gg>zSSKrSSKJrJr \"SS5\"SS5\"SS5/rS rS rS rS rS r Sr Sr g)N) MetricRequestRequestedMetricsParserz;sm__instruction_throughput.avg.pct_of_peak_sustained_activeinstruction_throughputz0sm__inst_issued.avg.pct_of_peak_sustained_activeinst_issued_avgz0sm__inst_issued.max.pct_of_peak_sustained_activeinst_issued_maxcg)NSlowPipeLimiterr 3nsight-compute-2025.1.1/sections/SlowPipeLimiter.pyget_identifierr #s r cg)NzSlow Pipe Limiterr r r r get_namer&s r cg)Nz&Slow pipe limiting compute utilizationr r r r get_descriptionr)s 3r cg)NComputeWorkloadAnalysisr r r r get_section_identifierr,s $r cS/$)NComputer r r r get_parent_rules_identifiersr0s ;r cUSR5nUSR5nX2- U- nSnXP;a'[RRnX@U-S-nXg4$[RRnUS-nXg4$)a+Estimate potential speedup from decreasing the usage of slow pipes. The performance improvement is approximated as the relative part of instructions not issued. In case the compute (SM) throughput was collected, the above approximation can be improved by weighing it with the achieved throughput. rrcompute_throughput_normalizedd)valueNvRules IFrontendSpeedupType_GLOBALSpeedupType_LOCAL)parent_weightsmetricsrrimprovement_localcompute_throughput_name speedup_typeimprovement_percents r get_estimated_speedupr&4s/0668O/0668O(:oM=0((;; /AX2YY\__  ,,((:: /#5  ,,r c d[R"U5nURS5RS5nUR 5n[ X5R [5nURS5nUSR5nUSR5nUSR5nSn Sn Sn S n X- n Xi:aXz:aX:aUR[RRS RX}X5S 5n[XT5unnURXU5 UR!XSR#5U[RR$S 5 UR!XSR#5U[RR&S RU55 gggg)NrrrrrPz See the @url:Kernel Profiling Guide:https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#metrics-decoder@ for the workloads handled by each pipeline.zIt is possible that a slow pipeline is preventing better workload performance. The average pipeline utilization of {:.1f}% is {:.1f}% lower than the maximum utilization of {:.1f}%. Try moving compute to other pipelines, e.g. from fp64 to fp32 or int.{}z Slow PipelinezVThe higher the instruction throughput the more likely is the impact of a slow pipelinezSIncrease the average pipeline utilization towards the maximum utilization ({:.1f}%))r get_context range_by_idx action_by_idxfrontendrparserequested_metricsreceive_dict_from_parentrmessagerMsgType_MSG_OPTIMIZATIONformatr&speedup focus_metricnameSeverity_SEVERITY_LOWSeverity_SEVERITY_HIGH)handlectxactionfer!r sm_busyrrno_bound_thresholdissued_avg_thresholddiff_thresholddoc_msg pipe_diffmsg_idr$ speedup_values r applyrFLs   f %C   a . .q 1F B$V4::;LMG00;N./557G/0668O/0668ONsG1I$)OT]TnG--FF VO I? \ 'rHs_2BOQijDFWXDFWX4%-0{r