arm: Implement integral_init4/8h/v_neon
[x264/x264-sandbox.git] / common / arm / mc-c.c
1 /*****************************************************************************
2  * mc-c.c: arm motion compensation
3  *****************************************************************************
4  * Copyright (C) 2009-2015 x264 project
5  *
6  * Authors: David Conrad <lessen42@gmail.com>
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License as published by
10  * the Free Software Foundation; either version 2 of the License, or
11  * (at your option) any later version.
12  *
13  * This program is distributed in the hope that it will be useful,
14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
16  * GNU General Public License for more details.
17  *
18  * You should have received a copy of the GNU General Public License
19  * along with this program; if not, write to the Free Software
20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02111, USA.
21  *
22  * This program is also available under a commercial proprietary license.
23  * For more information, contact us at licensing@x264.com.
24  *****************************************************************************/
25
26 #include "common/common.h"
27 #include "mc.h"
28
29 void x264_prefetch_ref_arm( uint8_t *, intptr_t, int );
30 void x264_prefetch_fenc_arm( uint8_t *, intptr_t, uint8_t *, intptr_t, int );
31
32 void *x264_memcpy_aligned_neon( void *dst, const void *src, size_t n );
33 void x264_memzero_aligned_neon( void *dst, size_t n );
34
35 void x264_pixel_avg_16x16_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
36 void x264_pixel_avg_16x8_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
37 void x264_pixel_avg_8x16_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
38 void x264_pixel_avg_8x8_neon  ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
39 void x264_pixel_avg_8x4_neon  ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
40 void x264_pixel_avg_4x16_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
41 void x264_pixel_avg_4x8_neon  ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
42 void x264_pixel_avg_4x4_neon  ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
43 void x264_pixel_avg_4x2_neon  ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, intptr_t, int );
44
45 void x264_pixel_avg2_w4_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, int );
46 void x264_pixel_avg2_w8_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, int );
47 void x264_pixel_avg2_w16_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, int );
48 void x264_pixel_avg2_w20_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, int );
49
50 void x264_plane_copy_deinterleave_neon(  pixel *dstu, intptr_t i_dstu,
51                                          pixel *dstv, intptr_t i_dstv,
52                                          pixel *src,  intptr_t i_src, int w, int h );
53 void x264_plane_copy_deinterleave_rgb_neon( pixel *dsta, intptr_t i_dsta,
54                                             pixel *dstb, intptr_t i_dstb,
55                                             pixel *dstc, intptr_t i_dstc,
56                                             pixel *src,  intptr_t i_src, int pw, int w, int h );
57 void x264_plane_copy_interleave_neon( pixel *dst,  intptr_t i_dst,
58                                       pixel *srcu, intptr_t i_srcu,
59                                       pixel *srcv, intptr_t i_srcv, int w, int h );
60 void x264_plane_copy_swap_neon( pixel *dst, intptr_t i_dst,
61                                 pixel *src, intptr_t i_src, int w, int h );
62
63 void x264_store_interleave_chroma_neon( pixel *dst, intptr_t i_dst, pixel *srcu, pixel *srcv, int height );
64 void x264_load_deinterleave_chroma_fdec_neon( pixel *dst, pixel *src, intptr_t i_src, int height );
65 void x264_load_deinterleave_chroma_fenc_neon( pixel *dst, pixel *src, intptr_t i_src, int height );
66
67 #define MC_WEIGHT(func)\
68 void x264_mc_weight_w20##func##_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, const x264_weight_t *, int );\
69 void x264_mc_weight_w16##func##_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, const x264_weight_t *, int );\
70 void x264_mc_weight_w8##func##_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, const x264_weight_t *, int );\
71 void x264_mc_weight_w4##func##_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, const x264_weight_t *, int );\
72 \
73 static weight_fn_t x264_mc##func##_wtab_neon[6] =\
74 {\
75     x264_mc_weight_w4##func##_neon,\
76     x264_mc_weight_w4##func##_neon,\
77     x264_mc_weight_w8##func##_neon,\
78     x264_mc_weight_w16##func##_neon,\
79     x264_mc_weight_w16##func##_neon,\
80     x264_mc_weight_w20##func##_neon,\
81 };
82
83 MC_WEIGHT()
84 MC_WEIGHT(_nodenom)
85 MC_WEIGHT(_offsetadd)
86 MC_WEIGHT(_offsetsub)
87
88 void x264_mc_copy_w4_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, int );
89 void x264_mc_copy_w8_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t, int );
90 void x264_mc_copy_w16_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, int );
91 void x264_mc_copy_w16_aligned_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, int );
92
93 void x264_mc_chroma_neon( uint8_t *, uint8_t *, intptr_t, uint8_t *, intptr_t, int, int, int, int );
94 void x264_frame_init_lowres_core_neon( uint8_t *, uint8_t *, uint8_t *, uint8_t *, uint8_t *, intptr_t, intptr_t, int, int );
95
96 void x264_hpel_filter_v_neon( uint8_t *, uint8_t *, int16_t *, intptr_t, int );
97 void x264_hpel_filter_c_neon( uint8_t *, int16_t *, int );
98 void x264_hpel_filter_h_neon( uint8_t *, uint8_t *, int );
99
100 void integral_init4h_neon( uint16_t *, uint8_t *, intptr_t );
101 void integral_init4v_neon( uint16_t *, uint16_t *, intptr_t );
102 void integral_init8h_neon( uint16_t *, uint8_t *, intptr_t );
103 void integral_init8v_neon( uint16_t *, intptr_t );
104
105 #if !HIGH_BIT_DEPTH
106 static void x264_weight_cache_neon( x264_t *h, x264_weight_t *w )
107 {
108     if( w->i_scale == 1<<w->i_denom )
109     {
110         if( w->i_offset < 0 )
111         {
112             w->weightfn = x264_mc_offsetsub_wtab_neon;
113             w->cachea[0] = -w->i_offset;
114         }
115         else
116         {
117             w->weightfn = x264_mc_offsetadd_wtab_neon;
118             w->cachea[0] = w->i_offset;
119         }
120     }
121     else if( !w->i_denom )
122         w->weightfn = x264_mc_nodenom_wtab_neon;
123     else
124         w->weightfn = x264_mc_wtab_neon;
125 }
126
127 static void (* const x264_pixel_avg_wtab_neon[6])( uint8_t *, intptr_t, uint8_t *, intptr_t, uint8_t *, int ) =
128 {
129     NULL,
130     x264_pixel_avg2_w4_neon,
131     x264_pixel_avg2_w8_neon,
132     x264_pixel_avg2_w16_neon,   // no slower than w12, so no point in a separate function
133     x264_pixel_avg2_w16_neon,
134     x264_pixel_avg2_w20_neon,
135 };
136
137 static void (* const x264_mc_copy_wtab_neon[5])( uint8_t *, intptr_t, uint8_t *, intptr_t, int ) =
138 {
139     NULL,
140     x264_mc_copy_w4_neon,
141     x264_mc_copy_w8_neon,
142     NULL,
143     x264_mc_copy_w16_neon,
144 };
145
146 static void mc_luma_neon( uint8_t *dst,    intptr_t i_dst_stride,
147                           uint8_t *src[4], intptr_t i_src_stride,
148                           int mvx, int mvy,
149                           int i_width, int i_height, const x264_weight_t *weight )
150 {
151     int qpel_idx = ((mvy&3)<<2) + (mvx&3);
152     intptr_t offset = (mvy>>2)*i_src_stride + (mvx>>2);
153     uint8_t *src1 = src[x264_hpel_ref0[qpel_idx]] + offset;
154     if ( (mvy&3) == 3 )             // explict if() to force conditional add
155         src1 += i_src_stride;
156
157     if( qpel_idx & 5 ) /* qpel interpolation needed */
158     {
159         uint8_t *src2 = src[x264_hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);
160         x264_pixel_avg_wtab_neon[i_width>>2](
161                 dst, i_dst_stride, src1, i_src_stride,
162                 src2, i_height );
163         if( weight->weightfn )
164             weight->weightfn[i_width>>2]( dst, i_dst_stride, dst, i_dst_stride, weight, i_height );
165     }
166     else if( weight->weightfn )
167         weight->weightfn[i_width>>2]( dst, i_dst_stride, src1, i_src_stride, weight, i_height );
168     else
169         x264_mc_copy_wtab_neon[i_width>>2]( dst, i_dst_stride, src1, i_src_stride, i_height );
170 }
171
172 static uint8_t *get_ref_neon( uint8_t *dst,   intptr_t *i_dst_stride,
173                               uint8_t *src[4], intptr_t i_src_stride,
174                               int mvx, int mvy,
175                               int i_width, int i_height, const x264_weight_t *weight )
176 {
177     int qpel_idx = ((mvy&3)<<2) + (mvx&3);
178     intptr_t offset = (mvy>>2)*i_src_stride + (mvx>>2);
179     uint8_t *src1 = src[x264_hpel_ref0[qpel_idx]] + offset;
180     if ( (mvy&3) == 3 )             // explict if() to force conditional add
181         src1 += i_src_stride;
182
183     if( qpel_idx & 5 ) /* qpel interpolation needed */
184     {
185         uint8_t *src2 = src[x264_hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);
186         x264_pixel_avg_wtab_neon[i_width>>2](
187                 dst, *i_dst_stride, src1, i_src_stride,
188                 src2, i_height );
189         if( weight->weightfn )
190             weight->weightfn[i_width>>2]( dst, *i_dst_stride, dst, *i_dst_stride, weight, i_height );
191         return dst;
192     }
193     else if( weight->weightfn )
194     {
195         weight->weightfn[i_width>>2]( dst, *i_dst_stride, src1, i_src_stride, weight, i_height );
196         return dst;
197     }
198     else
199     {
200         *i_dst_stride = i_src_stride;
201         return src1;
202     }
203 }
204
205 static void hpel_filter_neon( uint8_t *dsth, uint8_t *dstv, uint8_t *dstc, uint8_t *src,
206                               intptr_t stride, int width, int height, int16_t *buf )
207 {
208     intptr_t realign = (intptr_t)src & 15;
209     src -= realign;
210     dstv -= realign;
211     dstc -= realign;
212     dsth -= realign;
213     width += realign;
214     while( height-- )
215     {
216         x264_hpel_filter_v_neon( dstv, src, buf+8, stride, width );
217         x264_hpel_filter_c_neon( dstc, buf+8, width );
218         x264_hpel_filter_h_neon( dsth, src, width );
219         dsth += stride;
220         dstv += stride;
221         dstc += stride;
222         src  += stride;
223     }
224 }
225 #endif // !HIGH_BIT_DEPTH
226
227 void x264_mc_init_arm( int cpu, x264_mc_functions_t *pf )
228 {
229     if( !(cpu&X264_CPU_ARMV6) )
230         return;
231
232 #if !HIGH_BIT_DEPTH
233     pf->prefetch_fenc_420 = x264_prefetch_fenc_arm;
234     pf->prefetch_fenc_422 = x264_prefetch_fenc_arm; /* FIXME */
235     pf->prefetch_ref  = x264_prefetch_ref_arm;
236 #endif // !HIGH_BIT_DEPTH
237
238     if( !(cpu&X264_CPU_NEON) )
239         return;
240
241 #if !HIGH_BIT_DEPTH
242     pf->copy_16x16_unaligned = x264_mc_copy_w16_neon;
243     pf->copy[PIXEL_16x16] = x264_mc_copy_w16_aligned_neon;
244     pf->copy[PIXEL_8x8]   = x264_mc_copy_w8_neon;
245     pf->copy[PIXEL_4x4]   = x264_mc_copy_w4_neon;
246
247     pf->plane_copy_deinterleave = x264_plane_copy_deinterleave_neon;
248     pf->plane_copy_deinterleave_rgb = x264_plane_copy_deinterleave_rgb_neon;
249     pf->plane_copy_interleave = x264_plane_copy_interleave_neon;
250     pf->plane_copy_swap = x264_plane_copy_swap_neon;
251
252     pf->store_interleave_chroma = x264_store_interleave_chroma_neon;
253     pf->load_deinterleave_chroma_fdec = x264_load_deinterleave_chroma_fdec_neon;
254     pf->load_deinterleave_chroma_fenc = x264_load_deinterleave_chroma_fenc_neon;
255
256     pf->avg[PIXEL_16x16] = x264_pixel_avg_16x16_neon;
257     pf->avg[PIXEL_16x8]  = x264_pixel_avg_16x8_neon;
258     pf->avg[PIXEL_8x16]  = x264_pixel_avg_8x16_neon;
259     pf->avg[PIXEL_8x8]   = x264_pixel_avg_8x8_neon;
260     pf->avg[PIXEL_8x4]   = x264_pixel_avg_8x4_neon;
261     pf->avg[PIXEL_4x16]  = x264_pixel_avg_4x16_neon;
262     pf->avg[PIXEL_4x8]   = x264_pixel_avg_4x8_neon;
263     pf->avg[PIXEL_4x4]   = x264_pixel_avg_4x4_neon;
264     pf->avg[PIXEL_4x2]   = x264_pixel_avg_4x2_neon;
265
266     pf->weight    = x264_mc_wtab_neon;
267     pf->offsetadd = x264_mc_offsetadd_wtab_neon;
268     pf->offsetsub = x264_mc_offsetsub_wtab_neon;
269     pf->weight_cache = x264_weight_cache_neon;
270
271     pf->mc_chroma = x264_mc_chroma_neon;
272     pf->mc_luma = mc_luma_neon;
273     pf->get_ref = get_ref_neon;
274     pf->hpel_filter = hpel_filter_neon;
275     pf->frame_init_lowres_core = x264_frame_init_lowres_core_neon;
276
277     pf->integral_init4h = integral_init4h_neon;
278     pf->integral_init8h = integral_init8h_neon;
279     pf->integral_init4v = integral_init4v_neon;
280     pf->integral_init8v = integral_init8v_neon;
281 #endif // !HIGH_BIT_DEPTH
282
283 // Apple's gcc stupidly cannot align stack variables, and ALIGNED_ARRAY can't work on structs
284 #ifndef SYS_MACOSX
285     pf->memcpy_aligned  = x264_memcpy_aligned_neon;
286 #endif
287     pf->memzero_aligned = x264_memzero_aligned_neon;
288 }