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, int i_dst_stride,
41 pixel *src1, int i_src1_stride,
42 pixel *src2, int i_src2_stride,
43 int i_width, int i_height )
45 for( int y = 0; y < i_height; y++ )
47 for( int x = 0; x < i_width; x++ )
48 dst[x] = ( src1[x] + src2[x] + 1 ) >> 1;
50 src1 += i_src1_stride;
51 src2 += i_src2_stride;
55 static inline void pixel_avg_wxh( pixel *dst, int i_dst, pixel *src1, int i_src1, pixel *src2, int i_src2, int width, int height )
57 for( int y = 0; y < height; y++ )
59 for( int x = 0; x < width; x++ )
60 dst[x] = ( src1[x] + src2[x] + 1 ) >> 1;
67 /* Implicit weighted bipred only:
68 * assumes log2_denom = 5, offset = 0, weight1 + weight2 = 64 */
69 static inline void pixel_avg_weight_wxh( pixel *dst, int i_dst, pixel *src1, int i_src1, pixel *src2, int i_src2, int width, int height, int i_weight1 )
71 const int i_weight2 = 64 - i_weight1;
72 for( int y = 0; y<height; y++, dst += i_dst, src1 += i_src1, src2 += i_src2 )
73 for( int x = 0; x<width; x++ )
74 dst[x] = x264_clip_pixel( (src1[x]*i_weight1 + src2[x]*i_weight2 + (1<<5)) >> 6 );
78 #define PIXEL_AVG_C( name, width, height ) \
79 static void name( pixel *pix1, int i_stride_pix1, \
80 pixel *pix2, int i_stride_pix2, \
81 pixel *pix3, int i_stride_pix3, int weight ) \
84 pixel_avg_wxh( pix1, i_stride_pix1, pix2, i_stride_pix2, pix3, i_stride_pix3, width, height ); \
86 pixel_avg_weight_wxh( pix1, i_stride_pix1, pix2, i_stride_pix2, pix3, i_stride_pix3, width, height, weight ); \
88 PIXEL_AVG_C( pixel_avg_16x16, 16, 16 )
89 PIXEL_AVG_C( pixel_avg_16x8, 16, 8 )
90 PIXEL_AVG_C( pixel_avg_8x16, 8, 16 )
91 PIXEL_AVG_C( pixel_avg_8x8, 8, 8 )
92 PIXEL_AVG_C( pixel_avg_8x4, 8, 4 )
93 PIXEL_AVG_C( pixel_avg_4x16, 4, 16 )
94 PIXEL_AVG_C( pixel_avg_4x8, 4, 8 )
95 PIXEL_AVG_C( pixel_avg_4x4, 4, 4 )
96 PIXEL_AVG_C( pixel_avg_4x2, 4, 2 )
97 PIXEL_AVG_C( pixel_avg_2x8, 2, 8 )
98 PIXEL_AVG_C( pixel_avg_2x4, 2, 4 )
99 PIXEL_AVG_C( pixel_avg_2x2, 2, 2 )
101 static void x264_weight_cache( x264_t *h, x264_weight_t *w )
103 w->weightfn = h->mc.weight;
105 #define opscale(x) dst[x] = x264_clip_pixel( ((src[x] * scale + (1<<(denom - 1))) >> denom) + offset )
106 #define opscale_noden(x) dst[x] = x264_clip_pixel( src[x] * scale + offset )
107 static void mc_weight( pixel *dst, int i_dst_stride, pixel *src, int i_src_stride, const x264_weight_t *weight, int i_width, int i_height )
109 int offset = weight->i_offset << (BIT_DEPTH-8);
110 int scale = weight->i_scale;
111 int denom = weight->i_denom;
114 for( int y = 0; y < i_height; y++, dst += i_dst_stride, src += i_src_stride )
115 for( int x = 0; x < i_width; x++ )
120 for( int y = 0; y < i_height; y++, dst += i_dst_stride, src += i_src_stride )
121 for( int x = 0; x < i_width; x++ )
126 #define MC_WEIGHT_C( name, width ) \
127 static void name( pixel *dst, int i_dst_stride, pixel *src, int i_src_stride, const x264_weight_t *weight, int height ) \
129 mc_weight( dst, i_dst_stride, src, i_src_stride, weight, width, height );\
132 MC_WEIGHT_C( mc_weight_w20, 20 )
133 MC_WEIGHT_C( mc_weight_w16, 16 )
134 MC_WEIGHT_C( mc_weight_w12, 12 )
135 MC_WEIGHT_C( mc_weight_w8, 8 )
136 MC_WEIGHT_C( mc_weight_w4, 4 )
137 MC_WEIGHT_C( mc_weight_w2, 2 )
139 static weight_fn_t x264_mc_weight_wtab[6] =
148 const x264_weight_t x264_weight_none[3] = { {{0}} };
149 static void mc_copy( pixel *src, int i_src_stride, pixel *dst, int i_dst_stride, int i_width, int i_height )
151 for( int y = 0; y < i_height; y++ )
153 memcpy( dst, src, i_width * sizeof(pixel) );
160 #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]))
161 static void hpel_filter( pixel *dsth, pixel *dstv, pixel *dstc, pixel *src,
162 int stride, int width, int height, int16_t *buf )
164 const int pad = (BIT_DEPTH > 9) ? (-10 * PIXEL_MAX) : 0;
165 for( int y = 0; y < height; y++ )
167 for( int x = -2; x < width+3; x++ )
169 int v = TAPFILTER(src,stride);
170 dstv[x] = x264_clip_pixel( (v + 16) >> 5 );
171 /* transform v for storage in a 16-bit integer */
174 for( int x = 0; x < width; x++ )
175 dstc[x] = x264_clip_pixel( (TAPFILTER(buf+2,1) - 32*pad + 512) >> 10 );
176 for( int x = 0; x < width; x++ )
177 dsth[x] = x264_clip_pixel( (TAPFILTER(src,1) + 16) >> 5 );
185 static const uint8_t hpel_ref0[16] = {0,1,1,1,0,1,1,1,2,3,3,3,0,1,1,1};
186 static const uint8_t hpel_ref1[16] = {0,0,0,0,2,2,3,2,2,2,3,2,2,2,3,2};
188 static void mc_luma( pixel *dst, int i_dst_stride,
189 pixel *src[4], int i_src_stride,
191 int i_width, int i_height, const x264_weight_t *weight )
193 int qpel_idx = ((mvy&3)<<2) + (mvx&3);
194 int offset = (mvy>>2)*i_src_stride + (mvx>>2);
195 pixel *src1 = src[hpel_ref0[qpel_idx]] + offset + ((mvy&3) == 3) * i_src_stride;
197 if( qpel_idx & 5 ) /* qpel interpolation needed */
199 pixel *src2 = src[hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);
200 pixel_avg( dst, i_dst_stride, src1, i_src_stride,
201 src2, i_src_stride, i_width, i_height );
202 if( weight->weightfn )
203 mc_weight( dst, i_dst_stride, dst, i_dst_stride, weight, i_width, i_height );
205 else if( weight->weightfn )
206 mc_weight( dst, i_dst_stride, src1, i_src_stride, weight, i_width, i_height );
208 mc_copy( src1, i_src_stride, dst, i_dst_stride, i_width, i_height );
211 static pixel *get_ref( pixel *dst, int *i_dst_stride,
212 pixel *src[4], int i_src_stride,
214 int i_width, int i_height, const x264_weight_t *weight )
216 int qpel_idx = ((mvy&3)<<2) + (mvx&3);
217 int offset = (mvy>>2)*i_src_stride + (mvx>>2);
218 pixel *src1 = src[hpel_ref0[qpel_idx]] + offset + ((mvy&3) == 3) * i_src_stride;
220 if( qpel_idx & 5 ) /* qpel interpolation needed */
222 pixel *src2 = src[hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);
223 pixel_avg( dst, *i_dst_stride, src1, i_src_stride,
224 src2, i_src_stride, i_width, i_height );
225 if( weight->weightfn )
226 mc_weight( dst, *i_dst_stride, dst, *i_dst_stride, weight, i_width, i_height );
229 else if( weight->weightfn )
231 mc_weight( dst, *i_dst_stride, src1, i_src_stride, weight, i_width, i_height );
236 *i_dst_stride = i_src_stride;
241 /* full chroma mc (ie until 1/8 pixel)*/
242 static void mc_chroma( pixel *dstu, pixel *dstv, int i_dst_stride,
243 pixel *src, int i_src_stride,
245 int i_width, int i_height )
251 int cA = (8-d8x)*(8-d8y);
252 int cB = d8x *(8-d8y);
253 int cC = (8-d8x)*d8y;
256 src += (mvy >> 3) * i_src_stride + (mvx >> 3)*2;
257 srcp = &src[i_src_stride];
259 for( int y = 0; y < i_height; y++ )
261 for( int x = 0; x < i_width; x++ )
263 dstu[x] = ( cA*src[2*x] + cB*src[2*x+2] +
264 cC*srcp[2*x] + cD*srcp[2*x+2] + 32 ) >> 6;
265 dstv[x] = ( cA*src[2*x+1] + cB*src[2*x+3] +
266 cC*srcp[2*x+1] + cD*srcp[2*x+3] + 32 ) >> 6;
268 dstu += i_dst_stride;
269 dstv += i_dst_stride;
271 srcp += i_src_stride;
276 static void mc_copy_w##W( pixel *dst, int i_dst, pixel *src, int i_src, int i_height ) \
278 mc_copy( src, i_src, dst, i_dst, W, i_height ); \
284 void x264_plane_copy_c( pixel *dst, int i_dst,
285 pixel *src, int i_src, int w, int h )
289 memcpy( dst, src, w * sizeof(pixel) );
295 void x264_plane_copy_interleave_c( pixel *dst, int i_dst,
296 pixel *srcu, int i_srcu,
297 pixel *srcv, int i_srcv, int w, int h )
299 for( int y=0; y<h; y++, dst+=i_dst, srcu+=i_srcu, srcv+=i_srcv )
300 for( int x=0; x<w; x++ )
303 dst[2*x+1] = srcv[x];
307 static void x264_plane_copy_deinterleave_c( pixel *dstu, int i_dstu,
308 pixel *dstv, int i_dstv,
309 pixel *src, int i_src, int w, int h )
311 for( int y=0; y<h; y++, dstu+=i_dstu, dstv+=i_dstv, src+=i_src )
312 for( int x=0; x<w; x++ )
315 dstv[x] = src[2*x+1];
319 static void x264_plane_copy_deinterleave_rgb_c( pixel *dsta, int i_dsta,
320 pixel *dstb, int i_dstb,
321 pixel *dstc, int i_dstc,
322 pixel *src, int i_src, int pw, int w, int h )
324 for( int y=0; y<h; y++, dsta+=i_dsta, dstb+=i_dstb, dstc+=i_dstc, src+=i_src )
326 for( int x=0; x<w; x++ )
329 dstb[x] = src[x*pw+1];
330 dstc[x] = src[x*pw+2];
335 static void store_interleave_chroma( pixel *dst, int i_dst, pixel *srcu, pixel *srcv, int height )
337 for( int y=0; y<height; y++, dst+=i_dst, srcu+=FDEC_STRIDE, srcv+=FDEC_STRIDE )
338 for( int x=0; x<8; x++ )
341 dst[2*x+1] = srcv[x];
345 static void load_deinterleave_chroma_fenc( pixel *dst, pixel *src, int i_src, int height )
347 x264_plane_copy_deinterleave_c( dst, FENC_STRIDE, dst+FENC_STRIDE/2, FENC_STRIDE, src, i_src, 8, height );
350 static void load_deinterleave_chroma_fdec( pixel *dst, pixel *src, int i_src, int height )
352 x264_plane_copy_deinterleave_c( dst, FDEC_STRIDE, dst+FDEC_STRIDE/2, FDEC_STRIDE, src, i_src, 8, height );
355 static void prefetch_fenc_null( pixel *pix_y, int stride_y,
356 pixel *pix_uv, int stride_uv, int mb_x )
359 static void prefetch_ref_null( pixel *pix, int stride, int parity )
362 static void memzero_aligned( void * dst, int n )
367 static void integral_init4h( uint16_t *sum, pixel *pix, int stride )
369 int v = pix[0]+pix[1]+pix[2]+pix[3];
370 for( int x = 0; x < stride-4; x++ )
372 sum[x] = v + sum[x-stride];
373 v += pix[x+4] - pix[x];
377 static void integral_init8h( uint16_t *sum, pixel *pix, int stride )
379 int v = pix[0]+pix[1]+pix[2]+pix[3]+pix[4]+pix[5]+pix[6]+pix[7];
380 for( int x = 0; x < stride-8; x++ )
382 sum[x] = v + sum[x-stride];
383 v += pix[x+8] - pix[x];
387 static void integral_init4v( uint16_t *sum8, uint16_t *sum4, int stride )
389 for( int x = 0; x < stride-8; x++ )
390 sum4[x] = sum8[x+4*stride] - sum8[x];
391 for( int x = 0; x < stride-8; x++ )
392 sum8[x] = sum8[x+8*stride] + sum8[x+8*stride+4] - sum8[x] - sum8[x+4];
395 static void integral_init8v( uint16_t *sum8, int stride )
397 for( int x = 0; x < stride-8; x++ )
398 sum8[x] = sum8[x+8*stride] - sum8[x];
401 void x264_frame_init_lowres( x264_t *h, x264_frame_t *frame )
403 pixel *src = frame->plane[0];
404 int i_stride = frame->i_stride[0];
405 int i_height = frame->i_lines[0];
406 int i_width = frame->i_width[0];
408 // duplicate last row and column so that their interpolation doesn't have to be special-cased
409 for( int y = 0; y < i_height; y++ )
410 src[i_width+y*i_stride] = src[i_width-1+y*i_stride];
411 memcpy( src+i_stride*i_height, src+i_stride*(i_height-1), (i_width+1) * sizeof(pixel) );
412 h->mc.frame_init_lowres_core( src, frame->lowres[0], frame->lowres[1], frame->lowres[2], frame->lowres[3],
413 i_stride, frame->i_stride_lowres, frame->i_width_lowres, frame->i_lines_lowres );
414 x264_frame_expand_border_lowres( frame );
416 memset( frame->i_cost_est, -1, sizeof(frame->i_cost_est) );
418 for( int y = 0; y < h->param.i_bframe + 2; y++ )
419 for( int x = 0; x < h->param.i_bframe + 2; x++ )
420 frame->i_row_satds[y][x][0] = -1;
422 for( int y = 0; y <= !!h->param.i_bframe; y++ )
423 for( int x = 0; x <= h->param.i_bframe; x++ )
424 frame->lowres_mvs[y][x][0][0] = 0x7FFF;
427 static void frame_init_lowres_core( pixel *src0, pixel *dst0, pixel *dsth, pixel *dstv, pixel *dstc,
428 int src_stride, int dst_stride, int width, int height )
430 for( int y = 0; y < height; y++ )
432 pixel *src1 = src0+src_stride;
433 pixel *src2 = src1+src_stride;
434 for( int x = 0; x<width; x++ )
436 // slower than naive bilinear, but matches asm
437 #define FILTER(a,b,c,d) ((((a+b+1)>>1)+((c+d+1)>>1)+1)>>1)
438 dst0[x] = FILTER(src0[2*x ], src1[2*x ], src0[2*x+1], src1[2*x+1]);
439 dsth[x] = FILTER(src0[2*x+1], src1[2*x+1], src0[2*x+2], src1[2*x+2]);
440 dstv[x] = FILTER(src1[2*x ], src2[2*x ], src1[2*x+1], src2[2*x+1]);
441 dstc[x] = FILTER(src1[2*x+1], src2[2*x+1], src1[2*x+2], src2[2*x+2]);
444 src0 += src_stride*2;
452 /* Estimate the total amount of influence on future quality that could be had if we
453 * were to improve the reference samples used to inter predict any given macroblock. */
454 static void mbtree_propagate_cost( int *dst, uint16_t *propagate_in, uint16_t *intra_costs,
455 uint16_t *inter_costs, uint16_t *inv_qscales, float *fps_factor, int len )
457 float fps = *fps_factor / 256.f;
458 for( int i = 0; i < len; i++ )
460 float intra_cost = intra_costs[i] * inv_qscales[i];
461 float propagate_amount = propagate_in[i] + intra_cost*fps;
462 float propagate_num = intra_costs[i] - (inter_costs[i] & LOWRES_COST_MASK);
463 float propagate_denom = intra_costs[i];
464 dst[i] = (int)(propagate_amount * propagate_num / propagate_denom + 0.5f);
468 void x264_mc_init( int cpu, x264_mc_functions_t *pf )
470 pf->mc_luma = mc_luma;
471 pf->get_ref = get_ref;
473 pf->mc_chroma = mc_chroma;
475 pf->avg[PIXEL_16x16]= pixel_avg_16x16;
476 pf->avg[PIXEL_16x8] = pixel_avg_16x8;
477 pf->avg[PIXEL_8x16] = pixel_avg_8x16;
478 pf->avg[PIXEL_8x8] = pixel_avg_8x8;
479 pf->avg[PIXEL_8x4] = pixel_avg_8x4;
480 pf->avg[PIXEL_4x16] = pixel_avg_4x16;
481 pf->avg[PIXEL_4x8] = pixel_avg_4x8;
482 pf->avg[PIXEL_4x4] = pixel_avg_4x4;
483 pf->avg[PIXEL_4x2] = pixel_avg_4x2;
484 pf->avg[PIXEL_2x8] = pixel_avg_2x8;
485 pf->avg[PIXEL_2x4] = pixel_avg_2x4;
486 pf->avg[PIXEL_2x2] = pixel_avg_2x2;
488 pf->weight = x264_mc_weight_wtab;
489 pf->offsetadd = x264_mc_weight_wtab;
490 pf->offsetsub = x264_mc_weight_wtab;
491 pf->weight_cache = x264_weight_cache;
493 pf->copy_16x16_unaligned = mc_copy_w16;
494 pf->copy[PIXEL_16x16] = mc_copy_w16;
495 pf->copy[PIXEL_8x8] = mc_copy_w8;
496 pf->copy[PIXEL_4x4] = mc_copy_w4;
498 pf->store_interleave_chroma = store_interleave_chroma;
499 pf->load_deinterleave_chroma_fenc = load_deinterleave_chroma_fenc;
500 pf->load_deinterleave_chroma_fdec = load_deinterleave_chroma_fdec;
502 pf->plane_copy = x264_plane_copy_c;
503 pf->plane_copy_interleave = x264_plane_copy_interleave_c;
504 pf->plane_copy_deinterleave = x264_plane_copy_deinterleave_c;
505 pf->plane_copy_deinterleave_rgb = x264_plane_copy_deinterleave_rgb_c;
507 pf->hpel_filter = hpel_filter;
509 pf->prefetch_fenc_420 = prefetch_fenc_null;
510 pf->prefetch_fenc_422 = prefetch_fenc_null;
511 pf->prefetch_ref = prefetch_ref_null;
512 pf->memcpy_aligned = memcpy;
513 pf->memzero_aligned = memzero_aligned;
514 pf->frame_init_lowres_core = frame_init_lowres_core;
516 pf->integral_init4h = integral_init4h;
517 pf->integral_init8h = integral_init8h;
518 pf->integral_init4v = integral_init4v;
519 pf->integral_init8v = integral_init8v;
521 pf->mbtree_propagate_cost = mbtree_propagate_cost;
524 x264_mc_init_mmx( cpu, pf );
527 if( cpu&X264_CPU_ALTIVEC )
528 x264_mc_altivec_init( pf );
531 x264_mc_init_arm( cpu, pf );
535 void x264_frame_filter( x264_t *h, x264_frame_t *frame, int mb_y, int b_end )
537 const int b_interlaced = PARAM_INTERLACED;
538 int start = mb_y*16 - 8; // buffer = 4 for deblock + 3 for 6tap, rounded to 8
539 int height = (b_end ? frame->i_lines[0] + 16*PARAM_INTERLACED : (mb_y+b_interlaced)*16) + 8;
541 if( mb_y & b_interlaced )
544 for( int p = 0; p < (CHROMA444 ? 3 : 1); p++ )
546 int stride = frame->i_stride[p];
547 const int width = frame->i_width[p];
548 int offs = start*stride - 8; // buffer = 3 for 6tap, aligned to 8 for simd
550 if( !b_interlaced || h->mb.b_adaptive_mbaff )
552 frame->filtered[p][1] + offs,
553 frame->filtered[p][2] + offs,
554 frame->filtered[p][3] + offs,
555 frame->plane[p] + offs,
556 stride, width + 16, height - start,
561 /* MC must happen between pixels in the same field. */
562 stride = frame->i_stride[p] << 1;
563 start = (mb_y*16 >> 1) - 8;
564 int height_fld = ((b_end ? frame->i_lines[p] : mb_y*16) >> 1) + 8;
565 offs = start*stride - 8;
566 for( int i = 0; i < 2; i++, offs += frame->i_stride[p] )
569 frame->filtered_fld[p][1] + offs,
570 frame->filtered_fld[p][2] + offs,
571 frame->filtered_fld[p][3] + offs,
572 frame->plane_fld[p] + offs,
573 stride, width + 16, height_fld - start,
579 /* generate integral image:
580 * frame->integral contains 2 planes. in the upper plane, each element is
581 * the sum of an 8x8 pixel region with top-left corner on that point.
582 * in the lower plane, 4x4 sums (needed only with --partitions p4x4). */
584 if( frame->integral )
586 int stride = frame->i_stride[0];
589 memset( frame->integral - PADV * stride - PADH, 0, stride * sizeof(uint16_t) );
594 for( int y = start; y < height; y++ )
596 pixel *pix = frame->plane[0] + y * stride - PADH;
597 uint16_t *sum8 = frame->integral + (y+1) * stride - PADH;
599 if( h->frames.b_have_sub8x8_esa )
601 h->mc.integral_init4h( sum8, pix, stride );
603 sum4 = sum8 + stride * (frame->i_lines[0] + PADV*2);
605 h->mc.integral_init4v( sum8, sum4, stride );
609 h->mc.integral_init8h( sum8, pix, stride );
611 h->mc.integral_init8v( sum8-8*stride, stride );