-
Notifications
You must be signed in to change notification settings - Fork 688
/
CPULLVMIRGen.cpp
174 lines (154 loc) · 7.49 KB
/
CPULLVMIRGen.cpp
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
/**
* Copyright (c) Glow Contributors. See CONTRIBUTORS file.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "CPULLVMIRGen.h"
#include "glow/IR/Instrs.h"
#include "glow/LLVMIRCodeGen/LLVMBackend.h"
#include "glow/Quantization/Base/Base.h"
using namespace glow;
using llvm::cast;
CPULLVMIRGen::CPULLVMIRGen(const IRFunction *F,
AllocationsInfo &allocationsInfo,
std::string mainEntryName, llvm::StringRef libjitBC)
: LLVMIRGen(F, allocationsInfo, mainEntryName, libjitBC) {}
CPULLVMIRGen::CPULLVMIRGen(const IRFunction *F,
AllocationsInfo &allocationsInfo,
std::string mainEntryName, llvm::StringRef libjitBC,
llvm::ArrayRef<llvm::MemoryBufferRef> objectRegistry)
: LLVMIRGen(F, allocationsInfo, mainEntryName, libjitBC, objectRegistry) {}
void CPULLVMIRGen::generateLLVMIRForModule(llvm::IRBuilder<> &builder) {
// TODO: Add here any backend specific logic.
LLVMIRGen::generateLLVMIRForModule(builder);
}
void CPULLVMIRGen::generateLLVMIRForInstr(llvm::IRBuilder<> &builder,
const glow::Instruction *I) {
setCurrentDebugLocation(builder, I);
assert(!canBePartOfDataParallelKernel(I) &&
"data parallel instructions are not handled here");
// Perform any backend-specific code generation here and delegate everything
// else to LLVMIRGen.
switch (I->getKind()) {
case Kinded::Kind::CPUConvDKKC8InstKind: {
auto *CI = cast<CPUConvDKKC8Inst>(I);
auto *dest = CI->getDest();
auto *src = CI->getSrc();
auto *filter = CI->getFilter();
auto *bias = CI->getBias();
auto *destPtr = emitValueAddress(builder, dest);
auto *srcPtr = emitValueAddress(builder, src);
auto *filterPtr = emitValueAddress(builder, filter);
auto *biasPtr = emitValueAddress(builder, bias);
auto *destDims = emitValueDims(builder, dest);
auto *srcDims = emitValueDims(builder, src);
auto *filterDims = emitValueDims(builder, filter);
auto *biasDims = emitValueDims(builder, bias);
auto *kernels = emitConstDimTArray(builder, CI->getKernels());
auto *strides = emitConstDimTArray(builder, CI->getStrides());
auto *pads = emitConstDimTArray(builder, CI->getPads());
auto *group = emitConstDimT(builder, CI->getGroup());
size_t inChannels = src->dims()[3];
size_t outChannels = dest->dims()[3];
// Select a method for iterating on the image in the pixel (filter-first, or
// input-first). Perform convolutions with a high channel count by scanning
// the input image multiple times, once for each filter entry. Scan images
// with a low channel count by scanning the image once because the filter
// scan will fall in the cache.
bool pixelScanFirst = (inChannels < 16);
// The number of float8 registers that we use to process the depth channel.
unsigned numDepthRegs = (pixelScanFirst ? 8 : 2);
// The number of y pixels to process at once.
unsigned sizeGroupY = (pixelScanFirst ? 1 : 5);
// When producing output pixels process this many times of depth-strips,
// where each chunk is float8 * numDepthRegs. This is a form of tiling. It's
// profitable to scan multiple depth-strips of the filter if the scanned
// memory fits in the cahce and does not get evicted before the next
// iteration. By increasing the number strips (and using more cache memory)
// we reduce the number of times that we iterate over the input. However, we
// also increase the pressure on the cache that has to store the filter so
// we can't process too many strips at once.
unsigned depthStrips = 1;
unsigned stripSize = 8 * numDepthRegs * inChannels;
unsigned tileSize = 16384;
// Increase the number of strips until we reach the output-tensor depth size
// or until we exceed some threashold.
while (2 * depthStrips * stripSize <= tileSize &&
2 * depthStrips * numDepthRegs * 8 <= outChannels / CI->getGroup() &&
depthStrips < 8) {
depthStrips *= 2;
}
auto *pixelScanFirstVal = emitConstI32(builder, pixelScanFirst);
auto *numDepthRegsVal = emitConstI32(builder, numDepthRegs);
auto *sizeGroupYVal = emitConstI32(builder, sizeGroupY);
auto *depthStripsVal = emitConstI32(builder, depthStrips);
const char *kernelName = "convDKKC8";
auto *F = getFunction(kernelName, dest->getElementType());
createCall(builder, F,
{destPtr, srcPtr, filterPtr, biasPtr, destDims, srcDims,
filterDims, biasDims, kernels, strides, pads, group,
pixelScanFirstVal, numDepthRegsVal, sizeGroupYVal,
depthStripsVal});
break;
}
default:
LLVMIRGen::generateLLVMIRForInstr(builder, I);
}
}
void CPULLVMIRGen::generateLLVMIRForDataParallelInstr(
llvm::IRBuilder<> &builder, const glow::Instruction *I,
llvm::Function *kernel, llvm::DenseMap<Value *, int> &bufferToArgNum,
llvm::Value *loopCount) {
setCurrentDebugLocation(builder, I);
assert(canBePartOfDataParallelKernel(I) &&
"Expected a data parallel instruction");
// Perform any backend-specific code generation here and delegate everything
// else to LLVMIRGen.
switch (I->getKind()) {
case Kinded::Kind::CPUMaxSplatInstKind: {
auto *AN = cast<CPUMaxSplatInst>(I);
auto *dest = AN->getDest();
auto V = AN->getSplatValue();
auto *destPtr = emitBufferAddress(builder, dest, kernel, bufferToArgNum);
auto *lhs = AN->getSrc();
auto *lhsPtr = emitBufferAddress(builder, lhs, kernel, bufferToArgNum);
auto *F = getFunction("element_maxsplat_kernel", dest->getElementType());
auto *elementTy = getElementType(builder, dest);
auto *pointerNull =
llvm::ConstantPointerNull::get(elementTy->getPointerTo());
if (lhs->getType()->isQuantizedType()) {
// Quantize value from the splat to the {S,O} of the lhs param.
TensorQuantizationParams TQP{lhs->getType()->getScale(),
lhs->getType()->getOffset()};
auto quantizedValue = quantization::quantize(V, TQP);
auto *val = emitConst(builder, quantizedValue, lhs->getElementType());
auto *stackedOpCall = createUncheckedCall(
builder, F, {loopCount, val, lhsPtr, pointerNull});
auto *destAddr = builder.CreateGEP(builder.getInt8Ty(), destPtr,
loopCount, "buffer.element.addr");
builder.CreateStore(stackedOpCall, destAddr);
} else {
auto *val = emitConst(builder, V, lhs->getElementType());
auto *stackedOpCall = createUncheckedCall(
builder, F, {loopCount, val, lhsPtr, pointerNull});
auto *destAddr = builder.CreateGEP(builder.getFloatTy(), destPtr,
loopCount, "buffer.element.addr");
builder.CreateStore(stackedOpCall, destAddr);
}
break;
}
default:
LLVMIRGen::generateLLVMIRForDataParallelInstr(builder, I, kernel,
bufferToArgNum, loopCount);
}
}