gg SSKrSSKJrJr \"SS5\"SS5\"SS5\"S S 5\"S S 5\"S S5\"SS5\"SS5/rSrSrSrSrSr Sr Sr g)N) MetricRequestRequestedMetricsParserz@sm__sass_thread_inst_executed_op_ffma_pred_on.sum.peak_sustainedinst_executed_ffma_peakz@sm__sass_thread_inst_executed_op_dfma_pred_on.sum.peak_sustainedinst_executed_dfma_peakzEsmsp__sass_thread_inst_executed_op_fadd_pred_on.sum.per_cycle_elapsedinst_executed_faddzEsmsp__sass_thread_inst_executed_op_fmul_pred_on.sum.per_cycle_elapsedinst_executed_fmulzEsmsp__sass_thread_inst_executed_op_ffma_pred_on.sum.per_cycle_elapsedinst_executed_ffmazEsmsp__sass_thread_inst_executed_op_dadd_pred_on.sum.per_cycle_elapsedinst_executed_daddzEsmsp__sass_thread_inst_executed_op_dmul_pred_on.sum.per_cycle_elapsedinst_executed_dmulzEsmsp__sass_thread_inst_executed_op_dfma_pred_on.sum.per_cycle_elapsedinst_executed_dfmacg)N SOLFPRoofliner9nsight-compute-2025.1.1/sections/SpeedOfLight_Roofline.pyget_identifierr(s rcg)NRoofline Analysisrrrrget_namer+s rcg)Nz Floating Point Roofline Analysisrrrrget_descriptionr.s -rcg)NSpeedOfLight_RooflineChartrrrrget_section_identifierr1s 'rcS/$)NHighPipeUtilizationrrrrget_parent_rules_identifiersr4s ! ""rcXC- S:a[RRS4$X!U-- SXC- - -nSU;a$[RRnXPS-nXg4$[RRnUS-nXg4$)Nrfp64_pipeline_utilization_pctd)NvRules IFrontendSpeedupType_LOCALSpeedupType_GLOBAL)parent_weights achieved_fp32 achieved_fp64 peak_fp32 peak_fp64improvement_local speedup_typeimprovement_percents rget_estimated_speedupr.7s q   22A55&-*GH I !!'.8((;; /A`2aa  ,,((:: /#5  ,,rcB[R"U5nURS5RS5nUR 5n[ X5R [5nURS5nSUSR5-nSUSR5-nUSR5nUSR5n USR5n X-SU --n US R5n US R5n US R5nX-SU--nS nS nX- nUS:dUS:XaSOSnX- nUS:dUS:XaSOSnSRXg- 5nUSRUSU-USU-5- nSnUU:Ga!UU:GaUS- nUU- nUR[RRUS5n[X[XU5unnURUUU5 US:aUR!UUS R#5U [RR$S5 UR!UUS R#5U [RR$S5 UR!UUS R#5U[RR$S5 ggUU:a<UU:a6US- nUU- nUR[RRUS5ngUU- nUR[RR&US5ng)Nrrrrrrr r r r g333333?g333333?g{Gz?gz close to zWThe ratio of peak float (fp32) to double (fp64) performance on this device is {:.0f}:1.zs The workload achieved {}{:.0f}% of this device's fp32 peak performance and {}{:.0f}% of its fp64 peak performance.gY@z See the @url:Kernel Profiling Guide:https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#roofline@ for more details on roofline analysis.z If @section:ComputeWorkloadAnalysis:Compute Workload Analysis@ determines that this workload is fp64 bound, consider using 32-bit precision floating point operations to improve its performance.zFP64/32 UtilizationzDecrease fp64 ADD instructionszDecrease fp64 MUL instructionszDecrease fp64 FMA instructionsz If @section:SpeedOfLight:Speed Of Light@ analysis determines that this workload is compute bound, consider using integer arithmetic instead where applicable.zHigh FP Utilizationr)r" get_context range_by_idx action_by_idxfrontendrparserequested_metricsreceive_dict_from_parentvalueformatmessager#MsgType_MSG_OPTIMIZATIONr.speedup focus_metricnameSeverity_SEVERITY_HIGHMsgType_MSG_OK)handlectxactionfemetricsr&r)r*fp32_add_achievedfp32_mul_achievedfp32_fma_achievedr'fp64_add_achievedfp64_mul_achievedfp64_fma_achievedr(high_utilization_thresholdlow_utilization_thresholdachieved_fp64_pct fp64_prefixachieved_fp32_pct fp32_prefixr;message_profiling_guidemsg_idr, speedup_values rapplyrVLs   f %C   a . .q 1F B$V4::;LMG001FGNG56<<>>IG56<<>>I 45;;= 45;;= 45;;=%9A@QrXs2BTVopTVopY[opY[opY[opY[opY[opY[op .(#-*3\r