forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
mp8x7p7q-neon.c
311 lines (281 loc) · 10.6 KB
/
mp8x7p7q-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
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
/*
* 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_mp8x7p7q__neon(
size_t m,
size_t n,
const uint8_t* input,
size_t input_stride,
const uint8_t* zero,
int32_t* buffer,
uint8_t* output,
const union pytorch_qnnp_avgpool_quantization_params
quantization_params[restrict static 1]) {
assert(m > 7);
assert(n >= 8);
const uint8_t* i0 = input;
const uint8_t* i1 = i0 + input_stride;
const uint8_t* i2 = i1 + input_stride;
const uint8_t* i3 = i2 + input_stride;
const uint8_t* i4 = i3 + input_stride;
const uint8_t* i5 = i4 + input_stride;
const uint8_t* i6 = i5 + input_stride;
const size_t packed_n = (n + 7) & -8;
const size_t input_increment = 7 * input_stride - packed_n;
const int32x4_t vbias = vld1q_dup_s32(&quantization_params->neon.bias);
/* note: goes up to 7 elements over bound */
int32_t* acc = buffer;
for (size_t k = 0; k < n; k += 8) {
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));
vst1q_s32(acc, vacc_lo);
acc += 4;
vst1q_s32(acc, vacc_hi);
acc += 4;
}
for (m -= 7; m > 7; m -= 7) {
acc = buffer;
i0 = (const uint8_t*)((uintptr_t)i0 + input_increment);
i1 = (const uint8_t*)((uintptr_t)i1 + input_increment);
i2 = (const uint8_t*)((uintptr_t)i2 + input_increment);
i3 = (const uint8_t*)((uintptr_t)i3 + input_increment);
i4 = (const uint8_t*)((uintptr_t)i4 + input_increment);
i5 = (const uint8_t*)((uintptr_t)i5 + input_increment);
i6 = (const uint8_t*)((uintptr_t)i6 + input_increment);
/* note: goes up to 7 elements over bound */
for (size_t k = 0; k < n; k += 8) {
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;
int32x4_t vacc_lo = vld1q_s32(acc);
int32x4_t vacc_hi = vld1q_s32(acc + 4);
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));
vacc_lo = vaddw_s16(vacc_lo, vget_low_s16(vsum23));
vacc_hi = vaddw_s16(vacc_hi, 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));
vst1q_s32(acc, vacc_lo);
acc += 4;
vst1q_s32(acc, vacc_hi);
acc += 4;
}
}
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
i0 = (const uint8_t*)((uintptr_t)i0 + input_increment);
i1 = (const uint8_t*)((uintptr_t)i1 + input_increment);
if (m < 2) {
i1 = zero;
}
i2 = (const uint8_t*)((uintptr_t)i2 + input_increment);
if (m <= 2) {
i2 = zero;
}
i3 = (const uint8_t*)((uintptr_t)i3 + input_increment);
if (m < 4) {
i3 = zero;
}
i4 = (const uint8_t*)((uintptr_t)i4 + input_increment);
if (m <= 4) {
i4 = zero;
}
i5 = (const uint8_t*)((uintptr_t)i5 + input_increment);
if (m < 6) {
i5 = zero;
}
i6 = (const uint8_t*)((uintptr_t)i6 + input_increment);
if (m <= 6) {
i6 = zero;
}
acc = buffer;
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;
int32x4_t vacc_lo = vld1q_s32(acc);
acc += 4;
int32x4_t vacc_hi = vld1q_s32(acc);
acc += 4;
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));
vacc_lo = vaddw_s16(vacc_lo, vget_low_s16(vsum23));
vacc_hi = vaddw_s16(vacc_hi, 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));
int32x4_t vacc_lo = vld1q_s32(acc);
acc += 4;
int32x4_t vacc_hi = vld1q_s32(acc);
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));
vacc_lo = vaddw_s16(vacc_lo, vget_low_s16(vsum23));
vacc_hi = vaddw_s16(vacc_hi, 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);
}
}
}