1 /*****************************************************************************
2 * mc.c: h264 encoder library (Motion Compensation)
3 *****************************************************************************
4 * Copyright (C) 2003-2008 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.
22 *****************************************************************************/
37 static inline void pixel_avg( pixel *dst, int i_dst_stride,
38 pixel *src1, int i_src1_stride,
39 pixel *src2, int i_src2_stride,
40 int i_width, int i_height )
42 for( int y = 0; y < i_height; y++ )
44 for( int x = 0; x < i_width; x++ )
45 dst[x] = ( src1[x] + src2[x] + 1 ) >> 1;
47 src1 += i_src1_stride;
48 src2 += i_src2_stride;
52 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 )
54 for( int y = 0; y < height; y++ )
56 for( int x = 0; x < width; x++ )
57 dst[x] = ( src1[x] + src2[x] + 1 ) >> 1;
64 /* Implicit weighted bipred only:
65 * assumes log2_denom = 5, offset = 0, weight1 + weight2 = 64 */
66 #define op_scale2(x) dst[x] = x264_clip_pixel( (src1[x]*i_weight1 + src2[x]*i_weight2 + (1<<5)) >> 6 )
67 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 )
69 const int i_weight2 = 64 - i_weight1;
70 for( int y = 0; y<height; y++, dst += i_dst, src1 += i_src1, src2 += i_src2 )
74 if(width==2) continue;
77 if(width==4) continue;
82 if(width==8) continue;
95 #define PIXEL_AVG_C( name, width, height ) \
96 static void name( pixel *pix1, int i_stride_pix1, \
97 pixel *pix2, int i_stride_pix2, \
98 pixel *pix3, int i_stride_pix3, int weight ) \
101 pixel_avg_wxh( pix1, i_stride_pix1, pix2, i_stride_pix2, pix3, i_stride_pix3, width, height ); \
103 pixel_avg_weight_wxh( pix1, i_stride_pix1, pix2, i_stride_pix2, pix3, i_stride_pix3, width, height, weight ); \
105 PIXEL_AVG_C( pixel_avg_16x16, 16, 16 )
106 PIXEL_AVG_C( pixel_avg_16x8, 16, 8 )
107 PIXEL_AVG_C( pixel_avg_8x16, 8, 16 )
108 PIXEL_AVG_C( pixel_avg_8x8, 8, 8 )
109 PIXEL_AVG_C( pixel_avg_8x4, 8, 4 )
110 PIXEL_AVG_C( pixel_avg_4x8, 4, 8 )
111 PIXEL_AVG_C( pixel_avg_4x4, 4, 4 )
112 PIXEL_AVG_C( pixel_avg_4x2, 4, 2 )
113 PIXEL_AVG_C( pixel_avg_2x4, 2, 4 )
114 PIXEL_AVG_C( pixel_avg_2x2, 2, 2 )
116 static void x264_weight_cache( x264_t *h, x264_weight_t *w )
118 w->weightfn = h->mc.weight;
120 #define opscale(x) dst[x] = x264_clip_pixel( ((src[x] * scale + (1<<(denom - 1))) >> denom) + offset )
121 #define opscale_noden(x) dst[x] = x264_clip_pixel( src[x] * scale + offset )
122 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 )
124 int offset = weight->i_offset << (BIT_DEPTH-8);
125 int scale = weight->i_scale;
126 int denom = weight->i_denom;
129 for( int y = 0; y < i_height; y++, dst += i_dst_stride, src += i_src_stride )
130 for( int x = 0; x < i_width; x++ )
135 for( int y = 0; y < i_height; y++, dst += i_dst_stride, src += i_src_stride )
136 for( int x = 0; x < i_width; x++ )
141 #define MC_WEIGHT_C( name, width ) \
142 static void name( pixel *dst, int i_dst_stride, pixel *src, int i_src_stride, const x264_weight_t *weight, int height ) \
144 mc_weight( dst, i_dst_stride, src, i_src_stride, weight, width, height );\
147 MC_WEIGHT_C( mc_weight_w20, 20 )
148 MC_WEIGHT_C( mc_weight_w16, 16 )
149 MC_WEIGHT_C( mc_weight_w12, 12 )
150 MC_WEIGHT_C( mc_weight_w8, 8 )
151 MC_WEIGHT_C( mc_weight_w4, 4 )
152 MC_WEIGHT_C( mc_weight_w2, 2 )
154 static weight_fn_t x264_mc_weight_wtab[6] =
163 const x264_weight_t weight_none[3] = { {{0}} };
164 static void mc_copy( pixel *src, int i_src_stride, pixel *dst, int i_dst_stride, int i_width, int i_height )
166 for( int y = 0; y < i_height; y++ )
168 memcpy( dst, src, i_width * sizeof(pixel) );
175 #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]))
176 static void hpel_filter( pixel *dsth, pixel *dstv, pixel *dstc, pixel *src,
177 int stride, int width, int height, dctcoef *buf )
179 for( int y = 0; y < height; y++ )
181 for( int x = -2; x < width+3; x++ )
183 int v = TAPFILTER(src,stride);
184 dstv[x] = x264_clip_pixel( (v + 16) >> 5 );
187 for( int x = 0; x < width; x++ )
188 dstc[x] = x264_clip_pixel( (TAPFILTER(buf+2,1) + 512) >> 10 );
189 for( int x = 0; x < width; x++ )
190 dsth[x] = x264_clip_pixel( (TAPFILTER(src,1) + 16) >> 5 );
198 static const uint8_t hpel_ref0[16] = {0,1,1,1,0,1,1,1,2,3,3,3,0,1,1,1};
199 static const uint8_t hpel_ref1[16] = {0,0,0,0,2,2,3,2,2,2,3,2,2,2,3,2};
201 static void mc_luma( pixel *dst, int i_dst_stride,
202 pixel *src[4], int i_src_stride,
204 int i_width, int i_height, const x264_weight_t *weight )
206 int qpel_idx = ((mvy&3)<<2) + (mvx&3);
207 int offset = (mvy>>2)*i_src_stride + (mvx>>2);
208 pixel *src1 = src[hpel_ref0[qpel_idx]] + offset + ((mvy&3) == 3) * i_src_stride;
210 if( qpel_idx & 5 ) /* qpel interpolation needed */
212 pixel *src2 = src[hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);
213 pixel_avg( dst, i_dst_stride, src1, i_src_stride,
214 src2, i_src_stride, i_width, i_height );
215 if( weight->weightfn )
216 mc_weight( dst, i_dst_stride, dst, i_dst_stride, weight, i_width, i_height );
218 else if( weight->weightfn )
219 mc_weight( dst, i_dst_stride, src1, i_src_stride, weight, i_width, i_height );
221 mc_copy( src1, i_src_stride, dst, i_dst_stride, i_width, i_height );
224 static pixel *get_ref( pixel *dst, int *i_dst_stride,
225 pixel *src[4], int i_src_stride,
227 int i_width, int i_height, const x264_weight_t *weight )
229 int qpel_idx = ((mvy&3)<<2) + (mvx&3);
230 int offset = (mvy>>2)*i_src_stride + (mvx>>2);
231 pixel *src1 = src[hpel_ref0[qpel_idx]] + offset + ((mvy&3) == 3) * i_src_stride;
233 if( qpel_idx & 5 ) /* qpel interpolation needed */
235 pixel *src2 = src[hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);
236 pixel_avg( dst, *i_dst_stride, src1, i_src_stride,
237 src2, i_src_stride, i_width, i_height );
238 if( weight->weightfn )
239 mc_weight( dst, *i_dst_stride, dst, *i_dst_stride, weight, i_width, i_height );
242 else if( weight->weightfn )
244 mc_weight( dst, *i_dst_stride, src1, i_src_stride, weight, i_width, i_height );
249 *i_dst_stride = i_src_stride;
254 /* full chroma mc (ie until 1/8 pixel)*/
255 static void mc_chroma( pixel *dst, int i_dst_stride,
256 pixel *src, int i_src_stride,
258 int i_width, int i_height )
264 int cA = (8-d8x)*(8-d8y);
265 int cB = d8x *(8-d8y);
266 int cC = (8-d8x)*d8y;
269 src += (mvy >> 3) * i_src_stride + (mvx >> 3);
270 srcp = &src[i_src_stride];
272 for( int y = 0; y < i_height; y++ )
274 for( int x = 0; x < i_width; x++ )
275 dst[x] = ( cA*src[x] + cB*src[x+1] + cC*srcp[x] + cD*srcp[x+1] + 32 ) >> 6;
278 srcp += i_src_stride;
283 static void mc_copy_w##W( pixel *dst, int i_dst, pixel *src, int i_src, int i_height ) \
285 mc_copy( src, i_src, dst, i_dst, W, i_height ); \
291 void x264_plane_copy_c( pixel *dst, int i_dst,
292 uint8_t *src, int i_src, int w, int h)
296 #if X264_HIGH_BIT_DEPTH
297 for( int i = 0; i < w; i++ )
298 dst[i] = src[i] << (BIT_DEPTH-8);
300 memcpy( dst, src, w );
307 static void prefetch_fenc_null( pixel *pix_y, int stride_y,
308 pixel *pix_uv, int stride_uv, int mb_x )
311 static void prefetch_ref_null( pixel *pix, int stride, int parity )
314 static void memzero_aligned( void * dst, int n )
319 static void integral_init4h( uint16_t *sum, pixel *pix, int stride )
321 int v = pix[0]+pix[1]+pix[2]+pix[3];
322 for( int x = 0; x < stride-4; x++ )
324 sum[x] = v + sum[x-stride];
325 v += pix[x+4] - pix[x];
329 static void integral_init8h( uint16_t *sum, pixel *pix, int stride )
331 int v = pix[0]+pix[1]+pix[2]+pix[3]+pix[4]+pix[5]+pix[6]+pix[7];
332 for( int x = 0; x < stride-8; x++ )
334 sum[x] = v + sum[x-stride];
335 v += pix[x+8] - pix[x];
339 static void integral_init4v( uint16_t *sum8, uint16_t *sum4, int stride )
341 for( int x = 0; x < stride-8; x++ )
342 sum4[x] = sum8[x+4*stride] - sum8[x];
343 for( int x = 0; x < stride-8; x++ )
344 sum8[x] = sum8[x+8*stride] + sum8[x+8*stride+4] - sum8[x] - sum8[x+4];
347 static void integral_init8v( uint16_t *sum8, int stride )
349 for( int x = 0; x < stride-8; x++ )
350 sum8[x] = sum8[x+8*stride] - sum8[x];
353 void x264_frame_init_lowres( x264_t *h, x264_frame_t *frame )
355 pixel *src = frame->plane[0];
356 int i_stride = frame->i_stride[0];
357 int i_height = frame->i_lines[0];
358 int i_width = frame->i_width[0];
360 // duplicate last row and column so that their interpolation doesn't have to be special-cased
361 for( int y = 0; y < i_height; y++ )
362 src[i_width+y*i_stride] = src[i_width-1+y*i_stride];
363 memcpy( src+i_stride*i_height, src+i_stride*(i_height-1), (i_width+1) * sizeof(pixel) );
364 h->mc.frame_init_lowres_core( src, frame->lowres[0], frame->lowres[1], frame->lowres[2], frame->lowres[3],
365 i_stride, frame->i_stride_lowres, frame->i_width_lowres, frame->i_lines_lowres );
366 x264_frame_expand_border_lowres( frame );
368 memset( frame->i_cost_est, -1, sizeof(frame->i_cost_est) );
370 for( int y = 0; y < h->param.i_bframe + 2; y++ )
371 for( int x = 0; x < h->param.i_bframe + 2; x++ )
372 frame->i_row_satds[y][x][0] = -1;
374 for( int y = 0; y <= !!h->param.i_bframe; y++ )
375 for( int x = 0; x <= h->param.i_bframe; x++ )
376 frame->lowres_mvs[y][x][0][0] = 0x7FFF;
379 static void frame_init_lowres_core( pixel *src0, pixel *dst0, pixel *dsth, pixel *dstv, pixel *dstc,
380 int src_stride, int dst_stride, int width, int height )
382 for( int y = 0; y < height; y++ )
384 pixel *src1 = src0+src_stride;
385 pixel *src2 = src1+src_stride;
386 for( int x = 0; x<width; x++ )
388 // slower than naive bilinear, but matches asm
389 #define FILTER(a,b,c,d) ((((a+b+1)>>1)+((c+d+1)>>1)+1)>>1)
390 dst0[x] = FILTER(src0[2*x ], src1[2*x ], src0[2*x+1], src1[2*x+1]);
391 dsth[x] = FILTER(src0[2*x+1], src1[2*x+1], src0[2*x+2], src1[2*x+2]);
392 dstv[x] = FILTER(src1[2*x ], src2[2*x ], src1[2*x+1], src2[2*x+1]);
393 dstc[x] = FILTER(src1[2*x+1], src2[2*x+1], src1[2*x+2], src2[2*x+2]);
396 src0 += src_stride*2;
404 #if defined(__GNUC__) && (ARCH_X86 || ARCH_X86_64)
405 // gcc isn't smart enough to use the "idiv" instruction
406 static ALWAYS_INLINE int32_t div_64_32(int64_t x, int32_t y)
408 int32_t quotient, remainder;
410 :"=a"(quotient), "=d"(remainder)
411 :"a"((uint32_t)x), "d"((int32_t)(x>>32)), "r"(y)
416 #define div_64_32(x,y) ((x)/(y))
419 /* Estimate the total amount of influence on future quality that could be had if we
420 * were to improve the reference samples used to inter predict any given macroblock. */
421 static void mbtree_propagate_cost( int *dst, uint16_t *propagate_in, uint16_t *intra_costs,
422 uint16_t *inter_costs, uint16_t *inv_qscales, int len )
424 for( int i = 0; i < len; i++ )
426 int propagate_amount = propagate_in[i] + ((intra_costs[i] * inv_qscales[i] + 128)>>8);
427 dst[i] = div_64_32((int64_t)propagate_amount * (intra_costs[i] - (inter_costs[i] & LOWRES_COST_MASK)), intra_costs[i]);
431 void x264_mc_init( int cpu, x264_mc_functions_t *pf )
433 pf->mc_luma = mc_luma;
434 pf->get_ref = get_ref;
435 pf->mc_chroma = mc_chroma;
437 pf->avg[PIXEL_16x16]= pixel_avg_16x16;
438 pf->avg[PIXEL_16x8] = pixel_avg_16x8;
439 pf->avg[PIXEL_8x16] = pixel_avg_8x16;
440 pf->avg[PIXEL_8x8] = pixel_avg_8x8;
441 pf->avg[PIXEL_8x4] = pixel_avg_8x4;
442 pf->avg[PIXEL_4x8] = pixel_avg_4x8;
443 pf->avg[PIXEL_4x4] = pixel_avg_4x4;
444 pf->avg[PIXEL_4x2] = pixel_avg_4x2;
445 pf->avg[PIXEL_2x4] = pixel_avg_2x4;
446 pf->avg[PIXEL_2x2] = pixel_avg_2x2;
448 pf->weight = x264_mc_weight_wtab;
449 pf->offsetadd = x264_mc_weight_wtab;
450 pf->offsetsub = x264_mc_weight_wtab;
451 pf->weight_cache = x264_weight_cache;
453 pf->copy_16x16_unaligned = mc_copy_w16;
454 pf->copy[PIXEL_16x16] = mc_copy_w16;
455 pf->copy[PIXEL_8x8] = mc_copy_w8;
456 pf->copy[PIXEL_4x4] = mc_copy_w4;
458 pf->plane_copy = x264_plane_copy_c;
459 pf->hpel_filter = hpel_filter;
461 pf->prefetch_fenc = prefetch_fenc_null;
462 pf->prefetch_ref = prefetch_ref_null;
463 pf->memcpy_aligned = memcpy;
464 pf->memzero_aligned = memzero_aligned;
465 pf->frame_init_lowres_core = frame_init_lowres_core;
467 pf->integral_init4h = integral_init4h;
468 pf->integral_init8h = integral_init8h;
469 pf->integral_init4v = integral_init4v;
470 pf->integral_init8v = integral_init8v;
472 pf->mbtree_propagate_cost = mbtree_propagate_cost;
475 x264_mc_init_mmx( cpu, pf );
478 if( cpu&X264_CPU_ALTIVEC )
479 x264_mc_altivec_init( pf );
482 x264_mc_init_arm( cpu, pf );
486 void x264_frame_filter( x264_t *h, x264_frame_t *frame, int mb_y, int b_end )
488 const int b_interlaced = h->sh.b_mbaff;
489 const int stride = frame->i_stride[0] << b_interlaced;
490 const int width = frame->i_width[0];
491 int start = (mb_y*16 >> b_interlaced) - 8; // buffer = 4 for deblock + 3 for 6tap, rounded to 8
492 int height = ((b_end ? frame->i_lines[0] : mb_y*16) >> b_interlaced) + 8;
493 int offs = start*stride - 8; // buffer = 3 for 6tap, aligned to 8 for simd
495 if( mb_y & b_interlaced )
498 for( int y = 0; y <= b_interlaced; y++, offs += frame->i_stride[0] )
501 frame->filtered[1] + offs,
502 frame->filtered[2] + offs,
503 frame->filtered[3] + offs,
504 frame->plane[0] + offs,
505 stride, width + 16, height - start,
509 /* generate integral image:
510 * frame->integral contains 2 planes. in the upper plane, each element is
511 * the sum of an 8x8 pixel region with top-left corner on that point.
512 * in the lower plane, 4x4 sums (needed only with --partitions p4x4). */
514 if( frame->integral )
518 memset( frame->integral - PADV * stride - PADH, 0, stride * sizeof(uint16_t) );
523 for( int y = start; y < height; y++ )
525 pixel *pix = frame->plane[0] + y * stride - PADH;
526 uint16_t *sum8 = frame->integral + (y+1) * stride - PADH;
528 if( h->frames.b_have_sub8x8_esa )
530 h->mc.integral_init4h( sum8, pix, stride );
532 sum4 = sum8 + stride * (frame->i_lines[0] + PADV*2);
534 h->mc.integral_init4v( sum8, sum4, stride );
538 h->mc.integral_init8h( sum8, pix, stride );
540 h->mc.integral_init8v( sum8-8*stride, stride );