1 /*****************************************************************************
2 * mc.c: motion compensation
3 *****************************************************************************
4 * Copyright (C) 2003-2012 x264 project
6 * Authors: Laurent Aimar <fenrir@via.ecp.fr>
7 * Loren Merritt <lorenm@u.washington.edu>
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License as published by
11 * the Free Software Foundation; either version 2 of the License, or
12 * (at your option) any later version.
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
23 * This program is also available under a commercial proprietary license.
24 * For more information, contact us at licensing@x264.com.
25 *****************************************************************************/
40 static inline void pixel_avg( pixel *dst, intptr_t i_dst_stride,
41 pixel *src1, intptr_t i_src1_stride,
42 pixel *src2, intptr_t i_src2_stride, int i_width, int i_height )
44 for( int y = 0; y < i_height; y++ )
46 for( int x = 0; x < i_width; x++ )
47 dst[x] = ( src1[x] + src2[x] + 1 ) >> 1;
49 src1 += i_src1_stride;
50 src2 += i_src2_stride;
54 static inline void pixel_avg_wxh( pixel *dst, intptr_t i_dst,
55 pixel *src1, intptr_t i_src1,
56 pixel *src2, intptr_t i_src2, int width, int height )
58 for( int y = 0; y < height; y++ )
60 for( int x = 0; x < width; x++ )
61 dst[x] = ( src1[x] + src2[x] + 1 ) >> 1;
68 /* Implicit weighted bipred only:
69 * assumes log2_denom = 5, offset = 0, weight1 + weight2 = 64 */
70 static inline void pixel_avg_weight_wxh( pixel *dst, intptr_t i_dst,
71 pixel *src1, intptr_t i_src1,
72 pixel *src2, intptr_t i_src2, int width, int height, int i_weight1 )
74 int i_weight2 = 64 - i_weight1;
75 for( int y = 0; y<height; y++, dst += i_dst, src1 += i_src1, src2 += i_src2 )
76 for( int x = 0; x<width; x++ )
77 dst[x] = x264_clip_pixel( (src1[x]*i_weight1 + src2[x]*i_weight2 + (1<<5)) >> 6 );
81 #define PIXEL_AVG_C( name, width, height ) \
82 static void name( pixel *pix1, intptr_t i_stride_pix1, \
83 pixel *pix2, intptr_t i_stride_pix2, \
84 pixel *pix3, intptr_t i_stride_pix3, int weight ) \
87 pixel_avg_wxh( pix1, i_stride_pix1, pix2, i_stride_pix2, pix3, i_stride_pix3, width, height ); \
89 pixel_avg_weight_wxh( pix1, i_stride_pix1, pix2, i_stride_pix2, pix3, i_stride_pix3, width, height, weight ); \
91 PIXEL_AVG_C( pixel_avg_16x16, 16, 16 )
92 PIXEL_AVG_C( pixel_avg_16x8, 16, 8 )
93 PIXEL_AVG_C( pixel_avg_8x16, 8, 16 )
94 PIXEL_AVG_C( pixel_avg_8x8, 8, 8 )
95 PIXEL_AVG_C( pixel_avg_8x4, 8, 4 )
96 PIXEL_AVG_C( pixel_avg_4x16, 4, 16 )
97 PIXEL_AVG_C( pixel_avg_4x8, 4, 8 )
98 PIXEL_AVG_C( pixel_avg_4x4, 4, 4 )
99 PIXEL_AVG_C( pixel_avg_4x2, 4, 2 )
100 PIXEL_AVG_C( pixel_avg_2x8, 2, 8 )
101 PIXEL_AVG_C( pixel_avg_2x4, 2, 4 )
102 PIXEL_AVG_C( pixel_avg_2x2, 2, 2 )
104 static void x264_weight_cache( x264_t *h, x264_weight_t *w )
106 w->weightfn = h->mc.weight;
108 #define opscale(x) dst[x] = x264_clip_pixel( ((src[x] * scale + (1<<(denom - 1))) >> denom) + offset )
109 #define opscale_noden(x) dst[x] = x264_clip_pixel( src[x] * scale + offset )
110 static void mc_weight( pixel *dst, intptr_t i_dst_stride, pixel *src, intptr_t i_src_stride,
111 const x264_weight_t *weight, int i_width, int i_height )
113 int offset = weight->i_offset << (BIT_DEPTH-8);
114 int scale = weight->i_scale;
115 int denom = weight->i_denom;
118 for( int y = 0; y < i_height; y++, dst += i_dst_stride, src += i_src_stride )
119 for( int x = 0; x < i_width; x++ )
124 for( int y = 0; y < i_height; y++, dst += i_dst_stride, src += i_src_stride )
125 for( int x = 0; x < i_width; x++ )
130 #define MC_WEIGHT_C( name, width ) \
131 static void name( pixel *dst, intptr_t i_dst_stride, pixel *src, intptr_t i_src_stride, const x264_weight_t *weight, int height ) \
133 mc_weight( dst, i_dst_stride, src, i_src_stride, weight, width, height );\
136 MC_WEIGHT_C( mc_weight_w20, 20 )
137 MC_WEIGHT_C( mc_weight_w16, 16 )
138 MC_WEIGHT_C( mc_weight_w12, 12 )
139 MC_WEIGHT_C( mc_weight_w8, 8 )
140 MC_WEIGHT_C( mc_weight_w4, 4 )
141 MC_WEIGHT_C( mc_weight_w2, 2 )
143 static weight_fn_t x264_mc_weight_wtab[6] =
152 const x264_weight_t x264_weight_none[3] = { {{0}} };
153 static void mc_copy( pixel *src, intptr_t i_src_stride, pixel *dst, intptr_t i_dst_stride, int i_width, int i_height )
155 for( int y = 0; y < i_height; y++ )
157 memcpy( dst, src, i_width * sizeof(pixel) );
164 #define TAPFILTER(pix, d) ((pix)[x-2*d] + (pix)[x+3*d] - 5*((pix)[x-d] + (pix)[x+2*d]) + 20*((pix)[x] + (pix)[x+d]))
165 static void hpel_filter( pixel *dsth, pixel *dstv, pixel *dstc, pixel *src,
166 intptr_t stride, int width, int height, int16_t *buf )
168 const int pad = (BIT_DEPTH > 9) ? (-10 * PIXEL_MAX) : 0;
169 for( int y = 0; y < height; y++ )
171 for( int x = -2; x < width+3; x++ )
173 int v = TAPFILTER(src,stride);
174 dstv[x] = x264_clip_pixel( (v + 16) >> 5 );
175 /* transform v for storage in a 16-bit integer */
178 for( int x = 0; x < width; x++ )
179 dstc[x] = x264_clip_pixel( (TAPFILTER(buf+2,1) - 32*pad + 512) >> 10 );
180 for( int x = 0; x < width; x++ )
181 dsth[x] = x264_clip_pixel( (TAPFILTER(src,1) + 16) >> 5 );
189 static const uint8_t hpel_ref0[16] = {0,1,1,1,0,1,1,1,2,3,3,3,0,1,1,1};
190 static const uint8_t hpel_ref1[16] = {0,0,0,0,2,2,3,2,2,2,3,2,2,2,3,2};
192 static void mc_luma( pixel *dst, intptr_t i_dst_stride,
193 pixel *src[4], intptr_t i_src_stride,
195 int i_width, int i_height, const x264_weight_t *weight )
197 int qpel_idx = ((mvy&3)<<2) + (mvx&3);
198 int offset = (mvy>>2)*i_src_stride + (mvx>>2);
199 pixel *src1 = src[hpel_ref0[qpel_idx]] + offset + ((mvy&3) == 3) * i_src_stride;
201 if( qpel_idx & 5 ) /* qpel interpolation needed */
203 pixel *src2 = src[hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);
204 pixel_avg( dst, i_dst_stride, src1, i_src_stride,
205 src2, i_src_stride, i_width, i_height );
206 if( weight->weightfn )
207 mc_weight( dst, i_dst_stride, dst, i_dst_stride, weight, i_width, i_height );
209 else if( weight->weightfn )
210 mc_weight( dst, i_dst_stride, src1, i_src_stride, weight, i_width, i_height );
212 mc_copy( src1, i_src_stride, dst, i_dst_stride, i_width, i_height );
215 static pixel *get_ref( pixel *dst, intptr_t *i_dst_stride,
216 pixel *src[4], intptr_t i_src_stride,
218 int i_width, int i_height, const x264_weight_t *weight )
220 int qpel_idx = ((mvy&3)<<2) + (mvx&3);
221 int offset = (mvy>>2)*i_src_stride + (mvx>>2);
222 pixel *src1 = src[hpel_ref0[qpel_idx]] + offset + ((mvy&3) == 3) * i_src_stride;
224 if( qpel_idx & 5 ) /* qpel interpolation needed */
226 pixel *src2 = src[hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);
227 pixel_avg( dst, *i_dst_stride, src1, i_src_stride,
228 src2, i_src_stride, i_width, i_height );
229 if( weight->weightfn )
230 mc_weight( dst, *i_dst_stride, dst, *i_dst_stride, weight, i_width, i_height );
233 else if( weight->weightfn )
235 mc_weight( dst, *i_dst_stride, src1, i_src_stride, weight, i_width, i_height );
240 *i_dst_stride = i_src_stride;
245 /* full chroma mc (ie until 1/8 pixel)*/
246 static void mc_chroma( pixel *dstu, pixel *dstv, intptr_t i_dst_stride,
247 pixel *src, intptr_t i_src_stride,
249 int i_width, int i_height )
255 int cA = (8-d8x)*(8-d8y);
256 int cB = d8x *(8-d8y);
257 int cC = (8-d8x)*d8y;
260 src += (mvy >> 3) * i_src_stride + (mvx >> 3)*2;
261 srcp = &src[i_src_stride];
263 for( int y = 0; y < i_height; y++ )
265 for( int x = 0; x < i_width; x++ )
267 dstu[x] = ( cA*src[2*x] + cB*src[2*x+2] +
268 cC*srcp[2*x] + cD*srcp[2*x+2] + 32 ) >> 6;
269 dstv[x] = ( cA*src[2*x+1] + cB*src[2*x+3] +
270 cC*srcp[2*x+1] + cD*srcp[2*x+3] + 32 ) >> 6;
272 dstu += i_dst_stride;
273 dstv += i_dst_stride;
275 srcp += i_src_stride;
280 static void mc_copy_w##W( pixel *dst, intptr_t i_dst, pixel *src, intptr_t i_src, int i_height ) \
282 mc_copy( src, i_src, dst, i_dst, W, i_height ); \
288 void x264_plane_copy_c( pixel *dst, intptr_t i_dst,
289 pixel *src, intptr_t i_src, int w, int h )
293 memcpy( dst, src, w * sizeof(pixel) );
299 void x264_plane_copy_interleave_c( pixel *dst, intptr_t i_dst,
300 pixel *srcu, intptr_t i_srcu,
301 pixel *srcv, intptr_t i_srcv, int w, int h )
303 for( int y=0; y<h; y++, dst+=i_dst, srcu+=i_srcu, srcv+=i_srcv )
304 for( int x=0; x<w; x++ )
307 dst[2*x+1] = srcv[x];
311 static void x264_plane_copy_deinterleave_c( pixel *dstu, intptr_t i_dstu,
312 pixel *dstv, intptr_t i_dstv,
313 pixel *src, intptr_t i_src, int w, int h )
315 for( int y=0; y<h; y++, dstu+=i_dstu, dstv+=i_dstv, src+=i_src )
316 for( int x=0; x<w; x++ )
319 dstv[x] = src[2*x+1];
323 static void x264_plane_copy_deinterleave_rgb_c( pixel *dsta, intptr_t i_dsta,
324 pixel *dstb, intptr_t i_dstb,
325 pixel *dstc, intptr_t i_dstc,
326 pixel *src, intptr_t i_src, int pw, int w, int h )
328 for( int y=0; y<h; y++, dsta+=i_dsta, dstb+=i_dstb, dstc+=i_dstc, src+=i_src )
330 for( int x=0; x<w; x++ )
333 dstb[x] = src[x*pw+1];
334 dstc[x] = src[x*pw+2];
339 static void store_interleave_chroma( pixel *dst, intptr_t i_dst, pixel *srcu, pixel *srcv, int height )
341 for( int y=0; y<height; y++, dst+=i_dst, srcu+=FDEC_STRIDE, srcv+=FDEC_STRIDE )
342 for( int x=0; x<8; x++ )
345 dst[2*x+1] = srcv[x];
349 static void load_deinterleave_chroma_fenc( pixel *dst, pixel *src, intptr_t i_src, int height )
351 x264_plane_copy_deinterleave_c( dst, FENC_STRIDE, dst+FENC_STRIDE/2, FENC_STRIDE, src, i_src, 8, height );
354 static void load_deinterleave_chroma_fdec( pixel *dst, pixel *src, intptr_t i_src, int height )
356 x264_plane_copy_deinterleave_c( dst, FDEC_STRIDE, dst+FDEC_STRIDE/2, FDEC_STRIDE, src, i_src, 8, height );
359 static void prefetch_fenc_null( pixel *pix_y, intptr_t stride_y,
360 pixel *pix_uv, intptr_t stride_uv, int mb_x )
363 static void prefetch_ref_null( pixel *pix, intptr_t stride, int parity )
366 static void memzero_aligned( void * dst, size_t n )
371 static void integral_init4h( uint16_t *sum, pixel *pix, intptr_t stride )
373 int v = pix[0]+pix[1]+pix[2]+pix[3];
374 for( int x = 0; x < stride-4; x++ )
376 sum[x] = v + sum[x-stride];
377 v += pix[x+4] - pix[x];
381 static void integral_init8h( uint16_t *sum, pixel *pix, intptr_t stride )
383 int v = pix[0]+pix[1]+pix[2]+pix[3]+pix[4]+pix[5]+pix[6]+pix[7];
384 for( int x = 0; x < stride-8; x++ )
386 sum[x] = v + sum[x-stride];
387 v += pix[x+8] - pix[x];
391 static void integral_init4v( uint16_t *sum8, uint16_t *sum4, intptr_t stride )
393 for( int x = 0; x < stride-8; x++ )
394 sum4[x] = sum8[x+4*stride] - sum8[x];
395 for( int x = 0; x < stride-8; x++ )
396 sum8[x] = sum8[x+8*stride] + sum8[x+8*stride+4] - sum8[x] - sum8[x+4];
399 static void integral_init8v( uint16_t *sum8, intptr_t stride )
401 for( int x = 0; x < stride-8; x++ )
402 sum8[x] = sum8[x+8*stride] - sum8[x];
405 void x264_frame_init_lowres( x264_t *h, x264_frame_t *frame )
407 pixel *src = frame->plane[0];
408 int i_stride = frame->i_stride[0];
409 int i_height = frame->i_lines[0];
410 int i_width = frame->i_width[0];
412 // duplicate last row and column so that their interpolation doesn't have to be special-cased
413 for( int y = 0; y < i_height; y++ )
414 src[i_width+y*i_stride] = src[i_width-1+y*i_stride];
415 memcpy( src+i_stride*i_height, src+i_stride*(i_height-1), (i_width+1) * sizeof(pixel) );
416 h->mc.frame_init_lowres_core( src, frame->lowres[0], frame->lowres[1], frame->lowres[2], frame->lowres[3],
417 i_stride, frame->i_stride_lowres, frame->i_width_lowres, frame->i_lines_lowres );
418 x264_frame_expand_border_lowres( frame );
420 memset( frame->i_cost_est, -1, sizeof(frame->i_cost_est) );
422 for( int y = 0; y < h->param.i_bframe + 2; y++ )
423 for( int x = 0; x < h->param.i_bframe + 2; x++ )
424 frame->i_row_satds[y][x][0] = -1;
426 for( int y = 0; y <= !!h->param.i_bframe; y++ )
427 for( int x = 0; x <= h->param.i_bframe; x++ )
428 frame->lowres_mvs[y][x][0][0] = 0x7FFF;
431 static void frame_init_lowres_core( pixel *src0, pixel *dst0, pixel *dsth, pixel *dstv, pixel *dstc,
432 intptr_t src_stride, intptr_t dst_stride, int width, int height )
434 for( int y = 0; y < height; y++ )
436 pixel *src1 = src0+src_stride;
437 pixel *src2 = src1+src_stride;
438 for( int x = 0; x<width; x++ )
440 // slower than naive bilinear, but matches asm
441 #define FILTER(a,b,c,d) ((((a+b+1)>>1)+((c+d+1)>>1)+1)>>1)
442 dst0[x] = FILTER(src0[2*x ], src1[2*x ], src0[2*x+1], src1[2*x+1]);
443 dsth[x] = FILTER(src0[2*x+1], src1[2*x+1], src0[2*x+2], src1[2*x+2]);
444 dstv[x] = FILTER(src1[2*x ], src2[2*x ], src1[2*x+1], src2[2*x+1]);
445 dstc[x] = FILTER(src1[2*x+1], src2[2*x+1], src1[2*x+2], src2[2*x+2]);
448 src0 += src_stride*2;
456 /* Estimate the total amount of influence on future quality that could be had if we
457 * were to improve the reference samples used to inter predict any given macroblock. */
458 static void mbtree_propagate_cost( int *dst, uint16_t *propagate_in, uint16_t *intra_costs,
459 uint16_t *inter_costs, uint16_t *inv_qscales, float *fps_factor, int len )
461 float fps = *fps_factor / 256.f;
462 for( int i = 0; i < len; i++ )
464 float intra_cost = intra_costs[i] * inv_qscales[i];
465 float propagate_amount = propagate_in[i] + intra_cost*fps;
466 float propagate_num = intra_costs[i] - (inter_costs[i] & LOWRES_COST_MASK);
467 float propagate_denom = intra_costs[i];
468 dst[i] = (int)(propagate_amount * propagate_num / propagate_denom + 0.5f);
472 void x264_mc_init( int cpu, x264_mc_functions_t *pf )
474 pf->mc_luma = mc_luma;
475 pf->get_ref = get_ref;
477 pf->mc_chroma = mc_chroma;
479 pf->avg[PIXEL_16x16]= pixel_avg_16x16;
480 pf->avg[PIXEL_16x8] = pixel_avg_16x8;
481 pf->avg[PIXEL_8x16] = pixel_avg_8x16;
482 pf->avg[PIXEL_8x8] = pixel_avg_8x8;
483 pf->avg[PIXEL_8x4] = pixel_avg_8x4;
484 pf->avg[PIXEL_4x16] = pixel_avg_4x16;
485 pf->avg[PIXEL_4x8] = pixel_avg_4x8;
486 pf->avg[PIXEL_4x4] = pixel_avg_4x4;
487 pf->avg[PIXEL_4x2] = pixel_avg_4x2;
488 pf->avg[PIXEL_2x8] = pixel_avg_2x8;
489 pf->avg[PIXEL_2x4] = pixel_avg_2x4;
490 pf->avg[PIXEL_2x2] = pixel_avg_2x2;
492 pf->weight = x264_mc_weight_wtab;
493 pf->offsetadd = x264_mc_weight_wtab;
494 pf->offsetsub = x264_mc_weight_wtab;
495 pf->weight_cache = x264_weight_cache;
497 pf->copy_16x16_unaligned = mc_copy_w16;
498 pf->copy[PIXEL_16x16] = mc_copy_w16;
499 pf->copy[PIXEL_8x8] = mc_copy_w8;
500 pf->copy[PIXEL_4x4] = mc_copy_w4;
502 pf->store_interleave_chroma = store_interleave_chroma;
503 pf->load_deinterleave_chroma_fenc = load_deinterleave_chroma_fenc;
504 pf->load_deinterleave_chroma_fdec = load_deinterleave_chroma_fdec;
506 pf->plane_copy = x264_plane_copy_c;
507 pf->plane_copy_interleave = x264_plane_copy_interleave_c;
508 pf->plane_copy_deinterleave = x264_plane_copy_deinterleave_c;
509 pf->plane_copy_deinterleave_rgb = x264_plane_copy_deinterleave_rgb_c;
511 pf->hpel_filter = hpel_filter;
513 pf->prefetch_fenc_420 = prefetch_fenc_null;
514 pf->prefetch_fenc_422 = prefetch_fenc_null;
515 pf->prefetch_ref = prefetch_ref_null;
516 pf->memcpy_aligned = memcpy;
517 pf->memzero_aligned = memzero_aligned;
518 pf->frame_init_lowres_core = frame_init_lowres_core;
520 pf->integral_init4h = integral_init4h;
521 pf->integral_init8h = integral_init8h;
522 pf->integral_init4v = integral_init4v;
523 pf->integral_init8v = integral_init8v;
525 pf->mbtree_propagate_cost = mbtree_propagate_cost;
528 x264_mc_init_mmx( cpu, pf );
531 if( cpu&X264_CPU_ALTIVEC )
532 x264_mc_altivec_init( pf );
535 x264_mc_init_arm( cpu, pf );
539 void x264_frame_filter( x264_t *h, x264_frame_t *frame, int mb_y, int b_end )
541 const int b_interlaced = PARAM_INTERLACED;
542 int start = mb_y*16 - 8; // buffer = 4 for deblock + 3 for 6tap, rounded to 8
543 int height = (b_end ? frame->i_lines[0] + 16*PARAM_INTERLACED : (mb_y+b_interlaced)*16) + 8;
545 if( mb_y & b_interlaced )
548 for( int p = 0; p < (CHROMA444 ? 3 : 1); p++ )
550 int stride = frame->i_stride[p];
551 const int width = frame->i_width[p];
552 int offs = start*stride - 8; // buffer = 3 for 6tap, aligned to 8 for simd
554 if( !b_interlaced || h->mb.b_adaptive_mbaff )
556 frame->filtered[p][1] + offs,
557 frame->filtered[p][2] + offs,
558 frame->filtered[p][3] + offs,
559 frame->plane[p] + offs,
560 stride, width + 16, height - start,
565 /* MC must happen between pixels in the same field. */
566 stride = frame->i_stride[p] << 1;
567 start = (mb_y*16 >> 1) - 8;
568 int height_fld = ((b_end ? frame->i_lines[p] : mb_y*16) >> 1) + 8;
569 offs = start*stride - 8;
570 for( int i = 0; i < 2; i++, offs += frame->i_stride[p] )
573 frame->filtered_fld[p][1] + offs,
574 frame->filtered_fld[p][2] + offs,
575 frame->filtered_fld[p][3] + offs,
576 frame->plane_fld[p] + offs,
577 stride, width + 16, height_fld - start,
583 /* generate integral image:
584 * frame->integral contains 2 planes. in the upper plane, each element is
585 * the sum of an 8x8 pixel region with top-left corner on that point.
586 * in the lower plane, 4x4 sums (needed only with --partitions p4x4). */
588 if( frame->integral )
590 int stride = frame->i_stride[0];
593 memset( frame->integral - PADV * stride - PADH, 0, stride * sizeof(uint16_t) );
598 for( int y = start; y < height; y++ )
600 pixel *pix = frame->plane[0] + y * stride - PADH;
601 uint16_t *sum8 = frame->integral + (y+1) * stride - PADH;
603 if( h->frames.b_have_sub8x8_esa )
605 h->mc.integral_init4h( sum8, pix, stride );
607 sum4 = sum8 + stride * (frame->i_lines[0] + PADV*2);
609 h->mc.integral_init4v( sum8, sum4, stride );
613 h->mc.integral_init8h( sum8, pix, stride );
615 h->mc.integral_init8v( sum8-8*stride, stride );