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
|
#pragma once
#include <torch/csrc/Export.h>
#include <functional>
#include <vector>
#include <torch/csrc/jit/tensorexpr/expr.h>
#include <torch/csrc/jit/tensorexpr/reduction.h>
namespace torch {
namespace jit {
namespace tensorexpr {
class TORCH_API Tensor {
public:
// NOLINTNEXTLINE(cppcoreguidelines-pro-type-member-init)
Tensor(BufPtr buf, const std::vector<VarPtr>& args, ExprPtr body)
: buf_(buf) {
stmt_ = constructStmt(args, body, {}, {});
}
Tensor(BufHandle buf, const std::vector<VarHandle>& args, ExprHandle body)
: Tensor(buf.node(), VarHandleVectorToVarVector(args), body.node()) {}
// NOLINTNEXTLINE(cppcoreguidelines-pro-type-member-init)
Tensor(
BufPtr buf,
const std::vector<VarPtr>& args,
const std::vector<ExprPtr>& reduce_dims,
const std::vector<VarPtr>& reduce_args,
ExprPtr body)
: buf_(buf) {
stmt_ = constructStmt(args, body, reduce_dims, reduce_args);
}
Tensor(
BufHandle buf,
const std::vector<VarHandle>& args,
const std::vector<ExprHandle>& reduce_dims,
const std::vector<VarHandle>& reduce_args,
ExprHandle body)
: Tensor(
buf.node(),
VarHandleVectorToVarVector(args),
ExprHandleVectorToExprVector(reduce_dims),
VarHandleVectorToVarVector(reduce_args),
body.node()) {}
Tensor(BufPtr buf, StmtPtr stmt) : buf_(buf), stmt_(stmt) {}
BufPtr buf() const {
return buf_;
}
StmtPtr stmt() const {
return stmt_;
}
template <typename T>
inline ExprHandle load(const std::vector<T>& args) const;
template <typename... Ts>
inline ExprHandle load(const Ts&... ts) const;
private:
StmtPtr constructStmt(
const std::vector<VarPtr>& args,
ExprPtr body,
const std::vector<ExprPtr>& reduce_dims,
const std::vector<VarPtr>& reduce_args) const;
BufPtr buf_;
StmtPtr stmt_;
};
TORCH_API Tensor Compute(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
c10::optional<std::vector<ExprHandle>> strides,
const std::function<ExprHandle(const VarHandle&)>& body_func);
TORCH_API Tensor Compute(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
const std::function<ExprHandle(const VarHandle&)>& body_func);
TORCH_API Tensor Compute(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
c10::optional<std::vector<ExprHandle>> strides,
const std::function<ExprHandle(const VarHandle&, const VarHandle&)>&
body_func);
TORCH_API Tensor Compute(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
const std::function<ExprHandle(const VarHandle&, const VarHandle&)>&
body_func);
TORCH_API Tensor Compute(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
c10::optional<std::vector<ExprHandle>> strides,
const std::function<
ExprHandle(const VarHandle&, const VarHandle&, const VarHandle&)>&
body_func);
TORCH_API Tensor Compute(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
const std::function<
ExprHandle(const VarHandle&, const VarHandle&, const VarHandle&)>&
body_func);
TORCH_API Tensor Compute(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
c10::optional<std::vector<ExprHandle>> strides,
const std::function<ExprHandle(
const VarHandle&,
const VarHandle&,
const VarHandle&,
const VarHandle&)>& body_func);
TORCH_API Tensor Compute(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
const std::function<ExprHandle(
const VarHandle&,
const VarHandle&,
const VarHandle&,
const VarHandle&)>& body_func);
TORCH_API Tensor Compute(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
c10::optional<std::vector<ExprHandle>> strides,
const std::function<ExprHandle(const std::vector<VarHandle>&)>& body_func);
TORCH_API Tensor Compute(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
const std::function<ExprHandle(const std::vector<VarHandle>&)>& body_func);
inline std::vector<VarHandle> create_index_vars(
const std::vector<ExprHandle>& dims) {
std::vector<VarHandle> vars;
vars.reserve(dims.size());
for (const ExprHandle& dim : dims) {
vars.push_back(VarHandle(alloc<Var>(
"i", dim.dtype().scalar_type() == ScalarType::Long ? kLong : kInt)));
}
return vars;
}
// Handle reductions over a Reducer and a body_func which produces values.
template <typename InitFunc, typename BodyFunc>
Tensor Reduce(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
c10::optional<std::vector<ExprHandle>> strides,
const Reducer& reducer,
const InitFunc& init_func,
const BodyFunc& body_func,
const std::vector<ExprHandle>& reduce_dims) {
std::vector<VarHandle> vars = create_index_vars(dims);
std::vector<VarHandle> reduce_vars = create_index_vars(reduce_dims);
// If reduce_vars is empty, then it's not a reduction, but rather a simple
// copy
if (reduce_vars.empty()) {
ExprHandle body = Reducer::getReduceBody(body_func, vars);
BufHandle func_result =
Buf::make(func_name, dims, body.dtype(), c10::nullopt, strides);
return Tensor(func_result, vars, body);
}
std::vector<VarHandle> all_vars;
all_vars.insert(all_vars.end(), vars.begin(), vars.end());
all_vars.insert(all_vars.end(), reduce_vars.begin(), reduce_vars.end());
ExprHandle body = Reducer::getReduceBody(body_func, all_vars);
std::vector<ExprHandle> output_args(vars.begin(), vars.end());
ExprHandle init_expr = Cast::make(body.dtype(), init_func(vars));
BufHandle func_result = Buf::make(func_name, dims, body.dtype(), init_expr);
ExprHandle reduce_op = reducer(func_result, body, output_args, reduce_vars);
if (body.dtype() == kBFloat16) {
ExprHandle init_expr_acc = Cast::make(kFloat, init_func(vars));
BufHandle func_result_acc =
Buf::make(func_name + "_acc", dims, kFloat, init_expr_acc);
reduce_op =
reducer(func_result, func_result_acc, body, output_args, reduce_vars);
}
Tensor t = Tensor(func_result, vars, reduce_dims, reduce_vars, reduce_op);
return t;
}
template <typename InitFunc, typename BodyFunc>
Tensor Reduce(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
const Reducer& reducer,
const InitFunc& init_func,
const BodyFunc& body_func,
const std::vector<ExprHandle>& reduce_dims) {
return Reduce<InitFunc, BodyFunc>(
func_name,
dims,
c10::nullopt,
reducer,
init_func,
body_func,
reduce_dims);
}
template <typename BodyFunc>
Tensor Reduce(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
c10::optional<std::vector<ExprHandle>> strides,
const Reducer& reducer,
const BodyFunc& body_func,
const std::vector<ExprHandle>& reduce_dims) {
return Reduce(
func_name,
dims,
strides,
reducer,
[&](ParameterList p) { return ExprHandle(reducer.initializer()); },
body_func,
reduce_dims);
}
template <typename BodyFunc>
Tensor Reduce(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
const Reducer& reducer,
const BodyFunc& body_func,
const std::vector<ExprHandle>& reduce_dims) {
return Reduce<BodyFunc>(
func_name, dims, c10::nullopt, reducer, body_func, reduce_dims);
}
// Overload which allows inline lambda functions for the body_func.
template <typename BodyFunc>
Tensor Reduce(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
c10::optional<std::vector<ExprHandle>> strides,
const Reducer& reducer,
const BodyFunc&& body_func,
const std::vector<ExprHandle>& reduce_dims) {
return Reduce(func_name, dims, strides, reducer, body_func, reduce_dims);
}
template <typename BodyFunc>
Tensor Reduce(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
const Reducer& reducer,
const BodyFunc&& body_func,
const std::vector<ExprHandle>& reduce_dims) {
return Reduce(func_name, dims, c10::nullopt, reducer, body_func, reduce_dims);
}
TORCH_API Tensor Reduce(
const std::string& name,
const std::vector<ExprHandle>& dims,
c10::optional<std::vector<ExprHandle>> strides,
const Reducer& reducer,
const BufHandle& buffer,
const std::vector<ExprHandle>& reduce_dims);
TORCH_API Tensor Reduce(
const std::string& name,
const std::vector<ExprHandle>& dims,
const Reducer& reducer,
const BufHandle& buffer,
const std::vector<ExprHandle>& reduce_dims);
// Overload for the common case of all dimensions of a prevously Computed
// Tensor.
TORCH_API Tensor Reduce(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
c10::optional<std::vector<ExprHandle>> strides,
const Reducer& reducer,
Tensor tensor,
const std::vector<ExprHandle>& reduce_dims);
TORCH_API Tensor Reduce(
const std::string& func_name,
const std::vector<ExprHandle>& dims,
const Reducer& reducer,
Tensor tensor,
const std::vector<ExprHandle>& reduce_dims);
template <typename... Ts>
inline ExprHandle Tensor::load(const Ts&... ts) const {
// NOLINTNEXTLINE(cppcoreguidelines-init-variables)
std::vector<ExprHandle> params({ExprHandle(ts)...});
return Load::make(BufHandle(this->buf()), params);
}
template <typename T>
inline ExprHandle Tensor::load(const std::vector<T>& args) const {
// NOLINTNEXTLINE(cppcoreguidelines-init-variables)
std::vector<ExprHandle> params(args.begin(), args.end());
return Load::make(BufHandle(this->buf()), params);
}
template <typename... Ts>
inline ExprHandle BufHandle::load(const Ts&... ts) const {
// NOLINTNEXTLINE(cppcoreguidelines-init-variables)
std::vector<ExprHandle> params({ExprHandle(ts)...});
return ExprHandle(alloc<Load>(node(), ExprHandleVectorToExprVector(params)));
}
template <typename T>
inline ExprHandle BufHandle::load(const std::vector<T>& args) const {
// NOLINTNEXTLINE(cppcoreguidelines-init-variables)
std::vector<ExprHandle> params(args.begin(), args.end());
return ExprHandle(alloc<Load>(node(), ExprHandleVectorToExprVector(params)));
}
inline ExprHandle BufHandle::load(const std::vector<ExprHandle>& args) const {
return this->template load<ExprHandle>(args);
}
} // namespace tensorexpr
} // namespace jit
} // namespace torch
|
