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
|
#include <torch/csrc/jit/runtime/static/init.h>
#include <torch/csrc/jit/passes/freeze_module.h>
#include <torch/csrc/jit/runtime/static/fusion.h>
#include <torch/csrc/jit/runtime/static/impl.h>
// This number is a heuristic determined with pytorch/benchmark
#define DEFAULT_FUSION_SIZE 4
namespace torch {
namespace jit {
void initStaticModuleBindings(PyObject* module) {
auto m = py::handle(module).cast<py::module>();
py::class_<StaticModule> static_module(m, "StaticModule");
py::class_<StaticRuntime::IndividualMetrics>(
static_module, "IndividualMetrics")
.def_readonly("setup_time", &StaticRuntime::IndividualMetrics::setup_time)
.def_readonly(
"memory_alloc_time",
&StaticRuntime::IndividualMetrics::memory_alloc_time)
.def_readonly(
"memory_dealloc_time",
&StaticRuntime::IndividualMetrics::memory_dealloc_time)
.def_readonly(
"output_dealloc_time",
&StaticRuntime::IndividualMetrics::output_dealloc_time)
.def_readonly(
"first_iter_time", &StaticRuntime::IndividualMetrics::first_iter_time)
.def_readonly("total_time", &StaticRuntime::IndividualMetrics::total_time)
.def_readonly(
"out_nodes_count", &StaticRuntime::IndividualMetrics::out_nodes_count)
.def_readonly(
"total_nodes_count",
&StaticRuntime::IndividualMetrics::total_nodes_count)
.def_readonly(
"time_per_node", &StaticRuntime::IndividualMetrics::time_per_node)
.def_readonly(
"time_per_node_type",
&StaticRuntime::IndividualMetrics::time_per_node_type)
.def_readonly(
"percent_per_node_type",
&StaticRuntime::IndividualMetrics::percent_per_node_type)
.def_readonly(
"instances_per_node_type",
&StaticRuntime::IndividualMetrics::instances_per_node_type)
.def_readonly("out_nodes", &StaticRuntime::IndividualMetrics::out_nodes);
static_module
.def(
"__call__",
[](StaticModule& self,
const py::args& args,
const py::kwargs& kwargs) {
std::vector<c10::IValue> arg_ivalues;
std::unordered_map<std::string, c10::IValue> kwarg_ivalues;
for (size_t i = 0; i < args.size(); ++i) {
auto ivalue = torch::jit::toIValue(args[i], c10::AnyType::get());
arg_ivalues.push_back(ivalue);
}
for (const auto& kv : kwargs) {
kwarg_ivalues[py::cast<std::string>(kv.first)] =
torch::jit::toIValue(kv.second, c10::AnyType::get());
}
c10::IValue ret = self(arg_ivalues, kwarg_ivalues);
return toPyObject(ret);
})
.def(
"benchmark",
[](StaticModule& self,
const std::vector<at::Tensor>& args,
const std::unordered_map<std::string, at::Tensor>& kwargs,
const int warmup_runs,
const int main_runs) {
std::vector<c10::IValue> arg_ivalues{args.begin(), args.end()};
std::unordered_map<std::string, c10::IValue> kwarg_ivalues{
kwargs.begin(), kwargs.end()};
self.runtime().benchmark(
{arg_ivalues}, {kwarg_ivalues}, warmup_runs, main_runs);
})
.def(
"benchmark_individual_ops",
[](StaticModule& self,
const std::vector<at::Tensor>& args,
const std::unordered_map<std::string, at::Tensor>& kwargs,
const int warmup_runs,
const int main_runs) {
std::vector<c10::IValue> arg_ivalues{args.begin(), args.end()};
std::unordered_map<std::string, c10::IValue> kwarg_ivalues{
kwargs.begin(), kwargs.end()};
return self.runtime().benchmark_individual_ops(
{arg_ivalues}, {kwarg_ivalues}, warmup_runs, main_runs);
})
.def(
"runAsync",
[](StaticModule& self,
const py::tuple& args,
const py::dict& kwargs) {
std::vector<c10::IValue> arg_ivalues;
for (const auto& elem : args) {
arg_ivalues.push_back(
torch::jit::toIValue(elem, c10::AnyType::get()));
}
std::unordered_map<std::string, c10::IValue> kwarg_ivalues;
for (const auto& kv : kwargs) {
kwarg_ivalues[py::cast<std::string>(kv.first)] =
torch::jit::toIValue(kv.second, c10::AnyType::get());
}
// custom executor for async op execution
auto task_launcher = [](const std::function<void()>& f) {
at::launch(f);
};
return toPyObject(self.runtime().runAsync(
arg_ivalues, kwarg_ivalues, task_launcher));
});
m.def(
"_jit_to_static_module",
[](std::shared_ptr<torch::jit::Graph> g) { return StaticModule(g); })
.def(
"_jit_to_static_module",
[](const torch::jit::Module& module) { return StaticModule(module); })
.def(
"_fuse_to_static_module",
[](torch::jit::Module& module, size_t min_size) {
module.eval();
module = freeze_module(module);
Method method = module.get_method("forward");
auto graph = method.graph();
fuseStaticSubgraphs(graph, min_size);
},
py::arg("module"),
py::arg("min_size") = DEFAULT_FUSION_SIZE)
.def(
"_fuse_to_static_module",
[](std::shared_ptr<torch::jit::Graph> g, size_t min_size) {
fuseStaticSubgraphs(g, min_size);
},
py::arg("graph"),
py::arg("min_size") = DEFAULT_FUSION_SIZE);
}
} // namespace jit
} // namespace torch
|
