PhddlZddlZddlZddlZddlZddlmZmZda ddZ dZ dZ dZ dZ d Zgad ed efd Zdd Zy)N)profileProfilerActivitycy)NrkC:\Users\daisl\Desktop\realtime-object-detection\venv\Lib\site-packages\torch/_functorch/benchmark_utils.py synchronizer src |dg}|dgk7r8tjjrtjja|i}|i}|5tjdt dD]} ||fi|ttjdt j} t |D]} ||fi|tt j} ddd  z } tdd|i|5} |5ttjdt |D]} ||fi|t dddddd j|| S#1swYxYw#1swY0xYw#1swY4xYw)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#sa(5'ejj557jj,,  "   $qA e $| $ M $    xA e $| $ M!    "WF  >J >*= >$  M   d #8_%(<( % ? ^, M- \ ? >s2BE+FAE7F+E47F <FF cPt|}tj|}|d}|S)N traceEvents)openjsonload)filenamerdataeventss rget_chrome_trace_eventsr,?s' XA 99Q _eElh6 _45= _USW[\_M__rcjg}|D]}t|s|j|!t|dS)Nc |dS)Ntsr)xs rz'get_sorted_gpu_events..Qs1T7r)key)r4appendsorted)r+sorted_gpu_eventsr3s rget_sorted_gpu_eventsr>Ks<#E*   ' #): ;;rct|dk(ry|d}|d|dz}|d}|ddD]:}t|d|}|d|dz}|t||z dz}t||}<|S)Nrr7dur)lenmax)r=r3current_end_timetotal_duration start_timeend_times r get_durationrHTs " a ET{U5\15\N"12&t&67 ;u-'#h.CQ*GG/: ' rchd}t|}g}|D]}||s |j||S)NcRd|vxr"d|dvxsd|dvxsd|dvxsd|dvS)Nnamegemmconvcutlasswgradrr2s ris_mm_conv_eventz7get_sorted_gpu_mm_conv_events..is_mm_conv_eventcsX_FeFm$;%^vv?V%^'0E&M'A%^ELPUV\P]E] _r)r>r;)r+rP gpu_events sorted_eventsr3s rget_sorted_gpu_mm_conv_eventsrSbsD_'v.JM& U# rr) total_lengthct|}ga|D]3}d|vr|ddk(sd|ddvstj|d5|dz}t|}t ||z }t |}t ||z }||fS)a Process the chrome traces outputs by the pytorch profiler to compute GPU Utilization and percent of times spent on matmul and convolution Args: filename(str): Name of chrome traces file produced by pytorch profiler total_length(float): total length of the process without profiler in second Return: tuple: (GPU Utilization, percent of time spent on matmul and convolution) rKprocess_labelsGPUargslabelsr.g.A)r,r1r;r>rHrS)r)rTr+r3r= utilizationsorted_gpu_mm_conv_eventsmm_conv_utilizations rcompute_utilizationr]rs%X .FH    =, ,%-:Q1Q OOE%L )   #%L-f501L@K =f E&'@ALP + ++rc dtjj|}|s#tj|t d|z|t j }tjj||dz}t||||tjg|d}t||\} } | | fS)a Benchmark the GPU Utilization and percent of time spent on matmul and convolution operations of running f(input, **kwargs_for_f) with [optimize_ctx] [num_runs] times. It will produce a chrome trace file in trace_folder/trace_file_name.json Example: ``` def f(a): return a.sum() a = torch.rand(2**20, device="cuda") utilization, mm_conv_utilization = benchmark_utilization(f, a, "tmp", trace_file_name = "tmp_chrome_trace") ``` Args: f: function to benchmark input: input to :attr:`f` trace_folder: name of the folder to store the chrome trace optimize_ctx: the context in which f will run trace_file_name: name of the dumped chrome trace file, default to "tmp_chrome_trace" num_runs: number of times to run f, excluding the warm-up runs, default to 1. Return: tuple: (GPU Utilization, percent of time spent on matmul and convolution) zcreate folder z.jsonr )rr) ospathexistsmakedirsprint contextlib nullcontextjoinr#rCUDAr]) rr trace_folderrtrace_file_namerisExistchrome_trace_file_namerTrZr\s rbenchmark_utilizationrls@ggnn\*G  L! -.!--/ WW\\,'8QR$Q/E|&6&;&;%rHrSr1strfloatr]rlrrrrqsf 4 TUKO.b` <   ,#,U,D-,r