U Kc%@sddlZddlZddlZddlZddlmZmZddaGdddZddd Z d d Z d d Z ddZ ddZ ddZgaeedddZdddZdS)N)profileProfilerActivitycCsdSNrrrB/tmp/pip-unpacked-wheel-gikjz4vx/functorch/_src/benchmark_utils.py synchronize src@seZdZddZddZdS) NullContextcCsdSrr)selfrrr __enter__szNullContext.__enter__cCsdSrr)r exc_typeexc_valexc_tbrrr__exit__szNullContext.__exit__N)__name__ __module__ __qualname__r rrrrrr src  Cs*|dkrdg}|dgkr*tjr*tjja|dkr6i}|dkrBi}|jtdtdD]} ||f|tqZtdt} t|D]} ||f|tqt} W5QRX| | } tfd|i|F} |6ttdt|D]} ||f|tqW5QRXW5QRX| || 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 activities) torchrZ is_availablerZ manual_seedrangetime perf_counterrZexport_chrome_trace)finputZtrace_filename optimize_ctxrnum_runsdevicesZ kwargs_for_fZkwargs_for_profiler_t0t1ZtimingZprofrrrdump_chrome_traces:           r#cCst|}t|}|d}|S)NZ traceEvents)openjsonload)filenamerdataeventsrrrget_chrome_trace_eventsFs r*cCs(d|ko&|dtko&d|ko&|ddkS)NpidphX)gpu_pidseventrrris_gpu_compute_eventMsr1cCs2g}|D]}t|sq||qt|dddS)NcSs|dS)Ntsr)xrrrXz'get_sorted_gpu_events..)key)r1appendsorted)r)sorted_gpu_eventsr0rrrget_sorted_gpu_eventsRs  r:cCst|dkrdS|d}|d|d}|d}|ddD]>}t|d|}|d|d}|t||d}t||}q<|S)Nrr2Zdurr)lenmax)r9r0Zcurrent_end_timeZtotal_duration start_timeZend_timerrr get_duration[s  r>cCs6dd}t|}g}|D]}||s&q||q|S)NcSs8d|ko6d|dkp6d|dkp6d|dkp6d|dkS)NnameZgemmconvZcutlassZwgradrr/rrris_mm_conv_eventjs    z7get_sorted_gpu_mm_conv_events..is_mm_conv_event)r:r7)r)rAZ gpu_eventsZ sorted_eventsr0rrrget_sorted_gpu_mm_conv_eventsis rB)r' total_lengthcCst|}ga|D]8}d|krq|ddkrd|ddkrt|dq|d}t|}t||}t|}t||}||fS)a Process the chrome traces outputs by the pytorch profiler to compute GPU Utilization and percent of times spent on matmal 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 matmal and convolution) r?Zprocess_labelsZGPUargslabelsr+g.A)r*r.r7r:r>rB)r'rCr)r0r9 utilizationZsorted_gpu_mm_conv_eventsmm_conv_utilizationrrrcompute_utilizationys   rHtmp_chrome_tracec Csvtj|}|s&t|td||dkr4t}tj||d}t||||tj g|dd}t ||\} } | | fS)a Benchmark the GPU Utilization and percent of time spent on matmal 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 matmal and convolution) zcreate folder Nz.jsonr)rr) ospathexistsmakedirsprintrjoinr#rZCUDArH) rrZ trace_folderrZtrace_file_namerZisExistZchrome_trace_file_namerCrFrGrrrbenchmark_utilizations    rP)rNNN)NrIr)rrJr%rZtorch.profilerrrrrr#r*r1r:r>rBr.strfloatrHrPrrrrs& 1  "