9jSSKJr SSKrSSKrSSKrSSKr SSK J r "SS5r S0S4SS\RSS .S jjrS rg) ) annotationsN)runtimecp\rSrSrSrSrS Sjr\S Sj5r\S Sj5r \ S5r SSjr S r S rg )_PerfCaseResult aAn obscure object encompassing timing results recorded by :func:`~cupyx.profiler.benchmark`. Simple statistics can be obtained by converting an instance of this class to a string. .. warning:: This API is currently experimental and subject to change in future releases. cURS:XdeURS[U5S-:XdeURSS:deXlX lX0lg)Nr)ndimshapelenname_ts_devices)selfrtsdevicess N/home/wildlama/miniconda3/lib/python3.13/site-packages/cupyx/profiler/_time.py__init___PerfCaseResult.__init__sTww!||xx{c'lQ....xx{Q  c URSS9$)ziReturns a string representation of the object. Returns: str: A string representation of the object. Tshow_gputo_strrs r__repr___PerfCaseResult.__repr__s {{D{))rc URS$)zA :class:`numpy.ndarray` of shape ``(n_repeat,)``, holding times spent on CPU in seconds. These values are delta of the host-side performance counter (:func:`time.perf_counter`) between each repeat step. rrrs r cpu_times_PerfCaseResult.cpu_times'sxx{rc URSS$)zA :class:`numpy.ndarray` of shape ``(len(devices), n_repeat)``, holding times spent on GPU in seconds. These values are measured using ``cudaEventElapsedTime`` with events recorded before/after each repeat step. r Nr!rs r gpu_times_PerfCaseResult.gpu_times1sxx|rc6URS:XdeURS:deUS-nSRXR55nURS:a@USRUR 5UR 5UR 55- nU$)Nr rg.Az {}: {:9.03f} usz2 +/- {:6.03f} (min: {:9.03f} / max: {:9.03f}) us)r sizeformatmeanstdminmax) device_namett_usss r_to_str_per_item _PerfCaseResult._to_str_per_item;svv{{vvzz3w ! ( (iik B 66A: ELL DHHJ 4 4Arc bURSURS5/nU(a^[UR5HEup4UR URSR U5URSU-55 MG SR UR SRU55$)NCPUrzGPU-{}r z {:<20s}:{} )r2r enumeraterappendr)rjoin)rrresultsids rr_PerfCaseResult.to_strGs(( <= !$--0))(//!*<*.((1q5/;<1""499chhw.?@@rc URSS9$)NTrrrs r__str___PerfCaseResult.__str__Ps{{D{))r)rrrN)returnstr)rAz_numpy.ndarray)F)__name__ __module__ __qualname____firstlineno____doc__rrpropertyr"r% staticmethodr2rr?__static_attributes__rrrr sX *  A*rrrKi' )rn_warmup max_durationrc Uc URnUc[RR54n[ U5(d [ S5e[ U[5(d [ S5e[ U[5(d [ S5e[ U[5(d [ S5e[ U[5(d [ S5e[ U[5(d [ S5e[R"U5(d [ S5e[ U[5(d [ S5e[XX#XEXg5$) aTiming utility for measuring time spent by both CPU and GPU. This function is a very convenient helper for setting up a timing test. The GPU time is properly recorded by synchronizing internal streams. As a result, to time a multi-GPU function all participating devices must be passed as the ``devices`` argument so that this helper knows which devices to record. A simple example is given as follows: .. code-block:: py import cupy as cp from cupyx.profiler import benchmark def f(a, b): return 3 * cp.sin(-a) * b a = 0.5 - cp.random.random((100,)) b = cp.random.random((100,)) print(benchmark(f, (a, b), n_repeat=1000)) .. note:: For IPython/Jupyter users, the ``%gpu_timeit`` magic provides a convenient way to use this function. See the documentation for :doc:`/user_guide/performance` for details. Args: func (callable): a callable object to be timed. args (tuple): positional arguments to be passed to the callable. kwargs (dict): keyword arguments to be passed to the callable. n_repeat (int): number of times the callable is called. Increasing this value would improve the collected statistics at the cost of longer test time. name (str): the function name to be reported. If not given, the callable's ``__name__`` attribute is used. n_warmup (int): number of times the callable is called. The warm-up runs are not timed. max_duration (float): the maximum time (in seconds) that the entire test can use. If the taken time is longer than this limit, the test is stopped and the statistics collected up to the breakpoint is reported. devices (tuple): a tuple of device IDs (int) that will be timed during the timing test. If not given, the current device is used. Returns: :class:`~cupyx.profiler._time._PerfCaseResult`: an object collecting all test results. z#`func` should be a callable object.z`args` should be of tuple type.z `kwargs` should be of dict type.z `n_repeat` should be an integer.z`name` should be a string.z `n_warmup` should be an integer.z)`max_duration` should be given in secondsz!`devices` should be of tuple type)rC_cupycuda get_device_idcallable ValueError isinstancetupledictintrB_numpyisreal_repeat)funcargskwargsn_repeatrrMrNrs r benchmarkr`Ts j |}}::++-/ D>>>?? dE " ":;; fd # #;<< h $ $;<< dC 566 h $ $;<< == & &DEE gu % %<==  Fdl MMrcT/n/n UHn [R"5n [R"U 5 UR[R R R55 U R[R R R55 [R"U 5 M [U5H n U"U0UD6 M [X5Hgup[R"5n [R"U 5 U R5 [R"U 5 U R5 Mi /nUV s/sHn /PM nn Sn[U5GHn [X5HWup[R"5n [R"U 5 U R5 [R"U 5 MY [R"5nU"U0UD6 [R"5nUU- nURU5 [X5HWup[R"5n [R"U 5 U R5 [R"U 5 MY [X5HWup[R"5n [R"U 5 U R5 [R"U 5 MY [[X55H?un unn[R RUU5S-nXRU5 MA U[R"5U- - nUU:dGM O [ R""U/U-[ R$S9n['UUUS9$![R"U 5 f=f![R"U 5 f=fs sn f![R"U 5 f=f![R"U 5 f=f![R"U 5 f=f)NrgMbP?)dtype)r)r getDevice setDevicer8rPrQstreamEventrangeziprecord synchronize_time perf_counterr7get_elapsed_timerYasarrayfloat64r)r\r]r^r_rrMrNrevents_1events_2r; prev_deviceeventdevicer"r%durationt1t2cpu_timeev1ev2gpu_timers rr[r[sBHH '')  +   a OOEJJ--335 6 OOEJJ--335 6   k *8_ dfX/ '')  +   f % LLN   k * 0I%&XXI&H 8_ 3ME!++-K /!!&) !!+. 4   ! df    !7" 3ME!++-K /!!&) !!+. 4!3ME!++-K /!!&)!!#!!+. 4's8'>?MAzSzz223&N9 O&O&O4-&PN69OO14P P') __future__rmath_mathtimerknumpyrYcupyrPcupy_backends.cuda.apirrinfr`r[rKrrrsO"*E*E*Rb5MM BUYYMM`F6r