forked from pytorch/FBGEMM
-
Notifications
You must be signed in to change notification settings - Fork 0
/
GenerateKernelU8S8S32ACC16.cc
362 lines (311 loc) · 10.1 KB
/
GenerateKernelU8S8S32ACC16.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
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
/*
* 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 <iostream>
#include "GenerateKernel.h"
namespace fbgemm {
template <typename TA, typename TB, typename TC, typename accT>
thread_local asmjit::JitRuntime CodeGenBase<TA, TB, TC, accT>::rt_;
template <typename TA, typename TB, typename TC, typename accT>
thread_local asmjit::CodeHolder CodeGenBase<TA, TB, TC, accT>::code_;
template <typename TA, typename TB, typename TC, typename accT>
thread_local std::map<
std::tuple<bool, int, int, int, int, int, int, int>,
typename CodeGenBase<TA, TB, TC, accT>::jit_micro_kernel_fp>
CodeGenBase<TA, TB, TC, accT>::codeCache_;
namespace x86 = asmjit::x86;
/**
* Generate AVX2 instructions for initializing the C registers to 0 in 16-bit
* Accumulation kernel.
*/
template <>
template <>
void CodeGenBase<uint8_t, int8_t, int32_t, int16_t>::initCRegs<
inst_set_t::avx2>(
x86::Emitter* a,
int rowRegs,
int colRegs,
int leadingDimCReg) {
for (int i = 0; i < rowRegs; ++i) {
for (int j = 0; j < colRegs; ++j) {
a->vxorps(
CRegs_avx2_[i * leadingDimCReg + j],
CRegs_avx2_[i * leadingDimCReg + j],
CRegs_avx2_[i * leadingDimCReg + j]);
}
}
}
/**
* Generate AVX2 instructions for computing block in the rank-k update of 16-bit
* Accmulation kernel.
*/
template <>
template <>
void CodeGenBase<uint8_t, int8_t, int32_t, int16_t>::genComputeBlock<
inst_set_t::avx2>(
x86::Emitter* a,
x86::Gp buffer_A,
x86::Gp buffer_B,
x86::Gp /* unused (reserved for prefetching)*/,
int rowRegs,
int colRegs,
int lda,
int leadingDimCReg) {
// used for matrix A
x86::Ymm AReg = x86::ymm12;
x86::Ymm tmpReg = x86::ymm14;
for (int i = 0; i < rowRegs; ++i) {
// broadcast A
a->vpbroadcastw(
AReg, x86::dword_ptr(buffer_A, (i * lda) * sizeof(uint8_t)));
for (int j = 0; j < colRegs; ++j) {
a->vpmaddubsw(
tmpReg, AReg, x86::dword_ptr(buffer_B, j * VLEN_ * sizeof(int8_t)));
a->vpaddsw(
CRegs_avx2_[i * leadingDimCReg + j],
tmpReg,
CRegs_avx2_[i * leadingDimCReg + j]);
// Prefetching is hurting performance in some cases
// because prefetch instructions itself consumes a slot
// in pipeline issue thus slowing down the kernel.
// if((i == rowRegs - 1) && j % 2 == 0){
// a->prefetcht0(x86::dword_ptr(B_pf, j*VLEN_*sizeof(int8_t)));
//}
}
}
}
/**
* Generate AVX2 instructions for storing the C registers back to the memory in
* 16-bit Accumulation kernel.
*/
template <>
template <>
void CodeGenBase<uint8_t, int8_t, int32_t, int16_t>::storeCRegs<
inst_set_t::avx2>(
x86::Emitter* a,
int rowRegs,
int colRegs,
x86::Gp C_Offset,
x86::Gp ldcReg,
bool accum,
int leadingDimCReg) {
x86::Xmm extractDest128 = x86::xmm15;
x86::Ymm extractDest256 = x86::ymm15;
for (int i = 0; i < rowRegs; ++i) {
a->imul(C_Offset, ldcReg, static_cast<asmjit::Imm>(i * sizeof(int32_t)));
for (int j = 0; j < colRegs; ++j) {
for (int idx = 0; idx < 2; ++idx) {
a->vextracti128(
extractDest128, CRegs_avx2_[i * leadingDimCReg + j], idx);
a->vpmovsxwd(extractDest256, extractDest128);
x86::Mem destAddr = x86::dword_ptr(
a->zcx(), C_Offset, 0, (j * 2 + idx) * 8 * sizeof(int32_t));
if (accum) {
a->vpaddd(extractDest256, extractDest256, destAddr);
}
a->vmovups(destAddr, extractDest256);
}
}
}
}
/**
* Get or Create the AVX2 instructions for 16-bit Accumulation macro-kernel.
*
*/
template <>
template <>
CodeGenBase<uint8_t, int8_t, int32_t, int16_t>::jit_micro_kernel_fp
CodeGenBase<uint8_t, int8_t, int32_t, int16_t>::getOrCreate<inst_set_t::avx2>(
bool accum,
int32_t mc,
int32_t nc,
int32_t kc,
int32_t /* unused */) {
std::tuple<bool, int, int, int, int, int, int, int> kernelSig;
int kBlock;
int nBlock;
int mRegBlockSize;
int nRegBlockSize;
int nRegBlockSizeMin;
int row_interleave;
if (blocking_params) {
kBlock = blocking_params->KCB;
nBlock = blocking_params->NCB;
mRegBlockSize = blocking_params->MR;
nRegBlockSize = blocking_params->NR;
nRegBlockSizeMin = blocking_params->NR_MIN;
row_interleave = blocking_params->ROW_INTERLEAVE;
} else {
kBlock = PackingTraits<uint8_t, int16_t, inst_set_t::avx2>::KCB;
nBlock = PackingTraits<uint8_t, int16_t, inst_set_t::avx2>::NCB;
mRegBlockSize = PackingTraits<uint8_t, int16_t, inst_set_t::avx2>::MR;
nRegBlockSize = PackingTraits<uint8_t, int16_t, inst_set_t::avx2>::NR;
nRegBlockSizeMin =
PackingTraits<uint8_t, int16_t, inst_set_t::avx2>::NR_MIN;
row_interleave =
PackingTraits<uint8_t, int16_t, inst_set_t::avx2>::ROW_INTERLEAVE;
}
kernelSig = std::make_tuple(
accum,
mc,
nc,
nBlock,
kBlock,
mRegBlockSize,
nRegBlockSize,
nRegBlockSizeMin);
if (codeCache_.find(kernelSig) != codeCache_.end()) {
return codeCache_[kernelSig];
}
code_.reset(false);
code_.init(rt_.codeInfo());
x86::Assembler assembler(&code_);
x86::Emitter* a = assembler.as<x86::Emitter>();
#if defined(FBGEMM_LOG_CODE)
// generated code logging
FILE* codeLogfile = fopen(
getCodeLoggingFile<inst_set_t::avx2>(
accum,
mc,
nc,
nBlock,
kBlock,
mRegBlockSize,
nRegBlockSize,
nRegBlockSizeMin)
.c_str(),
"w");
asmjit::FileLogger* codeLogger = new asmjit::FileLogger(codeLogfile);
if (codeLogger) {
code_.setLogger(codeLogger);
}
#endif
int mRegBlocks = mc / mRegBlockSize;
int mRegBlocksRem = mc % mRegBlockSize;
assert(kc % row_interleave == 0 && "kc must be a multiple of row_interleave");
// assert((nc == nRegBlockSize) &&
//"nc must be equal to the number of register blocks");
// arguments to the function created
x86::Gp buffer_A = a->zdi();
x86::Gp buffer_B = a->zsi();
x86::Gp B_pf = a->zdx();
x86::Gp CBase = a->zcx();
x86::Gp kSize = a->gpz(8);
x86::Gp ldcReg = a->gpz(9);
asmjit::FuncDetail func;
func.init(
asmjit::
FuncSignatureT<void, uint8_t*, int8_t*, int8_t*, int32_t*, int, int>(
asmjit::CallConv::kIdHost));
asmjit::FuncFrame frame;
frame.init(func);
frame.setDirtyRegs(
x86::Reg::kGroupVec,
asmjit::Support::bitMask(0, 1, 2, 3, 4, 5, 6, 7) |
asmjit::Support::bitMask(8, 9, 10, 11, 12, 13, 14, 15));
frame.setDirtyRegs(
x86::Reg::kGroupGp, asmjit::Support::bitMask(8, 9, 10, 11, 12, 13, 14));
asmjit::FuncArgsAssignment args(&func);
args.assignAll(buffer_A, buffer_B, B_pf, CBase, kSize, ldcReg);
args.updateFuncFrame(frame);
frame.finalize();
a->emitProlog(frame);
a->emitArgsAssignment(frame, args);
asmjit::Label Loopk = a->newLabel();
asmjit::Label LoopMBlocks = a->newLabel();
x86::Gp buffer_B_saved = a->gpz(10);
x86::Gp C_Offset = a->gpz(11);
// x86::Gp B_pf_saved = a->gpz(12);
x86::Gp iIdx = a->gpz(13);
x86::Gp kIdx = a->gpz(14);
int colRegs = nc * row_interleave * sizeof(int8_t) / VLEN_;
if (mRegBlocks > 0) {
// move 0 to iteration variables
a->mov(iIdx, 0);
// save B_buffer address
a->mov(buffer_B_saved, buffer_B);
// a->mov(B_pf_saved, B_pf);
a->bind(LoopMBlocks);
a->inc(iIdx);
int rowRegs = mRegBlockSize;
// init C registers
initCRegs<inst_set_t::avx2>(a, rowRegs, colRegs);
// init k loop index
a->mov(kIdx, 0);
a->bind(Loopk);
// k is incremented by row_interleave
a->add(kIdx, static_cast<asmjit::Imm>(row_interleave));
genComputeBlock<inst_set_t::avx2>(
a, buffer_A, buffer_B, B_pf, rowRegs, colRegs, kBlock);
// update buffer_A address for next k iteration
a->add(
buffer_A, static_cast<asmjit::Imm>(row_interleave * sizeof(uint8_t)));
// update buffer_B address for next k iteration
a->add(
buffer_B,
static_cast<asmjit::Imm>(nBlock * row_interleave * sizeof(int8_t)));
// a->add(B_pf, static_cast<asmjit::Imm>(nBlock * row_interleave *
// sizeof(int8_t)));
a->cmp(kIdx, kSize);
a->jl(Loopk);
// store C matrix
storeCRegs<inst_set_t::avx2>(a, rowRegs, colRegs, C_Offset, ldcReg, accum);
// increment A for next block
a->sub(buffer_A, kSize);
a->add(
buffer_A, static_cast<asmjit::Imm>((rowRegs)*kBlock * sizeof(uint8_t)));
// increment C for next block
a->imul(
C_Offset, ldcReg, static_cast<asmjit::Imm>(rowRegs * sizeof(int32_t)));
a->add(CBase, C_Offset);
// reset B
a->mov(buffer_B, buffer_B_saved);
// a->mov(B_pf, B_pf_saved);
a->cmp(iIdx, mRegBlocks);
a->jl(LoopMBlocks);
}
// generate code for remainder
if (mRegBlocksRem > 0) {
asmjit::Label LoopkRem = a->newLabel();
int rowRegs = mRegBlocksRem;
// init C registers
initCRegs<inst_set_t::avx2>(a, rowRegs, colRegs);
// init k loop index
a->mov(kIdx, 0);
a->bind(LoopkRem);
// k is incremented by row_interleave
a->add(kIdx, static_cast<asmjit::Imm>(row_interleave));
genComputeBlock<inst_set_t::avx2>(
a, buffer_A, buffer_B, B_pf, rowRegs, colRegs, kBlock);
// update buffer_A address for next k iteration
a->add(
buffer_A, static_cast<asmjit::Imm>(row_interleave * sizeof(uint8_t)));
// update buffer_B address for next k iteration
a->add(
buffer_B,
static_cast<asmjit::Imm>(nBlock * row_interleave * sizeof(int8_t)));
// a->add(B_pf, static_cast<asmjit::Imm>(nBlock * row_interleave *
// sizeof(int8_t)));
a->cmp(kIdx, kSize);
a->jl(LoopkRem);
// store C matrix
storeCRegs<inst_set_t::avx2>(a, rowRegs, colRegs, C_Offset, ldcReg, accum);
}
a->emitEpilog(frame);
jit_micro_kernel_fp fn;
asmjit::Error err = rt_.add(&fn, &code_);
if (err) {
std::cout << "Error: in fn add" << std::endl;
return nullptr;
}
codeCache_[kernelSig] = fn;
#if defined(FBGEMM_LOG_CODE)
fclose(codeLogfile);
delete codeLogger;
#endif
return fn;
}
} // namespace fbgemm