h->mb.i_last_qp > orig_qp && direction == 1 ) ) + threshold++; h->mb.i_qp = orig_qp; failures = 0; prevcost = origcost; - while( h->mb.i_qp > 0 && h->mb.i_qp < 51 ) + h->mb.i_qp += direction; + while( h->mb.i_qp >= h->param.rc.i_qp_min && h->mb.i_qp <= h->param.rc.i_qp_max ) { - h->mb.i_qp += direction; + if( h->mb.i_last_qp == h->mb.i_qp ) + last_qp_tried = 1; h->mb.i_chroma_qp = h->chroma_qp_table[h->mb.i_qp]; cost = x264_rd_cost_mb( h, a->i_lambda2 ); COPY2_IF_LT( bcost, cost, bqp, h->mb.i_qp ); @@ -2223,25 +2294,29 @@ static inline void x264_mb_analyse_qp_rd( x264_t *h, x264_mb_analysis_t *a ) failures++; prevcost = cost; - /* Without psy-RD, require monotonicity when lowering - * quant, allow 1 failure when raising quant. - * With psy-RD, allow 1 failure when lowering quant, - * allow 2 failures when raising quant. - * Psy-RD generally seems to result in more chaotic - * RD score-vs-quantizer curves. */ - if( failures > ((direction + 1)>>1)+(!!h->mb.i_psy_rd) ) + if( failures > threshold ) break; if( direction == 1 && !h->mb.cbp[h->mb.i_mb_xy] ) break; + h->mb.i_qp += direction; } } + /* Always try the last block's QP. */ + if( !last_qp_tried ) + { + h->mb.i_qp = h->mb.i_last_qp; + h->mb.i_chroma_qp = h->chroma_qp_table[h->mb.i_qp]; + cost = x264_rd_cost_mb( h, a->i_lambda2 ); + COPY2_IF_LT( bcost, cost, bqp, h->mb.i_qp ); + } + h->mb.i_qp = bqp; h->mb.i_chroma_qp = h->chroma_qp_table[h->mb.i_qp]; /* Check transform again; decision from before may no longer be optimal. */ - if( h->mb.i_qp != orig_qp && x264_mb_transform_8x8_allowed( h ) && - h->param.analyse.b_transform_8x8 ) + if( h->mb.i_qp != orig_qp && h->param.analyse.b_transform_8x8 && + x264_mb_transform_8x8_allowed( h ) ) { h->mb.b_transform_8x8 ^= 1; cost = x264_rd_cost_mb( h, a->i_lambda2 ); @@ -2482,7 +2557,7 @@ void x264_macroblock_analyse( x264_t *h ) x264_mb_analyse_p_rd( h, &analysis, X264_MIN(i_satd_inter, i_satd_intra) ); i_type = P_L0; i_partition = D_16x16; - i_cost = analysis.l0.me16x16.cost; + i_cost = analysis.l0.i_rd16x16; COPY2_IF_LT( i_cost, analysis.l0.i_cost16x8, i_partition, D_16x8 ); COPY2_IF_LT( i_cost, analysis.l0.i_cost8x16, i_partition, D_8x16 ); COPY3_IF_LT( i_cost, analysis.l0.i_cost8x8, i_partition, D_8x8, i_type, P_8x8 ); @@ -2509,6 +2584,7 @@ void x264_macroblock_analyse( x264_t *h ) else if( i_partition == D_16x16 ) { x264_macroblock_cache_ref( h, 0, 0, 4, 4, 0, analysis.l0.me16x16.i_ref ); + analysis.l0.me16x16.cost = i_cost; x264_me_refine_qpel_rd( h, &analysis.l0.me16x16, analysis.i_lambda2, 0, 0 ); } else if( i_partition == D_16x8 ) @@ -2541,20 +2617,20 @@ void x264_macroblock_analyse( x264_t *h ) } else if( h->mb.i_sub_partition[i8x8] == D_L0_8x4 ) { - x264_me_refine_qpel_rd( h, &analysis.l0.me8x4[i8x8][0], analysis.i_lambda2, i8x8*4+0, 0 ); - x264_me_refine_qpel_rd( h, &analysis.l0.me8x4[i8x8][1], analysis.i_lambda2, i8x8*4+2, 0 ); + x264_me_refine_qpel_rd( h, &analysis.l0.me8x4[i8x8][0], analysis.i_lambda2, i8x8*4+0, 0 ); + x264_me_refine_qpel_rd( h, &analysis.l0.me8x4[i8x8][1], analysis.i_lambda2, i8x8*4+2, 0 ); } else if( h->mb.i_sub_partition[i8x8] == D_L0_4x8 ) { - x264_me_refine_qpel_rd( h, &analysis.l0.me4x8[i8x8][0], analysis.i_lambda2, i8x8*4+0, 0 ); - x264_me_refine_qpel_rd( h, &analysis.l0.me4x8[i8x8][1], analysis.i_lambda2, i8x8*4+1, 0 ); + x264_me_refine_qpel_rd( h, &analysis.l0.me4x8[i8x8][0], analysis.i_lambda2, i8x8*4+0, 0 ); + x264_me_refine_qpel_rd( h, &analysis.l0.me4x8[i8x8][1], analysis.i_lambda2, i8x8*4+1, 0 ); } else if( h->mb.i_sub_partition[i8x8] == D_L0_4x4 ) { - x264_me_refine_qpel_rd( h, &analysis.l0.me4x4[i8x8][0], analysis.i_lambda2, i8x8*4+0, 0 ); - x264_me_refine_qpel_rd( h, &analysis.l0.me4x4[i8x8][1], analysis.i_lambda2, i8x8*4+1, 0 ); - x264_me_refine_qpel_rd( h, &analysis.l0.me4x4[i8x8][2], analysis.i_lambda2, i8x8*4+2, 0 ); - x264_me_refine_qpel_rd( h, &analysis.l0.me4x4[i8x8][3], analysis.i_lambda2, i8x8*4+3, 0 ); + x264_me_refine_qpel_rd( h, &analysis.l0.me4x4[i8x8][0], analysis.i_lambda2, i8x8*4+0, 0 ); + x264_me_refine_qpel_rd( h, &analysis.l0.me4x4[i8x8][1], analysis.i_lambda2, i8x8*4+1, 0 ); + x264_me_refine_qpel_rd( h, &analysis.l0.me4x4[i8x8][2], analysis.i_lambda2, i8x8*4+2, 0 ); + x264_me_refine_qpel_rd( h, &analysis.l0.me4x4[i8x8][3], analysis.i_lambda2, i8x8*4+3, 0 ); } } } @@ -2617,7 +2693,7 @@ void x264_macroblock_analyse( x264_t *h ) const unsigned int flags = h->param.analyse.inter; int i_type; int i_partition; - int i_satd_inter = 0; // shut up uninitialized warning + int i_satd_inter; h->mb.b_skip_mc = 0; x264_mb_analyse_load_costs( h, &analysis ); @@ -2761,9 +2837,10 @@ void x264_macroblock_analyse( x264_t *h ) } } + i_satd_inter = i_cost; + if( analysis.i_mbrd ) { - i_satd_inter = i_cost; x264_mb_analyse_b_rd( h, &analysis, i_satd_inter ); i_type = B_SKIP; i_cost = i_bskip_cost; @@ -2809,9 +2886,15 @@ void x264_macroblock_analyse( x264_t *h ) if( i_partition == D_16x16 ) { if( i_type == B_L0_L0 ) + { + analysis.l0.me16x16.cost = i_cost; x264_me_refine_qpel_rd( h, &analysis.l0.me16x16, analysis.i_lambda2, 0, 0 ); + } else if( i_type == B_L1_L1 ) + { + analysis.l1.me16x16.cost = i_cost; x264_me_refine_qpel_rd( h, &analysis.l1.me16x16, analysis.i_lambda2, 0, 1 ); + } else if( i_type == B_BI_BI ) x264_me_refine_bidir_rd( h, &analysis.l0.me16x16, &analysis.l1.me16x16, i_biweight, 0, analysis.i_lambda2 ); } @@ -2859,6 +2942,19 @@ void x264_macroblock_analyse( x264_t *h ) x264_analyse_update_cache( h, &analysis ); + /* In rare cases we can end up qpel-RDing our way back to a larger partition size + * without realizing it. Check for this and account for it if necessary. */ + if( analysis.i_mbrd >= 2 ) + { + /* Don't bother with bipred or 8x8-and-below, the odds are incredibly low. */ + static const uint8_t check_mv_lists[X264_MBTYPE_MAX] = {[P_L0]=1, [B_L0_L0]=1, [B_L1_L1]=2}; + int list = check_mv_lists[h->mb.i_type] - 1; + if( list >= 0 && h->mb.i_partition != D_16x16 && + *(uint32_t*)&h->mb.cache.mv[list][x264_scan8[0]] == *(uint32_t*)&h->mb.cache.mv[list][x264_scan8[12]] && + h->mb.cache.ref[list][x264_scan8[0]] == h->mb.cache.ref[list][x264_scan8[12]] ) + h->mb.i_partition = D_16x16; + } + if( !analysis.i_mbrd ) x264_mb_analyse_transform( h );