-
Notifications
You must be signed in to change notification settings - Fork 81
Expand file tree
/
Copy pathprofile.cpp
More file actions
265 lines (249 loc) · 8.33 KB
/
profile.cpp
File metadata and controls
265 lines (249 loc) · 8.33 KB
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
// clang-format off
/*
* SPDX-FileCopyrightText: Copyright (c) 2025-present NVIDIA CORPORATION & AFFILIATES.
* All rights reserved.
* SPDX-License-Identifier: BSD-3-Clause
*/
// clang-format on
#include <bindings.h>
#include <fusion_profiler.h>
namespace nvfuser::python {
namespace {
void bindFusionProfile(py::module& nvfuser) {
py::class_<KernelProfile> kernel_prof(
nvfuser, "KernelProfile", py::module_local());
kernel_prof.def("__repr__", [](KernelProfile& self) {
std::stringstream ss;
ss << self;
return ss.str();
});
kernel_prof.def_property_readonly(
"name", [](KernelProfile& self) { return self.name; }, R"(
Returns the kernel name.
)");
kernel_prof.def_property_readonly(
"segment_id", [](KernelProfile& self) { return self.segment_id; }, R"(
Returns the segment id of the kernel.
)");
kernel_prof.def_property_readonly(
"device", [](KernelProfile& self) { return self.device; }, R"(
Returns the device id where the kernel was executed.
)");
kernel_prof.def_property_readonly(
"stream", [](KernelProfile& self) { return self.stream; }, R"(
Returns the CUDA stream id used for kernel execution.
)");
kernel_prof.def_property_readonly(
"correlation_id",
[](KernelProfile& self) { return self.correlation_id; },
R"(
Returns the CUPTI correlation id for the kernel.
)");
kernel_prof.def_property_readonly(
"compile_time_ms",
[](KernelProfile& self) { return self.compile_time_ms; },
R"(
Returns the compile time in milliseconds of the kernel.
)");
kernel_prof.def_property_readonly(
"time_ms", [](KernelProfile& self) { return self.time_ms; }, R"(
Returns the execution time in milliseconds of the kernel.
)");
kernel_prof.def_property_readonly(
"effective_bandwidth_gbs",
[](KernelProfile& self) { return self.effective_bandwidth_gbs; },
R"(
Returns the effective bandwidth in gigabytes per second of the kernel.
)");
kernel_prof.def_property_readonly(
"percentage_peak_bandwidth",
[](KernelProfile& self) { return self.percentage_peak_bandwidth; },
R"(
Returns the percentage of peak bandwidth achieved by the kernel.
)");
kernel_prof.def_property_readonly(
"grid_str", [](KernelProfile& self) { return self.grid_str; }, R"(
Returns the grid dimensions as a string.
)");
kernel_prof.def_property_readonly(
"block_str", [](KernelProfile& self) { return self.block_str; }, R"(
Returns the block dimensions as a string.
)");
kernel_prof.def_property_readonly(
"cluster_str", [](KernelProfile& self) { return self.cluster_str; }, R"(
Returns the cluster dimensions as a string.
)");
kernel_prof.def_property_readonly(
"shared_mem_str",
[](KernelProfile& self) { return self.shared_mem_str; },
R"(
Returns the shared memory usage as a string.
)");
kernel_prof.def_property_readonly(
"registers", [](KernelProfile& self) { return self.registers; }, R"(
Returns the number of registers used by the kernel.
)");
kernel_prof.def_property_readonly(
"input_bytes", [](KernelProfile& self) { return self.input_bytes; }, R"(
Returns the input bytes processed by the kernel.
)");
kernel_prof.def_property_readonly(
"output_bytes", [](KernelProfile& self) { return self.output_bytes; }, R"(
Returns the output bytes generated by the kernel.
)");
kernel_prof.def_property_readonly(
"scheduler", [](KernelProfile& self) { return self.scheduler; }, R"(
Returns the scheduler type used for the kernel.
)");
py::class_<FusionProfile> fusion_prof(
nvfuser, "FusionProfile", py::module_local());
kernel_prof.def("__repr__", [](KernelProfile& self) {
std::stringstream ss;
ss << self;
return ss.str();
});
fusion_prof.def_property_readonly(
"verbose", [](FusionProfile& self) { return self.verbose; }, R"(
Returns the verbosity of the fusion profile.
)");
fusion_prof.def_property_readonly(
"fusion_id", [](FusionProfile& self) { return self.fusion_id; }, R"(
Returns the fusion id of the fusion profile.
)");
fusion_prof.def_property_readonly(
"segments", [](FusionProfile& self) { return self.segments; }, R"(
Returns the segments in the fusion profile.
)");
fusion_prof.def_property_readonly(
"cuda_evt_time_ms",
[](FusionProfile& self) { return self.cuda_evt_time_ms; },
R"(
Returns the CUDA event time in milliseconds of the fusion profile.
)");
fusion_prof.def_property_readonly(
"host_time_ms", [](FusionProfile& self) { return self.host_time_ms; }, R"(
Returns the host time in milliseconds of the fusion profile.
)");
fusion_prof.def_property_readonly(
"compile_time_ms",
[](FusionProfile& self) { return self.compile_time_ms; },
R"(
Returns the compile time in milliseconds of the fusion profile.
)");
fusion_prof.def_property_readonly(
"kernel_time_ms",
[](FusionProfile& self) { return self.kernel_time_ms; },
R"(
Returns the kernel time in milliseconds of the fusion profile.
)");
fusion_prof.def_property_readonly(
"effective_bandwidth_gbs",
[](FusionProfile& self) { return self.effective_bandwidth_gbs; },
R"(
Returns the effective bandwidth in gigabytes per second of the fusion profile.
)");
fusion_prof.def_property_readonly(
"percentage_peak_bandwith",
[](FusionProfile& self) { return self.percentage_peak_bandwidth; },
R"(
Returns the percentage of peak bandwidth of the fusion profile.
)");
fusion_prof.def_property_readonly(
"input_bytes", [](FusionProfile& self) { return self.input_bytes; }, R"(
Returns the input bytes of the fusion profile.
)");
fusion_prof.def_property_readonly(
"output_bytes", [](FusionProfile& self) { return self.output_bytes; }, R"(
Returns the output bytes of the fusion profile.
)");
fusion_prof.def_property_readonly(
"kernel_profiles",
[](FusionProfile& self) { return self.kernel_profiles; },
R"(
Returns the kernel profiles of the fusion profile.
)");
}
const FusionProfile& get_fusion_profile() {
const FusionProfile& profile = FusionProfiler::profile();
NVF_ERROR(
profile.fusion_id != -1,
"Something went wrong with Fusion Profiling as an illegal fusion_id "
"was returned!")
NVF_ERROR(
profile.segments > 0,
"Something went wrong with Fusion Profiling as no kernel segments were "
"profiled!")
return profile;
}
class PythonProfiler {
public:
PythonProfiler(bool auto_scheduled = false)
: auto_scheduled_(auto_scheduled) {}
PythonProfiler* start() {
ProfilerOptionsGuard::getCurOptions().set(ProfilerOption::Enable);
if (!auto_scheduled_) {
FusionProfiler::start();
FusionProfiler::createSegments(1);
}
return this;
}
void stop() {
if (!auto_scheduled_) {
FusionProfiler::segment(0).scheduler("user");
FusionProfiler::stop();
}
ProfilerOptionsGuard::getCurOptions().unset(ProfilerOption::Enable);
}
private:
//! Automatically scheduled fusions use the FusionExecutorCache.
//! The FusionExecutorCache will trigger FusionProfiler, if it is enabled.
bool auto_scheduled_ = true;
};
void bindProfiler(py::module& nvfuser) {
py::class_<PythonProfiler> profiler(nvfuser, "PythonProfiler");
profiler.def(py::init<bool>(), py::arg("auto_scheduled") = true, R"(
Create a new PythonProfiler.
Parameters
----------
auto_scheduled : bool, optional
Whether the fusion is automatically scheduled.
)");
profiler.def(
"__enter__",
&PythonProfiler::start,
py::return_value_policy::reference_internal);
profiler.def(
"__exit__",
[&](PythonProfiler& self,
py::object exc_type,
py::object exc_value,
py::object traceback) { self.stop(); });
profiler.def_property_readonly(
"profile",
[&](PythonProfiler& self) { return get_fusion_profile(); },
py::return_value_policy::reference,
R"(
Returns the FusionProfile for the profiled fusion.
Returns:
FusionProfile
)");
nvfuser.def(
"get_fusion_profile",
&get_fusion_profile,
py::return_value_policy::reference,
R"(
Returns the FusionProfile for the profiled fusion.
Returns:
FusionProfile
)");
nvfuser.def(
"reset_profiler",
&FusionProfiler::reset,
R"(Resets FusionProfiler so it can be used again.)");
}
} // namespace
void bindProfile(py::module& nvfuser) {
bindFusionProfile(nvfuser);
bindProfiler(nvfuser);
}
} // namespace nvfuser::python