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

   1 /*****************************************************************************
   2  * mc-c.c: arm motion compensation
   3  *****************************************************************************
   4  * Copyright (C) 2009-2016 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_arm( uint8_t *, intptr_t, int );
  31 void x264_prefetch_fenc_arm( uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  32
  33 void *x264_memcpy_aligned_neon( void *dst, const void *src, size_t n );
  34 void x264_memzero_aligned_neon( void *dst, size_t n );
  35
  36 void x264_pixel_avg_16x16_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  37 void x264_pixel_avg_16x8_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  38 void x264_pixel_avg_8x16_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  39 void x264_pixel_avg_8x8_neon  ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  40 void x264_pixel_avg_8x4_neon  ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  41 void x264_pixel_avg_4x16_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  42 void x264_pixel_avg_4x8_neon  ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  43 void x264_pixel_avg_4x4_neon  ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  44 void x264_pixel_avg_4x2_neon  ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  45
  46 void x264_pixel_avg2_w4_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, int );
  47 void x264_pixel_avg2_w8_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, int );
  48 void x264_pixel_avg2_w16_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, int );
  49 void x264_pixel_avg2_w20_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, int );
  50
  51 void x264_plane_copy_neon( pixel *dst, intptr_t i_dst,
  52                            pixel *src, intptr_t i_src, int w, int h );
  53 void x264_plane_copy_deinterleave_neon(  pixel *dstu, intptr_t i_dstu,
  54                                          pixel *dstv, intptr_t i_dstv,
  55                                          pixel *src,  intptr_t i_src, int w, int h );
  56 void x264_plane_copy_deinterleave_rgb_neon( pixel *dsta, intptr_t i_dsta,
  57                                             pixel *dstb, intptr_t i_dstb,
  58                                             pixel *dstc, intptr_t i_dstc,
  59                                             pixel *src,  intptr_t i_src, int pw, int w, int h );
  60 void x264_plane_copy_interleave_neon( pixel *dst,  intptr_t i_dst,
  61                                       pixel *srcu, intptr_t i_srcu,
  62                                       pixel *srcv, intptr_t i_srcv, int w, int h );
  63 void x264_plane_copy_swap_neon( pixel *dst, intptr_t i_dst,
  64                                 pixel *src, intptr_t i_src, int w, int h );
  65
  66 void x264_store_interleave_chroma_neon( pixel *dst, intptr_t i_dst, pixel *srcu, pixel *srcv, int height );
  67 void x264_load_deinterleave_chroma_fdec_neon( pixel *dst, pixel *src, intptr_t i_src, int height );
  68 void x264_load_deinterleave_chroma_fenc_neon( pixel *dst, pixel *src, intptr_t i_src, int height );
  69
  70 #if !HIGH_BIT_DEPTH
  71 #define MC_WEIGHT(func)\
  72 void x264_mc_weight_w20##func##_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, const x264_weight_t *, int );\
  73 void x264_mc_weight_w16##func##_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, const x264_weight_t *, int );\
  74 void x264_mc_weight_w8##func##_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, const x264_weight_t *, int );\
  75 void x264_mc_weight_w4##func##_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, const x264_weight_t *, int );\
  76 \
  77 static weight_fn_t x264_mc##func##_wtab_neon[6] =\
  78 {\
  79     x264_mc_weight_w4##func##_neon,\
  80     x264_mc_weight_w4##func##_neon,\
  81     x264_mc_weight_w8##func##_neon,\
  82     x264_mc_weight_w16##func##_neon,\
  83     x264_mc_weight_w16##func##_neon,\
  84     x264_mc_weight_w20##func##_neon,\
  85 };
  86
  87 MC_WEIGHT()
  88 MC_WEIGHT(_nodenom)
  89 MC_WEIGHT(_offsetadd)
  90 MC_WEIGHT(_offsetsub)
  91 #endif
  92
  93 void x264_mc_copy_w4_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  94 void x264_mc_copy_w8_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  95 void x264_mc_copy_w16_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  96 void x264_mc_copy_w16_aligned_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, int );
  97
  98 void x264_mc_chroma_neon( uint8_t *, uint8_t *, intptr_t, uint8_t *, intptr_t, int, int, int, int );
  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_hpel_filter_v_neon( uint8_t *, uint8_t *, int16_t *, intptr_t, int );
 102 void x264_hpel_filter_c_neon( uint8_t *, int16_t *, int );
 103 void x264_hpel_filter_h_neon( uint8_t *, uint8_t *, int );
 104
 105 void integral_init4h_neon( uint16_t *, uint8_t *, intptr_t );
 106 void integral_init4v_neon( uint16_t *, uint16_t *, intptr_t );
 107 void integral_init8h_neon( uint16_t *, uint8_t *, intptr_t );
 108 void integral_init8v_neon( uint16_t *, intptr_t );
 109
 110 void x264_mbtree_propagate_cost_neon( int16_t *, uint16_t *, uint16_t *, uint16_t *, uint16_t *, float *, int );
 111
 112 #if !HIGH_BIT_DEPTH
 113 static void x264_weight_cache_neon( x264_t *h, x264_weight_t *w )
 114 {
 115     if( w->i_scale == 1<<w->i_denom )
 116     {
 117         if( w->i_offset < 0 )
 118         {
 119             w->weightfn = x264_mc_offsetsub_wtab_neon;
 120             w->cachea[0] = -w->i_offset;
 121         }
 122         else
 123         {
 124             w->weightfn = x264_mc_offsetadd_wtab_neon;
 125             w->cachea[0] = w->i_offset;
 126         }
 127     }
 128     else if( !w->i_denom )
 129         w->weightfn = x264_mc_nodenom_wtab_neon;
 130     else
 131         w->weightfn = x264_mc_wtab_neon;
 132 }
 133
 134 static void (* const x264_pixel_avg_wtab_neon[6])( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, int ) =
 135 {
 136     NULL,
 137     x264_pixel_avg2_w4_neon,
 138     x264_pixel_avg2_w8_neon,
 139     x264_pixel_avg2_w16_neon,   // no slower than w12, so no point in a separate function
 140     x264_pixel_avg2_w16_neon,
 141     x264_pixel_avg2_w20_neon,
 142 };
 143
 144 static void (* const x264_mc_copy_wtab_neon[5])( uint8_t *, intptr_t, uint8_t *, intptr_t, int ) =
 145 {
 146     NULL,
 147     x264_mc_copy_w4_neon,
 148     x264_mc_copy_w8_neon,
 149     NULL,
 150     x264_mc_copy_w16_neon,
 151 };
 152
 153 static void mc_luma_neon( uint8_t *dst,    intptr_t i_dst_stride,
 154                           uint8_t *src[4], intptr_t i_src_stride,
 155                           int mvx, int mvy,
 156                           int i_width, int i_height, const x264_weight_t *weight )
 157 {
 158     int qpel_idx = ((mvy&3)<<2) + (mvx&3);
 159     intptr_t offset = (mvy>>2)*i_src_stride + (mvx>>2);
 160     uint8_t *src1 = src[x264_hpel_ref0[qpel_idx]] + offset;
 161     if ( (mvy&3) == 3 )             // explict if() to force conditional add
 162         src1 += i_src_stride;
 163
 164     if( qpel_idx & 5 ) /* qpel interpolation needed */
 165     {
 166         uint8_t *src2 = src[x264_hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);
 167         x264_pixel_avg_wtab_neon[i_width>>2](
 168                 dst, i_dst_stride, src1, i_src_stride,
 169                 src2, i_height );
 170         if( weight->weightfn )
 171             weight->weightfn[i_width>>2]( dst, i_dst_stride, dst, i_dst_stride, weight, i_height );
 172     }
 173     else if( weight->weightfn )
 174         weight->weightfn[i_width>>2]( dst, i_dst_stride, src1, i_src_stride, weight, i_height );
 175     else
 176         x264_mc_copy_wtab_neon[i_width>>2]( dst, i_dst_stride, src1, i_src_stride, i_height );
 177 }
 178
 179 static uint8_t *get_ref_neon( uint8_t *dst,   intptr_t *i_dst_stride,
 180                               uint8_t *src[4], intptr_t i_src_stride,
 181                               int mvx, int mvy,
 182                               int i_width, int i_height, const x264_weight_t *weight )
 183 {
 184     int qpel_idx = ((mvy&3)<<2) + (mvx&3);
 185     intptr_t offset = (mvy>>2)*i_src_stride + (mvx>>2);
 186     uint8_t *src1 = src[x264_hpel_ref0[qpel_idx]] + offset;
 187     if ( (mvy&3) == 3 )             // explict if() to force conditional add
 188         src1 += i_src_stride;
 189
 190     if( qpel_idx & 5 ) /* qpel interpolation needed */
 191     {
 192         uint8_t *src2 = src[x264_hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);
 193         x264_pixel_avg_wtab_neon[i_width>>2](
 194                 dst, *i_dst_stride, src1, i_src_stride,
 195                 src2, i_height );
 196         if( weight->weightfn )
 197             weight->weightfn[i_width>>2]( dst, *i_dst_stride, dst, *i_dst_stride, weight, i_height );
 198         return dst;
 199     }
 200     else if( weight->weightfn )
 201     {
 202         weight->weightfn[i_width>>2]( dst, *i_dst_stride, src1, i_src_stride, weight, i_height );
 203         return dst;
 204     }
 205     else
 206     {
 207         *i_dst_stride = i_src_stride;
 208         return src1;
 209     }
 210 }
 211
 212 static void hpel_filter_neon( uint8_t *dsth, uint8_t *dstv, uint8_t *dstc, uint8_t *src,
 213                               intptr_t stride, int width, int height, int16_t *buf )
 214 {
 215     intptr_t realign = (intptr_t)src & 15;
 216     src -= realign;
 217     dstv -= realign;
 218     dstc -= realign;
 219     dsth -= realign;
 220     width += realign;
 221     while( height-- )
 222     {
 223         x264_hpel_filter_v_neon( dstv, src, buf+8, stride, width );
 224         x264_hpel_filter_c_neon( dstc, buf+8, width );
 225         x264_hpel_filter_h_neon( dsth, src, width );
 226         dsth += stride;
 227         dstv += stride;
 228         dstc += stride;
 229         src  += stride;
 230     }
 231 }
 232 #endif // !HIGH_BIT_DEPTH
 233
 234 PROPAGATE_LIST(neon)
 235
 236 void x264_mc_init_arm( int cpu, x264_mc_functions_t *pf )
 237 {
 238     if( !(cpu&X264_CPU_ARMV6) )
 239         return;
 240
 241 #if !HIGH_BIT_DEPTH
 242     pf->prefetch_fenc_420 = x264_prefetch_fenc_arm;
 243     pf->prefetch_fenc_422 = x264_prefetch_fenc_arm; /* FIXME */
 244     pf->prefetch_ref  = x264_prefetch_ref_arm;
 245 #endif // !HIGH_BIT_DEPTH
 246
 247     if( !(cpu&X264_CPU_NEON) )
 248         return;
 249
 250 #if !HIGH_BIT_DEPTH
 251     pf->copy_16x16_unaligned = x264_mc_copy_w16_neon;
 252     pf->copy[PIXEL_16x16] = x264_mc_copy_w16_aligned_neon;
 253     pf->copy[PIXEL_8x8]   = x264_mc_copy_w8_neon;
 254     pf->copy[PIXEL_4x4]   = x264_mc_copy_w4_neon;
 255
 256     pf->plane_copy              = x264_plane_copy_neon;
 257     pf->plane_copy_deinterleave = x264_plane_copy_deinterleave_neon;
 258     pf->plane_copy_deinterleave_rgb = x264_plane_copy_deinterleave_rgb_neon;
 259     pf->plane_copy_interleave = x264_plane_copy_interleave_neon;
 260     pf->plane_copy_swap = x264_plane_copy_swap_neon;
 261
 262     pf->store_interleave_chroma = x264_store_interleave_chroma_neon;
 263     pf->load_deinterleave_chroma_fdec = x264_load_deinterleave_chroma_fdec_neon;
 264     pf->load_deinterleave_chroma_fenc = x264_load_deinterleave_chroma_fenc_neon;
 265
 266     pf->avg[PIXEL_16x16] = x264_pixel_avg_16x16_neon;
 267     pf->avg[PIXEL_16x8]  = x264_pixel_avg_16x8_neon;
 268     pf->avg[PIXEL_8x16]  = x264_pixel_avg_8x16_neon;
 269     pf->avg[PIXEL_8x8]   = x264_pixel_avg_8x8_neon;
 270     pf->avg[PIXEL_8x4]   = x264_pixel_avg_8x4_neon;
 271     pf->avg[PIXEL_4x16]  = x264_pixel_avg_4x16_neon;
 272     pf->avg[PIXEL_4x8]   = x264_pixel_avg_4x8_neon;
 273     pf->avg[PIXEL_4x4]   = x264_pixel_avg_4x4_neon;
 274     pf->avg[PIXEL_4x2]   = x264_pixel_avg_4x2_neon;
 275
 276     pf->weight    = x264_mc_wtab_neon;
 277     pf->offsetadd = x264_mc_offsetadd_wtab_neon;
 278     pf->offsetsub = x264_mc_offsetsub_wtab_neon;
 279     pf->weight_cache = x264_weight_cache_neon;
 280
 281     pf->mc_chroma = x264_mc_chroma_neon;
 282     pf->mc_luma = mc_luma_neon;
 283     pf->get_ref = get_ref_neon;
 284     pf->hpel_filter = hpel_filter_neon;
 285     pf->frame_init_lowres_core = x264_frame_init_lowres_core_neon;
 286
 287     pf->integral_init4h = integral_init4h_neon;
 288     pf->integral_init8h = integral_init8h_neon;
 289     pf->integral_init4v = integral_init4v_neon;
 290     pf->integral_init8v = integral_init8v_neon;
 291
 292     pf->mbtree_propagate_cost = x264_mbtree_propagate_cost_neon;
 293     pf->mbtree_propagate_list = x264_mbtree_propagate_list_neon;
 294 #endif // !HIGH_BIT_DEPTH
 295
 296 // Apple's gcc stupidly cannot align stack variables, and ALIGNED_ARRAY can't work on structs
 297 #ifndef SYS_MACOSX
 298     pf->memcpy_aligned  = x264_memcpy_aligned_neon;
 299 #endif
 300     pf->memzero_aligned = x264_memzero_aligned_neon;
 301 }