• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2018, Alliance for Open Media. All rights reserved
3  *
4  * This source code is subject to the terms of the BSD 2 Clause License and
5  * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6  * was not distributed with this source code in the LICENSE file, you can
7  * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8  * Media Patent License 1.0 was not distributed with this source code in the
9  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10  */
11 
12 #include "warp_plane_neon.h"
13 
horizontal_filter_4x1_f4(const uint8x16_t in,int sx,int alpha)14 static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in,
15                                                            int sx, int alpha) {
16   const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1));
17 
18   // Loading the 8 filter taps
19   int16x8_t f[4];
20   load_filters_4(f, sx, alpha);
21 
22   int16x8_t in16_lo = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(in)));
23   int16x8_t in16_hi = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(in)));
24 
25   int16x8_t m0 = vmulq_s16(f[0], in16_lo);
26   int16x8_t m1 = vmulq_s16(f[1], vextq_s16(in16_lo, in16_hi, 1));
27   int16x8_t m2 = vmulq_s16(f[2], vextq_s16(in16_lo, in16_hi, 2));
28   int16x8_t m3 = vmulq_s16(f[3], vextq_s16(in16_lo, in16_hi, 3));
29 
30   int32x4_t m0123_pairs[] = { vpaddlq_s16(m0), vpaddlq_s16(m1), vpaddlq_s16(m2),
31                               vpaddlq_s16(m3) };
32 
33   int32x4_t tmp_res_low = horizontal_add_4d_s32x4(m0123_pairs);
34 
35   tmp_res_low = vaddq_s32(tmp_res_low, add_const);
36 
37   uint16x8_t res =
38       vcombine_u16(vqrshrun_n_s32(tmp_res_low, ROUND0_BITS), vdup_n_u16(0));
39   return vreinterpretq_s16_u16(res);
40 }
41 
horizontal_filter_8x1_f8(const uint8x16_t in,int sx,int alpha)42 static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f8(const uint8x16_t in,
43                                                            int sx, int alpha) {
44   const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1));
45 
46   // Loading the 8 filter taps
47   int16x8_t f[8];
48   load_filters_8(f, sx, alpha);
49 
50   int16x8_t in16_lo = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(in)));
51   int16x8_t in16_hi = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(in)));
52 
53   int16x8_t m0 = vmulq_s16(f[0], in16_lo);
54   int16x8_t m1 = vmulq_s16(f[1], vextq_s16(in16_lo, in16_hi, 1));
55   int16x8_t m2 = vmulq_s16(f[2], vextq_s16(in16_lo, in16_hi, 2));
56   int16x8_t m3 = vmulq_s16(f[3], vextq_s16(in16_lo, in16_hi, 3));
57   int16x8_t m4 = vmulq_s16(f[4], vextq_s16(in16_lo, in16_hi, 4));
58   int16x8_t m5 = vmulq_s16(f[5], vextq_s16(in16_lo, in16_hi, 5));
59   int16x8_t m6 = vmulq_s16(f[6], vextq_s16(in16_lo, in16_hi, 6));
60   int16x8_t m7 = vmulq_s16(f[7], vextq_s16(in16_lo, in16_hi, 7));
61 
62   int32x4_t m0123_pairs[] = { vpaddlq_s16(m0), vpaddlq_s16(m1), vpaddlq_s16(m2),
63                               vpaddlq_s16(m3) };
64   int32x4_t m4567_pairs[] = { vpaddlq_s16(m4), vpaddlq_s16(m5), vpaddlq_s16(m6),
65                               vpaddlq_s16(m7) };
66 
67   int32x4_t tmp_res_low = horizontal_add_4d_s32x4(m0123_pairs);
68   int32x4_t tmp_res_high = horizontal_add_4d_s32x4(m4567_pairs);
69 
70   tmp_res_low = vaddq_s32(tmp_res_low, add_const);
71   tmp_res_high = vaddq_s32(tmp_res_high, add_const);
72 
73   uint16x8_t res = vcombine_u16(vqrshrun_n_s32(tmp_res_low, ROUND0_BITS),
74                                 vqrshrun_n_s32(tmp_res_high, ROUND0_BITS));
75   return vreinterpretq_s16_u16(res);
76 }
77 
horizontal_filter_4x1_f1(const uint8x16_t in,int sx)78 static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f1(const uint8x16_t in,
79                                                            int sx) {
80   const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1));
81 
82   int16x8_t f_s16 =
83       vld1q_s16((int16_t *)(av1_warped_filter + (sx >> WARPEDDIFF_PREC_BITS)));
84 
85   int16x8_t in16_lo = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(in)));
86   int16x8_t in16_hi = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(in)));
87 
88   int16x8_t m0 = vmulq_s16(f_s16, in16_lo);
89   int16x8_t m1 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 1));
90   int16x8_t m2 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 2));
91   int16x8_t m3 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 3));
92 
93   int32x4_t m0123_pairs[] = { vpaddlq_s16(m0), vpaddlq_s16(m1), vpaddlq_s16(m2),
94                               vpaddlq_s16(m3) };
95 
96   int32x4_t tmp_res_low = horizontal_add_4d_s32x4(m0123_pairs);
97 
98   tmp_res_low = vaddq_s32(tmp_res_low, add_const);
99 
100   uint16x8_t res =
101       vcombine_u16(vqrshrun_n_s32(tmp_res_low, ROUND0_BITS), vdup_n_u16(0));
102   return vreinterpretq_s16_u16(res);
103 }
104 
horizontal_filter_8x1_f1(const uint8x16_t in,int sx)105 static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f1(const uint8x16_t in,
106                                                            int sx) {
107   const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1));
108 
109   int16x8_t f_s16 =
110       vld1q_s16((int16_t *)(av1_warped_filter + (sx >> WARPEDDIFF_PREC_BITS)));
111 
112   int16x8_t in16_lo = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(in)));
113   int16x8_t in16_hi = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(in)));
114 
115   int16x8_t m0 = vmulq_s16(f_s16, in16_lo);
116   int16x8_t m1 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 1));
117   int16x8_t m2 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 2));
118   int16x8_t m3 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 3));
119   int16x8_t m4 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 4));
120   int16x8_t m5 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 5));
121   int16x8_t m6 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 6));
122   int16x8_t m7 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 7));
123 
124   int32x4_t m0123_pairs[] = { vpaddlq_s16(m0), vpaddlq_s16(m1), vpaddlq_s16(m2),
125                               vpaddlq_s16(m3) };
126   int32x4_t m4567_pairs[] = { vpaddlq_s16(m4), vpaddlq_s16(m5), vpaddlq_s16(m6),
127                               vpaddlq_s16(m7) };
128 
129   int32x4_t tmp_res_low = horizontal_add_4d_s32x4(m0123_pairs);
130   int32x4_t tmp_res_high = horizontal_add_4d_s32x4(m4567_pairs);
131 
132   tmp_res_low = vaddq_s32(tmp_res_low, add_const);
133   tmp_res_high = vaddq_s32(tmp_res_high, add_const);
134 
135   uint16x8_t res = vcombine_u16(vqrshrun_n_s32(tmp_res_low, ROUND0_BITS),
136                                 vqrshrun_n_s32(tmp_res_high, ROUND0_BITS));
137   return vreinterpretq_s16_u16(res);
138 }
139 
vertical_filter_4x1_f1(const int16x8_t * src,int32x4_t * res,int sy)140 static AOM_FORCE_INLINE void vertical_filter_4x1_f1(const int16x8_t *src,
141                                                     int32x4_t *res, int sy) {
142   int16x4_t s0 = vget_low_s16(src[0]);
143   int16x4_t s1 = vget_low_s16(src[1]);
144   int16x4_t s2 = vget_low_s16(src[2]);
145   int16x4_t s3 = vget_low_s16(src[3]);
146   int16x4_t s4 = vget_low_s16(src[4]);
147   int16x4_t s5 = vget_low_s16(src[5]);
148   int16x4_t s6 = vget_low_s16(src[6]);
149   int16x4_t s7 = vget_low_s16(src[7]);
150 
151   int16x8_t f =
152       vld1q_s16((int16_t *)(av1_warped_filter + (sy >> WARPEDDIFF_PREC_BITS)));
153 
154   int32x4_t m0123 = vmull_lane_s16(s0, vget_low_s16(f), 0);
155   m0123 = vmlal_lane_s16(m0123, s1, vget_low_s16(f), 1);
156   m0123 = vmlal_lane_s16(m0123, s2, vget_low_s16(f), 2);
157   m0123 = vmlal_lane_s16(m0123, s3, vget_low_s16(f), 3);
158   m0123 = vmlal_lane_s16(m0123, s4, vget_high_s16(f), 0);
159   m0123 = vmlal_lane_s16(m0123, s5, vget_high_s16(f), 1);
160   m0123 = vmlal_lane_s16(m0123, s6, vget_high_s16(f), 2);
161   m0123 = vmlal_lane_s16(m0123, s7, vget_high_s16(f), 3);
162 
163   *res = m0123;
164 }
165 
vertical_filter_4x1_f4(const int16x8_t * src,int32x4_t * res,int sy,int gamma)166 static AOM_FORCE_INLINE void vertical_filter_4x1_f4(const int16x8_t *src,
167                                                     int32x4_t *res, int sy,
168                                                     int gamma) {
169   int16x8_t s0, s1, s2, s3;
170   transpose_elems_s16_4x8(
171       vget_low_s16(src[0]), vget_low_s16(src[1]), vget_low_s16(src[2]),
172       vget_low_s16(src[3]), vget_low_s16(src[4]), vget_low_s16(src[5]),
173       vget_low_s16(src[6]), vget_low_s16(src[7]), &s0, &s1, &s2, &s3);
174 
175   int16x8_t f[4];
176   load_filters_4(f, sy, gamma);
177 
178   int32x4_t m0 = vmull_s16(vget_low_s16(s0), vget_low_s16(f[0]));
179   m0 = vmlal_s16(m0, vget_high_s16(s0), vget_high_s16(f[0]));
180   int32x4_t m1 = vmull_s16(vget_low_s16(s1), vget_low_s16(f[1]));
181   m1 = vmlal_s16(m1, vget_high_s16(s1), vget_high_s16(f[1]));
182   int32x4_t m2 = vmull_s16(vget_low_s16(s2), vget_low_s16(f[2]));
183   m2 = vmlal_s16(m2, vget_high_s16(s2), vget_high_s16(f[2]));
184   int32x4_t m3 = vmull_s16(vget_low_s16(s3), vget_low_s16(f[3]));
185   m3 = vmlal_s16(m3, vget_high_s16(s3), vget_high_s16(f[3]));
186 
187   int32x4_t m0123_pairs[] = { m0, m1, m2, m3 };
188 
189   *res = horizontal_add_4d_s32x4(m0123_pairs);
190 }
191 
vertical_filter_8x1_f1(const int16x8_t * src,int32x4_t * res_low,int32x4_t * res_high,int sy)192 static AOM_FORCE_INLINE void vertical_filter_8x1_f1(const int16x8_t *src,
193                                                     int32x4_t *res_low,
194                                                     int32x4_t *res_high,
195                                                     int sy) {
196   int16x8_t s0 = src[0];
197   int16x8_t s1 = src[1];
198   int16x8_t s2 = src[2];
199   int16x8_t s3 = src[3];
200   int16x8_t s4 = src[4];
201   int16x8_t s5 = src[5];
202   int16x8_t s6 = src[6];
203   int16x8_t s7 = src[7];
204 
205   int16x8_t f =
206       vld1q_s16((int16_t *)(av1_warped_filter + (sy >> WARPEDDIFF_PREC_BITS)));
207 
208   int32x4_t m0123 = vmull_lane_s16(vget_low_s16(s0), vget_low_s16(f), 0);
209   m0123 = vmlal_lane_s16(m0123, vget_low_s16(s1), vget_low_s16(f), 1);
210   m0123 = vmlal_lane_s16(m0123, vget_low_s16(s2), vget_low_s16(f), 2);
211   m0123 = vmlal_lane_s16(m0123, vget_low_s16(s3), vget_low_s16(f), 3);
212   m0123 = vmlal_lane_s16(m0123, vget_low_s16(s4), vget_high_s16(f), 0);
213   m0123 = vmlal_lane_s16(m0123, vget_low_s16(s5), vget_high_s16(f), 1);
214   m0123 = vmlal_lane_s16(m0123, vget_low_s16(s6), vget_high_s16(f), 2);
215   m0123 = vmlal_lane_s16(m0123, vget_low_s16(s7), vget_high_s16(f), 3);
216 
217   int32x4_t m4567 = vmull_lane_s16(vget_high_s16(s0), vget_low_s16(f), 0);
218   m4567 = vmlal_lane_s16(m4567, vget_high_s16(s1), vget_low_s16(f), 1);
219   m4567 = vmlal_lane_s16(m4567, vget_high_s16(s2), vget_low_s16(f), 2);
220   m4567 = vmlal_lane_s16(m4567, vget_high_s16(s3), vget_low_s16(f), 3);
221   m4567 = vmlal_lane_s16(m4567, vget_high_s16(s4), vget_high_s16(f), 0);
222   m4567 = vmlal_lane_s16(m4567, vget_high_s16(s5), vget_high_s16(f), 1);
223   m4567 = vmlal_lane_s16(m4567, vget_high_s16(s6), vget_high_s16(f), 2);
224   m4567 = vmlal_lane_s16(m4567, vget_high_s16(s7), vget_high_s16(f), 3);
225 
226   *res_low = m0123;
227   *res_high = m4567;
228 }
229 
vertical_filter_8x1_f8(const int16x8_t * src,int32x4_t * res_low,int32x4_t * res_high,int sy,int gamma)230 static AOM_FORCE_INLINE void vertical_filter_8x1_f8(const int16x8_t *src,
231                                                     int32x4_t *res_low,
232                                                     int32x4_t *res_high, int sy,
233                                                     int gamma) {
234   int16x8_t s0 = src[0];
235   int16x8_t s1 = src[1];
236   int16x8_t s2 = src[2];
237   int16x8_t s3 = src[3];
238   int16x8_t s4 = src[4];
239   int16x8_t s5 = src[5];
240   int16x8_t s6 = src[6];
241   int16x8_t s7 = src[7];
242   transpose_elems_inplace_s16_8x8(&s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
243 
244   int16x8_t f[8];
245   load_filters_8(f, sy, gamma);
246 
247   int32x4_t m0 = vmull_s16(vget_low_s16(s0), vget_low_s16(f[0]));
248   m0 = vmlal_s16(m0, vget_high_s16(s0), vget_high_s16(f[0]));
249   int32x4_t m1 = vmull_s16(vget_low_s16(s1), vget_low_s16(f[1]));
250   m1 = vmlal_s16(m1, vget_high_s16(s1), vget_high_s16(f[1]));
251   int32x4_t m2 = vmull_s16(vget_low_s16(s2), vget_low_s16(f[2]));
252   m2 = vmlal_s16(m2, vget_high_s16(s2), vget_high_s16(f[2]));
253   int32x4_t m3 = vmull_s16(vget_low_s16(s3), vget_low_s16(f[3]));
254   m3 = vmlal_s16(m3, vget_high_s16(s3), vget_high_s16(f[3]));
255   int32x4_t m4 = vmull_s16(vget_low_s16(s4), vget_low_s16(f[4]));
256   m4 = vmlal_s16(m4, vget_high_s16(s4), vget_high_s16(f[4]));
257   int32x4_t m5 = vmull_s16(vget_low_s16(s5), vget_low_s16(f[5]));
258   m5 = vmlal_s16(m5, vget_high_s16(s5), vget_high_s16(f[5]));
259   int32x4_t m6 = vmull_s16(vget_low_s16(s6), vget_low_s16(f[6]));
260   m6 = vmlal_s16(m6, vget_high_s16(s6), vget_high_s16(f[6]));
261   int32x4_t m7 = vmull_s16(vget_low_s16(s7), vget_low_s16(f[7]));
262   m7 = vmlal_s16(m7, vget_high_s16(s7), vget_high_s16(f[7]));
263 
264   int32x4_t m0123_pairs[] = { m0, m1, m2, m3 };
265   int32x4_t m4567_pairs[] = { m4, m5, m6, m7 };
266 
267   *res_low = horizontal_add_4d_s32x4(m0123_pairs);
268   *res_high = horizontal_add_4d_s32x4(m4567_pairs);
269 }
270 
av1_warp_affine_neon(const int32_t * mat,const uint8_t * ref,int width,int height,int stride,uint8_t * pred,int p_col,int p_row,int p_width,int p_height,int p_stride,int subsampling_x,int subsampling_y,ConvolveParams * conv_params,int16_t alpha,int16_t beta,int16_t gamma,int16_t delta)271 void av1_warp_affine_neon(const int32_t *mat, const uint8_t *ref, int width,
272                           int height, int stride, uint8_t *pred, int p_col,
273                           int p_row, int p_width, int p_height, int p_stride,
274                           int subsampling_x, int subsampling_y,
275                           ConvolveParams *conv_params, int16_t alpha,
276                           int16_t beta, int16_t gamma, int16_t delta) {
277   av1_warp_affine_common(mat, ref, width, height, stride, pred, p_col, p_row,
278                          p_width, p_height, p_stride, subsampling_x,
279                          subsampling_y, conv_params, alpha, beta, gamma, delta);
280 }
281