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
|
#include <torch/csrc/jit/passes/graph_rewrite_helper.h>
#include <torch/csrc/jit/ir/subgraph_matcher.h>
#include <torch/csrc/jit/passes/constant_propagation.h>
#include <torch/csrc/jit/passes/subgraph_rewrite.h>
namespace torch {
namespace jit {
namespace graph_rewrite_helper {
std::string getFuncName(Value* func_value) {
auto func = func_value->type()->expectRef<FunctionType>().function();
const auto& qname = func->qualname();
const auto& name = qname.qualifiedName();
auto rdot_idx = name.rfind('.');
if (rdot_idx != std::string::npos) {
return name.substr(rdot_idx + 1, name.length());
} else {
return name;
}
}
Value* getValue(
const std::string& name,
const std::unordered_map<const Value*, Value*>& match_vmap,
const std::unordered_map<std::string, Value*>& vmap) {
return match_vmap.at(vmap.at(name));
}
c10::optional<IValue> getIValue(
const std::string& name,
const std::unordered_map<const Value*, Value*>& match_vmap,
const std::unordered_map<std::string, Value*>& vmap) {
return toIValue(getValue(name, match_vmap, vmap));
}
std::unordered_map<std::string, c10::IValue> getConvParams(
const Match& match,
const std::unordered_map<std::string, Value*>& vmap) {
std::unordered_map<std::string, c10::IValue> calc_values;
const auto& match_vmap = match.values_map;
auto transposed_value = getIValue("transposed", match_vmap, vmap).value();
calc_values["transposed"] = transposed_value;
auto output_padding_value =
getIValue("output_padding", match_vmap, vmap).value();
calc_values["output_padding"] = output_padding_value;
auto stride_value = getIValue("stride", match_vmap, vmap).value();
calc_values["stride"] = stride_value;
auto padding_value = getIValue("padding", match_vmap, vmap).value();
calc_values["padding"] = padding_value;
auto dilation_value = getIValue("dilation", match_vmap, vmap).value();
calc_values["dilation"] = dilation_value;
return calc_values;
}
void replaceConvolutionWithAtenConv(std::shared_ptr<Graph>& graph) {
// TODO: remove constant prop in the pass
ConstantPropagation(graph);
std::string convolution_deprecated = R"(
graph(%a, %w, %b, %stride:int[], %padding:int[], %dilation:int[],
%transposed:bool, %output_padding:int[], %groups:int, %benchmark:bool,
%deterministic:bool, %cudnn_enabled:bool):
%r = aten::_convolution(%a, %w, %b, %stride, %padding, %dilation,
%transposed, %output_padding, %groups, %benchmark, %deterministic, %cudnn_enabled)
return (%r) )";
std::string convolution = R"(
graph(%a, %w, %b, %stride:int[], %padding:int[], %dilation:int[],
%transposed:bool, %output_padding:int[], %groups:int, %benchmark:bool,
%deterministic:bool, %cudnn_enabled:bool, %allow_tf32:bool):
%r = aten::_convolution(%a, %w, %b, %stride, %padding, %dilation,
%transposed, %output_padding, %groups, %benchmark, %deterministic, %cudnn_enabled, %allow_tf32)
return (%r) )";
std::string conv2d_for_deprecated_conv = R"(
graph(%a, %w, %b, %stride:int[], %padding:int[], %dilation:int[],
%transposed:bool, %output_padding:int[], %groups:int, %benchmark:bool,
%deterministic:bool, %cudnn_enabled:bool):
%r = aten::conv2d(%a, %w, %b, %stride, %padding, %dilation, %groups)
return (%r) )";
std::string conv2d = R"(
graph(%a, %w, %b, %stride:int[], %padding:int[], %dilation:int[],
%transposed:bool, %output_padding:int[], %groups:int, %benchmark:bool,
%deterministic:bool, %cudnn_enabled:bool, %allow_tf32:bool):
%r = aten::conv2d(%a, %w, %b, %stride, %padding, %dilation, %groups)
return (%r) )";
std::string conv1d_for_deprecated_conv = R"(
graph(%a, %w, %b, %stride:int[], %padding:int[], %dilation:int[],
%transposed:bool, %output_padding:int[], %groups:int, %benchmark:bool,
%deterministic:bool, %cudnn_enabled:bool):
%r = aten::conv1d(%a, %w, %b, %stride, %padding, %dilation, %groups)
return (%r) )";
std::string conv1d = R"(
graph(%a, %w, %b, %stride:int[], %padding:int[], %dilation:int[],
%transposed:bool, %output_padding:int[], %groups:int, %benchmark:bool,
%deterministic:bool, %cudnn_enabled:bool, %allow_tf32:bool):
%r = aten::conv1d(%a, %w, %b, %stride, %padding, %dilation, %groups)
return (%r) )";
std::string conv3d_for_deprecated_conv = R"(
graph(%a, %w, %b, %stride:int[], %padding:int[], %dilation:int[],
%transposed:bool, %output_padding:int[], %groups:int, %benchmark:bool,
%deterministic:bool, %cudnn_enabled:bool):
%r = aten::conv3d(%a, %w, %b, %stride, %padding, %dilation, %groups)
return (%r) )";
std::string conv3d = R"(
graph(%a, %w, %b, %stride:int[], %padding:int[], %dilation:int[],
%transposed:bool, %output_padding:int[], %groups:int, %benchmark:bool,
%deterministic:bool, %cudnn_enabled:bool, %allow_tf32:bool):
%r = aten::conv3d(%a, %w, %b, %stride, %padding, %dilation, %groups)
return (%r) )";
std::string conv_transpose1d_for_deprecated_conv = R"(
graph(%a, %w, %b, %stride:int[], %padding:int[], %dilation:int[],
%transposed:bool, %output_padding:int[], %groups:int, %benchmark:bool,
%deterministic:bool, %cudnn_enabled:bool):
%r = aten::conv_transpose1d(%a, %w, %b, %stride, %padding, %output_padding, %groups, %dilation)
return (%r) )";
std::string conv_transpose1d = R"(
graph(%a, %w, %b, %stride:int[], %padding:int[], %dilation:int[],
%transposed:bool, %output_padding:int[], %groups:int, %benchmark:bool,
%deterministic:bool, %cudnn_enabled:bool, %allow_tf32:bool):
%r = aten::conv_transpose1d(%a, %w, %b, %stride, %padding, %output_padding, %groups, %dilation)
return (%r) )";
std::string conv_transpose2d_for_deprecated_conv = R"(
graph(%a, %w, %b, %stride:int[], %padding:int[], %dilation:int[],
%transposed:bool, %output_padding:int[], %groups:int, %benchmark:bool,
%deterministic:bool, %cudnn_enabled:bool):
%r = aten::conv_transpose2d(%a, %w, %b, %stride, %padding, %output_padding, %groups, %dilation)
return (%r) )";
std::string conv_transpose2d = R"(
graph(%a, %w, %b, %stride:int[], %padding:int[], %dilation:int[],
%transposed:bool, %output_padding:int[], %groups:int, %benchmark:bool,
%deterministic:bool, %cudnn_enabled:bool, %allow_tf32:bool):
%r = aten::conv_transpose2d(%a, %w, %b, %stride, %padding, %output_padding, %groups, %dilation)
return (%r) )";
std::string conv_transpose3d_for_deprecated_conv = R"(
graph(%a, %w, %b, %stride:int[], %padding:int[], %dilation:int[],
%transposed:bool, %output_padding:int[], %groups:int, %benchmark:bool,
%deterministic:bool, %cudnn_enabled:bool):
%r = aten::conv_transpose3d(%a, %w, %b, %stride, %padding, %output_padding, %groups, %dilation)
return (%r) )";
std::string conv_transpose3d = R"(
graph(%a, %w, %b, %stride:int[], %padding:int[], %dilation:int[],
%transposed:bool, %output_padding:int[], %groups:int, %benchmark:bool,
%deterministic:bool, %cudnn_enabled:bool, %allow_tf32:bool):
%r = aten::conv_transpose3d(%a, %w, %b, %stride, %padding, %output_padding, %groups, %dilation)
return (%r) )";
// Filter the unsupported case
auto filter_conv1d = [](const Match& match,
const std::unordered_map<std::string, Value*>& vmap) {
auto calc_value_map = getConvParams(match, vmap);
if (calc_value_map["output_padding"].toIntList().size() != 1 ||
calc_value_map["stride"].toIntList().size() != 1 ||
calc_value_map["padding"].toIntList().size() != 1 ||
calc_value_map["dilation"].toIntList().size() != 1) {
return false;
}
return !calc_value_map["transposed"].toBool();
};
auto filter_conv2d = [](const Match& match,
const std::unordered_map<std::string, Value*>& vmap) {
auto calc_value_map = getConvParams(match, vmap);
if (calc_value_map["output_padding"].toIntList().size() != 2 ||
calc_value_map["stride"].toIntList().size() != 2 ||
calc_value_map["padding"].toIntList().size() != 2 ||
calc_value_map["dilation"].toIntList().size() != 2) {
return false;
}
return !calc_value_map["transposed"].toBool();
};
auto filter_conv3d = [](const Match& match,
const std::unordered_map<std::string, Value*>& vmap) {
auto calc_value_map = getConvParams(match, vmap);
if (calc_value_map["output_padding"].toIntList().size() != 3 ||
calc_value_map["stride"].toIntList().size() != 3 ||
calc_value_map["padding"].toIntList().size() != 3 ||
calc_value_map["dilation"].toIntList().size() != 3) {
return false;
}
return !calc_value_map["transposed"].toBool();
};
auto filter_conv_transpose1d =
[](const Match& match,
const std::unordered_map<std::string, Value*>& vmap) {
auto calc_value_map = getConvParams(match, vmap);
if (calc_value_map["output_padding"].toIntList().size() != 1 ||
calc_value_map["stride"].toIntList().size() != 1 ||
calc_value_map["padding"].toIntList().size() != 1 ||
calc_value_map["dilation"].toIntList().size() != 1) {
return false;
}
return calc_value_map["transposed"].toBool();
};
auto filter_conv_transpose2d =
[](const Match& match,
const std::unordered_map<std::string, Value*>& vmap) {
auto calc_value_map = getConvParams(match, vmap);
if (calc_value_map["output_padding"].toIntList().size() != 2 ||
calc_value_map["stride"].toIntList().size() != 2 ||
calc_value_map["padding"].toIntList().size() != 2 ||
calc_value_map["dilation"].toIntList().size() != 2) {
return false;
}
return calc_value_map["transposed"].toBool();
};
auto filter_conv_transpose3d =
[](const Match& match,
const std::unordered_map<std::string, Value*>& vmap) {
auto calc_value_map = getConvParams(match, vmap);
if (calc_value_map["output_padding"].toIntList().size() != 3 ||
calc_value_map["stride"].toIntList().size() != 3 ||
calc_value_map["padding"].toIntList().size() != 3 ||
calc_value_map["dilation"].toIntList().size() != 3) {
return false;
}
return calc_value_map["transposed"].toBool();
};
SubgraphRewriter rewriter_conv1d;
rewriter_conv1d.RegisterRewritePattern(convolution, conv1d);
rewriter_conv1d.RegisterRewritePattern(
convolution_deprecated, conv1d_for_deprecated_conv);
rewriter_conv1d.runOnGraph(graph, filter_conv1d);
SubgraphRewriter rewriter_conv2d;
rewriter_conv2d.RegisterRewritePattern(convolution, conv2d);
rewriter_conv2d.RegisterRewritePattern(
convolution_deprecated, conv2d_for_deprecated_conv);
rewriter_conv2d.runOnGraph(graph, filter_conv2d);
SubgraphRewriter rewriter_conv3d;
rewriter_conv3d.RegisterRewritePattern(convolution, conv3d);
rewriter_conv3d.RegisterRewritePattern(
convolution_deprecated, conv3d_for_deprecated_conv);
rewriter_conv3d.runOnGraph(graph, filter_conv3d);
SubgraphRewriter rewriter_conv_transpose1d;
rewriter_conv_transpose1d.RegisterRewritePattern(
convolution, conv_transpose1d);
rewriter_conv_transpose1d.RegisterRewritePattern(
convolution_deprecated, conv_transpose1d_for_deprecated_conv);
rewriter_conv_transpose1d.runOnGraph(graph, filter_conv_transpose1d);
SubgraphRewriter rewriter_conv_transpose2d;
rewriter_conv_transpose2d.RegisterRewritePattern(
convolution, conv_transpose2d);
rewriter_conv_transpose2d.RegisterRewritePattern(
convolution_deprecated, conv_transpose2d_for_deprecated_conv);
rewriter_conv_transpose2d.runOnGraph(graph, filter_conv_transpose2d);
SubgraphRewriter rewriter_conv_transpose3d;
rewriter_conv_transpose3d.RegisterRewritePattern(
convolution, conv_transpose3d);
rewriter_conv_transpose3d.RegisterRewritePattern(
convolution_deprecated, conv_transpose3d_for_deprecated_conv);
rewriter_conv_transpose3d.runOnGraph(graph, filter_conv_transpose3d);
}
bool isClampFusable(
const Match& match,
const std::unordered_map<std::string, Value*>& vmap) {
const auto& match_vmap = match.values_map;
TORCH_CHECK(
vmap.find("dummy_min_max") != vmap.end(),
"Expected to find dummy_min_max Value in the subgraph to be replaced.");
auto dummy_min_max =
graph_rewrite_helper::getIValue("dummy_min_max", match_vmap, vmap);
auto is_fusable = !dummy_min_max || dummy_min_max.value().isNone();
// Also check if the output_min and output_max values are actually constant.
// If hardtanh's min/max Value's are not actually constants, we will end up
// rerouting those values to prepack op. And if they are not constants
// we will not be able to remove prepacking ops.
if (vmap.find("output_min") != vmap.end()) {
// aten::relu pattern does not have output_min/output_max.
// aten::hardtanh/_ does.
TORCH_CHECK(
vmap.find("output_max") != vmap.end(),
"Expected to find output_max as well given "
"output_min exist in pattern graph.");
// If output_min/max are not constant, we get c10::nullopt.
auto output_min =
graph_rewrite_helper::getIValue("output_min", match_vmap, vmap);
auto output_max =
graph_rewrite_helper::getIValue("output_max", match_vmap, vmap);
is_fusable =
is_fusable && (output_min.has_value() && output_max.has_value());
}
return is_fusable;
}
} // namespace graph_rewrite_helper
} // namespace jit
} // namespace torch
|
