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
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
|
#include "caffe2/quantization/server/pool_dnnlowp_op_avx2.h"
#include <immintrin.h>
#include <algorithm>
#include <cmath>
namespace caffe2 {
using namespace std;
void max_pool_avx2(
const uint8_t* Xdata,
int n,
int height,
int width,
int channels,
int pooled_height,
int pooled_width,
int kernel_h,
int kernel_w,
int stride_h,
int stride_w,
int pad_t,
int pad_l,
uint8_t* Ydata) {
const uint8_t* Xdata_temp = Xdata + n * height * width * channels;
uint8_t* Ydata_temp = Ydata + n * pooled_height * pooled_width * channels;
for (int ph = 0; ph < pooled_height; ++ph) {
int hstart = ph * stride_h - pad_t;
int hend = hstart + kernel_h < height ? hstart + kernel_h : height;
hstart = hstart > 0 ? hstart : 0;
for (int pw = 0; pw < pooled_width; ++pw) {
int wstart = pw * stride_w - pad_l;
int wend = wstart + kernel_w < width ? wstart + kernel_w : width;
wstart = wstart > 0 ? wstart : 0;
uint8_t* Yh = Ydata_temp + (ph * pooled_width + pw) * channels;
constexpr int VLEN = 32;
// vectorized loop
for (int c = 0; c < channels / VLEN * VLEN; c += VLEN) {
__m256i Y_v = _mm256_setzero_si256();
for (int h = hstart; h < hend; ++h) {
for (int w = wstart; w < wend; ++w) {
const int input_idx = (h * width + w) * channels + c;
Y_v = _mm256_max_epu8(
_mm256_loadu_si256(
reinterpret_cast<const __m256i*>(Xdata_temp + input_idx)),
Y_v);
}
}
_mm256_storeu_si256(reinterpret_cast<__m256i*>(Yh + c), Y_v);
}
// remainder
for (int c = channels / VLEN * VLEN; c < channels; ++c) {
Yh[c] = 0;
}
for (int h = hstart; h < hend; ++h) {
for (int w = wstart; w < wend; ++w) {
for (int c = channels / VLEN * VLEN; c < channels; ++c) {
const int input_idx = (h * width + w) * channels + c;
Yh[c] =
Xdata_temp[input_idx] > Yh[c] ? Xdata_temp[input_idx] : Yh[c];
}
}
}
} // pw loop
} // ph loop
}
void average_pool_avx2(
const uint8_t* Xdata,
int n,
int height,
int width,
int channels,
int pooled_height,
int pooled_width,
int kernel_h,
int kernel_w,
int stride_h,
int stride_w,
int pad_t,
int pad_l,
uint8_t* Ydata,
float in_scale,
float out_scale,
int32_t in_zero_point,
int32_t out_zero_point,
int32_t minimum,
int32_t maximum) {
const uint8_t* Xdata_temp = Xdata + n * height * width * channels;
uint8_t* Ydata_temp = Ydata + n * pooled_height * pooled_width * channels;
const __m256i shuffle_mask_v = _mm256_set_epi8(
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
0x0c,
0x08,
0x04,
0x00,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
0x0c,
0x08,
0x04,
0x00);
const __m256i permute_mask_v =
_mm256_set_epi32(0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00);
const __m256i min_v = _mm256_set1_epi32(minimum);
const __m256i max_v = _mm256_set1_epi32(maximum);
__m256 out_zero_point_v = _mm256_set1_ps(out_zero_point);
for (int ph = 0; ph < pooled_height; ++ph) {
int hstart = ph * stride_h - pad_t;
int hend = hstart + kernel_h < height ? hstart + kernel_h : height;
hstart = hstart > 0 ? hstart : 0;
for (int pw = 0; pw < pooled_width; ++pw) {
int wstart = pw * stride_w - pad_l;
int wend = wstart + kernel_w < width ? wstart + kernel_w : width;
wstart = wstart > 0 ? wstart : 0;
int size = (hend - hstart) * (wend - wstart);
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
float multiplier = in_scale / out_scale / size;
__m256 multiplier_v = _mm256_set1_ps(multiplier);
uint8_t* Yh = Ydata_temp + (ph * pooled_width + pw) * channels;
constexpr int VLEN = 8;
int32_t Yh0 = -in_zero_point * size;
// vectorized loop
for (int c = 0; c < channels / VLEN * VLEN; c += VLEN) {
__m256i Yh0_v = _mm256_set1_epi32(Yh0);
for (int h = hstart; h < hend; ++h) {
for (int w = wstart; w < wend; ++w) {
const int input_idx = (h * width + w) * channels + c;
const __m256i temp_v = _mm256_cvtepu8_epi32(_mm_loadl_epi64(
reinterpret_cast<const __m128i*>(Xdata_temp + input_idx)));
Yh0_v = _mm256_add_epi32(Yh0_v, temp_v);
}
}
__m256 Yh0_fp = _mm256_cvtepi32_ps(Yh0_v);
__m256 Y_float_v =
_mm256_fmadd_ps(Yh0_fp, multiplier_v, out_zero_point_v);
__m256i Y_rounded_v = _mm256_cvtps_epi32(Y_float_v);
__m256i Y_clipped_v =
_mm256_max_epi32(min_v, _mm256_min_epi32(max_v, Y_rounded_v));
Y_clipped_v = _mm256_shuffle_epi8(Y_clipped_v, shuffle_mask_v);
Y_clipped_v = _mm256_permutevar8x32_epi32(Y_clipped_v, permute_mask_v);
*reinterpret_cast<int64_t*>(Yh + c) =
_mm256_extract_epi64(Y_clipped_v, 0);
}
// remainder
for (int c = channels / VLEN * VLEN; c < channels; ++c) {
Yh[c] = 0;
}
for (int c = channels / VLEN * VLEN; c < channels; ++c) {
const int pool_idx = (ph * pooled_width + pw) * channels + c;
int32_t Yh_t = -in_zero_point * size;
for (int h = hstart; h < hend; ++h) {
for (int w = wstart; w < wend; ++w) {
const int input_idx = (h * width + w) * channels + c;
Yh_t += Xdata_temp[input_idx];
}
}
Ydata_temp[pool_idx] = std::min<int32_t>(
std::max<int32_t>(
nearbyint(Yh_t * multiplier + out_zero_point), minimum),
maximum);
}
} // pw loop
} // ph loop
}
void average_pool_3d_avx2(
const uint8_t* Xdata,
int n,
int height,
int width,
int depth,
int channels,
int pooled_height,
int pooled_width,
int pooled_depth,
int kernel_h,
int kernel_w,
int kernel_d,
int stride_h,
int stride_w,
int stride_d,
int pad_t,
int pad_l,
int pad_d,
uint8_t* Ydata,
float in_scale,
float out_scale,
int32_t in_zero_point,
int32_t out_zero_point,
int32_t minimum,
int32_t maximum) {
const uint8_t* Xdata_temp = Xdata + n * height * width * depth * channels;
uint8_t* Ydata_temp =
Ydata + n * pooled_height * pooled_width * pooled_depth * channels;
const __m256i shuffle_mask_v = _mm256_set_epi8(
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
0x0c,
0x08,
0x04,
0x00,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
0xff,
0x0c,
0x08,
0x04,
0x00);
const __m256i permute_mask_v =
_mm256_set_epi32(0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00);
const __m256i min_v = _mm256_set1_epi32(minimum);
const __m256i max_v = _mm256_set1_epi32(maximum);
__m256 out_zero_point_v = _mm256_set1_ps(out_zero_point);
for (int ph = 0; ph < pooled_height; ++ph) {
int hstart = ph * stride_h - pad_t;
int hend = hstart + kernel_h < height ? hstart + kernel_h : height;
hstart = hstart > 0 ? hstart : 0;
for (int pw = 0; pw < pooled_width; ++pw) {
int wstart = pw * stride_w - pad_l;
int wend = wstart + kernel_w < width ? wstart + kernel_w : width;
wstart = wstart > 0 ? wstart : 0;
for (int pd = 0; pd < pooled_depth; ++pd) {
int dstart = pd * stride_d - pad_d;
int dend = dstart + kernel_d < depth ? dstart + kernel_d : depth;
dstart = max(dstart, 0);
int size = (hend - hstart) * (wend - wstart) * (dend - dstart);
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
float multiplier = in_scale / out_scale / size;
__m256 multiplier_v = _mm256_set1_ps(multiplier);
uint8_t* Yh = Ydata_temp +
((ph * pooled_width + pw) * pooled_depth + pd) * channels;
constexpr int VLEN = 8;
int32_t Yh0 = -in_zero_point * size;
// vectorized loop
for (int c = 0; c < channels / VLEN * VLEN; c += VLEN) {
__m256i Yh0_v = _mm256_set1_epi32(Yh0);
for (int h = hstart; h < hend; ++h) {
for (int w = wstart; w < wend; ++w) {
for (int d = dstart; d < dend; ++d) {
const int input_idx =
((h * width + w) * depth + d) * channels + c;
const __m256i temp_v = _mm256_cvtepu8_epi32(_mm_loadl_epi64(
reinterpret_cast<const __m128i*>(Xdata_temp + input_idx)));
Yh0_v = _mm256_add_epi32(Yh0_v, temp_v);
}
}
}
__m256 Yh0_fp = _mm256_cvtepi32_ps(Yh0_v);
__m256 Y_float_v =
_mm256_fmadd_ps(Yh0_fp, multiplier_v, out_zero_point_v);
__m256i Y_rounded_v = _mm256_cvtps_epi32(Y_float_v);
__m256i Y_clipped_v =
_mm256_max_epi32(min_v, _mm256_min_epi32(max_v, Y_rounded_v));
Y_clipped_v = _mm256_shuffle_epi8(Y_clipped_v, shuffle_mask_v);
Y_clipped_v =
_mm256_permutevar8x32_epi32(Y_clipped_v, permute_mask_v);
*reinterpret_cast<int64_t*>(Yh + c) =
_mm256_extract_epi64(Y_clipped_v, 0);
}
// remainder
for (int c = channels / VLEN * VLEN; c < channels; ++c) {
Yh[c] = 0;
}
for (int c = channels / VLEN * VLEN; c < channels; ++c) {
const int pool_idx =
((ph * pooled_width + pw) * pooled_depth + pd) * channels + c;
int32_t Yh_t = -in_zero_point * size;
for (int h = hstart; h < hend; ++h) {
for (int w = wstart; w < wend; ++w) {
for (int d = dstart; d < dend; ++d) {
const int input_idx =
((h * width + w) * depth + d) * channels + c;
Yh_t += Xdata_temp[input_idx];
}
}
}
Ydata_temp[pool_idx] = std::min<int32_t>(
std::max<int32_t>(
nearbyint(Yh_t * multiplier + out_zero_point), minimum),
maximum);
}
} // pd loop
} // pw loop
} // ph loop
}
} // namespace caffe2
|
