git.sesse.net Git - x264/blob - common/aarch64/mc-c.c

   1 /*****************************************************************************
   2  * mc-c.c: aarch64 motion compensation
   3  *****************************************************************************
   4  * Copyright (C) 2009-2015 x264 project
   5  *
   6  * Authors: David Conrad <lessen42@gmail.com>
   7  *          Janne Grunau <janne-x264@jannau.net>
   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  * This program is also available under a commercial proprietary license.
  24  * For more information, contact us at licensing@x264.com.
  25  *****************************************************************************/
  26
  27 #include "common/common.h"
  28 #include "mc.h"
  29
  30 void x264_prefetch_ref_aarch64( uint8_t *, intptr_t, int );
  31 void x264_prefetch_fenc_420_aarch64( uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  32 void x264_prefetch_fenc_422_aarch64( uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  33
  34 void *x264_memcpy_aligned_neon( void *dst, const void *src, size_t n );
  35 void x264_memzero_aligned_neon( void *dst, size_t n );
  36
  37 void x264_pixel_avg_16x16_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  38 void x264_pixel_avg_16x8_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  39 void x264_pixel_avg_8x16_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  40 void x264_pixel_avg_8x8_neon  ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  41 void x264_pixel_avg_8x4_neon  ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  42 void x264_pixel_avg_4x16_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  43 void x264_pixel_avg_4x8_neon  ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  44 void x264_pixel_avg_4x4_neon  ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  45 void x264_pixel_avg_4x2_neon  ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  46
  47 void x264_pixel_avg2_w4_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, int );
  48 void x264_pixel_avg2_w8_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, int );
  49 void x264_pixel_avg2_w16_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, int );
  50 void x264_pixel_avg2_w20_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, int );
  51
  52 void x264_plane_copy_neon( pixel *dst, intptr_t i_dst,
  53                            pixel *src, intptr_t i_src, int w, int h );
  54 void x264_plane_copy_deinterleave_neon(  pixel *dstu, intptr_t i_dstu,
  55                                          pixel *dstv, intptr_t i_dstv,
  56                                          pixel *src,  intptr_t i_src, int w, int h );
  57 void x264_plane_copy_deinterleave_rgb_neon( pixel *dsta, intptr_t i_dsta,
  58                                             pixel *dstb, intptr_t i_dstb,
  59                                             pixel *dstc, intptr_t i_dstc,
  60                                             pixel *src,  intptr_t i_src, int pw, int w, int h );
  61 void x264_plane_copy_interleave_neon( pixel *dst,  intptr_t i_dst,
  62                                       pixel *srcu, intptr_t i_srcu,
  63                                       pixel *srcv, intptr_t i_srcv, int w, int h );
  64
  65 void x264_store_interleave_chroma_neon( pixel *dst, intptr_t i_dst, pixel *srcu, pixel *srcv, int height );
  66 void x264_load_deinterleave_chroma_fdec_neon( pixel *dst, pixel *src, intptr_t i_src, int height );
  67 void x264_load_deinterleave_chroma_fenc_neon( pixel *dst, pixel *src, intptr_t i_src, int height );
  68
  69 #define MC_WEIGHT(func)\
  70 void x264_mc_weight_w20##func##_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, const x264_weight_t *, int );\
  71 void x264_mc_weight_w16##func##_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, const x264_weight_t *, int );\
  72 void x264_mc_weight_w8##func##_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, const x264_weight_t *, int );\
  73 void x264_mc_weight_w4##func##_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, const x264_weight_t *, int );\
  74 \
  75 static void (* x264_mc##func##_wtab_neon[6])( uint8_t *, intptr_t, uint8_t *, intptr_t, const x264_weight_t *, int ) =\
  76 {\
  77     x264_mc_weight_w4##func##_neon,\
  78     x264_mc_weight_w4##func##_neon,\
  79     x264_mc_weight_w8##func##_neon,\
  80     x264_mc_weight_w16##func##_neon,\
  81     x264_mc_weight_w16##func##_neon,\
  82     x264_mc_weight_w20##func##_neon,\
  83 };
  84
  85 MC_WEIGHT()
  86 MC_WEIGHT(_nodenom)
  87 MC_WEIGHT(_offsetadd)
  88 MC_WEIGHT(_offsetsub)
  89
  90 void x264_mc_copy_w4_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  91 void x264_mc_copy_w8_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  92 void x264_mc_copy_w16_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  93
  94 void x264_mc_chroma_neon( uint8_t *, uint8_t *, intptr_t, uint8_t *, intptr_t, int, int, int, int );
  95 void integral_init4h_neon( uint16_t *, uint8_t *, intptr_t );
  96 void integral_init4v_neon( uint16_t *, uint16_t *, intptr_t );
  97 void integral_init8h_neon( uint16_t *, uint8_t *, intptr_t );
  98 void integral_init8v_neon( uint16_t *, intptr_t );
  99 void x264_frame_init_lowres_core_neon( uint8_t *, uint8_t *, uint8_t *, uint8_t *, uint8_t *, intptr_t, intptr_t, int, int );
 100
 101 void x264_mbtree_propagate_cost_neon( int16_t *, uint16_t *, uint16_t *, uint16_t *, uint16_t *, float *, int );
 102
 103 #if !HIGH_BIT_DEPTH
 104 static void x264_weight_cache_neon( x264_t *h, x264_weight_t *w )
 105 {
 106     if( w->i_scale == 1<<w->i_denom )
 107     {
 108         if( w->i_offset < 0 )
 109         {
 110             w->weightfn = x264_mc_offsetsub_wtab_neon;
 111             w->cachea[0] = -w->i_offset;
 112         }
 113         else
 114         {
 115             w->weightfn = x264_mc_offsetadd_wtab_neon;
 116             w->cachea[0] = w->i_offset;
 117         }
 118     }
 119     else if( !w->i_denom )
 120         w->weightfn = x264_mc_nodenom_wtab_neon;
 121     else
 122         w->weightfn = x264_mc_wtab_neon;
 123 }
 124
 125 static void (* const x264_pixel_avg_wtab_neon[6])( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, int ) =
 126 {
 127     NULL,
 128     x264_pixel_avg2_w4_neon,
 129     x264_pixel_avg2_w8_neon,
 130     x264_pixel_avg2_w16_neon,   // no slower than w12, so no point in a separate function
 131     x264_pixel_avg2_w16_neon,
 132     x264_pixel_avg2_w20_neon,
 133 };
 134
 135 static void (* const x264_mc_copy_wtab_neon[5])( uint8_t *, intptr_t, uint8_t *, intptr_t, int ) =
 136 {
 137     NULL,
 138     x264_mc_copy_w4_neon,
 139     x264_mc_copy_w8_neon,
 140     NULL,
 141     x264_mc_copy_w16_neon,
 142 };
 143
 144 static void mc_luma_neon( uint8_t *dst,    intptr_t i_dst_stride,
 145                           uint8_t *src[4], intptr_t i_src_stride,
 146                           int mvx, int mvy,
 147                           int i_width, int i_height, const x264_weight_t *weight )
 148 {
 149     int qpel_idx = ((mvy&3)<<2) + (mvx&3);
 150     intptr_t offset = (mvy>>2)*i_src_stride + (mvx>>2);
 151     uint8_t *src1 = src[x264_hpel_ref0[qpel_idx]] + offset;
 152     if ( (mvy&3) == 3 )             // explict if() to force conditional add
 153         src1 += i_src_stride;
 154
 155     if( qpel_idx & 5 ) /* qpel interpolation needed */
 156     {
 157         uint8_t *src2 = src[x264_hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);
 158         x264_pixel_avg_wtab_neon[i_width>>2](
 159                 dst, i_dst_stride, src1, i_src_stride,
 160                 src2, i_height );
 161         if( weight->weightfn )
 162             weight->weightfn[i_width>>2]( dst, i_dst_stride, dst, i_dst_stride, weight, i_height );
 163     }
 164     else if( weight->weightfn )
 165         weight->weightfn[i_width>>2]( dst, i_dst_stride, src1, i_src_stride, weight, i_height );
 166     else
 167         x264_mc_copy_wtab_neon[i_width>>2]( dst, i_dst_stride, src1, i_src_stride, i_height );
 168 }
 169
 170 static uint8_t *get_ref_neon( uint8_t *dst,   intptr_t *i_dst_stride,
 171                               uint8_t *src[4], intptr_t i_src_stride,
 172                               int mvx, int mvy,
 173                               int i_width, int i_height, const x264_weight_t *weight )
 174 {
 175     int qpel_idx = ((mvy&3)<<2) + (mvx&3);
 176     intptr_t offset = (mvy>>2)*i_src_stride + (mvx>>2);
 177     uint8_t *src1 = src[x264_hpel_ref0[qpel_idx]] + offset;
 178     if ( (mvy&3) == 3 )             // explict if() to force conditional add
 179         src1 += i_src_stride;
 180
 181     if( qpel_idx & 5 ) /* qpel interpolation needed */
 182     {
 183         uint8_t *src2 = src[x264_hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);
 184         x264_pixel_avg_wtab_neon[i_width>>2](
 185                 dst, *i_dst_stride, src1, i_src_stride,
 186                 src2, i_height );
 187         if( weight->weightfn )
 188             weight->weightfn[i_width>>2]( dst, *i_dst_stride, dst, *i_dst_stride, weight, i_height );
 189         return dst;
 190     }
 191     else if( weight->weightfn )
 192     {
 193         weight->weightfn[i_width>>2]( dst, *i_dst_stride, src1, i_src_stride, weight, i_height );
 194         return dst;
 195     }
 196     else
 197     {
 198         *i_dst_stride = i_src_stride;
 199         return src1;
 200     }
 201 }
 202
 203 void x264_hpel_filter_neon( uint8_t *dsth, uint8_t *dstv, uint8_t *dstc,
 204                             uint8_t *src, intptr_t stride, int width,
 205                             int height, int16_t *buf );
 206 #endif // !HIGH_BIT_DEPTH
 207
 208 #define CLIP_ADD(s,x) (s) = X264_MIN((s)+(x),(1<<15)-1)
 209 #define CLIP_ADD2(s,x)\
 210 do\
 211 {\
 212     CLIP_ADD((s)[0], (x)[0]);\
 213     CLIP_ADD((s)[1], (x)[1]);\
 214 } while(0)
 215
 216 void x264_mbtree_propagate_list_internal_neon( int16_t (*mvs)[2],
 217                                                int16_t *propagate_amount,
 218                                                uint16_t *lowres_costs,
 219                                                int16_t *output,
 220                                                int bipred_weight, int mb_y,
 221                                                int len );
 222
 223 static void x264_mbtree_propagate_list_neon( x264_t *h, uint16_t *ref_costs,
 224                                              int16_t (*mvs)[2],
 225                                              int16_t *propagate_amount,
 226                                              uint16_t *lowres_costs,
 227                                              int bipred_weight, int mb_y,
 228                                              int len, int list )
 229 {
 230     int16_t *current = h->scratch_buffer2;
 231
 232     x264_mbtree_propagate_list_internal_neon( mvs, propagate_amount,
 233                                               lowres_costs, current,
 234                                               bipred_weight, mb_y, len );
 235
 236     unsigned stride = h->mb.i_mb_stride;
 237     unsigned width = h->mb.i_mb_width;
 238     unsigned height = h->mb.i_mb_height;
 239
 240     for( unsigned i = 0; i < len; current += 32 )
 241     {
 242         int end = X264_MIN( i+8, len );
 243         for( ; i < end; i++, current += 2 )
 244         {
 245             if( !(lowres_costs[i] & (1 << (list+LOWRES_COST_SHIFT))) )
 246                 continue;
 247
 248             unsigned mbx = current[0];
 249             unsigned mby = current[1];
 250             unsigned idx0 = mbx + mby * stride;
 251             unsigned idx2 = idx0 + stride;
 252
 253             /* Shortcut for the simple/common case of zero MV */
 254             if( !M32( mvs[i] ) )
 255             {
 256                 CLIP_ADD( ref_costs[idx0], current[16] );
 257                 continue;
 258             }
 259
 260             if( mbx < width-1 && mby < height-1 )
 261             {
 262                 CLIP_ADD2( ref_costs+idx0, current+16 );
 263                 CLIP_ADD2( ref_costs+idx2, current+32 );
 264             }
 265             else
 266             {
 267                 /* Note: this takes advantage of unsigned representation to
 268                  * catch negative mbx/mby. */
 269                 if( mby < height )
 270                 {
 271                     if( mbx < width )
 272                         CLIP_ADD( ref_costs[idx0+0], current[16] );
 273                     if( mbx+1 < width )
 274                         CLIP_ADD( ref_costs[idx0+1], current[17] );
 275                 }
 276                 if( mby+1 < height )
 277                 {
 278                     if( mbx < width )
 279                         CLIP_ADD( ref_costs[idx2+0], current[32] );
 280                     if( mbx+1 < width )
 281                         CLIP_ADD( ref_costs[idx2+1], current[33] );
 282                 }
 283             }
 284         }
 285     }
 286 }
 287
 288 #undef CLIP_ADD
 289 #undef CLIP_ADD2
 290
 291 void x264_mc_init_aarch64( int cpu, x264_mc_functions_t *pf )
 292 {
 293 #if !HIGH_BIT_DEPTH
 294     if( cpu&X264_CPU_ARMV8 )
 295     {
 296         pf->prefetch_fenc_420 = x264_prefetch_fenc_420_aarch64;
 297         pf->prefetch_fenc_422 = x264_prefetch_fenc_422_aarch64;
 298         pf->prefetch_ref      = x264_prefetch_ref_aarch64;
 299     }
 300
 301     if( !(cpu&X264_CPU_NEON) )
 302         return;
 303
 304     pf->copy_16x16_unaligned = x264_mc_copy_w16_neon;
 305     pf->copy[PIXEL_16x16]    = x264_mc_copy_w16_neon;
 306     pf->copy[PIXEL_8x8]      = x264_mc_copy_w8_neon;
 307     pf->copy[PIXEL_4x4]      = x264_mc_copy_w4_neon;
 308
 309     pf->plane_copy                  = x264_plane_copy_neon;
 310     pf->plane_copy_deinterleave     = x264_plane_copy_deinterleave_neon;
 311     pf->plane_copy_deinterleave_rgb = x264_plane_copy_deinterleave_rgb_neon;
 312     pf->plane_copy_interleave       = x264_plane_copy_interleave_neon;
 313
 314     pf->load_deinterleave_chroma_fdec = x264_load_deinterleave_chroma_fdec_neon;
 315     pf->load_deinterleave_chroma_fenc = x264_load_deinterleave_chroma_fenc_neon;
 316     pf->store_interleave_chroma       = x264_store_interleave_chroma_neon;
 317
 318     pf->avg[PIXEL_16x16] = x264_pixel_avg_16x16_neon;
 319     pf->avg[PIXEL_16x8]  = x264_pixel_avg_16x8_neon;
 320     pf->avg[PIXEL_8x16]  = x264_pixel_avg_8x16_neon;
 321     pf->avg[PIXEL_8x8]   = x264_pixel_avg_8x8_neon;
 322     pf->avg[PIXEL_8x4]   = x264_pixel_avg_8x4_neon;
 323     pf->avg[PIXEL_4x16]  = x264_pixel_avg_4x16_neon;
 324     pf->avg[PIXEL_4x8]   = x264_pixel_avg_4x8_neon;
 325     pf->avg[PIXEL_4x4]   = x264_pixel_avg_4x4_neon;
 326     pf->avg[PIXEL_4x2]   = x264_pixel_avg_4x2_neon;
 327
 328     pf->weight       = x264_mc_wtab_neon;
 329     pf->offsetadd    = x264_mc_offsetadd_wtab_neon;
 330     pf->offsetsub    = x264_mc_offsetsub_wtab_neon;
 331     pf->weight_cache = x264_weight_cache_neon;
 332
 333     pf->mc_chroma = x264_mc_chroma_neon;
 334     pf->mc_luma = mc_luma_neon;
 335     pf->get_ref = get_ref_neon;
 336     pf->hpel_filter = x264_hpel_filter_neon;
 337     pf->frame_init_lowres_core = x264_frame_init_lowres_core_neon;
 338
 339     pf->integral_init4h = integral_init4h_neon;
 340     pf->integral_init8h = integral_init8h_neon;
 341     pf->integral_init4v = integral_init4v_neon;
 342     pf->integral_init8v = integral_init8v_neon;
 343
 344     pf->mbtree_propagate_cost = x264_mbtree_propagate_cost_neon;
 345     pf->mbtree_propagate_list = x264_mbtree_propagate_list_neon;
 346
 347     pf->memcpy_aligned  = x264_memcpy_aligned_neon;
 348     pf->memzero_aligned = x264_memzero_aligned_neon;
 349 #endif // !HIGH_BIT_DEPTH
 350 }