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
|
#pragma once
#include <torch/csrc/jit/ir/ir.h>
#include <torch/csrc/jit/runtime/interpreter.h>
#include <torch/csrc/jit/tensorexpr/analysis.h>
#include <torch/csrc/jit/tensorexpr/codegen.h>
#include <torch/csrc/jit/tensorexpr/tensor.h>
namespace torch {
namespace jit {
namespace tensorexpr {
template <typename T>
inline std::vector<int64_t> bufferSizes(const T& t) {
std::vector<int64_t> sizes;
for (size_t i = 0; i < t->buf()->ndim(); i++) {
sizes.push_back(dynamic_cast<const IntImm*>(t->buf()->dim(i))->value());
}
return sizes;
}
class TORCH_API TensorExprKernel {
public:
explicit TensorExprKernel(const std::shared_ptr<Graph>& subgraph);
void run(Stack& stack);
void fallback(Stack& stack) {
InterpreterState(code_).run(stack);
}
Stmt* getCodeGenStmt();
std::string getCodeText() {
return codegen_->getCodeText();
}
private:
enum BackendType {
kUninitialized,
kSimpleIREval,
kLLVMCodeGen,
kCudaCodeGen,
kBlockCodeGen,
};
void compile();
void runKernel(Stack& stack);
std::vector<DimArg> dimsFromSizes(const std::vector<ExprHandle>& sizes);
std::vector<ExprHandle> sizesForValue(const torch::jit::Value* v);
std::vector<ExprHandle> inferSizesForValue(const torch::jit::Value* v);
std::vector<ExprHandle> sizesFromVaryingShape(
const c10::VaryingShape<int64_t>& shape);
std::vector<ExprHandle> broadcastShapes(
const std::vector<ExprHandle>& a,
const std::vector<ExprHandle>& b);
std::vector<ExprHandle> broadcastShapes(
std::vector<std::vector<ExprHandle>> shapes);
ExprHandle constant(const torch::jit::Value* v);
ExprHandle broadcast(Tensor* t, const std::vector<ExprHandle>& axes);
ExprHandle chunk(
Tensor* t,
size_t chunkIdx,
int64_t dim,
int64_t chunks,
const std::vector<ExprHandle>& axes);
std::vector<ExprHandle> valueShape(const torch::jit::Value* v);
void promoteInputs(std::vector<ExprHandle>& inputs);
ExprHandle demoteOutput(const ExprHandle& e, const torch::jit::Value* v);
ExprHandle tensorOrConstant(
const torch::jit::Value* v,
const std::vector<ExprHandle>& axes);
Tensor* computeOneOperand(
const std::string& name,
const torch::jit::Value* v,
const std::function<ExprHandle(const ExprHandle&)>& innerExpr);
Tensor* computeTwoOperand(
const std::string& name,
const torch::jit::Value* v,
const std::function<ExprHandle(const ExprHandle&, const ExprHandle&)>&
innerExpr);
Tensor* computeTwoOperandWithAlpha(
const std::string& name,
const torch::jit::Value* v,
const std::function<ExprHandle(const ExprHandle&, const ExprHandle&)>&
innerExpr);
Tensor* computeThreeOperand(
const std::string& name,
const torch::jit::Value* v,
const std::function<
ExprHandle(const ExprHandle&, const ExprHandle&, const ExprHandle&)>&
innerExpr);
Tensor* computeConditionWithTwoOperand(
const std::string& name,
const torch::jit::Value* v,
const std::function<
ExprHandle(const ExprHandle&, const ExprHandle&, const ExprHandle&)>&
innerExpr);
Tensor* computeFourOperand(
const std::string& name,
const torch::jit::Value* v,
const std::function<ExprHandle(
const ExprHandle&,
const ExprHandle&,
const ExprHandle&,
const ExprHandle&)>& innerExpr);
Tensor* computeSum(const torch::jit::Value* v);
Tensor* computeValue(const torch::jit::Value* v);
void flattenTensors(BackendType backendType);
Stmt* generateStmt(BackendType backendType);
std::vector<CodeGen::BufferArg> prepareBufferArgs();
std::string getCodeGenName(BackendType backendType);
std::vector<CodeGen::CallArg> prepareRunArgs(
const at::ArrayRef<IValue>& inputs,
std::vector<at::Tensor>& outputs);
BackendType inferBackendTypeFromDevice(at::Device device);
void bindInput(const torch::jit::Value* input);
// Captures the information for reduction operation nodes.
struct ReductionInfo {
std::vector<DimArg> reductionDims;
std::vector<DimArg> outputDims;
std::vector<size_t> axes;
bool keepdim;
c10::optional<Dtype> dtype;
};
// Get the reduction info for the given node, based on properties and inputs.
ReductionInfo getReductionInfo(const torch::jit::Node* node);
// Get the reduction axes for the given node, based on properties and inputs.
std::vector<int64_t> getReductionAxes(const torch::jit::Node* node);
private:
struct ShapeArg {
size_t idx;
VarHandle var;
ShapeArg(size_t i, VarHandle v) : idx(i), var(v) {}
};
struct KernelArg {
template <typename B>
KernelArg(B&& b) : bufferArg_(std::forward<B>(b)) {}
template <typename B, typename T>
KernelArg(B&& b, T&& sizes, T&& strides)
: bufferArg_(b),
sizeArgs_(std::forward<T>(sizes)),
strideArgs_(std::forward<T>(strides)) {}
const CodeGen::BufferArg& buffer() const {
return bufferArg_;
}
const std::vector<ShapeArg>& sizes() const {
return sizeArgs_;
}
const std::vector<ShapeArg>& strides() const {
return strideArgs_;
}
CodeGen::BufferArg bufferArg_;
std::vector<ShapeArg> sizeArgs_;
std::vector<ShapeArg> strideArgs_;
};
int64_t nInputs_ = 0;
std::vector<KernelArg> kernelArgs_;
std::vector<Tensor*> tensorOutputs_;
std::vector<Tensor*> flatTensorOutputs_;
std::unordered_map<int64_t, Tensor*> tensors_;
std::unordered_map<int64_t, VarHandle> scalars_;
std::unique_ptr<CodeGen> codegen_;
at::Device device_ = at::kCPU;
KernelArena kernelArena_;
std::vector<TypePtr> inputTypes_;
std::shared_ptr<Graph> graph_;
Code code_;
bool fallback_{false};
bool hasRandom_{false};
bool hasBroadcast_{false};
std::unordered_map<const torch::jit::Value*, std::vector<ExprHandle>>
known_sizes_;
};
TORCH_API int& getTECudaPointwiseLoopLevels();
TORCH_API int& getTECudaPointwiseBlockCount();
TORCH_API int& getTECudaPointwiseBlockSize();
TORCH_API bool& getTEGenerateBlockCode();
TORCH_API bool fallbackAllowed();
TORCH_API bool setFallbackAllowed(bool value);
TORCH_API c10::optional<at::Device> pickDeviceType(
const at::ArrayRef<torch::jit::Value*>& inputs);
} // namespace tensorexpr
} // namespace jit
} // namespace torch
|
