U h%@sddlZddlmZddlZddlZddlZddlm Z m Z ddZ ddZ dd Z d d Zed d dgZeddddddZedddddZedddddZddejDddejDZddZd d!Zd"d#Zd$d%Zed&kredS)'N) namedtuple)COMPUTE_METADATA_SCOPE_NAME TritonHookcCslg}|rZ|D]J}|}|D]8}|dd}|||fkr||dq qq n|dgd}|S)N(r (inc))lowersplitstripappend)metrics raw_metricsretmetric raw_metricraw_metric_no_unitrH/var/www/html/venv/lib/python3.8/site-packages/triton/profiler/viewer.pymatch_available_metrics s  rcCs.t|}|d}tj|}|||fS)N)jsonloadpophtZ GraphFrameZ from_literalZshow_metric_columns)fileZdatabase device_infogfrrrget_raw_metricss   rc Csvtjd|jdgd}|D]V}||D]F}|||d}|||d}|||d}tjD]}|d|k} || } d|| jkrqbd } |d kr|d krd |d } n>|dkrd|d } n(|dkr|d|ddd|d } nR|dkr4|dkrd|d } n |dks&|dkrBd|d } ntd||j| df| d|d | 7<qbq&q|S)Nmin_timeindexcolumnsarchnum_sms clock_rateDeviceIdflopsrCUDA80g2CZ89gbBZ90rgx:AimBZHIPZgfx90ag//bBZgfx941Zgfx942g?y"CzUnsupported device type: ) pd DataFramer rZ flops_widthr! ValueErrorlocZfillna) dfrmin_time_flops device_type device_indexr"r#r$widthidx device_framesZ max_flopsrrrget_min_time_flops s6   "  0r7c Cstjd|jdgd}|D]x}||D]j}|d|k}||}|||d}|||d}d||dd } |j|df|d | 7<q$q|S) Nrrrr%memory_clock_rate bus_widthg@@r)bytes)r,r-r r/) r0rmin_time_bytesr2r3r5r6r8r9Zpeak_bandwidthrrrget_min_time_bytesAs  "r= FactorDictnamefactortimergMbP?gư>g& .>)time/sztime/msztime/usztime/nsr&geAr+)zflop/szgflop/sztflop/sr;)zbyte/szgbyte/sztbyte/scCsi|] }|tqSr)flops_factor_dict.0keyrrr TsrGcCsi|] }|tqSr)bytes_factor_dictrDrrrrGVscCsg}g}ttjg|d}tjd|dddd}|D]H}|dkrt|j|} t|j|} |j|tj|tjd} | d| dt | |jd <| d q@|t kr t |} | j} | j}t| g|d}|j||j|tj||||j|d <| |d q@|tjkrtjd|dd}|j|tj|tj||j|d <| |d q@| |q@|rt||}||S) Nr/rr)utilrBrz util (inc)r) rtime_factor_dictr?rr=Z dataframer7r@combinemaxr derivable_metrics)rr r rZderived_metricsZoriginal_metricsZtime_metric_nameZ time_unitrr<r1Ztime_secZderiveable_metricZ metric_nameZmetric_factor_dictZmatched_metric_nameZmetric_time_unitrrrderive_metrics[sD"        rPc Cst|d}t|\}}} t|dks.td|t|||| }|sL|rz|r\d|dn|} dd| ig} |j| dd }ddtdig} |j| dd }|rd|dd |ig} |j| dd }t|j |d|d d W5QRXdS) Nrrz"No metrics found in the input filez^(?!z).**r?T)Zsquashz>= F)Z metric_columnZ expand_namedepthZ render_header) openrlenAssertionErrorZupdate_inclusive_columnsrPfilterrprinttree) r filenameincludeexclude thresholdrSfrr rZ name_filterqueryrrrparse~s  r`c Cslt|dX}t|\}}}td|rT|D](}|dd}td|q*W5QRdSQRXdS)NrQzAvailable metrics:rrz- )rTrrXrr r) file_namer^_r rrrrr show_metricss rcc Cstjdtjd}|jddddd|jdd tdd d |jd d tddd |jddtddd |jddtddd |jddtddd |\}}t|dkst d|d}|j r|j dnd}|j }|j }|j}|j}|r|rtd|jrt|n|rt||||||dS)Nz,Performance data viewer for proton profiles.) descriptionformatter_classz-lz--list store_truea`List available metrics. Metric names are case insensitive and ignore units. Derived metrics can be created when source metrics are available. - time/s, time/ms, time/us, time/ns: time - flop/s, gflop/s, tflop/s: flops / time - byte/s, gbyte/s, tbyte/s: bytes / time - util: max(sum(flops) / peak_flops_time, bytes / peak_bandwidth_time)) )actionhelpz-mz --metricszAt maximum two metrics can be specified, separated by comma. There are two modes: 1) Choose the output metric to display. It's case insensitive and ignore units. 2) Derive a new metric from existing metrics. )typedefaultrhz-iz --includez?Include frames(kernels) that match the given regular expressionz-ez --excludez?Exclude frames(kernels) that match the given regular expressionz-tz --thresholdzrExclude frames(kernels) whose metrics are below the given threshold. This filter only applies on the first metric.z-dz--depthdz The depth of the tree to displayrzMust specify a file to readr,z'Cannot specify both include and exclude)argparseArgumentParserRawTextHelpFormatter add_argumentstrfloatintparse_known_argsrUrVr rr[r\r]rSr.listrcr`) Z argparserargsZ target_argsrar r[r\r]rSrrrmainsv     rw__main__)rm collectionsrrZpandasr,ZhatchetrZtriton.profiler.hookrrrrr7r=r>rLrCrHr@keysrOrPr`rcrw__name__rrrrs4 ! # M