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
|
#include <cstdint>
#include <limits>
#include <immintrin.h>
#ifdef _OPENMP
#include <omp.h>
#endif
#include <fbgemm/QuantUtils.h>
namespace caffe2 {
namespace internal {
template <typename T>
void SegmentMomentsAVX2(
const int N,
const T* src,
int64_t* sum,
int64_t* sumsq);
template <>
void SegmentMomentsAVX2<uint8_t>(
const int N,
const uint8_t* src,
int64_t* sum,
int64_t* sumsq) {
constexpr int kVLen = 16;
const int n = N / kVLen * kVLen;
const int r = N % kVLen;
const __m256i kOneInt16 = _mm256_set1_epi16(0x01);
__m256i sum_v = _mm256_setzero_si256();
__m256i sumsq_v = _mm256_setzero_si256();
for (int i = 0; i < n; i += kVLen) {
const __m256i cur_v = _mm256_cvtepu8_epi16(
_mm_loadu_si128(reinterpret_cast<const __m128i*>(src + i)));
sum_v = _mm256_add_epi32(sum_v, _mm256_madd_epi16(cur_v, kOneInt16));
sumsq_v = _mm256_add_epi32(sumsq_v, _mm256_madd_epi16(cur_v, cur_v));
}
// NOLINTNEXTLINE(modernize-avoid-c-arrays,cppcoreguidelines-avoid-c-arrays,cppcoreguidelines-avoid-magic-numbers)
int32_t sum_arr[8];
// NOLINTNEXTLINE(modernize-avoid-c-arrays,cppcoreguidelines-avoid-c-arrays,cppcoreguidelines-avoid-magic-numbers)
int32_t sumsq_arr[8];
_mm256_storeu_si256(reinterpret_cast<__m256i*>(sum_arr), sum_v);
_mm256_storeu_si256(reinterpret_cast<__m256i*>(sumsq_arr), sumsq_v);
for (int i = 0; i < 8; ++i) {
*sum += static_cast<int64_t>(sum_arr[i]);
*sumsq += static_cast<int64_t>(sumsq_arr[i]);
}
for (int i = 0; i < r; ++i) {
*sum += static_cast<int64_t>(src[n + i]);
*sumsq +=
static_cast<int64_t>(src[n + i]) * static_cast<int64_t>(src[n + i]);
}
}
template <typename T>
void VectorMomentsAVX2(const int N, const T* src, int64_t* sum, int64_t* sumsq);
template <>
void VectorMomentsAVX2<uint8_t>(
const int N,
const uint8_t* src,
int64_t* sum,
int64_t* sumsq) {
constexpr int kVLen = 32768;
const int n = N / kVLen * kVLen;
const int r = N % kVLen;
for (int i = 0; i < n; i += kVLen) {
SegmentMomentsAVX2<uint8_t>(kVLen, src + i, sum, sumsq);
}
if (r > 0) {
SegmentMomentsAVX2<uint8_t>(r, src + n, sum, sumsq);
}
}
void ComputeQuantizedFusedParamsAVX2(
const int N,
const int G,
const int K,
const int32_t X_zero_point,
const int32_t* mu,
const int32_t* rsig,
const int32_t* gamma,
int32_t* scale,
int32_t* bias) {
constexpr int kVLen = 8;
const int k = K / kVLen * kVLen;
const int r = K % kVLen;
for (int n = N - 1; n >= 0; --n) {
#ifdef _OPENMP
#pragma omp parallel for
#endif
for (int g = 0; g < G; ++g) {
const __m256i mu_v = _mm256_set1_epi32(mu[n * G + g] + X_zero_point);
const __m256i rsig_v = _mm256_set1_epi32(rsig[n * G + g]);
for (int i = 0; i < k; i += kVLen) {
const __m256i gamma_v =
_mm256_loadu_si256((const __m256i*)(gamma + g * K + i));
const __m256i beta_v =
_mm256_loadu_si256((const __m256i*)(bias + g * K + i));
__m256i scale_v = _mm256_mullo_epi32(gamma_v, rsig_v);
__m256i bias_v =
_mm256_sub_epi32(beta_v, _mm256_mullo_epi32(scale_v, mu_v));
const int offset = (n * G + g) * K + i;
_mm256_storeu_si256((__m256i*)(scale + offset), scale_v);
_mm256_storeu_si256((__m256i*)(bias + offset), bias_v);
}
for (int i = 0; i < r; ++i) {
const int offset = (n * G + g) * K + k + i;
scale[offset] = gamma[g * K + k + i] * rsig[n * G + g];
bias[offset] = bias[g * K + k + i] -
scale[offset] * (mu[n * G + g] + X_zero_point);
}
}
}
}
#define INIT_REQUANTIZE_AVX2 \
const __m256i b = _mm256_set1_epi32(params.multiplier); \
const __m256i prev_shift_nudge = _mm256_set1_epi64x( \
(1ll << (params.right_shift - 1)) + 0x8000000000000000ULL); \
const __m256i post_shift_nudge = _mm256_set1_epi64x( \
params.target_qparams.zero_point - \
(0x8000000000000000ULL >> params.right_shift)); \
const __m256i min_v = \
_mm256_set1_epi32(std::numeric_limits<uint8_t>::min()); \
const __m256i max_v = \
_mm256_set1_epi32(std::numeric_limits<uint8_t>::max()); \
const __m256i shuffle_mask_v = _mm256_set_epi8( \
0xff, \
0xff, \
0xff, \
0xff, \
0xff, \
0xff, \
0xff, \
0xff, \
0xff, \
0xff, \
0xff, \
0xff, \
0x0c, \
0x08, \
0x04, \
0x00, \
0xff, \
0xff, \
0xff, \
0xff, \
0xff, \
0xff, \
0xff, \
0xff, \
0xff, \
0xff, \
0xff, \
0xff, \
0x0c, \
0x08, \
0x04, \
0x00); \
const __m256i permute_mask_v = \
_mm256_set_epi32(0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00);
#define REQUANTIZE_AVX2(params, src, dst) \
do { \
__m256i a_v = (src); \
__m256i a_even_v = a_v; \
__m256i a_odd_v = _mm256_srli_si256(a_v, 4); \
__m256i ab_even_v = _mm256_mul_epi32(a_even_v, b); \
__m256i ab_odd_v = _mm256_mul_epi32(a_odd_v, b); \
__m256i even_rounded_v = _mm256_add_epi64(ab_even_v, prev_shift_nudge); \
__m256i odd_rounded_v = _mm256_add_epi64(ab_odd_v, prev_shift_nudge); \
__m256i even_result_v = _mm256_add_epi64( \
_mm256_srli_epi64(even_rounded_v, params.right_shift), \
post_shift_nudge); \
__m256i odd_result_v = _mm256_add_epi64( \
_mm256_srli_epi64(odd_rounded_v, params.right_shift), \
post_shift_nudge); \
odd_result_v = _mm256_slli_si256(odd_result_v, 4); \
__m256i result_v = _mm256_blend_epi32(even_result_v, odd_result_v, 0xaa); \
__m256i clipped_v = \
_mm256_max_epi32(min_v, _mm256_min_epi32(max_v, result_v)); \
clipped_v = _mm256_shuffle_epi8(clipped_v, shuffle_mask_v); \
clipped_v = _mm256_permutevar8x32_epi32(clipped_v, permute_mask_v); \
*(int64_t*)(dst) = _mm256_extract_epi64(clipped_v, 0); \
} while (false)
template <typename T>
void AffineBatchChannelAndRequantizeNCHWAVX2(
const int N,
const int C,
const int HxW,
const fbgemm::RequantizationParams& params,
const T* X,
const int32_t* scale,
const int32_t* bias,
T* Y);
template <>
void AffineBatchChannelAndRequantizeNCHWAVX2<uint8_t>(
const int N,
const int C,
const int HxW,
const fbgemm::RequantizationParams& params,
const uint8_t* X,
const int32_t* scale,
const int32_t* bias,
uint8_t* Y) {
INIT_REQUANTIZE_AVX2;
constexpr int kVLen = 8;
const int outer_size = N * C;
const int n = HxW / kVLen * kVLen;
const int r = HxW % kVLen;
#ifdef _OPENMP
#pragma omp parallel for
#endif
for (int i = 0; i < outer_size; ++i) {
const uint8_t* X_ptr = X + i * HxW;
uint8_t* Y_ptr = Y + i * HxW;
const __m256i scale_v = _mm256_set1_epi32(scale[i]);
const __m256i bias_v = _mm256_set1_epi32(bias[i]);
for (int j = 0; j < n; j += kVLen) {
const __m256i cur_v =
_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*)(X_ptr + j)));
REQUANTIZE_AVX2(
params,
_mm256_add_epi32(_mm256_mullo_epi32(cur_v, scale_v), bias_v),
(Y_ptr + j));
}
for (int j = 0; j < r; ++j) {
Y_ptr[n + j] = fbgemm::Requantize<uint8_t>(
static_cast<int32_t>(X_ptr[n + j]) * scale[i] + bias[i], params);
}
}
}
template <typename T>
void AffineBatchChannelAndRequantizeNHWCAVX2(
const int N,
const int C,
const int HxW,
const fbgemm::RequantizationParams& params,
const T* X,
const int32_t* scale,
const int32_t* bias,
T* Y);
template <>
void AffineBatchChannelAndRequantizeNHWCAVX2<uint8_t>(
const int N,
const int C,
const int HxW,
const fbgemm::RequantizationParams& params,
const uint8_t* X,
const int32_t* scale,
const int32_t* bias,
uint8_t* Y) {
INIT_REQUANTIZE_AVX2;
constexpr int kVLen = 8;
const int outer_size = N * HxW;
#ifdef _OPENMP
#pragma omp parallel for
#endif
for (int i = 0; i < outer_size; ++i) {
const int c = i / HxW * C;
const int n = C / kVLen * kVLen;
const int r = C % kVLen;
const uint8_t* X_ptr = X + i * C;
uint8_t* Y_ptr = Y + i * C;
for (int j = 0; j < n; j += kVLen) {
const __m256i cur_v =
_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*)(X_ptr + j)));
const __m256i scale_v =
_mm256_loadu_si256((const __m256i*)(scale + c + j));
const __m256i bias_v = _mm256_loadu_si256((const __m256i*)(bias + c + j));
REQUANTIZE_AVX2(
params,
_mm256_add_epi32(_mm256_mullo_epi32(cur_v, scale_v), bias_v),
(Y_ptr + j));
}
for (int j = 0; j < r; ++j) {
Y_ptr[n + j] = fbgemm::Requantize<uint8_t>(
static_cast<int32_t>(X_ptr[n + j]) * scale[c + n + j] +
bias[c + n + j],
params);
}
}
}
#undef REQUANTIZE_AVX2
#undef INIT_REQUANTIZE_AVX2
} // namespace internal
} // namespace caffe2
|
