common/mc.c

   1 /*****************************************************************************
   2  * mc.c: h264 encoder library (Motion Compensation)
   3  *****************************************************************************
   4  * Copyright (C) 2003-2008 x264 project
   5  *
   6  * Authors: Laurent Aimar <fenrir@via.ecp.fr>
   7  *          Loren Merritt <lorenm@u.washington.edu>
   8  *
   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.
  13  *
  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.
  18  *
  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  *****************************************************************************/
  23
  24 #include "common.h"
  25
  26 #ifdef HAVE_MMX
  27 #include "x86/mc.h"
  28 #endif
  29 #ifdef ARCH_PPC
  30 #include "ppc/mc.h"
  31 #endif
  32
  33
  34 static inline void pixel_avg( uint8_t *dst,  int i_dst_stride,
  35                               uint8_t *src1, int i_src1_stride,
  36                               uint8_t *src2, int i_src2_stride,
  37                               int i_width, int i_height )
  38 {
  39     int x, y;
  40     for( y = 0; y < i_height; y++ )
  41     {
  42         for( x = 0; x < i_width; x++ )
  43         {
  44             dst[x] = ( src1[x] + src2[x] + 1 ) >> 1;
  45         }
  46         dst  += i_dst_stride;
  47         src1 += i_src1_stride;
  48         src2 += i_src2_stride;
  49     }
  50 }
  51
  52 static inline void pixel_avg_wxh( uint8_t *dst, int i_dst, uint8_t *src1, int i_src1, uint8_t *src2, int i_src2, int width, int height )
  53 {
  54     int x, y;
  55     for( y = 0; y < height; y++ )
  56     {
  57         for( x = 0; x < width; x++ )
  58         {
  59             dst[x] = ( src1[x] + src2[x] + 1 ) >> 1;
  60         }
  61         src1 += i_src1;
  62         src2 += i_src2;
  63         dst += i_dst;
  64     }
  65 }
  66
  67 /* Implicit weighted bipred only:
  68  * assumes log2_denom = 5, offset = 0, weight1 + weight2 = 64 */
  69 #define op_scale2(x) dst[x] = x264_clip_uint8( (src1[x]*i_weight1 + src2[x]*i_weight2 + (1<<5)) >> 6 )
  70 static inline void pixel_avg_weight_wxh( uint8_t *dst, int i_dst, uint8_t *src1, int i_src1, uint8_t *src2, int i_src2, int width, int height, int i_weight1 )
  71 {
  72     int y;
  73     const int i_weight2 = 64 - i_weight1;
  74     for( y = 0; y<height; y++, dst += i_dst, src1 += i_src1, src2 += i_src2 )
  75     {
  76         op_scale2(0);
  77         op_scale2(1);
  78         if(width==2) continue;
  79         op_scale2(2);
  80         op_scale2(3);
  81         if(width==4) continue;
  82         op_scale2(4);
  83         op_scale2(5);
  84         op_scale2(6);
  85         op_scale2(7);
  86         if(width==8) continue;
  87         op_scale2(8);
  88         op_scale2(9);
  89         op_scale2(10);
  90         op_scale2(11);
  91         op_scale2(12);
  92         op_scale2(13);
  93         op_scale2(14);
  94         op_scale2(15);
  95     }
  96 }
  97 #undef op_scale2
  98
  99 #define PIXEL_AVG_C( name, width, height ) \
 100 static void name( uint8_t *pix1, int i_stride_pix1, \
 101                   uint8_t *pix2, int i_stride_pix2, \
 102                   uint8_t *pix3, int i_stride_pix3, int weight ) \
 103 { \
 104     if( weight == 32 )\
 105         pixel_avg_wxh( pix1, i_stride_pix1, pix2, i_stride_pix2, pix3, i_stride_pix3, width, height ); \
 106     else\
 107         pixel_avg_weight_wxh( pix1, i_stride_pix1, pix2, i_stride_pix2, pix3, i_stride_pix3, width, height, weight ); \
 108 }
 109 PIXEL_AVG_C( pixel_avg_16x16, 16, 16 )
 110 PIXEL_AVG_C( pixel_avg_16x8,  16, 8 )
 111 PIXEL_AVG_C( pixel_avg_8x16,  8, 16 )
 112 PIXEL_AVG_C( pixel_avg_8x8,   8, 8 )
 113 PIXEL_AVG_C( pixel_avg_8x4,   8, 4 )
 114 PIXEL_AVG_C( pixel_avg_4x8,   4, 8 )
 115 PIXEL_AVG_C( pixel_avg_4x4,   4, 4 )
 116 PIXEL_AVG_C( pixel_avg_4x2,   4, 2 )
 117 PIXEL_AVG_C( pixel_avg_2x4,   2, 4 )
 118 PIXEL_AVG_C( pixel_avg_2x2,   2, 2 )
 119
 120 static void mc_copy( uint8_t *src, int i_src_stride, uint8_t *dst, int i_dst_stride, int i_width, int i_height )
 121 {
 122     int y;
 123
 124     for( y = 0; y < i_height; y++ )
 125     {
 126         memcpy( dst, src, i_width );
 127
 128         src += i_src_stride;
 129         dst += i_dst_stride;
 130     }
 131 }
 132
 133 #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]))
 134 static void hpel_filter( uint8_t *dsth, uint8_t *dstv, uint8_t *dstc, uint8_t *src,
 135                          int stride, int width, int height )
 136 {
 137     int16_t *buf = x264_malloc((width+5)*sizeof(int16_t));
 138     int x, y;
 139     for( y=0; y<height; y++ )
 140     {
 141         for( x=-2; x<width+3; x++ )
 142         {
 143             int v = TAPFILTER(src,stride);
 144             dstv[x] = x264_clip_uint8((v + 16) >> 5);
 145             buf[x+2] = v;
 146         }
 147         for( x=0; x<width; x++ )
 148             dstc[x] = x264_clip_uint8((TAPFILTER(buf+2,1) + 512) >> 10);
 149         for( x=0; x<width; x++ )
 150             dsth[x] = x264_clip_uint8((TAPFILTER(src,1) + 16) >> 5);
 151         dsth += stride;
 152         dstv += stride;
 153         dstc += stride;
 154         src += stride;
 155     }
 156     x264_free(buf);
 157 }
 158
 159 static const int hpel_ref0[16] = {0,1,1,1,0,1,1,1,2,3,3,3,0,1,1,1};
 160 static const int hpel_ref1[16] = {0,0,0,0,2,2,3,2,2,2,3,2,2,2,3,2};
 161
 162 static void mc_luma( uint8_t *dst,    int i_dst_stride,
 163                      uint8_t *src[4], int i_src_stride,
 164                      int mvx, int mvy,
 165                      int i_width, int i_height )
 166 {
 167     int qpel_idx = ((mvy&3)<<2) + (mvx&3);
 168     int offset = (mvy>>2)*i_src_stride + (mvx>>2);
 169     uint8_t *src1 = src[hpel_ref0[qpel_idx]] + offset + ((mvy&3) == 3) * i_src_stride;
 170
 171     if( qpel_idx & 5 ) /* qpel interpolation needed */
 172     {
 173         uint8_t *src2 = src[hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);
 174         pixel_avg( dst, i_dst_stride, src1, i_src_stride,
 175                    src2, i_src_stride, i_width, i_height );
 176     }
 177     else
 178     {
 179         mc_copy( src1, i_src_stride, dst, i_dst_stride, i_width, i_height );
 180     }
 181 }
 182
 183 static uint8_t *get_ref( uint8_t *dst,   int *i_dst_stride,
 184                          uint8_t *src[4], int i_src_stride,
 185                          int mvx, int mvy,
 186                          int i_width, int i_height )
 187 {
 188     int qpel_idx = ((mvy&3)<<2) + (mvx&3);
 189     int offset = (mvy>>2)*i_src_stride + (mvx>>2);
 190     uint8_t *src1 = src[hpel_ref0[qpel_idx]] + offset + ((mvy&3) == 3) * i_src_stride;
 191
 192     if( qpel_idx & 5 ) /* qpel interpolation needed */
 193     {
 194         uint8_t *src2 = src[hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);
 195         pixel_avg( dst, *i_dst_stride, src1, i_src_stride,
 196                    src2, i_src_stride, i_width, i_height );
 197         return dst;
 198     }
 199     else
 200     {
 201         *i_dst_stride = i_src_stride;
 202         return src1;
 203     }
 204 }
 205
 206 /* full chroma mc (ie until 1/8 pixel)*/
 207 static void mc_chroma( uint8_t *dst, int i_dst_stride,
 208                        uint8_t *src, int i_src_stride,
 209                        int mvx, int mvy,
 210                        int i_width, int i_height )
 211 {
 212     uint8_t *srcp;
 213     int x, y;
 214
 215     const int d8x = mvx&0x07;
 216     const int d8y = mvy&0x07;
 217
 218     const int cA = (8-d8x)*(8-d8y);
 219     const int cB = d8x    *(8-d8y);
 220     const int cC = (8-d8x)*d8y;
 221     const int cD = d8x    *d8y;
 222
 223     src  += (mvy >> 3) * i_src_stride + (mvx >> 3);
 224     srcp = &src[i_src_stride];
 225
 226     for( y = 0; y < i_height; y++ )
 227     {
 228         for( x = 0; x < i_width; x++ )
 229         {
 230             dst[x] = ( cA*src[x]  + cB*src[x+1] +
 231                        cC*srcp[x] + cD*srcp[x+1] + 32 ) >> 6;
 232         }
 233         dst  += i_dst_stride;
 234
 235         src   = srcp;
 236         srcp += i_src_stride;
 237     }
 238 }
 239
 240 #define MC_COPY(W) \
 241 static void mc_copy_w##W( uint8_t *dst, int i_dst, uint8_t *src, int i_src, int i_height ) \
 242 { \
 243     mc_copy( src, i_src, dst, i_dst, W, i_height ); \
 244 }
 245 MC_COPY( 16 )
 246 MC_COPY( 8 )
 247 MC_COPY( 4 )
 248
 249 static void plane_copy( uint8_t *dst, int i_dst,
 250                         uint8_t *src, int i_src, int w, int h)
 251 {
 252     while( h-- )
 253     {
 254         memcpy( dst, src, w );
 255         dst += i_dst;
 256         src += i_src;
 257     }
 258 }
 259
 260 static void prefetch_fenc_null( uint8_t *pix_y, int stride_y,
 261                                 uint8_t *pix_uv, int stride_uv, int mb_x )
 262 {}
 263
 264 static void prefetch_ref_null( uint8_t *pix, int stride, int parity )
 265 {}
 266
 267 static void memzero_aligned( void * dst, int n )
 268 {
 269     memset( dst, 0, n );
 270 }
 271
 272 void x264_frame_init_lowres( x264_t *h, x264_frame_t *frame )
 273 {
 274     uint8_t *src = frame->plane[0];
 275     int i_stride = frame->i_stride[0];
 276     int i_height = frame->i_lines[0];
 277     int i_width  = frame->i_width[0];
 278     int x, y;
 279
 280     // duplicate last row and column so that their interpolation doesn't have to be special-cased
 281     for( y=0; y<i_height; y++ )
 282         src[i_width+y*i_stride] = src[i_width-1+y*i_stride];
 283     h->mc.memcpy_aligned( src+i_stride*i_height, src+i_stride*(i_height-1), i_width );
 284     h->mc.frame_init_lowres_core( src, frame->lowres[0], frame->lowres[1], frame->lowres[2], frame->lowres[3],
 285                                   i_stride, frame->i_stride_lowres, frame->i_width_lowres, frame->i_lines_lowres );
 286     x264_frame_expand_border_lowres( frame );
 287
 288     memset( frame->i_cost_est, -1, sizeof(frame->i_cost_est) );
 289
 290     for( x = 0; x < h->param.i_bframe + 2; x++ )
 291         for( y = 0; y < h->param.i_bframe + 2; y++ )
 292             frame->i_row_satds[y][x][0] = -1;
 293
 294     for( y = 0; y <= !!h->param.i_bframe; y++ )
 295         for( x = 0; x <= h->param.i_bframe; x++ )
 296             frame->lowres_mvs[y][x][0][0] = 0x7FFF;
 297 }
 298
 299 static void frame_init_lowres_core( uint8_t *src0, uint8_t *dst0, uint8_t *dsth, uint8_t *dstv, uint8_t *dstc,
 300                                     int src_stride, int dst_stride, int width, int height )
 301 {
 302     int x,y;
 303     for( y=0; y<height; y++ )
 304     {
 305         uint8_t *src1 = src0+src_stride;
 306         uint8_t *src2 = src1+src_stride;
 307         for( x=0; x<width; x++ )
 308         {
 309             // slower than naive bilinear, but matches asm
 310 #define FILTER(a,b,c,d) ((((a+b+1)>>1)+((c+d+1)>>1)+1)>>1)
 311             dst0[x] = FILTER(src0[2*x  ], src1[2*x  ], src0[2*x+1], src1[2*x+1]);
 312             dsth[x] = FILTER(src0[2*x+1], src1[2*x+1], src0[2*x+2], src1[2*x+2]);
 313             dstv[x] = FILTER(src1[2*x  ], src2[2*x  ], src1[2*x+1], src2[2*x+1]);
 314             dstc[x] = FILTER(src1[2*x+1], src2[2*x+1], src1[2*x+2], src2[2*x+2]);
 315 #undef FILTER
 316         }
 317         src0 += src_stride*2;
 318         dst0 += dst_stride;
 319         dsth += dst_stride;
 320         dstv += dst_stride;
 321         dstc += dst_stride;
 322     }
 323 }
 324
 325 void x264_mc_init( int cpu, x264_mc_functions_t *pf )
 326 {
 327     pf->mc_luma   = mc_luma;
 328     pf->get_ref   = get_ref;
 329     pf->mc_chroma = mc_chroma;
 330
 331     pf->avg[PIXEL_16x16]= pixel_avg_16x16;
 332     pf->avg[PIXEL_16x8] = pixel_avg_16x8;
 333     pf->avg[PIXEL_8x16] = pixel_avg_8x16;
 334     pf->avg[PIXEL_8x8]  = pixel_avg_8x8;
 335     pf->avg[PIXEL_8x4]  = pixel_avg_8x4;
 336     pf->avg[PIXEL_4x8]  = pixel_avg_4x8;
 337     pf->avg[PIXEL_4x4]  = pixel_avg_4x4;
 338     pf->avg[PIXEL_4x2]  = pixel_avg_4x2;
 339     pf->avg[PIXEL_2x4]  = pixel_avg_2x4;
 340     pf->avg[PIXEL_2x2]  = pixel_avg_2x2;
 341
 342     pf->copy_16x16_unaligned = mc_copy_w16;
 343     pf->copy[PIXEL_16x16] = mc_copy_w16;
 344     pf->copy[PIXEL_8x8]   = mc_copy_w8;
 345     pf->copy[PIXEL_4x4]   = mc_copy_w4;
 346
 347     pf->plane_copy = plane_copy;
 348     pf->hpel_filter = hpel_filter;
 349
 350     pf->prefetch_fenc = prefetch_fenc_null;
 351     pf->prefetch_ref  = prefetch_ref_null;
 352     pf->memcpy_aligned = memcpy;
 353     pf->memzero_aligned = memzero_aligned;
 354     pf->frame_init_lowres_core = frame_init_lowres_core;
 355
 356 #ifdef HAVE_MMX
 357     x264_mc_init_mmx( cpu, pf );
 358 #endif
 359 #ifdef ARCH_PPC
 360     if( cpu&X264_CPU_ALTIVEC )
 361         x264_mc_altivec_init( pf );
 362 #endif
 363 }
 364
 365 void x264_frame_filter( x264_t *h, x264_frame_t *frame, int mb_y, int b_end )
 366 {
 367     const int b_interlaced = h->sh.b_mbaff;
 368     const int stride = frame->i_stride[0] << b_interlaced;
 369     const int width = frame->i_width[0];
 370     int start = (mb_y*16 >> b_interlaced) - 8; // buffer = 4 for deblock + 3 for 6tap, rounded to 8
 371     int height = ((b_end ? frame->i_lines[0] : mb_y*16) >> b_interlaced) + 8;
 372     int offs = start*stride - 8; // buffer = 3 for 6tap, aligned to 8 for simd
 373     int x, y;
 374
 375     if( mb_y & b_interlaced )
 376         return;
 377
 378     for( y=0; y<=b_interlaced; y++, offs+=frame->i_stride[0] )
 379     {
 380         h->mc.hpel_filter(
 381             frame->filtered[1] + offs,
 382             frame->filtered[2] + offs,
 383             frame->filtered[3] + offs,
 384             frame->plane[0] + offs,
 385             stride, width + 16, height - start );
 386     }
 387
 388     /* generate integral image:
 389      * frame->integral contains 2 planes. in the upper plane, each element is
 390      * the sum of an 8x8 pixel region with top-left corner on that point.
 391      * in the lower plane, 4x4 sums (needed only with --partitions p4x4). */
 392
 393     if( frame->integral )
 394     {
 395         if( start < 0 )
 396         {
 397             memset( frame->integral - PADV * stride - PADH, 0, stride * sizeof(uint16_t) );
 398             start = -PADV;
 399         }
 400         if( b_end )
 401             height += PADV-8;
 402         for( y = start; y < height; y++ )
 403         {
 404             uint8_t  *ref  = frame->plane[0] + y * stride - PADH;
 405             uint16_t *line = frame->integral + (y+1) * stride - PADH + 1;
 406             uint16_t v = line[0] = 0;
 407             for( x = 1; x < stride-1; x++ )
 408                 line[x] = v += ref[x] + line[x-stride] - line[x-stride-1];
 409             line -= 8*stride;
 410             if( y >= 9-PADV )
 411             {
 412                 uint16_t *sum4 = line + stride * (frame->i_lines[0] + PADV*2);
 413                 for( x = 1; x < stride-8; x++, line++, sum4++ )
 414                 {
 415                     sum4[0] =  line[4+4*stride] - line[4] - line[4*stride] + line[0];
 416                     line[0] += line[8+8*stride] - line[8] - line[8*stride];
 417                 }
 418             }
 419         }
 420     }
 421 }