1 /*
2 * Copyright (c) 2023, 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 <arm_neon.h>
13
14 #include "aom_dsp/arm/aom_neon_sve_bridge.h"
15 #include "warp_plane_neon.h"
16
17 DECLARE_ALIGNED(16, static const uint8_t, usdot_permute_idx[48]) = {
18 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6,
19 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10,
20 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14
21 };
22
horizontal_filter_4x1_f4(const uint8x16_t in,int sx,int alpha)23 static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in,
24 int sx, int alpha) {
25 const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1));
26
27 // Loading the 8 filter taps
28 int16x8_t f[4];
29 load_filters_4(f, sx, alpha);
30
31 int8x16_t f01_u8 = vcombine_s8(vmovn_s16(f[0]), vmovn_s16(f[1]));
32 int8x16_t f23_u8 = vcombine_s8(vmovn_s16(f[2]), vmovn_s16(f[3]));
33
34 uint8x8_t in0 = vget_low_u8(in);
35 uint8x8_t in1 = vget_low_u8(vextq_u8(in, in, 1));
36 uint8x8_t in2 = vget_low_u8(vextq_u8(in, in, 2));
37 uint8x8_t in3 = vget_low_u8(vextq_u8(in, in, 3));
38
39 int32x4_t m01 = vusdotq_s32(vdupq_n_s32(0), vcombine_u8(in0, in1), f01_u8);
40 int32x4_t m23 = vusdotq_s32(vdupq_n_s32(0), vcombine_u8(in2, in3), f23_u8);
41
42 int32x4_t tmp_res_low = vpaddq_s32(m01, m23);
43
44 tmp_res_low = vaddq_s32(tmp_res_low, add_const);
45
46 uint16x8_t res =
47 vcombine_u16(vqrshrun_n_s32(tmp_res_low, ROUND0_BITS), vdup_n_u16(0));
48 return vreinterpretq_s16_u16(res);
49 }
50
horizontal_filter_8x1_f8(const uint8x16_t in,int sx,int alpha)51 static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f8(const uint8x16_t in,
52 int sx, int alpha) {
53 const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1));
54
55 // Loading the 8 filter taps
56 int16x8_t f[8];
57 load_filters_8(f, sx, alpha);
58
59 int8x16_t f01_u8 = vcombine_s8(vmovn_s16(f[0]), vmovn_s16(f[1]));
60 int8x16_t f23_u8 = vcombine_s8(vmovn_s16(f[2]), vmovn_s16(f[3]));
61 int8x16_t f45_u8 = vcombine_s8(vmovn_s16(f[4]), vmovn_s16(f[5]));
62 int8x16_t f67_u8 = vcombine_s8(vmovn_s16(f[6]), vmovn_s16(f[7]));
63
64 uint8x8_t in0 = vget_low_u8(in);
65 uint8x8_t in1 = vget_low_u8(vextq_u8(in, in, 1));
66 uint8x8_t in2 = vget_low_u8(vextq_u8(in, in, 2));
67 uint8x8_t in3 = vget_low_u8(vextq_u8(in, in, 3));
68 uint8x8_t in4 = vget_low_u8(vextq_u8(in, in, 4));
69 uint8x8_t in5 = vget_low_u8(vextq_u8(in, in, 5));
70 uint8x8_t in6 = vget_low_u8(vextq_u8(in, in, 6));
71 uint8x8_t in7 = vget_low_u8(vextq_u8(in, in, 7));
72
73 int32x4_t m01 = vusdotq_s32(vdupq_n_s32(0), vcombine_u8(in0, in1), f01_u8);
74 int32x4_t m23 = vusdotq_s32(vdupq_n_s32(0), vcombine_u8(in2, in3), f23_u8);
75 int32x4_t m45 = vusdotq_s32(vdupq_n_s32(0), vcombine_u8(in4, in5), f45_u8);
76 int32x4_t m67 = vusdotq_s32(vdupq_n_s32(0), vcombine_u8(in6, in7), f67_u8);
77
78 int32x4_t tmp_res_low = vpaddq_s32(m01, m23);
79 int32x4_t tmp_res_high = vpaddq_s32(m45, m67);
80
81 tmp_res_low = vaddq_s32(tmp_res_low, add_const);
82 tmp_res_high = vaddq_s32(tmp_res_high, add_const);
83
84 uint16x8_t res = vcombine_u16(vqrshrun_n_s32(tmp_res_low, ROUND0_BITS),
85 vqrshrun_n_s32(tmp_res_high, ROUND0_BITS));
86 return vreinterpretq_s16_u16(res);
87 }
88
horizontal_filter_4x1_f1(const uint8x16_t in,int sx)89 static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f1(const uint8x16_t in,
90 int sx) {
91 const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1));
92
93 int16x8_t f_s16 =
94 vld1q_s16((int16_t *)(av1_warped_filter + (sx >> WARPEDDIFF_PREC_BITS)));
95
96 int8x16_t f_s8 = vcombine_s8(vmovn_s16(f_s16), vmovn_s16(f_s16));
97
98 uint8x16_t perm0 = vld1q_u8(&usdot_permute_idx[0]);
99 uint8x16_t perm1 = vld1q_u8(&usdot_permute_idx[16]);
100
101 // Permute samples ready for dot product.
102 // { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 }
103 // { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 }
104 uint8x16_t in_0123 = vqtbl1q_u8(in, perm0);
105 uint8x16_t in_4567 = vqtbl1q_u8(in, perm1);
106
107 int32x4_t m0123 = vusdotq_laneq_s32(vdupq_n_s32(0), in_0123, f_s8, 0);
108 m0123 = vusdotq_laneq_s32(m0123, in_4567, f_s8, 1);
109
110 int32x4_t tmp_res_low = m0123;
111
112 tmp_res_low = vaddq_s32(tmp_res_low, add_const);
113
114 uint16x8_t res =
115 vcombine_u16(vqrshrun_n_s32(tmp_res_low, ROUND0_BITS), vdup_n_u16(0));
116 return vreinterpretq_s16_u16(res);
117 }
118
horizontal_filter_8x1_f1(const uint8x16_t in,int sx)119 static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f1(const uint8x16_t in,
120 int sx) {
121 const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1));
122
123 int16x8_t f_s16 =
124 vld1q_s16((int16_t *)(av1_warped_filter + (sx >> WARPEDDIFF_PREC_BITS)));
125
126 int8x16_t f_s8 = vcombine_s8(vmovn_s16(f_s16), vmovn_s16(f_s16));
127
128 uint8x16_t perm0 = vld1q_u8(&usdot_permute_idx[0]);
129 uint8x16_t perm1 = vld1q_u8(&usdot_permute_idx[16]);
130 uint8x16_t perm2 = vld1q_u8(&usdot_permute_idx[32]);
131
132 // Permute samples ready for dot product.
133 // { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 }
134 // { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 }
135 // { 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 }
136 uint8x16_t in_0123 = vqtbl1q_u8(in, perm0);
137 uint8x16_t in_4567 = vqtbl1q_u8(in, perm1);
138 uint8x16_t in_89ab = vqtbl1q_u8(in, perm2);
139
140 int32x4_t m0123 = vusdotq_laneq_s32(vdupq_n_s32(0), in_0123, f_s8, 0);
141 m0123 = vusdotq_laneq_s32(m0123, in_4567, f_s8, 1);
142
143 int32x4_t m4567 = vusdotq_laneq_s32(vdupq_n_s32(0), in_4567, f_s8, 0);
144 m4567 = vusdotq_laneq_s32(m4567, in_89ab, f_s8, 1);
145
146 int32x4_t tmp_res_low = m0123;
147 int32x4_t tmp_res_high = m4567;
148
149 tmp_res_low = vaddq_s32(tmp_res_low, add_const);
150 tmp_res_high = vaddq_s32(tmp_res_high, add_const);
151
152 uint16x8_t res = vcombine_u16(vqrshrun_n_s32(tmp_res_low, ROUND0_BITS),
153 vqrshrun_n_s32(tmp_res_high, ROUND0_BITS));
154 return vreinterpretq_s16_u16(res);
155 }
156
vertical_filter_4x1_f1(const int16x8_t * src,int32x4_t * res,int sy)157 static AOM_FORCE_INLINE void vertical_filter_4x1_f1(const int16x8_t *src,
158 int32x4_t *res, int sy) {
159 int16x4_t s0 = vget_low_s16(src[0]);
160 int16x4_t s1 = vget_low_s16(src[1]);
161 int16x4_t s2 = vget_low_s16(src[2]);
162 int16x4_t s3 = vget_low_s16(src[3]);
163 int16x4_t s4 = vget_low_s16(src[4]);
164 int16x4_t s5 = vget_low_s16(src[5]);
165 int16x4_t s6 = vget_low_s16(src[6]);
166 int16x4_t s7 = vget_low_s16(src[7]);
167
168 int16x8_t f =
169 vld1q_s16((int16_t *)(av1_warped_filter + (sy >> WARPEDDIFF_PREC_BITS)));
170
171 int32x4_t m0123 = vmull_lane_s16(s0, vget_low_s16(f), 0);
172 m0123 = vmlal_lane_s16(m0123, s1, vget_low_s16(f), 1);
173 m0123 = vmlal_lane_s16(m0123, s2, vget_low_s16(f), 2);
174 m0123 = vmlal_lane_s16(m0123, s3, vget_low_s16(f), 3);
175 m0123 = vmlal_lane_s16(m0123, s4, vget_high_s16(f), 0);
176 m0123 = vmlal_lane_s16(m0123, s5, vget_high_s16(f), 1);
177 m0123 = vmlal_lane_s16(m0123, s6, vget_high_s16(f), 2);
178 m0123 = vmlal_lane_s16(m0123, s7, vget_high_s16(f), 3);
179
180 *res = m0123;
181 }
182
vertical_filter_4x1_f4(const int16x8_t * src,int32x4_t * res,int sy,int gamma)183 static AOM_FORCE_INLINE void vertical_filter_4x1_f4(const int16x8_t *src,
184 int32x4_t *res, int sy,
185 int gamma) {
186 int16x8_t s0, s1, s2, s3;
187 transpose_elems_s16_4x8(
188 vget_low_s16(src[0]), vget_low_s16(src[1]), vget_low_s16(src[2]),
189 vget_low_s16(src[3]), vget_low_s16(src[4]), vget_low_s16(src[5]),
190 vget_low_s16(src[6]), vget_low_s16(src[7]), &s0, &s1, &s2, &s3);
191
192 int16x8_t f[4];
193 load_filters_4(f, sy, gamma);
194
195 int64x2_t m0 = aom_sdotq_s16(vdupq_n_s64(0), s0, f[0]);
196 int64x2_t m1 = aom_sdotq_s16(vdupq_n_s64(0), s1, f[1]);
197 int64x2_t m2 = aom_sdotq_s16(vdupq_n_s64(0), s2, f[2]);
198 int64x2_t m3 = aom_sdotq_s16(vdupq_n_s64(0), s3, f[3]);
199
200 int64x2_t m01 = vpaddq_s64(m0, m1);
201 int64x2_t m23 = vpaddq_s64(m2, m3);
202
203 *res = vcombine_s32(vmovn_s64(m01), vmovn_s64(m23));
204 }
205
vertical_filter_8x1_f1(const int16x8_t * src,int32x4_t * res_low,int32x4_t * res_high,int sy)206 static AOM_FORCE_INLINE void vertical_filter_8x1_f1(const int16x8_t *src,
207 int32x4_t *res_low,
208 int32x4_t *res_high,
209 int sy) {
210 int16x8_t s0 = src[0];
211 int16x8_t s1 = src[1];
212 int16x8_t s2 = src[2];
213 int16x8_t s3 = src[3];
214 int16x8_t s4 = src[4];
215 int16x8_t s5 = src[5];
216 int16x8_t s6 = src[6];
217 int16x8_t s7 = src[7];
218
219 int16x8_t f =
220 vld1q_s16((int16_t *)(av1_warped_filter + (sy >> WARPEDDIFF_PREC_BITS)));
221
222 int32x4_t m0123 = vmull_lane_s16(vget_low_s16(s0), vget_low_s16(f), 0);
223 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s1), vget_low_s16(f), 1);
224 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s2), vget_low_s16(f), 2);
225 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s3), vget_low_s16(f), 3);
226 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s4), vget_high_s16(f), 0);
227 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s5), vget_high_s16(f), 1);
228 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s6), vget_high_s16(f), 2);
229 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s7), vget_high_s16(f), 3);
230
231 int32x4_t m4567 = vmull_lane_s16(vget_high_s16(s0), vget_low_s16(f), 0);
232 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s1), vget_low_s16(f), 1);
233 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s2), vget_low_s16(f), 2);
234 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s3), vget_low_s16(f), 3);
235 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s4), vget_high_s16(f), 0);
236 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s5), vget_high_s16(f), 1);
237 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s6), vget_high_s16(f), 2);
238 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s7), vget_high_s16(f), 3);
239
240 *res_low = m0123;
241 *res_high = m4567;
242 }
243
vertical_filter_8x1_f8(const int16x8_t * src,int32x4_t * res_low,int32x4_t * res_high,int sy,int gamma)244 static AOM_FORCE_INLINE void vertical_filter_8x1_f8(const int16x8_t *src,
245 int32x4_t *res_low,
246 int32x4_t *res_high, int sy,
247 int gamma) {
248 int16x8_t s0 = src[0];
249 int16x8_t s1 = src[1];
250 int16x8_t s2 = src[2];
251 int16x8_t s3 = src[3];
252 int16x8_t s4 = src[4];
253 int16x8_t s5 = src[5];
254 int16x8_t s6 = src[6];
255 int16x8_t s7 = src[7];
256 transpose_elems_inplace_s16_8x8(&s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
257
258 int16x8_t f[8];
259 load_filters_8(f, sy, gamma);
260
261 int64x2_t m0 = aom_sdotq_s16(vdupq_n_s64(0), s0, f[0]);
262 int64x2_t m1 = aom_sdotq_s16(vdupq_n_s64(0), s1, f[1]);
263 int64x2_t m2 = aom_sdotq_s16(vdupq_n_s64(0), s2, f[2]);
264 int64x2_t m3 = aom_sdotq_s16(vdupq_n_s64(0), s3, f[3]);
265 int64x2_t m4 = aom_sdotq_s16(vdupq_n_s64(0), s4, f[4]);
266 int64x2_t m5 = aom_sdotq_s16(vdupq_n_s64(0), s5, f[5]);
267 int64x2_t m6 = aom_sdotq_s16(vdupq_n_s64(0), s6, f[6]);
268 int64x2_t m7 = aom_sdotq_s16(vdupq_n_s64(0), s7, f[7]);
269
270 int64x2_t m01 = vpaddq_s64(m0, m1);
271 int64x2_t m23 = vpaddq_s64(m2, m3);
272 int64x2_t m45 = vpaddq_s64(m4, m5);
273 int64x2_t m67 = vpaddq_s64(m6, m7);
274
275 *res_low = vcombine_s32(vmovn_s64(m01), vmovn_s64(m23));
276 *res_high = vcombine_s32(vmovn_s64(m45), vmovn_s64(m67));
277 }
278
av1_warp_affine_sve(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)279 void av1_warp_affine_sve(const int32_t *mat, const uint8_t *ref, int width,
280 int height, int stride, uint8_t *pred, int p_col,
281 int p_row, int p_width, int p_height, int p_stride,
282 int subsampling_x, int subsampling_y,
283 ConvolveParams *conv_params, int16_t alpha,
284 int16_t beta, int16_t gamma, int16_t delta) {
285 av1_warp_affine_common(mat, ref, width, height, stride, pred, p_col, p_row,
286 p_width, p_height, p_stride, subsampling_x,
287 subsampling_y, conv_params, alpha, beta, gamma, delta);
288 }
289