o o hEG@sddlZddlmZddlZddlZz ddlZddlm Z Wn e y)e dwddl Z ddl mZmZddZdefd d Zd d Zd dZddZedddgZeddddddZedddejDZeddddddZedd dejDZed!dd"d#d$Zid%dejDZdd"d#d&Ze d'deDej!D]'Z"d(e"Z#d)e"d*dd+e"d*d"d,e"d*d#iZ$e d-de$Dqd.d/Z%d0d1Z&dAd2d3Z'd4d5Z(dBd8d9Z)dCd:d;Z*dd?Z,e-d@kre,dSdS)DN) namedtuple) NegationQueryzPFailed to import hatchet. `pip install llnl-hatchet` to get the correct version.)COMPUTE_METADATA_SCOPE_NAME TritonHookcCsg}t|ts |g}|r>|D]/}|}||D]$}||vr dnd}|dd}|||fvr<|||nqqt|dkrLtd|d|S)N (inc)(rzMetric z< is not found. Use the --list flag to list available metrics) isinstancelistlowersplitstripappendlen RuntimeError)metricsinclusive_metricsexclusive_metricsretmetric raw_metricsuffixraw_metric_no_unitrj/var/www/html/construction_image-detection-poc/venv/lib/python3.10/site-packages/triton/profiler/viewer.pymatch_available_metricss"    rdatabasecs8fddg}|D]}|}|dur||q |S)Ncsd|vr|S|ddtkrdSt|ddkr"t|ddkr"dS|dg}g}|D]}|}|dur;||q,t|dksHt|dkrN||d<|SdS)Nframenamerrchildren)rrgetr)noder new_childrenchild new_childremove_frame_helperrrr&&s"   z*remove_frames..remove_frame_helper)r)r new_databaser!new_noderr%r remove_frames s  r)csRt|}t|}|d}tj|}|fdd|jj D}|||fS)Ncsg|]}|vr|qSrr).0rrrr Fz#get_raw_metrics..) jsonloadr)popht GraphFrame from_literalshow_metric_columns dataframecolumns)filer device_infogfrrr,rget_raw_metrics@s    r;c Cstjd|jdgd}|D]}||D]}|||d}|||d}|||d}tjD]}|d|k} || } d|| jvrDq/d } |d kr|d krUd |d } nX|dkr`d|d } nM|dkru|d|ddd|d } n8|dkr|d|dd|d } n%|dkr|dkrd|d } n|dks|dkrd|d } ntd||j| df| d|d | 7<q/qq |S) Nmin_timeindexr7archnum_sms clock_rate device_idflopsrCUDA80g2C89gbB90rgx:AimB100i@@@g.AHIPgfx90ag//bBgfx941gfx942g?y"CzUnsupported device type: ) pd DataFramer?r flops_widthr7 ValueErrorlocfillna) dfr9min_time_flops device_type device_indexr@rArBwidthidx device_frames max_flopsrrrget_min_time_flopsJsB   " ,r`c Cstjd|jdgd}|D]<}||D]5}|d|k}||}|||d}|||d}d||dd } |j|df|d | 7<qq |S) Nr<r=r>rCmemory_clock_rate bus_widthrMrGbytes)rRrSr?rV) rXr9min_time_bytesrZr[r]r^rarbpeak_bandwidthrrrget_min_time_bytesms   rg FactorDictrfactortimer*gMbP?gư>g& .>)ztime/sztime/msztime/usztime/nsavg_timecCi|] \}}d||qSavg_rr+keyvaluerrr |srrcpu_time)z cpu_time/sz cpu_time/msz cpu_time/usz cpu_time/ns avg_cpu_timecCrlrmrrorrrrrsrdgeArK)zbyte/szgbyte/sztbyte/scCsi|]}|tqSr)bytes_factor_dictr+rprrrrrs )zflop/szgflop/sztflop/scCsi|]}|tdtqS)rD)rhdefault_flop_factor_dictrvrrrrrsrDflopz/sgfloptflopcCsi|]}|tttqSr)rh factor_name factor_dictrvrrrrrr.csg}fdd}|D]G}|dkrJt|j|}t|j|} ||jdt} |jd} | d|dt| |jd<tj|jj | df<| dq |t vrt |} | j } | j }t| d}|j|||jdt|||j|d<| |dq |tj vs|tj vs|tj vs|tj vr|tj vp|tj v}|tj vp|tj v}|r|rtntn|rtnt}|rd nd} |j d |d d }||j| |}|r||jd }||j ||j|d<| |dq |d }|d} t|d krF|d }|d kr td|t| d}|j|}d}d|vr,d}|j|jd}n|j|}||d|j||<| ||q t| d}| |q |D]}|drf|j |qV|j |qV|S)NcsDt|d}|jd|dddd}|||j|S)Nr/rr*))rrr ri)rXrr|time_metric_name time_unitrrrrget_time_secondss"z(derive_metrics..get_time_secondsutilrjrCr=rrrsr}r*z count (inc)%zUnsupported unit rz(inc)gY@)rgr6r`time_factor_dictisnacombinemaxnpnanrVrderivable_metricsrrircpu_time_factor_dictavg_time_factor_dictavg_cpu_time_factor_dictr rrUilocsumendswith inc_metrics exc_metrics)r:rrrr9derived_metricsrrrerYtime_secinternal_frame_indicesderivable_metric metric_namemetric_factor_dictmatched_metric_nameis_cpuis_avgr|metric_time_unit time_valuemetric_name_and_unit metric_unit single_framertotalderived_metricrrrderive_metricsst             rcCsr|dkr|jddd|jd<|S|dkr&|jddd|jd<|S|dkr7|jddd|jd<|S) Nfile_function_linercS|ddS)Nr}r xrrrzformat_frames.. function_linecSr)N:rrrrrrrr file_functioncSs6|ddddd|ddddS)Nr}rrr@rrrrrrs6)r6apply)r:formatrrr format_framess  rcCsl|rd|d}|j|dd}|r"d|d}t|}|j|dd}|r4d|d|ig}|j|dd}|S)Nz1 MATCH ("*")->(".", p)->("*") WHERE p."name" =~ "z" T)squashz* MATCH (".", p)->("*") WHERE p."name" =~ "*z>= )filterr)r:includeexclude thresholdrqueryinclusion_queryrrr filter_framessrcCs:d|vr|jdj}t|}|dkrtddSdSdS)Nz bytes (inc)rzZWarning: Negative byte values detected, this is usually the result of a datatype overflow )r6valuesrnanminprint)r:r byte_valuesmin_byte_valuerrr emit_warningss   rdFcCst||}t|j|d|dd|r]td|d|jj|dgdd}tdt|D]0}t|j|dd krF|j|ddd d n|j|d}td ||j||dq,t ||dS) NTF) metric_column expand_namedepth render_headerzSorted kernels by metric r)by ascendingr*rrz...z {:105} {:.4}) rrtreer6 sort_valuesrangerrrr)r:rrr print_sorted sorted_dfrow kernel_namerrr print_trees    rc Cst|d6}t|\}}}} t||dksJd|t||||| }t|||||d}||fWdS1s>wYdS)Nrrz"No metrics found in the input file)openr;rupdate_inclusive_columnsrr) rfilenamerrrfr:rrr9rrrparses $rcCst|dS}t|\}}}}td|r+|D]}|dd}td|qtd|rH|D]}|dd}td|q3WddSWddS1s[wYdS)NrzAvailable inclusive metrics:rrz- zAvailable exclusive metrics:)rr;rr r r ) file_namer_rrrrrrr show_metricss  "rcCstjdtjd}|jddddd|jdd tdd d |jd d tddd |jddtddd |jddtddd |jddtddd |jddtgdddd|jddd d!d"|jd#d$tdd%d |\}}t|d&kssJd'|d(}|j r|j d)nd}|j }|j }|j }|j}|j} |j} |j} |r|rtd*|jrt|dS|rt|||||\} } | rt|| |||\}}| |} t| | || | 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 starting from 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. )rchoicesrrz--print-sortedFz6Sort output by metric value instead of chronologically)rrrz--diff-profilez-diffzCompare two profiles. When used as 'proton-viewer -m time -diff file1.log file2.log', computes the difference: file2['time'] - file1['time']r*zMust specify a file to readr,z'Cannot specify both include and exclude)argparseArgumentParserRawTextHelpFormatter add_argumentstrfloatintparse_known_argsrrr rrrrr diff_profilerrUr rrsubr) argparserargs target_argsrrrrrrrdiffrr:rgf2rrrrmain,s       r__main__)NNNN)rNF)NNN).r collectionsrr/pandasrRhatchetr2 hatchet.queryr ImportErrornumpyrtriton.profiler.hookrrrr)r;r`rgrhrriitemsrrrrukeysrrwupdaterTr\r{r|rrrrrrrr__name__rrrrsn    #     (I     q