/* conduct the analysis using this lamda and QP */
int i_lambda;
int i_qp;
+ int16_t *p_cost_mv;
/* I: Intra part */
+ /* Take some shortcuts in intra search if intra is deemed unlikely */
+ int b_fast_intra;
+
/* Luma part 16x16 and 4x4 modes stats */
int i_sad_i16x16;
int i_predict16x16;
5, 3, 3, 1
};
+/* 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 )
+{
+ static int16_t *p_cost_mv[52];
+
+ if( !p_cost_mv[a->i_qp] )
+ {
+ /* 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 */
+ int i;
+ p_cost_mv[a->i_qp] = x264_malloc( (4*4*h->param.analyse.i_mv_range + 1) * sizeof(int16_t) );
+ p_cost_mv[a->i_qp] += 2*4*h->param.analyse.i_mv_range;
+ for( i = 0; i <= 2*4*h->param.analyse.i_mv_range; i++ )
+ {
+ p_cost_mv[a->i_qp][-i] =
+ p_cost_mv[a->i_qp][i] = a->i_lambda * bs_size_se( i );
+ }
+ }
+
+ a->p_cost_mv = p_cost_mv[a->i_qp];
+}
+
static void x264_mb_analyse_init( x264_t *h, x264_mb_analysis_t *a, int i_qp )
{
memset( a, 0, sizeof( x264_mb_analysis_t ) );
a->i_qp = i_qp;
a->i_lambda = i_qp0_cost_table[i_qp];
+ h->mb.i_subpel_refine = h->param.analyse.i_subpel_refine;
+ h->mb.b_chroma_me = h->param.analyse.b_chroma_me && h->sh.i_type == SLICE_TYPE_P
+ && h->mb.i_subpel_refine >= 5;
+ a->b_fast_intra = 0;
+
/* I: Intra part */
- a->i_sad_i16x16 = -1;
- a->i_sad_i4x4 = -1;
- a->i_sad_i8x8 = -1;
+ a->i_sad_i16x16 =
+ a->i_sad_i4x4 =
+ a->i_sad_i8x8 = COST_MAX;
/* II: Inter part P/B frame */
if( h->sh.i_type != SLICE_TYPE_I )
{
int i;
+ int i_fmv_range = h->param.analyse.i_mv_range - 16;
/* Calculate max allowed MV range */
- h->mb.mv_min_fpel[0] = -16*h->mb.i_mb_x - 8;
- h->mb.mv_max_fpel[0] = 16*( h->sps->i_mb_width - h->mb.i_mb_x ) - 8;
+#define CLIP_FMV(mv) x264_clip3( mv, -i_fmv_range, i_fmv_range )
+ h->mb.mv_min_fpel[0] = CLIP_FMV( -16*h->mb.i_mb_x - 8 );
+ h->mb.mv_max_fpel[0] = CLIP_FMV( 16*( h->sps->i_mb_width - h->mb.i_mb_x ) - 8 );
h->mb.mv_min[0] = 4*( h->mb.mv_min_fpel[0] - 16 );
h->mb.mv_max[0] = 4*( h->mb.mv_max_fpel[0] + 16 );
if( h->mb.i_mb_x == 0)
{
- h->mb.mv_min_fpel[1] = -16*h->mb.i_mb_y - 8;
- h->mb.mv_max_fpel[1] = 16*( h->sps->i_mb_height - h->mb.i_mb_y ) - 8;
+ h->mb.mv_min_fpel[1] = CLIP_FMV( -16*h->mb.i_mb_y - 8 );
+ h->mb.mv_max_fpel[1] = CLIP_FMV( 16*( h->sps->i_mb_height - h->mb.i_mb_y ) - 8 );
h->mb.mv_min[1] = 4*( h->mb.mv_min_fpel[1] - 16 );
h->mb.mv_max[1] = 4*( h->mb.mv_max_fpel[1] + 16 );
}
+#undef CLIP_FMV
- a->l0.me16x16.cost = -1;
- a->l0.i_cost8x8 = -1;
+ a->l0.me16x16.cost =
+ a->l0.i_cost8x8 = COST_MAX;
for( i = 0; i < 4; i++ )
{
- a->l0.i_cost4x4[i] = -1;
- a->l0.i_cost8x4[i] = -1;
- a->l0.i_cost4x8[i] = -1;
+ a->l0.i_cost4x4[i] =
+ a->l0.i_cost8x4[i] =
+ a->l0.i_cost4x8[i] = COST_MAX;
}
- a->l0.i_cost16x8 = -1;
- a->l0.i_cost8x16 = -1;
+ a->l0.i_cost16x8 =
+ a->l0.i_cost8x16 = COST_MAX;
if( h->sh.i_type == SLICE_TYPE_B )
{
- a->l1.me16x16.cost = -1;
- a->l1.i_cost8x8 = -1;
+ a->l1.me16x16.cost =
+ a->l1.i_cost8x8 = COST_MAX;
for( i = 0; i < 4; i++ )
{
- a->l1.i_cost4x4[i] = -1;
- a->l1.i_cost8x4[i] = -1;
- a->l1.i_cost4x8[i] = -1;
- a->i_cost8x8direct[i] = -1;
+ a->l1.i_cost4x4[i] =
+ a->l1.i_cost8x4[i] =
+ a->l1.i_cost4x8[i] =
+ a->i_cost8x8direct[i] = COST_MAX;
}
- a->l1.i_cost16x8 = -1;
- a->l1.i_cost8x16 = -1;
+ a->l1.i_cost16x8 =
+ a->l1.i_cost8x16 =
+
+ a->i_cost16x16bi =
+ a->i_cost16x16direct =
+ a->i_cost8x8bi =
+ a->i_cost16x8bi =
+ a->i_cost8x16bi = COST_MAX;
+ }
- a->i_cost16x16bi = -1;
- a->i_cost16x16direct = -1;
- a->i_cost8x8bi = -1;
- a->i_cost16x8bi = -1;
- a->i_cost8x16bi = -1;
+ /* Fast intra decision */
+ if( h->mb.i_mb_xy > 4 )
+ {
+ const unsigned int i_neighbour = h->mb.i_neighbour;
+ if( ((i_neighbour&MB_LEFT) && IS_INTRA( h->mb.type[h->mb.i_mb_xy - 1] ))
+ || ((i_neighbour&MB_TOP) && IS_INTRA( h->mb.type[h->mb.i_mb_xy - h->mb.i_mb_stride] ))
+ || (((i_neighbour&(MB_TOP|MB_LEFT)) == (MB_TOP|MB_LEFT)) && IS_INTRA( h->mb.type[h->mb.i_mb_xy - h->mb.i_mb_stride-1 ] ))
+ || ((i_neighbour&MB_TOPRIGHT) && IS_INTRA( h->mb.type[h->mb.i_mb_xy - h->mb.i_mb_stride+1 ] ))
+ || (h->sh.i_type == SLICE_TYPE_P && IS_INTRA( h->fref0[0]->mb_type[h->mb.i_mb_xy] ))
+ || (h->mb.i_mb_xy < 3*(h->stat.frame.i_mb_count[I_4x4] + h->stat.frame.i_mb_count[I_16x16])) )
+ { /* intra is likely */ }
+ else
+ {
+ a->b_fast_intra = 1;
+ }
}
}
}
}
}
-static void x264_mb_analyse_intra( x264_t *h, x264_mb_analysis_t *res )
+static void x264_mb_analyse_intra( x264_t *h, x264_mb_analysis_t *res, int i_cost_inter )
{
const unsigned int flags = h->sh.i_type == SLICE_TYPE_I ? h->param.analyse.intra : h->param.analyse.inter;
const int i_stride = h->mb.pic.i_stride[0];
i_sad = h->pixf.satd[PIXEL_16x16]( p_dst, i_stride, p_src, i_stride ) +
res->i_lambda * bs_size_ue( x264_mb_pred_mode16x16_fix[i_mode] );
/* if i_score is lower it is better */
- if( res->i_sad_i16x16 == -1 || res->i_sad_i16x16 > i_sad )
+ if( res->i_sad_i16x16 > i_sad )
{
res->i_predict16x16 = i_mode;
res->i_sad_i16x16 = i_sad;
if( h->sh.i_type == SLICE_TYPE_B )
res->i_sad_i16x16 += res->i_lambda * i_mb_b_cost_table[I_16x16];
+ if( res->b_fast_intra )
+ {
+ if( res->i_sad_i16x16 > 2*i_cost_inter )
+ return;
+ }
+
/* 4x4 prediction selection */
if( flags & X264_ANALYSE_I4x4 )
{
p_src_by = p_src + 4 * x + 4 * y * i_stride;
p_dst_by = p_dst + 4 * x + 4 * y * i_stride;
- i_best = -1;
+ i_best = COST_MAX;
predict_4x4_mode_available( h->mb.i_neighbour, idx, predict_mode, &i_max );
for( i = 0; i < i_max; i++ )
{
i_sad += res->i_lambda * (i_pred_mode == x264_mb_pred_mode4x4_fix[i_mode] ? 1 : 4);
/* if i_score is lower it is better */
- if( i_best == -1 || i_best > i_sad )
+ if( i_best > i_sad )
{
res->i_predict4x4[x][y] = i_mode;
i_best = i_sad;
uint8_t *p_dstc[2], *p_srcc[2];
int i_stride[2];
+ if( res->i_sad_i8x8 < COST_MAX )
+ return;
+
/* 8x8 prediction selection for chroma */
p_dstc[0] = h->mb.pic.p_fdec[1];
p_dstc[1] = h->mb.pic.p_fdec[2];
i_stride[1] = h->mb.pic.i_stride[2];
predict_8x8_mode_available( h->mb.i_neighbour, predict_mode, &i_max );
- res->i_sad_i8x8 = -1;
+ res->i_sad_i8x8 = COST_MAX;
for( i = 0; i < i_max; i++ )
{
int i_sad;
res->i_lambda * bs_size_ue( x264_mb_pred_mode8x8_fix[i_mode] );
/* if i_score is lower it is better */
- if( res->i_sad_i8x8 == -1 || res->i_sad_i8x8 > i_sad )
+ if( res->i_sad_i8x8 > i_sad )
{
res->i_predict8x8 = i_mode;
res->i_sad_i8x8 = i_sad;
}
}
+#define LOAD_FENC( m, src, xoff, yoff) \
+ (m)->i_stride[0] = h->mb.pic.i_stride[0]; \
+ (m)->i_stride[1] = h->mb.pic.i_stride[1]; \
+ (m)->p_fenc[0] = &(src)[0][(xoff)+(yoff)*(m)->i_stride[0]]; \
+ (m)->p_fenc[1] = &(src)[1][((xoff)>>1)+((yoff)>>1)*(m)->i_stride[1]]; \
+ (m)->p_fenc[2] = &(src)[2][((xoff)>>1)+((yoff)>>1)*(m)->i_stride[1]];
+#define LOAD_HPELS(m, src, xoff, yoff) \
+ (m)->p_fref[0] = &(src)[0][(xoff)+(yoff)*(m)->i_stride[0]]; \
+ (m)->p_fref[1] = &(src)[1][(xoff)+(yoff)*(m)->i_stride[0]]; \
+ (m)->p_fref[2] = &(src)[2][(xoff)+(yoff)*(m)->i_stride[0]]; \
+ (m)->p_fref[3] = &(src)[3][(xoff)+(yoff)*(m)->i_stride[0]]; \
+ (m)->p_fref[4] = &(src)[4][((xoff)>>1)+((yoff)>>1)*(m)->i_stride[1]]; \
+ (m)->p_fref[5] = &(src)[5][((xoff)>>1)+((yoff)>>1)*(m)->i_stride[1]];
+
static void x264_mb_analyse_inter_p16x16( x264_t *h, x264_mb_analysis_t *a )
{
x264_me_t m;
/* 16x16 Search on all ref frame */
m.i_pixel = PIXEL_16x16;
- m.lm = a->i_lambda;
- m.p_fenc = h->mb.pic.p_fenc[0];
- m.i_stride= h->mb.pic.i_stride[0];
+ m.p_cost_mv = a->p_cost_mv;
+ LOAD_FENC( &m, h->mb.pic.p_fenc, 0, 0 );
a->l0.me16x16.cost = INT_MAX;
for( i_ref = 0; i_ref < h->i_ref0; i_ref++ )
{
- const int i_ref_cost = m.lm * bs_size_te( h->sh.i_num_ref_idx_l0_active - 1, i_ref );
+ const int i_ref_cost = a->i_lambda * bs_size_te( h->sh.i_num_ref_idx_l0_active - 1, i_ref );
i_fullpel_thresh -= i_ref_cost;
/* search with ref */
- m.p_fref = h->mb.pic.p_fref[0][i_ref][0];
+ LOAD_HPELS( &m, h->mb.pic.p_fref[0][i_ref], 0, 0 );
x264_mb_predict_mv_16x16( h, 0, i_ref, m.mvp );
x264_mb_predict_mv_ref16x16( h, 0, i_ref, mvc, &i_mvc );
x264_me_search_ref( h, &m, mvc, i_mvc, p_fullpel_thresh );
}
/* subtract ref cost, so we don't have to add it for the other P types */
- a->l0.me16x16.cost -= m.lm * bs_size_te( h->sh.i_num_ref_idx_l0_active - 1, a->l0.i_ref );
+ a->l0.me16x16.cost -= a->i_lambda * bs_size_te( h->sh.i_num_ref_idx_l0_active - 1, a->l0.i_ref );
/* Set global ref, needed for all others modes */
x264_macroblock_cache_ref( h, 0, 0, 4, 4, 0, a->l0.i_ref );
static void x264_mb_analyse_inter_p8x8( x264_t *h, x264_mb_analysis_t *a )
{
- uint8_t *p_fref = h->mb.pic.p_fref[0][a->l0.i_ref][0];
- uint8_t *p_fenc = h->mb.pic.p_fenc[0];
+ uint8_t **p_fref = h->mb.pic.p_fref[0][a->l0.i_ref];
+ uint8_t **p_fenc = h->mb.pic.p_fenc;
int mvc[5][2], i_mvc;
int i;
const int y8 = i/2;
m->i_pixel = PIXEL_8x8;
- m->lm = a->i_lambda;
+ m->p_cost_mv = a->p_cost_mv;
- m->p_fenc = &p_fenc[8*(y8*h->mb.pic.i_stride[0]+x8)];
- m->p_fref = &p_fref[8*(y8*h->mb.pic.i_stride[0]+x8)];
- m->i_stride= h->mb.pic.i_stride[0];
+ LOAD_FENC( m, p_fenc, 8*x8, 8*y8 );
+ LOAD_HPELS( m, p_fref, 8*x8, 8*y8 );
x264_mb_predict_mv( h, 0, 4*i, 2, m->mvp );
x264_me_search( h, m, mvc, i_mvc );
static void x264_mb_analyse_inter_p16x8( x264_t *h, x264_mb_analysis_t *a )
{
- uint8_t *p_fref = h->mb.pic.p_fref[0][a->l0.i_ref][0];
- uint8_t *p_fenc = h->mb.pic.p_fenc[0];
+ uint8_t **p_fref = h->mb.pic.p_fref[0][a->l0.i_ref];
+ uint8_t **p_fenc = h->mb.pic.p_fenc;
int mvc[2][2];
int i;
x264_me_t *m = &a->l0.me16x8[i];
m->i_pixel = PIXEL_16x8;
- m->lm = a->i_lambda;
+ m->p_cost_mv = a->p_cost_mv;
- m->p_fenc = &p_fenc[8*i*h->mb.pic.i_stride[0]];
- m->p_fref = &p_fref[8*i*h->mb.pic.i_stride[0]];
- m->i_stride= h->mb.pic.i_stride[0];
+ LOAD_FENC( m, p_fenc, 0, 8*i );
+ LOAD_HPELS( m, p_fref, 0, 8*i );
mvc[0][0] = a->l0.me8x8[2*i].mv[0];
mvc[0][1] = a->l0.me8x8[2*i].mv[1];
static void x264_mb_analyse_inter_p8x16( x264_t *h, x264_mb_analysis_t *a )
{
- uint8_t *p_fref = h->mb.pic.p_fref[0][a->l0.i_ref][0];
- uint8_t *p_fenc = h->mb.pic.p_fenc[0];
+ uint8_t **p_fref = h->mb.pic.p_fref[0][a->l0.i_ref];
+ uint8_t **p_fenc = h->mb.pic.p_fenc;
int mvc[2][2];
int i;
x264_me_t *m = &a->l0.me8x16[i];
m->i_pixel = PIXEL_8x16;
- m->lm = a->i_lambda;
+ m->p_cost_mv = a->p_cost_mv;
- m->p_fenc = &p_fenc[8*i];
- m->p_fref = &p_fref[8*i];
- m->i_stride= h->mb.pic.i_stride[0];
+ LOAD_FENC( m, p_fenc, 8*i, 0 );
+ LOAD_HPELS( m, p_fref, 8*i, 0 );
mvc[0][0] = a->l0.me8x8[i].mv[0];
mvc[0][1] = a->l0.me8x8[i].mv[1];
a->l0.i_cost8x16 = a->l0.me8x16[0].cost + a->l0.me8x16[1].cost;
}
+static int x264_mb_analyse_inter_p4x4_chroma( x264_t *h, x264_mb_analysis_t *a, uint8_t **p_fref, int i8x8, int pixel )
+{
+ uint8_t pix1[8*8], pix2[8*8];
+ const int i_stride = h->mb.pic.i_stride[1];
+ const int off = 4*(i8x8&1) + 2*(i8x8&2)*i_stride;
+
+#define CHROMA4x4MC( width, height, me, x, y ) \
+ h->mc.mc_chroma( &p_fref[4][off+x+y*i_stride], i_stride, &pix1[x+y*8], 8, (me).mv[0], (me).mv[1], width, height ); \
+ h->mc.mc_chroma( &p_fref[5][off+x+y*i_stride], i_stride, &pix2[x+y*8], 8, (me).mv[0], (me).mv[1], width, height );
+
+ if( pixel == PIXEL_4x4 )
+ {
+ CHROMA4x4MC( 2,2, a->l0.me4x4[i8x8][0], 0,0 );
+ CHROMA4x4MC( 2,2, a->l0.me4x4[i8x8][1], 0,2 );
+ CHROMA4x4MC( 2,2, a->l0.me4x4[i8x8][2], 2,0 );
+ CHROMA4x4MC( 2,2, a->l0.me4x4[i8x8][3], 2,2 );
+ }
+ else if( pixel == PIXEL_8x4 )
+ {
+ CHROMA4x4MC( 4,2, a->l0.me8x4[i8x8][0], 0,0 );
+ CHROMA4x4MC( 4,2, a->l0.me8x4[i8x8][1], 0,2 );
+ }
+ else
+ {
+ CHROMA4x4MC( 2,4, a->l0.me4x8[i8x8][0], 0,0 );
+ CHROMA4x4MC( 2,4, a->l0.me4x8[i8x8][1], 2,0 );
+ }
+
+ return h->pixf.satd[PIXEL_4x4]( &h->mb.pic.p_fenc[1][off], i_stride, pix1, 8 )
+ + h->pixf.satd[PIXEL_4x4]( &h->mb.pic.p_fenc[2][off], i_stride, pix2, 8 );
+}
+
static void x264_mb_analyse_inter_p4x4( x264_t *h, x264_mb_analysis_t *a, int i8x8 )
{
- uint8_t *p_fref = h->mb.pic.p_fref[0][a->l0.i_ref][0];
- uint8_t *p_fenc = h->mb.pic.p_fenc[0];
+ uint8_t **p_fref = h->mb.pic.p_fref[0][a->l0.i_ref];
+ uint8_t **p_fenc = h->mb.pic.p_fenc;
int i4x4;
x264_me_t *m = &a->l0.me4x4[i8x8][i4x4];
m->i_pixel = PIXEL_4x4;
- m->lm = a->i_lambda;
+ m->p_cost_mv = a->p_cost_mv;
- m->p_fenc = &p_fenc[4*(y4*h->mb.pic.i_stride[0]+x4)];
- m->p_fref = &p_fref[4*(y4*h->mb.pic.i_stride[0]+x4)];
- m->i_stride= h->mb.pic.i_stride[0];
+ LOAD_FENC( m, p_fenc, 4*x4, 4*y4 );
+ LOAD_HPELS( m, p_fref, 4*x4, 4*y4 );
x264_mb_predict_mv( h, 0, idx, 1, m->mvp );
x264_me_search( h, m, &a->l0.me8x8[i8x8].mv, i_mvc );
a->l0.me4x4[i8x8][2].cost +
a->l0.me4x4[i8x8][3].cost +
a->i_lambda * i_sub_mb_p_cost_table[D_L0_4x4];
+ if( h->mb.b_chroma_me )
+ a->l0.i_cost4x4[i8x8] += x264_mb_analyse_inter_p4x4_chroma( h, a, p_fref, i8x8, PIXEL_4x4 );
}
static void x264_mb_analyse_inter_p8x4( x264_t *h, x264_mb_analysis_t *a, int i8x8 )
{
- uint8_t *p_fref = h->mb.pic.p_fref[0][a->l0.i_ref][0];
- uint8_t *p_fenc = h->mb.pic.p_fenc[0];
+ uint8_t **p_fref = h->mb.pic.p_fref[0][a->l0.i_ref];
+ uint8_t **p_fenc = h->mb.pic.p_fenc;
int i8x4;
x264_me_t *m = &a->l0.me8x4[i8x8][i8x4];
m->i_pixel = PIXEL_8x4;
- m->lm = a->i_lambda;
+ m->p_cost_mv = a->p_cost_mv;
- m->p_fenc = &p_fenc[4*(y4*h->mb.pic.i_stride[0]+x4)];
- m->p_fref = &p_fref[4*(y4*h->mb.pic.i_stride[0]+x4)];
- m->i_stride= h->mb.pic.i_stride[0];
+ LOAD_FENC( m, p_fenc, 4*x4, 4*y4 );
+ LOAD_HPELS( m, p_fref, 4*x4, 4*y4 );
x264_mb_predict_mv( h, 0, idx, 2, m->mvp );
x264_me_search( h, m, &a->l0.me4x4[i8x8][0].mv, i_mvc );
a->l0.i_cost8x4[i8x8] = a->l0.me8x4[i8x8][0].cost + a->l0.me8x4[i8x8][1].cost +
a->i_lambda * i_sub_mb_p_cost_table[D_L0_8x4];
+ if( h->mb.b_chroma_me )
+ a->l0.i_cost8x4[i8x8] += x264_mb_analyse_inter_p4x4_chroma( h, a, p_fref, i8x8, PIXEL_8x4 );
}
static void x264_mb_analyse_inter_p4x8( x264_t *h, x264_mb_analysis_t *a, int i8x8 )
{
- uint8_t *p_fref = h->mb.pic.p_fref[0][a->l0.i_ref][0];
- uint8_t *p_fenc = h->mb.pic.p_fenc[0];
+ uint8_t **p_fref = h->mb.pic.p_fref[0][a->l0.i_ref];
+ uint8_t **p_fenc = h->mb.pic.p_fenc;
int i4x8;
x264_me_t *m = &a->l0.me4x8[i8x8][i4x8];
m->i_pixel = PIXEL_4x8;
- m->lm = a->i_lambda;
+ m->p_cost_mv = a->p_cost_mv;
- m->p_fenc = &p_fenc[4*(y4*h->mb.pic.i_stride[0]+x4)];
- m->p_fref = &p_fref[4*(y4*h->mb.pic.i_stride[0]+x4)];
- m->i_stride= h->mb.pic.i_stride[0];
+ LOAD_FENC( m, p_fenc, 4*x4, 4*y4 );
+ LOAD_HPELS( m, p_fref, 4*x4, 4*y4 );
x264_mb_predict_mv( h, 0, idx, 1, m->mvp );
x264_me_search( h, m, &a->l0.me4x4[i8x8][0].mv, i_mvc );
a->l0.i_cost4x8[i8x8] = a->l0.me4x8[i8x8][0].cost + a->l0.me4x8[i8x8][1].cost +
a->i_lambda * i_sub_mb_p_cost_table[D_L0_4x8];
+ if( h->mb.b_chroma_me )
+ a->l0.i_cost4x8[i8x8] += x264_mb_analyse_inter_p4x4_chroma( h, a, p_fref, i8x8, PIXEL_4x8 );
}
static void x264_mb_analyse_inter_direct( x264_t *h, x264_mb_analysis_t *a )
/* Assumes that fdec still contains the results of
* x264_mb_predict_mv_direct16x16 and x264_mb_mc */
- uint8_t *p_fenc = h->mb.pic.p_fenc[0];
- uint8_t *p_fdec = h->mb.pic.p_fdec[0];
+ uint8_t **p_fenc = h->mb.pic.p_fenc;
+ uint8_t **p_fdec = h->mb.pic.p_fdec;
int i_stride= h->mb.pic.i_stride[0];
int i;
const int off = 8 * x8 + 8 * i_stride * y8;
a->i_cost16x16direct +=
a->i_cost8x8direct[i] =
- h->pixf.satd[PIXEL_8x8]( &p_fenc[off], i_stride, &p_fdec[off], i_stride );
+ h->pixf.satd[PIXEL_8x8]( &p_fenc[0][off], i_stride, &p_fdec[0][off], i_stride );
/* mb type cost */
a->i_cost8x8direct[i] += a->i_lambda * i_sub_mb_b_cost_table[D_DIRECT_8x8];
a->i_cost16x16direct += a->i_lambda * i_mb_b_cost_table[B_DIRECT];
}
+#define WEIGHTED_AVG( size, pix1, stride1, src2, stride2 ) \
+ { \
+ if( h->param.analyse.b_weighted_bipred ) \
+ h->pixf.avg_weight[size]( pix1, stride1, src2, stride2, \
+ h->mb.bipred_weight[a->l0.i_ref][a->l1.i_ref] ); \
+ else \
+ h->pixf.avg[size]( pix1, stride1, src2, stride2 ); \
+ }
+
static void x264_mb_analyse_inter_b16x16( x264_t *h, x264_mb_analysis_t *a )
{
uint8_t pix1[16*16], pix2[16*16];
+ uint8_t *src2;
+ int stride2 = 16;
+ int src2_ref, pix1_ref;
x264_me_t m;
int i_ref;
/* 16x16 Search on all ref frame */
m.i_pixel = PIXEL_16x16;
- m.lm = a->i_lambda;
- m.p_fenc = h->mb.pic.p_fenc[0];
- m.i_stride= h->mb.pic.i_stride[0];
+ m.p_cost_mv = a->p_cost_mv;
+ LOAD_FENC( &m, h->mb.pic.p_fenc, 0, 0 );
/* ME for List 0 */
a->l0.me16x16.cost = INT_MAX;
for( i_ref = 0; i_ref < h->i_ref0; i_ref++ )
{
/* search with ref */
- m.p_fref = h->mb.pic.p_fref[0][i_ref][0];
+ LOAD_HPELS( &m, h->mb.pic.p_fref[0][i_ref], 0, 0 );
x264_mb_predict_mv_16x16( h, 0, i_ref, m.mvp );
x264_mb_predict_mv_ref16x16( h, 0, i_ref, mvc, &i_mvc );
x264_me_search_ref( h, &m, mvc, i_mvc, p_fullpel_thresh );
/* add ref cost */
- m.cost += m.lm * bs_size_te( h->sh.i_num_ref_idx_l0_active - 1, i_ref );
+ m.cost += a->i_lambda * bs_size_te( h->sh.i_num_ref_idx_l0_active - 1, i_ref );
if( m.cost < a->l0.me16x16.cost )
{
h->mb.mvr[0][i_ref][h->mb.i_mb_xy][1] = m.mv[1];
}
/* subtract ref cost, so we don't have to add it for the other MB types */
- a->l0.me16x16.cost -= m.lm * bs_size_te( h->sh.i_num_ref_idx_l0_active - 1, a->l0.i_ref );
+ a->l0.me16x16.cost -= a->i_lambda * bs_size_te( h->sh.i_num_ref_idx_l0_active - 1, a->l0.i_ref );
/* ME for list 1 */
- /* not using fullpel_thresh since we don't yet do more than 1 list 1 ref */
+ i_fullpel_thresh = INT_MAX;
+ p_fullpel_thresh = h->i_ref1>1 ? &i_fullpel_thresh : NULL;
a->l1.me16x16.cost = INT_MAX;
for( i_ref = 0; i_ref < h->i_ref1; i_ref++ )
{
/* search with ref */
- m.p_fref = h->mb.pic.p_fref[1][i_ref][0];
+ LOAD_HPELS( &m, h->mb.pic.p_fref[1][i_ref], 0, 0 );
x264_mb_predict_mv_16x16( h, 1, i_ref, m.mvp );
x264_mb_predict_mv_ref16x16( h, 1, i_ref, mvc, &i_mvc );
- x264_me_search( h, &m, mvc, i_mvc );
+ x264_me_search_ref( h, &m, mvc, i_mvc, p_fullpel_thresh );
/* add ref cost */
- m.cost += m.lm * bs_size_te( h->sh.i_num_ref_idx_l1_active - 1, i_ref );
+ m.cost += a->i_lambda * bs_size_te( h->sh.i_num_ref_idx_l1_active - 1, i_ref );
if( m.cost < a->l1.me16x16.cost )
{
h->mb.mvr[1][i_ref][h->mb.i_mb_xy][1] = m.mv[1];
}
/* subtract ref cost, so we don't have to add it for the other MB types */
- a->l1.me16x16.cost -= m.lm * bs_size_te( h->sh.i_num_ref_idx_l1_active - 1, a->l1.i_ref );
+ a->l1.me16x16.cost -= a->i_lambda * bs_size_te( h->sh.i_num_ref_idx_l1_active - 1, a->l1.i_ref );
/* Set global ref, needed for other modes? */
x264_macroblock_cache_ref( h, 0, 0, 4, 4, 0, a->l0.i_ref );
x264_macroblock_cache_ref( h, 0, 0, 4, 4, 1, a->l1.i_ref );
/* get cost of BI mode */
- h->mc[MC_LUMA]( h->mb.pic.p_fref[0][a->l0.i_ref][0], h->mb.pic.i_stride[0],
- pix1, 16,
- a->l0.me16x16.mv[0], a->l0.me16x16.mv[1],
- 16, 16 );
- h->mc[MC_LUMA]( h->mb.pic.p_fref[1][a->l1.i_ref][0], h->mb.pic.i_stride[0],
- pix2, 16,
- a->l1.me16x16.mv[0], a->l1.me16x16.mv[1],
- 16, 16 );
- h->pixf.avg[PIXEL_16x16]( pix1, 16, pix2, 16 );
-
- a->i_cost16x16bi = h->pixf.satd[PIXEL_16x16]( h->mb.pic.p_fenc[0], h->mb.pic.i_stride[0], pix1, 16 ) +
- a->i_lambda * ( bs_size_te( h->sh.i_num_ref_idx_l0_active - 1, a->l0.i_ref ) +
- bs_size_te( h->sh.i_num_ref_idx_l1_active - 1, a->l1.i_ref ) +
- bs_size_se( a->l0.me16x16.mv[0] - a->l0.me16x16.mvp[0] ) +
- bs_size_se( a->l0.me16x16.mv[1] - a->l0.me16x16.mvp[1] ) +
- bs_size_se( a->l1.me16x16.mv[0] - a->l1.me16x16.mvp[0] ) +
- bs_size_se( a->l1.me16x16.mv[1] - a->l1.me16x16.mvp[1] ) );
+ if ( ((a->l0.me16x16.mv[0] | a->l0.me16x16.mv[1]) & 1) == 0 )
+ {
+ /* l0 reference is halfpel, so get_ref on it will make it faster */
+ src2 = h->mc.get_ref( h->mb.pic.p_fref[0][a->l0.i_ref], h->mb.pic.i_stride[0],
+ pix2, &stride2,
+ a->l0.me16x16.mv[0], a->l0.me16x16.mv[1],
+ 16, 16 );
+ h->mc.mc_luma( h->mb.pic.p_fref[1][a->l1.i_ref], h->mb.pic.i_stride[0],
+ pix1, 16,
+ a->l1.me16x16.mv[0], a->l1.me16x16.mv[1],
+ 16, 16 );
+ src2_ref = a->l0.i_ref;
+ pix1_ref = a->l1.i_ref;
+ }
+ else
+ {
+ /* if l0 was qpel, we'll use get_ref on l1 instead */
+ h->mc.mc_luma( h->mb.pic.p_fref[0][a->l0.i_ref], h->mb.pic.i_stride[0],
+ pix1, 16,
+ a->l0.me16x16.mv[0], a->l0.me16x16.mv[1],
+ 16, 16 );
+ src2 = h->mc.get_ref( h->mb.pic.p_fref[1][a->l1.i_ref], h->mb.pic.i_stride[0],
+ pix2, &stride2,
+ a->l1.me16x16.mv[0], a->l1.me16x16.mv[1],
+ 16, 16 );
+ src2_ref = a->l1.i_ref;
+ pix1_ref = a->l0.i_ref;
+ }
+
+ if( h->param.analyse.b_weighted_bipred )
+ h->pixf.avg_weight[PIXEL_16x16]( pix1, 16, src2, stride2,
+ h->mb.bipred_weight[pix1_ref][src2_ref] );
+ else
+ h->pixf.avg[PIXEL_16x16]( pix1, 16, src2, stride2 );
+
+ a->i_cost16x16bi = h->pixf.satd[PIXEL_16x16]( h->mb.pic.p_fenc[0], h->mb.pic.i_stride[0], pix1, 16 )
+ + a->i_lambda * ( bs_size_te( h->sh.i_num_ref_idx_l0_active - 1, a->l0.i_ref )
+ + bs_size_te( h->sh.i_num_ref_idx_l1_active - 1, a->l1.i_ref ) )
+ + a->l0.me16x16.cost_mv
+ + a->l1.me16x16.cost_mv;
/* mb type cost */
a->i_cost16x16bi += a->i_lambda * i_mb_b_cost_table[B_BI_BI];
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][0],
- h->mb.pic.p_fref[1][a->l1.i_ref][0] };
- uint8_t *p_fenc = h->mb.pic.p_fenc[0];
+ 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] };
uint8_t pix[2][8*8];
int i, l;
{
const int x8 = i%2;
const int y8 = i/2;
- uint8_t *p_fenc_i = &p_fenc[8*(y8*h->mb.pic.i_stride[0]+x8)];
int i_part_cost;
int i_part_cost_bi = 0;
x264_me_t *m = &lX->me8x8[i];
m->i_pixel = PIXEL_8x8;
- m->lm = a->i_lambda;
+ m->p_cost_mv = a->p_cost_mv;
- m->p_fenc = p_fenc_i;
- m->p_fref = &p_fref[l][8*(y8*h->mb.pic.i_stride[0]+x8)];
- m->i_stride = h->mb.pic.i_stride[0];
+ LOAD_FENC( m, h->mb.pic.p_fenc, 8*x8, 8*y8 );
+ LOAD_HPELS( m, p_fref[l], 8*x8, 8*y8 );
x264_mb_predict_mv( h, l, 4*i, 2, m->mvp );
x264_me_search( h, m, &lX->me16x16.mv, 1 );
x264_macroblock_cache_mv( h, 2*x8, 2*y8, 2, 2, l, m->mv[0], m->mv[1] );
/* BI mode */
- h->mc[MC_LUMA]( m->p_fref, m->i_stride, pix[l], 8,
+ h->mc.mc_luma( m->p_fref, m->i_stride[0], pix[l], 8,
m->mv[0], m->mv[1], 8, 8 );
- i_part_cost_bi += a->i_lambda * ( bs_size_se( m->mv[0] - m->mvp[0] ) +
- bs_size_se( m->mv[1] - m->mvp[1] ) );
+ i_part_cost_bi += m->cost_mv;
/* FIXME: ref cost */
}
- h->pixf.avg[PIXEL_8x8]( pix[0], 8, pix[1], 8 );
- i_part_cost_bi += h->pixf.satd[PIXEL_8x8]( p_fenc_i, h->mb.pic.i_stride[0], pix[0], 8 )
+ WEIGHTED_AVG( PIXEL_8x8, pix[0], 8, pix[1], 8 );
+ i_part_cost_bi += h->pixf.satd[PIXEL_8x8]( a->l0.me8x8[i].p_fenc[0], h->mb.pic.i_stride[0], pix[0], 8 )
+ a->i_lambda * i_sub_mb_b_cost_table[D_BI_8x8];
a->l0.me8x8[i].cost += a->i_lambda * i_sub_mb_b_cost_table[D_L0_8x8];
a->l1.me8x8[i].cost += a->i_lambda * i_sub_mb_b_cost_table[D_L1_8x8];
i_part_cost = i_part_cost_bi;
h->mb.i_sub_partition[i] = D_BI_8x8;
}
- if( a->i_cost8x8direct[i] < i_part_cost && a->i_cost8x8direct[i] >= 0)
+ if( a->i_cost8x8direct[i] < i_part_cost )
{
i_part_cost = a->i_cost8x8direct[i];
h->mb.i_sub_partition[i] = D_DIRECT_8x8;
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][0],
- h->mb.pic.p_fref[1][a->l1.i_ref][0] };
- uint8_t *p_fenc = h->mb.pic.p_fenc[0];
- uint8_t pix[2][8*8];
- int i_ref_stride = h->mb.pic.i_stride[0];
+ 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] };
+ uint8_t pix[2][16*8];
int mvc[2][2];
int i, l;
for( i = 0; i < 2; i++ )
{
- uint8_t *p_fenc_i = &p_fenc[8*i*i_ref_stride];
int i_part_cost;
int i_part_cost_bi = 0;
x264_me_t *m = &lX->me16x8[i];
m->i_pixel = PIXEL_16x8;
- m->lm = a->i_lambda;
+ m->p_cost_mv = a->p_cost_mv;
- m->p_fenc = p_fenc_i;
- m->i_stride= i_ref_stride;
- m->p_fref = &p_fref[l][8*i*i_ref_stride];
+ LOAD_FENC( m, h->mb.pic.p_fenc, 0, 8*i );
+ LOAD_HPELS( m, p_fref[l], 0, 8*i );
mvc[0][0] = lX->me8x8[2*i].mv[0];
mvc[0][1] = lX->me8x8[2*i].mv[1];
x264_me_search( h, m, mvc, 2 );
/* BI mode */
- h->mc[MC_LUMA]( m->p_fref, m->i_stride, pix[l], 8,
- m->mv[0], m->mv[1], 8, 8 );
+ h->mc.mc_luma( m->p_fref, m->i_stride[0], pix[l], 16,
+ m->mv[0], m->mv[1], 16, 8 );
/* FIXME: ref cost */
- i_part_cost_bi += a->i_lambda * ( bs_size_se( m->mv[0] - m->mvp[0] ) +
- bs_size_se( m->mv[1] - m->mvp[1] ) );
+ i_part_cost_bi += m->cost_mv;
}
- h->pixf.avg[PIXEL_16x8]( pix[0], 8, pix[1], 8 );
- i_part_cost_bi += h->pixf.satd[PIXEL_16x8]( p_fenc_i, h->mb.pic.i_stride[0], pix[0], 8 );
+ WEIGHTED_AVG( PIXEL_16x8, pix[0], 16, pix[1], 16 );
+ i_part_cost_bi += h->pixf.satd[PIXEL_16x8]( a->l0.me16x8[i].p_fenc[0], h->mb.pic.i_stride[0], pix[0], 16 );
i_part_cost = a->l0.me16x8[i].cost;
a->i_mb_partition16x8[i] = D_L0_8x8; /* not actually 8x8, only the L0 matters */
}
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][0],
- h->mb.pic.p_fref[1][a->l1.i_ref][0] };
- uint8_t *p_fenc = h->mb.pic.p_fenc[0];
- uint8_t pix[2][8*8];
- int i_ref_stride = h->mb.pic.i_stride[0];
+ 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] };
+ uint8_t pix[2][8*16];
int mvc[2][2];
int i, l;
for( i = 0; i < 2; i++ )
{
- uint8_t *p_fenc_i = &p_fenc[8*i];
int i_part_cost;
int i_part_cost_bi = 0;
x264_me_t *m = &lX->me8x16[i];
m->i_pixel = PIXEL_8x16;
- m->lm = a->i_lambda;
+ m->p_cost_mv = a->p_cost_mv;
- m->p_fenc = p_fenc_i;
- m->p_fref = &p_fref[l][8*i];
- m->i_stride= i_ref_stride;
+ LOAD_FENC( m, h->mb.pic.p_fenc, 8*i, 0 );
+ LOAD_HPELS( m, p_fref[l], 8*i, 0 );
mvc[0][0] = lX->me8x8[i].mv[0];
mvc[0][1] = lX->me8x8[i].mv[1];
x264_me_search( h, m, mvc, 2 );
/* BI mode */
- h->mc[MC_LUMA]( m->p_fref, m->i_stride, pix[l], 8,
- m->mv[0], m->mv[1], 8, 8 );
+ h->mc.mc_luma( m->p_fref, m->i_stride[0], pix[l], 8,
+ m->mv[0], m->mv[1], 8, 16 );
/* FIXME: ref cost */
- i_part_cost_bi += a->i_lambda * ( bs_size_se( m->mv[0] - m->mvp[0] ) +
- bs_size_se( m->mv[1] - m->mvp[1] ) );
+ i_part_cost_bi += m->cost_mv;
}
- h->pixf.avg[PIXEL_8x16]( pix[0], 8, pix[1], 8 );
- i_part_cost_bi += h->pixf.satd[PIXEL_8x16]( p_fenc_i, h->mb.pic.i_stride[0], pix[0], 8 );
+ WEIGHTED_AVG( PIXEL_8x16, pix[0], 8, pix[1], 8 );
+ i_part_cost_bi += h->pixf.satd[PIXEL_8x16]( a->l0.me8x16[i].p_fenc[0], h->mb.pic.i_stride[0], pix[0], 8 );
i_part_cost = a->l0.me8x16[i].cost;
a->i_mb_partition8x16[i] = D_L0_8x8;
/*--------------------------- Do the analysis ---------------------------*/
if( h->sh.i_type == SLICE_TYPE_I )
{
- x264_mb_analyse_intra( h, &analysis );
+ x264_mb_analyse_intra( h, &analysis, COST_MAX );
- if( analysis.i_sad_i4x4 >= 0 && analysis.i_sad_i4x4 < analysis.i_sad_i16x16 )
+ if( analysis.i_sad_i4x4 < analysis.i_sad_i16x16 )
h->mb.i_type = I_4x4;
else
h->mb.i_type = I_16x16;
int i_type;
int i_partition;
+ x264_mb_analyse_load_costs( h, &analysis );
+
x264_mb_analyse_inter_p16x16( h, &analysis );
if( flags & X264_ANALYSE_PSUB16x16 )
x264_mb_analyse_inter_p8x8( h, &analysis );
}
}
- x264_mb_analyse_intra( h, &analysis );
+ x264_mb_analyse_intra( h, &analysis, i_cost );
+ if( h->mb.b_chroma_me &&
+ ( analysis.i_sad_i16x16 < i_cost
+ || ( analysis.i_sad_i4x4 < i_cost )))
+ {
+ x264_mb_analyse_intra_chroma( h, &analysis );
+ analysis.i_sad_i16x16 += analysis.i_sad_i8x8;
+ analysis.i_sad_i4x4 += analysis.i_sad_i8x8;
+ }
+
i_intra_type = I_16x16;
i_intra_cost = analysis.i_sad_i16x16;
- if( analysis.i_sad_i4x4 >=0 && analysis.i_sad_i4x4 < i_intra_cost )
+ if( analysis.i_sad_i4x4 < i_intra_cost )
{
i_intra_type = I_4x4;
i_intra_cost = analysis.i_sad_i4x4;
}
- if( i_intra_cost >= 0 && i_intra_cost < i_cost )
+ if( i_intra_cost < i_cost )
{
h->mb.i_type = i_intra_type;
i_cost = i_intra_cost;
int i_partition;
int i_cost;
+ x264_mb_analyse_load_costs( h, &analysis );
+
/* select best inter mode */
/* direct must be first */
if( analysis.b_direct_available )
h->mb.i_type = B_BI_BI;
i_cost = analysis.i_cost16x16bi;
}
- if( analysis.i_cost16x16direct < i_cost && analysis.i_cost16x16direct >= 0 )
+ if( analysis.i_cost16x16direct < i_cost )
{
h->mb.i_type = B_DIRECT;
i_cost = analysis.i_cost16x16direct;
}
/* best intra mode */
- x264_mb_analyse_intra( h, &analysis );
+ x264_mb_analyse_intra( h, &analysis, i_cost );
- if( analysis.i_sad_i16x16 >= 0 && analysis.i_sad_i16x16 < i_cost )
+ if( analysis.i_sad_i16x16 < i_cost )
{
h->mb.i_type = I_16x16;
i_cost = analysis.i_sad_i16x16;
}
- if( analysis.i_sad_i4x4 >=0 && analysis.i_sad_i4x4 < i_cost )
+ if( analysis.i_sad_i4x4 < i_cost )
{
h->mb.i_type = I_4x4;
i_cost = analysis.i_sad_i4x4;
}
}
+#include "slicetype_decision.c"
+