forked from pytorch/FBGEMM
-
Notifications
You must be signed in to change notification settings - Fork 0
/
FbgemmSparseDenseVectorInt8Avx512.cc
256 lines (234 loc) · 9.39 KB
/
FbgemmSparseDenseVectorInt8Avx512.cc
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
/*
* Copyright (c) Meta Platforms, Inc. and 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.
*/
#define FBGEMM_EXPORTS
#include "fbgemm/FbgemmSparse.h"
#include "fbgemm/Utils.h"
#include "fbgemm/spmmUtilsAvx2.h"
#include <immintrin.h>
#include <cassert>
namespace fbgemm {
namespace internal {
static inline int32_t horizontal_add(__m256i a) {
__m256i t1 = _mm256_hadd_epi32(a, a);
__m256i t2 = _mm256_hadd_epi32(t1, t1);
__m128i t3 = _mm256_extracti128_si256(t2, 1);
__m128i t4 = _mm_add_epi32(_mm256_castsi256_si128(t2), t3);
return _mm_cvtsi128_si32(t4);
}
template <
bool FUSE_RELU,
bool ACT_ZP_0, // is activation zero point 0?
bool HAS_BIAS,
QuantizationGranularity Q_GRAN>
static inline void requantizeForMV(
uint8_t* dst,
int32_t* src,
int len,
trRequantizationParams_t& rParams) {
constexpr int VLEN_INT32 = 16;
__m512i C_zero_point_epi8_v = _mm512_set1_epi8(rParams.C_zero_point);
__m512i C_zero_point_epi32_v = _mm512_set1_epi32(rParams.C_zero_point);
// clang-format off
__m512i permute_mask_v = _mm512_set_epi32(
0x0F, 0x0B, 0x07, 0x03,
0x0E, 0x0A, 0x06, 0x02,
0x0D, 0x09, 0x05, 0x01,
0x0C, 0x08, 0x04, 0x00);
// clang-format on
int i = 0;
for (; i < len / VLEN_INT32 * VLEN_INT32; i += VLEN_INT32) {
__m512i x_v = _mm512_loadu_si512(src + i);
if (!ACT_ZP_0) {
__m512i weight_row_offset_v =
_mm512_loadu_si512(rParams.weight_row_offsets + i);
__m512i act_zero_point_v = _mm512_set1_epi32(rParams.act_zero_point);
x_v = _mm512_sub_epi32(
x_v, _mm512_mullo_epi32(act_zero_point_v, weight_row_offset_v));
}
__m512 act_times_w_scale_v;
if (Q_GRAN == QuantizationGranularity::OUT_CHANNEL) {
act_times_w_scale_v = _mm512_loadu_ps(rParams.act_times_w_scale + i);
} else {
act_times_w_scale_v = _mm512_set1_ps(rParams.act_times_w_scale[0]);
}
__m512 c_scale_v = _mm512_set1_ps(rParams.C_scale);
__m512 act_times_w_div_c_v = _mm512_div_ps(act_times_w_scale_v, c_scale_v);
__m512 xf_v;
if (HAS_BIAS) {
__m512 bias_v = _mm512_loadu_ps(rParams.bias + i);
bias_v = _mm512_div_ps(bias_v, act_times_w_scale_v);
xf_v = _mm512_add_ps(_mm512_cvtepi32_ps(x_v), bias_v);
} else {
xf_v = _mm512_cvtepi32_ps(x_v);
}
__m512 x_scaled_v = _mm512_mul_ps(xf_v, act_times_w_div_c_v);
__m512i x_rounded_v = _mm512_cvtps_epi32(x_scaled_v);
__m512i x_added_v = _mm512_add_epi32(x_rounded_v, C_zero_point_epi32_v);
__m512i x_clamped_v = _mm512_packs_epi32(x_added_v, _mm512_setzero_si512());
x_clamped_v = _mm512_packus_epi16(x_clamped_v, _mm512_setzero_si512());
if (FUSE_RELU) {
x_clamped_v = _mm512_max_epu8(C_zero_point_epi8_v, x_clamped_v);
}
x_clamped_v = _mm512_permutexvar_epi32(permute_mask_v, x_clamped_v);
_mm_store_si128(
reinterpret_cast<__m128i*>(dst + i),
_mm512_castsi512_si128(x_clamped_v));
}
int rem_int32 = len - i;
if (rem_int32 > 0) {
__mmask64 mask_int8_v = (((long long)1) << rem_int32) - 1;
__mmask16 mask_int32_v = (((long long)1) << rem_int32) - 1;
__m512i x_v = _mm512_maskz_loadu_epi32(mask_int32_v, src + i);
if (!ACT_ZP_0) {
__m512i weight_row_offset_v = _mm512_maskz_loadu_epi32(
mask_int32_v, rParams.weight_row_offsets + i);
__m512i act_zero_point_v = _mm512_set1_epi32(rParams.act_zero_point);
x_v = _mm512_sub_epi32(
x_v, _mm512_mullo_epi32(act_zero_point_v, weight_row_offset_v));
}
__m512 act_times_w_scale_v;
if (Q_GRAN == QuantizationGranularity::OUT_CHANNEL) {
act_times_w_scale_v =
_mm512_maskz_loadu_ps(mask_int32_v, rParams.act_times_w_scale + i);
} else {
act_times_w_scale_v = _mm512_set1_ps(rParams.act_times_w_scale[0]);
}
__m512 c_scale_v = _mm512_set1_ps(rParams.C_scale);
__m512 act_times_w_div_c_v = _mm512_div_ps(act_times_w_scale_v, c_scale_v);
__m512 xf_v;
if (HAS_BIAS) {
__m512 bias_v = _mm512_maskz_loadu_ps(mask_int32_v, rParams.bias + i);
bias_v = _mm512_div_ps(bias_v, act_times_w_scale_v);
xf_v = _mm512_add_ps(_mm512_cvtepi32_ps(x_v), bias_v);
} else {
xf_v = _mm512_cvtepi32_ps(x_v);
}
__m512 x_scaled_v = _mm512_mul_ps(xf_v, act_times_w_div_c_v);
__m512i x_rounded_v = _mm512_cvtps_epi32(x_scaled_v);
__m512i x_added_v = _mm512_add_epi32(x_rounded_v, C_zero_point_epi32_v);
__m512i x_clamped_v = _mm512_packs_epi32(x_added_v, _mm512_setzero_si512());
x_clamped_v = _mm512_packus_epi16(x_clamped_v, _mm512_setzero_si512());
if (FUSE_RELU) {
x_clamped_v = _mm512_max_epu8(C_zero_point_epi8_v, x_clamped_v);
}
x_clamped_v = _mm512_permutexvar_epi32(permute_mask_v, x_clamped_v);
_mm512_mask_storeu_epi8(dst + i, mask_int8_v, x_clamped_v);
}
}
// matrix-vector product
// i.e., produces same results as SparseDenseInt8MMAvx512 with N == 1
template <bool FUSE_RELU, QuantizationGranularity Q_GRAN>
void SparseDenseInt8MVAvx512(
const std::unique_ptr<BCSRMatrix<>>& bcsr,
const uint8_t* B,
int ldb,
int32_t* C_i32,
uint8_t* C_u8,
trRequantizationParams_t& rParams,
bool accum,
int thread_id,
int num_threads) {
(void)num_threads; // Suppress unused variable warning
// Calcualtes accum ? C += A * B : C = A * B
constexpr int VLEN_INT32 = 16;
constexpr int block_size = BCSRMatrix<>::CB;
constexpr int colTileSize = BCSRMatrix<>::COLTILE;
// all work is done by thread 0 for now
assert(num_threads > 0 && "Numbers of threads should be > 0");
if (thread_id > 0) {
return;
}
assert(ldb == 1 && "ldb should be 1");
__m512i one_16bit_v = _mm512_set1_epi16(1);
// Number of columns in the sparse matrix A
int K = bcsr->C;
int M = bcsr->R;
assert(K % 4 == 0 && "K should be multiple of 4");
assert((K > 0) && "K needs to be positive");
int kTiles = (K + colTileSize - 1) / colTileSize;
const int* row_ptr = bcsr->rowBPtr.data();
const int* col_idx = bcsr->colBIdx.data();
const int8_t* values = bcsr->values.data();
for (int kt = 0; kt < kTiles; ++kt) {
const int* cur_row_ptr = row_ptr + kt * M;
const uint8_t* cur_B = B + kt * colTileSize * ldb;
// TODO: unroll this loop?
for (int i = 0; i < M; ++i) {
__m512i res = _mm512_set1_epi32(0);
int r = cur_row_ptr[i];
int r_end_aligned = cur_row_ptr[i] +
(cur_row_ptr[i + 1] - cur_row_ptr[i]) / VLEN_INT32 * VLEN_INT32;
for (; r < r_end_aligned; r += VLEN_INT32) {
__m512i a_v = _mm512_loadu_si512(values + r * block_size);
__m512i b_idx = _mm512_loadu_si512(col_idx + r);
__m512i b_v = _mm512_i32gather_epi32(
b_idx, reinterpret_cast<const int32_t*>(cur_B), block_size);
__m512i c_i16_v = _mm512_maddubs_epi16(b_v, a_v);
__m512i c_i32_v = _mm512_madd_epi16(one_16bit_v, c_i16_v);
res = _mm512_add_epi32(res, c_i32_v);
}
int rem = cur_row_ptr[i + 1] - r;
if (rem > 0) {
__mmask16 mask_int32_v = (((long long)1) << (rem)) - 1;
__m512i a_v =
_mm512_maskz_loadu_epi32(mask_int32_v, values + r * block_size);
__m512i b_idx = _mm512_maskz_loadu_epi32(mask_int32_v, col_idx + r);
__m512i b_v = _mm512_i32gather_epi32(
b_idx, reinterpret_cast<const int32_t*>(cur_B), block_size);
__m512i c_i16_v = _mm512_maddubs_epi16(b_v, a_v);
__m512i c_i32_v = _mm512_madd_epi16(one_16bit_v, c_i16_v);
res = _mm512_add_epi32(res, c_i32_v);
}
// Horizontal reduce
// _mm512_reduce_add_epi32 is only available for gcc version > 7
#if __GNUC__ >= 7
int32_t res_i32 = _mm512_reduce_add_epi32(res);
#else
__m256i low = _mm512_castsi512_si256(res);
__m256i high = _mm512_extracti64x4_epi64(res, 1);
int32_t res_i32 = horizontal_add(_mm256_add_epi32(low, high));
#endif
// store the results
if (accum || kt > 0) {
C_i32[i] += res_i32;
} else {
C_i32[i] = res_i32;
}
}
}
if (rParams.bias == nullptr) {
if (rParams.act_zero_point) {
requantizeForMV<FUSE_RELU, false, false, Q_GRAN>(C_u8, C_i32, M, rParams);
} else {
requantizeForMV<FUSE_RELU, true, false, Q_GRAN>(C_u8, C_i32, M, rParams);
}
} else {
if (rParams.act_zero_point) {
requantizeForMV<FUSE_RELU, false, true, Q_GRAN>(C_u8, C_i32, M, rParams);
} else {
requantizeForMV<FUSE_RELU, true, true, Q_GRAN>(C_u8, C_i32, M, rParams);
}
}
}
#define CREATE_INSTANCE(FUSE_RELU, QGRAN) \
template void SparseDenseInt8MVAvx512<FUSE_RELU, QGRAN>( \
const std::unique_ptr<BCSRMatrix<>>& bcsr, \
const uint8_t* B, \
int ldb, \
int32_t* C_i32, \
uint8_t* C_u8, \
trRequantizationParams_t& rParams, \
bool accum, \
int thread_id, \
int num_threads);
CREATE_INSTANCE(true, QuantizationGranularity::TENSOR)
CREATE_INSTANCE(true, QuantizationGranularity::OUT_CHANNEL)
CREATE_INSTANCE(false, QuantizationGranularity::TENSOR)
CREATE_INSTANCE(false, QuantizationGranularity::OUT_CHANNEL)
#undef CREATE_INSTANCE
} // namespace internal
} // namespace fbgemm