gg} SSKrSSKJrJrJr \"SS5\"SS5/r\"SS\R SS 5\"S 5\"S S \R S5/r\"S S\R SS 5\"SS\R SS 5/r\"SS\R SS 5\"SS\R SS 5/r Sr Sr Sr Sr SrSrSrg)N) Importance MetricRequestRequestedMetricsParser*device__attribute_compute_capability_majorcc_major*device__attribute_compute_capability_minorcc_minorzlts__t_sectors.sum input_sectorsF#l1tex__m_l1tex2xbar_write_bytes.sumz2dram__bytes_read.sum.pct_of_peak_sustained_elapseddram_read_bandwidthzlts__gcomp_input_sectors.sumcompression_input_sectorsz0lts__average_gcomp_input_sector_success_rate.pctcompression_success_ratedzlrc__ilc_input_sectors.sumz.lrc__average_ilc_input_sector_success_rate.pctcg)NMemoryL2Compressionr7nsight-compute-2025.1.1/sections/MemoryL2Compression.pyget_identifierrQs rcg)NzMemory L2 Compressionrrrrget_namerTs "rcg)Nz;Detection of inefficient use of L2 Compute Data Compressionrrrrget_descriptionrWs Hrcg)NMemoryWorkloadAnalysis_Chartrrrrget_section_identifierrZs )rcS/$)NMemoryrrrrget_parent_rules_identifiersr]s :rc6USR5nUSR5nUSR5nUSR5nUS:Xa[RRS4$SU- SX- - -U-n[RRU4$)Nr r rr r)valueNvRules IFrontendSpeedupType_LOCALSpeedupType_GLOBAL)metricsr r rdram_bandwidth_percentimprovement_percents rget_estimated_speedupr*`s '(C D J J LO,224M&'ABHHJ$%:;AAC  22A55 % % (8 8 :  !    / /1D DDrc z[R"U5nURS5RS5nUR 5n[ X5R [5nUSR5S-USR5-nUS:ag[nUS:a[nU[- n[ X5R U5nUSnUcgSn US R5n U S:GaS U -n US R5n UR5n X:Ga:S RX5nURRUS 5n[!U5unnUR#UUU5 UR%XS R'5U [RR(S5 U S:aLUR%XS R'5U [RR(SRU 55 UR%XSR'5USR5[RR*S5 USR5nUS:GaISU-U :Ga>SRU U- 5nURRUS5n[!U5unnUR#UUU5 UR%XS R'5U [RR(S5 U S:aLUR%XS R'5U [RR(SRU 55 UR%XSR'5USR5[RR*S5 gggg)Nrr r PZr r g@@ra Out of the {:.1f} bytes sent to the L2 Compression unit only {:.2f}% were successfully compressed. To increase this success rate, consider marking only those memory regions as compressible that contain the most zero values and/or expose the most homogeneous values.zLow Compression Ratez%Increase the compression success ratezPIncrease the number of compressed sectors towards all L2 input sectors ({:,.0f})r zOThe higher the DRAM peak read utilization the more can be gained by compressionr g@aThe access patterns for writes to compressible memory are not well suited for the L2 Compression unit. As a consequence, {:.1f}x the data written to the L2 cache has to be communicated to the L2 Compression unit. Try maximizing local coherence for the write operations to compressible memory. For example, avoid writes with large strides as they lead to partial accesses to many L2 cache lines. Instead, try accessing fewer overall cache lines by modifying many values per cache line with each warp's execution of a write operation.zAccess Pattern)r# get_context range_by_idx action_by_idxfrontendrparserequested_metrics_baser"%requested_metrics_compression_generic#requested_metrics_compression_gh100requested_metricsformatmessager$MsgType_MSG_OPTIMIZATIONr*speedup focus_metricnameSeverity_SEVERITY_HIGHSeverity_SEVERITY_LOW)handlectxactionfe cc_metricsccrequested_metrics_for_chipr'input_sectors_metriclow_compression_thresholdr compression_input_bytesrr r:msg_id speedup_type speedup_valuel1tex_write_bytess rapplyrOqs   f %C   a . .q 1F B'7==>TUJ J  % % '" ,z*/E/K/K/M MB Bw!F Rx%H""33$V4::;UVG"?3# " '(C D J J L 1$"&)B"B#*+E#F#L#L#N ,224  $ ?biijA\GZZ 1 1 J JGUklF*?*H 'L- JJv|] ; OOF,F$G$L$L$NPhjqj{j{kSkSU| }q 0K(L(Q(Q(SUnpwqBqBqYqYfmmn{|~ OOF,A$B$G$G$I7ShKiKoKoKqsztEtEt[t[a c$$IJPPR  !.?(?BY(YmttuL M ^ u^ _ GZZ 1 1 J JGUefF*?*H 'L- JJv|] ; OOF,F$G$L$L$NPhjqj{j{kSkSU| }q 0K(L(Q(Q(SUnpwqBqBqYqYfmmn{|~ OOF,A$B$G$G$I7ShKiKoKoKqsztEtEt[t[a c)Z !-%r)r#RequestedMetricsrrrr5OPTIONALr8r6r7rrrrrr*rOrrrrRs'2NN> K> K &9L9LdTYZ78<  &#  :"  )%($#  8"  '#$!#I*E"9cr