wi5ddlZddlmZddlZddlZ ddlZddlm Z n#e $r e dwxYwddl Z ddl mZmZdZdZdZd Zed d d gZed dddddZeddejDZedddddZidejDZddddddiZedeDejD]LZdeZ ded dd!ed dd"ed diZ!ed#e!DMd$Z"d%Z#d-d&Z$d.d(Z%d)Z&d*Z'd+Z(e)d,kr e(dSdS)/N) namedtuple) NegationQueryzPFailed to import hatchet. `pip install llnl-hatchet` to get the correct version.)COMPUTE_METADATA_SCOPE_NAME TritonHookcng}|r~|D]z}|}|D]a}|dd}|||fvr||dznb{n |ddzg}t |dkrt d|d|S)N(r (inc)zMetric z< is not found. Use the --list flag to list available metrics)lowersplitstripappendlen RuntimeError)metrics raw_metricsretmetric raw_metricraw_metric_no_units f/root/.openclaw/workspace/chatterbox_venv_py311/lib/python3.11/site-packages/triton/profiler/viewer.pymatch_available_metricsrs C *  F\\^^F)   %/%5%5c%:%:1%=%C%C%E%E%K%K%M%M"j*<===JJzH4555E>  1~() 3xx1}}iViiijjj Jctj|}|d}tj|}|||fS)N)jsonloadpopht GraphFrame from_literalshow_metric_columns)filedatabase device_infogfs rget_raw_metricsr&sLyH,,q//K  # #H - -B r%%'' 44rc tjd|jdg}|D]2}||D]%}|||d}|||d}|||d}tjD]}|d|k} || } d|| jvr#d } |d kr<|d kr d |d z z } ni|dkr d|d z z } nZ|dkr|dz |zdz dzdz|d z z } n<|dkr$|dkr d|d z z } n'|dks|dkrd|d z z } nt d||j| dfxx| d|d | z z cc<ڐ'4|S)Nmin_timeindexcolumnsarchnum_sms clock_rate device_idflopsrCUDA80g2C89gbB90rgx:AimBHIPgfx90ag//bBgfx941gfx942g?y"CzUnsupported device type: ) pd DataFramer+r flops_widthr, ValueErrorlocfillna) dfr$min_time_flops device_type device_indexr-r.r/widthidx device_frames max_flopss rget_min_time_flopsrK&s\#RX |LLLN"ll ' 4 l lL{+L9&AD!+.|r+rA) rCr$min_time_bytesrErFrHrIrMrNpeak_bandwidths rget_min_time_bytesrSGs\#RX |LLLN"[[ ' 4 [ [L[/\1CsGM +K 8 FGZ [ #K0>{KI]->>DqHN  sJ / / /=3IN3Z Z / / / /  [ r FactorDictnamefactortimergMbP?gư>g& .>)ztime/sztime/msztime/usztime/nsavg_timec i|] \}}d|| S)avg_).0keyvalues r r_Vs$.u.u.uzsE|c||U.u.u.urrPgeAr8)zbyte/szgbyte/sztbyte/sci|] }|t Sr[)bytes_factor_dictr\r]s rr_r_Zs+333  333rzflop/szgflop/sztflop/sc:i|]}|tdtS)r1)rTdefault_flop_factor_dictrbs rr_r_as5111  *W6 7 7111rr1flopz/sgfloptflopcDi|]}|tttSr[)rT factor_name factor_dictrbs rr_r_fs$fffCc:k;#G#Gfffrcg}|jd}fd}|D]@}|dkrt|j|}t|j|} ||j} | d|dt | z |jd<t j|jj|df<| d|tvr~t|} | j } | j } t| gd}|j|||jz | |z |j|d<| |d-|tj vrstj dz|dd z}||jtj |z |j|d<| |d|t j vrt j dz|dd z}||j|jd z t j |z |j|d<t j|jj||df<| |d[|d}|d} t#|d kr|d }|d krt%d |t| gd}|j|}|j|jd}||z d z|j| d<| | dt| gd}| |B|S)Nr0cttjgd}tjdz|ddddz}||tj|zS)Nr/rr))rtime_factor_dictrUr rV)rCtime_metric_name time_unitrs rget_time_secondsz(derive_metrics..get_time_secondsmsu24D4I3JKXXYZ[%*S03C3I3I#3N3Nq3Q3W3WX[3\3\]^3__ "#&6&=i&HHHrutilr)z util (inc)rr rmrcount%zUnsupported unit gY@z/% (inc)) dataframeisnarSrKcombinemaxnpnanrAr derivable_metricsrUrVrror avg_time_factor_dictrr@iloc)r%rrr$derived_metricsinternal_frame_indicesrrrrQrDtime_secderivable_metric metric_namemetric_factor_dictmatched_metric_namemetric_time_unitmetric_name_and_unit metric_unit single_frametotals ` rderive_metricsrisO\+6;;==IIIII )<)< V  / kJJN/ kJJN'' 55H)7 )C)K)KN[eLfhk)l)low)wBL &EGVBL 3\A B  " "< 0 0 0 0 ( ( (08 */K!1!8 "9;-"U"UVW"X /1|cARARSTAUU /?/? /M/MPRP\]dPe/e/C/JK[/\0]BLF*** +JL&BL 35F5F5FF G  " "f#4#4#4 5 5 5 5#)<<#4#4 .q1K'((1,,215 #%%$%F%F%FGGG&={m[&Y&YZ[&\#!|,?@  %89>qA:F:NRW9W 5556&&+'?'?'?@@@@&={m[&Y&YZ[&\#&&':;;;; rc$|dkr*|jdd|jd<n_|dkr*|jdd|jd<n/|dkr)|jdd|jd<|S)Nfile_function_linerUc8|ddS)Nrmr xs rzformat_frames..AGGCLLQSDTr function_linec8|ddS)N:rrrs rrzformat_frames..rr file_functioncj|ddddS)Nrmr@rrrs rrzformat_frames..s+AGGCLLQSDTDZDZ[^D_D_`aDbr)rvapply)r%formats r format_framesrs %%%!|F399:T:TUU V ? " "!|F399:T:TUU V ? " "!|F399:b:bcc V Irc.|rd|d}||d}|r,d|d}t|}||d}ddtdig}||d}|r d |d |ig}||d}|S) Nz1 MATCH ("*")->(".", p)->("*") WHERE p."name" =~ "z" T)squashz* MATCH (".", p)->("*") WHERE p."name" =~ "rUz^(?!z).**z>= )filterrr)r%includeexclude thresholdrqueryinclusion_querys r filter_framesrs+YYuTY * *+o.. YYuTY * *=8=== > ?E 5 & &B+v0Y0012 YYuTY * * Irdc t|d5}t|\}} } t||}t| dks Jd|t ||| | }t |||||d}t||d|dt||ddddS#1swxYwYdS)Nrrz"No metrics found in the input fileTF) metric_column expand_namedepth render_header) openr&rrupdate_inclusive_columnsrrprinttree emit_warnings) rfilenamerrrrrfr%rr$s rparsers) h   #'6q'9'9$K 2v & &;!###%I### ##%%% Wk;GG 2wGAJ G G bggGUZ_g``aaab'""" # # # # # # # # # # # # # # # # # #sB6CCCcd|vr=|jdj}tj|}|dkrt ddSdSdS)Nz bytes (inc)rzZWarning: Negative byte values detected, this is usually the result of a datatype overflow )rvvaluesrznanminr)r%r byte_valuesmin_byte_values rrrs[l=18 ;// A   o p p p p p    rcLt|d5}t|\}}}td|rV|D]S}|dd}td|TddddS#1swxYwYdS)NrzAvailable metrics:rrz- )rr&rr r r ) file_namer_rrrs r show_metricsrs i  1+A..; "###  1) 1 1 %/%5%5c%:%:1%=%C%C%E%E%K%K%M%M"/-//0000 111111111111111111sA;BB Bc ~tjdtj}|dddd|dd tdd |d d tdd |ddtdd |ddt dd |ddt dd |ddtgddd|\}}t|dks Jd |d!}|j r|j d"nd}|j }|j }|j }|j}|j} |r|rt!d#|jrt%|dS|rt'||||||| dSdS)$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 - avg_time/s, avg_time/ms, avg_time/us, avg_time/ns: time / count - flop[<8/16/32/64>]/s, gflop[<8/16/32/64>]/s, tflop[<8/16/32/64>]/s: flops / time - byte/s, gbyte/s, tbyte/s: bytes / time - util: max(sum(flops) / peak_flops_time, sum(bytes) / peak_bandwidth_time) - /%%: frame(metric) / sum(metric). Only availble for inclusive metrics (e.g. 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. )typedefaultrz-iz --includeaFind frames that match the given regular expression and return all nodes in the paths that pass through the matching frames. For example, the following command will display all paths that contain frames that contains "test": ``` proton-viewer -i ".*test.*" path/to/file.json ``` z-ez --excludezExclude frames that match the given regular expression and their children. For example, the following command will exclude all paths that contain frames that contains "test": ``` proton-viewer -e ".*test.*" path/to/file.json ``` z-tz --thresholdzrExclude frames(kernels) whose metrics are below the given threshold. This filter only applies on the first metric.z-dz--depthrz The depth of the tree to displayz-fz--format)fullrrrra!Formatting the frame name. - full: include the path, file name, function name and line number. - file_function_line: include the file name, function name and line number. - function_line: include the function name and line number. - file_function: include the file name and function name. )rchoicesrrrzMust specify a file to readr,z'Cannot specify both include and exclude)argparseArgumentParserRawTextHelpFormatter add_argumentstrfloatintparse_known_argsrrr rrrrrr@listrr) argparserargs target_argsrrrrrrrs rmainrsm'B 5I                          |    /  js,l,l,l"2244D+ {  q "? AI)-?dl  %%%4GlGlGI JE [FD7DBCCC yNY N gy'7IufMMMMMNNr__main__)NNNN)NNNrN)*r collectionsrrpandasr=hatchetr hatchet.queryr ImportErrornumpyrztriton.profiler.hookrrrr&rKrSrTrorVitemsr}rakeysr|rdupdater?rGrirjrrrrrrr__name__r[rrrs,"""""" j+++++++jjj +h i iijHHHHHHHH"555B   Z vx&8 9 9 :ftPTae&f&fgg!z*.u.uUeUlUrUrUtUt.u.u.uvvJw1QU(V(VWW33#*//11333 &q*c:tL11(--//111222 #hhE!%//K#%###Q(9(9(9(93@Q@Q@Q@QSWXKffS^ScScSeSefffgggg333l. # # # #qqq111^N^N^NB zDFFFFFs 0