ggSSKrSSKJrJrJr \"SS\R SS5\"SS\R SS5/rSrS rS r S r S r S r Sr g)N) Importance MetricRequestRequestedMetricsParser&sm__maximum_warps_per_active_cycle_pcttheoretical_occupancyFz1sm__warps_active.avg.pct_of_peak_sustained_activeachieved_occupancycg)NAchievedOccupancyr 5nsight-compute-2025.1.1/sections/AchievedOccupancy.pyget_identifierr"s r cg)NzAchieved Occupancyr r r r get_namer%s r cg)Nz"Analysis of the Achieved Occupancyr r r r get_descriptionr(s /r cg)N Occupancyr r r r get_section_identifierr+s r cS/$)NIssueSlotUtilizationr r r r get_parent_rules_identifiersr.s " ##r cUSR5nUSR5nX#- U- nSnXP;a0[RRn[ XU5nUS-nXh4$[RR nUS-nXh4$)a;Estimate potential speedup from increasing the achieved occupancy. The performance improvement is approximated as relative proportion of the difference of theoretical and achieved occupancy. In case it's available, the performance improvement can be upper-bounded by the speedup estimate of IssueSlotUtilization. rr"issue_slot_util_speedup_normalizedd)valueNvRules IFrontendSpeedupType_GLOBALminSpeedupType_LOCAL) parent_weightsmetricsrrimprovement_localparent_speedup_name speedup_typeimprovement_globalimprovement_percents r get_estimated_speedupr)2s$$;<BBD !56<<>.CG\\>,((;;   /1B 136  ,,((:: /#5  ,,r c [R"U5nURS5RS5nUR 5n[ X5R [5n[SUR555(agURS5nSnSnUR5[RR:XaUSR5nUSR5n X- n X::agSRXU5n UR![R"R$U 5n ['XT5upUR)XU5 UR+XSR-5U [R"R.S RU55 gUSnUSR5n /n[1UR355HQnURU5nX- n X:dMUR55RU5nUR7UU45 MS [9U5S:XagUR;S S S 9 S R[9U5S:XaSRUSS5O)SRSR=SUSS555U5n UR![R"R$U 5n g)Nrc3(# UHoSLv M g7f)Nr ).0metrics r apply..Ss 9(8fT>(8sraLoad imbalances can occur between warps within a block as well as across blocks of the same kernel. See the @url:CUDA Best Practices Guide:https://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html#occupancy@ for more details on optimizing occupancy. rrzThe difference between calculated theoretical ({:.1f}%) and measured achieved occupancy ({:.1f}%) can be the result of warp scheduling overheads or workload imbalances during the kernel execution. {}zGIncrease the achieved occupancy towards the theoretical limit ({:.1f}%)Tc US$)Nr )xs r apply..s!r )reversekeyaThe large difference between the calculated theoretical occupancy (per launch) and the measured achieved occupancy (of the entire workload) for some launches of this workload (e.g., {}) can be the result of warp scheduling overheads or workload imbalances during the kernel execution. {}r2z launch ID {}z launch IDs {}z, c3># UHn[US5v M g7f)rN)str)r,launchs r r.r/s5nSmc&)nnSms)r get_context range_by_idx action_by_idxfrontendrparserequested_metricsanyvaluesreceive_dict_from_parent workload_typeIActionWorkloadType_KERNELrformatmessagerMsgType_MSG_OPTIMIZATIONr)speedup focus_metricnameSeverity_SEVERITY_DEFAULTrange num_instancescorrelation_idsappendlensortjoin)handlectxactionfer#r"load_imbalance_adviceoccupancy_difference_thresholdrroccupancy_differencerImsg_idr& speedup_valuetheoretical_occupancy_metriclow_occupancy_launches instance_id launch_ids r applyrcMs   f %C   a . .q 1F B$V4::;LMG 9(8 999 001GHN 5&(" !D!DD '(? @ F F H$%9:@@B4I  A  &%;P G--FFP&;N&T#  67 (< = B B DFXZaZkZk[F[FHQHXHXYnHo p(//F'G$$%9:@@B!# !=!K!K!MNK$@$F$F{$S !#8#M $D8HHJPPQ\] &--y:O.PQO % &! +  ##Dn#E PQWPVGJKaGbfgGg%%&rfsu2NN: