`iddlZddlZddlZddlZddlZddlZddlmZmZda ddZ dZ dZ dZ d Zd Zgad ed efd Z ddZdS)N)profileProfilerActivitycdS)Nrt/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/torch/_functorch/benchmark_utils.py synchronizer sDrc L|dg}|dgkr/tjrtjja|i}|i}|5tjdt dD]} ||fi|ttjdt j} t |D]} ||fi|tt j} dddn #1swxYwY| | z } tdd|i|5} |5ttjdt |D]} ||fi|t dddn #1swxYwYdddn #1swxYwY| || S)a0 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$s*(5'ej5577j,  "   ! ! $q  A Ae $ $| $ $ $ MMMM $   x  A Ae $ $| $ $ $ MMMM    ! ! ! ! ! ! ! ! ! ! ! ! ! ! !"WF  > >J >*= > >$    MMM  d # # #8__  %((<(((                  ^,,, MsJB!C33C7:C7FA E-! F-E1 1F4E1 5FF Fc\t|}tj|}|d}|S)N traceEvents)openjsonload)filenamerdataeventss rget_chrome_trace_eventsr-Ks) XA 9Q<\s`((#E**    '''' #).is_mm_conv_eventtsT eFm # (v& (E&M) (%-'  r)r>r9)r,rO gpu_events sorted_eventsr4s rget_sorted_gpu_mm_conv_eventsrRssg   'v..JM$$&&  U#### rr* total_lengthcFt|}ga|D]C}d|vr|ddkr0d|ddvr t|dD|dz}t|}t ||z }t |}t ||z }||fS)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-r2r9r>rGrR)r*rSr,r4r= utilizationsorted_gpu_mm_conv_eventsmm_conv_utilizations rcompute_utilizationr\s%X . .FH**     =, , ,%-:Q1Q1Q OOE%L ) ) )#%L-f55011L@K =f E E&'@AALP + ++rtmp_chrome_tracec rtj|}|s&tj|t d|z|t j}tj||dz}t||||tj g|dg}t||\} } | | fS)a 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 Nz.jsonr )rr) ospathexistsmakedirsprint contextlib nullcontextjoinr$rCUDAr\) rr trace_folderrtrace_file_namerisExistchrome_trace_file_namerSrYr[s rbenchmark_utilizationrlsVgnn\**G / L!!! -...!-// W\\,'8QRR$    L(; (($K$ + ++r)r NNN)Nr]r )rdr(r;r_rrtorch.profilerrrr r$r-r5r>rGrRr2strfloatr\rlrrrrps1  44444444    7777tDDD   $ ,#,U,,,,L&  A,A,A,A,A,A,r