1 | n/a | #! /usr/bin/env python3 |
---|
2 | n/a | # |
---|
3 | n/a | # Class for profiling python code. rev 1.0 6/2/94 |
---|
4 | n/a | # |
---|
5 | n/a | # Written by James Roskind |
---|
6 | n/a | # Based on prior profile module by Sjoerd Mullender... |
---|
7 | n/a | # which was hacked somewhat by: Guido van Rossum |
---|
8 | n/a | |
---|
9 | n/a | """Class for profiling Python code.""" |
---|
10 | n/a | |
---|
11 | n/a | # Copyright Disney Enterprises, Inc. All Rights Reserved. |
---|
12 | n/a | # Licensed to PSF under a Contributor Agreement |
---|
13 | n/a | # |
---|
14 | n/a | # Licensed under the Apache License, Version 2.0 (the "License"); |
---|
15 | n/a | # you may not use this file except in compliance with the License. |
---|
16 | n/a | # You may obtain a copy of the License at |
---|
17 | n/a | # |
---|
18 | n/a | # http://www.apache.org/licenses/LICENSE-2.0 |
---|
19 | n/a | # |
---|
20 | n/a | # Unless required by applicable law or agreed to in writing, software |
---|
21 | n/a | # distributed under the License is distributed on an "AS IS" BASIS, |
---|
22 | n/a | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, |
---|
23 | n/a | # either express or implied. See the License for the specific language |
---|
24 | n/a | # governing permissions and limitations under the License. |
---|
25 | n/a | |
---|
26 | n/a | |
---|
27 | n/a | import sys |
---|
28 | n/a | import os |
---|
29 | n/a | import time |
---|
30 | n/a | import marshal |
---|
31 | n/a | from optparse import OptionParser |
---|
32 | n/a | |
---|
33 | n/a | __all__ = ["run", "runctx", "Profile"] |
---|
34 | n/a | |
---|
35 | n/a | # Sample timer for use with |
---|
36 | n/a | #i_count = 0 |
---|
37 | n/a | #def integer_timer(): |
---|
38 | n/a | # global i_count |
---|
39 | n/a | # i_count = i_count + 1 |
---|
40 | n/a | # return i_count |
---|
41 | n/a | #itimes = integer_timer # replace with C coded timer returning integers |
---|
42 | n/a | |
---|
43 | n/a | class _Utils: |
---|
44 | n/a | """Support class for utility functions which are shared by |
---|
45 | n/a | profile.py and cProfile.py modules. |
---|
46 | n/a | Not supposed to be used directly. |
---|
47 | n/a | """ |
---|
48 | n/a | |
---|
49 | n/a | def __init__(self, profiler): |
---|
50 | n/a | self.profiler = profiler |
---|
51 | n/a | |
---|
52 | n/a | def run(self, statement, filename, sort): |
---|
53 | n/a | prof = self.profiler() |
---|
54 | n/a | try: |
---|
55 | n/a | prof.run(statement) |
---|
56 | n/a | except SystemExit: |
---|
57 | n/a | pass |
---|
58 | n/a | finally: |
---|
59 | n/a | self._show(prof, filename, sort) |
---|
60 | n/a | |
---|
61 | n/a | def runctx(self, statement, globals, locals, filename, sort): |
---|
62 | n/a | prof = self.profiler() |
---|
63 | n/a | try: |
---|
64 | n/a | prof.runctx(statement, globals, locals) |
---|
65 | n/a | except SystemExit: |
---|
66 | n/a | pass |
---|
67 | n/a | finally: |
---|
68 | n/a | self._show(prof, filename, sort) |
---|
69 | n/a | |
---|
70 | n/a | def _show(self, prof, filename, sort): |
---|
71 | n/a | if filename is not None: |
---|
72 | n/a | prof.dump_stats(filename) |
---|
73 | n/a | else: |
---|
74 | n/a | prof.print_stats(sort) |
---|
75 | n/a | |
---|
76 | n/a | |
---|
77 | n/a | #************************************************************************** |
---|
78 | n/a | # The following are the static member functions for the profiler class |
---|
79 | n/a | # Note that an instance of Profile() is *not* needed to call them. |
---|
80 | n/a | #************************************************************************** |
---|
81 | n/a | |
---|
82 | n/a | def run(statement, filename=None, sort=-1): |
---|
83 | n/a | """Run statement under profiler optionally saving results in filename |
---|
84 | n/a | |
---|
85 | n/a | This function takes a single argument that can be passed to the |
---|
86 | n/a | "exec" statement, and an optional file name. In all cases this |
---|
87 | n/a | routine attempts to "exec" its first argument and gather profiling |
---|
88 | n/a | statistics from the execution. If no file name is present, then this |
---|
89 | n/a | function automatically prints a simple profiling report, sorted by the |
---|
90 | n/a | standard name string (file/line/function-name) that is presented in |
---|
91 | n/a | each line. |
---|
92 | n/a | """ |
---|
93 | n/a | return _Utils(Profile).run(statement, filename, sort) |
---|
94 | n/a | |
---|
95 | n/a | def runctx(statement, globals, locals, filename=None, sort=-1): |
---|
96 | n/a | """Run statement under profiler, supplying your own globals and locals, |
---|
97 | n/a | optionally saving results in filename. |
---|
98 | n/a | |
---|
99 | n/a | statement and filename have the same semantics as profile.run |
---|
100 | n/a | """ |
---|
101 | n/a | return _Utils(Profile).runctx(statement, globals, locals, filename, sort) |
---|
102 | n/a | |
---|
103 | n/a | |
---|
104 | n/a | class Profile: |
---|
105 | n/a | """Profiler class. |
---|
106 | n/a | |
---|
107 | n/a | self.cur is always a tuple. Each such tuple corresponds to a stack |
---|
108 | n/a | frame that is currently active (self.cur[-2]). The following are the |
---|
109 | n/a | definitions of its members. We use this external "parallel stack" to |
---|
110 | n/a | avoid contaminating the program that we are profiling. (old profiler |
---|
111 | n/a | used to write into the frames local dictionary!!) Derived classes |
---|
112 | n/a | can change the definition of some entries, as long as they leave |
---|
113 | n/a | [-2:] intact (frame and previous tuple). In case an internal error is |
---|
114 | n/a | detected, the -3 element is used as the function name. |
---|
115 | n/a | |
---|
116 | n/a | [ 0] = Time that needs to be charged to the parent frame's function. |
---|
117 | n/a | It is used so that a function call will not have to access the |
---|
118 | n/a | timing data for the parent frame. |
---|
119 | n/a | [ 1] = Total time spent in this frame's function, excluding time in |
---|
120 | n/a | subfunctions (this latter is tallied in cur[2]). |
---|
121 | n/a | [ 2] = Total time spent in subfunctions, excluding time executing the |
---|
122 | n/a | frame's function (this latter is tallied in cur[1]). |
---|
123 | n/a | [-3] = Name of the function that corresponds to this frame. |
---|
124 | n/a | [-2] = Actual frame that we correspond to (used to sync exception handling). |
---|
125 | n/a | [-1] = Our parent 6-tuple (corresponds to frame.f_back). |
---|
126 | n/a | |
---|
127 | n/a | Timing data for each function is stored as a 5-tuple in the dictionary |
---|
128 | n/a | self.timings[]. The index is always the name stored in self.cur[-3]. |
---|
129 | n/a | The following are the definitions of the members: |
---|
130 | n/a | |
---|
131 | n/a | [0] = The number of times this function was called, not counting direct |
---|
132 | n/a | or indirect recursion, |
---|
133 | n/a | [1] = Number of times this function appears on the stack, minus one |
---|
134 | n/a | [2] = Total time spent internal to this function |
---|
135 | n/a | [3] = Cumulative time that this function was present on the stack. In |
---|
136 | n/a | non-recursive functions, this is the total execution time from start |
---|
137 | n/a | to finish of each invocation of a function, including time spent in |
---|
138 | n/a | all subfunctions. |
---|
139 | n/a | [4] = A dictionary indicating for each function name, the number of times |
---|
140 | n/a | it was called by us. |
---|
141 | n/a | """ |
---|
142 | n/a | |
---|
143 | n/a | bias = 0 # calibration constant |
---|
144 | n/a | |
---|
145 | n/a | def __init__(self, timer=None, bias=None): |
---|
146 | n/a | self.timings = {} |
---|
147 | n/a | self.cur = None |
---|
148 | n/a | self.cmd = "" |
---|
149 | n/a | self.c_func_name = "" |
---|
150 | n/a | |
---|
151 | n/a | if bias is None: |
---|
152 | n/a | bias = self.bias |
---|
153 | n/a | self.bias = bias # Materialize in local dict for lookup speed. |
---|
154 | n/a | |
---|
155 | n/a | if not timer: |
---|
156 | n/a | self.timer = self.get_time = time.process_time |
---|
157 | n/a | self.dispatcher = self.trace_dispatch_i |
---|
158 | n/a | else: |
---|
159 | n/a | self.timer = timer |
---|
160 | n/a | t = self.timer() # test out timer function |
---|
161 | n/a | try: |
---|
162 | n/a | length = len(t) |
---|
163 | n/a | except TypeError: |
---|
164 | n/a | self.get_time = timer |
---|
165 | n/a | self.dispatcher = self.trace_dispatch_i |
---|
166 | n/a | else: |
---|
167 | n/a | if length == 2: |
---|
168 | n/a | self.dispatcher = self.trace_dispatch |
---|
169 | n/a | else: |
---|
170 | n/a | self.dispatcher = self.trace_dispatch_l |
---|
171 | n/a | # This get_time() implementation needs to be defined |
---|
172 | n/a | # here to capture the passed-in timer in the parameter |
---|
173 | n/a | # list (for performance). Note that we can't assume |
---|
174 | n/a | # the timer() result contains two values in all |
---|
175 | n/a | # cases. |
---|
176 | n/a | def get_time_timer(timer=timer, sum=sum): |
---|
177 | n/a | return sum(timer()) |
---|
178 | n/a | self.get_time = get_time_timer |
---|
179 | n/a | self.t = self.get_time() |
---|
180 | n/a | self.simulate_call('profiler') |
---|
181 | n/a | |
---|
182 | n/a | # Heavily optimized dispatch routine for os.times() timer |
---|
183 | n/a | |
---|
184 | n/a | def trace_dispatch(self, frame, event, arg): |
---|
185 | n/a | timer = self.timer |
---|
186 | n/a | t = timer() |
---|
187 | n/a | t = t[0] + t[1] - self.t - self.bias |
---|
188 | n/a | |
---|
189 | n/a | if event == "c_call": |
---|
190 | n/a | self.c_func_name = arg.__name__ |
---|
191 | n/a | |
---|
192 | n/a | if self.dispatch[event](self, frame,t): |
---|
193 | n/a | t = timer() |
---|
194 | n/a | self.t = t[0] + t[1] |
---|
195 | n/a | else: |
---|
196 | n/a | r = timer() |
---|
197 | n/a | self.t = r[0] + r[1] - t # put back unrecorded delta |
---|
198 | n/a | |
---|
199 | n/a | # Dispatch routine for best timer program (return = scalar, fastest if |
---|
200 | n/a | # an integer but float works too -- and time.clock() relies on that). |
---|
201 | n/a | |
---|
202 | n/a | def trace_dispatch_i(self, frame, event, arg): |
---|
203 | n/a | timer = self.timer |
---|
204 | n/a | t = timer() - self.t - self.bias |
---|
205 | n/a | |
---|
206 | n/a | if event == "c_call": |
---|
207 | n/a | self.c_func_name = arg.__name__ |
---|
208 | n/a | |
---|
209 | n/a | if self.dispatch[event](self, frame, t): |
---|
210 | n/a | self.t = timer() |
---|
211 | n/a | else: |
---|
212 | n/a | self.t = timer() - t # put back unrecorded delta |
---|
213 | n/a | |
---|
214 | n/a | # Dispatch routine for macintosh (timer returns time in ticks of |
---|
215 | n/a | # 1/60th second) |
---|
216 | n/a | |
---|
217 | n/a | def trace_dispatch_mac(self, frame, event, arg): |
---|
218 | n/a | timer = self.timer |
---|
219 | n/a | t = timer()/60.0 - self.t - self.bias |
---|
220 | n/a | |
---|
221 | n/a | if event == "c_call": |
---|
222 | n/a | self.c_func_name = arg.__name__ |
---|
223 | n/a | |
---|
224 | n/a | if self.dispatch[event](self, frame, t): |
---|
225 | n/a | self.t = timer()/60.0 |
---|
226 | n/a | else: |
---|
227 | n/a | self.t = timer()/60.0 - t # put back unrecorded delta |
---|
228 | n/a | |
---|
229 | n/a | # SLOW generic dispatch routine for timer returning lists of numbers |
---|
230 | n/a | |
---|
231 | n/a | def trace_dispatch_l(self, frame, event, arg): |
---|
232 | n/a | get_time = self.get_time |
---|
233 | n/a | t = get_time() - self.t - self.bias |
---|
234 | n/a | |
---|
235 | n/a | if event == "c_call": |
---|
236 | n/a | self.c_func_name = arg.__name__ |
---|
237 | n/a | |
---|
238 | n/a | if self.dispatch[event](self, frame, t): |
---|
239 | n/a | self.t = get_time() |
---|
240 | n/a | else: |
---|
241 | n/a | self.t = get_time() - t # put back unrecorded delta |
---|
242 | n/a | |
---|
243 | n/a | # In the event handlers, the first 3 elements of self.cur are unpacked |
---|
244 | n/a | # into vrbls w/ 3-letter names. The last two characters are meant to be |
---|
245 | n/a | # mnemonic: |
---|
246 | n/a | # _pt self.cur[0] "parent time" time to be charged to parent frame |
---|
247 | n/a | # _it self.cur[1] "internal time" time spent directly in the function |
---|
248 | n/a | # _et self.cur[2] "external time" time spent in subfunctions |
---|
249 | n/a | |
---|
250 | n/a | def trace_dispatch_exception(self, frame, t): |
---|
251 | n/a | rpt, rit, ret, rfn, rframe, rcur = self.cur |
---|
252 | n/a | if (rframe is not frame) and rcur: |
---|
253 | n/a | return self.trace_dispatch_return(rframe, t) |
---|
254 | n/a | self.cur = rpt, rit+t, ret, rfn, rframe, rcur |
---|
255 | n/a | return 1 |
---|
256 | n/a | |
---|
257 | n/a | |
---|
258 | n/a | def trace_dispatch_call(self, frame, t): |
---|
259 | n/a | if self.cur and frame.f_back is not self.cur[-2]: |
---|
260 | n/a | rpt, rit, ret, rfn, rframe, rcur = self.cur |
---|
261 | n/a | if not isinstance(rframe, Profile.fake_frame): |
---|
262 | n/a | assert rframe.f_back is frame.f_back, ("Bad call", rfn, |
---|
263 | n/a | rframe, rframe.f_back, |
---|
264 | n/a | frame, frame.f_back) |
---|
265 | n/a | self.trace_dispatch_return(rframe, 0) |
---|
266 | n/a | assert (self.cur is None or \ |
---|
267 | n/a | frame.f_back is self.cur[-2]), ("Bad call", |
---|
268 | n/a | self.cur[-3]) |
---|
269 | n/a | fcode = frame.f_code |
---|
270 | n/a | fn = (fcode.co_filename, fcode.co_firstlineno, fcode.co_name) |
---|
271 | n/a | self.cur = (t, 0, 0, fn, frame, self.cur) |
---|
272 | n/a | timings = self.timings |
---|
273 | n/a | if fn in timings: |
---|
274 | n/a | cc, ns, tt, ct, callers = timings[fn] |
---|
275 | n/a | timings[fn] = cc, ns + 1, tt, ct, callers |
---|
276 | n/a | else: |
---|
277 | n/a | timings[fn] = 0, 0, 0, 0, {} |
---|
278 | n/a | return 1 |
---|
279 | n/a | |
---|
280 | n/a | def trace_dispatch_c_call (self, frame, t): |
---|
281 | n/a | fn = ("", 0, self.c_func_name) |
---|
282 | n/a | self.cur = (t, 0, 0, fn, frame, self.cur) |
---|
283 | n/a | timings = self.timings |
---|
284 | n/a | if fn in timings: |
---|
285 | n/a | cc, ns, tt, ct, callers = timings[fn] |
---|
286 | n/a | timings[fn] = cc, ns+1, tt, ct, callers |
---|
287 | n/a | else: |
---|
288 | n/a | timings[fn] = 0, 0, 0, 0, {} |
---|
289 | n/a | return 1 |
---|
290 | n/a | |
---|
291 | n/a | def trace_dispatch_return(self, frame, t): |
---|
292 | n/a | if frame is not self.cur[-2]: |
---|
293 | n/a | assert frame is self.cur[-2].f_back, ("Bad return", self.cur[-3]) |
---|
294 | n/a | self.trace_dispatch_return(self.cur[-2], 0) |
---|
295 | n/a | |
---|
296 | n/a | # Prefix "r" means part of the Returning or exiting frame. |
---|
297 | n/a | # Prefix "p" means part of the Previous or Parent or older frame. |
---|
298 | n/a | |
---|
299 | n/a | rpt, rit, ret, rfn, frame, rcur = self.cur |
---|
300 | n/a | rit = rit + t |
---|
301 | n/a | frame_total = rit + ret |
---|
302 | n/a | |
---|
303 | n/a | ppt, pit, pet, pfn, pframe, pcur = rcur |
---|
304 | n/a | self.cur = ppt, pit + rpt, pet + frame_total, pfn, pframe, pcur |
---|
305 | n/a | |
---|
306 | n/a | timings = self.timings |
---|
307 | n/a | cc, ns, tt, ct, callers = timings[rfn] |
---|
308 | n/a | if not ns: |
---|
309 | n/a | # This is the only occurrence of the function on the stack. |
---|
310 | n/a | # Else this is a (directly or indirectly) recursive call, and |
---|
311 | n/a | # its cumulative time will get updated when the topmost call to |
---|
312 | n/a | # it returns. |
---|
313 | n/a | ct = ct + frame_total |
---|
314 | n/a | cc = cc + 1 |
---|
315 | n/a | |
---|
316 | n/a | if pfn in callers: |
---|
317 | n/a | callers[pfn] = callers[pfn] + 1 # hack: gather more |
---|
318 | n/a | # stats such as the amount of time added to ct courtesy |
---|
319 | n/a | # of this specific call, and the contribution to cc |
---|
320 | n/a | # courtesy of this call. |
---|
321 | n/a | else: |
---|
322 | n/a | callers[pfn] = 1 |
---|
323 | n/a | |
---|
324 | n/a | timings[rfn] = cc, ns - 1, tt + rit, ct, callers |
---|
325 | n/a | |
---|
326 | n/a | return 1 |
---|
327 | n/a | |
---|
328 | n/a | |
---|
329 | n/a | dispatch = { |
---|
330 | n/a | "call": trace_dispatch_call, |
---|
331 | n/a | "exception": trace_dispatch_exception, |
---|
332 | n/a | "return": trace_dispatch_return, |
---|
333 | n/a | "c_call": trace_dispatch_c_call, |
---|
334 | n/a | "c_exception": trace_dispatch_return, # the C function returned |
---|
335 | n/a | "c_return": trace_dispatch_return, |
---|
336 | n/a | } |
---|
337 | n/a | |
---|
338 | n/a | |
---|
339 | n/a | # The next few functions play with self.cmd. By carefully preloading |
---|
340 | n/a | # our parallel stack, we can force the profiled result to include |
---|
341 | n/a | # an arbitrary string as the name of the calling function. |
---|
342 | n/a | # We use self.cmd as that string, and the resulting stats look |
---|
343 | n/a | # very nice :-). |
---|
344 | n/a | |
---|
345 | n/a | def set_cmd(self, cmd): |
---|
346 | n/a | if self.cur[-1]: return # already set |
---|
347 | n/a | self.cmd = cmd |
---|
348 | n/a | self.simulate_call(cmd) |
---|
349 | n/a | |
---|
350 | n/a | class fake_code: |
---|
351 | n/a | def __init__(self, filename, line, name): |
---|
352 | n/a | self.co_filename = filename |
---|
353 | n/a | self.co_line = line |
---|
354 | n/a | self.co_name = name |
---|
355 | n/a | self.co_firstlineno = 0 |
---|
356 | n/a | |
---|
357 | n/a | def __repr__(self): |
---|
358 | n/a | return repr((self.co_filename, self.co_line, self.co_name)) |
---|
359 | n/a | |
---|
360 | n/a | class fake_frame: |
---|
361 | n/a | def __init__(self, code, prior): |
---|
362 | n/a | self.f_code = code |
---|
363 | n/a | self.f_back = prior |
---|
364 | n/a | |
---|
365 | n/a | def simulate_call(self, name): |
---|
366 | n/a | code = self.fake_code('profile', 0, name) |
---|
367 | n/a | if self.cur: |
---|
368 | n/a | pframe = self.cur[-2] |
---|
369 | n/a | else: |
---|
370 | n/a | pframe = None |
---|
371 | n/a | frame = self.fake_frame(code, pframe) |
---|
372 | n/a | self.dispatch['call'](self, frame, 0) |
---|
373 | n/a | |
---|
374 | n/a | # collect stats from pending stack, including getting final |
---|
375 | n/a | # timings for self.cmd frame. |
---|
376 | n/a | |
---|
377 | n/a | def simulate_cmd_complete(self): |
---|
378 | n/a | get_time = self.get_time |
---|
379 | n/a | t = get_time() - self.t |
---|
380 | n/a | while self.cur[-1]: |
---|
381 | n/a | # We *can* cause assertion errors here if |
---|
382 | n/a | # dispatch_trace_return checks for a frame match! |
---|
383 | n/a | self.dispatch['return'](self, self.cur[-2], t) |
---|
384 | n/a | t = 0 |
---|
385 | n/a | self.t = get_time() - t |
---|
386 | n/a | |
---|
387 | n/a | |
---|
388 | n/a | def print_stats(self, sort=-1): |
---|
389 | n/a | import pstats |
---|
390 | n/a | pstats.Stats(self).strip_dirs().sort_stats(sort). \ |
---|
391 | n/a | print_stats() |
---|
392 | n/a | |
---|
393 | n/a | def dump_stats(self, file): |
---|
394 | n/a | with open(file, 'wb') as f: |
---|
395 | n/a | self.create_stats() |
---|
396 | n/a | marshal.dump(self.stats, f) |
---|
397 | n/a | |
---|
398 | n/a | def create_stats(self): |
---|
399 | n/a | self.simulate_cmd_complete() |
---|
400 | n/a | self.snapshot_stats() |
---|
401 | n/a | |
---|
402 | n/a | def snapshot_stats(self): |
---|
403 | n/a | self.stats = {} |
---|
404 | n/a | for func, (cc, ns, tt, ct, callers) in self.timings.items(): |
---|
405 | n/a | callers = callers.copy() |
---|
406 | n/a | nc = 0 |
---|
407 | n/a | for callcnt in callers.values(): |
---|
408 | n/a | nc += callcnt |
---|
409 | n/a | self.stats[func] = cc, nc, tt, ct, callers |
---|
410 | n/a | |
---|
411 | n/a | |
---|
412 | n/a | # The following two methods can be called by clients to use |
---|
413 | n/a | # a profiler to profile a statement, given as a string. |
---|
414 | n/a | |
---|
415 | n/a | def run(self, cmd): |
---|
416 | n/a | import __main__ |
---|
417 | n/a | dict = __main__.__dict__ |
---|
418 | n/a | return self.runctx(cmd, dict, dict) |
---|
419 | n/a | |
---|
420 | n/a | def runctx(self, cmd, globals, locals): |
---|
421 | n/a | self.set_cmd(cmd) |
---|
422 | n/a | sys.setprofile(self.dispatcher) |
---|
423 | n/a | try: |
---|
424 | n/a | exec(cmd, globals, locals) |
---|
425 | n/a | finally: |
---|
426 | n/a | sys.setprofile(None) |
---|
427 | n/a | return self |
---|
428 | n/a | |
---|
429 | n/a | # This method is more useful to profile a single function call. |
---|
430 | n/a | def runcall(self, func, *args, **kw): |
---|
431 | n/a | self.set_cmd(repr(func)) |
---|
432 | n/a | sys.setprofile(self.dispatcher) |
---|
433 | n/a | try: |
---|
434 | n/a | return func(*args, **kw) |
---|
435 | n/a | finally: |
---|
436 | n/a | sys.setprofile(None) |
---|
437 | n/a | |
---|
438 | n/a | |
---|
439 | n/a | #****************************************************************** |
---|
440 | n/a | # The following calculates the overhead for using a profiler. The |
---|
441 | n/a | # problem is that it takes a fair amount of time for the profiler |
---|
442 | n/a | # to stop the stopwatch (from the time it receives an event). |
---|
443 | n/a | # Similarly, there is a delay from the time that the profiler |
---|
444 | n/a | # re-starts the stopwatch before the user's code really gets to |
---|
445 | n/a | # continue. The following code tries to measure the difference on |
---|
446 | n/a | # a per-event basis. |
---|
447 | n/a | # |
---|
448 | n/a | # Note that this difference is only significant if there are a lot of |
---|
449 | n/a | # events, and relatively little user code per event. For example, |
---|
450 | n/a | # code with small functions will typically benefit from having the |
---|
451 | n/a | # profiler calibrated for the current platform. This *could* be |
---|
452 | n/a | # done on the fly during init() time, but it is not worth the |
---|
453 | n/a | # effort. Also note that if too large a value specified, then |
---|
454 | n/a | # execution time on some functions will actually appear as a |
---|
455 | n/a | # negative number. It is *normal* for some functions (with very |
---|
456 | n/a | # low call counts) to have such negative stats, even if the |
---|
457 | n/a | # calibration figure is "correct." |
---|
458 | n/a | # |
---|
459 | n/a | # One alternative to profile-time calibration adjustments (i.e., |
---|
460 | n/a | # adding in the magic little delta during each event) is to track |
---|
461 | n/a | # more carefully the number of events (and cumulatively, the number |
---|
462 | n/a | # of events during sub functions) that are seen. If this were |
---|
463 | n/a | # done, then the arithmetic could be done after the fact (i.e., at |
---|
464 | n/a | # display time). Currently, we track only call/return events. |
---|
465 | n/a | # These values can be deduced by examining the callees and callers |
---|
466 | n/a | # vectors for each functions. Hence we *can* almost correct the |
---|
467 | n/a | # internal time figure at print time (note that we currently don't |
---|
468 | n/a | # track exception event processing counts). Unfortunately, there |
---|
469 | n/a | # is currently no similar information for cumulative sub-function |
---|
470 | n/a | # time. It would not be hard to "get all this info" at profiler |
---|
471 | n/a | # time. Specifically, we would have to extend the tuples to keep |
---|
472 | n/a | # counts of this in each frame, and then extend the defs of timing |
---|
473 | n/a | # tuples to include the significant two figures. I'm a bit fearful |
---|
474 | n/a | # that this additional feature will slow the heavily optimized |
---|
475 | n/a | # event/time ratio (i.e., the profiler would run slower, fur a very |
---|
476 | n/a | # low "value added" feature.) |
---|
477 | n/a | #************************************************************** |
---|
478 | n/a | |
---|
479 | n/a | def calibrate(self, m, verbose=0): |
---|
480 | n/a | if self.__class__ is not Profile: |
---|
481 | n/a | raise TypeError("Subclasses must override .calibrate().") |
---|
482 | n/a | |
---|
483 | n/a | saved_bias = self.bias |
---|
484 | n/a | self.bias = 0 |
---|
485 | n/a | try: |
---|
486 | n/a | return self._calibrate_inner(m, verbose) |
---|
487 | n/a | finally: |
---|
488 | n/a | self.bias = saved_bias |
---|
489 | n/a | |
---|
490 | n/a | def _calibrate_inner(self, m, verbose): |
---|
491 | n/a | get_time = self.get_time |
---|
492 | n/a | |
---|
493 | n/a | # Set up a test case to be run with and without profiling. Include |
---|
494 | n/a | # lots of calls, because we're trying to quantify stopwatch overhead. |
---|
495 | n/a | # Do not raise any exceptions, though, because we want to know |
---|
496 | n/a | # exactly how many profile events are generated (one call event, + |
---|
497 | n/a | # one return event, per Python-level call). |
---|
498 | n/a | |
---|
499 | n/a | def f1(n): |
---|
500 | n/a | for i in range(n): |
---|
501 | n/a | x = 1 |
---|
502 | n/a | |
---|
503 | n/a | def f(m, f1=f1): |
---|
504 | n/a | for i in range(m): |
---|
505 | n/a | f1(100) |
---|
506 | n/a | |
---|
507 | n/a | f(m) # warm up the cache |
---|
508 | n/a | |
---|
509 | n/a | # elapsed_noprofile <- time f(m) takes without profiling. |
---|
510 | n/a | t0 = get_time() |
---|
511 | n/a | f(m) |
---|
512 | n/a | t1 = get_time() |
---|
513 | n/a | elapsed_noprofile = t1 - t0 |
---|
514 | n/a | if verbose: |
---|
515 | n/a | print("elapsed time without profiling =", elapsed_noprofile) |
---|
516 | n/a | |
---|
517 | n/a | # elapsed_profile <- time f(m) takes with profiling. The difference |
---|
518 | n/a | # is profiling overhead, only some of which the profiler subtracts |
---|
519 | n/a | # out on its own. |
---|
520 | n/a | p = Profile() |
---|
521 | n/a | t0 = get_time() |
---|
522 | n/a | p.runctx('f(m)', globals(), locals()) |
---|
523 | n/a | t1 = get_time() |
---|
524 | n/a | elapsed_profile = t1 - t0 |
---|
525 | n/a | if verbose: |
---|
526 | n/a | print("elapsed time with profiling =", elapsed_profile) |
---|
527 | n/a | |
---|
528 | n/a | # reported_time <- "CPU seconds" the profiler charged to f and f1. |
---|
529 | n/a | total_calls = 0.0 |
---|
530 | n/a | reported_time = 0.0 |
---|
531 | n/a | for (filename, line, funcname), (cc, ns, tt, ct, callers) in \ |
---|
532 | n/a | p.timings.items(): |
---|
533 | n/a | if funcname in ("f", "f1"): |
---|
534 | n/a | total_calls += cc |
---|
535 | n/a | reported_time += tt |
---|
536 | n/a | |
---|
537 | n/a | if verbose: |
---|
538 | n/a | print("'CPU seconds' profiler reported =", reported_time) |
---|
539 | n/a | print("total # calls =", total_calls) |
---|
540 | n/a | if total_calls != m + 1: |
---|
541 | n/a | raise ValueError("internal error: total calls = %d" % total_calls) |
---|
542 | n/a | |
---|
543 | n/a | # reported_time - elapsed_noprofile = overhead the profiler wasn't |
---|
544 | n/a | # able to measure. Divide by twice the number of calls (since there |
---|
545 | n/a | # are two profiler events per call in this test) to get the hidden |
---|
546 | n/a | # overhead per event. |
---|
547 | n/a | mean = (reported_time - elapsed_noprofile) / 2.0 / total_calls |
---|
548 | n/a | if verbose: |
---|
549 | n/a | print("mean stopwatch overhead per profile event =", mean) |
---|
550 | n/a | return mean |
---|
551 | n/a | |
---|
552 | n/a | #**************************************************************************** |
---|
553 | n/a | |
---|
554 | n/a | def main(): |
---|
555 | n/a | usage = "profile.py [-o output_file_path] [-s sort] scriptfile [arg] ..." |
---|
556 | n/a | parser = OptionParser(usage=usage) |
---|
557 | n/a | parser.allow_interspersed_args = False |
---|
558 | n/a | parser.add_option('-o', '--outfile', dest="outfile", |
---|
559 | n/a | help="Save stats to <outfile>", default=None) |
---|
560 | n/a | parser.add_option('-s', '--sort', dest="sort", |
---|
561 | n/a | help="Sort order when printing to stdout, based on pstats.Stats class", |
---|
562 | n/a | default=-1) |
---|
563 | n/a | |
---|
564 | n/a | if not sys.argv[1:]: |
---|
565 | n/a | parser.print_usage() |
---|
566 | n/a | sys.exit(2) |
---|
567 | n/a | |
---|
568 | n/a | (options, args) = parser.parse_args() |
---|
569 | n/a | sys.argv[:] = args |
---|
570 | n/a | |
---|
571 | n/a | if len(args) > 0: |
---|
572 | n/a | progname = args[0] |
---|
573 | n/a | sys.path.insert(0, os.path.dirname(progname)) |
---|
574 | n/a | with open(progname, 'rb') as fp: |
---|
575 | n/a | code = compile(fp.read(), progname, 'exec') |
---|
576 | n/a | globs = { |
---|
577 | n/a | '__file__': progname, |
---|
578 | n/a | '__name__': '__main__', |
---|
579 | n/a | '__package__': None, |
---|
580 | n/a | '__cached__': None, |
---|
581 | n/a | } |
---|
582 | n/a | runctx(code, globs, None, options.outfile, options.sort) |
---|
583 | n/a | else: |
---|
584 | n/a | parser.print_usage() |
---|
585 | n/a | return parser |
---|
586 | n/a | |
---|
587 | n/a | # When invoked as main program, invoke the profiler on a script |
---|
588 | n/a | if __name__ == '__main__': |
---|
589 | n/a | main() |
---|