shSSKrSSKrSSKrSSKrSSKrSSKrSSKJrJr Sq SSjr Sr Sr Sr SrS r/qS \S \4S jrSS jrg)N)profileProfilerActivitycg)Nrt/Users/tiagomarins/Projetos/claudeai/copy_bank/venv/lib/python3.13/site-packages/torch/_functorch/benchmark_utils.py synchronizer src bUcS/nUS/:wa=[RR5(a[RRqUc0nUc0nU [R"S5 [ S5Hn U"U40UD6 [5 M [R"S5 [ R"5n [ U5Hn U"U40UD6 [5 M [ R"5n SSS5 W W - n [SSU0UD6n U [5 [R"S5 [ U5Hn U"U40UD6 [5 M SSS5 SSS5 W RU5 U $!,(df  N=f!,(df  N:=f!,(df  NC=f)a Output the chrome trace of running f(input, **kwargs_for_f) with [optimize_ctx] [num_runs] times to [trace_filename]. [activities] are the activities that the profiler will record, e.g. ProfilerActivity.CUDA. Return total runtime without the profiler Outputs to trace_filename Ncudacpui9 activitiesr) torchr is_availabler manual_seedrangetime perf_counterrexport_chrome_trace)finputtrace_filename optimize_ctxrnum_runsdevices kwargs_for_fkwargs_for_profiler_t0t1timingprofs rdump_chrome_tracer#sd*(5'ejj5577jj,,  "   $qA e $| $ M $    xA e $| $ M!    "WF  >J >*= >$  M   d #8_%(<( % ? ^, M- \ ? >s2B!E>F AFF > F  F F  F.cR[U5n[R"U5nUSnU$)N traceEvents)openjsonload)filenamerdataeventss rget_chrome_trace_eventsr,Ks' XA 99Q.is_mm_conv_eventtsT eFm # (v& (E&M) (%-'  r)r=r8)r+rO gpu_events sorted_eventsr3s rget_sorted_gpu_mm_conv_eventsrSssF 'v.JM&& U# rr) total_lengthc[U5n/qUH<nSU;aM USS:XdMSUSS;dM$[RUS5 M> US-n[U5n[ U5U- n[ U5n[ U5U- nXW4$)a Process the chrome traces outputs by the pytorch profiler to compute GPU Utilization and percent of times spent on matmul and convolution Args: filename(str): Name of chrome traces file produced by pytorch profiler total_length(float): total length of the process without profiler in second Return: tuple: (GPU Utilization, percent of time spent on matmul and convolution) rJprocess_labelsGPUargslabelsr.g.A)r,r1r8r=rGrS)r)rTr+r3r< utilizationsorted_gpu_mm_conv_eventsmm_conv_utilizations rcompute_utilizationr]s%X .FH    =, ,%-:Q1Q OOE%L )   #%L-f501L@K =f E&'@ALP  ++rc h[RRU5nU(d$[R"U5 [ SU-5 Uc[ R "5n[RRX$S-5n[UUUU[R/US/S9n[Xx5upX4$)am Benchmark the GPU Utilization and percent of time spent on matmul and convolution operations of running f(input, **kwargs_for_f) with [optimize_ctx] [num_runs] times. It will produce a chrome trace file in trace_folder/trace_file_name.json Example: ``` def f(a): return a.sum() a = torch.rand(2**20, device="cuda") utilization, mm_conv_utilization = benchmark_utilization(f, a, "tmp", trace_file_name = "tmp_chrome_trace") ``` Args: f: function to benchmark input: input to :attr:`f` trace_folder: name of the folder to store the chrome trace optimize_ctx: the context in which f will run trace_file_name: name of the dumped chrome trace file, default to "tmp_chrome_trace" num_runs: number of times to run f, excluding the warm-up runs, default to 1. Return: tuple: (GPU Utilization, percent of time spent on matmul and convolution) zcreate folder z.jsonr )rr) ospathexistsmakedirsprint contextlib nullcontextjoinr#rCUDAr]) rr trace_folderrtrace_file_namerisExistchrome_trace_file_namerTrZr\s rbenchmark_utilizationrlsNggnn\*G  L! -.!--/ WW\\,'8QR$     L(;($K  ++r)r@NNN)Ntmp_chrome_tracer@)rdr'r:r_rrtorch.profilerrrr r#r,r4r=rGrSr1strfloatr]rlrrrrqsz  4  7tD $ ,#,U,L&  =,r