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
|
#include <torch/csrc/profiler/python/init.h>
#include <ATen/record_function.h>
#include <torch/csrc/DynamicTypes.h>
#include <torch/csrc/autograd/utils/wrap_outputs.h>
#include <torch/csrc/jit/python/pybind_utils.h>
#include <torch/csrc/profiler/collection.h>
#include <torch/csrc/utils/pybind.h>
namespace torch {
namespace profiler {
void initPythonBindings(PyObject* module) {
auto rootModule = py::handle(module).cast<py::module>();
auto m = rootModule.def_submodule("_profiler");
using namespace torch::profiler::impl;
py::enum_<at::RecordScope>(m, "RecordScope")
.value("FUNCTION", at::RecordScope::FUNCTION)
.value("BACKWARD_FUNCTION", at::RecordScope::BACKWARD_FUNCTION)
.value("TORCHSCRIPT_FUNCTION", at::RecordScope::TORCHSCRIPT_FUNCTION)
.value("KERNEL_FUNCTION_DTYPE", at::RecordScope::KERNEL_FUNCTION_DTYPE)
.value("CUSTOM_CLASS", at::RecordScope::CUSTOM_CLASS)
.value("BUILD_FEATURE", at::RecordScope::BUILD_FEATURE)
.value("LITE_INTERPRETER", at::RecordScope::LITE_INTERPRETER)
.value("USER_SCOPE", at::RecordScope::USER_SCOPE)
.value("STATIC_RUNTIME_OP", at::RecordScope::STATIC_RUNTIME_OP)
.value("STATIC_RUNTIME_MODEL", at::RecordScope::STATIC_RUNTIME_MODEL);
py::enum_<ProfilerState>(m, "ProfilerState")
.value("Disabled", ProfilerState::Disabled)
.value("CPU", ProfilerState::CPU)
.value("CUDA", ProfilerState::CUDA)
.value("NVTX", ProfilerState::NVTX)
.value("ITT", ProfilerState::ITT)
.value("KINETO", ProfilerState::KINETO)
.value("KINETO_GPU_FALLBACK", ProfilerState::KINETO_GPU_FALLBACK);
py::enum_<ActiveProfilerType>(m, "ActiveProfilerType")
.value("NONE", ActiveProfilerType::NONE)
.value("LEGACY", ActiveProfilerType::LEGACY)
.value("KINETO", ActiveProfilerType::KINETO)
.value("NVTX", ActiveProfilerType::NVTX)
.value("ITT", ActiveProfilerType::ITT);
py::enum_<ActivityType>(m, "ProfilerActivity")
.value("CPU", ActivityType::CPU)
.value("CUDA", ActivityType::CUDA);
py::class_<ExperimentalConfig>(m, "_ExperimentalConfig")
.def(
py::init<
std::vector<std::string> /* profiler_metrics */,
bool /* profiler_measure_per_kernel */,
bool /* verbose */
>(),
"An experimental config for Kineto features. Please note that"
"backward compatibility is not guaranteed.\n"
" profiler_metrics : a list of CUPTI profiler metrics used\n"
" to measure GPU performance events.\n"
" If this list contains values Kineto runs in CUPTI profiler mode\n"
" profiler_measure_per_kernel (bool) : whether to profile metrics per kernel\n"
" or for the entire measurement duration.\n"
" verbose (bool) : whether the trace file has `Call stack` field or not.",
py::arg("profiler_metrics") = std::vector<std::string>(),
py::arg("profiler_measure_per_kernel") = false,
py::arg("verbose") = true)
.def(py::pickle(
[](const ExperimentalConfig& p) { // __getstate__
py::list py_metrics;
for (const auto& metric : p.profiler_metrics) {
py::bytes mbytes(metric);
py_metrics.append(mbytes);
}
/* Return a tuple that fully encodes the state of the config */
return py::make_tuple(
py_metrics, p.profiler_measure_per_kernel, p.verbose);
},
[](py::tuple t) { // __setstate__
if (t.size() != 3) {
throw std::runtime_error("Expected 3 values in state");
}
py::list py_metrics = t[0].cast<py::list>();
std::vector<std::string> metrics{py_metrics.size()};
for (const auto& py_metric : py_metrics) {
metrics.push_back(py::str(py_metric));
}
return ExperimentalConfig(
std::move(metrics), t[1].cast<bool>(), t[2].cast<bool>());
}));
py::class_<ProfilerConfig>(m, "ProfilerConfig")
.def(py::init<
ProfilerState,
bool, /* record_input_shapes */
bool, /* profile_memory */
bool, /* with_stack */
bool, /* with_flops */
bool, /* with_modules */
ExperimentalConfig /* experimental_config */
>());
py::enum_<EventType>(m, "_EventType")
.value("TorchOp", EventType::TorchOp)
.value("Backend", EventType::Backend)
.value("Allocation", EventType::Allocation)
.value("PyCall", EventType::PyCall)
.value("PyCCall", EventType::PyCCall)
.value("Kineto", EventType::Kineto);
py::class_<Inputs>(m, "_Inputs")
.def_readonly("shapes", &Inputs::shapes_)
.def_readonly("dtypes", &Inputs::dtypes_)
.def_readonly("strides", &Inputs::strides_)
.def_property_readonly(
"ivalues",
[](const Inputs& inputs) {
py::list list;
for (auto& v : inputs.ivalues_) {
list.append(torch::jit::toPyObject(v));
}
return list;
})
.def_readonly("tensor_metadata", &Inputs::tensor_metadata_);
py::class_<TensorMetadata>(m, "_TensorMetadata")
.def_readonly("impl_ptr", &TensorMetadata::impl_)
.def_readonly("storage_data_ptr", &TensorMetadata::data_)
.def_readonly("id", &TensorMetadata::id_)
.def_property_readonly(
"layout",
[](const TensorMetadata& metadata) {
PyObject* layout_obj =
torch::autograd::utils::wrap(metadata.layout_);
return py::reinterpret_borrow<py::object>(layout_obj);
})
.def_property_readonly("device", &TensorMetadata::device)
.def_property_readonly("dtype", [](const TensorMetadata& metadata) {
return py::reinterpret_borrow<py::object>(
torch::autograd::utils::wrap(torch::getTHPDtype(metadata.dtype_)));
});
using torch_op_t = ExtraFields<EventType::TorchOp>;
py::class_<torch_op_t>(m, "_ExtraFields_TorchOp")
.def_readonly("inputs", &torch_op_t::inputs_)
.def_readonly("scope", &torch_op_t::scope_)
.def_readonly("sequence_number", &torch_op_t::sequence_number_)
.def_readonly("allow_tf32_cublas", &torch_op_t::allow_tf32_cublas_);
py::class_<ExtraFields<EventType::Backend>>(m, "_ExtraFields_Backend");
using allocation_t = ExtraFields<EventType::Allocation>;
py::class_<allocation_t>(m, "_ExtraFields_Allocation")
.def_property_readonly(
"ptr",
[](const allocation_t& a) {
return reinterpret_cast<intptr_t>(a.ptr_);
})
.def_readonly("id", &allocation_t::id_)
.def_readonly("alloc_size", &allocation_t::alloc_size_)
.def_readonly("total_allocated", &allocation_t::total_allocated_)
.def_readonly("total_reserved", &allocation_t::total_reserved_)
.def_property_readonly("device", &allocation_t::device);
py::class_<PyFrameState>(m, "_PyFrameState")
.def_readonly("line_number", &PyFrameState::line_no_)
.def_property_readonly(
"file_name", [](const PyFrameState& s) { return s.filename_.str(); })
.def_property_readonly("function_name", [](const PyFrameState& s) {
return s.funcname_.str();
});
py::class_<NNModuleInfo>(m, "_NNModuleInfo")
.def_property_readonly(
"params",
[](const NNModuleInfo& s) {
py::list list;
for (auto& p : s.params_) {
list.append(std::make_pair(
p.first, reinterpret_cast<intptr_t>(p.second)));
}
return list;
})
.def_property_readonly(
"cls_name", [](const NNModuleInfo& s) { return s.cls_name_.str(); });
py::class_<OptimizerInfo>(m, "_OptInfo")
.def_property_readonly(
"self",
[](const OptimizerInfo& a) {
return reinterpret_cast<intptr_t>(a.self_.value_of());
})
.def_property_readonly(
"param_addrs",
[](const OptimizerInfo& s) {
py::list params_addrs;
for (auto& addr : s.params_addr_) {
params_addrs.append(reinterpret_cast<intptr_t>(addr));
}
return params_addrs;
})
.def_property_readonly("opt_state", [](const OptimizerInfo& s) {
py::list states;
for (auto& a : s.opt_state_) {
states.append(
std::make_pair(a.first, reinterpret_cast<intptr_t>(a.second)));
}
return states;
});
py::class_<ExtraFields<EventType::PyCall>>(m, "_ExtraFields_PyCall")
.def_readonly("opt", &ExtraFields<EventType::PyCall>::opt_)
.def_readonly("callsite", &ExtraFields<EventType::PyCall>::callsite_)
.def_readonly("caller", &ExtraFields<EventType::PyCall>::caller_)
.def_readonly("module", &ExtraFields<EventType::PyCall>::module_);
py::class_<ExtraFields<EventType::PyCCall>>(m, "_ExtraFields_PyCCall")
.def_readonly("caller", &ExtraFields<EventType::PyCall>::caller_);
py::class_<ExtraFields<EventType::OutOfMemory>>(
m, "_ExtraFields_OutOfMemory");
py::class_<ExtraFields<EventType::Kineto>>(m, "_ExtraFields_Kineto");
py::class_<Result, std::shared_ptr<Result>>(m, "_ProfilerEvent")
.def_property_readonly("name", &Result::name)
.def_property_readonly("tag", &Result::tag)
.def_readonly("extra_fields", &Result::extra_fields_)
.def_property_readonly(
"id",
[](const Result& r) {
return reinterpret_cast<intptr_t>(r.shared_from_this().get());
})
.def_property_readonly(
"parent", [](const Result& r) { return r.parent_.lock(); })
.def_readonly("children", &Result::children_)
.def_readonly("start_time_ns", &Result::start_time_ns_)
.def_readonly("start_tid", &Result::start_tid_)
.def_property_readonly("correlation_id", &Result::correlationID)
.def_property_readonly("end_time_ns", &Result::endTimeNS)
.def_property_readonly("duration_time_ns", [](const Result& r) {
return r.endTimeNS() - r.start_time_ns_;
});
}
} // namespace profiler
} // namespace torch
|
