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
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
|
/*
* 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 "ExecuteKernelU8S8.h"
#include <cpuinfo.h>
#include <chrono>
#ifdef FBGEMM_MEASURE_TIME_BREAKDOWN
double kernel_time = 0.0;
double postprocessing_time = 0.0;
#endif
namespace fbgemm {
template <typename packingAMatrix, typename cT, typename processOutputType>
ExecuteKernel<
packingAMatrix,
PackBMatrix<int8_t, typename packingAMatrix::accType>,
cT,
processOutputType>::
ExecuteKernel(
PackMatrix<packingAMatrix, uint8_t, typename packingAMatrix::accType>&
packA,
PackMatrix<
PackBMatrix<int8_t, typename packingAMatrix::accType>,
int8_t,
typename packingAMatrix::accType>& packB,
int32_t kBlock,
cT* matC,
int32_t* C_buffer,
int32_t ldc,
const processOutputType& outputProcess,
int thread_id,
int num_threads)
: packedA_(packA),
packedB_(packB),
kBlock_(kBlock),
matC_(matC),
C_buffer_(C_buffer),
ldc_(ldc),
outputProcess_(outputProcess),
thread_id_(thread_id),
num_threads_(num_threads) {
if (cpuinfo_has_x86_avx512f()) {
mbSize_ = PackingTraits<
int8_t,
typename packingAMatrix::accType,
inst_set_t::avx512>::MCB;
nbSize_ = PackingTraits<
int8_t,
typename packingAMatrix::accType,
inst_set_t::avx512>::NCB;
} else if (cpuinfo_has_x86_avx2()) {
mbSize_ = PackingTraits<
int8_t,
typename packingAMatrix::accType,
inst_set_t::avx2>::MCB;
nbSize_ = PackingTraits<
int8_t,
typename packingAMatrix::accType,
inst_set_t::avx2>::NCB;
} else {
assert(0 && "unsupported architecure");
}
C_tile_ = new int32_t[mbSize_ * nbSize_];
}
template <typename packingAMatrix, typename cT, typename processOutputType>
void ExecuteKernel<
packingAMatrix,
PackBMatrix<int8_t, typename packingAMatrix::accType>,
cT,
processOutputType>::execute(int kBlock) {
// packedA_.printPackedMatrix("packedA from kernel");
// packedB_.printPackedMatrix("packedB from kernel");
int32_t bColBlocks = packedB_.blockCols();
int8_t* bBuf;
int8_t* bBuf_pf;
uint8_t* aBuf = packedA_.getBuf(0);
int32_t packed_rows_A = packedA_.numPackedRows();
int32_t row_start_A = packedA_.packedRowStart();
int group = kBlock / packedB_.blockRows();
int NDim = packedB_.numCols();
bool lastKBlock = packedB_.isThisLastKBlock(kBlock % packedB_.blockRows());
bool accum = (kBlock % packedB_.blockRows()) > 0;
typename BaseType::jit_micro_kernel_fp fn;
if (cpuinfo_initialize()) {
if (cpuinfo_has_x86_avx512f()) {
fn = BaseType::template getOrCreate<inst_set_t::avx512>(
accum,
packed_rows_A,
packedB_.blockColSize(),
packedA_.numPackedCols(),
nbSize_);
} else if (cpuinfo_has_x86_avx2()) {
fn = BaseType::template getOrCreate<inst_set_t::avx2>(
accum,
packed_rows_A,
packedB_.blockColSize(),
packedA_.numPackedCols(),
nbSize_);
} else {
// TODO: Have default slower path
assert(0 && "unsupported architecture");
return;
}
} else {
throw std::runtime_error("Failed to initialize cpuinfo!");
}
#ifdef FBGEMM_MEASURE_TIME_BREAKDOWN
std::chrono::time_point<std::chrono::high_resolution_clock> t_start, t_end;
double dt;
t_start = std::chrono::high_resolution_clock::now();
#endif
for (int jb = 0; jb < bColBlocks; ++jb) {
bBuf = packedB_.getBuf(jb, kBlock);
// prefetch addr of the next packed block of B matrix
bBuf_pf = packedB_.getBuf(jb == bColBlocks - 1 ? jb : jb + 1, kBlock);
// If the accumulation buffer C_buffer_ is the same as matC_ (inplace output
// processing), then each thread use the different parts of output buffer
// matC_;
// Otherwise, each thread uses different portions of the accumulation
// buffer C_buffer_. If m is large enough (m >= nthreads * MC), then we only
// need to use (nthreads * MC) x n portion of C_buffer_, each thread access
// the C_buffer_row_start as tid * MC * ldc_; else when m is very small, we
// juse use the whole m x n C_buffer_: each thread use the different
// portion.
int32_t* C_buffer_row_start = C_buffer_ +
((C_buffer_ == reinterpret_cast<int32_t*>(matC_) ||
num_threads_ * mbSize_ > packedA_.numRows())
? row_start_A * ldc_ + NDim * group
: thread_id_ * mbSize_ * ldc_ + NDim * group);
int32_t* C_buffer_start = C_buffer_row_start + jb * nbSize_;
int32_t leadingDim = ldc_;
if (packedB_.isThereColRemainder() && (jb == bColBlocks - 1)) {
// In case we will access memory past C_buffer_, we use C_tile_ scratchpad
// instead.
C_buffer_start = C_tile_;
leadingDim = nbSize_;
}
fn(aBuf,
bBuf,
bBuf_pf,
C_buffer_start,
packedA_.numPackedCols(),
leadingDim);
#ifdef FBGEMM_MEASURE_TIME_BREAKDOWN
t_end = std::chrono::high_resolution_clock::now();
dt = std::chrono::duration_cast<std::chrono::nanoseconds>(t_end - t_start)
.count();
kernel_time += (dt);
t_start = std::chrono::high_resolution_clock::now();
#endif
// Output processing is done only once per rowblock to amortize overhead
// and for better spatial locality.
if (lastKBlock && jb == bColBlocks - 1) {
// When C_tile_ is used for the last column block, we need a separate
// handling for the last column block.
int32_t nSize =
C_buffer_start == C_tile_ ? jb * nbSize_ : packedB_.numCols();
if (nSize) {
if (cpuinfo_has_x86_avx512f()) {
// TODO: avx512 path
// Currently use avx2 code
outputProcess_.template f<inst_set_t::avx2>(
matC_,
C_buffer_row_start,
{row_start_A, packed_rows_A, NDim * group, nSize},
ldc_,
ldc_);
} else if (cpuinfo_has_x86_avx2()) {
outputProcess_.template f<inst_set_t::avx2>(
matC_,
C_buffer_row_start,
{row_start_A, packed_rows_A, NDim * group, nSize},
ldc_,
ldc_);
} else {
// TODO: Have default slower path
assert(0 && "unsupported architecure");
}
}
if (C_buffer_start == C_tile_) {
// When C_tile_ scratchpad was used to avoid accessing memory past
// C_buffer_ .
if (cpuinfo_has_x86_avx512f()) {
// TODO: avx512 path
// Currently use avx2 code
outputProcess_.template f<inst_set_t::avx2>(
matC_,
C_tile_,
{row_start_A,
packed_rows_A,
NDim * group + jb * nbSize_,
packedB_.lastBcol()},
ldc_,
leadingDim);
} else if (cpuinfo_has_x86_avx2()) {
outputProcess_.template f<inst_set_t::avx2>(
matC_,
C_tile_,
{row_start_A,
packed_rows_A,
NDim * group + jb * nbSize_,
packedB_.lastBcol()},
ldc_,
leadingDim);
} else {
// TODO: Have default slower path
assert(0 && "unsupported architecure");
}
}
} // output processing
#ifdef FBGEMM_MEASURE_TIME_BREAKDOWN
t_end = std::chrono::high_resolution_clock::now();
dt = std::chrono::duration_cast<std::chrono::nanoseconds>(t_end - t_start)
.count();
postprocessing_time += (dt);
t_start = std::chrono::high_resolution_clock::now();
#endif
} // for each j block
}
////////////////////////////////////////////////////////////////////////////////
// ReQuantizeOutput
#define INSTANTIATE_BASE(PACK_A, ACC_T, RELU, Q_GRAN) \
template class ExecuteKernel< \
PACK_A<uint8_t, ACC_T>, \
PackBMatrix<int8_t, ACC_T>, \
uint8_t, \
ReQuantizeOutput<RELU, Q_GRAN>>;
#define INSTANTIATE_Q_GRANS(PACK_A, ACC_T, RELU) \
INSTANTIATE_BASE(PACK_A, ACC_T, RELU, QuantizationGranularity::TENSOR); \
INSTANTIATE_BASE(PACK_A, ACC_T, RELU, QuantizationGranularity::GROUP); \
INSTANTIATE_BASE(PACK_A, ACC_T, RELU, QuantizationGranularity::OUT_CHANNEL);
#define INSTANTIATE_RELU(PACK_A, ACC_T) \
INSTANTIATE_Q_GRANS(PACK_A, ACC_T, false); \
INSTANTIATE_Q_GRANS(PACK_A, ACC_T, true);
#define INSTANTIATE_ACC_T(PACK_A) \
INSTANTIATE_RELU(PACK_A, int32_t); \
INSTANTIATE_RELU(PACK_A, int16_t);
INSTANTIATE_ACC_T(PackAMatrix);
INSTANTIATE_ACC_T(PackAWithRowOffset);
#undef INSTANTIATE_ACC_T
#undef INSTANTIATE_RELU
#undef INSTANTIATE_Q_GRANS
#undef INSTANTIATE_BASE
#define INSTANTIATE_BASE(ACC_T, RELU, SPATIAL_DIM, Q_GRAN) \
template class ExecuteKernel< \
PackAWithIm2Col<uint8_t, ACC_T, SPATIAL_DIM>, \
PackBMatrix<int8_t, ACC_T>, \
uint8_t, \
ReQuantizeOutput<RELU, Q_GRAN>>;
#define INSTANTIATE_Q_GRANS(ACC_T, RELU, SPATIAL_DIM) \
INSTANTIATE_BASE(ACC_T, RELU, SPATIAL_DIM, QuantizationGranularity::TENSOR); \
INSTANTIATE_BASE(ACC_T, RELU, SPATIAL_DIM, QuantizationGranularity::GROUP); \
INSTANTIATE_BASE( \
ACC_T, RELU, SPATIAL_DIM, QuantizationGranularity::OUT_CHANNEL);
#define INSTANTIATE_SPATIAL_DIM(ACC_T, RELU) \
INSTANTIATE_Q_GRANS(ACC_T, RELU, 2); \
INSTANTIATE_Q_GRANS(ACC_T, RELU, 3);
#define INSTANTIATE_RELU(ACC_T) \
INSTANTIATE_SPATIAL_DIM(ACC_T, false); \
INSTANTIATE_SPATIAL_DIM(ACC_T, true);
INSTANTIATE_RELU(int32_t);
INSTANTIATE_RELU(int16_t);
#undef INSTANTIATE_RELU
#undef INSTANTIATE_SPATIAL_DIM
#undef INSTANTIATE_Q_GRANS
#undef INSTANTIATE_BASE
////////////////////////////////////////////////////////////////////////////////
// ReQuantizeForFloat
#define INSTANTIATE_BASE(PACK_A, RELU, Q_GRAN) \
template class ExecuteKernel< \
PACK_A<uint8_t, int32_t>, \
PackBMatrix<int8_t, int32_t>, \
float, \
ReQuantizeForFloat<RELU, Q_GRAN>>;
#define INSTANTIATE_Q_GRANS(PACK_A, RELU) \
INSTANTIATE_BASE(PACK_A, RELU, QuantizationGranularity::TENSOR); \
INSTANTIATE_BASE(PACK_A, RELU, QuantizationGranularity::GROUP); \
INSTANTIATE_BASE(PACK_A, RELU, QuantizationGranularity::OUT_CHANNEL);
#define INSTANTIATE_RELU(PACK_A) \
INSTANTIATE_Q_GRANS(PACK_A, false); \
INSTANTIATE_Q_GRANS(PACK_A, true);
INSTANTIATE_RELU(PackAWithRowOffset);
INSTANTIATE_RELU(PackAWithQuantRowOffset);
#undef INSTANTIATE_RELU
#undef INSTANTIATE_Q_GRANS
#undef INSTANTIATE_BASE
#define INSTANTIATE_BASE(ACC_T, RELU, SPATIAL_DIM, Q_GRAN) \
template class ExecuteKernel< \
PackAWithIm2Col<uint8_t, ACC_T, SPATIAL_DIM>, \
PackBMatrix<int8_t, ACC_T>, \
float, \
ReQuantizeForFloat<RELU, Q_GRAN>>;
#define INSTANTIATE_Q_GRANS(ACC_T, RELU, SPATIAL_DIM) \
INSTANTIATE_BASE(ACC_T, RELU, SPATIAL_DIM, QuantizationGranularity::TENSOR); \
INSTANTIATE_BASE(ACC_T, RELU, SPATIAL_DIM, QuantizationGranularity::GROUP); \
INSTANTIATE_BASE( \
ACC_T, RELU, SPATIAL_DIM, QuantizationGranularity::OUT_CHANNEL);
#define INSTANTIATE_SPATIAL_DIM(ACC_T, RELU) \
INSTANTIATE_Q_GRANS(ACC_T, RELU, 2); \
INSTANTIATE_Q_GRANS(ACC_T, RELU, 3);
#define INSTANTIATE_RELU(ACC_T) \
INSTANTIATE_SPATIAL_DIM(ACC_T, false); \
INSTANTIATE_SPATIAL_DIM(ACC_T, true);
INSTANTIATE_RELU(int32_t);
INSTANTIATE_RELU(int16_t);
#undef INSTANTIATE_RELU
#undef INSTANTIATE_SPATIAL_DIM
#undef INSTANTIATE_Q_GRANS
#undef INSTANTIATE_BASE
template class ExecuteKernel<
PackAWithRowOffset<uint8_t, int16_t>,
PackBMatrix<int8_t, int16_t>,
float,
ReQuantizeForFloat<false /* FUSE_RELU*/>>;
////////////////////////////////////////////////////////////////////////////////
// DoSpmdmOnInpBuffer
#define INSTANTIATE_BASE(RELU, Q_GRAN) \
template class ExecuteKernel< \
PackAWithRowOffset<uint8_t, int16_t>, \
PackBMatrix<int8_t, int16_t>, \
uint8_t, \
DoSpmdmOnInpBuffer<uint8_t, int32_t, ReQuantizeOutput<RELU, Q_GRAN>>>;
#define INSTANTIATE_Q_GRANS(RELU) \
INSTANTIATE_BASE(RELU, QuantizationGranularity::TENSOR); \
INSTANTIATE_BASE(RELU, QuantizationGranularity::GROUP); \
INSTANTIATE_BASE(RELU, QuantizationGranularity::OUT_CHANNEL);
INSTANTIATE_Q_GRANS(false);
INSTANTIATE_Q_GRANS(true);
#undef INSTANTIATE_Q_GRANS
#undef INSTANTIATE_BASE
#define INSTANTIATE_BASE(RELU, Q_GRAN) \
template class ExecuteKernel< \
PackAWithIm2Col<uint8_t, int16_t>, \
PackBMatrix<int8_t, int16_t>, \
uint8_t, \
DoSConvOnInpBuffer<uint8_t, int32_t, ReQuantizeOutput<RELU, Q_GRAN>>>;
#define INSTANTIATE_Q_GRANS(RELU) \
INSTANTIATE_BASE(RELU, QuantizationGranularity::TENSOR); \
INSTANTIATE_BASE(RELU, QuantizationGranularity::GROUP); \
INSTANTIATE_BASE(RELU, QuantizationGranularity::OUT_CHANNEL);
INSTANTIATE_Q_GRANS(false);
INSTANTIATE_Q_GRANS(true);
#undef INSTANTIATE_Q_GRANS
#undef INSTANTIATE_BASE
template class ExecuteKernel<
PackAWithRowOffset<uint8_t, int16_t>,
PackBMatrix<int8_t, int16_t>,
float,
DoSpmdmOnInpBuffer<float, int32_t, ReQuantizeForFloat<false>>>;
////////////////////////////////////////////////////////////////////////////////
// memCopy
#define INSTANTIATE_BASE(PACK_A, ACC_T) \
template class ExecuteKernel< \
PACK_A<uint8_t, ACC_T>, \
PackBMatrix<int8_t, ACC_T>, \
int32_t, \
memCopy<>>;
#define INSTANTIATE_ACC_T(PACK_A) \
INSTANTIATE_BASE(PACK_A, int32_t) \
INSTANTIATE_BASE(PACK_A, int16_t)
INSTANTIATE_ACC_T(PackAMatrix);
INSTANTIATE_ACC_T(PackAWithRowOffset);
#undef INSTANTIATE_ACC_T
#undef INSTANTIATE_BASE
#define INSTANTIATE_BASE(ACC_T, SPATIAL_DIM) \
template class ExecuteKernel< \
PackAWithIm2Col<uint8_t, ACC_T, SPATIAL_DIM>, \
PackBMatrix<int8_t, ACC_T>, \
int32_t, \
memCopy<>>;
#define INSTANTIATE_SPATIAL_DIM(ACC_T) \
INSTANTIATE_BASE(ACC_T, 2); \
INSTANTIATE_BASE(ACC_T, 3);
INSTANTIATE_SPATIAL_DIM(int32_t);
INSTANTIATE_SPATIAL_DIM(int16_t);
#undef INSTANTIATE_SPATIAL_DIM
#undef INSTANTIATE_BASE
template class ExecuteKernel<
PackAWithQuantRowOffset<uint8_t, int32_t>,
PackBMatrix<int8_t, int32_t>,
int32_t,
memCopy<>>;
template class ExecuteKernel<
PackAMatrix<uint8_t, int16_t>,
PackBMatrix<int8_t, int16_t>,
int32_t,
DoNothing<int32_t, int32_t>>;
} // namespace fbgemm
|
