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
|
/*
* 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 "fbgemm/Fbgemm.h"
#include <cpuinfo.h>
#include <stdexcept>
#include "ExecuteKernel.h"
#ifdef FBGEMM_MEASURE_TIME_BREAKDOWN
double packing_time = 0.0;
double computing_time = 0.0;
double run_time = 0.0;
#endif
using namespace fbgemm2;
namespace fbgemm2 {
template <
typename packingAMatrix,
typename packingBMatrix,
typename cT,
typename processOutputType>
void fbgemmPacked(
PackMatrix<
packingAMatrix,
typename packingAMatrix::inpType,
typename packingAMatrix::accType>& packA,
PackMatrix<
packingBMatrix,
typename packingBMatrix::inpType,
typename packingBMatrix::accType>& packB,
cT* C,
int32_t* C_buffer,
uint32_t ldc,
const processOutputType& outProcess,
int thread_id,
int /* num_threads */) {
static_assert(
std::is_same<
typename packingAMatrix::accType,
typename packingBMatrix::accType>::value,
"Accumulation type of both matrices should be the same");
int MCB, KCB;
// Run time CPU detection
if (cpuinfo_initialize()) {
if (cpuinfo_has_x86_avx512f()) {
MCB = PackingTraits<
typename packingAMatrix::inpType,
typename packingAMatrix::accType,
inst_set_t::avx512>::MCB;
KCB = PackingTraits<
typename packingAMatrix::inpType,
typename packingAMatrix::accType,
inst_set_t::avx512>::KCB;
} else if (cpuinfo_has_x86_avx2()) {
MCB = PackingTraits<
typename packingAMatrix::inpType,
typename packingAMatrix::accType,
inst_set_t::avx2>::MCB;
KCB = PackingTraits<
typename packingAMatrix::inpType,
typename packingAMatrix::accType,
inst_set_t::avx2>::KCB;
} else {
// TODO: Have default slower path
assert(0 && "unsupported architecture");
return;
}
} else {
throw std::runtime_error("Failed to initialize cpuinfo!");
}
int MDim = packA.numRows();
int KDim = packB.numRows();
int mBlocks = (MDim + MCB - 1) / MCB;
int kBlocks = (KDim + KCB - 1) / KCB;
// remainders
int _mc = MDim % MCB;
int _kc = KDim % KCB;
int kc, mc;
block_type_t blockA{0, 0, 0, 0};
// B must be prepacked
assert(packB.isPrePacked() && "B matrix must be prepacked");
#ifdef FBGEMM_MEASURE_TIME_BREAKDOWN
std::chrono::time_point<std::chrono::high_resolution_clock> t_very_start,
t_start, t_end;
double dt;
t_start = std::chrono::high_resolution_clock::now();
t_very_start = std::chrono::high_resolution_clock::now();
#endif
ExecuteKernel<packingAMatrix, packingBMatrix, cT, processOutputType>
exeKernelObj(packA, packB, 0, C, C_buffer, ldc, outProcess);
// ToDo: thread based work division
for (int i = 0; i < mBlocks; ++i) {
mc = (i != mBlocks - 1 || _mc == 0) ? MCB : _mc;
for (int k = 0; k < kBlocks; ++k) {
kc = (k != kBlocks - 1 || _kc == 0) ? KCB : _kc;
// pack A matrix
blockA = {i * MCB, mc, k * KCB, kc};
packA.pack(blockA);
#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();
packing_time += (dt);
t_start = std::chrono::high_resolution_clock::now();
#endif
exeKernelObj.execute(k);
#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();
computing_time += (dt);
t_start = std::chrono::high_resolution_clock::now();
#endif
}
}
#ifdef FBGEMM_MEASURE_TIME_BREAKDOWN
t_end = std::chrono::high_resolution_clock::now();
dt =
std::chrono::duration_cast<std::chrono::nanoseconds>(t_end - t_very_start)
.count();
run_time += (dt);
t_start = std::chrono::high_resolution_clock::now();
#endif
}
template void fbgemmPacked(
PackMatrix<PackAWithRowOffset<uint8_t, int32_t>, uint8_t, int32_t>& packA,
PackMatrix<PackBMatrix<int8_t, int32_t>, int8_t, int32_t>& packB,
uint8_t* C,
int32_t* C_buffer,
uint32_t ldc,
const ReQuantizeOutput<false>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithRowOffset<uint8_t, int32_t>, uint8_t, int32_t>& packA,
PackMatrix<PackBMatrix<int8_t, int32_t>, int8_t, int32_t>& packB,
uint8_t* C,
int32_t* C_buffer,
uint32_t ldc,
const ReQuantizeOutput<true>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithQuantRowOffset<uint8_t, int32_t>, uint8_t, int32_t>&
packA,
PackMatrix<PackBMatrix<int8_t, int32_t>, int8_t, int32_t>& packB,
float* C,
int32_t* C_buffer,
uint32_t ldc,
const ReQuantizeForFloat<false>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithQuantRowOffset<uint8_t, int32_t>, uint8_t, int32_t>&
packA,
PackMatrix<PackBMatrix<int8_t, int32_t>, int8_t, int32_t>& packB,
float* C,
int32_t* C_buffer,
uint32_t ldc,
const ReQuantizeForFloat<true>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAMatrix<uint8_t, int32_t>, uint8_t, int32_t>& packA,
PackMatrix<PackBMatrix<int8_t, int32_t>, int8_t, int32_t>& packB,
int32_t* C,
int32_t* C_buffer,
uint32_t ldc,
const memCopy<>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithRowOffset<uint8_t, int32_t>, uint8_t, int32_t>& packA,
PackMatrix<PackBMatrix<int8_t, int32_t>, int8_t, int32_t>& packB,
float* C,
int32_t* C_buffer,
uint32_t ldc,
const ReQuantizeForFloat<false>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithRowOffset<uint8_t, int32_t>, uint8_t, int32_t>& packA,
PackMatrix<PackBMatrix<int8_t, int32_t>, int8_t, int32_t>& packB,
float* C,
int32_t* C_buffer,
uint32_t ldc,
const ReQuantizeForFloat<true>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithRowOffset<uint8_t, int32_t>, uint8_t, int32_t>& packA,
PackMatrix<PackBMatrix<int8_t, int32_t>, int8_t, int32_t>& packB,
int32_t* C,
int32_t* C_buffer,
uint32_t ldc,
const memCopy<>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithIm2Col<uint8_t, int32_t>, uint8_t, int32_t>& packA,
PackMatrix<PackBMatrix<int8_t, int32_t>, int8_t, int32_t>& packB,
int32_t* C,
int32_t* C_buffer,
uint32_t ldc,
const memCopy<>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithIm2Col<uint8_t, int32_t, 3>, uint8_t, int32_t>& packA,
PackMatrix<PackBMatrix<int8_t, int32_t>, int8_t, int32_t>& packB,
int32_t* C,
int32_t* C_buffer,
uint32_t ldc,
const memCopy<>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithQuantRowOffset<uint8_t, int32_t>, uint8_t, int32_t>&
packA,
PackMatrix<PackBMatrix<int8_t, int32_t>, int8_t, int32_t>& packB,
int32_t* C,
int32_t* C_buffer,
uint32_t ldc,
const memCopy<>& outProcess,
int thread_id,
int num_threads);
// 16 bit accumulation functions
template void fbgemmPacked(
PackMatrix<PackAMatrix<uint8_t, int16_t>, uint8_t, int16_t>& packA,
PackMatrix<PackBMatrix<int8_t, int16_t>, int8_t, int16_t>& packB,
int32_t* C,
int32_t* C_buffer,
uint32_t ldc,
const memCopy<>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAMatrix<uint8_t, int16_t>, uint8_t, int16_t>& packA,
PackMatrix<PackBMatrix<int8_t, int16_t>, int8_t, int16_t>& packB,
uint8_t* C,
int32_t* C_buffer,
uint32_t ldc,
const ReQuantizeOutput<false>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithRowOffset<uint8_t, int16_t>, uint8_t, int16_t>& packA,
PackMatrix<PackBMatrix<int8_t, int16_t>, int8_t, int16_t>& packB,
uint8_t* C,
int32_t* C_buffer,
uint32_t ldc,
const DoSpmdmOnInpBuffer<uint8_t, int32_t, ReQuantizeOutput<false>>&
outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithRowOffset<uint8_t, int16_t>, uint8_t, int16_t>& packA,
PackMatrix<PackBMatrix<int8_t, int16_t>, int8_t, int16_t>& packB,
uint8_t* C,
int32_t* C_buffer,
uint32_t ldc,
const DoSpmdmOnInpBuffer<uint8_t, int32_t, ReQuantizeOutput<true>>&
outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithRowOffset<uint8_t, int16_t>, uint8_t, int16_t>& packA,
PackMatrix<PackBMatrix<int8_t, int16_t>, int8_t, int16_t>& packB,
float* C,
int32_t* C_buffer,
uint32_t ldc,
const DoSpmdmOnInpBuffer<float, int32_t, ReQuantizeForFloat<false>>&
outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithRowOffset<uint8_t, int16_t>, uint8_t, int16_t>& packA,
PackMatrix<PackBMatrix<int8_t, int16_t>, int8_t, int16_t>& packB,
uint8_t* C,
int32_t* C_buffer,
uint32_t ldc,
const ReQuantizeOutput<false>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithRowOffset<uint8_t, int16_t>, uint8_t, int16_t>& packA,
PackMatrix<PackBMatrix<int8_t, int16_t>, int8_t, int16_t>& packB,
uint8_t* C,
int32_t* C_buffer,
uint32_t ldc,
const ReQuantizeOutput<true>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithRowOffset<uint8_t, int16_t>, uint8_t, int16_t>& packA,
PackMatrix<PackBMatrix<int8_t, int16_t>, int8_t, int16_t>& packB,
int32_t* C,
int32_t* C_buffer,
uint32_t ldc,
const memCopy<>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithIm2Col<uint8_t, int16_t>, uint8_t, int16_t>& packA,
PackMatrix<PackBMatrix<int8_t, int16_t>, int8_t, int16_t>& packB,
int32_t* C,
int32_t* C_buffer,
uint32_t ldc,
const memCopy<>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithIm2Col<uint8_t, int16_t, 3>, uint8_t, int16_t>& packA,
PackMatrix<PackBMatrix<int8_t, int16_t>, int8_t, int16_t>& packB,
int32_t* C,
int32_t* C_buffer,
uint32_t ldc,
const memCopy<>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAMatrix<uint8_t, int16_t>, uint8_t, int16_t>& packA,
PackMatrix<PackBMatrix<int8_t, int16_t>, int8_t, int16_t>& packB,
int32_t* C,
int32_t* C_buffer,
uint32_t ldc,
const DoNothing<int32_t, int32_t>& outProcess,
int thread_id,
int num_threads);
template void fbgemmPacked(
PackMatrix<PackAWithRowOffset<uint8_t, int16_t>, uint8_t, int16_t>& packA,
PackMatrix<PackBMatrix<int8_t, int16_t>, int8_t, int16_t>& packB,
float* C,
int32_t* C_buffer,
uint32_t ldc,
const ReQuantizeForFloat<false>& outProcess,
int thread_id,
int num_threads);
} // namespace fbgemm2
|
