forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
up8x7-neon.c
210 lines (188 loc) · 7.1 KB
/
up8x7-neon.c
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
/*
* Copyright (c) Facebook, Inc. and its affiliates.
* All rights reserved.
*
* This source code is licensed under the BSD-style license found in the
* LICENSE file in the root directory of this source tree.
*/
#include <assert.h>
#include <arm_neon.h>
#include <qnnpack/q8gavgpool.h>
void pytorch_q8gavgpool_ukernel_up8x7__neon(
size_t m,
size_t n,
const uint8_t* input,
size_t input_stride,
const uint8_t* zero,
uint8_t* output,
const union pytorch_qnnp_avgpool_quantization_params
quantization_params[restrict static 1]) {
assert(m >= 1);
assert(m <= 7);
assert(n >= 8);
const uint8_t* i0 = input;
const uint8_t* i1 = i0 + input_stride;
if (m < 2) {
i1 = zero;
}
const uint8_t* i2 = i1 + input_stride;
if (m <= 2) {
i2 = zero;
}
const uint8_t* i3 = i2 + input_stride;
if (m < 4) {
i3 = zero;
}
const uint8_t* i4 = i3 + input_stride;
if (m <= 4) {
i4 = zero;
}
const uint8_t* i5 = i4 + input_stride;
if (m < 6) {
i5 = zero;
}
const uint8_t* i6 = i5 + input_stride;
if (m <= 6) {
i6 = zero;
}
const int32x4_t vbias = vld1q_dup_s32(&quantization_params->neon.bias);
const float32x4_t vscale = vdupq_n_f32(quantization_params->neon.scale);
#if defined(__aarch64__)
const int16x8_t voutput_zero_point =
vld1q_dup_s16(&quantization_params->neon.output_zero_point);
const uint8x8_t voutput_min =
vld1_dup_u8(&quantization_params->neon.output_min);
const uint8x8_t voutput_max =
vld1_dup_u8(&quantization_params->neon.output_max);
#else
const float32x4_t vfmin = vdupq_n_f32(quantization_params->neon.vfmin);
const float32x4_t vfmax = vdupq_n_f32(quantization_params->neon.vfmax);
const float32x4_t vfmagic = vdupq_n_f32(quantization_params->neon.vfmagic);
const int32x4_t vimagic = vdupq_n_s32(quantization_params->neon.vimagic);
#endif
do {
const uint8x8_t vi0 = vld1_u8(i0);
i0 += 8;
const uint8x8_t vi1 = vld1_u8(i1);
i1 += 8;
const uint8x8_t vi2 = vld1_u8(i2);
i2 += 8;
const uint8x8_t vi3 = vld1_u8(i3);
i3 += 8;
const uint8x8_t vi4 = vld1_u8(i4);
i4 += 8;
const uint8x8_t vi5 = vld1_u8(i5);
i5 += 8;
const uint8x8_t vi6 = vld1_u8(i6);
i6 += 8;
const int16x8_t vsum016 =
vreinterpretq_s16_u16(vaddw_u8(vaddl_u8(vi0, vi1), vi6));
const int16x8_t vsum23 = vreinterpretq_s16_u16(vaddl_u8(vi2, vi3));
const int16x8_t vsum45 = vreinterpretq_s16_u16(vaddl_u8(vi4, vi5));
int32x4_t vacc_lo = vaddw_s16(vbias, vget_low_s16(vsum23));
int32x4_t vacc_hi = vaddw_s16(vbias, vget_high_s16(vsum23));
vacc_lo = vaddw_s16(vacc_lo, vget_low_s16(vsum45));
vacc_hi = vaddw_s16(vacc_hi, vget_high_s16(vsum45));
vacc_lo = vaddw_s16(vacc_lo, vget_low_s16(vsum016));
vacc_hi = vaddw_s16(vacc_hi, vget_high_s16(vsum016));
float32x4_t vacc_lo_f = vcvtq_f32_s32(vacc_lo);
float32x4_t vacc_hi_f = vcvtq_f32_s32(vacc_hi);
vacc_lo_f = vmulq_f32(vacc_lo_f, vscale);
vacc_hi_f = vmulq_f32(vacc_hi_f, vscale);
#if defined(__aarch64__)
vacc_lo = vcvtnq_s32_f32(vacc_lo_f);
vacc_hi = vcvtnq_s32_f32(vacc_hi_f);
const int16x8_t vacc = vqaddq_s16(
vqmovn_high_s32(vqmovn_s32(vacc_lo), vacc_hi), voutput_zero_point);
uint8x8_t vout = vqmovun_s16(vacc);
vout = vmax_u8(vout, voutput_min);
vout = vmin_u8(vout, voutput_max);
#else
vacc_lo_f = vminq_f32(vmaxq_f32(vacc_lo_f, vfmin), vfmax);
vacc_hi_f = vminq_f32(vmaxq_f32(vacc_hi_f, vfmin), vfmax);
vacc_lo = vsubq_s32(
vreinterpretq_s32_f32(vaddq_f32(vacc_lo_f, vfmagic)), vimagic);
vacc_hi = vsubq_s32(
vreinterpretq_s32_f32(vaddq_f32(vacc_hi_f, vfmagic)), vimagic);
const int16x8_t vacc =
vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi));
uint8x8_t vout = vqmovun_s16(vacc);
#endif
vst1_u8(output, vout);
output += 8;
n -= 8;
} while (n >= 8);
if (n != 0) {
const size_t address_increment = n - 8;
i0 = (const uint8_t*)((uintptr_t)i0 + address_increment);
i1 = (const uint8_t*)((uintptr_t)i1 + address_increment);
i2 = (const uint8_t*)((uintptr_t)i2 + address_increment);
i3 = (const uint8_t*)((uintptr_t)i3 + address_increment);
i4 = (const uint8_t*)((uintptr_t)i4 + address_increment);
i5 = (const uint8_t*)((uintptr_t)i5 + address_increment);
i6 = (const uint8_t*)((uintptr_t)i6 + address_increment);
const int64x1_t vshift = vmov_n_s64(8 * address_increment);
const uint8x8_t vi0 =
vreinterpret_u8_u64(vshl_u64(vreinterpret_u64_u8(vld1_u8(i0)), vshift));
const uint8x8_t vi1 =
vreinterpret_u8_u64(vshl_u64(vreinterpret_u64_u8(vld1_u8(i1)), vshift));
const uint8x8_t vi2 =
vreinterpret_u8_u64(vshl_u64(vreinterpret_u64_u8(vld1_u8(i2)), vshift));
const uint8x8_t vi3 =
vreinterpret_u8_u64(vshl_u64(vreinterpret_u64_u8(vld1_u8(i3)), vshift));
const uint8x8_t vi4 =
vreinterpret_u8_u64(vshl_u64(vreinterpret_u64_u8(vld1_u8(i4)), vshift));
const uint8x8_t vi5 =
vreinterpret_u8_u64(vshl_u64(vreinterpret_u64_u8(vld1_u8(i5)), vshift));
const uint8x8_t vi6 =
vreinterpret_u8_u64(vshl_u64(vreinterpret_u64_u8(vld1_u8(i6)), vshift));
const int16x8_t vsum016 =
vreinterpretq_s16_u16(vaddw_u8(vaddl_u8(vi0, vi1), vi6));
const int16x8_t vsum23 = vreinterpretq_s16_u16(vaddl_u8(vi2, vi3));
const int16x8_t vsum45 = vreinterpretq_s16_u16(vaddl_u8(vi4, vi5));
int32x4_t vacc_lo = vaddw_s16(vbias, vget_low_s16(vsum23));
int32x4_t vacc_hi = vaddw_s16(vbias, vget_high_s16(vsum23));
vacc_lo = vaddw_s16(vacc_lo, vget_low_s16(vsum45));
vacc_hi = vaddw_s16(vacc_hi, vget_high_s16(vsum45));
vacc_lo = vaddw_s16(vacc_lo, vget_low_s16(vsum016));
vacc_hi = vaddw_s16(vacc_hi, vget_high_s16(vsum016));
float32x4_t vacc_lo_f = vcvtq_f32_s32(vacc_lo);
float32x4_t vacc_hi_f = vcvtq_f32_s32(vacc_hi);
vacc_lo_f = vmulq_f32(vacc_lo_f, vscale);
vacc_hi_f = vmulq_f32(vacc_hi_f, vscale);
#if defined(__aarch64__)
vacc_lo = vcvtnq_s32_f32(vacc_lo_f);
vacc_hi = vcvtnq_s32_f32(vacc_hi_f);
const int16x8_t vacc = vqaddq_s16(
vqmovn_high_s32(vqmovn_s32(vacc_lo), vacc_hi), voutput_zero_point);
uint8x8_t vout = vqmovun_s16(vacc);
vout = vmax_u8(vout, voutput_min);
vout = vmin_u8(vout, voutput_max);
#else
vacc_lo_f = vminq_f32(vmaxq_f32(vacc_lo_f, vfmin), vfmax);
vacc_hi_f = vminq_f32(vmaxq_f32(vacc_hi_f, vfmin), vfmax);
vacc_lo = vsubq_s32(
vreinterpretq_s32_f32(vaddq_f32(vacc_lo_f, vfmagic)), vimagic);
vacc_hi = vsubq_s32(
vreinterpretq_s32_f32(vaddq_f32(vacc_hi_f, vfmagic)), vimagic);
const int16x8_t vacc =
vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi));
uint8x8_t vout = vqmovun_s16(vacc);
#endif
if (n & 4) {
vst1_lane_u32(
__builtin_assume_aligned(output, 1), vreinterpret_u32_u8(vout), 0);
output += 4;
vout = vext_u8(vout, vout, 4);
}
if (n & 2) {
vst1_lane_u16(
__builtin_assume_aligned(output, 1), vreinterpret_u16_u8(vout), 0);
output += 2;
vout = vext_u8(vout, vout, 2);
}
if (n & 1) {
vst1_lane_u8(output, vout, 0);
}
}
}