1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
|
/*
* Copyright (c) 2018, Alliance for Open Media. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef AV1_COMMON_ARM_CONVOLVE_NEON_H_
#define AV1_COMMON_ARM_CONVOLVE_NEON_H_
#include <arm_neon.h>
#define HORIZ_EXTRA_ROWS ((SUBPEL_TAPS + 7) & ~0x07)
static INLINE uint8x8_t wiener_convolve8_vert_4x8(
const int16x8_t s0, const int16x8_t s1, const int16x8_t s2,
const int16x8_t s3, const int16x8_t s4, const int16x8_t s5,
const int16x8_t s6, int16_t *filter_y, const int bd,
const int round1_bits) {
int16x8_t ss0, ss1, ss2;
int32x4_t sum0, sum1;
uint16x4_t tmp0, tmp1;
uint16x8_t tmp;
uint8x8_t res;
const int32_t round_const = (1 << (bd + round1_bits - 1));
const int32x4_t round_bits = vdupq_n_s32(-round1_bits);
const int32x4_t zero = vdupq_n_s32(0);
const int32x4_t round_vec = vdupq_n_s32(round_const);
ss0 = vaddq_s16(s0, s6);
ss1 = vaddq_s16(s1, s5);
ss2 = vaddq_s16(s2, s4);
sum0 = vmull_n_s16(vget_low_s16(ss0), filter_y[0]);
sum0 = vmlal_n_s16(sum0, vget_low_s16(ss1), filter_y[1]);
sum0 = vmlal_n_s16(sum0, vget_low_s16(ss2), filter_y[2]);
sum0 = vmlal_n_s16(sum0, vget_low_s16(s3), filter_y[3]);
sum1 = vmull_n_s16(vget_high_s16(ss0), filter_y[0]);
sum1 = vmlal_n_s16(sum1, vget_high_s16(ss1), filter_y[1]);
sum1 = vmlal_n_s16(sum1, vget_high_s16(ss2), filter_y[2]);
sum1 = vmlal_n_s16(sum1, vget_high_s16(s3), filter_y[3]);
sum0 = vsubq_s32(sum0, round_vec);
sum1 = vsubq_s32(sum1, round_vec);
/* right shift & rounding */
sum0 = vrshlq_s32(sum0, round_bits);
sum1 = vrshlq_s32(sum1, round_bits);
sum0 = vmaxq_s32(sum0, zero);
sum1 = vmaxq_s32(sum1, zero);
/* from int32x4_t to uint8x8_t */
tmp0 = vqmovn_u32(vreinterpretq_u32_s32(sum0));
tmp1 = vqmovn_u32(vreinterpretq_u32_s32(sum1));
tmp = vcombine_u16(tmp0, tmp1);
res = vqmovn_u16(tmp);
return res;
}
static INLINE uint16x8_t wiener_convolve8_horiz_8x8(
const int16x8_t s0, const int16x8_t s1, const int16x8_t s2,
const int16x8_t s3, int16_t *filter_x, const int bd,
const int round0_bits) {
int16x8_t sum;
uint16x8_t res;
int32x4_t sum_0, sum_1;
int32x4_t s3_0, s3_1;
const int32_t round_const_0 = (1 << (bd + FILTER_BITS - 1));
const int32_t round_const_1 = (1 << ((bd) + 1 + FILTER_BITS - round0_bits));
/* for the purpose of right shift by { conv_params->round_0 } */
const int32x4_t round_bits = vdupq_n_s32(-round0_bits);
const int32x4_t round_vec_0 = vdupq_n_s32(round_const_0);
const int32x4_t round_vec_1 = vdupq_n_s32(round_const_1);
sum = vmulq_n_s16(s0, filter_x[0]);
sum = vmlaq_n_s16(sum, s1, filter_x[1]);
sum = vmlaq_n_s16(sum, s2, filter_x[2]);
/* sum from 16x8 to 2 32x4 registers */
sum_0 = vmovl_s16(vget_low_s16(sum));
sum_1 = vmovl_s16(vget_high_s16(sum));
/* s[3]*128 -- and filter coef max can be 128
* then max value possible = 128*128*255 exceeding 16 bit
*/
s3_0 = vmull_n_s16(vget_low_s16(s3), filter_x[3]);
s3_1 = vmull_n_s16(vget_high_s16(s3), filter_x[3]);
sum_0 = vaddq_s32(sum_0, s3_0);
sum_1 = vaddq_s32(sum_1, s3_1);
/* Add the constant value */
sum_0 = vaddq_s32(sum_0, round_vec_0);
sum_1 = vaddq_s32(sum_1, round_vec_0);
/* right shift & rounding & saturating */
sum_0 = vqrshlq_s32(sum_0, round_bits);
sum_1 = vqrshlq_s32(sum_1, round_bits);
/* Clipping to max value */
sum_0 = vminq_s32(sum_0, round_vec_1);
sum_1 = vminq_s32(sum_1, round_vec_1);
res = vcombine_u16(vqmovun_s32(sum_0), vqmovun_s32(sum_1));
return res;
}
static INLINE uint16x4_t wiener_convolve8_horiz_4x8(
const int16x4_t s0, const int16x4_t s1, const int16x4_t s2,
const int16x4_t s3, const int16x4_t s4, const int16x4_t s5,
const int16x4_t s6, int16_t *filter_x, const int bd,
const int round0_bits) {
uint16x4_t res;
int32x4_t sum_0, s3_0;
int16x4_t sum, temp0, temp1, temp2;
const int32_t round_const_0 = (1 << (bd + FILTER_BITS - 1));
const int32_t round_const_1 = (1 << ((bd) + 1 + FILTER_BITS - round0_bits));
const int32x4_t round_bits = vdupq_n_s32(-round0_bits);
const int32x4_t zero = vdupq_n_s32(0);
const int32x4_t round_vec_0 = vdupq_n_s32(round_const_0);
const int32x4_t round_vec_1 = vdupq_n_s32(round_const_1);
temp0 = vadd_s16(s0, s6);
temp1 = vadd_s16(s1, s5);
temp2 = vadd_s16(s2, s4);
sum = vmul_n_s16(temp0, filter_x[0]);
sum = vmla_n_s16(sum, temp1, filter_x[1]);
sum = vmla_n_s16(sum, temp2, filter_x[2]);
sum_0 = vmovl_s16(sum);
/* s[3]*128 -- and filter coff max can be 128.
* then max value possible = 128*128*255 Therefore, 32 bits are required to
* hold the result.
*/
s3_0 = vmull_n_s16(s3, filter_x[3]);
sum_0 = vaddq_s32(sum_0, s3_0);
sum_0 = vaddq_s32(sum_0, round_vec_0);
sum_0 = vrshlq_s32(sum_0, round_bits);
sum_0 = vmaxq_s32(sum_0, zero);
sum_0 = vminq_s32(sum_0, round_vec_1);
res = vqmovun_s32(sum_0);
return res;
}
static INLINE int16x8_t
convolve8_8x8_s16(const int16x8_t s0, const int16x8_t s1, const int16x8_t s2,
const int16x8_t s3, const int16x8_t s4, const int16x8_t s5,
const int16x8_t s6, const int16x8_t s7, const int16_t *filter,
const int16x8_t horiz_const, const int16x8_t shift_round_0) {
int16x8_t sum;
int16x8_t res;
sum = horiz_const;
sum = vmlaq_n_s16(sum, s0, filter[0]);
sum = vmlaq_n_s16(sum, s1, filter[1]);
sum = vmlaq_n_s16(sum, s2, filter[2]);
sum = vmlaq_n_s16(sum, s3, filter[3]);
sum = vmlaq_n_s16(sum, s4, filter[4]);
sum = vmlaq_n_s16(sum, s5, filter[5]);
sum = vmlaq_n_s16(sum, s6, filter[6]);
sum = vmlaq_n_s16(sum, s7, filter[7]);
res = vqrshlq_s16(sum, shift_round_0);
return res;
}
static INLINE int16x4_t
convolve8_4x4_s16(const int16x4_t s0, const int16x4_t s1, const int16x4_t s2,
const int16x4_t s3, const int16x4_t s4, const int16x4_t s5,
const int16x4_t s6, const int16x4_t s7, const int16_t *filter,
const int16x4_t horiz_const, const int16x4_t shift_round_0) {
int16x4_t sum;
sum = horiz_const;
sum = vmla_n_s16(sum, s0, filter[0]);
sum = vmla_n_s16(sum, s1, filter[1]);
sum = vmla_n_s16(sum, s2, filter[2]);
sum = vmla_n_s16(sum, s3, filter[3]);
sum = vmla_n_s16(sum, s4, filter[4]);
sum = vmla_n_s16(sum, s5, filter[5]);
sum = vmla_n_s16(sum, s6, filter[6]);
sum = vmla_n_s16(sum, s7, filter[7]);
sum = vqrshl_s16(sum, shift_round_0);
return sum;
}
static INLINE uint16x4_t convolve8_4x4_s32(
const int16x4_t s0, const int16x4_t s1, const int16x4_t s2,
const int16x4_t s3, const int16x4_t s4, const int16x4_t s5,
const int16x4_t s6, const int16x4_t s7, const int16_t *y_filter,
const int32x4_t round_shift_vec, const int32x4_t offset_const) {
int32x4_t sum0;
uint16x4_t res;
const int32x4_t zero = vdupq_n_s32(0);
sum0 = vmull_n_s16(s0, y_filter[0]);
sum0 = vmlal_n_s16(sum0, s1, y_filter[1]);
sum0 = vmlal_n_s16(sum0, s2, y_filter[2]);
sum0 = vmlal_n_s16(sum0, s3, y_filter[3]);
sum0 = vmlal_n_s16(sum0, s4, y_filter[4]);
sum0 = vmlal_n_s16(sum0, s5, y_filter[5]);
sum0 = vmlal_n_s16(sum0, s6, y_filter[6]);
sum0 = vmlal_n_s16(sum0, s7, y_filter[7]);
sum0 = vaddq_s32(sum0, offset_const);
sum0 = vqrshlq_s32(sum0, round_shift_vec);
sum0 = vmaxq_s32(sum0, zero);
res = vmovn_u32(vreinterpretq_u32_s32(sum0));
return res;
}
#endif // AV1_COMMON_ARM_CONVOLVE_NEON_H_
|
