git.sesse.net Git - x264/blob - common/x86/mc-c.c

   1 /*****************************************************************************
   2  * mc-c.c: x86 motion compensation
   3  *****************************************************************************
   4  * Copyright (C) 2003-2012 x264 project
   5  *
   6  * Authors: Laurent Aimar <fenrir@via.ecp.fr>
   7  *          Loren Merritt <lorenm@u.washington.edu>
   8  *          Fiona Glaser <fiona@x264.com>
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License as published by
  12  * the Free Software Foundation; either version 2 of the License, or
  13  * (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18  * GNU General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public License
  21  * along with this program; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02111, USA.
  23  *
  24  * This program is also available under a commercial proprietary license.
  25  * For more information, contact us at licensing@x264.com.
  26  *****************************************************************************/
  27
  28 #include <stdlib.h>
  29 #include <stdio.h>
  30 #include <string.h>
  31
  32 #include "common/common.h"
  33 #include "mc.h"
  34
  35 #define DECL_SUF( func, args )\
  36     void func##_mmx2 args;\
  37     void func##_sse2 args;\
  38     void func##_ssse3 args;
  39
  40 DECL_SUF( x264_pixel_avg_16x16, ( pixel *, intptr_t, pixel *, intptr_t, pixel *, intptr_t, int ))
  41 DECL_SUF( x264_pixel_avg_16x8,  ( pixel *, intptr_t, pixel *, intptr_t, pixel *, intptr_t, int ))
  42 DECL_SUF( x264_pixel_avg_8x16,  ( pixel *, intptr_t, pixel *, intptr_t, pixel *, intptr_t, int ))
  43 DECL_SUF( x264_pixel_avg_8x8,   ( pixel *, intptr_t, pixel *, intptr_t, pixel *, intptr_t, int ))
  44 DECL_SUF( x264_pixel_avg_8x4,   ( pixel *, intptr_t, pixel *, intptr_t, pixel *, intptr_t, int ))
  45 DECL_SUF( x264_pixel_avg_4x16,  ( pixel *, intptr_t, pixel *, intptr_t, pixel *, intptr_t, int ))
  46 DECL_SUF( x264_pixel_avg_4x8,   ( pixel *, intptr_t, pixel *, intptr_t, pixel *, intptr_t, int ))
  47 DECL_SUF( x264_pixel_avg_4x4,   ( pixel *, intptr_t, pixel *, intptr_t, pixel *, intptr_t, int ))
  48 DECL_SUF( x264_pixel_avg_4x2,   ( pixel *, intptr_t, pixel *, intptr_t, pixel *, intptr_t, int ))
  49
  50 #define MC_WEIGHT(w,type) \
  51     void x264_mc_weight_w##w##_##type( pixel *, intptr_t, pixel *, intptr_t, const x264_weight_t *, int );
  52
  53 #define MC_WEIGHT_OFFSET(w,type) \
  54     void x264_mc_offsetadd_w##w##_##type( pixel *, intptr_t, pixel *, intptr_t, const x264_weight_t *, int ); \
  55     void x264_mc_offsetsub_w##w##_##type( pixel *, intptr_t, pixel *, intptr_t, const x264_weight_t *, int ); \
  56     MC_WEIGHT(w,type)
  57
  58 MC_WEIGHT_OFFSET( 4, mmx2 )
  59 MC_WEIGHT_OFFSET( 8, mmx2 )
  60 MC_WEIGHT_OFFSET( 12, mmx2 )
  61 MC_WEIGHT_OFFSET( 16, mmx2 )
  62 MC_WEIGHT_OFFSET( 20, mmx2 )
  63 MC_WEIGHT_OFFSET( 12, sse2 )
  64 MC_WEIGHT_OFFSET( 16, sse2 )
  65 MC_WEIGHT_OFFSET( 20, sse2 )
  66 #if HIGH_BIT_DEPTH
  67 MC_WEIGHT_OFFSET( 8, sse2 )
  68 #endif
  69 MC_WEIGHT( 8, sse2  )
  70 MC_WEIGHT( 4, ssse3 )
  71 MC_WEIGHT( 8, ssse3 )
  72 MC_WEIGHT( 12, ssse3 )
  73 MC_WEIGHT( 16, ssse3 )
  74 MC_WEIGHT( 20, ssse3 )
  75 #undef MC_OFFSET
  76 #undef MC_WEIGHT
  77
  78 void x264_mc_copy_w4_mmx  ( pixel *, intptr_t, pixel *, intptr_t, int );
  79 void x264_mc_copy_w8_mmx  ( pixel *, intptr_t, pixel *, intptr_t, int );
  80 void x264_mc_copy_w8_sse2 ( pixel *, intptr_t, pixel *, intptr_t, int );
  81 void x264_mc_copy_w16_mmx ( pixel *, intptr_t, pixel *, intptr_t, int );
  82 void x264_mc_copy_w16_sse2( pixel *, intptr_t, pixel *, intptr_t, int );
  83 void x264_mc_copy_w16_aligned_sse2( pixel *, intptr_t, pixel *, intptr_t, int );
  84 void x264_prefetch_fenc_420_mmx2( pixel *, intptr_t, pixel *, intptr_t, int );
  85 void x264_prefetch_fenc_422_mmx2( pixel *, intptr_t, pixel *, intptr_t, int );
  86 void x264_prefetch_ref_mmx2( pixel *, intptr_t, int );
  87 void x264_plane_copy_core_mmx2( pixel *, intptr_t, pixel *, intptr_t, int w, int h );
  88 void x264_plane_copy_c( pixel *, intptr_t, pixel *, intptr_t, int w, int h );
  89 void x264_plane_copy_interleave_core_mmx2( pixel *dst,  intptr_t i_dst,
  90                                            pixel *srcu, intptr_t i_srcu,
  91                                            pixel *srcv, intptr_t i_srcv, int w, int h );
  92 void x264_plane_copy_interleave_core_sse2( pixel *dst,  intptr_t i_dst,
  93                                            pixel *srcu, intptr_t i_srcu,
  94                                            pixel *srcv, intptr_t i_srcv, int w, int h );
  95 void x264_plane_copy_interleave_core_avx( pixel *dst,  intptr_t i_dst,
  96                                           pixel *srcu, intptr_t i_srcu,
  97                                           pixel *srcv, intptr_t i_srcv, int w, int h );
  98 void x264_plane_copy_interleave_c( pixel *dst,  intptr_t i_dst,
  99                                    pixel *srcu, intptr_t i_srcu,
 100                                    pixel *srcv, intptr_t i_srcv, int w, int h );
 101 void x264_plane_copy_deinterleave_mmx( pixel *dstu, intptr_t i_dstu,
 102                                        pixel *dstv, intptr_t i_dstv,
 103                                        pixel *src,  intptr_t i_src, int w, int h );
 104 void x264_plane_copy_deinterleave_sse2( pixel *dstu, intptr_t i_dstu,
 105                                         pixel *dstv, intptr_t i_dstv,
 106                                         pixel *src,  intptr_t i_src, int w, int h );
 107 void x264_plane_copy_deinterleave_ssse3( uint8_t *dstu, intptr_t i_dstu,
 108                                          uint8_t *dstv, intptr_t i_dstv,
 109                                          uint8_t *src,  intptr_t i_src, int w, int h );
 110 void x264_plane_copy_deinterleave_avx( uint16_t *dstu, intptr_t i_dstu,
 111                                        uint16_t *dstv, intptr_t i_dstv,
 112                                        uint16_t *src,  intptr_t i_src, int w, int h );
 113 void x264_store_interleave_chroma_mmx2( pixel *dst, intptr_t i_dst, pixel *srcu, pixel *srcv, int height );
 114 void x264_store_interleave_chroma_sse2( pixel *dst, intptr_t i_dst, pixel *srcu, pixel *srcv, int height );
 115 void x264_store_interleave_chroma_avx ( pixel *dst, intptr_t i_dst, pixel *srcu, pixel *srcv, int height );
 116 void x264_load_deinterleave_chroma_fenc_mmx ( pixel *dst, pixel *src, intptr_t i_src, int height );
 117 void x264_load_deinterleave_chroma_fenc_sse2( pixel *dst, pixel *src, intptr_t i_src, int height );
 118 void x264_load_deinterleave_chroma_fenc_ssse3( uint8_t *dst, uint8_t *src, intptr_t i_src, int height );
 119 void x264_load_deinterleave_chroma_fenc_avx( uint16_t *dst, uint16_t *src, intptr_t i_src, int height );
 120 void x264_load_deinterleave_chroma_fdec_mmx ( pixel *dst, pixel *src, intptr_t i_src, int height );
 121 void x264_load_deinterleave_chroma_fdec_sse2( pixel *dst, pixel *src, intptr_t i_src, int height );
 122 void x264_load_deinterleave_chroma_fdec_ssse3( uint8_t *dst, uint8_t *src, intptr_t i_src, int height );
 123 void x264_load_deinterleave_chroma_fdec_avx( uint16_t *dst, uint16_t *src, intptr_t i_src, int height );
 124 void *x264_memcpy_aligned_mmx ( void *dst, const void *src, size_t n );
 125 void *x264_memcpy_aligned_sse2( void *dst, const void *src, size_t n );
 126 void x264_memzero_aligned_mmx ( void *dst, size_t n );
 127 void x264_memzero_aligned_sse2( void *dst, size_t n );
 128 void x264_integral_init4h_sse4( uint16_t *sum, uint8_t *pix, intptr_t stride );
 129 void x264_integral_init8h_sse4( uint16_t *sum, uint8_t *pix, intptr_t stride );
 130 void x264_integral_init8h_avx ( uint16_t *sum, uint8_t *pix, intptr_t stride );
 131 void x264_integral_init4v_mmx  ( uint16_t *sum8, uint16_t *sum4, intptr_t stride );
 132 void x264_integral_init4v_sse2 ( uint16_t *sum8, uint16_t *sum4, intptr_t stride );
 133 void x264_integral_init4v_ssse3( uint16_t *sum8, uint16_t *sum4, intptr_t stride );
 134 void x264_integral_init8v_mmx ( uint16_t *sum8, intptr_t stride );
 135 void x264_integral_init8v_sse2( uint16_t *sum8, intptr_t stride );
 136 void x264_mbtree_propagate_cost_sse2( int *dst, uint16_t *propagate_in, uint16_t *intra_costs,
 137                                       uint16_t *inter_costs, uint16_t *inv_qscales, float *fps_factor, int len );
 138 void x264_mbtree_propagate_cost_avx ( int *dst, uint16_t *propagate_in, uint16_t *intra_costs,
 139                                       uint16_t *inter_costs, uint16_t *inv_qscales, float *fps_factor, int len );
 140 void x264_mbtree_propagate_cost_fma4( int *dst, uint16_t *propagate_in, uint16_t *intra_costs,
 141                                       uint16_t *inter_costs, uint16_t *inv_qscales, float *fps_factor, int len );
 142
 143 #define MC_CHROMA(cpu)\
 144 void x264_mc_chroma_##cpu( pixel *dstu, pixel *dstv, intptr_t i_dst, pixel *src, intptr_t i_src,\
 145                            int dx, int dy, int i_width, int i_height );
 146 MC_CHROMA(mmx2)
 147 MC_CHROMA(sse2)
 148 MC_CHROMA(sse2_misalign)
 149 MC_CHROMA(ssse3)
 150 MC_CHROMA(ssse3_cache64)
 151 MC_CHROMA(avx)
 152 MC_CHROMA(avx_cache64)
 153
 154 #define LOWRES(cpu)\
 155 void x264_frame_init_lowres_core_##cpu( pixel *src0, pixel *dst0, pixel *dsth, pixel *dstv, pixel *dstc,\
 156                                         intptr_t src_stride, intptr_t dst_stride, int width, int height );
 157 LOWRES(mmx2)
 158 LOWRES(cache32_mmx2)
 159 LOWRES(sse2)
 160 LOWRES(ssse3)
 161 LOWRES(avx)
 162 LOWRES(xop)
 163
 164 #define PIXEL_AVG_W(width,cpu)\
 165 void x264_pixel_avg2_w##width##_##cpu( pixel *, intptr_t, pixel *, intptr_t, pixel *, intptr_t );
 166 /* This declares some functions that don't exist, but that isn't a problem. */
 167 #define PIXEL_AVG_WALL(cpu)\
 168 PIXEL_AVG_W(4,cpu); PIXEL_AVG_W(8,cpu); PIXEL_AVG_W(10,cpu); PIXEL_AVG_W(12,cpu); PIXEL_AVG_W(16,cpu); PIXEL_AVG_W(18,cpu); PIXEL_AVG_W(20,cpu);
 169
 170 PIXEL_AVG_WALL(mmx2)
 171 PIXEL_AVG_WALL(cache32_mmx2)
 172 PIXEL_AVG_WALL(cache64_mmx2)
 173 PIXEL_AVG_WALL(cache64_sse2)
 174 PIXEL_AVG_WALL(sse2)
 175 PIXEL_AVG_WALL(sse2_misalign)
 176 PIXEL_AVG_WALL(cache64_ssse3)
 177
 178 #define PIXEL_AVG_WTAB(instr, name1, name2, name3, name4, name5)\
 179 static void (* const x264_pixel_avg_wtab_##instr[6])( pixel *, intptr_t, pixel *, intptr_t, pixel *, intptr_t ) =\
 180 {\
 181     NULL,\
 182     x264_pixel_avg2_w4_##name1,\
 183     x264_pixel_avg2_w8_##name2,\
 184     x264_pixel_avg2_w12_##name3,\
 185     x264_pixel_avg2_w16_##name4,\
 186     x264_pixel_avg2_w20_##name5,\
 187 };
 188
 189 #if HIGH_BIT_DEPTH
 190 /* we can replace w12/w20 with w10/w18 as only 9/17 pixels in fact are important */
 191 #define x264_pixel_avg2_w12_mmx2       x264_pixel_avg2_w10_mmx2
 192 #define x264_pixel_avg2_w20_mmx2       x264_pixel_avg2_w18_mmx2
 193 #define x264_pixel_avg2_w12_sse2         x264_pixel_avg2_w10_sse2
 194 #define x264_pixel_avg2_w20_sse2         x264_pixel_avg2_w18_sse2
 195 #else
 196 /* w16 sse2 is faster than w12 mmx as long as the cacheline issue is resolved */
 197 #define x264_pixel_avg2_w12_cache64_ssse3 x264_pixel_avg2_w16_cache64_ssse3
 198 #define x264_pixel_avg2_w12_cache64_sse2 x264_pixel_avg2_w16_cache64_sse2
 199 #define x264_pixel_avg2_w12_sse3         x264_pixel_avg2_w16_sse3
 200 #define x264_pixel_avg2_w12_sse2         x264_pixel_avg2_w16_sse2
 201 #endif // HIGH_BIT_DEPTH
 202
 203 PIXEL_AVG_WTAB(mmx2, mmx2, mmx2, mmx2, mmx2, mmx2)
 204 #if HIGH_BIT_DEPTH
 205 PIXEL_AVG_WTAB(sse2, mmx2, sse2, sse2, sse2, sse2)
 206 #else // !HIGH_BIT_DEPTH
 207 #if ARCH_X86
 208 PIXEL_AVG_WTAB(cache32_mmx2, mmx2, cache32_mmx2, cache32_mmx2, cache32_mmx2, cache32_mmx2)
 209 PIXEL_AVG_WTAB(cache64_mmx2, mmx2, cache64_mmx2, cache64_mmx2, cache64_mmx2, cache64_mmx2)
 210 #endif
 211 PIXEL_AVG_WTAB(sse2, mmx2, mmx2, sse2, sse2, sse2)
 212 PIXEL_AVG_WTAB(sse2_misalign, mmx2, mmx2, sse2, sse2, sse2_misalign)
 213 PIXEL_AVG_WTAB(cache64_sse2, mmx2, cache64_mmx2, cache64_sse2, cache64_sse2, cache64_sse2)
 214 PIXEL_AVG_WTAB(cache64_ssse3, mmx2, cache64_mmx2, cache64_ssse3, cache64_ssse3, cache64_sse2)
 215 #endif // HIGH_BIT_DEPTH
 216
 217 #define MC_COPY_WTAB(instr, name1, name2, name3)\
 218 static void (* const x264_mc_copy_wtab_##instr[5])( pixel *, intptr_t, pixel *, intptr_t, int ) =\
 219 {\
 220     NULL,\
 221     x264_mc_copy_w4_##name1,\
 222     x264_mc_copy_w8_##name2,\
 223     NULL,\
 224     x264_mc_copy_w16_##name3,\
 225 };
 226
 227 MC_COPY_WTAB(mmx,mmx,mmx,mmx)
 228 #if HIGH_BIT_DEPTH
 229 MC_COPY_WTAB(sse2,mmx,sse2,sse2)
 230 #else
 231 MC_COPY_WTAB(sse2,mmx,mmx,sse2)
 232 #endif
 233
 234 #define MC_WEIGHT_WTAB(function, instr, name1, name2, w12version)\
 235     static void (* x264_mc_##function##_wtab_##instr[6])( pixel *, intptr_t, pixel *, intptr_t, const x264_weight_t *, int ) =\
 236 {\
 237     x264_mc_##function##_w4_##name1,\
 238     x264_mc_##function##_w4_##name1,\
 239     x264_mc_##function##_w8_##name2,\
 240     x264_mc_##function##_w##w12version##_##instr,\
 241     x264_mc_##function##_w16_##instr,\
 242     x264_mc_##function##_w20_##instr,\
 243 };
 244
 245 #if HIGH_BIT_DEPTH
 246 MC_WEIGHT_WTAB(weight,mmx2,mmx2,mmx2,12)
 247 MC_WEIGHT_WTAB(offsetadd,mmx2,mmx2,mmx2,12)
 248 MC_WEIGHT_WTAB(offsetsub,mmx2,mmx2,mmx2,12)
 249 MC_WEIGHT_WTAB(weight,sse2,mmx2,sse2,12)
 250 MC_WEIGHT_WTAB(offsetadd,sse2,mmx2,sse2,16)
 251 MC_WEIGHT_WTAB(offsetsub,sse2,mmx2,sse2,16)
 252
 253 static void x264_weight_cache_mmx2( x264_t *h, x264_weight_t *w )
 254 {
 255     if( w->i_scale == 1<<w->i_denom )
 256     {
 257         if( w->i_offset < 0 )
 258             w->weightfn = h->mc.offsetsub;
 259         else
 260             w->weightfn = h->mc.offsetadd;
 261         for( int i = 0; i < 8; i++ )
 262             w->cachea[i] = abs(w->i_offset<<(BIT_DEPTH-8));
 263         return;
 264     }
 265     w->weightfn = h->mc.weight;
 266     int den1 = 1<<w->i_denom;
 267     int den2 = w->i_scale<<1;
 268     int den3 = 1+(w->i_offset<<(BIT_DEPTH-8+1));
 269     for( int i = 0; i < 8; i++ )
 270     {
 271         w->cachea[i] = den1;
 272         w->cacheb[i] = i&1 ? den3 : den2;
 273     }
 274 }
 275 #else
 276 MC_WEIGHT_WTAB(weight,mmx2,mmx2,mmx2,12)
 277 MC_WEIGHT_WTAB(offsetadd,mmx2,mmx2,mmx2,12)
 278 MC_WEIGHT_WTAB(offsetsub,mmx2,mmx2,mmx2,12)
 279 MC_WEIGHT_WTAB(weight,sse2,mmx2,sse2,16)
 280 MC_WEIGHT_WTAB(offsetadd,sse2,mmx2,mmx2,16)
 281 MC_WEIGHT_WTAB(offsetsub,sse2,mmx2,mmx2,16)
 282 MC_WEIGHT_WTAB(weight,ssse3,ssse3,ssse3,16)
 283
 284 static void x264_weight_cache_mmx2( x264_t *h, x264_weight_t *w )
 285 {
 286     int i;
 287     int16_t den1;
 288
 289     if( w->i_scale == 1<<w->i_denom )
 290     {
 291         if( w->i_offset < 0 )
 292             w->weightfn = h->mc.offsetsub;
 293         else
 294             w->weightfn = h->mc.offsetadd;
 295         memset( w->cachea, abs(w->i_offset), sizeof(w->cachea) );
 296         return;
 297     }
 298     w->weightfn = h->mc.weight;
 299     den1 = 1 << (w->i_denom - 1) | w->i_offset << w->i_denom;
 300     for( i = 0; i < 8; i++ )
 301     {
 302         w->cachea[i] = w->i_scale;
 303         w->cacheb[i] = den1;
 304     }
 305 }
 306
 307 static void x264_weight_cache_ssse3( x264_t *h, x264_weight_t *w )
 308 {
 309     int i, den1;
 310     if( w->i_scale == 1<<w->i_denom )
 311     {
 312         if( w->i_offset < 0 )
 313             w->weightfn = h->mc.offsetsub;
 314         else
 315             w->weightfn = h->mc.offsetadd;
 316
 317         memset( w->cachea, abs( w->i_offset ), sizeof(w->cachea) );
 318         return;
 319     }
 320     w->weightfn = h->mc.weight;
 321     den1 = w->i_scale << (8 - w->i_denom);
 322     for( i = 0; i < 8; i++ )
 323     {
 324         w->cachea[i] = den1;
 325         w->cacheb[i] = w->i_offset;
 326     }
 327 }
 328 #endif // !HIGH_BIT_DEPTH
 329
 330 static const uint8_t hpel_ref0[16] = {0,1,1,1,0,1,1,1,2,3,3,3,0,1,1,1};
 331 static const uint8_t hpel_ref1[16] = {0,0,0,0,2,2,3,2,2,2,3,2,2,2,3,2};
 332
 333 #define MC_LUMA(name,instr1,instr2)\
 334 static void mc_luma_##name( pixel *dst,    intptr_t i_dst_stride,\
 335                             pixel *src[4], intptr_t i_src_stride,\
 336                             int mvx, int mvy,\
 337                             int i_width, int i_height, const x264_weight_t *weight )\
 338 {\
 339     int qpel_idx = ((mvy&3)<<2) + (mvx&3);\
 340     int offset = (mvy>>2)*i_src_stride + (mvx>>2);\
 341     pixel *src1 = src[hpel_ref0[qpel_idx]] + offset + ((mvy&3) == 3) * i_src_stride;\
 342     if( qpel_idx & 5 ) /* qpel interpolation needed */\
 343     {\
 344         pixel *src2 = src[hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);\
 345         x264_pixel_avg_wtab_##instr1[i_width>>2](\
 346                 dst, i_dst_stride, src1, i_src_stride,\
 347                 src2, i_height );\
 348         if( weight->weightfn )\
 349             weight->weightfn[i_width>>2]( dst, i_dst_stride, dst, i_dst_stride, weight, i_height );\
 350     }\
 351     else if( weight->weightfn )\
 352         weight->weightfn[i_width>>2]( dst, i_dst_stride, src1, i_src_stride, weight, i_height );\
 353     else\
 354         x264_mc_copy_wtab_##instr2[i_width>>2](dst, i_dst_stride, src1, i_src_stride, i_height );\
 355 }
 356
 357 MC_LUMA(mmx2,mmx2,mmx)
 358 MC_LUMA(sse2,sse2,sse2)
 359 #if !HIGH_BIT_DEPTH
 360 #if ARCH_X86
 361 MC_LUMA(cache32_mmx2,cache32_mmx2,mmx)
 362 MC_LUMA(cache64_mmx2,cache64_mmx2,mmx)
 363 #endif
 364 MC_LUMA(cache64_sse2,cache64_sse2,sse2)
 365 MC_LUMA(cache64_ssse3,cache64_ssse3,sse2)
 366 #endif // !HIGH_BIT_DEPTH
 367
 368 #define GET_REF(name)\
 369 static pixel *get_ref_##name( pixel *dst,   intptr_t *i_dst_stride,\
 370                               pixel *src[4], intptr_t i_src_stride,\
 371                               int mvx, int mvy,\
 372                               int i_width, int i_height, const x264_weight_t *weight )\
 373 {\
 374     int qpel_idx = ((mvy&3)<<2) + (mvx&3);\
 375     int offset = (mvy>>2)*i_src_stride + (mvx>>2);\
 376     pixel *src1 = src[hpel_ref0[qpel_idx]] + offset + ((mvy&3) == 3) * i_src_stride;\
 377     if( qpel_idx & 5 ) /* qpel interpolation needed */\
 378     {\
 379         pixel *src2 = src[hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);\
 380         x264_pixel_avg_wtab_##name[i_width>>2](\
 381                 dst, *i_dst_stride, src1, i_src_stride,\
 382                 src2, i_height );\
 383         if( weight->weightfn )\
 384             weight->weightfn[i_width>>2]( dst, *i_dst_stride, dst, *i_dst_stride, weight, i_height );\
 385         return dst;\
 386     }\
 387     else if( weight->weightfn )\
 388     {\
 389         weight->weightfn[i_width>>2]( dst, *i_dst_stride, src1, i_src_stride, weight, i_height );\
 390         return dst;\
 391     }\
 392     else\
 393     {\
 394         *i_dst_stride = i_src_stride;\
 395         return src1;\
 396     }\
 397 }
 398
 399 GET_REF(mmx2)
 400 GET_REF(sse2)
 401 #if !HIGH_BIT_DEPTH
 402 #if ARCH_X86
 403 GET_REF(cache32_mmx2)
 404 GET_REF(cache64_mmx2)
 405 #endif
 406 GET_REF(sse2_misalign)
 407 GET_REF(cache64_sse2)
 408 GET_REF(cache64_ssse3)
 409 #endif // !HIGH_BIT_DEPTH
 410
 411 #define HPEL(align, cpu, cpuv, cpuc, cpuh)\
 412 void x264_hpel_filter_v_##cpuv( pixel *dst, pixel *src, int16_t *buf, intptr_t stride, intptr_t width);\
 413 void x264_hpel_filter_c_##cpuc( pixel *dst, int16_t *buf, intptr_t width );\
 414 void x264_hpel_filter_h_##cpuh( pixel *dst, pixel *src, intptr_t width );\
 415 static void x264_hpel_filter_##cpu( pixel *dsth, pixel *dstv, pixel *dstc, pixel *src,\
 416                                     intptr_t stride, int width, int height, int16_t *buf )\
 417 {\
 418     intptr_t realign = (intptr_t)src & (align-1);\
 419     src -= realign;\
 420     dstv -= realign;\
 421     dstc -= realign;\
 422     dsth -= realign;\
 423     width += realign;\
 424     while( height-- )\
 425     {\
 426         x264_hpel_filter_v_##cpuv( dstv, src, buf+8, stride, width );\
 427         x264_hpel_filter_c_##cpuc( dstc, buf+8, width );\
 428         x264_hpel_filter_h_##cpuh( dsth, src, width );\
 429         dsth += stride;\
 430         dstv += stride;\
 431         dstc += stride;\
 432         src  += stride;\
 433     }\
 434     x264_sfence();\
 435 }
 436
 437 HPEL(8, mmx2, mmx2, mmx2, mmx2)
 438 #if HIGH_BIT_DEPTH
 439 HPEL(16, sse2, sse2, sse2, sse2)
 440 #else // !HIGH_BIT_DEPTH
 441 HPEL(16, sse2_amd, mmx2, mmx2, sse2)
 442 #if ARCH_X86_64
 443 void x264_hpel_filter_sse2 ( uint8_t *dsth, uint8_t *dstv, uint8_t *dstc, uint8_t *src, intptr_t stride, int width, int height, int16_t *buf );
 444 void x264_hpel_filter_ssse3( uint8_t *dsth, uint8_t *dstv, uint8_t *dstc, uint8_t *src, intptr_t stride, int width, int height, int16_t *buf );
 445 void x264_hpel_filter_avx  ( uint8_t *dsth, uint8_t *dstv, uint8_t *dstc, uint8_t *src, intptr_t stride, int width, int height, int16_t *buf );
 446 #else
 447 HPEL(16, sse2, sse2, sse2, sse2)
 448 HPEL(16, ssse3, ssse3, ssse3, ssse3)
 449 HPEL(16, avx, avx, avx, avx)
 450 #endif
 451 HPEL(16, sse2_misalign, sse2, sse2_misalign, sse2)
 452 #endif // HIGH_BIT_DEPTH
 453
 454 static void x264_plane_copy_mmx2( pixel *dst, intptr_t i_dst, pixel *src, intptr_t i_src, int w, int h )
 455 {
 456     int c_w = 16/sizeof(pixel) - 1;
 457     if( w < 256 ) { // tiny resolutions don't want non-temporal hints. dunno the exact threshold.
 458         x264_plane_copy_c( dst, i_dst, src, i_src, w, h );
 459     } else if( !(w&c_w) ) {
 460         x264_plane_copy_core_mmx2( dst, i_dst, src, i_src, w, h );
 461     } else if( i_src > 0 ) {
 462         // have to use plain memcpy on the last line (in memory order) to avoid overreading src
 463         x264_plane_copy_core_mmx2( dst, i_dst, src, i_src, (w+c_w)&~c_w, h-1 );
 464         memcpy( dst+i_dst*(h-1), src+i_src*(h-1), w*sizeof(pixel) );
 465     } else {
 466         memcpy( dst, src, w*sizeof(pixel) );
 467         x264_plane_copy_core_mmx2( dst+i_dst, i_dst, src+i_src, i_src, (w+c_w)&~c_w, h-1 );
 468     }
 469 }
 470
 471 #define PLANE_INTERLEAVE(cpu) \
 472 static void x264_plane_copy_interleave_##cpu( pixel *dst,  intptr_t i_dst,\
 473                                               pixel *srcu, intptr_t i_srcu,\
 474                                               pixel *srcv, intptr_t i_srcv, int w, int h )\
 475 {\
 476     if( !(w&15) ) {\
 477         x264_plane_copy_interleave_core_##cpu( dst, i_dst, srcu, i_srcu, srcv, i_srcv, w, h );\
 478     } else if( w < 16 || (i_srcu ^ i_srcv) ) {\
 479         x264_plane_copy_interleave_c( dst, i_dst, srcu, i_srcu, srcv, i_srcv, w, h );\
 480     } else if( i_srcu > 0 ) {\
 481         x264_plane_copy_interleave_core_##cpu( dst, i_dst, srcu, i_srcu, srcv, i_srcv, (w+15)&~15, h-1 );\
 482         x264_plane_copy_interleave_c( dst+i_dst*(h-1), 0, srcu+i_srcu*(h-1), 0, srcv+i_srcv*(h-1), 0, w, 1 );\
 483     } else {\
 484         x264_plane_copy_interleave_c( dst, 0, srcu, 0, srcv, 0, w, 1 );\
 485         x264_plane_copy_interleave_core_##cpu( dst+i_dst, i_dst, srcu+i_srcu, i_srcu, srcv+i_srcv, i_srcv, (w+15)&~15, h-1 );\
 486     }\
 487 }
 488
 489 PLANE_INTERLEAVE(mmx2)
 490 PLANE_INTERLEAVE(sse2)
 491 #if HIGH_BIT_DEPTH
 492 PLANE_INTERLEAVE(avx)
 493 #endif
 494
 495 void x264_mc_init_mmx( int cpu, x264_mc_functions_t *pf )
 496 {
 497     if( !(cpu&X264_CPU_MMX) )
 498         return;
 499
 500     pf->load_deinterleave_chroma_fenc = x264_load_deinterleave_chroma_fenc_mmx;
 501     pf->load_deinterleave_chroma_fdec = x264_load_deinterleave_chroma_fdec_mmx;
 502
 503     pf->plane_copy_deinterleave = x264_plane_copy_deinterleave_mmx;
 504
 505     pf->copy_16x16_unaligned = x264_mc_copy_w16_mmx;
 506     pf->copy[PIXEL_16x16] = x264_mc_copy_w16_mmx;
 507     pf->copy[PIXEL_8x8]   = x264_mc_copy_w8_mmx;
 508     pf->copy[PIXEL_4x4]   = x264_mc_copy_w4_mmx;
 509     pf->memcpy_aligned  = x264_memcpy_aligned_mmx;
 510     pf->memzero_aligned = x264_memzero_aligned_mmx;
 511     pf->integral_init4v = x264_integral_init4v_mmx;
 512     pf->integral_init8v = x264_integral_init8v_mmx;
 513
 514     if( !(cpu&X264_CPU_MMX2) )
 515         return;
 516
 517     pf->prefetch_fenc_420 = x264_prefetch_fenc_420_mmx2;
 518     pf->prefetch_fenc_422 = x264_prefetch_fenc_422_mmx2;
 519     pf->prefetch_ref  = x264_prefetch_ref_mmx2;
 520
 521     pf->plane_copy = x264_plane_copy_mmx2;
 522     pf->plane_copy_interleave = x264_plane_copy_interleave_mmx2;
 523     pf->store_interleave_chroma = x264_store_interleave_chroma_mmx2;
 524
 525     pf->avg[PIXEL_16x16] = x264_pixel_avg_16x16_mmx2;
 526     pf->avg[PIXEL_16x8]  = x264_pixel_avg_16x8_mmx2;
 527     pf->avg[PIXEL_8x16]  = x264_pixel_avg_8x16_mmx2;
 528     pf->avg[PIXEL_8x8]   = x264_pixel_avg_8x8_mmx2;
 529     pf->avg[PIXEL_8x4]   = x264_pixel_avg_8x4_mmx2;
 530     pf->avg[PIXEL_4x16]  = x264_pixel_avg_4x16_mmx2;
 531     pf->avg[PIXEL_4x8]   = x264_pixel_avg_4x8_mmx2;
 532     pf->avg[PIXEL_4x4]   = x264_pixel_avg_4x4_mmx2;
 533     pf->avg[PIXEL_4x2]   = x264_pixel_avg_4x2_mmx2;
 534
 535     pf->mc_luma = mc_luma_mmx2;
 536     pf->get_ref = get_ref_mmx2;
 537     pf->mc_chroma = x264_mc_chroma_mmx2;
 538     pf->hpel_filter = x264_hpel_filter_mmx2;
 539     pf->weight = x264_mc_weight_wtab_mmx2;
 540     pf->weight_cache = x264_weight_cache_mmx2;
 541     pf->offsetadd = x264_mc_offsetadd_wtab_mmx2;
 542     pf->offsetsub = x264_mc_offsetsub_wtab_mmx2;
 543
 544     pf->frame_init_lowres_core = x264_frame_init_lowres_core_mmx2;
 545
 546 #if HIGH_BIT_DEPTH
 547 #if ARCH_X86 // all x86_64 cpus with cacheline split issues use sse2 instead
 548     if( cpu&(X264_CPU_CACHELINE_32|X264_CPU_CACHELINE_64) )
 549         pf->frame_init_lowres_core = x264_frame_init_lowres_core_cache32_mmx2;
 550 #endif
 551
 552     if( !(cpu&X264_CPU_SSE2) )
 553         return;
 554
 555     pf->frame_init_lowres_core = x264_frame_init_lowres_core_sse2;
 556
 557     pf->load_deinterleave_chroma_fenc = x264_load_deinterleave_chroma_fenc_sse2;
 558     pf->load_deinterleave_chroma_fdec = x264_load_deinterleave_chroma_fdec_sse2;
 559
 560     pf->plane_copy_interleave   = x264_plane_copy_interleave_sse2;
 561     pf->plane_copy_deinterleave = x264_plane_copy_deinterleave_sse2;
 562
 563     if( cpu&X264_CPU_SSE2_IS_FAST )
 564     {
 565         pf->get_ref = get_ref_sse2;
 566         pf->mc_luma = mc_luma_sse2;
 567         pf->hpel_filter = x264_hpel_filter_sse2;
 568     }
 569
 570     pf->memcpy_aligned  = x264_memcpy_aligned_sse2;
 571     pf->memzero_aligned = x264_memzero_aligned_sse2;
 572     pf->integral_init4v = x264_integral_init4v_sse2;
 573     pf->integral_init8v = x264_integral_init8v_sse2;
 574     pf->mbtree_propagate_cost = x264_mbtree_propagate_cost_sse2;
 575     pf->store_interleave_chroma = x264_store_interleave_chroma_sse2;
 576     pf->offsetadd = x264_mc_offsetadd_wtab_sse2;
 577     pf->offsetsub = x264_mc_offsetsub_wtab_sse2;
 578
 579     if( cpu&X264_CPU_SSE2_IS_SLOW )
 580         return;
 581
 582     pf->avg[PIXEL_16x16] = x264_pixel_avg_16x16_sse2;
 583     pf->avg[PIXEL_16x8]  = x264_pixel_avg_16x8_sse2;
 584     pf->avg[PIXEL_8x16]  = x264_pixel_avg_8x16_sse2;
 585     pf->avg[PIXEL_8x8]   = x264_pixel_avg_8x8_sse2;
 586     pf->avg[PIXEL_8x4]   = x264_pixel_avg_8x4_sse2;
 587     pf->avg[PIXEL_4x16]  = x264_pixel_avg_4x16_sse2;
 588     pf->avg[PIXEL_4x8]   = x264_pixel_avg_4x8_sse2;
 589     pf->avg[PIXEL_4x4]   = x264_pixel_avg_4x4_sse2;
 590     pf->avg[PIXEL_4x2]   = x264_pixel_avg_4x2_sse2;
 591
 592     pf->copy[PIXEL_16x16] = x264_mc_copy_w16_aligned_sse2;
 593     pf->weight = x264_mc_weight_wtab_sse2;
 594
 595     if( !(cpu&X264_CPU_STACK_MOD4) )
 596         pf->mc_chroma = x264_mc_chroma_sse2;
 597
 598     if( !(cpu&X264_CPU_SSSE3) )
 599         return;
 600
 601     pf->frame_init_lowres_core = x264_frame_init_lowres_core_ssse3;
 602
 603     if( (cpu&X264_CPU_SHUFFLE_IS_FAST) && !(cpu&X264_CPU_SLOW_ATOM) )
 604         pf->integral_init4v = x264_integral_init4v_ssse3;
 605
 606     if( !(cpu&X264_CPU_AVX) )
 607         return;
 608
 609     pf->frame_init_lowres_core = x264_frame_init_lowres_core_avx;
 610     pf->load_deinterleave_chroma_fenc = x264_load_deinterleave_chroma_fenc_avx;
 611     pf->load_deinterleave_chroma_fdec = x264_load_deinterleave_chroma_fdec_avx;
 612     pf->plane_copy_interleave        = x264_plane_copy_interleave_avx;
 613     pf->plane_copy_deinterleave      = x264_plane_copy_deinterleave_avx;
 614     pf->store_interleave_chroma      = x264_store_interleave_chroma_avx;
 615
 616     if( !(cpu&X264_CPU_STACK_MOD4) )
 617         pf->mc_chroma = x264_mc_chroma_avx;
 618
 619     if( cpu&X264_CPU_XOP )
 620         pf->frame_init_lowres_core = x264_frame_init_lowres_core_xop;
 621 #else // !HIGH_BIT_DEPTH
 622
 623 #if ARCH_X86 // all x86_64 cpus with cacheline split issues use sse2 instead
 624     if( cpu&X264_CPU_CACHELINE_32 )
 625     {
 626         pf->mc_luma = mc_luma_cache32_mmx2;
 627         pf->get_ref = get_ref_cache32_mmx2;
 628         pf->frame_init_lowres_core = x264_frame_init_lowres_core_cache32_mmx2;
 629     }
 630     else if( cpu&X264_CPU_CACHELINE_64 )
 631     {
 632         pf->mc_luma = mc_luma_cache64_mmx2;
 633         pf->get_ref = get_ref_cache64_mmx2;
 634         pf->frame_init_lowres_core = x264_frame_init_lowres_core_cache32_mmx2;
 635     }
 636 #endif
 637
 638     if( !(cpu&X264_CPU_SSE2) )
 639         return;
 640
 641     pf->memcpy_aligned = x264_memcpy_aligned_sse2;
 642     pf->memzero_aligned = x264_memzero_aligned_sse2;
 643     pf->integral_init4v = x264_integral_init4v_sse2;
 644     pf->integral_init8v = x264_integral_init8v_sse2;
 645     pf->hpel_filter = x264_hpel_filter_sse2_amd;
 646     pf->mbtree_propagate_cost = x264_mbtree_propagate_cost_sse2;
 647
 648     if( cpu&X264_CPU_SSE2_IS_SLOW )
 649         return;
 650
 651     pf->weight = x264_mc_weight_wtab_sse2;
 652     if( !(cpu&X264_CPU_SLOW_ATOM) )
 653     {
 654         pf->offsetadd = x264_mc_offsetadd_wtab_sse2;
 655         pf->offsetsub = x264_mc_offsetsub_wtab_sse2;
 656     }
 657
 658     pf->copy[PIXEL_16x16] = x264_mc_copy_w16_aligned_sse2;
 659     pf->avg[PIXEL_16x16] = x264_pixel_avg_16x16_sse2;
 660     pf->avg[PIXEL_16x8]  = x264_pixel_avg_16x8_sse2;
 661     pf->avg[PIXEL_8x16] = x264_pixel_avg_8x16_sse2;
 662     pf->avg[PIXEL_8x8]  = x264_pixel_avg_8x8_sse2;
 663     pf->avg[PIXEL_8x4]  = x264_pixel_avg_8x4_sse2;
 664     pf->hpel_filter = x264_hpel_filter_sse2;
 665     if( cpu&X264_CPU_SSE_MISALIGN )
 666         pf->hpel_filter = x264_hpel_filter_sse2_misalign;
 667     pf->frame_init_lowres_core = x264_frame_init_lowres_core_sse2;
 668     if( !(cpu&X264_CPU_STACK_MOD4) )
 669         pf->mc_chroma = x264_mc_chroma_sse2;
 670
 671     if( cpu&X264_CPU_SSE2_IS_FAST )
 672     {
 673         pf->store_interleave_chroma = x264_store_interleave_chroma_sse2; // FIXME sse2fast? sse2medium?
 674         pf->load_deinterleave_chroma_fenc = x264_load_deinterleave_chroma_fenc_sse2;
 675         pf->load_deinterleave_chroma_fdec = x264_load_deinterleave_chroma_fdec_sse2;
 676         pf->plane_copy_interleave   = x264_plane_copy_interleave_sse2;
 677         pf->plane_copy_deinterleave = x264_plane_copy_deinterleave_sse2;
 678         pf->mc_luma = mc_luma_sse2;
 679         pf->get_ref = get_ref_sse2;
 680         if( cpu&X264_CPU_CACHELINE_64 )
 681         {
 682             pf->mc_luma = mc_luma_cache64_sse2;
 683             pf->get_ref = get_ref_cache64_sse2;
 684         }
 685         if( cpu&X264_CPU_SSE_MISALIGN )
 686         {
 687             pf->get_ref = get_ref_sse2_misalign;
 688             if( !(cpu&X264_CPU_STACK_MOD4) )
 689                 pf->mc_chroma = x264_mc_chroma_sse2_misalign;
 690         }
 691     }
 692
 693     if( !(cpu&X264_CPU_SSSE3) )
 694         return;
 695
 696     pf->avg[PIXEL_16x16] = x264_pixel_avg_16x16_ssse3;
 697     pf->avg[PIXEL_16x8]  = x264_pixel_avg_16x8_ssse3;
 698     pf->avg[PIXEL_8x16]  = x264_pixel_avg_8x16_ssse3;
 699     pf->avg[PIXEL_8x8]   = x264_pixel_avg_8x8_ssse3;
 700     pf->avg[PIXEL_8x4]   = x264_pixel_avg_8x4_ssse3;
 701     pf->avg[PIXEL_4x16]  = x264_pixel_avg_4x16_ssse3;
 702     pf->avg[PIXEL_4x8]   = x264_pixel_avg_4x8_ssse3;
 703     pf->avg[PIXEL_4x4]   = x264_pixel_avg_4x4_ssse3;
 704     pf->avg[PIXEL_4x2]   = x264_pixel_avg_4x2_ssse3;
 705
 706     pf->load_deinterleave_chroma_fenc = x264_load_deinterleave_chroma_fenc_ssse3;
 707     pf->load_deinterleave_chroma_fdec = x264_load_deinterleave_chroma_fdec_ssse3;
 708     pf->plane_copy_deinterleave = x264_plane_copy_deinterleave_ssse3;
 709
 710     pf->hpel_filter = x264_hpel_filter_ssse3;
 711     pf->frame_init_lowres_core = x264_frame_init_lowres_core_ssse3;
 712     if( !(cpu&X264_CPU_STACK_MOD4) )
 713         pf->mc_chroma = x264_mc_chroma_ssse3;
 714
 715     if( cpu&X264_CPU_CACHELINE_64 )
 716     {
 717         if( !(cpu&X264_CPU_STACK_MOD4) )
 718             pf->mc_chroma = x264_mc_chroma_ssse3_cache64;
 719         pf->mc_luma = mc_luma_cache64_ssse3;
 720         pf->get_ref = get_ref_cache64_ssse3;
 721
 722         /* ssse3 weight is slower on Nehalem, so only assign here. */
 723         pf->weight_cache = x264_weight_cache_ssse3;
 724         pf->weight = x264_mc_weight_wtab_ssse3;
 725     }
 726
 727     if( (cpu&X264_CPU_SHUFFLE_IS_FAST) && !(cpu&X264_CPU_SLOW_ATOM) )
 728         pf->integral_init4v = x264_integral_init4v_ssse3;
 729
 730     if( !(cpu&X264_CPU_SSE4) )
 731         return;
 732
 733     pf->integral_init4h = x264_integral_init4h_sse4;
 734     pf->integral_init8h = x264_integral_init8h_sse4;
 735
 736     if( !(cpu&X264_CPU_AVX) )
 737         return;
 738
 739     pf->frame_init_lowres_core = x264_frame_init_lowres_core_avx;
 740     pf->integral_init8h = x264_integral_init8h_avx;
 741     pf->hpel_filter = x264_hpel_filter_avx;
 742
 743     /* ssse3 weight seems to be faster again on Sandy Bridge and Bulldozer. */
 744     pf->weight_cache = x264_weight_cache_ssse3;
 745     pf->weight = x264_mc_weight_wtab_ssse3;
 746     if( !(cpu&X264_CPU_STACK_MOD4) )
 747         pf->mc_chroma = x264_mc_chroma_avx;
 748
 749     if( cpu&X264_CPU_XOP )
 750         pf->frame_init_lowres_core = x264_frame_init_lowres_core_xop;
 751 #endif // HIGH_BIT_DEPTH
 752
 753     if( !(cpu&X264_CPU_AVX) )
 754         return;
 755     pf->mbtree_propagate_cost = x264_mbtree_propagate_cost_avx;
 756
 757     if( !(cpu&X264_CPU_FMA4) )
 758         return;
 759     pf->mbtree_propagate_cost = x264_mbtree_propagate_cost_fma4;
 760 }