From 6bbdeeaf12c2bc896a855dc96cdc067f0a297a99 Mon Sep 17 00:00:00 2001 From: Christian Schulz Date: Sat, 21 Nov 2026 16:29:47 +0100 Subject: [PATCH] Directory edition of the plugin in claude-plugin/ Anthropic's plugin directory accepts only Markdown, text, JSON, YAML and images from organizations that are not Verified, no symbolic links, and no commands that download code. claude-plugin/ is such an edition, generated by scripts/build_claude_plugin.py: - skills/aae/SKILL.md: AAE.md without the Codex section, without curl downloads, and with knowledge-base contributions written to report/kb-entry.md for the user to submit instead of gh pull requests - skills/aae/report-template.md: the LaTeX skeleton plus the generated files of a three-experiment example (produced by build_report.py), so the agent can write the report without a bundled program - skills/aae/knowledge-base/: copy of the knowledge base - .claude-plugin/plugin.json with displayName and a 1024 px icon The plugin at the repository root is unchanged. A workflow checks that claude-plugin/ is up to date. Two knowledge-base entries now point to the tools' own install instructions instead of pip install. Co-Authored-By: Claude Opus 5.5 --- .github/workflows/claude-plugin.yml | 17 + claude-plugin/.claude-plugin/icon.png | Bin 0 -> 98293 bytes claude-plugin/.claude-plugin/plugin.json | 20 + claude-plugin/LICENSE | 21 + claude-plugin/README.md | 11 + claude-plugin/skills/aae/SKILL.md | 363 ++++++++++++ .../001-blocked-recursive-sorting.md | 55 ++ .../002-batch-size-tuning-transformer.md | 31 + .../003-arena-allocation-json-parsing.md | 15 + .../004-soa-vs-aos-cache-efficiency.md | 46 ++ .../005-branchless-programming.md | 51 ++ .../006-small-buffer-optimization.md | 43 ++ .../007-move-semantics-avoiding-copies.md | 49 ++ .../008-constexpr-compile-time-computation.md | 42 ++ .../009-false-sharing-avoidance.md | 50 ++ .../010-pgo-lto-compiler-optimizations.md | 51 ++ .../011-mmap-large-file-processing.md | 49 ++ .../012-open-addressing-hash-maps.md | 34 ++ .../013-simd-vectorization-batch-ops.md | 59 ++ .../014-loop-tiling-matrix-operations.md | 47 ++ .../015-builtin-intrinsics-bit-operations.md | 47 ++ .../016-reserve-preallocate-containers.md | 47 ++ .../017-hot-cold-data-splitting.md | 56 ++ .../018-string-view-avoiding-copies.md | 49 ++ .../019-numpy-vectorization-over-loops.md | 53 ++ .../020-slots-memory-reduction.md | 56 ++ .../021-generator-expressions-memory.md | 52 ++ .../022-numba-jit-numerical-computation.md | 78 +++ .../023-multiprocessing-cpu-bound.md | 61 ++ .../024-dict-set-lookup-vs-list.md | 56 ++ .../025-mmap-large-datasets-python.md | 60 ++ .../026-local-variable-caching.md | 60 ++ .../027-itertools-lazy-pipelines.md | 48 ++ .../028-struct-binary-data-packing.md | 53 ++ .../029-preallocation-patterns.md | 68 +++ .../030-string-join-vs-concatenation.md | 56 ++ .../031-deque-vs-list-queue-ops.md | 61 ++ .../032-array-module-typed-data.md | 57 ++ .../033-cython-c-speed-hot-loops.md | 72 +++ .../034-graph-clustering-lp-refinement.md | 146 +++++ ...035-streaming-hypergraph-multipass-eval.md | 123 ++++ .../036-kahip-bit-identical-speedup.md | 364 ++++++++++++ .../037-cascading-reduction-rules.md | 120 ++++ .../038-kahip-quality-at-fixed-time-budget.md | 155 +++++ .../039-pseudoflow-solve-direction.md | 116 ++++ .../040-parhip-bit-identical-speedup.md | 561 ++++++++++++++++++ .../skills/aae/knowledge-base/CONTRIBUTING.md | 57 ++ .../skills/aae/knowledge-base/INDEX.md | 46 ++ claude-plugin/skills/aae/report-template.md | 450 ++++++++++++++ img/icon.png | Bin 0 -> 98293 bytes .../022-numba-jit-numerical-computation.md | 2 +- .../033-cython-c-speed-hot-loops.md | 2 +- scripts/build_claude_plugin.py | 271 +++++++++ 53 files changed, 4555 insertions(+), 2 deletions(-) create mode 100644 .github/workflows/claude-plugin.yml create mode 100644 claude-plugin/.claude-plugin/icon.png create mode 100644 claude-plugin/.claude-plugin/plugin.json create mode 100644 claude-plugin/LICENSE create mode 100644 claude-plugin/README.md create mode 100644 claude-plugin/skills/aae/SKILL.md create mode 100644 claude-plugin/skills/aae/knowledge-base/001-blocked-recursive-sorting.md create mode 100644 claude-plugin/skills/aae/knowledge-base/002-batch-size-tuning-transformer.md create mode 100644 claude-plugin/skills/aae/knowledge-base/003-arena-allocation-json-parsing.md create mode 100644 claude-plugin/skills/aae/knowledge-base/004-soa-vs-aos-cache-efficiency.md create mode 100644 claude-plugin/skills/aae/knowledge-base/005-branchless-programming.md create mode 100644 claude-plugin/skills/aae/knowledge-base/006-small-buffer-optimization.md create mode 100644 claude-plugin/skills/aae/knowledge-base/007-move-semantics-avoiding-copies.md create mode 100644 claude-plugin/skills/aae/knowledge-base/008-constexpr-compile-time-computation.md create mode 100644 claude-plugin/skills/aae/knowledge-base/009-false-sharing-avoidance.md create mode 100644 claude-plugin/skills/aae/knowledge-base/010-pgo-lto-compiler-optimizations.md create mode 100644 claude-plugin/skills/aae/knowledge-base/011-mmap-large-file-processing.md create mode 100644 claude-plugin/skills/aae/knowledge-base/012-open-addressing-hash-maps.md create mode 100644 claude-plugin/skills/aae/knowledge-base/013-simd-vectorization-batch-ops.md create mode 100644 claude-plugin/skills/aae/knowledge-base/014-loop-tiling-matrix-operations.md create mode 100644 claude-plugin/skills/aae/knowledge-base/015-builtin-intrinsics-bit-operations.md create mode 100644 claude-plugin/skills/aae/knowledge-base/016-reserve-preallocate-containers.md create mode 100644 claude-plugin/skills/aae/knowledge-base/017-hot-cold-data-splitting.md create mode 100644 claude-plugin/skills/aae/knowledge-base/018-string-view-avoiding-copies.md create mode 100644 claude-plugin/skills/aae/knowledge-base/019-numpy-vectorization-over-loops.md create mode 100644 claude-plugin/skills/aae/knowledge-base/020-slots-memory-reduction.md create mode 100644 claude-plugin/skills/aae/knowledge-base/021-generator-expressions-memory.md create mode 100644 claude-plugin/skills/aae/knowledge-base/022-numba-jit-numerical-computation.md create mode 100644 claude-plugin/skills/aae/knowledge-base/023-multiprocessing-cpu-bound.md create mode 100644 claude-plugin/skills/aae/knowledge-base/024-dict-set-lookup-vs-list.md create mode 100644 claude-plugin/skills/aae/knowledge-base/025-mmap-large-datasets-python.md create mode 100644 claude-plugin/skills/aae/knowledge-base/026-local-variable-caching.md create mode 100644 claude-plugin/skills/aae/knowledge-base/027-itertools-lazy-pipelines.md create mode 100644 claude-plugin/skills/aae/knowledge-base/028-struct-binary-data-packing.md create mode 100644 claude-plugin/skills/aae/knowledge-base/029-preallocation-patterns.md create mode 100644 claude-plugin/skills/aae/knowledge-base/030-string-join-vs-concatenation.md create mode 100644 claude-plugin/skills/aae/knowledge-base/031-deque-vs-list-queue-ops.md create mode 100644 claude-plugin/skills/aae/knowledge-base/032-array-module-typed-data.md create mode 100644 claude-plugin/skills/aae/knowledge-base/033-cython-c-speed-hot-loops.md create mode 100644 claude-plugin/skills/aae/knowledge-base/034-graph-clustering-lp-refinement.md create mode 100644 claude-plugin/skills/aae/knowledge-base/035-streaming-hypergraph-multipass-eval.md create mode 100644 claude-plugin/skills/aae/knowledge-base/036-kahip-bit-identical-speedup.md create mode 100644 claude-plugin/skills/aae/knowledge-base/037-cascading-reduction-rules.md create mode 100644 claude-plugin/skills/aae/knowledge-base/038-kahip-quality-at-fixed-time-budget.md create mode 100644 claude-plugin/skills/aae/knowledge-base/039-pseudoflow-solve-direction.md create mode 100644 claude-plugin/skills/aae/knowledge-base/040-parhip-bit-identical-speedup.md create mode 100644 claude-plugin/skills/aae/knowledge-base/CONTRIBUTING.md create mode 100644 claude-plugin/skills/aae/knowledge-base/INDEX.md create mode 100644 claude-plugin/skills/aae/report-template.md create mode 100644 img/icon.png create mode 100644 scripts/build_claude_plugin.py diff --git a/.github/workflows/claude-plugin.yml b/.github/workflows/claude-plugin.yml new file mode 100644 index 0000000..6419774 --- /dev/null +++ b/.github/workflows/claude-plugin.yml @@ -0,0 +1,17 @@ +name: Claude plugin directory edition + +on: + push: + branches: [main] + pull_request: + branches: [main] + +jobs: + check: + runs-on: ubuntu-latest + steps: + - uses: actions/checkout@v5 + - name: claude-plugin/ matches AAE.md, the knowledge base and the report template + run: python3 scripts/build_claude_plugin.py --check + - name: no symbolic links in claude-plugin/ + run: test -z "$(find claude-plugin -type l)" diff --git a/claude-plugin/.claude-plugin/icon.png b/claude-plugin/.claude-plugin/icon.png new file mode 100644 index 0000000000000000000000000000000000000000..b4b67a04f0ccffb65c2dddcca048b8099bfc9895 GIT binary patch literal 98293 zcmc$_1x#FF*Dg9(DNvxeOK~m5tw@V|ahD>+-F1-SPK&!!+#LpYcX!v~I?SE6-?=xr z$$#>nlarhY$)04`?|x;iXFbnaYlnW3mq0}(L+-dGBODH%CQ=wO)JFET*$7MidbZInT@TM;)!ukWg z;0iynR@MZrbb7PQ5{De&xHJDH7#v%k$IWe?vPZ++lzDf|IY%tCpW%6Sl~z&J-kbXx z`Zxe!D!^Cf?-$=cvw(hL{#^i=|E|J+W(EC7{<{E>{$0tTbNut6`gi#+ZvLPD`oA^w zfA7KnYeWApt?9ov^#9g&{a+jUzxUw(;*kG~oBvPP&;M`e|C6TqpP~Q9nx=vdoc6aE z0K(`uS14ZHj|_Esd+4E!gi!N_9%WV{BfZM|bexaJJ->UtD&~$@3K&ua`OE!tfC(7_ zDnUn`?;&g?kKa4191DYj?xI=6RZ(+_S#`+%^FJk`h9}SF*&;wTad(2p3(A~*d8E5x zW$gH3BGIu}q-(btF&he{P>SygSbz-;sYZCfgd^u;Z;JT4{<3~K?g6i*gM->hMcXg; zE^YRcT31nEw)g7j@eFK(|YHX+iD%eq(Og zsI%5hjr^NTXTy8pfJV2n!;*8k*+{pun-5K+Db5=Bj9*ngr-@Q*iRNLd+UsR22g=8% zsoE9K&@VIJ7vQDivp|mb2faI0M9-m z&vLFZeW5_w>7=6lYrw__H>_6xf#uuF+>&MX+Dx|EOb0NppELiP@VVp!Tedz9C*q$| z4g^BQy97oE5yJYzKjH6OaL;!pIO49T=NKkPmbl1~`N z?n!be@4yi>u!q@S<{JM4;=BUcLN0T&h!9af<4_`FHljh}jp`oiSok>$;{H&2i)YzH zIWm4pUQ=m1Cwp0JHxq$A>|^=Gh|I-6lUZDGZv;Y?DRa!xgnNU&)rezJJ$YgQ6c}M! zKL%mIL`~Ca%5p)noDOEA-A{}ZIy(z$euIceCN*(fMj-*(O2d+`lUmbD>924`16y?7 zUz{z~rJLnQVu?{~eI4UIw#pSJa#9P7g8>|DPN32Ou0Os~BvUMnAR{K>rkmqdr=$p0 zJ=gfMhmPd&L@^>9!n|mvRXbOkKjIsmHJrwDd2EuQ6s$6Ao;x;YSt}(0X9)P?k}rZ_@MG}*1~8{rEn-AMbX+!eR#s9>-PsNGNt0QreAtl^IJlwCA8eL2-!W@h zYSrQzg;7u1ns^&3G&XidR$`%YG7|f_921!|u!~~er$~SA3j7m6{vU3kV{2j(jS?cS_ z{pKW=xaiZ4g~T0`(aP6DN!X51ugPScQX&yg-m*!v zeDPFXWjknihGMBN4eU1Ms~&k`g7l9cQ%-02TS;2P@c6I|AG-a^UH1q$BkAkS(tp)r zM#4u)y}F>4YsE{GOfm})FX5=iJq?2S6S(u31B#-_2>E^7oTe&^R#%WdW%T9^lN2+N zhSA`oW-0%imkggv!7b;Ub)S^Xq6|WLwt+jBGtI-Wo-?nLE^(dt>8B;s7 zs*{c%fCv1U&h?P5sTvYiq@E(5=l>k zW!OHqYO&Fe^ifVe!LdEE##U7;o4pq6FQQCb8Mn+3Z0E4|e}5u10Sfl#bf5adVSJiy z-ej0{`))gzj90;)pqkrc2a8U4?Q^KElw;SfhL>RGRW_z#3=8-ZIF!;>bL7ad(phPs zz&3T`$xY-OFcXW@&5Ens%s#cGO%1g%oyNJ90}01Xvg-pu;b`5MY2 zgp)raC8L0K{cYblCj|`8h)sIC(xNM2?-%laK`uNLR>;Q9>2~LLQe0BZ=j_+68t(mQ z{dkyeM`EdKS-(#`wsY$CNS>z4V>_H;*QqFBB*{k}Fe8~K76hX>{#nN5u4!>08D0jb z;8yJsy_*|X&3l)kGuaOmMOsICw{WOiTXh8$?cA!|(&i3@x=crSS>{fuE{E6geF?>M zZ~vZ5419H-0s&64054k(C~m0Voa-P6O)L^4P|QexJZwpghDFQr*(G<7>}Yp15cuv` zd4i&u|!q7R;{R!e$jVW7=T&icc|CeGl_#|1~+TuU+*^Gkh=aR zG)?21;)ugQf;|*CjVPnxGcoM3*<8LQz^=Pn`OdW!^1^=)HgQVnQTSn&I<`+=D{V@w zj`=r?TUP5HxlezlRS41WaJUOkktZPv7shT}gL_)hsU5psXo=FB(>5dhJZHqEl9er9 zy!q@H&oZNdkMPe>Q;1l}Jl-bX|#p4)`?{h8rXeWjL^`7$x)sp}de3E-1K z%VjiamfQF$DX|F4NF@RcC$j=7*eizd_f6^=d89Mzx{Kq!_aN>Om z!AEN)jk8mIZ#mj*ksEi>*g`i;E&1!~;9K`$=C&uS5i!veFL`l1$!DurU>OP--OeY% zzaH+P%CoD+l8%$c`<^1df1)@=-k~28@NRbexXLvh%3(b2`b}xhgqfC9jr+cZdnI&$ zrWS%3ME-MgYnRpT)Q`{8ed0Y%kDBVL;v~fzKZlTf3#J1qTGXsqZy%%4Z-C9zs64tk zniLSN?f*<#VOB_?>11aEiIRc|90apPV!cZKIt@Q}#*#SO>|~QW?eHxaOOaFy;EDy8 zTfniw7zz9i^}j!mN=_YEH7t)LPfAEhAw8UF2gYB5k@?*JgdrU#86ohy*b@9+B*n|C zsWgD3>9c7*jH#qk^#GpPvq=0y%&WhMs>0yCRSw)TfKCn18TSS3JaicilGls&4 zt~2k+Ht)=50O&=t1DX)bjr}^%`V~I(p!}zs* z+C1CW^(@(`eLT5TM4*eG@@1#yQ{B`Ty7ic=z zei%k1w&E+HlaSN#Mkv{8_O@fQ$2;5oX-vf@vsEa4oj)X`q6ru${=+RN@| zyVM<=*5Lny9=_QDn90mE0w#JPw7+sk0Zwhl{z*96 z5$K>p(Q5^Kgi>0pkZ++*{#l@k>IXCYzwQ(@7s0w|xd+ciVs_{$t5TMWB#h?*Uu%|r zSmjER5@i-@yrtRrMj$P@iqV))Y^#FaD?Z8#?T=yCimVgwCgME=T`SttN2Qm$R(@`p zQMDibgF!;Ijx%XA6gk8?{(4u6Y>$yJsDGWmBX(i*F}@H{J{24Hx8tJJOq5F!mD8YG z$3N)l?st+TUBkP353$4-wq^K^@$aV1O`2`fjJNGuKYeE-aAVR#9Sa=S{Jw`_gi@Ff zp4i>Qk6Sz9Xgc9|`4XP?MO_te0tf8Rv-Xl}BZ=~&S6 zKfC$qDIOA`^8UeUU%k<9Qx!tXLY7{6!?$J*Z|Dgj>Yv!=Prcp=R5Yzkx^#$AA(Na- z(8*Kvk-E*+`6R-u`_GX-DsZcba}RnI4`fDFy~0KSI9X$&0d$R*l754L+UxO?LKrvc z&pcRiP2Dyo{%Q50HL1Ad(TQV1J@X;PRzC($o~4-^;hHLS#PsnPB(>ik#{6 z@Df0bl}+j+Fwv7GgMf{o`5N!N=>qfFOXAyDa!qGu;!^F{jCp+nowG0Qw6q09{f~xh z_+Np~(VkC&1p43WBXP`2Lxy9bjC1TMA~DTVgVbOEx|cN}GJqIkuQB^hAf(=BYC;3= ze8G%yjO68*EY7c!8(AqOiZ|bt;}kQ{NN%C^(`J&6n6Ts})nz9Me@9r$f>u$L{J z`~ui-hTfkJGcouD;dK@}0TuXY8m>X-WVrHp)sss4-j~{_bW-qxO>-|No8^Rp?AT@x zK893ZX^5-mOhfKneg!te)+oSPRTukL{*J6)hp^z<dS4@G;CFwX3xd+6!X-b}PW+p58&3jV)EOHfgIpit1XN}!r1RIGu+y;JjA9Y97#=ZiIE-ycM{#xpA+SdkS^kzM_ zI<20{cC9b%LKz;QkdwtT(|uGE($*@Z*P;}6mB`>C6rF1q1j?u#QY}qkk4|FE=L*D zSL|38;@00esg8sRAdgEMgK=GXXKL~--M;V4^96scrf}NTzMv=g_4PRZtWhbz7wr=f zz&9$1gX41lSRimwYl!q~PYq|u4q;yp7aKgpui~ej>Z6xM+K(y`5eHWB$vJ-5l!mV7 z9?y`M=-Fs$-5!)xz3iR#rxT<8x3R23LU+$s$y5(&!1Y7VK?4ruJY;#q-?MPgG((Kl zD^1YD$Mrf#|DYMtRaNEISK2}~mc~TG_+{RN!)tPd3eg0{YkuTEFw6<3v2aL7bgjHM zvG_MRcUWIJ_Kf%04Yv6| zM}vi`RLQt4*j*N(n+SSEkDf{KkJ?agl#%_d&JyLul}a42{ewUoVsIL!VW#f#&LV(knNUNiUOV4<`0&XT4eft12#-K>@ zIxAgKZhcF!5gA}T>}1xUyGb&~qN?}2>`<^5d0)o))U$`>%&gLI)eA>-W@%m4Wsh!* z+T6Z-y&BGfrULvyLzYixrybm9F~R#F(v^JxA+zPjD`Vd*iKSxO4}`^GWiV*{Vh|Xy zk9<%=G0RIMNBGcnMMcWnU7jig9M^v-o|$)jWcA9fR##P0#QsA?Ht^FwP&f>W=~)iJ zdj@WkjW9IT*EMbSp?W6hBv+bO7>|&;>^Xc2V@XHnE?urNa7pn>f!)51O4z*r47;Cq z2$ftRFD=_+w0cxus)XU-1%=5~+N3$Rg`Q?+D)jmB)QE^=$h<1gekkyRKinQ%N8^`> zp#dgnRA6Au;+YPGgvMQO2tJ8ew0dG2dk&UHL(#L=d1OyV`xfaYzoITJ-}0TJi)IZ? zg9bIr(K5z`wLzG$hqWB zcR!hjBY`hiqFvEAp5fQI9(0;-X}cv(;uEIJNtWmROhH zKt9hxc2a|mV&yLG_~}&Ki(lzek_NS@wo|=|X{N=AOnm)8>1We-hHP3GPRvh9tkieZ zUWBhty^e%#Ysqy6EySzwwsp=-c4+Wz#< zaAW&u5NO7hajB`2+DH~EL>h51M1L^v_cR{Hu~Oo$;9XeB2cxSS=m;vsMvW!sR3!~j zu)f1I3LEl*iVwV%)I0p&r;>7thN@}S+g43udzL4T@h-GuHHZEu-&_H?R=Rj?gP#);9 zE-`{{t63P1WzMIpf~y>87a@jIU?DqAurDgQmSs2Zc(#gj@T$V!Hfp(Z9Xrv}tiUugfsc&HZKfp}Hw+JWqo8OwljJmi^W9!0LGy znU7{}-K_SDUi%Su-E!s}b~Ll8#qcb((JQ0RS$i$_lIqc{pXztz!-2T?D>^6$$7)>Q zWTD55v}|tf6Ze`s?;$?g+|3Qf&VnqV;@6w_%Gl7w*(x4wqauTVOS*ai2cSZwP?8O+ z+%a5>6Xn2qd$mF=qaCWKs^I9@{5<0ceNqD@|2t|=Qm=tKFghH-QOx}Q@wj7-c`Ou5 ztfbBaZ3CaZz~fA3ioa<^l_G^n=~Zqd;o*HF3{>^^5b%;ecA#eDSTW<jp<>*198v-w$XQ^&!zMS zwr$#3|G3%^pVOyB8M6pg$NJ=&Ucx6)6Mh~PyK2_zcWfsyUm-6IRasfPPUn|6&^1=> zUJtCb)n$5ZrEv_#{J???QBsP$r9Qvh9po8fU1`9~3wsC?(78b=z$9|st z1;dKqLul1Bjncnn>VWq2PR~|@cEBaqn;hdX47mJ!DBZ|rQ+D!>8!zI)Xt?YFjcMWP zG+Il3oY2X8vk@`M?L+g{`b+O%m4T)L#ETMnxmrT6Ww_F(@T+aS*OhzwbTkT?ec6!5 z#o7L99qAjz<)>YCL9Hx%STdRzn0JOx*e^S@V4tT{`0>auW~&Hz;zPsB9-Eg(j8PZc z+tW3|2EF$EwCBFgIj$_P?<^Bz4x$B+4`NTJEZxM3B(ems_ z&-KG>GMfMejn=j6KD>I`^#)*DU$u8;C)V3gy6InXd`b{`s1R?-7&iO;?oj~+#;}D~ zJEum0bE$xg3Ad8+i&eUNhi@mk(8vzF>h{pfOLU(t$lCL!k6OlI>?3>m`xt`{-|tS4 zPM6~IjJ}CdA{-4Lc&Z@6QbBbWc{&ar9q_??JBxx-EJ{K>AjhiVo4l2^A2f)mf5r)= zOiIa5hATHP^rDTQ8@Z_C&oHmzQ8$V#%Wlhg_4I`K|DmPUWn7%u5$QPb9Y9C_&~ ztXEx@jlYTWaccqp&xN7TN{*WD zqjPXe=1xLFg2yHGtQyW14x}2o{qA_|(pDs1j0MSJ5R|*&y;O*V#0{e<)<2ct_R2-O zHE6NZ5?F(md6Mk5A9cf7=5X|nfGLE9WhR%j>zWyE>n~>%%R#F_9+{G5&mB+xT?o@6 zISU@wkXA-wvFpq;*q(aF+q%B?ISpphSAA>~p>HuTc6D7e*Gv*=-1OguTQ^;pS{NC5 za$yR5iHB_8BZtM!syz%b4hs4p<`PhxPW(dS1>`Y)93q#~wJ(P3T`84UjMKy)zsxK# z)i|0FG0vTP(0@%`J?Z35N$b{@XVJEIB)WMblGa~-xZ%~R6^s|^PwG;$;(P&H_s8Fv ztY66XA+j+nF+L!@>K9<* zLlhCI+-?9d&o+tnSwzA#-0b>^o|iolqZV5qD&K>$;l-qNkf@c z;yD-of>L=^6v@*kE5QN}-#zz65~f9#8!B2hx^kChE=tsp6Yg215~A=I;|pt4atpBw za_9h4Z;An*JO(Q*fnd$;5SU+jxdJnl?;SUyHN$z<3{JO8o^>K;YU^__%j1|IaPN{D z53BQqk<uo3u*x)M1;qQ06MiS-efw!}LWzrYilewj1L&&z+(adDKtfaDoR=sjp$Po(ry+MuLBiO;G0h(9$#~)<#7^<#lQ*kecGRk5r2h4 z@2Y=lzj!?(G*bjLT^-x630@%<3P5ttwONOv-?GMC$Kz9aps8eEY2zV6?%DNEU=icX z{TCW7{hHH3Y!G=1V?MO_gImGQdCy;{oiR1k?ZJ`^RTVJAn{@OM%+rf(=~k?Q zsyXHqEN5-*zG=yekh=zIpOe;G3%ldF+$F7H8)0<8Sg)7&f)$41=mZY|tV;p3fIQTn z0O95&HU@`rPSWsSsg_=Blg9SZY`nq9{));0J=&+fbYB`Mav{_?Cn2dewRljsW7vc% zD;u~eDt|?Sa8c%mFi1hf?iH_pouzOp9B{wLc7=7mxc|P|dfD`~tKRIhK~lQC*Mjke zD2Hp+O`yWb_dB578^egxtrN>Xpz&R~KzL5iRV4oT!RM(6em02G#AMkAOf(f`gWcN> zYVQ4LApBe~G|Z^)RduW|I$#zgTF18@ObnQUlLnBJ{4P;I*W;mFh?FpJKpQ7>4qg<{ z;@W?CYvwSjURRlY?QGj?iiD;Wu2K$aSVtPKlcl&6e`$BRpte3K_qiJ;0JGE?%ZH?9 zC_Zlowb?>RL%s)7e&snA0Z7ASskJ{TPxm10v^=H{J*w>?{$><`%S$S&Mg_xD+sL-W ztEi(e+$l{mK@UTzSs^9a`SB3xo7qyi(Auz3Rqj$sTI^;9SK03M@3Ylae01=NaDd87 zwZ9^4)}06v;MxHy_>QH|>9HtRZ%Np{;YPK9%(_Hq`VAE8Np4e;k;FZ#O3vNKo|OY` zEN9HDSDG=_T`c!RbD3YLlPBkdwpIymPOXeU80QHiFo_u9Koc96$Wj(>P7|SuHq|M9 zeQn4~`S$Y7-tG0iBhzI~Ehqn*b0*Pibg2#vlB#oOE`?l%F`Wo#MM{!>$*MckN4l2! zzFMX2$!0dymh|-16Zlc>X`za_P#yGC4qSb-`8e$d8~OIoGt!U_^ZOb&U~oueHnQBj)Hn+7IypvMtd3YLso${V+n7q*s?cvq=0e1r zQbJz9wH_Ueub!pZ`hvdCT+;DbJq^g6@}K)q=y-3sE=eM$ zbnaw(3)Ea4z{>jN%9?2T&0N_CUAs;Ulc=bU<-*fm2K79=Wn$q>e{IQUpKZotqjSCl z8@74R-trYrzGuJk9ol*|(<>FV{=w0KRmU^I`B2(h2ndoZ>OLH`Y>R%r;#^BaeBGfx z0*6?icJ^?)oN}j>bN7`o_ZO!1#6URlpBm9Mc|U6O^95n_;(|5e1`}Ti)G#EL%V*+v z*UV0j%aU9*i1k4XUo8z}`jYxC(QA_m>&ol}ic4T5I?ewG#zZ}J&~bjf%$sJV4_)m# zCO%sN{mc5xF6sjzz3=4AB2G7 zP$n&UnbWWP56+UC1v6*$GyEYceS@Y^f zi#M#fO;4)5uQRCIBJ0bP&X?!na>mzO|yNPEvqUnR(bM;&l#kkPqURk z`;#|U`q6O@(cMfJXiBwgk}OCQ`~m6w-o4ats_o;rQsTj>G~IYEAk(*ykpBFEMw7jN z{@kbOq326?O6eA3@>_lVjuhT1o4aA_2-=BLl}OxjcJLKRzvsFJ<1v*}1NsC-%D4D5 z=DCGC!odif{E@6Ko%vDkr=;+3tT9ewlU*GcJY13ZP3pKx>#~JFQt4f2P z7I>j+DFJR7Qh$M_B5&Db!`+k2YE3fVLP%^REH}=!P?5=)ScRSou2P-9&D_H9P@q@M z@@Z!&{nU6zsq89|>OrT%4fHX)d*9z^pEH|6?ifq8zcYAtI1QJ|4Gq24Ru)i$*wb~e z@6wdk1=s68e~uPce#j-D`vUY+JI>q2d37B%k9P11+XJRW69jqe+w%YEhO$cls5ROmaM?Qr{xtjiu(Zc=DRhkHjfV?26%jg@{VyXa{Z=rzUb-V!m)R#X%G6^V_(z1ah!R3t`BLW zfs^@9=2kF3U;SC%OD!~PB~}?MX499ql)Ch4&{H%qhZUHR;}`*ueV zVy%YJUB>+!+e3Qb+<60$y*8{wQ9vW~0cY3+$d*=Pa<>+{#=YpXZvSsBxCtS*32|36mWfX_?rF`yh|;vD1+PMp#)=;)Z@n>>(3Au zZ*P}M8-ofW|P-R8c=zg`Axmv{aaz2V+)SYpk2&L1vwx4bu>sLKWYDtEF?}wDtL`CZ-WoaIZR-!&v2Gtcgu#N6Pl_(s<8F*LYD^mScP*CC z6!SX@L&Lp}lE?JlD$ukRmC;=1`arYl*YUo+5;N9la(vmb&Um7iTHwujb3*_GN++V* zx7}Q`%bGE)ysl;H$q!JY$voApXHrw1?o}85#vzy2mR1bq&EwxO0SB@2$@nk32zm)|0@Wty zFPfbT_cWkp!nahGI_OD}CXRPNh^H)wwy@8 zaaV1d2&FVXp6yKB*M>FjFmXs_sw^L-K+yXUe9N6&nBk#fN#X{+5dc0|UeWX7>3ng3 zc$YDc8&z%GlH~MS9W5sgUW}(fNggVf>{N3L`J?M1% za|6}XS>!sC{dyl-YD^&(@wulifwcSFF)|0ZQ7~lO1_?lR=D2WdG}P<2@Dzdmo1Vd$ zZSq|E+?h(v_n5$&E$pWIV^;#b4BJ&zzvdUfFB7I|7OpS$4dU6!H42QadxXmelgy^} zVi!yg<&WSyqV~t)-IdPh7iMTS5X(w&x*ojOfQo`!yD}LgLHh0@*dllm`~(sLzu8)y zBRz2r^!A4$0ya`2z$1n%^9W?DRut}PQHyhz{M^WCskGuGvj-3TGo4I+Qd`5Ehf@G& zsA2g&u2!7rJMv$^rzZ?0BIg-KXueg=nc(}h<pr{88v@VmesCt5efS?TS9*5`8!U zT`!!p@KD~#UxKo$Ew!99}{Gd#*0A~3wsSvFdh%56KpdUvpSPYawoEJXLY zwhd-}I5g$AX({s@-^Gg8>z@do5ZRCq%-HO6Yo~2I=(u~Vjx#?%qx(Cfqfm)BAN3jS(USK@l z^r;=5)W7!pDE@S#BBTM`pgIg_x6d4bX@t)e>Q%>9ntI zkke|c`V08(rC6-$OFktqA$Mb@(Y`aJ$fdt#v>)ME`pGh%8&vJpvL=duikW3v05j}V z|8b}=#K~{uO8Lg~Bq=5mZ;ms$0Wnt^T% zw@LO|Cn98X6Nw&m8z5`}3Bv%LBA<`gPukqu;IDwAJmZi~<>F8iKIfKYHf+_1C_~E) zO}KUY@(OO>s`EP%9CK>y>IYeq>JHcjTn##3WM zw^PpgZhjKv^A**nLE2w-N^tN)7-o^;#Lzs`2WZqpE-ZuvfhE?qgCtr^&ZC}xB%AS_ zAokU;s(^w*`yyTvsdai+&vP^kY-?D(e|=j9Q5lKf#-v7D%>OgA((qhh`gCH+XvIj) zLw6Cu>j}?;h|y`YWE!1!SLXCwdYRrTUW}rU#~bIXeJn-0r`vZx;7rU9N4p?$Y#qls z@9P8WhBfI3lT2@ym9XOVM817aU8)*C2)j6+13+@|EebTLT0IFNE5DQ_#D(g@0n$r$ zXeDzR)P%eEv^pDqe11c6Kc&vPu71%Z;F@A)*Pt^D-BHXHvRkXi&AU>Ms!knIg7fc;XyBK7I)ZN?fe>4LR8#=Si*G&s0I34mWQUgZN#fT zAJ1bcmKGP^){LLziW8t9qm6Rz?VFA#8dsJA(^Jk!QW2tqC=Y648Px`FgdBOmK*OWxcA-`{%avxsyMemmXl_ zLux-$lFZUW`9}u^QpCD_zU$P4O~?!CoX=iFRSMyT)*mSy+_%G^{ZiBT-kYO$zvQ_~ z)hURMe+xJZKd5O}T;bqH2+m$Vt~7N7pCxkin!2FkQCY4qV2tm1m)MjoU1wpIpH%-) z<$N><-L0tLbu}_z@vZG7SRP&c97!B;W0dB+{$wtI&CvS7M?<29({!1i7{9c%r~Gl< z8*fc}FM}^|uq;YY57LWQN9Mgk(ScFc9n>Vx^UiaO_~~FWZkdFBUBV;LCGzKRk!Y65 z01Ww!0WmpyxsZg~# zXAaI=;YY%rduO)udG9hG@wUfe@8`xn<=dCj$ss&_tEN16gNw?3GVYu6OasQ0`rFp4 z?)?1GLd{?_g2x4;zMP!yH4nEo&U0Ypc_^(G8dHplCFi4dM797R4;$)>7Ge-kdk?p_ zY@#>>TE|9Dx(%a*U=Jo;bmNKCvu$I_M7_PcBU7yVi%q)l_e}nzAkS zM~wZxrhjV}et<2Tn&77UBaeL{dzkAZ_PR^!z}tsk7?G5X(D4V4X)+MhEHyo!iQ&DnNT{t6hG6MKlV$OCb1$5riuD^MS6yvj zlE7V+^`sVx4LgA6vuwx1ASOtb!V)j&S9gZbkbw14tP+CrldaG6@0)o!Afbapkl;&_ zl0N8WA7s|9F_*?0K<3^EZdrR8pp0i}TZvb{cZ_VyBS-W-3x<+upZWAd{-+*0yz^Rv zk=6P%^Hcq`4EFr5=PmZ$Fr54JkbziknR@KnwA}$A&Xq>iqxIFBd6!wPZn6y5b2{!V zE;wHoV?gQ8nhMTptef*Rp;akM8wt$1E`BobhCK~ge^XyM z5a`QlEk2;Y?Xw>d$-yrwU8)(EcXejDooW5w;Z=9U7QYVSPdDDBbJZV~Jg>~|mwrzN zYqR&wo_e)2DokwSz3|RoJSN^I`S-O0?63`c+UFO)rHhB?Dq?rw)ZI=ijP~~0Pr7@d zkxM1le0xwU3sY6c!mwYfX3uO5f6s<~K*nv~4_U=6D+*Cln*&}1nz1hTM_Ok5HkQrA zU8>_R$I8wleE;R=G}=rbJQX0O{=E>d-Q>s@^)p4b-2;m+_1JsS%~umWP!`jImvK%H z$*Qul5n0snT#NXw8t`V$C$96f;<33d@>q-R%T;>Ixx2rag`fZ--Rjk8yjE%Uu)O1@ z)-;%<$whsru^*6SbQI!U#>##1qI-#!{SwYX@pQK=5s7VGKbL^pXE8JuK3%6l+Cc9CW8E$hB8_l%~Cea%Sg;(=7DtQPruC65cC}#eZ7P8Apt0z^bm*0B$+wk zjzvAXE?&u7r0CTh#&&GBUE49AK%=21hXeCLH`wM12i9<|OXQ<_|l z7tWP*qRubI_V+xW#vb-dn1DXLs(Jx<5Ws{vnq9p{KCC#xAPe$F*w({Y%Tbdf<Kk6(h$?7>-0z0yXwSQTyMCFlU2g#!j3S=v-Syo}Ozdj5+w*Qu->1od$I8i(w|msN z&E4xQ7PUdREF<8HoOD~LYLuVh2p0ZI;v*IhSc)F+d3tfJS6vCei`jl!smLYI*(Jz`h+G0U{_?Z3t$)yF_($)vj zh#;H4A8n96?TphK-5>tAi!E}C-rJhfyY=8Ab=~22OPpcAg$~>^pFIyhS@mQp9K=th zo3cqSj)fCW6gd_?u;@h?T2=SVH)_9G@n#|Zu-og$zK6_gbp0)&e(O`A*EfeqvztV5 zB&sx0l#|t(9YS5HFS4DYSHAQtgd_pW3+H~Qz;X>wbvzw~^yy%k5pIZpdEYy6hUi}xwB5C=q*sT4Q&O+ zXPq>lciI`h{0W@VLQK~h*&Yz;HL$J=XtLX03bFO<(zM9qJ6YT~awIjns8d9kJQ)yL ztU34AMRDG~vHFb!?QMTFj29CrbO6M7y)z6~MaDwKYh3USSYp#;c$m&iDSpJSu67s` zNIXsfj%ra%tqX+ejz(2Cq50&Y&Kuk`g+4f_+9F0A6;eBB<0#L z(;dQGZ;lrHijBg$@kwYWrQ`j-_w$g$_^u6@Op~@&v39Q^O#(BefQ1l*-D>*&SYTMg z7Q3$8O5TF@MRLvW+PS`GHVtIL+t$gYc4#70x@bz%jjzs)Bm7k&wA-n#x=B}gurObB z`OV$O^YWrZJ4+a0Jn?laT2n?2E?II@?!Zz9aq-D5cr(E|w+y)5bJS@(8HQnW8J@z7m9EJYnafbJn_&PTu2a2Sm^ zIr|;>#B)L^m}lw2{)7x+7Negxh_+TtBYbmR+Ywu2cggp3Brno>Y$h$p`y4;pJDa8) zaLW6=;!&XF5(7&WH;x&>qLFW-sNjPo?*pYY4flS2NQeWfB>jo%AwqF$G=M$%RuX0H!b?+CBZ(uctx69XRa7n zm?}RTcI9u#G_u4kw6z7kw9FWs8o5Nu#w;wzP#GmaI{e?aBthkJOQvhtnoXBC0$??S z9oqGCLl}7ychQBm&dt*JSlqbgBYobYf8&w*LQ+zUX?>ee=EqD;{o;H~vAc_S^}S5H zWPc3DSftSj8c(JucRpIqFZD;UFEKXasI)|Zbr=^? zaA-wG|nx|6$equMyaoXN8xQrDAlj~}|!(F^@#Z$_WZ3azA* zVf@!si2ANR*XgPv!S;X{X^y_RBdb*LTroE1*()grzO)~-WVhefK|+=GMm72lF?7{vt{8WHT&> zKkvW@EImy7CypHLZ2C0e-u?uxsGx0>-CM0+l3DoUMtPEZ4nVtnc-O)nkQ{x#bn_|J zLz`~v(@ZLF*fo&7;xikX^#%6Ip1W4}nG^+igp3J1MvmhFVL%k=SPY4ogI z1%RYtT@S7$helTE`%v8gewQ76AI&Q*;R1jE6m&r=!K6ebb=oA^C?qT`!Lj3gpihtP zYw>Mpf3OQF&}*+$bF#R2quzSd5gfR3q0>^hs{k~pdlW&_)u2Xl z`fe~+Jlu!AA$3U9fkDIPhupIfQU`X7(CThYo@L+$BeaO+pbMO(2=|Pb6s+83s`2EQ zH|lK0DLYkH4}ezhpyq{=BZYLz0^XbY;_V<2J?n+84-5@C~T~g9Wh;(;%cZmo{ry?EF-Ho)QEC@>@ zNOw2vJO2Lf2Y_dGX70?r=bm$>z_3@h%K0d&vVU*%W9KR)hw)Fe!*89I0mcYAhMN|G z*=EkY`YF0W97ON5&!4$0N1xiyUQ1@bR(#ofyW70kqnKK@lDW{{E$)Br#2M#(C^*<> zDUBL!p_skgPA(oNmFQ&~ZCxuz#$(&pOO@yzcfm#(RtCvrA1M(DeA_WLDV_iEWMF34M2K%90<}7bm_)w@O<4Zs#S)#2J1< zV0uDr5_zXC3J&GXQF_ZXpQLT)(39Ewxk}~S~n1EGak=M4*njMHvC&00`-Z#ZKH*?f)J6fcZ&*S;P@eb?Pe zF=EZ~!}up=@rj*l0CZs=z*0j0x~kmB#@_2BGPP~ImZ)w0JbtRdrk;PYbkDD`>2f9b z1?_!Q11kXR$VN6@*(UqHa1dJpanD^_jOC&{WFnubFV(t_7|Se!zQXD5XQ%7;Tu!qA zciSuW(v<*yldkNxc!zX zy-VmE{#e^x#zB?NWNq1_cJ*u0#VzKHsna=TQvI3ozoa9ZxUjZ;sY{r`Kk}<4{a+b` z;N{>NHX)V``wwzMWSyYXYQIyPV1pc}vFmmr zRZzUIErqi?)y;II0TaNi9er)^zm+Y%EKMYbKHVD_xlL4%ZH-a^y%OllkbQX!)84kX zA|t87jLOku6s(UST}~^pE064&6mUsuEAwdo6S36pP&Kw&Cob(AU1f`H{HGu>Kqrl_ z`p2U9de2t|nserxLEKSQ$%=E9Th+IdTIq#uBAZM>D_PuZF4)xHKyHBjn_a6X#ib6R z;cX1S)}$()DvrNsB}JdEQrIPN8DF`bzdX9iBwaoMB3^uIP#1L4Y%m_YASx>nJic7; z6)+_eQh&u*EIn6k+?|>pwCuXbE=X&({+2@rC8fafe8s$7ZV$@z%$<@4;nVnPV+s7) z!t4|6PJ;y~3+d}b6UThQ3-|R-&Ok~fhIE_w-TQW7rhEuZ3q}yMlSqYX_oeBYIReqD z-EJfemtSyyPP_H`H1z-#C_4ry1E*?y)6u-1L(YA-c>S zd;J-`TI)I2HvTo^qZVN|_Q7WgbP*cAAr^N}I#pSxAwI1QIA+S`r=lM9GM1RDC3ELZ zZl5Z#!GFBYE6XEHd3-``5KVZ^#?@3PPR?Rw73X_Huhtb zKKe8AwiSa-3bj17qzQwiMf$R7JB1biy!8UDEFBj_= zV}^C(lW*u>gDMS|S~veKn`2STiVjk)1p^*#@5{*5%DFG#zsG)<0tXz8M(5{VSye|k56KZ7y6?5rh(^dR*KyMMv!>V zH+_U9$t>@(9et5N=T4^MJh8-S1zQSU$)tA5-u3z_FA6>I{%<+AvKOUSJ0K8N+o-|F zeL}<$poHrXXM51NjU9>?tqFIu#c$n}NESTa%|pZC&t3!2f$PHq;C%I6HBsYyz5OQQ zWZBcAAGLpTs2em5bD5%7izO|QU)FIR&wl(V>LfUDgAnY_ueYu>+#Yo&%1uZRb$3wYa&T zNm)hcuB)?mJy~OU=AKt=#}n=BVLDrrz_J(y`bC&d4P@U&1m(9uBg107Stu$ zFM~&Nn4fozE?u>M5^=J?ZjPSq=K31|8r-z9yV>D;Spihqi}`vZPco$407~p4D4uxT zyPrk02d>s%twBYQG|FWQYP8C5l;yNLTvuWfaeRtJE$7FVUKJRnJlg&{=G}iqx z3d*dG^Z{D}QEEHFB{Yv{4AZkawpJl9QDnW4>QHNl`RAo%oDOLFwKer!v3HEu9d-2m zvg?i)xweuQNtjgF;a9O#=gsM*0ys;1-(h4e?-2RhPmwf91^J&JQ5St=FLkUaJghBh z@K;>?i1h1iXfjz{7hg|20c@50)SU^+=3Hp}J<(jH=+;Hr`F1Y47~9Tg{&#KXF9VG) zgKTTRe9h=xD+I03w?syx7uJc&=1c51KoXN15s{7x$>Ak6d#f@0Y&LkkJwu(04B$gn zV_9rp6%QgG0e_=&yGYEzFMW0CyCVJ?pPba@$}~(qyXqZM7<_04l@F1(90cC z&>r7aI*pD@v>6+LrkxH?INGNf!fIf#aR^tB;w4wYK8}RW!}dvitQCHD*d8dVO)(x4W9^?HdKfqZpf=xul^^x(f|H&RC> zERZz&;oyqI(@ca3oon z8yYS!2>Wd9xGcx*vqsf_WiB9p85u2>GF~Z_B7Q*(E=#Tq0j|>|!H~+ueB8>@L-=o1 z=UQ`TMtyNEOC9b3zEp}O$r(LP6~2w(&Hbe)cJ()RNiOT8MU2aC{>@044~P)*2CnhT zZ60}QPBc!3`aGIypQw^*>oY7d2NKXMUR4QP%i`l*ZEZX1@LCaj9Lh&C{D|CSUlT)W z^SCn6z-YS5NPU z+M~%_ZT54DrJA8l1bY3nXvEl;`ypBxdtSM1DnPMG_Vbzo#qr>@+H(zNMB<9qPRZ7B z^jO-o;%Fufn(1aMVzUcUpt$nbGBPIPofL7bMgVE*;w~6(f<|!$S8G*;{V>6}*4~ z?9zB}^(v;;*-2>r#4loSjEV|N{1UT(Uk`?889DeFeD4$m`-eU}G~&ZT;|_Eiw4l6{ z((-Qo!BQsr&PLaqAh1m{eN-p{hn&#J5zw`g=eTb@cLsFDT=UU-^JqJIRANh)xJ%7Z z+DgyACY09Cg5dRSdD#CJ!W}`Z#)344;3rhW#9{RUrlX*+hGRB1Z{W4Elr?>~=&q(# zB>tg0-VZ9Rm>T(0#KE@N-pP7vynwPL1o||D(9HdnHR`z{f-Ilkd+Xt!*39Jc{ojPK z?VGzAN=3oth%)Z2W&gcFoM=|#1^h5FS{Fy?4i&P^gVpy)Y~j~diZrjT{@c}bgQcq8 z%3yCIH*w1675Oh=`}WnuyCW~ftc?B8BgP3VE*TVnLObuPB)#4!+PYTT^0sXhQEj|0 zoVncx*F7cA9_`%w-xsKj7FGUt-0VTbD@>4JWkU0P&>VDy2nq{}K6%L+osL^pS_(fM zxwL0HsvXgdX^I{&iF=shGMB5Zn&wF;ME?&}B}Yt>=(=N$%P@7S7r9#oMWJkKqIa@K zzS(h3k{EBBjQhq5U~zGH6US2c2Ll-rk8fgn)jxZvJobdH$?trxioh_);4D%@QqsdtF7=pxNqN z`VLTP0zs~{71OT7w?(fn`tIxiFd&##OFy~+TF{U@)vRlwht7G>)tRFdxQ>R#pB%=a zEi@=E`~jR9K0-o(J29O72nlD98NTXgWB8^dyPWzG+wT+&>geRq88kK@qh&}ir?res zG~6>=fif=O3bB6pBu#5d}#=DV}`FEVn#!>^tc9wQV|$_Mn~q^Cyn_4 z{$L#{{P#X7q3vQZK$Tfa6Lfz)s*w9I$`6dDrd#Aem#!AAjh*`1Dlqcka+pvti1?(9 z6u6u8bIQ#Ew+R6U>5!fX?|Bm4;5KyH4Zgs$q^`dHRp6xh-H_=?g$~)$%1jQYB>W~W zTBlAoeF1c_^qe|%a6Yx}zp>zYAL-Eh^B<-kI}DPZ6~jZk#G7s74>du~uhEWmS#Opq(OL*(3O0XF_?BX}>i7@t8a>2ohYG z&wZIGBPo>{VBhsF^zrlu4wk2zYsAkcTY~3dR|hQ3yQR(9cnj6J6)_NM!A!}=M{6=D zJOpOA;IU;;;N>AUFDb2C=8EHs^g-%bvp9lhlqW33kg3EA6%SQ0U})*8=kh!HCm_7T zO6!_My6*}*1J(+Kfll~t1>R~kuGo&F#PV>4==XSQ+@wz*2@F%X@@SsX zMaR9ueN}%xXh#Yhy68QK9+bUjrneUJ4P)$3`9!qqE>2S2&<7vr|ItM8L^h~W+3sw( zVuPGg-0!##-y!{5xU2X``1lW5uWe5I-6^a4`)ht#AS0bI@>ZSd8XOIP0m|nhLMdzy zF!Smi{LX(o*{e30$(76m4byq{b7$`@1Fz|zM`^Nflb=J8)4%C1m;)-HZq|7yk*AWu zy`$u0%AV)hKu_@z8pNtS&`=EJjL>DPHGh6|`H`D>W|;)zkcIe6))Y0{J!IEuICUkt zU>ja(_59z-#TYi*t9ly~5=Rv0YtU_KBM3^(;_n4?^>#$h9Ul5Q!bMK1n#sjoh3b^W zGL^+1A{PFWnz}}T{jXW~Wq-dLM})4;;6J)v;&r<4+kzwBP8nHgY5~l7$PbeYzJp~u zGOQs*ObfSV&_Q^b*K#9ev~e%;wz6{m{M}L|(VkaF`Nu1ufR_s8c3dhQ!{JTtoS9FR zyeP{Ld2iKS5-q>_$>aNq{=mg}P-!Ym&P(*KOb`gyrC9Dp{ro2g8&n&#Iwe&c|M4U) z?O>o14B=n4zHK6Ws@`>0nUP{cCU|Nc$NXNi@#dE7gI-f(W3b(D^7CIJb|mUFR+gl- zb)-b>3cDJlEED{s(i8u@1I6teyLuKlRNn5gpUNx~!X&JK7=AeQK{1M#=b^+d_g(Ki z$apQWo8QEjhZCM1!s|`PC{W?XriaHyplIwD^XymvVtS=miSH|oLhpWX%D}x|@12HA zEduL=k|e})g5u+sZjixiG_VKLRnQz42>JhL{ate+C1l6PlyU0{R7_o{n*-WoK;e9$ ztCQxuZ#28I`;_&=V)$9eLtQKMFtUBxBSJEU`sC=FySw_2g3s0^fBgPJfh>k!C*!{W*c_5tl2UzAYT$;q7o75Bd(E%i z);?0HqIQ0SL34`=kz+&H;wQUK&i8q3HI~MGoe-X`qT12PYn4&ys~6xy*RljOhs^DF6HZkdM}o(1fPpk)jV5*rfU$<>TB(PuLi z=&3xJACzP!ceb4!@p^-*Df-dv)6(SYTNnX+`KZ$A^P_ZVnfc~oA=k%@Zhp_X3&D3B z|BXpW$V_Em_FMjrP2FMLuQ$S;e{O3g10!^o+z82}n1coAo$Ob(G{+qJ~ITN;%SX~GoF=W`Wvr|6qbCWP;O0k}BPLE}G zT@0k>5fB(S2V_JmN|8;_KFc9{M@>_t)V6AG1lx})TWw~l6ty}s@?&&~iTFx=m3^BC zntwbnEx|6|k(KmEw)8UqFr0W*?<;to2MPbXCManJW{phGt!RHa871pIu6F^7fKK?; z{@hj}S=}$8I+qhu(@~Fmw9SzIii%3#fCa~n2A{?W{c)cy?B>7NP)obN5Rg~b2OS5N z7Y@|9>nvGihh6Qp^$Dv50Fo0FW{v zzL;UN3Nf*GLzA5^n*cFDpE1LD^p}&Us;kgJzU+c*Tq^fsMFpT_1f5;LQ28Bf1>aVG zMP0rt5*6#xc||N&O!bp<}C9`tm19(yVe7423h;0ZfA~Det9Pg#dQt7E{=s^D z+bdVZiNOpmj9+J^+93ecPW(uZ& z3+&YubxV0=zA&Qh;Jf{98or65=(m&*G@L6oZN0RKBkO+q+&1IT58B{a^xW$yL_Cf! z{=^S+aiUXPtNLsB>(dB8?&nK|=0hj%;QD|T{(c8L;$K@@8x_;w z+iGYG2ISVyyiH9Rj$_}F#cpxNihdsj&KY$SPsJS{ISB(pz*TJE+p;wAAvm* zH0L7=(AbAd!{-J8()F488;+7{z6XGtHs9f`j-O7x-n&oHTzmn6bV%}A-FXNp+aI+# zg3tU@9?0g~=F6Vl(fW-|)>bj#bGhUvo^0zwSk>+rf0CEQr^x#Ev{dL%CUHw4-%A`z?%lLeCgHWB zNmO>+^4etO29&niEa_*UE+r4ELxxWJhdkH?8Ug*?xJaoF%cx4gpC1QLAQK~sL^n#V zTX7-uR+iYedq(yqU| zd^w{L<71UGi4T{-9R&mh^m^!LGRrgb2mqT*jn4tsKzy^E zcv4mF+s6;*BwPgmTcfc?d-h1s7&w zwt4o>J*o>N0`aD%)1w}$rg~2bu%KCLL9HuM4_pU+m2~*Pe%x2j!cM5|YUDUe8aGMeod&<)Crhl` zNM5_dA;F9YAa4EL#JxXi>?Wh*$JdQf|YQH*@@ls`P`ofZe_##O_%O(+Kz7W2{9YpXo0RH z-e{83r&k8LZIM;LM7ld1iMU_OY=Mtx+ik^a{{82p1h2l>?&pma1eLSQsl z7DqGoeq4mtCiQNi4&u=3E*>hxW?s!31+x91-v_r{ zuN=qr8Voqk_4y+XXd2Ltxch@*(xXj7f9MatBP9Vl7|5Odk4=d5QyIe)zfwg4h> zyGD_@mcjC3_pQ3agAJ5LN(jUz59jX{iO-hjr9Bf9$;0@9g2O?h(F1d{gP$-AmvuAb zYB4F`at>|FR&-^UgxjZ_Wqvme_dsY3T70}Kcy2We(B((=lBE}7!w;He+q9W#;luvh_|`uqDmyK#$j=mw}U zd`bGx7h*COu_t*wJ_jANUTWRdU{bzbcM9UhVRdu}5y264|1+)Cw*QTFdbSJypLa!$ z*Up!s*MV2$Z~* zu^3p)?-QE>yjIR|l|irps|=RK=8!!opJvn6`1ZnPj&Y`=R@nJ<(0>!4>!)d=cfr1t z4+zGySC3>ci+67Nsw41+-!tishp8Be-P<&11|JHM&;2LwNX#EvxkKfb!sbNy92E{O zi+0}{f$kZFA`iXZ!Cy;iigj z=P#84ZleSL#+U}uZI4p}J^EV1sY~w~b|lu&C>zInxRZQ~NBBs@v`zaBoaw5^ZErwMZoOsWUsiuyPh#?rs7z`3nScm-%~>{U@jt(Y$dii~I4>gR?PK~iJ;r0c z3@LURe6Cdvgzobczv{sqni1cg&sLQ{888^q2Atz{ZWD~LkyYV`g^`N( zM>)Eve4@ly$pcE}c|+^O-=g!eWM&^Ygw4<$y@6BqK0(;9P(#1|-=-B5J_m)@DpSlx z{w`C~FmR#0KeN@RPKt{E?{{pr+jO@_n}c@PRwigF+0rDkg@MlF^McR|+pp9ksXdqW zU`Hd8QCbTca9O(FLS4tH{Fij5gGH+#kQ@zCvp8nf#Yhb!ft8asxLQYN^H3?Q?O;al z>hVBTVbzvTCXdDZzgU)aS@``{MOkyfNTL>S%9}RT%-xoAg8=$fk+~4*-9+!J>-p%7 z?8IJOu@d&F3j%RQYRC|oE3xQLI6Ykox;Hyx3W*`(*Z0-wg<+_{Z58oKKE6QxM{^nE zb^B0QKz|4v(&E{qE5^pQy=sO8+eH>s-=whh1c|IIOt;9k*`}{nR6&o4AdvYDzs#|t z-_Jknb=}wAF0}ydH5E`_V2Rzj6tuS&>BsDuD_dhlgX8a`g!HidUmE%OW@~3*49lhl zB)q7i4%;DOuRbN6I>tez$XhR)ngQk4^T0Hm)=3Br@f!$*wvNK;aR3^EA#<>aT< z(i=p^->=8ng@?nNSAA#bDs9?4^aS$9I&g>BPkm<{+RbO+;jz^2BXVM^&ya*x5#mPy z-5KxK*jBPsLC@wBO#E%IUD{Q%Z@{Z-H6oN18M!yZ-TjAIXlkFFwA84K=>!c8PYJQG z)Babcr}kQ&G=NGq{RWUa=r=HKrXQcp*edan< z_+MVi&_#kfbpGwdR2aD&Hu9!ol-YL`zk#>D_Y@3{dpUhM43#{*r%F~k8 zQi16TFzYh0qF4su848+Z-oddHIaHO!sfxq0{ zyYG$sOAF0N9Wo}3LR%lrrQx+L;yjiN85(U5B@{;E*waBPHEeDMB}*rYJ55^dpnvwX zrmOQ1m|;L5gK}2;M>Wchm=ls%c-E{_E)>34>Q8n>4qJzTj(X-xo%7-Y!`6iOSmH;w z25rgSEEt&Ea{1^6ul9=KATPiV8`gT52W-x@%$c2eNhg6Ta`EXP@T>=t2@lG3b|(Vc z#Ib9|mmjmm9euz(yZ4R>Y_K6t9$=~RZLMtX^lN2n%j@I5-lFqp{CPV;+N>pxChkQ2D)4IB zFkMh5EIA+!-UQ_ne>T>-i!oDSYbMXDtJSrIHGl+Mjt6uDOlZZXs$?G~pVHUcrMo45(aW&>(VGn34nxK_5?d{>z9 z(&vNhZ1tu+F^BV0CQ_gr3~GfZ6Va2fqu$KGW}IfyUAV#FQ`=8mHZ$2l6E8)_)%axT$%VPM&3sZg}eK-gtffl6sEiP~#2v!ZID@9()v?@ zomJ`Xvv|?l*P-v80!Y5KcYo-4lSoVN>5eYYVREw-ZDBr0%QhCGfPLS3zQCf(KoHKc z>9mES5ku+Qo4ZO6yujUaiOfdRm!Va1?wq>7Z#1})G4ht45xLD+7I=<1kW?+4S(3Zd zSIu1G_%}E_o(}}Tl?1fVv^2`^%>1aj&+G{rxExVH5Ggzv$cp`bC;gRORa6pt!c0w| z{>gUkxDe-S3-x-I%RY~_R*|EE-XbpOpYua_bIgVL+k`pZ#ngaypdKyT-j9Crh3udo z@B*d!8X~;lMbCd(>K)=s*cyg}AL#{}fqsU_V1q;U`MMuAP5UD%=ox5Amj|n6^905X zM{p6dNxtyMdAR=H5#@rJ_n0lbN1j?a(hj;gntVikN{gaFLY%6N>tw89wM<%E8$VP1 z=f1L~ftF6-NQ^!Fz97PD#00IFvkvfhGo-Ef9_E0g5HRv>ew~2Vmo~esW_+X&x7sfUFQ$dXVtTwy#w}RM)VmQyMBS7 zDB0C@U$`%6ovcKG#Ep6QUx@euY$*7UCv~Tbzwm-Uy9yeLzMFLQns}mRi!IKlGl@#xxKE+EY@DDX+tMxwxK>@~F{K{z`)R07B3?s)>hdGVBi+j)^b0|29V zlipz1k$QjvDXN;86tkw_f9c(1rU-9x!LS(0LNFkxaCJ}Y=v6M8Xe1vzOK!b2hGZq9w?NR=>2bVWMC7aWeeE;F=Iq>_?uu#WCw_E;#pBxd0~Nn_u>Objd}Fm z@h-?P|IS^76khZ8FL)Zj%}y@g*yU5DfPi~c+FY-`|a`Mnm zxCr7P{#)u!Pf4pNt3jEjBEgLx^K_mtHOP8f#?$A!Ves1Ta%cxmW>2hEGpenkj(%u? ze~qTavT953FOH@GES5fww0$0L1c{DlKd=TT_x1oP&?K*0xbLq@lV%C9FeV&*vyu#1 zS<=(E&^a1-+-IZRED85kHy48U+5H78kF-_p=QsqR4nL~7Q+4YYSXOs|HE@qb-eYD~ zz|kGHKS}Wo`(m@`?#g$WQjd!1Nff1%mfH+Ge&{}_VcVHcUcv8ox;4`RUI|!D;G!wu z65~Y8PjqKP+Gd=fNul{jo`Wh*`zLGl!;VB!E^Up+;e^pEiN>u=kRdI07x^C^dO>Tb zqP^gQ&sQ3oBWu6EDqEj7&uGh=Ou8G=X}K*b@V99dwaeqAENL>mQH-6ioedmds|6 z@KQaY{C_c+Ts2itJq@1F0sT?MXE;s*BNv}0`N6*c3`!Zq?`&r==qXj6^047(vKJ;sd?D@d5&9dCrQnO#X9p$jFqLwenZJ;>t3TT~_-GLKca?e~j#BlAEJ7Gt#0Q9f8D{L^UTebQjaKJf=M&kMuwnq; zagf7do`+?QgJ-&OSj6wP0p_hXK|hZG1M!)W-L)7io9!zxpZPYy5Zo zM6Mx{Zs(&`4=Rwvq4tXzn@RUn?YPL!bY0UMS*Iua-(V6+elUgE(7L}DAez@?$rUjB-nO|LHtcohF{fGz_rFY*@?szAM; zJsAGGW}r2bgD&ap_^PGW-`FC-V z*-L*ep)*MY$N|5HdVyJJ&<5rma!s$OTBWqE-QynX>wF8n)kY)u+d_S>6%6c$A*B4W zBjh>Bz!@FJ$9bX|XZ$N*MRv}NZgy0`PyeRo+& zSfSGF5A@9gV!8j;t#fIBy%rT79U0F5Ys2>w9R;Zg|e=9 zXi$*;x)sboM2T-td%qV0YyaoS@dxGjE z;7O(T7psshC5-d9p{%_8>C$0!`pp5Z&|KZk_e00qcd-vpo%QTk6s7w508?R53Dg8n zXDW8V=>z(2Ji81^Gzor22H1f0|POd5M~8OK!@7+?HKa* zJt}w|<>vyr&5qYh-2}csig84vx(Vm5nv6%mO;t(1v0=sL3UZMky0V>Mhv%J|1BLb` zlHPc7bL2&vp>2Ne#ATW7Gop(AJ&-oY(V`0Im7nO?T*~;<{nd$vt)>{+g3_*3z{9oK zY%HjZPptN0Y6w`dH_og#Ipp4Igv@_;V2$Tq-*QoN?>B1O3sV-{1Co5#o;tAOCP&)P z)u#57Nz96J2*HII!M)Y&91e>?9=WV~T$)i~gxUMO*YcNbaOA0OifNJ2tEt-_1>^EF z1WoUnY*!8NBqxO-90bGs$+J66rM5az`E^XC5Vg?%UL|#J12R>h)Q)qZZ6ySFM&-iA zhe{;HOH)f(kW1Ml9aT=EB4vfr17`i(@2WAM;btu0Bm zC`bl!xs;3E?r1!9G_O}FLQlR23(cT4hJ8M=mq93@uLZ2m?7TX2KH(7C2hf|%JlA?I zRR>&wJck-&?tT<^smIJvwF^e40b77jwDA7ztN^$JYr5b5ppGG}$U{ixF+M2i>hWn( z2fe#^%08Qz)JqMjh=JE1YW3>%*iSJs?>+ti&qS~wpS_!gHvL1tA4jmflHh8SF~@2j zP82l6(&~S+;X@k;Sofq~fH&F+2CuHy{@~T9Y-7OxHdMKO(bQ!5GwLrV{X!V47e-!^ z9O6qy$!O1+Y&an8cVZ`p@IF8f0=NdQ=ab(wIL|0UMe*ci#{c_*WOBmcWkKgzF&AG0QbrQ6YB9i(SP&dJQ0*qFT6oQqW~Iv1)MVe z1GCmm`?^uQnhw6+BSm;Duq8ZPF9j*uKN+iHy9<7iG?Vlzc!sSnQE9Jmpj46i(o;eppp)c41`vRl4t$mZi7G;2bPYf8L{ zf-u47E^7rzk6mKvo0RudFkX#m9s`)W(v$~QHn#X+d;UbTdno}eZYK4_8;b1LxWV=A zlg)Crqq=hXT0{B#5Ix&!T^IZ(HBkdEQ{QqST>s?T841sJ@g*h)7BvUrOSc zl&Wdo?75RfJd!H@*B{zennZ-)S)r;n?`<&(|5)eBozlvk)=Hair8eIx%UTO!Eiw3SV7y8& z4<#tC!L){fh6Zn8>F2_dW)_I3rgbuB%(z$ckIcS~2z0QZ=}h_gD1rnl;bCO3qlZJ? z{0=&+S&pTE@r_p9wzC$q75kUNo!XI|@E1$32-{4=C2k0eOUhU zFE;@f@D+OSnLOriq-q+-1fQvw?ALn&4LB6i2NBKeOXafHqGFM#mqy8nls9@oJ0zv* zUSoP7oR!2(c*4s1tx-8oc)qnecE{@z!WmJ7U>fv=Hl=wlzm*p2Dyo`(?Cj0eN-vtk z`otr&!bAgC2R6O9!R~9-wq|cllTYc56q+aXROir6P>XQBPiL!o7)5t>^ZHeqtWn9P>SFpHGU)ej1r)c80UwPRlTi9;iFDAwe_)>j(2+oS(?6&41!L$s8(hYr5E+6WZ zhw&(hz3v37J5?WQ(pBgQLkz`(8#Reb6*Qc^K?D0DkNPM4EvA$NKC-H#?Zo10=yPOM zo^LVdlFB~>wSH@}l157R)@UOUgy+brMFf-_pKb66ZS2qlq5dmxCI?5=ci=@)?QPFO zgCu_WOMR}j=OKw^2<@hOoB58fs+dzrgMhvi>?dZhpIZmw&a*nLI@tf{!_*Db2=e55 z+@2%JKcm-4l$KAV3#A*XUF;r+i{=cq=lLm9D;Oy^jZaS+DbowpR+n6boQr4M?c1jn zsW|f~<$1+7fh+Hd{ruke5PM?r`hLji0Lj?*c95F}?4jfT<-H#{SY zrYK(M6@>9fYY=Eii1AU}J^fl~SM!xn_%%zsdFacbTDqu2|5qZ<3Or}B@sXhBjo(H! zX(n4L`|@hE)s%8yZV?APeGg+KO=teD>UUK_b0DP!5~@jv7*oR?d`Ev8YweUx#(y{7 zbwzRt|7isjtPDCW9<*TolD0r%$D(#?h0=QT*v@Z+>QPE4*xKpq= z3r2*IS@NMY6}?wDXyQkn_s2Y?pG%ADjR{cT9}Z)Q1;MsT8nWln!IGVtRo`kYmCCP) zz>Q&6z(<@ZTn`7Vk5%h)i5R>LC6Xo8we7;4HgZP#O`Y6P*#3;0ER{huu+O|ar9kQxjL?(Exr#G@b2fSiVDDQ^Bmx0uBo9hst%pvgi>{+vB`x`R}~ zOK72Dj?bQXF#BfODE3T%CL^e=lH0LD7k1@$Cv+z@O5D7W1ot5SsY=kL+P9U~o)*&8)v$NXFa3nd|#8Pu7F zH1=a?@kI9W9E7)l{P=lfo;jCZj7(j(wPXHRFm(iB-I!_Pb5&MCG~|ce>TCg5S-HF_ z9zJ=+FMjs5avXn3qLFg6hFxtCPp38xHPvr@W6-FC=pu*AdG#D*^1q7&T(Xqn#*o`D zMve=;h%GGX4(2C5#^nr!nsezyGrV2M7v=Y$S$DROp~-drNfqsG;VZY31E;xRa{Jk7}mRt;K(*y~kw+TMJD(KUebM4T5`cWl3%h2mbl< z1gJIN*HMOija00P#q7MX&HA~F&$t-j539!FQ_uUwx3r6EqRg`xs-_sw?j78bJ|g9O zsY$;uc`&u%qMPZobr}@pfrpv(N2HbWMa?2K+W!n7;i+L zdCmM7p`TqZ&Pc69D)+%_>I$&M;E*k*qcIthAO!#&O_e!fJOn{;9w|FQv35)>J}-}9&9M3$s%}mL=R!P z1OT31yhW*-c5IxL%B_%18o#in!^cLZN%)f+kD-Oriejc1DcxQZ0aJr{nz@Id5k@VU z*4TX`MZfFK4KD3-ezIy^p;T1(15f%f^3#?e7=4fZ@4@7P5x1Zn88bzRGa~`?47a*| z%eod!6t#t@e$#|^FQ3U=Ma7fSBpB1-nk&9acd(2avsS!6l*FfR8*^xC$**yqIKm1? zl+#&#nH3)$P;r7WMQcbl+8pBzX$S^;DKXm9T3sQ%<~3o3^ilRi;HYWCf_-Hps@a@RS9oT;N(*+ zkS<0Bi|h53ht9wd?uXFRYGD0{ao|0S@nL$mcAnh9eY?cQR)G9zIKsBUb6+HbDTF|k zA4drm`E#b=e8~IyM^-ZxlU*xzDPfN15wa!3PC>rflrqaOF7tiR?gS0N=%cPUIFIBJ zUa=bVe37rfKGf)|M|FIcC^Bh<;xB9LD#_#cMIXA|5hz)25AS zqcf>qZ7(CRWj~D_qKDr<=|ZK@mRLx&wND0bW`K?&SD!Z~P7obLWa-C;uhYrE(IaXm z_YtRFfsfdMuW88KhOHc?&im~ABZ8=M#zZ}lufTvdEAD6MH|r*vWmC3p_4{J##^8t6 zYm+C)z4~LPu3nG(sV~*ey+d*5UAi_6Df_VpF)lHhxC*~xbs~e6?lg6(=J0jj3+$~$ z`^LF_hlZ0FCTPwI9y;=~Gs&uZY$U`i;~!>x3l5$|tYmtH=$3V3_Kni{tKvWKB${$3 z-h7-TjZsmEyZmS-r5jGe`ETUc574|r=jte#~{^lvK~F(aiubO$q%#yop& zO1Zt%V-h5=&rhmEn7&|NTd)7)zdKes5*Z=O`DYvp$`~7Tw&0xjnK6DVt`x!q5P1g7 z2D>vo(Y_Yhh5wLHz8b+uzdL!xWQR=it>tzab(_ShAqpC zYj~)t$>pA3zQ^!&w$9+sYzz7NhMs>f{oGVDa70kSS2Vq8Em&OrQUP1GaO@BpIUpoh z7IqvU??r0H;$b4=i;oaZX#IlA2+RbmSz>20&+bJlT*WUM4XkzuO%Cj8tBC@JdwW1-R5;Ix}INIp5{%ZSTqO*EZ@q} zMX;oPavw;wxj*NLSSpYHjLaenJxyOpJmDv70K@NDHOcF4+irP zIe;Y}b*Q8sqiN5#_E(uDk6JK;#4@U5gLT~J<^Gl5o(Dxr$`c{wDfOzR@gt{tVQw}{ zJ$;*CsPXtX6No!eqx?!*1qIr+=80lE$J{HT;@Icc;SQci>OJSUAuYrm-}Agl(8PzI z5mRx$Vp|?7U$BWNDTXJp6LY5bIy4C4d7k3Mgo(s7kosOCZPT?@`jFXu2#fhYDeF?fXxlgC*-RgR3M<&`kj?R#Rn_=K}<_`MYWD0E6#C<{kpv z1-;?ZcwhfGH?wKBfo+6-J2A;XQLcHj&$EpoM5vZT?hueh8l+PJX+&u$>23t+lJ1s}7NkMCn?rYZANoAo&-eGP^*-xa z>-=?g-8pm3Tr+dO>6ZaN(pw~ll9O=>h93esP#$7sRqc9qCgO4J-L%z+ECnE*0}oA> zRuWu14b7XLOYv_9jz_EKGmO*>*Eowq*#5}dcaBXp%8b^5S22S3j3|4b!6iy?wU|lr zLFqAttj1<|kRviUNX$Mn{07>ZJo-sqjuu#5+0fSHi=I42s?%P#hh(PxZF-OB z+yTlR8P^qOTfLxP&p7Qob83s4mrElt2torst?du#T_^nnn`FrSy1L&wW5-T^NmEI- z9y?js;0jwvbUw@4G zkGb@M$^>it0?d8Af(0Q`&tRNpQ$E3Lx({J10YYF8tOJQ)>+=1JI{An0gYZ{e>p9}2 zSdB~b4ivYFc^QdXI2i7&E1u3c(=Wl`%8!HR8#Ef##alglI*Y|mA z?cyv6x@CGW4zpj<0p-h99Tv+|VeuJpaS%s$P8CQ?R+GG-zY^lO^nTN<`bRVVMe>UB zKfRqF(x;4h1*WMaTH7dL(&ZC`j=RW@2K0C&3M#nbxF*Xu=Y25E-zgWFitc>h5UIB# zS@xgM{PQM@M7hNdDMvoL2xg>?EWY4J3q{E4)r&kv8h;Vo6=2D`<-DCT5=w~FUqd6@ z5ChR5!paT&h+UtJ$O*~pdzt+L<~72}uXSX#@{^jc$=fm&V3)a#0f%SH?5C^KDW~>w zo!ke_A8pqTw*$( z5|_$WPrT0dGJ=;^bDwmj47ie7|4RKb;t!7_W=L0>eX0t-bVYSrg`$?pA4Sm##wc_0 zYM$|;aM8C?8q1WqUf=}3h>N>v=iSwhqWJ_0J;xBJGy))R2oCM~uxY3`RA@IqB7bc0 z`|}f==W7)WYirj3q{vK5S}612eB(Dz+@U`CTuYE$b>oWD#eaX>OTQ6$eY>uYwJn&j zN7hm-1s@w2QYc9zv655OfQ|`xQi4dsrn*l>lMmcFdSNnPs@auiWE?1%qk@-68b_L+ z%;^6Z2!VF%lukt1bnWf;#Qm0RiO5TEXW=0IRWpwlo=;_4o={?zsDNT-t$SgD_JCO>ECbEWNbP6I{h+>Nvuj zASbR>j)ZHkt+?jVJ_D|w_lY7s&c5X{D4O{&6%y{ng+bu`le5vZe|VJWTUXoB8;DO< z6V4j|PF^YF5EE1DRWm%O8s>#$FfH+7<4s2+kVu)&0^k$5263@u9F^sR+%{YwzkN?N`M>j6+=5F+l6Mjsw(Xv{R-E2?OH^#LYznfD@kXu%%5uQh_Be>XbF za>YCQ_OGNtp61v1oaFB|)SFLw`**So8?6_={S4qcz$_sn&>$BYk9lvwsv3l7>{dt+ zzE-~vC&(1aHk27hM)cGN_CC!K%$$;VTZ&iqYo03N8+HUnN^D%*?pJe1XY4fia`*?o zn-vkUii;+g--++(QGRNixv4^u=}_T}`V zCq~-@UaLua1^8F=4jnA2uZ_dekTK?Z=W1Rp#xr~xh%b-h&7Zdb-*}2`%ab3+uFEqi z@3g{c6t;UXYB+j1+`4-yvYWB{;j_W1U+f(zQ9cbl9X&MxB<^QE)Ofq;2#yDsH$4x$ zJJsZ?>o|iM0Me?(p`S3U(oN343hsq^e*^$oloz}skX4h~=15tk3)h4J*;wbVIYczE zaJ2($=pl#q()dX$>N^qQ#dHN&&&70nzjx74`7F7fQQ@g!TvcgZ+%WuN(n^V^>=w7@ z*GucHg)EHMC8eS8Pa;$Zr&NIdYivvctUFWPss)1qvp}I)|7H<5Kh7WMDTYQL9O#Nx zm7>eKe+ZEw6wnng2tDCn)h4hqN@({}<`mZS)`Z`QKUVJl+-F7ge|%QJ53Y5WOoLW_ z32%pyVk;M{4Xr)^cW*T5kv)AxPW=q6Z*~U?Oo+UMp}^q)=}62a|NBG z+(qz;c;E$$8n{T3DZCWt|Fx`yj(3&%_>G!H+kr%;=!aE2;Q+)|VSjrOr0Qp`byc!Oabh&U+KS{~Z!nphn|D6R4hmd*9#&$B!h6};eJ54z>im-e2&OEX;A0+N~62KN)1iA!ZaNth}To9zK zlXGB$*@k3DG$+t&i8BrW(70UT9XUOG+5I-W*&i26o2BckarXQSAf8lRo7($V%jc%@ zm9#Ool&%LtE5bHR64+uw{=im0<|lASI}B_wuvWufEt827o^_xcH&-X&Ve&0BFQ~f)M?I=Bpdec z%kJsOQn-LCfCU2m`slRJCR27 zQ*q)wCHkBtCIHZ-i~Ek;_XR)!XP7ahs-39XT7i_B%?cREeXK4Z&8+1^UoFq|{SJeJ z1K^EcHlmVr+#!y9l~HxOh!+i7X{YY)6a)ZMqJWq}D}cO95@9O{`L;e7iY+h{b+?Bf ze6cHyHNP19`~sfZ{&iJav1OLrPZ^K$MHHEw%Yg^C$n)o*r!_F1k86t5-@H?>R$g6h zHK{6AY->=hmk3v4r>X>OQ8)FZrmA*g<`*W0uQV= z!NkFengCR>6*^yh=HG`@w@70p+SjBF#Od$pDi!;4;@@@H6@X!#sL_lItZA_Gv19CCM_`oq4g%e< zVKL$mVS_?JS$`-tAn|Moz%!ALzQ;g_`OM##X9A=C?BG*m?Yt9>DuqevTxtMkWcZSq zGlT>%W_$PDO zhgikwMX6g+)t1b*nX1O@HuJq+C#|(B3DqAI>h=qZv3YN6w%`&*M`COsEv4Yznqm8G1}tn*+|DNZ-g--Oy!=sB>3hWn zg_xci9L2Y)iPiUJ)ybTAZ|)bb9wJ#KNN-egDX7aPxR}0BslCaI?~veRmlTZPwVH!6 zTw&^EqtNK~w{h1CEhnb3^9;srQcF}@{@UZkSYOvC@fJ&_-nyeUPH)!) z>qiQLEbR_dBx6KsqF^bpzycjtXD#Ds9^GG~@}@u}PXw>!Tnxk2G&ux#dh-PyE>_7{ z$F1oHp%i=vA5M5JngthGRMOU~ylxZ(u3=Zvxr<*7YFoFoLUY?h9jKS9s1kM^T)Mo zlh4$ADVr_+3GU^{1K`dDA8T?2*RK7cU`vI+If6cyYwyU~KYPH*J$4TP6*hLg2&c?t z3&#ap&HRM}&#P~&6W0~Rfaap5h5RTg59#2C^lX(@6+)McH?2dX`3%u$nimZZ3htYc zxc)v0z{lz%;6SR*#ta)v@(x)Q#)dMNcM%llbkwRKc}`6(C5Vuu6V zaR57wikNTDzG9$@j15|Axe1C6nE5#uU}CuQVIs>iUVo+2nl6kD#JO+{ zZlQ3xf0laHx5^-T)c=Y#u-$A}{dp@Y=ay_$jIu7F%Zp91YG_DW0G=_Zs1(kxvi6U& zsfS~H&XEiQoFDdE)=>`5vOl;}rwqF~yqG^=C~#aTI$W#P{=Q2HnWnB>`r$ggu`LYb zr0S+<>^$?uWJ5(!-YgS8xjy<@VWZBpg4fNAu~h z%S>3}fHQNX%|)Kuyc>HUNIyA3oyhNBfH{a zaRe%u7Jvz{x!aE`sEZYAu=6mtc=Btk+z}^91aDwQ9MqoT-_+rMcD=|nl?CmR{`KDT zcV;D4dd3|QIkbAJRNqEj9w?W!t=mWKew5iAS{#08pj?1$JP4!-D>Xrm7Tv~JA>A6* zKjyV+_n*_6?gg@hzm_YdKtpI%z|y-{*t2W-&)yh z1)e27v$=`hd_!x2;U2koQTo_uGZ56a<7#T5A>Ai5aPeAmFLrX*uJWS#sQn$?f!^LL z5F%`2!W(0qEjV)mheRB51HK>X__kaAP+B=cM;}|aXORgHnBy4A zxrTa~9+1-%Fu$i40g_VD6^9|)W-9bj1toK%g{@D4ZR4>CzkDa0Oo2X^9qyDL0w zT)VB-7IXBB@I!13Pp!lz3%7<*6F$je5ZXMg9cx6rTBe!Q{axVulyt3S6&O@D@3C7m zyUQ6|bnslc)asQeI<+4yPOh+v~qmJA-BYB!FQvF29ts(6HWVQlAeA2lZu?#TyQm zS+p~onoBZiGbxb60_SUK>{OxQm%=iy^F4w6012wVIatXZtn&p*cAGzrRDl-+JUD|* z#Q{o+^>{T-NHfAxX}F;%ioEK1Hvcj63frT%?S%iM+ah86#dMc-0)_mbuDJzSpoWs; zE2XHXbxul;yUZP(!RrQ!)pjb;n?h0WtcMH9N(fcHeDDtgl248Y43Lm;Bj00f6jf*J zRJz%{Lran6s_O|kyGqq}1Un4r)`lZ}J{n=+bj#QJFN?#78h39ivxJuT-J90t>9jd` z%aFAvQjXs@YniBuYaPap`fBZwSIr%%jfLp zv3T%7>ikc23ew`y{e8XRs67b3PoAInh9E}TNqA0M^^l_&+)Om`B&)2bnX<@nj-?L z9E1B(0|Ob`#MZ{wDW3d=(CJPDMLv?F{Sn{wi55Q1y7TRKr+(e<_Qib9EJw6&G#L&V z2Myc1T?cowTRzQ&Ums2d`rXcUJ;c-%W~iQLi(fgq%@xX*gEmCLf+DjQ0mOgXUVZXYvF^DmfLjT;`iKL3L*N7}{*O~Z3v3F|Mvm1* zRb9AT`#04^utvS*c>*WJO!o0dN_%t9>awHjMytD|1anVi z;UAV~5%&N!;gH7X0DF!VkHW_Fx6&p`a+dBt49@woF9?%zy?3MCy16MvA>O5C7FvcE za4Z(T7q8m(7XdlZ|2iXk=#dBK$?L3z0;;W?syNM-TB_3{HIi!DudUr;KT$N@ z;<)Lt!!=ngY=7?mL@@`<5%VV$ zz$WA#jpD=RkSK9#!%XL$UH;C9Jumm~&W@GVJTo^AxA5|fNaj#_a z8E|9zYfyU!@GJJ_+!(HyuF9?ap)^lKh(N3`9tnQ5!U)d~?HtdH(!N~xaNc4=2CivL zk`5tRh>QL^hl`hz-<9wX$%fg>Ev68#Z8@mM1~!8XGbvPd^}jJRpBg6eZ}~`12!*$= zcf{zaE!ADtk6bGmCY%=<9^R~YZLj*eyeSygmX4zro}lUzJPW5)megdP!lR~Fh~L7j z=i>^dM+?NU(B15RS+=}=co^6j{g|uJ)y9@9UU;1fX=at8l6=HQ1&oWbhtPe0W4A9P zKPP-#iNbR*%};|LCU_hvX^qLJA)$Ju+?3PAv|p@d#J>$@b&?RZmmKzKnOC zhE69j%WFSPSr*;F&Y~c3(K&c0Tb{7e{;;>xmUWRf)jCJswCD+dhfGk&la7?%+c8U# zXJM{KY_RYFItnPF=sgtEDr9KkxQ3L);cU0f>ZesYFXdt^beNO3=8i-;bkm!sj~!C; zv`6bPwepERw?&(Ol)D_D`|iTA_{VGPByD=~nvqN(B7nJ(pFq3Ot{TjcA~`rts4sov zw(#}s@FH1SIB<5|tu9j|q{5LLW)l$iM8b$ivW8I<7oS6~02-lhY0;Ag*rKh8nesW` z@zkNV`l|9UcB6RGYx2T?C&Abb8hz z4FdpZcBQ@#h0(18lb!6f>$ANrG?aziOBypEL=X1f@9?5{9gOqYaAjbem9# zwy^r#z`F7?tG?04Y!-n9w?>dBNzJd0-aP#cSIe%L z{`Br~yZHr$h%|L2;7xq4Xnd-!T#YsrBgMzD=C)vc!|6uhRqOE~!hbX)cbM>6Kzlmo zVb{2e?9^FVlwNLM)Z=z9p7RLA<%v57BT(^bk4O7~6)1N0g;bG*t3r1dV?tuy9t~$O zb(E}<2qp5+7R>|UaB!}E7#I@(B~~HEjYr$fKqZar0 z7OE~^?kpFA|CCIN_V9iF7A*j6cmJDi^*M86*wst+lu9PuNXWfKDWwm@v0}bO7I^E( zU3s3S;I<`$;fS%L)GEcdBe)%&sRSQ0&@vIUglEy!ls)t8|>Rl7@yW!pi6Pd z*gn&u$Hv|av#v+;RD7>i8=d;X7tpTAC<)YX?dWQCM1XuFgDL&?$t8jzj!zG+sIxxSYij8-Ws&y?2?Df zq~EnyLKf#U4;O?B26pO%ZF|5FE;BmZu-LrYCjB=H!cVUYxiDqYL=v?KRnwOv^Te_n zn)7)Ctj}xeYr?scmrZNZb;b6ecRpf%m7K^y&pcZ`#&&Ak{b>Md?(;K;Oi~Be-NUog zo|;uFzY?j0++G>&EH)45D+Y3Ks^%jwLnd%n-9^Z;?mF24fPnf#^N=O+FE~;)kQbv$ zVaWc3d#@y0HW)IsQjpu8NR95p4|j;rq5L%WHG`)>lR&Ne6B0S7Kz_Vs--hdVV8lSk zrmw`sx%a}g9_uCV_|LBmUPc@>?j7HwQ%maE;sw3o{((|S-?l8Mqt}u1NrvNhyK!!y zw63G$LaCjsX%1KxOtUu)HA+f;Y?37Ft9u4lde(xgdzOXb%-1+5KNkG zraW~p9n4=!4s!Qa(M3B5F3*1l`RUD#8=fmdp$8qvVy$@xf?xRCk#he*N#YS?eE_Mj zI%aBww{RmQ|9QYQW&tWzR7TLo1tcUT-myzYk6I7aFte$jkb-IDt>o}BlLQJux;yCF z&PP;lZ4}|2`uM;uf!pgF+z?5;c(3s-6#y8gTxl8y;o5LF@OHQ{C}%B={^z9y{5&FQMt&A;}2l{;Xj$Q zTEVt5f8|w4+@AO;pJx7=JR?C!7F7H0c%*BfpB1!R03@k}vvy{wI}`B51oUxg&!`E2 zQ`{!$jfqtr+;bV}!Nh+k6(INy7)O?z3W3SM0KoD+KK6cY3v7TU!lN=Jq$l(rV+Q1? zkGo`k(pWLgiv|AWve75-BLE&9PRR|LHhx8s#)P@pn#9llzWNPLhDT5}()))Q8s5Jb zPLA{PH41V^kpWqwm1wh_m4*0bW{< zy}?SL?+3sP{eFpi3;>A-aLURansXqOM$OcG?S8jSMxO!zj?6&(Y+bUM!c-0_I13WU zyee7*VdTE%lf0Pxsh76@x6d8iuLHo*gTHvTLhh1f5~TrK2pGe^eMgK75dUQfk2Av^ z<}QARd!O~)6l25h0t*!m-rqp7xEI!eXAItoPhD6H*FoRm4YYtp7@)3GkXWr0@f8WY z6p!2U!|c7hJebu5Ug6rMjz{z_Dv1TJ09f@BA`3m-yFuXJD%r^ZM~E(&dG5IZkiWQ)%%-%GXdb*@6H(ej_i2l9CUM1_NI z22NjTX>>4lU_3mZ@VTwjT5B}w+W8%;aG`K9$@^+<5UYa4ud@IET$?F8Yo>!SY#ur< z0LQ2jo8<~ow@Q}X4uXy)DUJ6_K7lzBV8j2r52s|`It`r%1S1_uwmw&Ji)EcqlpWgA zBu3z?E^-+laT7pq4D_QBM;D?4DR6)}^5lmT_*1Hs{DLT`iJ&`3q5d0?akUp+8mnKp zz~KY9_~86E5$6yA@FN`=a9!_L^jU+e!MhA%>cR)C=)gxzUw}pD{g*5cr&FylXW{za zl(&}GEZK7YPQT4r-90@xy6H3b%lgLKIK;Zdkv!CUPLsX5I6rZWZ+9dCI>-$VTpNd( zNh6V@#q8be)z!_fj55p!&HmjfBP~i(={5+JN(29~OI|kTagv%QLJw)tuxE__J#WBY zylmdktV~XI`ORC>@25CHxXLnV!GPhr9f$c0*IFFR-aHym-b0v_?*a`}!JDaCDs#q& zu0p+Y)wz(bHHxFMk~qFeW2yrHxcHpp0b^oRe*@9_j3KwxW;dTip(ku_wXSu9P_h;l zJL3KM>uxIh>w+UGJ??3sNd11+Dd4=nJ3lD_x-!4x(CMGLQUps z$3=vr#`+NguQy~Yr2-9FntIcYygcsNI#`b`!myvmJ_vB3UBrZOXCpX=R%8ED?D9j^ht#v8ajN&-v&d5GCoP z4%=aO&JTSNC@-UbSa^d{ps5ySSoLM|{jSS$;oOXdMa5fW`<}Y5%S|#YNpUSvyV8VE zk^BDWo_b{?4inDGkImE5s|!0sbq+|=vvO^>x@F3Wd3&$5a-ym#*h_ zZ|ropxK-}iqG?sp?zK`4jIbEs09$EK0auM{u zj##vv-FexSIy>U0q!>p=*DPQj4*rNv8b#fDTEc0Xup!fZE2BY}6%+6d_$RoBaZhKA z83$+}GE>)rsQ&Dz0D8l#6?dm!K2l7{Y$av<&RO;(>aqm>{@TBko?iA@QDxgisx?HMCyd%ZW;xG?5@mL~_o%LrUYrKmpYyZAakXc*&jGAje&l

k zOlu_=s%T4@P`FLisQKP?rY}v?d!d^d{;8v1aL#2)fZFM~t`bqU%y$_;%`DkVesK~m z=r#4pGRKYbVf&msdp?+=7~Q|7W$^3{J~W=C*vo-=PV2VCxaY;0tg_XDjx<+ zx^T&VsPYPD(pa1Py;DDPGI1-*Fg-cmSxiwz(A4XrosdvR@q-CCAa0~3596ektnaz)vdH6q*Uu0ktXYE zt6@VB-Ik4ttRS6}aP}LtKb#Qqqt>+I#%CuJ7l6vo&c6>I*xI4suOJ1ARL%xVHc$C@ zUG51-4Qm~CD_-sI13!w}2A|(m+FLSfqq*`LtXm(%_XM??H>TU}DHC)f7Y;0?(i)Aa*(7|*ca}k)o`pgRRjCMX%?CKYsKR<`l4Eub#--NVRB_X zDlV%u?VyLR(8|rKqSbzN^m#YN84VR=PTqZNfY{OETC=H+QLP2Fr9|RLg|9)?KI`b6 z*uS}cE(u9GiM3%XTl7B6GSKCF6K zs5F`W@KHIXdRTMT|LzPaUu-6Hp z%jTnPc!ccypa}a#l)FRj&6Mp#)BTpuk<#JNTy(vreM+77tk>itEV_4eAeg$p%_-DK zL@@s8ZlW5JdF^((*vxYM)nH^N>ZDh@Nlp!98kam0EFvM^&mEU2YB8pTS)$SMix;5H z@2oP&3#ZlQ3k?s+J=)T_WEMeBMNPKTOj=<+F3Q**4&r_n0V<1;$UTD;?Pi zdsH<~J)GftmA*Hy8r8KkQ45W?FYq1O;Eke%lEY?rLPPb~+;935UtqF9#qPhhit6?r z>8!V18J(WZRNpwN4ZhpD5x}MtrBZD2WSG64$h9Cm8F-MhOThN@^7{N;+VU!Jt|ltI zY;wu3?J3IQhKlTm^3UUGHruFT`zBbmSHj4)Uc>L3%oB~yb^Q4HnrRlq=Z(1JRV&pK ze8sBc+vlgfYR==5B3aJQrLZgSa>Sh6#7X!sBO;UK7WPR36ZnK75JTo^ij>N zQb6NLWoqNPY%;GjZx@@K_~u#8?l>QUrIjd}L6223c)wjA0u4PLt2Q`5av=Pzb+glJ zynu9_^~PPIgfU38wROdtSzg#p!PoeJv^*CJXuQh}Km(|m;HiDq&>7)s)-zl)?fTV? z--X?ITfKfJuvWdMGg*#Go`SzS)?!;`@c;^bebLfQgROdcazD3Mo*(b1BWH6?U)?y^ z>qWXd=r;OnP3k}|%=W8cPgJevu5OGmMth)d#a(SdR(qtBTjNG~_@kBOMRY4Z!xm=Y zGRi<||2zEJlkJG9Jipt~-dT((EGg)|`NzSR7FySqbuAsm!edL5UfK6$m7?e46?{q= zHWv(F@luzj#;)|n`?iuieo=1rh1pmuZYJl#YSA=Sp(Og$A4Iox_$h}}w>z-=L42=< zit;*!kFk|^(*rMT4cDI3LhJV5Q*?rr;I# zvXg?|nXeX&46M{!Tm4EuCMJnw8cFXgxy$3RJAe}X->DMsr zyhKK7pqW$jM$~e}t0~v3#fh~ws<@dyxkwp)Zm7)0M!$YG2i-7qV zWz@Dx!-#u&mHImY=jh@x-7tSdW}Y>J{Q~=>9I?6O-In6mK~-|)f?nGfHYx9_+UBy_ z)?bDBt+m3uCx!`($vnrz-fGR_mq>TbJ{N-pEoc>gmU^=&-uziR-P?sMOgT{-%vjr1 z8aW9|dHZS0Y0Euq-`ucKw5~mG@q~)2B{=V^v!5O3CHY`&Cqe?nD^WCF(GYHFt$N=KY814I4`Q zFe!AY{-djc2bg4~0B;*nNpMRmjb;|Nm9cA+;B+PhTt{E82A)Ypm|`{8A#Y&`0>euR zi*L)XJx7jM8>&a7Euvin3@|E|OsRI!g}h{aL+E%uS>fdT2fg)i{v zBi-!7aW)ePTudCjhE4m%w?qF&V5Rjn=WJXofIk;~T(Hj?UBZ3QnTNHO!|q1e%DpF&UYb;V2U{s} zO3<6o!U;w*xUuB*as5m`oJl`l=kG?}_c#%DA`^CxDCEMN#anG!AFo0vGio!a%%~b3 z#iYyM8i}j@aF*?TxIU0MYr_*o$u@hlv(BSLWn|rc0%7*LzV|AFoE}&SI-d2HRBWyL zB!MO9j}IxIYF!eVDF?TB6D$RNS)BL#<02BF5EXNWpoMhk%!>D}=BzLB`;PM;GcS!d z2mr@xeFS_!f@54`*(cB+>9}@1lXcK}$)LxtBv*hhei@7vV6Ip)`kOFu+Hl`|6~L)?&-jWE#0AZLWM?=R>dP`P+|lQ#@wHs6z~kG_&4&`yYMg2GlNw z4EEE&%{bY0m@(j*xq?)vO3m^d@+9uFU#FI})iCmp$GYooGvmJN6lGZAb?2B)8KrRp#4r_cQwp$6%Gj?dyk{416QAMt2{5q-PU zKM>qCHOB~P-xmbL_V_1|CF+EWWY$JEvN6ng?=h{y&I2}8L?;F){u+6$u8dGRUh!{P zkubErjN7mmJj*|Qy4}X+d|)h5G60FIW~?Om^cq0kN3%#&9uDY#dG;0FpB&BqM*tfh zqJ%HAzxv@mR@)m72WD-R%u7q{enagfJ)gGb^GW(AzwQNF7Ct!*sv2{r``gEU>bAV@ z&n2C^-?@bv!!|tRXZaZCeE;GUkZ{|CC=u)W9Nn)*-|;L5jC}tO-XdYKzM!B}i{7~jm*f62Fe1dSRKL*tWW^?m|I#*f}wJ|4WL!7o}pOAo68FOu+ zZNxp3FmtAKbMNCjlEQLta+^1|Q$h1OIn+D0)@St~3yd1aSX^A*(p2v|Gv^YAE7foC zZj+q#djVKUfp_R$mD2tGn%#XrPiQKEQRky_|HR9jOZWjZ1d+)JQ&!c_kO$$H5#x-pkCxGt+56fS7<}TE@xyhQ12tOlwNTVy(~573O~0PF=#^ z>o|@4c27jDlzgiBFUMQab5;HH_}(S-4X76D;aNK+H;Hm+Tn*0G=RO+K>>OI&xbS9+ z1nTV8L^VT?Dil-<`R4^ef${~OdjDMsona^|=(zxK;SLUQNJ%(%{0164bv8lAe*Ic0 zpF(t+^>MVOc>Jkd3tn0{#Jpm;dSUruCB?z3-exncuCPa4P3~duLZ|w~iDE+Qngu=Q zwrJ6fS1kEcdD^RHGw9A_T~u4qRKnLUPAgBrD@iANj9!yb>L)Hk8BBc(!OnKxd)30r zfn%r8YgZRn4=+zIpXXfkOq=<8FV_kdx8?}B!HY|k$M&ZKZ)QUW#6{Isk5b>eECSI2 zR63W>-ztiX+GaC&*~he|A%8eBNNf2?^xF<6_p7eyJxcHrV0Sz(v;wm}@YQ;rz+r6gq%dw)lL1)c8P8 z*So;+`SAtu39Jb=%;U}yiiFZXPXVoL83VrjR{E|`I&6CdeLdxqH_Nd{_~igU8sSO~O96+3k*xnii!n@p{ckbtB1W7cC2E>w_ruLVf7c`=l)3 zQ|XeoQdBBWGi&J=h;5tCDMGf_0h`)V?Hun>XF|o`#YTZlBaU{n*KUonHATYCl{2c4 zV3T?V-jQWXJSl0<%u$QeZ{ZQVf?wuC2aY@z92}Lpzjt<)p<-uq_e$oS5=rVD{=KVAsuQO;_kXZG$ zz+`e;%X3?1*Qi}sL-q-(%G1vZ7i^LBQX!X$Au6r8$ww$kt^Ywx_ZZ*F6SJ1<{Q z^sU(ZTGdWCc;mROPx&+$KRcT=%U}7=J}XV0`4NrAy}KAg`6fN9v>a9H{!fWdmvl@O zkrz^T6YJb;&_j&Nb`D+do9K3kk(Zj*MT*zV2~%6m)oq*r*@Bl8ycJuSe*1Af7+yB< z;0P&!z6n8|CiI)~velcX+=0c*Q1Ip=pgZH%V34o+Kn z?r%1=u7T%)EI)J$e2JzMo1NZ{IWMpEoWG|KhI}Y!KK{j|h6y<%@>I+I(r%`vrI=T5 zu_*lzj1U?f!yzI2KMrzia9V#;m~1is%qu-mPU^MJ~nYe%;^c=7_u=b1i-Pkb>a@|gp#AxqD55PQ((k3##- zm&#)pyX0~-vV-%uRSc-SW)W4y>tqCC>8KG=0Y^r_l;w`@pDW#4@|=&~*__f4qn_7Q zZ9+6DXl0E)gf*xbyWN!~LHcJfDHgeVgOg%@X6GeJmbkmv+O#d=DpB1XD}uN%;SZ#eBi)uZkA0`nr%l2(*cD0Hh%Geab&x3DG9T;f_UQJ6lp3qiEqeajzP8*hvyO_3-U za*v(jc5nO1b)L>5NY=a`dw#>O;52NphaqSaqXsjd=k^?@pJUZ0n>%fW>BmKjY_ajR zOw&fKc{k9U=FGh6+Jg}&WWkNc6%|D1g5?oaS}_^>1@MLj?+Ul!g2RnO!z35>4BU7K z!$^G#C&2WLQY)!fQ54=ojC=6ny%a`She zAVL-6i)V}3{4F4x-sbt##P?f?XF4&)m7l-G?P%`KXx-E3cD*I_sZe* zc70tFp0)7JQA}FW=~APKjO{nk&AmU+jd>G$O6x0heCE0x?JQfbcE^#MOnr!!naJSU zafaW+-bBI(K8ZFOU;D+>X{J&2W&N*f7fL5tg_(KpgX8=@CcS$VFcRR5X*c@S%Af05 z3hTi>!#MUim-}0Nd*A7oh`TW497}#H012dhs6&kjXReLEJxb;)oKh|2YtT#e&$;(J z_gYmAWb~T5kzgdn#rU*pe=>V$3Ha5rJJ82pKw$f}4VvY9pBPs5C4N-D?QBbw<b@!sJ#(1-qMv&5q^IzSHA_y*9~@vA;c*PTm&}naNEUgNY-qu4>f<7*?zw zKYY}-ql?4C4EbDb?B^y28AdLw{r^&__N9pq^0Oz-O*XF0Bjye1@^~~dXs*A(aJ)0s zYxVNlrb!clRApEEzEDpvk>9?OlXqqYT|QyN%55Rs)-`!ZmKJ>A+@-;}&h0+;U@S^1 zR#QQtr@zl#1P|-3gB{$5#>mv|Yu1Cvt$xAkEp4Q`42RXSTtiDTcrrnsuIRcP!vr;v z2Jm6+b!|n0w&@M~2B*&X^@%$6L*G`CUQD$~&}+RpcCMTl+q>~|^!pXgiy!ARSF78zPcI^g=i5Bfrv4D()ZO?e3WTXqsK*dHYA&{7 zCNPI+1vpV^hYjDpj(dzV$1qNK^mm@M{}`wD7Vkf-Puz~=uS;`$&WsAWN&rNOPdTK- zxwJm=+%=6HHJ?a`Eh_3uN|_HV>DjuzR*~!Y{(&93o*JZ4maY0UkxcT#8S{fAp+Xv5 zETO`bCf5sK>r51tv7ZWWh!rPGGG3;Bb~DiQjUSRSFA*+M7B3c3B)*`eMr65!RYYp9 zX1!yzVc0pkRKY;)m&;kL{gS5d_QEb`kA6T3p=DZEf{|Z5fTRfrHftT#pD=OriN$$Q z&)^4UGR>AV2fBaogkTzqaxJnDGQ4pSm2*+)E9jDuWK?_TlFNmTH4ZxxPz&Ex!xA$x zy`}&FHCP~seBl_U>pI{2#1n#aW3yEzV*FDmlpYNvLa}VPH^K3A&g=AP-0dc_ck4=_ z1BR0T`Eb~Tiy_WG5>NslK(^X%PhCKN;qAy{=`R#oRN^@4nS{~x8TErlx&snURP^Zz z3K{mX_yXLCI53Pw95Utn9^Rjw)(`G6yq94Fi(9a|Ibz!bWBctORvBho)_F=c$~cG z`DvLENHHTHo!r8q*5LxvCO3Y8(}jYkG{n?is~QJZJaBrtXXe9!n;JpWlu0s+%S-8u z!RXo$tgzzh6AfanKG{iF9x+A*`Z>R56zor67%Sj)KA`xg-tGsWnNv{o-KEvA_{f-8 z!P<^%y5MKi+0MLewpNKl95P63K(yOI@_xh5opM$Y1(g5a0@ROZF@`nU3rTgs`@?5Y zku%8PsVS9wb|pco9RM~kSn_GSX|X_(7}OS$c@-DEZ9hqMSKE$DcwRTcgNC=bT^2ANJvR{2qGdVNQanA1S2J8V9bTm@tEXLL-$@Z6tCRG0J)o)@>w-w)xC| zoCN^0AehVD&;X5Mm|Eth4iSkP7%cy9*A_x?TSEpdd8lO#ubnb2R` z@f)N`fX74%=}1bDcl@mIoDS_jFD6a|=j(@}49pKc@4?&hTNemvr%7Y7na0486yYuj z{(y$YmL+D^z)?yOtbGLEKpdo3)?16ui#4{e4;HH}i&0Lp^nVGqjX7a~X19i2Hh0Zl z&oSYD@l#TCt#@kPi%q;-nVHevZ!z#v+F}BYb+%HI1*hN%D3hpQguS=zb0nAZ}>D|80m%KCKRi#>+~N8BJy~je=pKVpp)_DNRDYObR?=a4ir|C~GFq z;%VnIrDNnqFjt{sTc_Q;G~J@ zX>Iy-_@EL4PO0ArSzk#IWA`6P{CAj<4#_X3zSMTZ$UB^!y(wsbCTS+R*JSpa%>Qcs z@W3V?g$(3*k0)$?J*gw7RW>N5^RAZP3Y{oe2%7$Cl#Z4B37`ac;9GJ@4 zP~!+SNTgVaWJS5Vp{PK%6sVqOoDxocp|ACX#P3J7?TJ@^$XAGz3D$GF9))MmB4S6R zE(&j+lKC5U4h0eh3riZ%^7c;H?6_ws{uA@DS5 zTQnInvN&$6?LYGPr7h zml44tbFm?19<^qDoM_HM4Ey1vpQK)Yk(FGFaA2xFLCGPgjSc0JWrzSJ9d1DEg=>39 z1FE@)cgmdEZNySAAG@&U)TTEPBn5Q#-+$pp@#37nNCfr=g9BxR7q&~+_rGuVD7}$(< z=lS2NlfOhkrsfxH9_(R{XgRsYArz@uP;uxbQ(U2_#gf@WXFn4-k7-ZCOdD3k(zDjx0HNSY;#fLS8?}K|@TfRB>iOk5tdu0f3 zNzsL9#FTp5JiCMgUCrUOpzl}GlmMxe%r9OcWvRubk`&o|D68XHBj`w?gdD_Yk|dp{ zC2Zjlix@cc8YC!D1XBHc;41)`v=M8Kd-}UJgDL)S?z=JX$3&G?>uK>>#gwgtGerhM zmP-DxKxl8B#sai789sNc{cRrfG?)0sw*MwEb->CnPDZu=T1v0O)SLIKn4 zGmGvJrd9B}M-;LUfGb7cY=)E-Z)!vSnYy`ZjkKYj-EU(`-SZ1R#`oD)FOJ*dCCQQVMKa9=PEm|bY z$sqC#q&}CE+&U5HTYSN>xsu{JrM=#I&=xwIA&uo=t+(lD4dw2_oiHS&vx$G+egfi> z_Jx@qt=Z@TB0KBW8!R$;%N#sw`Z_Pgt08@B( zSlGo%h0Q@YB{J|9!JQ*~8BWAv@19zjQCxYn^`PV#0>KN~FvL}bquB{^zRcP_CBxAv z1og8h{)WJNb#5xsirGmg=s+Be3#g$T!qPrb_Tn!!!a4hdG2db^M6UUQ6|6KLIF!Ux zEvK7bjp*~`XaBscW(A7_aB+FJ2wz*bc;hSTD+=@B{Tx=gh*fmHs%*e5B@g?p71|dSuCS z(lNU@JyOxZ!Oq!TBJrVi^`=S4ppE)eCtV~VRg9;^3iT;dB;m&~opZEjQH0Wl>o~DNz(U&b}L(irst+NFT_0ldVMuS8Ku+BORrqCdQJ%2oflMu=h@rE2Yl>d zWmQX%mWcpH9bcj`XMUpOp=)Ru_8arBBcgKgCXy4U!pojg(ved0HXy}7Q7)gFdH)Vc ziQlbPfy=ou@*&z9^Z6^y?e6TgH`cgxo5MNI>`?XBB8Hto?872Og45y#hWEwh`MhbB zRpZLDLOqr&X!^=i!0;&9nN?cEF$*@MY0BJXnbMrFiprq{dmP`XK1HD3;k`r10hcsn zO)mQf)2`qPe>?Zx(cJadq(qG6Du+;tn<_U4jOU);ss{sF_KWmhZo*zXm2X`;Fe&8! zGIw1t_$_>G(H?^#zOioBs)_Ygtb9Zoy2{ifYR00#r{|CUWr?Oduw?2?zQs*)(MQ(Q zZbhifc)wR+p((b{G>0-9Imr#(e|$Vzk%#T-Fs2wpoujVQJv2%-fcBj)(ZBgiOTiG) z=dO1%DzjH|rn@t-tnxSMk8jfYy7eba4tir{YY$`Jt3R&xO`-Po+S*kKJL<&#C^tVv z+BWqkcwCyX|Li5F5d1cUE=7w!)IQvVKW7t>pvuLR@u9XdmT*8Xz1x-bIPKE|7q78S zYJ`GfHdc76 zVX5|ECGo<6@9Pd8`&&f>{#dw7cdgTrfs` zl&%#ed6L%T>b1vC+&{f{^2{)N zPPBJgJE%^sPXl3;orh;|GO}Hd3_8E~6mv$J5jgjt+l}OrojgH*PjS0Hj3NV;`I#WSg$X9o<1*I=^?$IZdr*?(k4N9 zdK_qG`{w06EyV~!BxJ5kZrJAr^AxA)8;<0~8Gq8Vd`_JQG9Bd1gF%2#3bapu`|`Oo zsw8=ED- zT}HX<}c6u!%Bqo6Y^7dtjYTwB}FR-RIfFqKrN< zV^zIkP)DkJ?MhZ(1<(Fbi{V^7RZ8R~EWW52KOXJriOFI0rNn0xUirXJc*(hWl-hWy zo+!*YKyD!80C5i}r3z^rhH19}$$i&Xqka`fZ5`cMTkpbkpAnthJ6(OjPj@!@WjB9j zgfhCKf#EKqq7<&|@46QjW(upy=HZwK_X|l@4g{#EK%kQWUY?>(3VP7+uy}0ta4VS^ z<;t<8q%h}{I`CHJU8Qk{Ac+<5Ox|4`yh#FE+79Pc`3wY;k`vq%w6!5m7UMFZ&;7TU z@9qw5CR4stFcSrPo=5Xh_P#Eo?6nV3?j~a6L=!P!M@+6TogP8D1Irr@6^BRpO+Io! zLt=|*${u^VC&?nwe>Ou`zoyc{|F$yiH=+z>{rzjIGrTv}5y->^H3&s4@S_(@=>knc ziX|ZYPx`{GG(@gH7%Wo|dCv|g@7z6?#+%tgW0Y(U?7YM-8-BK}y169@^dFUgcujh~(&;kCzn+ ztHJ*8+~1LwDTA`B z3%4J#W{J;&gXZAW!#B9vYKuO@?qkI(2 zI-TQc6+RJDvyLquR2GXs3J|8~kq*BYPEJQr{gR_lo34uJUQ?*99d0g(b?|TpJ~e|p z2SV*GBUEFBC?b8~-v~+`-s3($Q~`!=X9k*%O3_F*lYC9K0#laIiLN$n-c4fQiN=Gb zE6ve(Q+C&^g$7#);58$HK4o%?175RBVA0t`a8RBco=;}Gexinj%KSr}u4`Od*V%ko2^2oUt; zKzKu>r`*Z(j7Zhe?xCtxnbq6w>*9f?DAmy=M{R!Btm(@%V|LO}9Pl;g+66_p3toCM z5IIsVoRG0sGz>K7%uEugYt1tha+CPdQBqCewOj2xWb--`K+ZcX=b(k|^rd|4N$#2bJe zTx66M#c>f_1!zT5!KpR2Q5AtuVoC7bLs_!)R@+0MA7{W5b%9gMk|WVlF!QUu7U({! zY8%7HZ9d>GwO%$C6`n7La^UYW z7YYj*r(+G3&;F+59q7}L$ttlL1?YlF;z+CkYI#p-WlXqQlId_LH%8^b`&1!jN8D-) zMjPAXl+gtGT>Rmvy4mptl zK7axp4!TiN7Ff15ZV8a!@|C2)<(CO1eWg^cg2=9x&9k|w{^th#hxYXk>Bp33PwrVY ziX9Sd!wTD-QQ((pl{FBc@9q2h``ItaF_G#`uF|rW9s2}9yobutCQ;2*Jo)*=$cE8a zQ1hgY2D^m|h4ML&2>*%#Pkv71X`0ld8fDK~a^>;V80Bzb8BBOV21rZ8iyMgN@t@`TtwA||%JUzXa3Z(!b?g_DP%#9dI%{D)AeTC*F805Jz| zWgSX_IrAchrXQ*T&8fzjwAfuvN|*9dua`h*x;H}pQ~cn5BKchT-eZ^Y6pj#^!KJxZ z7_0InlrD9b^hZ@y>?rSRX5zSZF7bXT>RJ5n8BO}J61wJR^IbLUxu_CDpgXpa$dhWk zm(2U$+f0O-X2##gu$+X2t!!fSyF#(>QfMjIh0uH%crFFoFSb4~vTjNbh2xho+iBf&AYW%WU~W|_m&(M&dJH#WMXVn z3#ys}j=-B0xAhn?%$<$QR2mj%Nks!ReXXtC8J#tJWa$VQBQtI$BQGz$9_kBIR<(}| zi9sZwR$~9nFxKCDBS3p@J(}~KDn2pOGsCk zvdRNHwT=^k5Z)l@+E&dxM<&Im@ft50P(I2y6KRjH*l`YWENm%%^Zb7o^S}Kh1{*|LkzWVEtD&!v&YR~_}{!;hgp4Xu#d~FebG&D z=o1ou@V2KE`r@r-Kvy?uwWb{#za-q;Gqyfe{zLVZb!D{ojgp%<$Zd#PcwOI`>hZl- zdBkN=7HYA@J#MHSub?0Zt;ed_=88!T7s@EA0A+&D}{e8mnuF~alqM36>-VLS0;pm7~`~3}awzr+xu#7XC7R_m#nlj0es}t||l7tfHRH|p8s~IIeYW@OGXdX)&6nVmiZIow4ZWQgeelA9{ zOO~%DW37iBfEtbris_-5N-a~f(nH_%a7zkA!O>>W@bFL>%s@soRzAX(zV_(;JDy#K z`EF9E(I@@|@yzaf5ISg15I~wg35Xqo?mzh<$zG+akvX4q1RK;vp?^1c)RgB?X!tgX^yrk=(WYN!CcJ$3& zx}j^|xtYnnXE1r_hRi9BxJFuvCNa}Y49(n=)}xQ`si`agmoDD77eT5KoEHXPf!5OX z&8PnH6)`NMh(>zy?R+soJ}NVD@Cs8NdrsSS#bz+@f|QX2^mH9l25VSQ1*(cbFRH*l zTHTvT`l{>PDv%7f5;I>G~rtIq{pVXb`G<7JMlLv#Cp-R^ZQq5xgu<@MC%0)PT%3hSx;_mSKeD)p#IQ) zuVdNN310an2>IcmD>!hGV)`cF<%W2wiXyUG7+#;Zt!Qv}~pE=TjOt?%&SK^)yHVs{anTeiFVZ%}FlC zZZO@yGL(cRHH|gTG_c)H?muU;_6%-GO1X^JYoN`DF;_4+#nGx%e(-nva-;J~#DKco z_d$7`_3$&6U*Pg=OFYfgM7#Wh%wrse0NV*KESYh z9uhtFdZXb*21;E*UVw6e^u{kD||KO7>~BJoF=hObr~%vKN9{?j!9Os}ROwAx zyFKVS`VPWLA_okVD=LxWaim8BS_NS7B35`g4G86_VNZ~PQlvMxijwweg=X#Jx%#_n z@t^w-yAoTj?9JZZT^MjFawYCOS5%jS^-hDj*{3qeFhMDp)(+S_7wTS_&ZySoJz zPPr;gmfu{DZ4o(e02q!o7twPw$i3uM3J1Y-eINYro!|80Ff5TM17RSJ>qtD?HEYd& zh?>Xoa+hq8Yt+Ur(mEAyqF7+koqBI6w8w~0Yus~H#B}p0yF`!>04V9GIuZ;}ZlFlCUz&*(M9AJ?z7)r!swC^Ux zU(kbUVxeocW7m7wvS>0~E+`Hi03(QL7SesDqhR4EVs|XGk*O-rI??{SPAg7H-W`dXFRgSWhK5^GRko}xSeExT4aB>OH)6xq*M#1i|Yli81m zK1m?W$`1i(fuU7NbSd(u&+QFlL_f=f1%s7sml6sq6QtHWk*vLvV61Oc0`$tg%q*6! zj!I?(+1kraJ{vkxvx3qgtQdgic?e*hx1OqHu`$Ilkv|nf7w1yrjPoYTik21+OAd<# ziV@~d)2IF@GxNg)Ozo<` zVbuW&^Qft=5If~kunXbnUL}W(6?4*toE+=12nGH%5$Yl)fPsm0dXm+LQ2Ne2X3=}7 zW~7Ne>Vu-u`~ff~Fv0PXF#=tR5-5$OpvfvoyLalg$f~1|8)q-7B?M(;nl?%@3zsia z!K>yJH)~rq?RTr({4HM=RkiZ1!_e1hryK@(okjG*k~@shHrbVrgPnKqG@P-JuSxOnaIVNgm>l(Qntlygc;@w? zT~tT$Tac`@fm+ugs`F><9DpZn4OQiy;#_KNr94Se&TKEdq;zlwC*92>*_WF|CA_48OjdhUtXJ(RGxy1SzSgl1r9qI5IFx~-$`ORqOp--Sp@agbXiyop z_yMctO1wXIVDB_v}z000%pViT|a`4bT^9yr4AkNphB)WwOWL$ zhAd1pp)Wq10!yug?7My;dBk7>DgZ&rElf__2vjpzl6N963l*g0WDa|d`gjzCSoZ%K zkB&P}w+ndC%4G)LnH>AF1R+9;^XoC5bm9)I5}^Pd#L$1#h<@)0nkvZ(fp^QRcc0Mo zZOsQCA{kH^Za^qF!2(2^4U9zKtvzJwi+?Z3M6}vM<@SmyJPYX?)>2nPxT>~>;3X17 z77t1>$H|jCTy>M}MMNXKj*sSKz~3Bzx2;0i;^%X^ZhszqSHi=*8RFxV?4_yDAO;Ya z0LZDS)o*AziRRo56y+2^?x{4dMS9|?Je^Chp^LG8k;vwz6y^e_FXyJ;r@+Wt9&6%! z8jk*gl(LbZxAmyHAWI=%26>|eRBm(Y`U#?r==e5F3fM~HMg@w|QyYmz_G0ICf3s^v%Aln| zQCBmejfIy^YFRgY14Ow_9@V&)sQ6@~j5dROb3(9vSSp5hkVAaC zj==zI|4!FiU}g+qqH6kPVp#yq-{Y`P`@PoL5(kv>>rea-}cqxM?KSMhI5`!g<7ON#Qsp*&ZF?^>qpGRH z_@kOMQVfI>^P6XmcW#=SOD&^$!gQ|;l-b%Ph1WB}G^*zRbV9&_wWeO{9;RIupnyAPngXACvh zXCm4n;Y%NK5j2=n#!F){*>sI(qQ}1`lSB1Dk$hq{Pip<1Y)>gI!$ppnpj_sLd%_+orX;(Gq-q3~{>3S$CnWqwat)4wZ3dg-=ph9E6_b3d~`$ zVixa2MW!<@^pI=DL)KzogFv*#6+%Fjo`xMW#F4TfpfFiZ8m$&{?LG<}`93LE`h(`| z0_t1PNb6~&bIhM~QB^QwgaGq{N-ocqsUwp#MXGCwp=U$8Df7Qnsb3GmClQT}`LXJ< zma~LbOsKtyQiSDWP8-necenwJkt_yKtiRUT&X2@k3V8CQk*93v4EO9=>aaK6=yfU2 zcsVuAOd!H>oH*ZSvG7do>}si!6-ebOImy*cNC^huxi*RNkK|_Km)Ktil->!o78uJiygx;6=yo`sIZy_gDNK2>>C z_o^tuP!a=zG)JW;iOe3e6398Qv<60UM5bBdV(kRc9ivft2jNg#{_m8?3GQ|w9&+5_ zQgd|hSZ*d+^%t7Jb9uT{HGysCN(gwI`ozIoyOYhD5Gy}x9IS@oEG`@*1pUKG%k%{^ zVicynGNIYh?dQT1!pxyf-123N;zVc*7=Tu=q$od3jV(Q~G~3O4h9uUGg}>Ml{>yBp zk#z!>E8Ikei5R(YPO&Wl^J>-|Raw6=fsB#0!%dX)6tIX(!fA%KJsM#861)bNfU>gDS1hzM`c9fmQL{b_c>pQiA?K*tL(3x{3 z!#eIoT>dpgeUxPVc(>l(L5wbTXjK>$dPzC1NHP4^tfNwqQ6CM=cgSyjCGtX+xONEs z5K=G5%58|S+Y6V+!VJo?Vrx8B zSxb`~MRFX2FLN(RoZne5=`JxC<_(qF0kOYF%)fUZ>%`-%y|`X-2T=m!xMlst$m2$& z^Jks-mpI~HWTC)7ae~SV@D5tcrZCB<$sojKQk%ydY$-QJCAAMvB3-&v^l${NY6|7N6-;$k&eIF$B(abtH+Pz25a(qqhi`CQn7hW{#+&sbs z(LZ@oP@n_ojE^C1xi)Gx&nBq&pDJM@`9|Xg*)3vaUq*ajK-52TMjxY`#00bHYQXo5 z2}P8wAUW8$d#&IZ^;{{>$Znlyet?_6cKY6eN(5NssJwHvLYEc^-7n^F=h%#wMU{U)W!G5# zmf;CR#sfNK;QS&?{U}g%bp!IbKlJwE*5E6c69&tqe&eJtBgjLPJZQnF&q$rklF^cn z;Kp!&U)qCV7>F6_1=tLZujxD8X}yGabi$BhT;DMsg`=cPFL)_!S{^snJF>{xa>s=; zCY0KN%&tZrUhb);O-jkc8rt=17&?%{?cKR-4L4Lwqxy8Cr^8B)i-{UknjMNG(Ywi~ zvmy<4yFiiDd+j>M2;0v=L0}@}(L)*LRzhF?;)u*kY%sL)dqDNI?aiR6))yp=N9una zrJUp*G3SUN260I3nqi6LI-10R8OaJ{Wc<@%20;}? zFuZJjk-m04($Q`lG#-eBxD+tS{-MYg0?y1DN*R^v4&B}?HRCLN1F1hV+ zjidRB+Av09rsijG+Lt=9^5Lnjq2?mfh;0$(B$AwE{^4GJ%&AJHG=>vwDb;EmSu7L- zfB$sB0e7flvAuU!O90la;Om_YFnk#&R$c=bgssNGj10nl)R7)Ya&co}qas%NT1?_y z^mPf_;!3W@L<=(iI((%&XcnfJl z-xv;z%9Y=)nQO4=Uic4R5)I~EMT%?L@MN68!1B3p?65jS-8T&p;YVk}Qt=T$b~0*l z3j!UGKGvDvCZS;Gh{OW473h_S;jqd4KGnGmc|emn`LdSm{ zK;8F3^W{@2Y}tM^LGW4tFB!q-v%Ne?M3~b~tCOQhD>>rG4;)8n9FkLe{b*&BndW%h zmc<*R6&9L;!d*ye`A0BlZua z>2!y&JxZf_4Pr|p;oxc$!nUC3e6y#CO!nild1~H~DGY3~VOgwFX|jZ9r)^~nxfYzc zv@{1rU8AItkt|1JDn4PC1e1{=HlClqHublgl@R;~;f*`D7#-U*t$}gUp;lMTU(^Bp zjXw$d>slx&lAxGo0Ye)w>G7RF{DMm>E_fNFT4|+J$V>)g}L_Nh}m^n<1PjHCAN)BS5%cYdYeVWFhoCd@eUK1px z3qIdG(3#c~3{TNN2=!bEPsaywmy*^@}PC*E;KPlBHJ>L%{EG#14s29?H1Inz+{ zm{R>X8?3wg{mcVZnYovm>llR;QIVBW^1La$u7CV#s&ICK%~8PM>?!@W13eMNOUshY z^AQ(AR8(M;UQkVmg(+56c7bT15jw$_kXp({Mk7s3YF{i8fz4RXY?3LfL~afCg*Vl- zf{^Dj@>T7VPHB+J%yFv;q!Sz{iHCE?M>xtVHd;`mWWgqc$)~4k0~KR$if2LH<^iX5 z<1u+BB!uDKqgA8L)jE5}Xy?z+6C4YpGByL%HLRZlSz(W?d>UsHqx2&K*E(R?gMkjKoGz(ccuL7$#0I*O7P?X@i<}lI_zx+y{MppSfti;Eq$~z0ESlsP zv{ItHDLpe6T;nf|MJ}li?L+0UP&5c#)xh9BKq&n#FO$66n|O(K)aYy$RsyCeGK{9C z5XV`tZMdw9yx4%2!XDcf5U6Qe%V}8~Xj;oIJUHkk27t%!*psR13`MwS&e=4Dtr#GJ!|CLSbjg+>+IQ1P-9K#GIJ zrJbFcz}<3j<8F0{J=bRj^)f}OwPxnfV9G|^tl-29x10d!)Jo_lWnvbk6zBkJKz&Dj zO>M1xZEa0QZB1^eeQjk;Zf)(rL$*vGS4R>goB={S;+D^V9o~DqmMo%_f}TWV@HDlS zvxx(L+e$?#>^X$-=sxUeT#dVm)bgu-+>a=bQ0ac(JqHwQAkyLuMNYC=KN(prUH11~ z=00f%oH`!+rM;SRB)%IF?6gT#w?IIOFQ8dIP2BUMgv^Dk=I! zXv-Y$lgg#D$Z4wTPRp$h;$e;1sN2(Io>VT=N>10> zer7qkB&e8>(QY8z3&U$9cI9;D0Nj~6*pv+N>J4wcqRx;5X|0%X0I%S44GjdWvL>#a zi#bdtA`he;c3LlXI7q}n@NY z$gY^i5WRkYjryqvf}*CaqcblkrxP9MucITRLpf;_dQJ!+h*v95)G)jT0v}k>39!KO zBkYAAIvS{w+#azN_?5OmE*-!|jHRge5+hN&gKoeTWIdV@IF$O6_ zOZgY%tFb;R@aq9ARndonkE9~~vJH%kW0Ga#k=upNnKQt3<3w`drCd1756D)}=6c0h zeWMdqP4g$xyGYIz-sfa!mXU*s8%xBcQcnC#@Hu!00JL8u3q*;LT6XUbz+d{O0W7<| zncB6{aF%0$?s?FIZlOB=9rt(Ubs?sm4_HEl_kAhiG)m{EVfDo*KJNbpfhP?$FPdny zaYr@ONJDbzz)>O?t;8PO!j=Uv|CP=FKqw(U&ys;%TmA=cDsUPgG+{?V-rszFW;H8g z!kgs3aWqC}E^4Kz?-vAf4m$xg5Pw|KD$-W;24}yJk91idSB(H1h?IX@?imia9>*O* z?sGg;05<4bzp52R(t=ACj{*_^060|bZlsqCbzNUQI93(!NdbV2E4|sH=9h>_OR$`= zcn|cwhV~D>Y;T!yF~Bj?UJthO2ki>N>`H20W6lrep=3I>sZn3Cz@c#PK6iw2)^dHg z6#v&-4)_c&bbi!q^c7pI37a1r2E>^i&}u0I$I(Xr55P5n5*l-)Y8z6r>=iiwUnLVb z2+p|Hut+WL?|Q#OjQ_r}>iwqgswoK>U^4nu^1tQuHJwePb~*mW zA|b~FQAq(D)ZzBXR$rXo^%btd|E+~T!oFioIsj$G*Y`V@T@d&UVx^4Ps;2!d@Jh_> zH?&zC0w9cW_;4VMCF4AgU8VQYz;8~eBW%g+T}hWrWWWx0z^bT5AIzhr0sKjqhfuRP z_yxSE&-Jvl0K@m^zut4e2l4H>w-~+)TCoLYFLB=fqrs#ebZZ}b*f>-ef3JrD2;CoO zzD!y3{ck18w7K-)C)q{#6xQ`ihM|`{QT5XQ&GZ)m5h&;Ye^gJJL*n&%=D*{+2C>=6 z^EYvhhAj)&i_bw?rT4p*&{Axob{&uMc!{C}fpx(^wjLs5_T{^F;Fa-~jC2UDCm03k z-87V&2f66+AbZpzp-B%^6#NieJY4V95Tps*&sRUxf#SP=BMy6ZVAw~&i0PLX%nLEbq zCF8gG)+IR)o!2C9@>zv1&S%)a?hy)s;VEF!nLWOdd(kAj?eUq(=W-6uu&$`aC6!*B z>ziFdt6+m`ntd(U#p~D9Awf4)E)Dw4y>{1L`-ye;Zz*BnS-UI1*}tg~NfySwxO&FL zuX(~L2!3hN`~`E8vJv5dNas#@+3TmCeMN>Z1J+7-Ltah%X2oQd-u*M4`;3cx`CCc9 zk**yS;~(s;wdNAI|`ja+5-B8@^%qmP4Y+GFTuAk}B>-)T8W-h4^rt?7E(h8`_6| z?rwPsp|i@4`G0;k82MN#BsEm0#W~OX6SUBF3~L_Y4|=5A-2>}pAWu0pr+F==cghz|(^OLZgm z$|kbww!iD#H3$i}1zr$5SWQjI{}R>C_eL&jG_4L#Gx^cPE=B#{?BOL&wAS1+7GHCj z*!-uZ=|@8p;ZpZ9M`_2mBk#OaPB#6;TdFS(pZrmcKmQa{yWitre>;)>%4fJ=IIv?T zNyo`ixW0OZ)3(fQ`xKf9k&9nK*;@-#8;pXD{-ta86f*bmDvpL&g7%LnSLM zC}Bz^pb-BT{|za{-D?mt#JejQIJK9?f|9`g6ZXCIpbc~p#h+B`>?eu@VjR8f4| zHR#5q9f|O%_W_G~J^x7#gL)r`1E@}Zu->vyh6>tfnT*7x#})xyAN{wdW6VSJ zd)D$CUr`#*?~8VrWrhT{Xv0x$>@WRYa8KdPNgm_(U$tvs}X9O@65>@ixPo}vF+ zPtOm|vN`>CoD0DZ&)Z(Qm~!w-~6iZ`E)+ng;7f1L%T16`5f#Nr2m^05%m;(3wPuzw|u*SrRBSpUG`@ew;NNP z{9O_M3gG>8&~Fk)Iql#Vk@)N`IGd}zC3Y%;IZ#q zxlc;rzWBj}noF0*+hh-$Akx;Q)mVEFvGFRUUP?V%28@9Huw4D4cKi9{fNpM#>OS!9 z>*CZ|0m}!DEHRI4MahEy6vIWIy~puXNDNVXT+RK-%HL1FO`lXW4EkNyKX@so*l-9l zoE*>kjs(n*Y?G&;!7m$WdPaFUC%&=(Vdf8xQ&VPU-HUzh3lFI(>jcvUwa z2aj*C3oEmFCL$pu&Ps6lY1`rdc49;8xct2R^#1T;g?sIag!C@OE8{}GLo@9`-yNl0 zOi--zW$m^ksuBq|f&o2NwJAJ@Ex7-@}=^Ni2-?bpWLFAO-2xb%D`A!2PQk9pM8X!+pZ zBK-GuKP!4=0%q7g>Q_Hi^bB{fJ{TAZkpuvN$m@vsMPgvYU$o&7z?Qz(`C0Sifc=lj zZ%^h(hu_n$oPS*RBtJa%A%`rtCjH(Wk;4}K$M^6@h>X4tPalv%D7;UT`t%UhV7gf1W!Un~Xp3;6Xv7 z3a2OuGMCY|)x|#-u3Cv5+vu20M7#6pX?jS?3gHX8R7ZUK>GH$vb}kvu{~;pKYrZm( z71!MsC9xu+?d|^rXOjG)qg!NXYj^cgE$zOeHgd^mYbEzEO@;?-{))_@0v zPo6~|zCF7-RR68-`EVhJ|LEp*P`k25l1UIf671s(2*hpO;VOQTCa+IczjfTGdH223 zGL^UI^j9v${hh1rcXYF#8eOYDB0`RA_ggEi3A9D-PmK;Mj%d8$!3G#+Tgd;Ucv8%G z&tW)qpyKJJ(+l3~J6_&GwfKV_*h!80TbdxAxN!ngG2yF2HLpJnAvRM-C0q+UJkx*4 z9O7W$J?W1M-jni-18$a?nL~FyRX;n)lam4y6xV+i`nO%^N%;Ns_X_^lvz4?_6kPZ9 z&*asDVdrVh-}9$J(nr&HRA}a4xCDd+7gbLfYT8W7FSPK}(L5RXD7k)SZMGzC*3mX? zPpgLJ`{#W7qa-ThOB09d3z9sDg`G`k@4K5;YqK|lqAL+=U~sON8}T$4Te~U6j{0@Q zs}N5pBa7pbI&Xo&{hOlOsTv=)>eC3qj+VNnt@j~U;FUc5oslUPcO{Egs;sP_>K#!9o(E!WSJ;x-E=;?=m-tlmAbJNN}!FC_w#CZHqP#Zr7R zGJY{9^k&1d!EE8C$I*OhDEgz=yUS!k3&{HSJQwGyNfI>ziR+hlfAein86v|2p47^d z|0gXej*y6fuB)$I&ZEhTj{}T7Bd9}GAasNv-LH!sH zLX-ZwZhrggCc5lH1OV8y+{q4k{yKr?kB(6VzjHhc`FW10nH9T!#q;iI@Q3s1Q-b?G zFR1619u7NY7N|$hoX6-F~#w* zM#e)y{&V_Sy4Bt({ce(M`2)@W_W2Ank2tiCC-e8U0vMbpIx6aW?$_m$7<*jx-j0~l zEZ&2(_QN8!e@5avP94CICL|%1PICMhc3*TViQk)#r7d(Ct5Tv~H#O8Zp_)nWyd3mu zJ}uVA5$A7_v*=CH)XuFmDx;Og6m zba~6aH}@& zge?2>`FLy4kTZ?X-7)U|vr{B>;oSPRh990*RdKX>xJfB}|Muhm(e%|}QGU;SBe1~%k29$9G>c+BGx)4;BrkxlYtDhKVpNnd?+_%6UeC?;l>3%^NF710%52;; zoo%ZcFe~Yd{Q4V2}s&U9(fZg@YyZenXc-n%?Q? zdiML;?5g-p^v!Bo)D2|P&abwZ-QKiqD8W`_kBF@G#v;4tr`&@zreO)C`N?=jS`G~( zsJ-cmH*B@l+8f>{91g2p*2WjPbWZzar(G^ZOwOj#J5t+d1l;KZ1q()>2uRKAeFIl) z5?c_)7?YJFZMXO9U&V*;5QccoEfJP4bEKMX$UzcCzx%TQmiVs<@7Q^DSKCkj)XON-lYLlS?{U{PjOh4F0&euWa@)19B`^#`7~8h zr&UR{E+)iV#O*hs0-DA4FTV8GAkXG+{GoC9$zQ3pr>|%_J+Id~H0zzSxce{$*S&b$IV`>z&0$suBz$V_XLmtNwr0u!KcCzWB|e6?!t ziFO@Nc`PGY%Fd@}HAqZa_aH>^9yOVs6bmSDy^V?qIW6rV%Q7E}pZ{S)O%-fN+tq&4 zf8DcAsHv_{^F34BJD*a~Fboaf6&Jqrm)2lc3BQw22;~f`0A8xu;z=tvPM}lsDF&B5 z3P8v1G;lQQXKbNv)Y<&*s-L%bGG?&x#a(iDOv|WP6HB-RD86>CgL>{W|3BA!xt0Vg zmVU9AHl1jvtAD?Rydp#Y6N$638jAZzLXw1+dH!N&9}ORJx=jPkhRT!Fkpj~Ykk$@H_>!Au8QXRCvroZe*8zqY4IZ$uJb@2VyxCLRLCM&{|7y--ek?uGBiH5tU2qEFN#7*Z zsvkCW($Y3iV{d*=@LvM-kmx=$F#7pVxls_19L39pgGvLwn@G)!W}T=;+TcO2CI=Fp zss6(mP|{%QDr>#;fF3(ZjYGG?&DqNZJ_i3}eh(b_-7ji8Gm4nP#V^GXcn9ojuze>) zxbg^0?T|6*WF`d*zOCr`OAXc=|I(Ce=BkgIg>0&hm}J?ZiF+yvGHyre*?Lq@Z1ATU zv~Eswteo$goCw1TUw;+dJ%yd`zmE1`tMuAjSskgP{`L6AE+ukwk_9;QM|hq+@eV`F z_C0IvV%=U!GNIH#J#U@E{?S}rn4#EU^ummRLhxNIyZ&8l%aL<^ z$cyW*?=K?lQ~ej5QB@}~y=+TzB&fq{=*V$_kkVDm2C@ZwkSBm*@hG1B;B0i7vSCQs zEA>qHfwGEwomID}DfT08xws>xO`XdJNfMg{v<6d1*D+fXmTGB+Y2W zro-x&>w&W(wE^SEh=vvw2D%}YVniLV!Y=KCV({uzCF^780Q;e$<}810c7i zIAo4w6dYH*sabAuIp;v%$5d6wx;SM!Hwyfn#^`!DTtHV-x#~1)V&f5dKSdwiZr_q^ z%rw6nUCmNqczZ*WB_h--4%g)i(WdeKYF|hfM!&%gMS0(6MeEQvBiNoE1q;0)_+4|1h z&oiLHHhN`hbyC=+UD1jj8HujvcZh!kiSyjpW9h&{){7Bj+zK8>@oIA=0AVl7g%ubs{tj^5Pz^}pnLHNQhLXxd$g%R!v8N@aTKRx(= zKOc4Ap2}w1D+kc$Le99tn848vramy7sRlyi%GxQ?;WwNNeVG}XIoo-5Ig!c8XL<{n zvnT7Q8@9=OWV=VdzS^N~YPClSoQh25Yf9oB&7Jc}&C1iRU_p`BkoMxwIx9gjkAM|MFo2r_dKnFFjS$g0cU2oR zh@p({^&h$1yWG{wzJf%dUpFnwPKGm6bxS^*JZj~JO{AD79onHo3V_}vJpRLim#51h zpx<;a(v=E#$B#i%&edZeI!t!ALg`0UrS7EjNv`W-EXy~JhDptt?<>GputaV0J;}O? z!MsZ*;J4=(`L)D{VJ`?TZaPw_xKRO3%SOnR1hfWqJXb}+1e3xZ>vX^QZK&O5e3Z+p z5AK3(lJSIT?YxAxV?JE4k8e$DI&zymo&DZCkwfK9K$Du!H*MF!BJ*jT+jpY$mIC+(;x;gvY2X-LcA0|4M-5)eFC;e)Uv#}`?pR4 zKK=6v5|=jG0=zgUzI<~_MI6bEy4q`sVToUpU!?E$<}y@$=?8fJPfjlNQ?tm_v7T^Q zsxMzPruk80JT&Dx!$187%F#p09}U4)WAh);iK+JnxB9lt?0l~4Ei$3cTs{H>EY`_p z4r?_o;_7Ba>+TSJvwEDUZ@Knd_tf(g#qm#Tkj>`0t6|V`g)yLmKhj)rj{>lIFy^7Q zb@d|m6DFXr|Le|%I5UKtN21$UStQkYbNuf%=G${|tHeLa&b=6^2wBNo(nwut8q`5Ajbz76``aF+%Rf!OZ9)VISE zg$)xiYR^-`wg}X-){KU8^^#}1w zspFxS?0+PhEG-l*^U`S?pvb@~55ajSSjA&HbkE`QR|ajOwqM@9$g>E@+h-QH@P7kH zlNRb9+q

$5Q;(CjNJkp&H`nODIW*tye4{e)Jg&Hz+vp)IbP;GH=bOqjcA}tMm_K zS0v(Z&#}7<|1|o}1BpJ@ujzxXCWRe>gYP8pu(RuD z*i5BW?*2a==fLO300q^a*{r9leWu7o?^D*-a^LjHVAbUFhVrH#MX^zoO|9 zAW-F1Am}0?&;szle7%5q}-zL(X$R`S3T0wUo_Nr{Q=J7=o`<)(zD?sk< zJi{SwQq_s@?px_A3V;7Y`O&Fi#lxNzHsM#vpMX@qQwLWPo<&giZiyZx(x07*dqqM= zqwMbPEj4(EoS+;zwy;4jA^T0OQ$rwx_+9^g5!9D7NhI!k=e9=`j5&%yWRQqW@GD-+ zAS*KPRJRe4+(EOrWxyH00wsU7e6xr>!Jn%>Sf=9(RBn>xU%HU0{0b!7JrLKnF4!1(* zM9r!l@)&@CZcTY|HShmY@4fw>AapyM3TCt!P*ZQ5;#7Cg*`bK|_QpUohl(3gI+~^$ zbp^@YqiCW2o4eY0CEy`$lqdaM5oLau4J7UHAKJG?#OrJMnRQw9u4Ce=|ZXD zBXLG&YxaD}Cx=SnS+-q@w5a5qKe6aT-)44ooMvo4lAk_Aw46`Ag&AQf`vrXVb9>W; z#>VsWYMK`76{q?;`J@KDgVqPz%nDuIR(yB#0`-XevPsIu`0?2+m4nD_>+O1D9JXGj=<PjV^zSFT z0}iXTon~|Q7vYel+w-1z;4344?>!S;4?*2IGHapyugZIG0SfciulsQ?I(D-|D?_iF zt0=k0hS4FXo_8eR%aipW?h`_LpZ#Bixu^{tO^X(h_b*$JH!z5OZkb9RDS4}9!`_Gg zsDb3kIhPk%e#~+`j_@+{uR&gJ-PqBL!=Q8fU zXA(bNlE_cL2MgATimYFCmV0j2bjsRVYBnrW z5x}6GA?j4;@gGM!RmfG-ez!I~JMtWFB{~`;#L3{QPG;q)sQG!19#Dlu5Kg~34^!%V zgLC6R0Lra;ih=*R#Ga1meLDJX3~vJb-kOlD_G}$CBfg^4*I#Qmn+`Z z)%eT*NKc8DyBbCn6W+3-mcy@nb~6FyT^a4c#lO4tnCcmv{Q40{6xc3JpgP~YaWhMW zzMIbMT*K!3;O59&-{U!j-f(JlDV&>OC?42b`FIOYk+6?E6&^V@S>) z@kk=hWEM;5Wdl|lUrncSSR9sbSvl@)jF49P0#A%qe*0sTq3H7@oH+SC1zunny*T&| z!J$ho<(3}+1IP@DTtpY2K+@bu{8pE`9Q^JUhxOIZbipB2GpqKO6JG(?A+(9)u>jN( zU&%+qzGIhpy23U?A5Va#W+!loNx-I`cizn)bC5~nc7;h`n?fW_crgdQ&3SKWDXcTs z6u<-dm;CgS5FdjU<0TdMz#-9Fw(l0(R$O<&PB5TD2%kE6jyZQX{3rM$8YDo82H@C( zhrferF8&JGmSR^`Nx+CP_E4YI!?|V5G5hnfe5_EgQqNYv3Q*Axp*7?*;~)_Dg>-ZD zl<3(%ls6lDwbwK`*TH~V)=VmMtpkuxK*=qqbAE0iwMj>;?%LzCA>kz<1~3kQFCpI& zYp7D~Bwwv`sd?Cf9+?#p;&#D(N|l#uAPU=`!|Oaewu=p-S2zv}iS-gOwCixy)piSi zYtHQf|1Wq#UtHkH9Y)0Od_bT>zR3rpd7UD`FGr9zp#;IU!n z=Cq-K6QTI{ho|ayUGF%MroW$?YQBt=*#N2|lu^HM2&ww3FA*B9$f}cTvf9{<+CTmL zC)}eC1;W;0WE8jH95`poqBIO<>1cRy^?dGVN!!JxYtZgz`@PA3ybeiC^H4##$xq5&F9r$lDKCBUxQasVB2^EtKG$yl4R^J{lzg^HO>WnzP$shAJ6a}nyHsKQekb2B!=42 zVU=vJky{a&IUV*j$vN*k)@UaD&rV5dAk^#J9XtwvGsL+1Et9{J-;o9M-sG<7A3P<;b2@-6%tPl5y>6U+|lvAY&T{*-rq^uSyu)gTo-*qKp zMnBlzsfU4Vm%&)sS|jE<5s0WJNPdD`8O7V+K1&1iSl?kVZhdzvuUFCfCj$q^Hlj;~ zFQSxfH}MnfW#BfX^T(}*yhrU^RG$HXdsuzYQnu82V<*BOEpVNruOMagpFfT-p~x+2OeS^sd@R) zj$LnO%c9BP)!}u;J+_Cr2fDrPpK1r+?24_$P_ogr9_eD; zwesRd`i`+3mH6jZR#?mC$3_>6s-HXR z)^|5=sYw6q>@(vku2kwQ{N-L=28x{VPxyM76>aE8Ba97rL#w#Fv}1m^4pTS<{9ZV9 z7kq*|k{v8;pXE`oq+qJ6JPp}+wz}eaIDZVjT3<5$nza_$lB|w&pRmk2`eqUs$Op9b zPaV5Bn`bAJzEyoG`#03&`dhdc5WQg&S#u}0m=Szu6VcaFWw7RCrp2E|e|Bv9<}BYC zu=`pHH5h^Xa*BG#W~qzJuM}dfjL)nj_)bE`U-P}h3my1R<4xD)_@K5x!msC#EpH_E zK!Ct$M6LNMD4G1`4a{zl)^SZPfs@Baj&2@$uW(tLU#Z*p?8{EOJ^NVp*Z2FxlUMv= zB5e%3XUEg03TOTB@|14R-R>=^K70vci$m@XSkGW;u{FbtP13Nwo(swU0(!cioN`Q4 z?Bwh-ZJ{)JTeM4k!!P@4_OV(|&KTp&gJsMVVXra{FP~Tt3gvZ*5^((dBC6=muP=lb z{nsDgL>E>J1J3J4cHsXQaNe(yAtvMsUyD>(%C1`|JP;!(B8e4OC8AOI_q}a|v7KIL zcn7oUX`wj{`gMQMuKqxm3EiP2Uf-9pF6$3-ny%2)9fWU?q*VD|H!Ibrvs-w@%l&OM zUD+S_T>d6d?PP=9!!!fa52%Wb7<;e=7bm8gFIcHnrmtEQJt(xw{groG43KovMT5?l z@;CD!29`&kRH!sg29iGNtN(S3)sNq*>{9rZYMA_3^C#rtww3j|UPmT7%#12HqYY2m z(L=-n#+#Ke;>043?yHNN@tluXe)d|6kL~9c-?D7gtT~h}#6OhvfX6 zJ}tB^XP=8{=-cQUXkVVRQA>>>8fhY4wJS69Oa{ewp%!}*gHAh0;kWNENiYBs$(%DN z?RupzqWml46C#(84rW_!R2duPJZ!jTh0>tu#$`ll! z5#F@nrccK|FxtE9xf8b~4{au03Xhl;n^S-i5HUcRJJRK_NtRgzYAKI4xf(P7&o#2T zl4ZudrgyF7tWOl3)`x$9Mjgvt2t7kQWW2(?{OpH(fE6tUj7K5KJyXtbUCVaKy8Ty* zAAH|Wp0O1EW_bam*Jb?H!f&WE`ou@#7haJWF zGtE@k%#m6rSA`#c@+sa_cz`aA#v;yT!*XCZQ1ESk6!^a7*yF(x=j<2BF|+Vj5d8SU zBXU_jilAVJ-`k4+g91WcG+b|X?^!Fa-SpW#Lc~vAt=88k(eZTth&W#E8woO3d%U}d z_^+`22=j@YqHC<6hJlGcvTtTz4^mz6#L_=6^7S51&u8cM{C4U}40&L_uzX%YjT<|LL)oA;j|Fk)y7iGZ7}g(PpCNIj@}wXk-g_8w&<{+X&3)9m-GTUVs*qbDmTMQ4X%2)l(>wY7BaK&4wBxFjS64DxFyZ_6axE+{Q z`lvx0-9MfwEUEg<1*v>zwo#OG*JU~m2Mpn+c8&s{)~@`S@W zCl4Hn6A$>{yZk*=TAxC*HIkKV=h1rtnhyE-e@Q)FHHB802UDqB)$h4zydm$#mArao z)uUczc~S3kK7RHfSRWwRJ;8adz71)O_ylcNxR$QqcgJlN%HNDG@>F@?j)8dU-iS)&@lk_;)DwZ{f@ZzU!kVPLcx@8#PJW`sKZ zms5qk#&GqgY@B!|<|{iPzdE}fS*xF9@JwH!J!Fx>zMDL2u|j(8v88|6i?@6HK!&Bw zT%I4sAeCgF$Iivy zZ~ZGqJyXzs(VFpn#Bv?G?eZ_|=)lyBVy;*7j@E9xR3DoH$_RgA8Ij+7$eB9^@!igA zDAHwB*&8(Jww!?D@tr-~QRI?~^?hdHA18O+n}qC6-3-JQ=EVMz%#ls6E}t9yNAJu# zwd*AI{p?B=33buZ>-5bkLgirl|M~6u6&JDDDO{B{qCs3JZL5sU+Q2j zPsiFs%Sr)3_&fi_{l-(z-J&oK+?}8!+0g4XuLx=ABaAd0rS}Sdv+I+FTg85{u}q_{ zU=?(VuRmh7Eju7j&{?&@C0rOs@Yt$O7r)s(1_#IQhaT6V#rzdG6AdZY6PQ$tYSd|O zTp#^sO*ju#i@UNM(;oAphaBCpi7*Yn(Xl0)M_uP=bg=o$vZj^3C}fnaxmKYpe(?k& zEXw3_dl8;o;!D#`!)OV_c846zlELK|blb4Cz)GR4Zgpe7;q$oZUkm5aSzhCyxccQF zu&332Yp^4=BS?Erpg)kt!BfP@`qTEKrx^IyCW#ECb3p`Zb*opJaw+K39g$$dQG8#E zpO%LxF(kD|XEJF;3fx(>sH)2qV>;*&MwGPNzx(7|k`IZdovWcOTj(uY;3YN2qP=aE zx{vf`34~i7l_SHv<>P!fPbFXHEV{lDq{rh}Z z)7|t*yjkwJb_C5ak#&w)oHRTQb?JpV5hMd&1AIJ^S(NYYBvNDy-&O>-9_-{+ci88T z=AF9gLA}<9rD+uj&s@0cTHdQJ^ZjY{D+OU-%wefp-*3lqZKR2-sSHU$5VIjZ0y|;B zuL*Fsi_0CgQ{3~czYD5>bIRliXC0hfU&*IErTfHo_uE5&!gaHx3m5_z7m#OrWM$Mk zqH#m=vgtt#R%J>K>qA(QnS!}{t@b%w!9!HJ%_2V&I-8IiG`w z16i_t2r^V_<9p|5t`Z^t{>WYEzq;eZ-;qyq672H4w5(k6UINc`NEC)}+iwSQDPHHA ztD#OrG9$Q~5-STd!kTNwk0`&pUrSpJ#>x|Fr01p;M#n`;y#MKFkTT07xWH>1PHz1v z#QidELsPu7{U&gyj1qL{p~Eh&)Jl7RH(E;^`l;vpJA}12?1=y&bNdACdkKSO=T*y; zB>Bc)V9i|Apejue2N>`+7KV{o^67LL={+1we?#!PWvXI`68oEN!5^id>Mu657+PDN zIi(j4%4Suu=l&xug#LF`xG(qBrWcSf8bv?XG6CmHx;Ka;$GIxz2n1{L_AT0fe~Myo zk#X@N!F3EI+So{^q=e;#_Pv21z?Q{o4tiF5zj7tV{ng&4adkfWK|%N3(+WJ9Y9F=2 z-iHHAJ;qx})fN1tn}L`v!}E7V&9`pvznsZoyAf-%(+^P*Q%-%$^}A9@5iAHfN-L|F zAp9+TL3D0OpA4@996y1JI{pRIxvHHnIwLBn{%(puTk85Ip0BdwfmxO z8NS=Wl)-2>Pwcr<O7eGiiUEAWu${MmldK!@rDrjQ9C*LYh&}wi z@@*A&T4~I{K$FQ3#S*B2@14L#$)gzLY+WFvLPWsfPGN|Q<^*9E>@Vx4%SHgoGW0f{ z{I^3IW8U@*ck!c6e&fvNIvL!j7?G9d97#;9=q#;FGPpkI>PEnlln98_*Nd?lnU6os zlFoj}#s4+=uehj8Q@FBg2aAl6F@te*|Aqfe8KI`n(VSS}0Q*w#K-oC_@IZ8XUp$og zLBx}tnY8vS>OE~JgA zV=X(zqG673#MyMlS;@G)up##x^IIf_Q9S-1!mAM)l3_v<$YHLp+wpLIp%XUv&Uq#_ zR}5U={Gh5p{?l@kz#n9}R*F57narw(I&6L^e?WwmH8`z)dC5{C@gCy_;JsY*_xADy z(T>5>mL@zyf9#r=M~^W2o<%TCX(W~iMv{zQEKljwUCuThoo&bb0M7+zt{#%-B0SYq z4mv!t=n#hzkGia_ZD!PApYOyDmsp`s#;>-Ah@_B* z=iuhzNp)q3IEGu!sn}2wv=b>0`!*O3SuYZY afWTn4IxG*G$CZt;b(KnU z+G^5AU|7WG7b_Tu!Cex-R|Os`CsvF>|8y|M#u2$`}bZw#*dOH?l?%CK%WjG_xFflCB3%fv8|3VrY z79zG^cokGBvlIE9|3UdH1<~V3WUPP(B?&BkRii!3wgu)#agF1DpLbV74BJ!k#G`|( zGrXJYEUIPkfc@}B0r5qwGbx~+Fk%-U(t)Q7F-_V9ske`hx(libGOy6c=qu+Tro<<= zfk&F?!*>)blN$=o z6Wv)#{Xv{XR)B#p|8Ck->aaE&+1&0|DhuHC9my(_wjcCs8r@p*M^irCx2S?=%k zHSk4{6mP@LwSI^5ASCe3lxg+jA{in)^F2Ln!OQfqO8KaRrx>u;@1~imQT6|{$e9MG zr)+PY4{S|%9T%j;pp;gcG}!;1F2}hX7Z^0a(~+stD0G=$)1Zw!p}o6)**Q3O8S!t` zF=Bu#)Ytv6h~!0VBj`b(k*ZY+8DAvnaio;2LPVSh_uvf*@B7g%Xc-Jgk$?`4TbsX) za{4%~UVSt;Z({i1Y#!h~`kNyCBPo;lp3)4SuYV>7VfH03HW1IKNnd?m|1kw06uNbd zsWbyl5X)aZxSOqsL7F(uG%2Ci*UbwwQP4wE;r~$g$p=F%h*8~COXQEXpM&1#&&0Q+ z+#V)Ra88=r>hmA&bsM-leO9AtwNgyb%UbiS&Q_7pZ&V!&zgt zmAhK4CC2_%D)aXnkoljlpCvo^=LM2sL;AOb6i8zf3bpUNriqCz;^4io| z`VA6vtu|g;ycldI*(vfDPDc(ccE?r>Y%Ou&w)u=!hwGqM@1WtpJzNk|%fZzyzpfRh zU?u{v|DE)*T8VxXynBAblZum2i9`DoUN;t5_dUvp}yHY zZ$uGMBT@K%H{!ouX@?hOW~Is)_!Fe*1~f`g|AD(bmD3im3&N7a;PO^t?oIht{)7Vy zW!IuA+=gi(40OA%%BGp5gAQt$=^}T&F8K1$39m8J^qG;iEzk}BofG3D(Hq~G_f9JZ z0*7`1LB~*+?&p50!|K(+ER@b0;mdpKrWf6sYPjO+Q*Sq3uk{qSJzx;vepJ2{V9<(x z#VHD}=w?ki+`OZm(4$^0_qX%Oleo8@m~hd={52g&?poj2r*wh_Dxh57do#bx{%u_U z-IhVit}7+1Izz8HM{B9*|4IpdV$m)F!=#erZ5`YkvP+KS)Eh%-W#;rzw0+TEZ+i$Y z_TY}SqG&4=6Rd}aD7l7g73B73_s+%C?UvoY+hJrixn?C>xAU;4V)wa%i%L}IH={Po zuFiHTm=j+=ofJMsA8v;7u48Hudv*nqsg&WyVN~BRC95>~V1__K#@(Jod-!X?X=+I`SooPT#76lk{*w&63VQp$ zHUdkC?uOP(erQLVcR zycP?F3dTDimdeYK!)#haHER;0hn{Js<@vcDmW1zAGoWI>9IlFZ_Rn}fO@3+Ov~5Di z8TKdwe`v-4?9N>g_r(-0o98L#J)Xuu_xG?0cipxd`nM84VwPC%{tc(JxA0Cmr%-yIevTykNxpv+-jg3fzhR$qx>a<$b!nBy zX1EY*krkkps&DVZ5WK1dM62Dx1#h-VXr07H-QYg{aq&MzJNpRiSqc1EMo2$-fi_t*39XS6_6bCP5t%&*~CSAEh{Sj z3)zZlSLu)mtj~O}@x>Yt(RvWLz1>=aVS+3)E zxp3=~u#l7*4qxmxP6Z)UQw45|I%EgfDC z$d$1&ZIGb{e5bL}L%5niA|^xl39FB-&UI5V#AD}=OEljmKm0lyk4;Yg)hd*b>D95L zV-D=XZ04ns_N(KIh*9@{>=V;qM`Nbz zeMA7%;Zc<7dou|*ffs-JbDTypJ3>=;n2Psm61DT@j{G-T1|NTNa7W|km{zI#NU#?o zo&(ZjVNp^mLuORdK3TZ^w?sSGwJ_UA_klsTDBX&R>z$UZJlsX{>W}JAxgFG;PIZvL z`HemJqiB;gE1${NqLtgGDdG>#+fJ&QKQc21&pJP@u?=HcZ*5e6irD@(h+*{RFI{l+ z22aFgG7%NY!f#MV6DP;|F$X_x zjIRI5aL?kDS(S?D-xGGeOdy;W&Tx$=d3ojlx#dl5Q^9{gZqef9`T=FDo$tD{ZlHML zDbc%|X%o!bz%G=XV>D9cTqHp+ybR@64n$4r$I7F~-vU+WG)tn22HBKBx6Z`5OUm!J zX{F~$W-rEmKklG#!@;l!<}Fguyo>wBKZu7?cm6Bjo|;>I$`s^#KE4qIu|W6&;01^s z`-?VyZzl2@^tZjd3)m$lP(^p)0HB)5;X>Mrq9#R{L>if$fNFjD^h?+_>eAsiP(__y zkNfY?NWsBh@*{KR%A!aCreYA^Kd8f?Yhar<>Zl_qX zn#UMxUnc2-ixK`(0~u|{=So;Z+|0SC&@o%tsDY<?dapR{8@%1fGlGj`<8OIiM*BA%VIjW@_0!%jW7@@fup zaD2m+smGl4Kdv+uwJ(ZfvX6G+D(F0Jo6Y~u*XzAsPKoz$v6&^8wCQH-9SmXD9aG;y zrDhz`6+dSPLRejmt?G8a-H<-@XtDS4nb>0xV6g10cS}_+CT-cNep1uDu)@3hOwI?9 z5nFB+4J|dr-~?WweyNVq(4ozNz0)?#K?HJCcBe$4HEYzs+b0BVYBE?}S280UhB2a! z4(<^L0E85E(5QJ0RQfC+ccZu}uW1`&q# zEKk8SX@?2)&PhbLIC3AGWL0#9|8Ge}BrfFcdLWFZDU;Ihej(n<_2|&Xp#ePET&T!pF_NsY&~MW~_Xy){P*}%U)@M?goQ6*pG^f6->aUIY z;nb;{*GLZGz<}`XTTf%Db_p?ub-6J&iy9__m#M`_ZOqjKUgg|iC)0c~j4=69S-DbQ ztqeVCrZRu5w+7(SHXqNWj29*FCLv_v&0oi(#v%G&^~+cw_Ofv2XcdAYcGPPx`I(pj z59J3LdP?Z2o>b})@Q(NXOK?2#{O;%Med`PwS-CnH!EMe9?@lY3MS5OYe3~z`fYdcrS3Wt zU%wTWyHi|_m$UDxu$Z`elpKJ-u`lM5aA?_J%^ur#+^g;X8Q10K_FMB>)sq+-3s}+p zP-n;faHW~w(Ae{VxUmr zE{r;h7^8d+x}cBCTKh0&vTU#(_v8u}!&72tT62H+l%!0hpbnWuG=uiFoYp^VaMX=o znY8A?n=lhqkqlK}AW1co_7tTz+S_`1_nf3U@pughgHkZ7T!F_Zg%^QeQe3@W9k8c( zctT}YtG^OD3!=rq2%G;{M9%QX-d8uEZ>)FVWa zZRz~g$^_mV67VEvh&DNRnOvN7Zq%e!sYTKQ@n;ly+9k6$rUWym2u5lQY)V%>(Q5p1 zkNA^6yKA`{l7&WOx_#C8@b)tRw%XWrtMJwNG8gtCL8AJrhFRHhIfYp0_bl4xb#U^U zn~-%g`&EF2mv(+0JctoG?k45$ zS{D9?1@Wc1aToXXT%}*tFBj&&D z4^kU&Si}V0b+bsw=G*y~W=Z> zO-as;xxIM4fc8cGVxcjJ);L^Ubf{X8!`d->3QQ`g+GG$|yC}(2;OrOhQLj}ni~vkE zS@$xo+DZLb3^$bTkZYT4eM zlFXcQKceR_@?98O`44A(vbJaa6ZDBzPd;?lvQNMxHFZZmWh%P3^q2hcHDCHMqjb%J z*cF?IST9hS+E?&ML$pZqG{PbpTtmm*&lAOMPLDhcC{&Klp62)-{Hr^#5`*L6cxYD` z(244xfok7KhpT6XTH_raqfLGtGIB?}V!}t)`xTFk)*Gcin>5{CtsV9W94MoiVh$4N zoc3EY!pLJF0$q;#>qv2T#4+kPW(V+i z)rf&h-7#fbu)Yib%1hCtpU#?F&L(-KtFm>THqUDD&w>&xEB~e(zbR8SX&5w4Mfd8a z%I`BO#WPrEP}w$khc0I$RmXl$uk8M`w$b(v!m+#C&%htiG^EP##V$Fx7jK<}3T(xw zym%VY!K&_qMe_Ly$c`91QI86%{Ib+2WnzXA81^yinl%P(LVVNjRVn%0#P3lQiIsY# zGZ42yOZs>VZ1KNyZo3^mZ7){DRufe@9clw~=>4nf2B^az^rE~}GBX`4&R!8%@G7F% z09y=LnqZxbu0Lc{i+rT5D=TLWSRk=Quue3W`t_-LbNpxEb|0cs$>;MT)bhbo)&^u2;a>5yarvEDDYLFV=5cAfey17ME_ za*fZDEL3~rzJDj(41Bah70<7$+pHU3m+^YY1adQV6m&N0mD^vH6wY`r>t25VENQk^ z1iRdvsku8Ta(A2NF*3g>T% zr`c;J4?Z(IO(S_KThMHm%I)*yoyPbK7XbR3iZ*vo{JahIrE?oZ9e*vgYW3Mw%Cn6t zmUaufLMsuJ6Ernk&~`dD(NcWQ%aDkqO+LmB7W@&&qG>sGH8?TO;jZ#f!^^S~UXi#gT1rZb z&NL7yZTBQyo6V4h+qu4kl!+*%QlBw~P4I2eO(vW*2g$U^C+p-N5|Z(=i&*3!0N1_2vcj|ewyB89=Fb*U`epQ{^X)_3&MmUBN~P*tiSZ!g8+ zAgYw0$KLAk#(vte_J@CkKL zILxLjv&@X>g2K! zO#)R>_^gons%9HKmlS%4*Ya%&e1e3%)3sG7qY6B(fexvGydoyp*(MB%%HmFXJgSN^ z%Qm0;qR0w#I$!d>sBu`*FQ?$SxUKfM>7HBQ9%^A1J=-q2$2u(lUQk?|LwnAM`!w7-fR@*BCD_qZX3~V6!x}+<_(_qm?C8wK?obpTd(=dl=!%7Xanp6hOKKl-E zV<4ncw}2Ez*+*mKh}!8s6Tbrx1RA~^!N&j$ptKgvp*BIewg(Zu74dSx;QL)&o@m(? zf!hhR&;4?8!26Z!*0`!`K1k6sH-eoHKsUlme(gfG`ug@_+Dp07ADd156kDOm{6&bb z!mmFGV3C`18j5nJyJ74WsE?1g;_U_gAk~{f#`DMmJ*QKyy1iQcpqHN{OYzDIj^jY&9a zs6=VDNhklt!^y5k=Srn#Hpb=wwe6H*!XHl!I2bB&GZ9N$lYENtxw8=)gi)w+ukN?r zUzwgN)L*m=Yx;TYw};ULm*5agT^=q`dkR8anM2=Nq(Qyklazo8YlvJJVg0p|JA%-Fn#V1h z<_KwjynrEKtg+9wFT~-{Ca@vxlW~aJzYHx%{T6hzgj&nimMiwhA$pss!t>&-6UCFL zphr1Vc82cU+juW3U(xqtrg6oyOY7I_Hbr@m67_YWT(6aT%RUHqbduuSe1w@c?aFA{grKLedx_dwxq(ebEL_)f|q+5{g?uKFR!Qc1ZKkob9yMFgw?>+0R z8P*#1K6^js*-w6+y)zS$7s4NAn%D$93HYJGk}VRACHn!m;@mi~aDVJ>F%Gj#z+Nk^ z(5=VRS3^j921HDzG{W1J@BsbN@yC{(aa5a>u_HZ=Y(>A>@x}`vWm_WVQ|Ehfd%P(j zRamG!VSGetym*}Kmo`y?JX$uy#5sX~**++s+5YoYmxt<%ONu!G2mqI(HlD0LX-88( zEnZXb$L8{zi%@F=5f_uZ&g8nFP&yJ{hCfIl=1pqmUWR}Yg*nKEfm~+lV|J?JKchCt z`h^G23{0%|>F*YyF5Y!)B8>J=Oi2J(3Ee=d0qheD6-{p$%fZCH68yw zb?R$bPBLU|nY%%!g{EwYbK=J3vCfGgy6To5~zi?Db1KSAxNfZdRN-J}Gc@@!YTCuWn1Zb$TM$rX51?}uKF0M;a8^R~*# zif#dClWtnmn}=h3h*K;iJwjznyMd~i`LB)%zGMoT9@Dcu z^VrzJUhahY$i&#VdAH=}pfI+U;_{;b=n(6lacjf;(4Ok>i8EBVWOe-|uvYDy#B1q1 z$-bUm7PYyN&uNkFJBwA}4K{Pv_4<*#wvC%O7;UfauC94Gk9*b0&A5cp`3tkLoia;o z-mR7zaMB1&wE}M=*^78|FOO4m2@$+Z1`Di^w^a+u3F7p~B?5Uc)i3 zpY*uJBvL?>g3xCnfN^>)GTMNAv>>xCOXkn;pS=Oy$Mz&F>I}3q0A7#6WA#G;vU0!5 zuH6U7+B+fw$>lgt4Hp=(tjrzWOn!ZRMTwz=pr>Ta)=9lw{du;M_F2+lZTXup6j~?j zB6jJr_MKmM05AwWMKl^@sbm%GdHJzecJhs6joPum!F+D(qGeCGArHFlm54s#3iUI&Q%s)0Ao%|){{Oji95k{=8qPqql#H&}lse7->@*g}4hMUiWhZJlmlxNi*n0X71LLEq$5 zL)tac`na+yx})Nr6XE}w)r}KL#iBeupMsT%;^rKmYNf<@kWqgJqL!T}Jw6X?vW-D- zqJ));CZDsLp)X~k%bSm1)~NubFZ~FFCQ-uL6H{p>j9Vq1oT@wM4|Hfa&V7lwT8xRY z!)^X(iuCQCTXDWE%aG?GDteNw??K+8dMavijd1I<3Ac(VJHX#A3TXZzPT3UZ^@l+* z<`*_287zJ~rb)FzvmcUoT`%G-!CJZk^OutMuVMe)%`dlVogLF)EJr={&C){*=U!vo|-bet!6Pm6|I&@NeZW5V`Dtmz) z<_Cg-!r(=+W;WRGK%j zF?wnc$D(^xn!1I2=no@^m%Rlt^1$GQcheK2My0X2x|Pf#&vKUO%Xj4yD{8}<4ET+G z3ESsPUo zJ*0h+=v#@;jtl4&9>L63G`6&w+G9Q6UvwWLC=iI7#u|^fW=woyT&zCA;={*;(n!hj9eCT49z- zM|0{X(@7rwq604NU|-dEHYL7zD)}T>^r(e7pCDckzg|qPoT>uPd2RDE^f?OMGlNNs z_GAE30m3C+IcMF!8g}lswl%Qcz3_e?x?nTTB$ugaejX{e?y#rTB9MCU;Nk@`)^P0k z*eCJ56^gs|(aHkjtH6{`yf-s6NW1B4rXLcoi^)_^zfH|8J@Br|sIzjtW35?4uMn1; zOHEh4L5b0ht*CuDI;~O`jJD`*DWhAb5TgTW+AD@!WAF_fnlp_lM3||Y6i5|52dfn4 zOiAn>+mJ1n*@?e4OlUy9xS6F-wauT_u8#Wn{0p0o-T-4b*tc_irgTy2XJm)-`~W!W zX5`lX>*|<9D<14MqAAvkSY7F&hR(-2p9B{TNJJ%<#o~D)f%Er72_l9=S^4LhxIMie zrDQhOJI?*EnkILbj+9T$8I9W{nPSiLX?(W^pry&byRq~Qi-n^l8O3zz#x1=`V^AC? zf`4HHylR_(jBNKt_`nWmpx&&KO^SWqhrr2C4ZCMmtaw( zMmb;5qsKgH3Sl}uPfmOBL(uaHs1kZRdNbf)FQ(UH@!f0U&#MRxRKT{Uc)fRVX?FJh zd825yM%TN%vxL@WUR@D)E)X{J?_`aSxy=`EX70o)_Xp?**YhiZI5!bgh6F8y$~@vx z1d%;5W2#=?gsrb%vk>XoC+<((I8YTiSHt%#TNAwlh<=AP1YRgk`M)^L6{<0s z97A<-8jgu#P6e{@P}~lViR2@6=*4>)Nm^e;rl)Fugvm~fLC>o@-H)!K%LzBoDR==^ zf<+LNUJI+#ah>d@a$hDnV%lS%8cLqepH$C1t5b_V~ea7tRmcR+?nuD zAHs{cg`PNfIiB*0B2bB-`G1{csh0fm+44xO4i1wGm&D?mtr^|CpX*@*PEkIG>n!i5 z)a@d?|oDLpOGV{qVM(o|ja7Oux9XQ{|awKXAFA_9X72 z?cjR7ode|;L!2Xi@G}gS_g*N8l;qy6CM?##NYNk9IZZiU7&r2t*HlVFclW*iWM;j++t=2Ou4_=%@dJzLJA1E=CH%XAh3GW12v?Ns`B zZpeqfLB8vGbMhDNq(h#o3cHmeZ=@Y0di_=9_WZKUTH~EFoue*KVj7`{;!|7@ zGXw)c<-{cPzmXhvkfFg+NTy4=Esd_Q>OA# zDIv^R=)2ke=a-$0=4vG{QyLQd-_dH0O?TOxc-RUI8L;D0o77y(AXP!w$V$6M=pmUB zBeCge1RkntU#Bgd?ac)D#GOvW$VHO%Zo~lzJtqME2kK`OX9Y}M?@HH-o&3h(mIr;S z&Ubqw7v9cWtIa=VD>#J&^fTz?u7-|p$-iIC8XeeL4;oaZFneF%!s6GkiRTv)Rp+yn za42y+>kBFh5l9s`vYK0)6so@Z3rJBuf{rCer|XH8MJtQ4lX@g(yCr55L-UE)x2uB{ zDbx<4Yg~fQO-dJUA8O`paWQ0~lR&XDI6|*&!jSGN7o4FoY?r?;YYHzMy^te?^Hk_p z?<_BW={B`;bGU&3r99PL;}uf!z^;hjshBJLxeMU1A#h)wlS1Kw`UytE3fb&8^>eK| z70;TRfiza|vWG>a;{3H?JQ*YRa=e;x7phz5RZ6dadX&2;Uoz!83~>Xgn=ulERS2*{ zzzIsd?r&Rie7rD>qBbx$U5i79E6b?a4i|l9APIDYtTWn5zmiN^7R!tZRnlweZT~sB zBjL2Qp_w=Z*|U3<#VL&Knkk{izfbpbzAT6o*W@&;(@x`^4)Ww@>f`wB!!@K^Bj$;U zW)Zew&+%5OPCwRkP)Mmcv~`YF3>w z>|D$hDf&aye2qM681>;@7cZSed?ozAOIceK#X}m`3(J zC}FiG}|AVVqxiH&Qmw$DCA+>vdQ62U}a0E-YmnUu)(1VUCv%M%A83zJB>4c z_KLi&@2A&vub(O1cQtzw#wFKZ4BxOupwN;s$8cxaOl4QA}dpOduVQ|A(-% zw({p5!l5We_O7TPNZ>lS&n8kc#C@RZ2e;eO14AKdoUut~SMynY9TyD)M zN^LklvNW&JI^Sout@>FCbCW3*DUI4VU=2;teJ{;l25WPg#6R{~Pr~GGWBk;zPV`$3 zB~@c|vfDM<-qE>dR68HHDk{X>Ei%(w(1pL#=)s{UfTRUV{Xm(+GnU)L_;`eSdJOWa z153uYHz@;4BGMvh>7Kj%jH}(>g@IiDA^+XtZ1)m(iBK*O5O3IT75k$j&X9)^T=va? zeB4Lfg2WJ!bd;pVCanCq(jb{O0V8n(D_IUCH6l2JJfW;ZfD|cnCBR9_9b&tW8xEV6 zp(lDy6gmk>x5=zIOML)zM_DlHppswWiy=kRZ3K4HX;Jq|c+UVkGibSIt2JNJicXj{pWC;7*e{yd+k$@{`}Uye*2WrqqjtAkJBb96V2 zw(`BEOip%pcW3lSkG55X;s%G-&F-ler_3s$GR`cI!*GsJlh?LvevRP9DG(iYly*$gcxAd{>@)@ zZfin68f1we+ixLiEkYI;?(^2RzqG9!CeqDW-Rth!Hd)e^s?1jJc77=>^`#d+Yi1v% z^r<_q$_eSR!=(cw+do|ZU^o}il=SyS#KWfye1^Lxi*DzFSnri#PTaqiWZobFFVU1U zKK93>a=R8dQCc^|*NtVfO6PmpMJg?EH5kxl!Nv;k4OTkTf88$|iqA}s+f#hWhkdc?2{g?OG&0B$ad%+NT*lC))oIK!y)5fNPXkQDP!k~T@ z;3uvB!xEvA_}lce5Tj&oIKRjn4g08Z9p|qTmgYeSx2FW)mWCG6`8g-^ntdU@MHwql zKOR7kS%^*DBMfz#*0An&nC&O=8zl=k_xW;?#5bKvKuS>X3H+gE%by99 zM9y4-vO2*FkILGD-vM6KxNNK&;&67vD0?Xeke_NH9V-+VBKsR!uD6o3VBdidXp-|^ z!xynjri)}IZc)5A?e!u0RO8!czXkJXueU^?_75Hqhk}eIm#q@UymeS@_RBsPc5<0H zc#8oAbU&Th zswJiWM~|-M_*MGVdKqftzRJy4z`bh3&c*watgre_4$JTkh}hV1S#W|{07X}QXI2Ma ziyA0B^N~IQo8U?586gN1&gPfurbezYfxrobkiV4ZySS$saz@RZZ`}u|ajCn5E~VuG zrFBW-i|p`*aXpvbSxURLlDHR}2p^!B!_kLuaAg+_IC8Zkk9sVZ(~kL$FR*|2FU$`(ALSAY~rx=qn% zTo}^(?oyTo5-G!Tg&1%dHja`6Hb<~L8wL$77DelCk(`92TNBMxO`fjbbJJfL)}q;qmg z%fON`26be=2Dm+#T5jswQWR#~1}!IFQ<^O&Jo7LR5B5Vrx&5^?h=`U+|GcR3t~Vka zwF}~_bD}%bpO^0Jk|wjm@X&l19qVB(5Mgs9_(fA6;7w21}6zG zqZx9*zCdi(FUt#j6|bz;%xg4muYUA}VaT_;QCxsutC*%*PgdY61&+^2ZSdsO`K5sJ zY&iDrhc#9_>u^PDp0T)jg#l**?@)?pYO%pQtf4ZrsLZS(lfo*kW)jB~RWgvnFu`+T zP(QhxNBYvV^mbOIrqMoowa?<>&Om!Yh=dn;0Qdt01e}6`4QNwKd1rOCo<;}MUa6x# zDt&OC)#h#bF23yvPwoXu679WaPvl4oR-$n8dd5zQNjhP=f6ed=<*sqy>}m!M{oy_P zvhl-)a1=vFZ(W&DiQ(vcls(l;Kyy~Eo?Ghck%^yF<2(o}C&e#X*s_!`wH0eYhwx(D zYq!n_f}f)lvDp)35bN679%8|^z9575wX@dMIoz56=(uu?UvPqc6M7L(%k+Mib^&dPROcry{JEaq>j=NbK4Xd_>I02IsJ{W^I#Yea$#?B?oa$G)=nZm33L5Sg#MWSpu7hxkCj2K-zoi8-&s*NQ%t!0oL%XW zC2}ZyXm-_7qW?NbjqKjiGFs<|d}Ux6qrX*Z%W)BBN&+AoL6C!O2=jP?B-P5KB%^>2 zs#N}GMqQo z)SDu&qexKh|BOiKhjvp0&TgE}Bb6-F$;yE1{Dvj{MtaGuEedb@ z>jyL`*U(&u!`|l4RcP9L*kIL)XHVEDND7#TLkLq`Mb4|y(H-9Yk4600R%9HpPwRU5 zn#iP30vV0BnC(Wckt}FNoT16R@vspzaEmN8$^4&J8}CHg&>=Z=_k0wb1@Y~nsn1rL zON-&@u6wJ^DV^NWspR|_A4F6*<`_>nf+8nYX&?ekqSn{MW6bEBKiv?q^#O{W&~_@9 z?J@4hVMmQ|*wE0VKRLX_H{?{z_jZjWu6OqBZWv(HZ2fdvBUD%P9MJg$_&*V;I<`I$ z<$Guw)hp0{wPnC4ZOANyY5!R_sOqRKjE)UfWpi+j)fcn{v4o(fjOvpObj*FW6WHyN zRugRI$U|M3j_%gn+ZJoR8d#32qT{!AGRKjHl=CWaBdI`R# z11>AxA?A4ZHW_8gSDVmTGZ8G*x0<{3NdF-)mx3;UJ)}Q)*;_8~08>!_uDiGb6d6v< zdCkvx1rc7^4FG#TfqVmljxL2>7RFo9>XX;*=hS|A5&k%2pUydh5CjDklnow z4}{9aq{te5QJ+3pv~x)Z?(R5)zpF9>u<_Eb$IaHJF;wIdfE1r>-izcHI|Md{JtiFS zwbdLyCx@a;zS;@Xhc08ocx&uho?k=*a$!nS;gYCW<9>0K>J0Pb<-d=es=EFetIv&d zOD)F{JKfyQH!BP(^rr*z5GPx~Zjt?#aVV9(?%*RahSI9q`k`yyA2niNz=myMf)UHl z%=5oLAl%j(j7Ohu!%NFm<0*Ikq@Tp;%c0eJ?lyUi%%&s~FSjuTiy=3rqw|_4=bVDh zbUqrR+_NOGpWsrow9BHHN70~G@ZZrfW_Ad9B;(*33YhT&mbD)E$B9`4oe3cBGwBG$ZgtBF8 z4VseVz@pC4H7#f;Mh!bo#CBXr+XU3{VMFyBFM(KHFCN!pnClb4X64w0S;5)E;_rJa z(H&t8p1*z{!|Jio1M$0JOc7i(1}f`vqb$TbzM!IGsQSh#?Clgpk^= zxZRZ;z2uEn6`)yeLr1_ZR*-~vR}Ty{D)TN#C_gTvpxwWxk6ErzK>oOx74tu%=10qg zP7ddI9PFYXv5Dn#aqnF`?)Qe5&Ck$|9b0o+?&I6HT354$Ajg-Jm9U_yW4g@N&vuB=u;28Qy z1%f>~nvP&wr8`e5Y{<3dTcoL#1%q-Ob=zqH%_OHA$^^tVU0}uhW4@Fq#8!k~Vhy+S z-uYCj#xn^e02Q9$9CxDOYGk}$)W(02@@XNmpMg$wE>mrjjA5PD;(ckUW`~*O!#=s`4{WA5T8#(X2Myn*0IS$wJ+(REDsDnCZ zS(9X z)P)N&KY58`3otqtzGQ8OM$!`JARe0)#Zo%|gHmdPThPSAFH&g$FUW=GpoF*+II9R$ zNQq^&_^n8n+4y79d8PTe8~Px6HVr_3wP8A%lRU86^B^W_O3V-6?<>3B;0@lv_pH)_ zD_;Lx1>*FKyuoXm#Gw=et`ACuXJW0<=cJqb`sVN4Cv9M?;J=*=(R#uOvsv9k1sWAb zMM6FT0DxcR_}ipsPY*OM$=`NdvjRQYwZ$Z#4W{$G;b0u^&qY1?ycv#4aP^?6wQs)Z zI{G_<5X!w5+u|RFX88+1*#d-e&=5&>&A3aRk)~Wdqbd>kAteQ{K!=g?Bl(A!o5|rm zE7D+Z?QXEc^C*{ZtpKc{+KDlh9ds`XVMN$OB(f7gkkj>KRHxF@58)J2;>og4-S#>S zER*?}N6ZZX%LEIMCLzAMNs6a(q`v82#N>vL(AN@V$=Uk?~`R zaFGk;=jY9T#UI|xaYbk=@FDwsE#pXvMtG*xQn(D=p_mKm^aX091X4A2YPP=)oG zwk_;X>slAN6c;(`hp&Zw*r{@%@3w8vm);GMy!jH%D-;(Ji(@Gx_)PbU)r(e)hB0^Z zp_8J8$t_{w8;p{VVBta5`}u&9YI0Z;T~E{-6h?vIJUq?fv+u{A53c(V<>hynDA14$ zOYL>CEHskPw2`Na1xlD*fu0fy-xp?A8R~wOgTILWRr!#wY>BFiLyE^{G`G{e8E%g? zrQ9pojgeMg>MhYxm;yUgd6f7_IW^2L`?`ZU%*|JP1H>r9Z;6+rBPG&vgpA(f;$0B6 zEh&QU1+6bFxq6@1YYd-+k;Jr)uodCN{A16oYaP15R~3&Qf90vQ$8ZP*cm~p4X1e=y zF?xu&wBq{5dv6VR8+bJt9GBw{mnV^DoKP`;zS62Zj3}x_N@&t1#^BkH{A*L}5bN#eB zVGN9j?=M;B0-g#zLO##x#I990Nq%AYagJW1LiNd>9hz9T!?gICC^$|0H~H zD4@fIW*B1h_de_IF~Z@RKc%cQC50HHa|yRK9nQ?9AiHgD$*5u*40YQKb@Ei7$5nX% zfh(9ij^^aM@jCQvZmC@4;e(-a_y*PxDU>4=y2EhK&!CP0x`8F)r=n)~UxBhDp2NIN zRI3m_O|O=3Yv_vM1pXW7%7lZRdT&ty5j~};6sE;ojthu@B_$7NQV1=AXUG9=(k@9} zsL|*p#9{K6ZbUAYKC*wcBs=^Nj~udzQsC~BA`zXr!IHVT_sYmA^Kb=i!=>h#O8qkx z+|`iLRr!K4I59VpCf6>#iDt34N5DJH1aRvO^C6i=V56pX;H(E7M<@-fPNCF>xa)rJ~;dyPJ%u*0dA`gbToGvr0;J`X|;XDLa*9~ilIN8k4;t4$})%Fh3aqI01nn5V{Ba_xcT_0@kkuV z0*>pjDfi^$Zw+9{jG)L^VNH9erTrD+M)UKSdmsGflv3(ncB0Mz*gS}JfZWX1IVA~R zy;}BBq=gr1B5-qNXr^ca?;=dlN&{IX8xOq6zl7&$a9xbOxGxV%UB>p!Zf9dCtvW#+jJRri05!T9a0MAshk+`5Wx1RnW#0rZWQY~BNG zY^W$J|Lv+XS1BVX$wz0Q^!r{uaQ`nACHutqeka*~^`VxbV_;ep)_*>ND=q(8_B$3^ zA-c08IHlX4Kx5e0Ua*;eQsrr0l4@~Q;A6gE(-IKG@R9Woa1NlR12&~pjaeag0k?p@W^#(4*b6?b>H|XQik~c`|pan|Nh%0 zu+s1U2CjQRRfb!T)!5 z{a+&2zc=dt4_*Iz5B@*A}?>|KS&wu?l`R9LY{NKr%|8)J&oP1#`NKB_N2}=Df RyyN?k%fXai6hn>t{|9>E-&z0w literal 0 HcmV?d00001 diff --git a/claude-plugin/.claude-plugin/plugin.json b/claude-plugin/.claude-plugin/plugin.json new file mode 100644 index 0000000..1ad0e61 --- /dev/null +++ b/claude-plugin/.claude-plugin/plugin.json @@ -0,0 +1,20 @@ +{ + "name": "aae", + "displayName": "Agentic Algorithm Engineering", + "version": "1.12.0", + "description": "Agentic Algorithm Engineering: autonomous performance tuning through falsifiable hypotheses, benchmarks, numbered patches, and a PDF report.", + "author": { + "name": "Christian Schulz" + }, + "homepage": "https://chszlab.github.io/AgenticAlgorithmEngineering/", + "repository": "https://github.com/CHSZLab/AgenticAlgorithmEngineering", + "license": "MIT", + "icon": ".claude-plugin/icon.png", + "keywords": [ + "performance", + "optimization", + "algorithm engineering", + "benchmarking", + "profiling" + ] +} diff --git a/claude-plugin/LICENSE b/claude-plugin/LICENSE new file mode 100644 index 0000000..b2b9fea --- /dev/null +++ b/claude-plugin/LICENSE @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2026 Christian Schulz and contributors + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. diff --git a/claude-plugin/README.md b/claude-plugin/README.md new file mode 100644 index 0000000..7690ac0 --- /dev/null +++ b/claude-plugin/README.md @@ -0,0 +1,11 @@ +# AAE for the Claude plugin directory + +This folder is the edition of [Agentic Algorithm Engineering](https://github.com/CHSZLab/AgenticAlgorithmEngineering) +for Anthropic's plugin directory. It is generated by `scripts/build_claude_plugin.py` from `AAE.md`, +the knowledge base and the report template; do not edit it by hand. + +Compared with the plugin at the repository root, this edition ships no programs: the agent writes +the LaTeX report itself from the template in `skills/aae/report-template.md`, and knowledge-base +contributions are prepared as a file for the user to submit instead of being opened as pull requests. + +Homepage: https://chszlab.github.io/AgenticAlgorithmEngineering/ diff --git a/claude-plugin/skills/aae/SKILL.md b/claude-plugin/skills/aae/SKILL.md new file mode 100644 index 0000000..138f93d --- /dev/null +++ b/claude-plugin/skills/aae/SKILL.md @@ -0,0 +1,363 @@ +--- +name: aae +description: Use when asked to autonomously optimize, tune, or speed up a program (running time, throughput, quality, memory) with the Agentic Algorithm Engineering cycle of falsifiable hypotheses, benchmarks, and keep/discard decisions, or when the user invokes /aae:aae in Claude Code. +--- + +# Autonomous Performance Tuning via the Algorithm Engineering Cycle + +This program turns an autonomous agent into a performance engineer that follows the Algorithm Engineering (AE) methodology: a cycle of design, analysis, implementation, and experimental evaluation, driven by falsifiable hypotheses (cf. Sanders, "Algorithm Engineering - An Attempt at a Definition"). + +``` + ┌─────────────────┐ + │ Realistic Model │ + │ of the system │ + └───────┬─────────┘ + │ + ▼ + ┌──────────┐ ┌────────┐ ┌─────────────┐ + │ Analysis │◄──│ Design │──►│ Falsifiable │ + │ │ │ │ │ Hypothesis │ + └────┬─────┘ └────────┘ └──────┬──────┘ + │ (induction) │ + │ ◄────────────────────── │ + ▼ ▼ + ┌────────────────┐ ┌──────────────┐ + │ Performance │ │ Implement │ + │ Guarantees │ │ the change │ + │ (deduction) │ └──────┬───────┘ + └────────────────┘ │ + ▼ + ┌─────────────────────────────────────┐ + │ Experiment │ + │ Run, measure, compare to baseline │ + └──────────────┬──────────────────────┘ + │ + ┌───────▼───────┐ + │ Evaluate │ + │ keep/discard │ + └───────┬───────┘ + │ + (loop back to Design) +``` + +## Session Start + +When a session starts (for example when this program is invoked as a skill), first show the user the Algorithm Engineering cycle with AAE at its center, exactly as below and inside a code block, followed by one line naming the target program and metric if they are already known. Before continuing with [Setup](#setup), offer the user **multi-component AAE**: instrument the target code so one benchmark run can test multiple hypotheses (one per component), then use the per-component profile against an instrumented baseline to keep or discard each change independently. If selected, put each component change in its own Git commit before the shared benchmark run, so each can be reverted independently with `git revert ` without removing the other component changes. Record the component commit hash alongside its hypothesis, measurement, and verdict; revert discarded component commits individually and export each accepted component commit as its own numbered patch. Keep shared profiling instrumentation in a separate commit and keep the benchmark harness unchanged. If component interactions or profiling overhead prevent reliable attribution, test those hypotheses separately and validate the retained combination against the overall metric and correctness assertions. + +``` + ╭──────────────────╮ + ┌────────►│ DESIGN │─────────┐ + │ │ falsifiable │ │ + │ │ hypothesis │ ▼ + ╭────────┴─────────╮╰──────────────────╯╭──────────────────╮ + │ EVALUATE │ │ ANALYSIS │ + │ keep / discard │ ┏━━━━━━━━┓ │ predict outcome │ + │ patch + report │ ┃ AAE ┃ │ before measuring │ + ╰──────────────────╯ ┗━━━━━━━━┛ ╰────────┬─────────╯ + ▲ │ + │ ╭──────────────────╮ ╭──────────────▼───╮ + └──┤ EXPERIMENT │◄─┤ IMPLEMENTATION │ + │ run · measure │ │ minimal diff │ + │ geometric mean │ │ + git commit │ + ╰──────────────────╯ ╰──────────────────╯ +``` + +## Overview + +The agent receives: +- A **target program** to optimize (the code under test). +- A **metric** to improve (e.g. latency, throughput, val_bpb, memory usage). +- A **benchmark harness** that produces reproducible measurements. +- A **time budget** per experiment. + +The agent then runs an autonomous loop, applying the AE cycle to systematically improve performance. Each iteration produces a falsifiable hypothesis ("changing X will improve metric Y by roughly Z"), implements it, runs the experiment, and either keeps or discards the change based on the result. + +## Setup + +To set up a new tuning session, work with the user to: + +1. **Agree on a run tag**: propose a tag based on today's date (e.g. `mar21`). The branch `aae/` must not already exist. +2. **Create the branch**: `git checkout -b aae/` from current main/master. +3. **Read the in-scope files**: Understand the full context: + - The **target file(s)** the agent may modify. + - The **benchmark harness** (read-only) that defines the metric and evaluation procedure. + - Any **configuration or constants** that are fixed. +4. **Identify constraints**: Understand what the agent can and cannot change (see below). +5. **Formulate assertions**: After reading the code, derive correctness invariants and present them to the user for confirmation (see below). +6. **Verify benchmark works**: Run the benchmark once to confirm it produces output. +7. **Initialize results.tsv, patches, and report**: Create `results.tsv` with just the header row, an empty `patches/` folder (see [Patches](#patches)), and a `report/` folder with the report template, an `experiments/` subfolder, and `session.json` (see [Experiment Report](#experiment-report)). Add `results.tsv`, `patches/`, and `report/` to `.git/info/exclude`. +8. **Ask about knowledge-base contributions**: Ask the user once, explicitly, whether you may share the session's learnings with the [Knowledge Base](#knowledge-base) by writing a knowledge-base entry to `report/kb-entry.md` and updating it every 10 experiments, which the user can submit to the AAE repository (see [Contributing Back](#contributing-back)). Wait for a clear yes or no; no answer counts as no. Record the answer as `"kb_contributions": true` or `false` in `report/session.json`, so it survives context compaction and restarts. This is the only time you ask; do not ask again during the loop. +9. **Confirm and go**: Confirm setup looks good with the user, then begin. + +## Constraints + +The user defines these per project. The agent must respect them strictly. + +**What the agent CAN do:** +- Modify the designated target file(s). Everything within them is fair game: algorithms, data structures, parameters, control flow, memory layout, parallelism, etc. +- Replace entire subcomponents (e.g. swap one algorithm for a fundamentally different one, replace a data structure with an alternative, rewrite a module from scratch) if the agent has reason to believe this will improve the target metric. Optimizations are not limited to incremental tuning; architectural changes and algorithmic replacements are encouraged when the analysis supports them. +- The agents owns the codebase. + +**What the agent CANNOT do:** +- Modify the benchmark harness or evaluation code. +- Install new packages or add dependencies beyond what is already available. +- Modify the metric definition or measurement methodology. +- Change the time budget or input data. + +## Assertions + +After reading the target code and benchmark during setup, the agent must formulate correctness assertions: invariants that must hold after every change. These act as a safety net, ensuring that optimizations do not silently break the program's correctness. + +The agent derives these from its understanding of the code. Examples: +- "The output must be a valid partition (every node assigned to exactly one block, no block empty)." +- "The sorted output must be a permutation of the input." +- "The loss must be finite and non-negative after every training step." + +Present the assertions to the user for confirmation during setup. The user may add, modify, or remove assertions. Once agreed upon, verify them after every implementation, before running the full benchmark. If an assertion fails, the change is incorrect; discard it immediately (no need to benchmark) and log it as a crash. + +Assertions are distinct from the benchmark metric. The metric measures performance; assertions guard correctness. A change that improves the metric but violates an assertion is always discarded. + +## The AE Cycle in Practice + +Each iteration of the loop implements one full AE cycle: + +### 1. Realistic Model (Understand the System) + +Before proposing changes, the agent must have an accurate mental model of: +- The target program's architecture and hot paths. +- The hardware/runtime environment (CPU, GPU, memory hierarchy, I/O). +- Where time and resources are actually spent (profiling data, prior experiments). +- Which assumptions from previous iterations still hold. + +The model should be updated after every experiment. When experiments contradict expectations, the model is wrong and must be revised, not the experiment. + +### 2. Design (Formulate a Hypothesis) + +Propose a specific, falsifiable hypothesis. Good hypotheses are: +- **Specific**: "Increasing batch size from 32 to 64 will improve throughput by ~15% because the GPU is underutilized" (not "make it faster"). +- **Falsifiable**: There must be a conceivable experimental outcome that would disprove it. +- **Grounded**: Based on the current model, prior experimental results, or established algorithmic knowledge. +- **Minimal**: Change one thing at a time when possible, to isolate effects. + +Hypotheses can come from: +- **Induction**: Patterns observed in prior experiments (e.g. "larger batches consistently helped, so an even larger batch might help further"). +- **Creative insight**: Novel algorithmic ideas, known techniques from the literature, or architectural redesigns. +- **Analysis**: Deductive reasoning about algorithmic complexity, cache behavior, parallelism, etc. + +### 3. Analysis (Predict Before Measuring) + +Before implementing, reason about the expected outcome: +- What is the expected direction and magnitude of improvement? +- What could go wrong (OOM, numerical instability, regression on other metrics)? +- Is the complexity cost worth the expected gain? (Simplicity criterion: all else equal, simpler is better.) + +This step prevents wasted experiments and builds understanding even when hypotheses fail. + +### 4. Implementation (Make the Change) + +Implement the change in the target file(s). Principles: +- **Minimal diff**: Change only what the hypothesis requires. +- **Correctness first**: The program must still produce correct results. +- **Commit before running**: Every experiment has a corresponding git commit, so changes are traceable and revertable. + +### 5. Experiment (Run and Measure) + +Run the benchmark and collect results: +- Redirect all output to a log file to avoid flooding the agent's context. +- Extract the key metric(s) from the log. +- If the run crashes, diagnose from the log tail. + +**Parallel runs**: Whenever the benchmark involves multiple inputs, seeds, or configurations, prefer GNU `parallel` to run them concurrently. This dramatically reduces wall-clock time per iteration and should be the default approach unless runs compete for shared resources (e.g. a single GPU). Example: +```bash +parallel --jobs 4 --results run_results/ './bench.sh --input {} > logs/run_{}.log 2>&1' ::: input1 input2 input3 input4 +``` +Only parallelize when runs are independent and the system has sufficient resources (CPU cores, memory, GPU slots) to avoid contention that would distort measurements. + +**Aggregation**: When an experiment produces multiple measurements (e.g. across inputs or seeds), use the **geometric mean** as the default summary statistic. The geometric mean is appropriate for ratios and multiplicative quantities like speedups or normalized scores, and is less sensitive to outliers than the arithmetic mean. Compute it as: +``` +geometric_mean = (x1 * x2 * ... * xn) ^ (1/n) +``` +Use the arithmetic mean only when the metric is inherently additive (e.g. total time summed across phases) and the user explicitly requests it. + +### 6. Evaluate (Keep or Discard) + +Compare the result to the current best: +- **Improvement**: Keep the change. The branch advances, and the change is exported as the next numbered patch (see [Patches](#patches)). +- **No improvement or regression**: Discard. Git reset to the previous best. +- **Crash**: Attempt a quick fix if it's trivial (typo, missing import). Otherwise log as crash and move on. + +Apply the **simplicity criterion** when evaluating: +- A tiny improvement that adds significant complexity is not worth keeping. +- An equal result with simpler code is a win; keep it. +- Removing code while maintaining performance is a great outcome. + +## Output Format + +The benchmark should print a summary with at least the primary metric. The agent extracts key values via grep. Example: + +``` +grep "^primary_metric:" run.log +``` + +## Logging Results + +Every experiment is logged to `results.tsv` (tab-separated). The TSV has a header row and columns: + +``` +commit metric_value resource_usage status hypothesis +``` + +1. **commit**: git commit hash (short, 7 chars) +2. **metric_value**: the primary metric achieved (use 0.000000 for crashes) +3. **resource_usage**: peak resource consumption, e.g. memory in GB (use 0.0 for crashes) +4. **status**: `keep`, `discard`, or `crash` +5. **hypothesis**: the falsifiable hypothesis that motivated this experiment + +Example: + +``` +commit metric_value resource_usage status hypothesis +a1b2c3d 0.997900 44.0 keep baseline +b2c3d4e 0.993200 44.2 keep doubling LR will reduce val_bpb because current LR undershoots the loss basin +c3d4e5f 1.005000 44.0 discard GeLU activation will improve gradient flow in early layers +d4e5f6g 0.000000 0.0 crash doubling model width will improve capacity (OOM) +``` + +Note: do not commit `results.tsv`, `patches/`, or `report/`; leave them untracked. Add all three to `.git/info/exclude` during setup, so that `git reset --hard` to the previous best never touches them. + +## Patches + +Every accepted change is kept as a numbered patch in `patches/`, created during setup. Patches are numbered in order of acceptance (`0001`, `0002`, ...), not by experiment number; the baseline gets no patch. Right after a change is kept: + +```bash +git format-patch -1 HEAD --stdout > patches/$(printf '%04d' $N)-.patch +``` + +where `N` is the number of accepted changes so far and `` names the change in a few words (e.g. `0003-bucket-queue-refinement.patch`). The commit message of a kept commit states the hypothesis, so each patch is self-describing. Applying all patches in order to the baseline commit reproduces the current best version: + +```bash +git checkout && git am patches/*.patch +``` + +Never renumber or rewrite existing patches. If a later experiment undoes an earlier change, that undoing is itself a new patch. + +## Experiment Report + +As a side artifact, the agent maintains a PDF report of the whole session: a title page with summary and contents, the trajectory of the metric over all experiments, the list of accepted patches, and **one page per tested hypothesis** (kept, discarded, and crashed alike) with the hypothesis, intuition, prediction, measurements, verdict, reflection, and the code change. + +**Template.** This skill ships the report template in `report-template.md`, next to this file. During setup, create `report/experiments/` and copy the LaTeX skeleton from `report-template.md` into `report/report.tex`. + +**Session data.** During setup, write `report/session.json`: + +```json +{"title": "AAE session: ", "subtitle": "", + "target": "", "metric": "", "unit": "s", "direction": "lower", + "resource": "peak memory [GB]", "machine": "", "branch": "aae/", + "date": "", "summary": ""} +``` + +`direction` is `lower` or `higher` (which way the metric improves). Add `"kb_contributions": true` or `false` once the user has answered the contribution question in setup step 8; after the first contribution, also record `"kb_entry"` (the entry filename) and `"kb_pr"` (the PR URL). + +**Per experiment.** Experiments are numbered `001`, `002`, ... in the order they are run; the baseline is `001`. Each experiment gets a folder `report/experiments/NNN/` with two files: + +- `change.patch`: the code change, saved right after committing and **before** the benchmark runs, because a discarded change disappears with the reset: `git show HEAD > report/experiments/NNN/change.patch`. +- `meta.json`: written in two steps. At commit time, record what was predicted; after evaluation, add what happened: + +```json +{"exp": 7, "title": "Bucket queue in refinement", "commit": "b2c3d4e", + "hypothesis": "Replacing the binary heap ... reduces time by ~15% because ...", + "intuition": "Why this should work, from the current model of the system (profiles, prior results).", + "prediction": "Expected direction and magnitude, and what could go wrong.", + "status": "keep", "metric": 2.181, "resource": 1.2, + "measurements": [{"input": "rgg_n_2_20", "value": 3.483}, {"input": "road_eu", "value": 3.976}], + "verdict": "Confirmed: -15.2% geometric mean, matching the prediction.", + "reflection": "What the result says about the model and what to try next.", + "patch": "patches/0003-bucket-queue-refinement.patch"} +``` + +`status` is `baseline`, `keep`, `discard`, or `crash`; `patch` is set only for kept changes. `measurements` holds the per-input values of the primary metric (one entry per input or seed); if `metric` is omitted, use their geometric mean. Text fields are plain text; `` `code` `` is set in typewriter and `$...$` is passed through as math. The report compares every experiment against the baseline and the best result accepted before it, so only this experiment's values need to be recorded. + +**Build.** After every experiment (after logging), rewrite the files in `report/generated/` as described in `report-template.md` and compile the report: + +```bash +(cd report && latexmk -pdf -interaction=nonstopmode report.tex > /dev/null 2>&1) || true +``` + +This writes `report/report.pdf`. If `latexmk` is not available, run `pdflatex` twice instead, or keep the sources and tell the user that the PDF could not be built. Rebuilding is cheap and idempotent, so the report is always current when the user interrupts the loop. A failing build must never stop the loop; note it and fix the offending `meta.json` text later. Before handing the report to the user, check `report/report.log` for overfull boxes and look at the PDF. + +## The Experiment Loop + +The experiment runs on a dedicated branch (e.g. `aae/mar21`). + +**The first run** is always the unmodified baseline. + +LOOP FOREVER (DO NOT STOP EVER): + +1. **Model**: Review the current state: branch, recent results, what has worked and what hasn't. Update your mental model. +2. **Hypothesize**: Formulate a specific, falsifiable hypothesis for the next change. +3. **Analyze**: Predict the expected outcome and assess whether the experiment is worth running. +4. **Implement**: Modify the target file(s) according to the hypothesis. +5. **Commit**: `git commit` the change, with the hypothesis in the commit message. Save the change and the prediction for the report: `git show HEAD > report/experiments/NNN/change.patch` and the first half of `meta.json` (see [Experiment Report](#experiment-report)). +6. **Experiment**: Run the benchmark: ` > run.log 2>&1` (redirect everything; do NOT use tee or let output flood your context). +7. **Measure**: Extract results: `grep "^:" run.log` +8. **Evaluate**: + - If grep output is empty, the run crashed. Run `tail -n 50 run.log` to read the error. Attempt a fix if trivial; otherwise give up on this hypothesis. + - If the metric improved: keep the commit, advance the branch, and export it as the next numbered patch in `patches/` (see [Patches](#patches)). + - If the metric is equal or worse: `git reset` back to the previous best. +9. **Log**: Record the result in `results.tsv`. +10. **Reflect**: What did this result tell you? Update your model. Did it confirm, refute, or refine your hypothesis? Use induction from accumulated results to generate the next hypothesis. +11. **Report** (mandatory, every experiment): Complete `report/experiments/NNN/meta.json` with status, measurements, verdict, and reflection, then rebuild the report (see [Experiment Report](#experiment-report)). +12. **Contribute** (every 10 experiments): If `kb_contributions` is `true` in `report/session.json` and the experiment number is a multiple of 10 (`010`, `020`, ...), write or update `report/kb-entry.md` (see [Contributing Back](#contributing-back)). Do not ask the user; permission was given in setup. If it is `false` or missing, skip this step. +13. **Go to 1**. + +## Operational Rules + +**Timeout**: Each experiment should complete within the configured time budget (plus startup overhead). If a run exceeds 2x the budget, kill it and treat it as a failure. + +**Crashes**: Use judgment. Fix trivial bugs (typo, missing import) and re-run. If the idea is fundamentally broken, skip it, log "crash", and move on. + +**NEVER STOP**: Once the loop begins, do NOT pause to ask the user if you should continue. The user may be away and expects autonomous operation. If you run out of ideas, think harder: +- Re-read the target code and benchmark for new angles. +- Combine near-misses from previous experiments. +- Try more radical architectural changes. +- Revisit discarded ideas with modifications. +- Look at algorithmic alternatives for the core computation. + +The loop runs until the user manually interrupts. + +**Report after every experiment**: The report is updated after each experiment, without exception: kept, discarded, and crashed experiments all get their `meta.json` completed and the report rebuilt before the next hypothesis is formulated. Kept experiments additionally get their numbered patch in `patches/` at the same point. Never batch report updates for later; the report must reflect every experiment run so far whenever the user looks at it or interrupts the loop. + +**Simplicity criterion**: All else being equal, simpler is better. A small improvement that adds ugly complexity is not worth it. Removing code for equal or better results is a great outcome. Weigh complexity cost against improvement magnitude. + +**Resource usage** is a soft constraint. Some increase is acceptable for meaningful metric gains, but it should not blow up dramatically. + +## Knowledge Base + +This methodology is supported by a community-contributed knowledge base of optimization techniques, experiment results, and lessons learned from AAE sessions. A copy is bundled with this skill in `knowledge-base/`, next to this file; the latest version is at https://github.com/CHSZLab/AgenticAlgorithmEngineering/tree/main/knowledge-base + +Before or during a tuning session, consult the knowledge base for techniques relevant to your problem: + +1. Read `knowledge-base/INDEX.md` and scan titles, problem domains, and key techniques. +2. Read the individual entries that seem relevant. +3. Use the insights to inform hypothesis generation in the AE cycle. + +### Contributing Back + +Permission is asked once, during [Setup](#setup) (step 8), and stored as `kb_contributions` in `report/session.json`. Never write a contribution unless it is `true`. + +If permission was given, contribute every 10 experiments (loop step 12). Each session has **one** entry, which grows over the session: + +1. **First contribution** (experiment 010): write `report/kb-entry.md` following the entry template in `knowledge-base/CONTRIBUTING.md`, and record `kb_entry` in `report/session.json`. +2. **Later contributions** (experiments 020, 030, ...): update the same file with what the last 10 experiments taught (new techniques that worked, dead ends worth recording, updated baseline vs. best numbers). +3. **Nothing new**: if the last 10 experiments produced no transferable insight, skip the contribution and note that in the report summary. + +At the end of the session, tell the user where the entry is and that they can submit it as a pull request to https://github.com/CHSZLab/AgenticAlgorithmEngineering (new file in `knowledge-base/` plus a row in `INDEX.md`). Only share transferable insights, and never include proprietary code, data, or paths from the user's project beyond what the entry needs. + +## Philosophical Foundation + +This methodology mirrors Popper's scientific method as applied to algorithm engineering (Sanders, 2009): + +- **Falsifiable hypotheses** drive every experiment. "It might be faster" is not a hypothesis. "Replacing the O(n^2) inner loop with a hash-based lookup will reduce latency by ~40% for inputs >10k elements" is. +- **Induction** from experiments feeds back into new hypotheses. Each result, whether positive or negative, refines understanding. +- **Reproducibility** is ensured by git commits, deterministic benchmarks, and logged results. +- **The cycle never ends**: there is always another hypothesis to test, another angle to explore, another simplification to try. Algorithm engineering is not a linear process but a continuous spiral of refinement. diff --git a/claude-plugin/skills/aae/knowledge-base/001-blocked-recursive-sorting.md b/claude-plugin/skills/aae/knowledge-base/001-blocked-recursive-sorting.md new file mode 100644 index 0000000..32dee2c --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/001-blocked-recursive-sorting.md @@ -0,0 +1,55 @@ +# Cache-Friendly Blocked Recursive Sorting + +## Problem + +Sorting large integer arrays (10M+ elements) in C++. The metric was wall-clock time measured via `std::chrono::high_resolution_clock`. The baseline was `std::sort`, which uses introsort internally. + +## What Worked + +Replacing the default quicksort partitioning with a blocked approach that processes elements in cache-line-sized blocks (64 elements per block). Instead of scanning left-to-right and right-to-left with pointer swaps, the algorithm fills two buffers of indices that need swapping, then performs the swaps in bulk. This keeps memory accesses sequential within each block and reduces branch mispredictions because the buffer-fill loop has no data-dependent branches. + +The improvement was ~20% over `std::sort` on uniformly random 64-bit integers (10M elements, averaged over 10 runs). + +## Experiment Data + +| Variant | Time (ms) | Relative | +|---------|-----------|----------| +| std::sort baseline | 842 | 1.00x | +| Blocked partition, block size 32 | 713 | 1.18x | +| Blocked partition, block size 64 | 678 | 1.24x | +| Blocked partition, block size 128 | 691 | 1.22x | + +Block size 64 was the sweet spot, matching the L1 cache line width. + +## What Didn't Work + +- **Radix sort**: Faster in theory for integers, but the extra memory allocation and multiple passes made it slower for arrays that fit in L2 cache. Only won for 100M+ elements. +- **SIMD-based partitioning**: The overhead of gather/scatter instructions on the test hardware (Zen 3) negated the theoretical throughput gain. Might work better on Intel with AVX-512. + +## Code Example + +```cpp +// Core idea: buffer indices that need swapping, then swap in bulk +int buf_l[BLOCK], buf_r[BLOCK]; +int nl = 0, nr = 0; +while (left + BLOCK <= right - BLOCK) { + if (nl == 0) { + for (int i = 0; i < BLOCK; i++) + buf_l[nl] = i + (arr[left + i] >= pivot); // branchless + nl += (arr[left + i] >= pivot); + } + if (nr == 0) { + for (int i = 0; i < BLOCK; i++) + buf_r[nr] = i + (arr[right - 1 - i] < pivot); + nr += (arr[right - 1 - i] < pivot); + } + int swaps = min(nl, nr); + for (int i = 0; i < swaps; i++) + std::swap(arr[left + buf_l[i]], arr[right - 1 - buf_r[i]]); + // advance pointers... +} +``` + +## Environment + +C++17, GCC 12.2 with `-O3 -march=native`, AMD Ryzen 9 5900X, 64 GB DDR4-3600, Ubuntu 22.04. Array fits entirely in RAM, no disk I/O. diff --git a/claude-plugin/skills/aae/knowledge-base/002-batch-size-tuning-transformer.md b/claude-plugin/skills/aae/knowledge-base/002-batch-size-tuning-transformer.md new file mode 100644 index 0000000..cbb018d --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/002-batch-size-tuning-transformer.md @@ -0,0 +1,31 @@ +# Gradient Accumulation as Batch Size Proxy for Transformer Training + +## Problem + +Training a small transformer language model (125M parameters) on a single GPU with limited VRAM (24 GB). The metric was validation bits-per-byte (val_bpb), lower is better. Larger batch sizes are known to stabilize training, but the model with batch size > 32 caused OOM on the available hardware. + +## What Worked + +Using gradient accumulation to simulate large batch sizes without increasing memory usage. Instead of trying to fit batch size 128 into memory, accumulate gradients over 4 forward passes of batch size 32 before performing one optimizer step. This achieved the same effective batch size 128 while staying within VRAM limits. + +The key insight: gradient accumulation is not just a memory workaround, it is functionally equivalent to a larger batch (assuming no batch normalization). The model trained with accumulation steps=4 achieved identical val_bpb (within noise) to a machine with enough VRAM for true batch size 128, and val_bpb improved from 1.42 (batch 32) to 1.31 (effective batch 128). + +## Experiment Data + +``` +commit metric_value resource_usage status hypothesis +a1b2c3d 1.4200 22.1 keep baseline (batch_size=32, no accumulation) +b2c3d4e 1.3500 22.1 keep accumulation_steps=2 (effective batch 64) will reduce val_bpb by smoothing gradient noise +c3d4e5f 1.3100 22.3 keep accumulation_steps=4 (effective batch 128) will further reduce val_bpb +d4e5f6g 1.3150 22.3 discard accumulation_steps=8 (effective batch 256) will continue the trend +e5f6g7h 1.2950 22.3 keep accumulation_steps=4 with linear LR scaling (4x base LR) following the linear scaling rule +``` + +## What Didn't Work + +- **Effective batch 256** (accumulation_steps=8): No further improvement; the model is small enough that batch 128 already provides sufficient gradient signal. Diminishing returns. +- **Square root LR scaling**: Scaling LR by sqrt(accumulation_steps) instead of linearly underperformed. Linear scaling rule (Goyal et al., 2017) worked better for this model size. + +## Environment + +PyTorch 2.1, single NVIDIA RTX 4090 (24 GB VRAM), 125M parameter GPT-2-style transformer, OpenWebText dataset, trained for 50k steps. diff --git a/claude-plugin/skills/aae/knowledge-base/003-arena-allocation-json-parsing.md b/claude-plugin/skills/aae/knowledge-base/003-arena-allocation-json-parsing.md new file mode 100644 index 0000000..16c1a87 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/003-arena-allocation-json-parsing.md @@ -0,0 +1,15 @@ +# Arena Allocation for JSON Parsing + +## Problem + +Parsing large JSON files (100 MB+) in a Rust service. The metric was throughput in MB/s. The baseline used `serde_json::from_str` which allocates individually for every string and object in the JSON tree, causing heavy allocator pressure. + +## What Worked + +Switching from per-node heap allocation to an arena allocator (`bumpalo`). All parsed nodes, strings, and intermediate structures are allocated from a single contiguous memory region that is freed in one shot after processing. This eliminated thousands of individual `malloc`/`free` calls per parse and improved cache locality because related nodes end up adjacent in memory. + +Throughput improved from 380 MB/s to 620 MB/s (~63% improvement) on a 150 MB JSON file with deeply nested objects. + +## Environment + +Rust 1.74, `bumpalo` 3.14, AMD EPYC 7763 (server), 256 GB RAM, Linux 6.1. Input files are API response dumps with 3-5 levels of nesting and many small string values. diff --git a/claude-plugin/skills/aae/knowledge-base/004-soa-vs-aos-cache-efficiency.md b/claude-plugin/skills/aae/knowledge-base/004-soa-vs-aos-cache-efficiency.md new file mode 100644 index 0000000..ffed8ca --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/004-soa-vs-aos-cache-efficiency.md @@ -0,0 +1,46 @@ +# Structure of Arrays vs Array of Structures for Cache Efficiency + +## Problem + +Processing large collections of objects (1M+ entities) in C++ where each operation touches only a subset of fields. The metric was throughput (operations per second). Baseline used a traditional AoS layout (`std::vector` where `Particle` has position, velocity, mass, color, etc.) but the hot loop only read `x, y, z` positions. + +## What Worked + +Converting the data layout from Array of Structures (AoS) to Structure of Arrays (SoA). Instead of one `struct` with all fields packed together, store each field in its own contiguous array. This ensures that when the hot loop iterates over positions, every cache line contains only position data — no wasted bandwidth loading unused fields like color or material ID. + +For a particle simulation update loop touching only `x, y, z` on 4M particles, SoA achieved 2.1x throughput over AoS because each 64-byte cache line carried 8 useful `double` values instead of 2 (the struct was 48 bytes, so only 1.3 structs fit per line, wasting ~60% of fetched memory bandwidth on cold fields). + +## Experiment Data + +| Layout | Throughput (Mops/s) | L1 Cache Miss Rate | +|--------|--------------------|--------------------| +| AoS (baseline) | 142 | 12.3% | +| SoA | 298 | 3.1% | +| SoA + AVX2 vectorization | 487 | 2.8% | + +## What Didn't Work + +- **Hybrid AoSoA** (blocking 8 structs into a mini-array): Marginal improvement over AoS (~15%) but added complexity. Only useful when multiple fields are always accessed together. +- **`__attribute__((packed))`** on the struct: Reduced struct size but caused unaligned access penalties that negated the benefit. + +## Code Example + +```cpp +// AoS — cold fields pollute cache lines +struct Particle { double x, y, z, vx, vy, vz; int material; float color[4]; }; +std::vector particles(N); +for (auto& p : particles) p.x += p.vx * dt; // loads 80 bytes per particle, uses 16 + +// SoA — only position data in cache +struct Particles { + std::vector x, y, z, vx, vy, vz; + std::vector material; + std::vector> color; +}; +Particles ps(N); +for (size_t i = 0; i < N; i++) ps.x[i] += ps.vx[i] * dt; // loads 16 bytes, uses 16 +``` + +## Environment + +C++17, GCC 13.1 with `-O3 -march=native`, Intel Core i7-13700K, 32 GB DDR5-5600, Ubuntu 23.04. Measured with `perf stat` for cache miss rates. diff --git a/claude-plugin/skills/aae/knowledge-base/005-branchless-programming.md b/claude-plugin/skills/aae/knowledge-base/005-branchless-programming.md new file mode 100644 index 0000000..8bb7b22 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/005-branchless-programming.md @@ -0,0 +1,51 @@ +# Branchless Programming to Eliminate Branch Mispredictions + +## Problem + +Filtering or transforming large arrays in C++ where element-wise decisions depend on data values (e.g., clamping, conditional increment, partitioning). The metric was wall-clock time. Baseline used straightforward `if/else` branches, which suffer from branch misprediction when data is random or has low predictability. + +## What Worked + +Replacing conditional branches with arithmetic equivalents. Modern CPUs pipeline ~15-20 instructions ahead; a mispredicted branch flushes the entire pipeline (~12-15 cycle penalty on x86). For random data the mispredict rate approaches 50%, making branches extremely expensive in tight loops. + +Key patterns: +1. **Conditional move via arithmetic**: `x = (cond) ? a : b` → compiler emits `cmov` if written carefully, but sometimes needs manual help. +2. **Branchless min/max**: Use subtraction + sign-bit masking instead of `std::min`/`std::max` when the compiler doesn't optimize. +3. **Predicated accumulation**: `sum += arr[i] * (arr[i] > threshold)` — the boolean converts to 0/1, no branch needed. + +For a partitioning loop on 10M random integers, branchless code ran 2.8x faster than the branching version due to eliminating ~47% misprediction rate. + +## Experiment Data + +| Variant | Time (ms) | Branch Misses (perf) | +|---------|-----------|---------------------| +| if/else partition | 38.2 | 24.7M | +| Branchless (arithmetic) | 13.6 | 0.02M | +| Branchless + unrolled 4x | 11.1 | 0.01M | + +## What Didn't Work + +- **Branchless on already-sorted data**: When branch prediction accuracy is >95% (sorted/nearly-sorted input), the branch version is actually faster because `cmov` has a data dependency that serializes execution, while a correctly predicted branch allows speculative execution. Always profile with realistic data distributions. + +## Code Example + +```cpp +// Branching — ~50% mispredict on random data +for (int i = 0; i < N; i++) { + if (arr[i] < pivot) out[left++] = arr[i]; + else out[right--] = arr[i]; +} + +// Branchless — zero mispredictions +for (int i = 0; i < N; i++) { + int goes_left = (arr[i] < pivot); // 0 or 1 + out[left] = arr[i]; + out[right] = arr[i]; + left += goes_left; + right -= (1 - goes_left); +} +``` + +## Environment + +C++17, GCC 12.2 with `-O3 -march=native`, AMD Ryzen 9 5900X, measured with `perf stat -e branch-misses`. diff --git a/claude-plugin/skills/aae/knowledge-base/006-small-buffer-optimization.md b/claude-plugin/skills/aae/knowledge-base/006-small-buffer-optimization.md new file mode 100644 index 0000000..401ecea --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/006-small-buffer-optimization.md @@ -0,0 +1,43 @@ +# Small Buffer Optimization to Avoid Heap Allocations + +## Problem + +High-frequency allocation and deallocation of small, variable-sized objects in C++ (e.g., short strings, small vectors, temporary buffers). The metric was throughput of an operation that creates and destroys thousands of small containers per call. Baseline used `std::string` and `std::vector` which heap-allocate even for tiny sizes, causing allocator contention and cache misses from pointer chasing. + +## What Worked + +Small Buffer Optimization (SBO) embeds a fixed-size buffer directly inside the object. If the data fits in the inline buffer, no heap allocation occurs. This is the technique behind `std::string`'s SSO (Short String Optimization, typically 15-22 bytes inline depending on implementation) and can be applied to any container. + +For a tokenizer producing millions of short tokens (avg. 8 bytes), switching from `std::string` to a custom SBO string with 32-byte inline buffer eliminated 94% of heap allocations and improved throughput by 1.7x. The key insight: even when SSO is already active in `std::string`, the default inline size (typically 15 bytes on libstdc++) may be too small for the workload. A custom SBO with a tuned buffer size can outperform. + +## Experiment Data + +| Variant | Tokens/sec (M) | Heap Allocs (millions) | +|---------|----------------|----------------------| +| std::string (SSO=15) | 4.2 | 3.1 | +| SBO string (inline=32) | 7.1 | 0.19 | +| SBO string (inline=64) | 6.8 | 0.04 | + +Inline 64 had slightly lower throughput than 32 because the larger object size reduced cache density for the token array itself. + +## Code Example + +```cpp +template +class SmallString { + size_t size_; + union { + char inline_buf_[InlineSize]; + char* heap_ptr_; + }; + bool is_heap() const { return size_ >= InlineSize; } +public: + char* data() { return is_heap() ? heap_ptr_ : inline_buf_; } + // ... constructor allocates on heap only when size >= InlineSize +}; +// Same pattern applies to SmallVector: inline storage for N elements +``` + +## Environment + +C++17, GCC 13.1 with `-O3`, libstdc++, Intel Core i7-12700K, Ubuntu 22.04. diff --git a/claude-plugin/skills/aae/knowledge-base/007-move-semantics-avoiding-copies.md b/claude-plugin/skills/aae/knowledge-base/007-move-semantics-avoiding-copies.md new file mode 100644 index 0000000..f69819c --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/007-move-semantics-avoiding-copies.md @@ -0,0 +1,49 @@ +# Move Semantics to Eliminate Expensive Deep Copies + +## Problem + +C++ programs that return or transfer ownership of large containers (vectors, strings, maps) from functions, causing unnecessary deep copies. The metric was function call overhead for routines returning `std::vector` with 1M+ elements. + +## What Worked + +Three techniques at different levels: + +1. **Return value optimization (RVO/NRVO)**: The compiler elides the copy entirely when a function returns a local variable by name. This is guaranteed in C++17 for prvalues (copy elision). Ensure the function has a single, named return variable — multiple return paths can prevent NRVO. + +2. **`std::move` for transfers**: When moving an object to a new owner (e.g., pushing into a container or passing to a constructor), `std::move` converts the lvalue to an rvalue reference, triggering the move constructor. A moved `std::vector` transfers pointer ownership in O(1) instead of copying N elements. + +3. **Emplace instead of insert**: `container.emplace_back(args...)` constructs the object in-place, avoiding both copy and move. Particularly impactful when the object is expensive to construct. + +For a pipeline that passed `vector(1M)` through 5 transformation stages, ensuring moves instead of copies reduced stage-transition overhead from 4.2ms to <0.001ms per transfer. + +## What Didn't Work + +- **Moving from `const` references**: `std::move(const_ref)` silently falls back to a copy because the move constructor requires a non-const rvalue reference. This is a common silent performance bug — no compiler warning by default. +- **Moving small types**: For types smaller than two pointers (e.g., `std::pair`), move is identical to copy. The overhead of thinking about moves is wasted. + +## Code Example + +```cpp +// BAD: forces copy if NRVO fails (multiple return paths) +std::vector compute(bool flag) { + std::vector a = heavy_compute_a(); + std::vector b = heavy_compute_b(); + if (flag) return a; // NRVO may fail — two candidates + return b; +} + +// GOOD: single return variable, guaranteed NRVO +std::vector compute(bool flag) { + std::vector result; + if (flag) result = heavy_compute_a(); + else result = heavy_compute_b(); + return result; // NRVO: single named variable returned +} + +// Transfer ownership explicitly +pipeline.add_stage(std::move(large_vector)); // O(1) pointer swap +``` + +## Environment + +C++17 or later (mandatory copy elision for prvalues). Applies to all major compilers (GCC, Clang, MSVC). diff --git a/claude-plugin/skills/aae/knowledge-base/008-constexpr-compile-time-computation.md b/claude-plugin/skills/aae/knowledge-base/008-constexpr-compile-time-computation.md new file mode 100644 index 0000000..e432efc --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/008-constexpr-compile-time-computation.md @@ -0,0 +1,42 @@ +# constexpr for Compile-Time Computation + +## Problem + +C++ programs that repeatedly compute fixed lookup tables, mathematical constants, or configuration-derived values at runtime. The metric was startup latency and hot-loop throughput. Baseline computed sine/cosine tables, CRC tables, and hash seeds on every program start or once-per-call with static locals. + +## What Worked + +Moving deterministic computations to compile time using `constexpr` (and `consteval` in C++20). The compiler evaluates the expressions during compilation, embedding the results directly into the binary as constants. This eliminates: +- Runtime initialization cost (especially for large tables) +- Branch/guard overhead for lazy static initialization +- Cache misses from touching cold memory during init + +A 256-entry CRC32 lookup table computed at compile time saved 1.2μs of startup time per instantiation and allowed the table to live in `.rodata` (read-only, shareable across processes). For a packet processing loop using this table, throughput improved 8% because the optimizer could see the table contents and optimize access patterns. + +## Code Example + +```cpp +// C++17: constexpr lookup table generation +constexpr std::array make_crc_table() { + std::array table{}; + for (uint32_t i = 0; i < 256; i++) { + uint32_t crc = i; + for (int j = 0; j < 8; j++) + crc = (crc >> 1) ^ (0xEDB88320 & (-(crc & 1))); + table[i] = crc; + } + return table; +} +constexpr auto crc_table = make_crc_table(); // computed at compile time + +// C++20: consteval guarantees compile-time evaluation +consteval auto make_sin_table() { /* ... */ } +``` + +## What Didn't Work + +- **Very large constexpr tables** (>64KB): Some compilers hit constexpr evaluation step limits or produce enormous compile times. GCC's `-fconstexpr-ops-limit` may need increasing. For huge tables, code generation (offline script writing a `.cpp` file) is more practical. + +## Environment + +C++17 minimum (`constexpr` functions), C++20 for `consteval`. GCC 12+, Clang 15+, MSVC 19.30+. diff --git a/claude-plugin/skills/aae/knowledge-base/009-false-sharing-avoidance.md b/claude-plugin/skills/aae/knowledge-base/009-false-sharing-avoidance.md new file mode 100644 index 0000000..6099e2a --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/009-false-sharing-avoidance.md @@ -0,0 +1,50 @@ +# Avoiding False Sharing in Multithreaded Code + +## Problem + +Multithreaded C++ program with per-thread counters/accumulators showed poor scaling beyond 2 threads. Expected 4x speedup on 4 cores but observed only 1.3x. The metric was parallel speedup vs single-threaded baseline. Profiling with `perf c2c` revealed heavy cross-core cache line contention despite threads writing to independent variables. + +## What Worked + +False sharing occurs when independent variables used by different threads happen to reside on the same cache line (64 bytes on x86). When any thread writes to the line, the cache coherency protocol (MESI) invalidates the line in all other cores, forcing expensive cache-to-cache transfers (~50-100 cycles on modern hardware vs. ~4 cycles for L1 hit). + +The fix: pad per-thread data to cache line boundaries using `alignas(64)` or C++17's `std::hardware_destructive_interference_size`. + +After padding, the 4-thread parallel sum achieved 3.82x speedup (vs 1.3x before), matching the theoretical near-linear scaling. + +## Experiment Data + +| Threads | Speedup (packed) | Speedup (padded) | +|---------|------------------|------------------| +| 1 | 1.00x | 1.00x | +| 2 | 1.21x | 1.97x | +| 4 | 1.31x | 3.82x | +| 8 | 1.28x | 7.41x | + +## Code Example + +```cpp +// BAD: per-thread counters packed together — false sharing +struct Counters { int64_t count[NUM_THREADS]; }; // all on same/adjacent cache lines + +// GOOD: each counter on its own cache line +struct alignas(64) PaddedCounter { int64_t count; }; +PaddedCounter counters[NUM_THREADS]; // each on separate cache line + +// C++17 portable version +struct alignas(std::hardware_destructive_interference_size) PaddedCounter { + int64_t count; +}; + +// Alternative: thread-local accumulation + final reduction +thread_local int64_t local_count = 0; +// ... each thread accumulates locally, then merge once at the end +``` + +## What Didn't Work + +- **Over-padding with page alignment** (4096 bytes): Wasted too much memory and caused TLB pressure, actually hurting performance for >32 threads. Cache line alignment (64 bytes) is the sweet spot. + +## Environment + +C++17, GCC 12.2, AMD EPYC 7763 (64 cores), Linux 6.1. Diagnosed with `perf c2c record/report`. diff --git a/claude-plugin/skills/aae/knowledge-base/010-pgo-lto-compiler-optimizations.md b/claude-plugin/skills/aae/knowledge-base/010-pgo-lto-compiler-optimizations.md new file mode 100644 index 0000000..138db71 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/010-pgo-lto-compiler-optimizations.md @@ -0,0 +1,51 @@ +# Profile-Guided Optimization and Link-Time Optimization + +## Problem + +C++ application with complex control flow (many branches, virtual calls, deep call trees) where `-O3` alone leaves significant performance on the table. The metric was end-to-end throughput of a compiler-like workload (parsing + optimization + code generation). The optimizer makes guesses about branch probabilities and inlining without runtime data, often getting it wrong. + +## What Worked + +**PGO (Profile-Guided Optimization)** feeds actual runtime profiling data back into the compiler, enabling: +- Accurate branch probability annotations (hot paths get fall-through layout) +- Informed inlining decisions (inline functions on hot paths, skip cold ones) +- Hot/cold code splitting (frequently executed code packed together for better I-cache utilization) +- Better register allocation along hot paths + +**LTO (Link-Time Optimization)** performs whole-program optimization across translation units, enabling cross-module inlining, dead code elimination, and interprocedural constant propagation. + +Combined PGO+LTO achieved 22% throughput improvement on a real-world workload. PGO alone gave ~15%, LTO alone ~8%, but they compound because LTO exposes more inlining opportunities for PGO-guided decisions. + +## Experiment Data + +| Configuration | Throughput (ops/s) | Binary Size | +|--------------|-------------------|-------------| +| -O3 baseline | 1,000 | 12.1 MB | +| -O3 + LTO | 1,082 | 10.8 MB | +| -O3 + PGO | 1,148 | 12.4 MB | +| -O3 + PGO + LTO | 1,221 | 11.2 MB | + +## Code Example + +```bash +# GCC PGO workflow (three-step): +# 1. Build instrumented binary +g++ -O3 -fprofile-generate=./profdata -flto -o app_instrumented *.cpp + +# 2. Run with representative workload to collect profile +./app_instrumented < representative_input.txt + +# 3. Rebuild using profile data +g++ -O3 -fprofile-use=./profdata -flto -o app_optimized *.cpp + +# Clang uses -fprofile-instr-generate / -fprofile-instr-use instead +``` + +## What Didn't Work + +- **Non-representative training data**: PGO with synthetic benchmarks that don't match production traffic led to *worse* performance than baseline (-3%) because the optimizer optimized for the wrong hot paths. The training workload must closely match production. +- **PGO on very small programs**: The overhead of instrumentation and the three-step build process isn't worth it for programs under ~10K lines where `-O3` already does well. + +## Environment + +GCC 13.1 / Clang 17, Linux. PGO is supported by all major compilers. LTO requires all translation units to be compiled with the same compiler. diff --git a/claude-plugin/skills/aae/knowledge-base/011-mmap-large-file-processing.md b/claude-plugin/skills/aae/knowledge-base/011-mmap-large-file-processing.md new file mode 100644 index 0000000..0bdddd6 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/011-mmap-large-file-processing.md @@ -0,0 +1,49 @@ +# Memory-Mapped I/O for Large File Processing + +## Problem + +Processing large files (1GB+) in C++ by reading them into memory. The baseline used `std::ifstream::read()` into a `std::vector`, which required allocating a buffer equal to the file size, waiting for the entire read to complete before processing, and doubling memory usage if the file content needed to persist alongside processed results. + +## What Worked + +Using `mmap()` (POSIX) or `CreateFileMapping` (Windows) to map the file directly into the process's virtual address space. The OS loads pages on demand as they're accessed, and the kernel's page cache serves as a shared buffer — no explicit allocation or copying needed. + +Key advantages: +1. **Zero-copy access**: File data is accessed directly from the kernel page cache. No `read()` syscall per chunk, no userspace buffer. +2. **Lazy loading**: Only pages actually touched are loaded from disk. For sparse access patterns (e.g., searching a large file), this avoids loading irrelevant sections. +3. **Memory efficiency**: Multiple processes mapping the same file share physical pages. The OS can evict pages under memory pressure without the application managing a cache. + +For sequential processing of a 2GB CSV file, `mmap` was 1.4x faster than `fread` with 64KB buffers and used 50% less resident memory (RSS) because untouched pages were never loaded. + +## What Didn't Work + +- **mmap for random small reads on HDD**: On spinning disks, mmap's page-fault-driven I/O generates random seeks per fault. Explicit `read()` with `posix_fadvise(POSIX_FADV_RANDOM)` performed better by issuing larger batched reads. +- **mmap on 32-bit systems**: The 4GB virtual address limit makes mapping large files impossible. Use `read()` with streaming. + +## Code Example + +```cpp +#include +#include + +int fd = open("large_file.csv", O_RDONLY); +struct stat st; +fstat(fd, &st); + +// Map entire file; OS handles paging +const char* data = static_cast( + mmap(nullptr, st.st_size, PROT_READ, MAP_PRIVATE, fd, 0)); +madvise((void*)data, st.st_size, MADV_SEQUENTIAL); // hint for readahead + +// Process directly — no buffer allocation +for (size_t i = 0; i < st.st_size; i++) { + if (data[i] == '\n') { /* process line */ } +} + +munmap((void*)data, st.st_size); +close(fd); +``` + +## Environment + +POSIX systems (Linux, macOS). On Linux, `madvise` hints (`MADV_SEQUENTIAL`, `MADV_WILLNEED`, `MADV_HUGEPAGE`) significantly affect performance. Tested on Linux 6.1, ext4 filesystem, NVMe SSD. diff --git a/claude-plugin/skills/aae/knowledge-base/012-open-addressing-hash-maps.md b/claude-plugin/skills/aae/knowledge-base/012-open-addressing-hash-maps.md new file mode 100644 index 0000000..230f703 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/012-open-addressing-hash-maps.md @@ -0,0 +1,34 @@ +# Open-Addressing Hash Maps vs std::unordered_map + +## Problem + +Lookup-heavy workloads in C++ where `std::unordered_map` is the bottleneck. The metric was lookup throughput (queries per second) for a map with 1M+ entries and random access patterns. `std::unordered_map` uses separate chaining with linked list nodes, causing pointer-chasing cache misses on every lookup. + +## What Worked + +Replacing `std::unordered_map` with an open-addressing hash map that stores entries inline in a flat array. This eliminates pointer chasing — all probing hits contiguous memory, dramatically improving cache utilization. + +Recommended implementations: +- **`absl::flat_hash_map`** (Google): Swiss table design, SIMD-accelerated probing, 1 byte metadata per slot +- **`robin_hood::unordered_map`**: Robin Hood hashing with backward-shift deletion +- **`ankerl::unordered_dense::map`**: Dense storage, very fast for small-to-medium maps + +For 10M random lookups in a 2M-entry map of `int64_t → int64_t`, `absl::flat_hash_map` achieved 3.2x the throughput of `std::unordered_map`. + +## Experiment Data + +| Implementation | Lookup (Mops/s) | Memory (MB) | Insert (Mops/s) | +|---------------|-----------------|-------------|-----------------| +| std::unordered_map | 8.2 | 198 | 5.1 | +| absl::flat_hash_map | 26.4 | 89 | 14.7 | +| robin_hood::unordered_map | 24.1 | 72 | 12.3 | +| ankerl::unordered_dense | 22.8 | 68 | 15.9 | + +## What Didn't Work + +- **Open addressing with high load factor** (>0.9): Probe chains become long, degrading lookup to near-linear scan. Keep load factor at 0.5-0.7 for open addressing (vs. 1.0+ acceptable for chaining). +- **Flat maps for large values**: If the value type is large (>128 bytes), the flat layout wastes cache bandwidth probing past cold data. Use `absl::node_hash_map` (open addressing for metadata, nodes for values) for large value types. + +## Environment + +C++17, GCC 12.2, `-O3`. absl from Abseil 20230802, robin_hood v3.11.5. Benchmark: uniform random keys, int64_t → int64_t. AMD Ryzen 9 7950X. diff --git a/claude-plugin/skills/aae/knowledge-base/013-simd-vectorization-batch-ops.md b/claude-plugin/skills/aae/knowledge-base/013-simd-vectorization-batch-ops.md new file mode 100644 index 0000000..fa7fce1 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/013-simd-vectorization-batch-ops.md @@ -0,0 +1,59 @@ +# SIMD Vectorization for Batch Array Operations + +## Problem + +Processing large arrays of floats/doubles element-wise in C++ (e.g., scaling, dot products, distance computations). The metric was throughput (elements/second). The compiler auto-vectorizes simple loops but fails on anything with conditionals, non-trivial reductions, or non-contiguous access patterns. + +## What Worked + +Explicit SIMD intrinsics (SSE/AVX2/AVX-512) to process 4-16 elements per instruction. The key insight is that modern CPUs have 256-bit (AVX2) or 512-bit (AVX-512) SIMD registers that can operate on 8 floats or 4 doubles simultaneously, but the compiler often can't prove it's safe to vectorize due to aliasing, alignment, or control flow. + +Three-tier approach: +1. **Help auto-vectorization first**: Use `__restrict__`, `#pragma omp simd`, ensure aligned allocations, avoid loop-carried dependencies. +2. **Intrinsics for critical kernels**: Write SSE/AVX2 intrinsics for the top 2-3 hotspot loops identified by profiling. +3. **Use a SIMD wrapper library** (Highway, xsimd, Vc) for portable SIMD across architectures. + +For a Euclidean distance matrix computation (10K × 10K, float32), AVX2 intrinsics achieved 6.2x speedup over scalar code and 2.1x over auto-vectorized code (the compiler failed to vectorize the inner reduction). + +## Experiment Data + +| Variant | Throughput (Gflops) | vs Scalar | +|---------|-------------------|-----------| +| Scalar `-O3` | 2.1 | 1.0x | +| Auto-vectorized (`-O3 -march=native`) | 4.8 | 2.3x | +| AVX2 intrinsics | 13.0 | 6.2x | +| AVX-512 intrinsics | 18.7 | 8.9x | + +## Code Example + +```cpp +#include + +// AVX2: process 8 floats at a time +float dot_product_avx2(const float* a, const float* b, size_t n) { + __m256 sum = _mm256_setzero_ps(); + size_t i = 0; + for (; i + 8 <= n; i += 8) { + __m256 va = _mm256_loadu_ps(a + i); + __m256 vb = _mm256_loadu_ps(b + i); + sum = _mm256_fmadd_ps(va, vb, sum); // fused multiply-add + } + // Horizontal sum of 8 floats in sum register + __m128 hi = _mm256_extractf128_ps(sum, 1); + __m128 lo = _mm256_castps256_ps128(sum); + __m128 s = _mm_add_ps(lo, hi); + s = _mm_hadd_ps(s, s); + s = _mm_hadd_ps(s, s); + float result = _mm_cvtss_f32(s); + for (; i < n; i++) result += a[i] * b[i]; // scalar tail + return result; +} +``` + +## What Didn't Work + +- **AVX-512 on consumer CPUs (Intel 12th-13th gen)**: The CPU downclocks significantly when AVX-512 is used, sometimes making it slower than AVX2 for short bursts. Only beneficial for sustained computation on server chips (Xeon, EPYC). + +## Environment + +C++17, GCC 13.1, `-O3 -mavx2 -mfma`. Intel Xeon w5-3435X for AVX-512 results. Use `lscpu` to verify SIMD support before deploying. diff --git a/claude-plugin/skills/aae/knowledge-base/014-loop-tiling-matrix-operations.md b/claude-plugin/skills/aae/knowledge-base/014-loop-tiling-matrix-operations.md new file mode 100644 index 0000000..bb29641 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/014-loop-tiling-matrix-operations.md @@ -0,0 +1,47 @@ +# Loop Tiling for Cache-Efficient Matrix Operations + +## Problem + +Matrix multiplication or similar O(n³) nested-loop computations in C++ where the naive triple loop causes massive cache thrashing. For matrix multiply `C = A × B` with N=2048, the naive `ijk` loop order causes column-wise access to `B`, missing the cache on every access since rows are stored contiguously (row-major). + +## What Worked + +**Loop tiling** (also called loop blocking) partitions the iteration space into small tiles that fit in L1/L2 cache. Instead of processing entire rows/columns, the algorithm works on `T×T` sub-matrices that remain cache-resident throughout computation. + +The optimal tile size `T` depends on cache size: `3 × T² × sizeof(element) ≤ L1_cache_size` (three sub-matrices of A, B, C must fit simultaneously). For 32KB L1 with `double`, T≈36; in practice, T=32 or T=64 works well due to alignment. + +For N=2048 double-precision matrix multiply, tiled (T=64) was 4.8x faster than naive `ijk` and 2.1x faster than simply reordering loops to `ikj` (which fixes B's access pattern but still thrashes for large N). + +## Experiment Data + +| Variant | Time (s) | GFlops | +|---------|----------|--------| +| Naive ijk | 28.4 | 0.60 | +| Reordered ikj | 11.2 | 1.53 | +| Tiled T=32 | 6.8 | 2.52 | +| Tiled T=64 | 5.9 | 2.91 | +| OpenBLAS dgemm | 0.9 | 19.1 | + +OpenBLAS is still 6.5x faster due to AVX2 + micro-kernel + multi-level tiling, but loop tiling alone captured most of the cache efficiency gain. + +## Code Example + +```cpp +constexpr int T = 64; // tile size +for (int ii = 0; ii < N; ii += T) + for (int jj = 0; jj < N; jj += T) + for (int kk = 0; kk < N; kk += T) + for (int i = ii; i < std::min(ii+T, N); i++) + for (int k = kk; k < std::min(kk+T, N); k++) + for (int j = jj; j < std::min(jj+T, N); j++) + C[i][j] += A[i][k] * B[k][j]; +``` + +## What Didn't Work + +- **Tile size = cache line (T=8)**: Too small — the overhead of tile boundary logic dominates. Tiles should be sized to fit in L1 cache, not matched to cache line width. +- **Auto-tuning without pinning frequency**: CPU frequency scaling caused inconsistent results during tile size search. Always pin frequency during benchmarking (`cpupower frequency-set -g performance`). + +## Environment + +C++17, GCC 12.2, `-O3 -march=native`, double-precision (8 bytes). Intel Core i7-12700K, 48KB L1d, 1.25MB L2 per P-core. diff --git a/claude-plugin/skills/aae/knowledge-base/015-builtin-intrinsics-bit-operations.md b/claude-plugin/skills/aae/knowledge-base/015-builtin-intrinsics-bit-operations.md new file mode 100644 index 0000000..c3e8aba --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/015-builtin-intrinsics-bit-operations.md @@ -0,0 +1,47 @@ +# Compiler Built-in Intrinsics for Bit Operations + +## Problem + +Algorithms that frequently count set bits (popcount), find leading/trailing zeros (clz/ctz), or perform other bit manipulations in C++. The metric was throughput for operations like Hamming distance, bit-parallel set operations, or bitmap index scans. Baseline used manual loops to count bits. + +## What Worked + +Using compiler built-ins (`__builtin_popcount`, `__builtin_clz`, `__builtin_ctz`) and C++20 `` header functions (`std::popcount`, `std::countl_zero`, `std::countr_zero`). These compile to single hardware instructions (x86: `POPCNT`, `LZCNT`, `TZCNT`) that execute in 1 cycle with 1-cycle latency, replacing 10-20 instruction software emulations. + +For a Hamming distance computation over 1M 64-bit words, using `__builtin_popcountll` was 8.3x faster than a naive bit-counting loop and 2.1x faster than the Kernighan bit-counting trick (`n &= n - 1`). + +## Experiment Data + +| Method | Throughput (Gwords/s) | Cycles/word | +|--------|----------------------|-------------| +| Naive loop (check each bit) | 0.12 | ~40 | +| Kernighan trick | 0.47 | ~10 | +| `__builtin_popcountll` | 0.98 | ~5 | +| POPCNT + AVX2 (batch of 4) | 3.41 | ~1.4 | + +## Code Example + +```cpp +// Hamming distance: count differing bits between two arrays +uint64_t hamming(const uint64_t* a, const uint64_t* b, size_t n) { + uint64_t dist = 0; + for (size_t i = 0; i < n; i++) + dist += __builtin_popcountll(a[i] ^ b[i]); // single POPCNT instruction + return dist; +} + +// Find lowest set bit position (useful for bitmap scanning) +int next_set_bit(uint64_t mask) { + return __builtin_ctzll(mask); // single TZCNT instruction +} + +// C++20 portable equivalents +#include +int count = std::popcount(x); // popcount +int lz = std::countl_zero(x); // clz +int tz = std::countr_zero(x); // ctz +``` + +## Environment + +Requires hardware support: x86 with `POPCNT` (since Nehalem/Barcelona, ~2008), `LZCNT`/`TZCNT` (since Haswell/Piledriver). Compile with `-mpopcnt` or `-march=native`. C++20 `` header for portable code. diff --git a/claude-plugin/skills/aae/knowledge-base/016-reserve-preallocate-containers.md b/claude-plugin/skills/aae/knowledge-base/016-reserve-preallocate-containers.md new file mode 100644 index 0000000..90f32af --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/016-reserve-preallocate-containers.md @@ -0,0 +1,47 @@ +# Reserve and Preallocate STL Containers + +## Problem + +C++ programs that build up `std::vector`, `std::string`, or `std::unordered_map` incrementally without knowing the final size upfront. The metric was insertion throughput and peak memory usage. The default growth strategy (typically 2x reallocation) causes O(log N) reallocations, each copying all existing elements, and temporarily doubles memory usage during reallocation. + +## What Worked + +Calling `.reserve(N)` before bulk insertion when the final size is known or can be estimated. This performs a single allocation upfront, eliminating all intermediate reallocations and copies. For `std::vector` with 10M `push_back` calls, reserving upfront was 2.4x faster and used 33% less peak memory. + +Key rules: +1. **`vector::reserve(N)`**: Pre-allocates storage for N elements. Size stays 0; capacity becomes N. +2. **`string::reserve(N)`**: Same for strings. Essential before repeated `+=` concatenation. +3. **`unordered_map::reserve(N)`**: Pre-allocates buckets for N elements, preventing rehashing. Rehashing at high load is very expensive (reconstructs all hash chains). +4. **`resize` vs `reserve`**: Use `resize` when you need default-initialized elements to index into; use `reserve` when you'll `push_back`. + +## Experiment Data + +| Container | Operation | Without reserve | With reserve | Speedup | +|-----------|-----------|----------------|-------------|---------| +| vector 10M | push_back | 42ms | 17ms | 2.4x | +| string 100MB | += chars | 320ms | 95ms | 3.4x | +| unordered_map 1M | insert | 890ms | 410ms | 2.2x | + +## Code Example + +```cpp +// BAD: 24 reallocations for 10M elements (log2(10M) ≈ 24) +std::vector v; +for (int i = 0; i < 10'000'000; i++) v.push_back(i); + +// GOOD: single allocation +std::vector v; +v.reserve(10'000'000); +for (int i = 0; i < 10'000'000; i++) v.push_back(i); + +// For maps: prevent rehashing +std::unordered_map map; +map.reserve(expected_entries); // allocates buckets for expected_entries + +// When size is approximate: over-estimate by 10-20% rather than under-estimate +v.reserve(estimated_size * 1.1); +``` + +## Environment + +Applies to all C++ standard library implementations (libstdc++, libc++, MSVC STL). The growth factor varies (GCC: 2x, MSVC: 1.5x), but `reserve` bypasses it entirely. diff --git a/claude-plugin/skills/aae/knowledge-base/017-hot-cold-data-splitting.md b/claude-plugin/skills/aae/knowledge-base/017-hot-cold-data-splitting.md new file mode 100644 index 0000000..0cd0bd1 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/017-hot-cold-data-splitting.md @@ -0,0 +1,56 @@ +# Hot-Cold Data Splitting for Cache Efficiency + +## Problem + +C++ data structures where each object has frequently-accessed "hot" fields (e.g., key, pointer, status flag) and rarely-accessed "cold" fields (e.g., debug info, statistics, metadata). The metric was lookup throughput in a container of 1M+ objects. Iterating over the container loaded cold fields into cache, evicting useful data. + +## What Worked + +Splitting each struct into a hot part (stored contiguously in the main array) and a cold part (stored in a parallel array, accessed by index only when needed). This increases the density of useful data per cache line, allowing more hot records to fit in cache simultaneously. + +For a routing table with 2M entries where lookups touch only `prefix` and `next_hop` (16 bytes) but the full entry is 128 bytes (includes statistics, timestamps, debug strings), splitting hot/cold improved lookup throughput by 3.1x because 4 hot records fit per cache line instead of 0.5 full records. + +## Experiment Data + +| Layout | Lookup (Mops/s) | L1 Miss Rate | +|--------|-----------------|-------------| +| Monolithic struct (128B) | 12.4 | 18.2% | +| Hot/cold split (16B hot) | 38.5 | 4.7% | +| Hot/cold + prefetch | 42.1 | 3.1% | + +## Code Example + +```cpp +// BEFORE: monolithic — cold fields pollute cache +struct RouteEntry { + uint32_t prefix; // hot: always read + uint32_t next_hop; // hot: always read + uint64_t packet_count; // cold: only for stats + uint64_t byte_count; // cold: only for stats + time_t last_update; // cold: only for monitoring + char description[80]; // cold: only for debugging +}; +std::vector table; // 128 bytes per entry + +// AFTER: split hot and cold +struct RouteHot { + uint32_t prefix; + uint32_t next_hop; +}; +struct RouteCold { + uint64_t packet_count; + uint64_t byte_count; + time_t last_update; + char description[80]; +}; +std::vector table_hot; // 8 bytes per entry — dense +std::vector table_cold; // indexed in parallel, accessed rarely +``` + +## What Didn't Work + +- **Splitting when most operations touch all fields**: If >50% of accesses need both hot and cold data, the extra indirection (two memory accesses instead of one) makes it slower. Profile access patterns before splitting. + +## Environment + +C++17, GCC 13.1, `-O3 -march=native`. AMD EPYC 7763. Verified with `perf stat -e L1-dcache-load-misses`. diff --git a/claude-plugin/skills/aae/knowledge-base/018-string-view-avoiding-copies.md b/claude-plugin/skills/aae/knowledge-base/018-string-view-avoiding-copies.md new file mode 100644 index 0000000..1a2c951 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/018-string-view-avoiding-copies.md @@ -0,0 +1,49 @@ +# std::string_view to Eliminate String Copies + +## Problem + +C++ code that passes strings through multiple function layers (parsing, validation, transformation) where each function takes `const std::string&` or `std::string`. Even with `const&`, callers with string literals or substrings must construct a temporary `std::string`, triggering heap allocation. The metric was throughput of a text parsing pipeline processing millions of short strings. + +## What Worked + +Replacing `const std::string&` parameters with `std::string_view` in functions that only read (never store or modify) the string. `string_view` is a non-owning (pointer, length) pair — 16 bytes, always on the stack, zero allocation. It binds to `std::string`, `const char*`, string literals, and substrings without any copy or allocation. + +For a CSV parser calling 6 functions per field (validate, trim, classify, transform, hash, store), switching the first 5 functions from `const string&` to `string_view` eliminated 5 temporary `string` allocations per field. On a 500MB CSV with 50M fields, parsing throughput improved 1.9x. + +## Experiment Data + +| Parameter Type | Parse Throughput (MB/s) | Heap Allocs/field | +|---------------|------------------------|-------------------| +| const string& | 180 | 6 | +| string_view (read path) | 340 | 1 (final store only) | + +## Code Example + +```cpp +// BEFORE: every call site may allocate +bool validate(const std::string& field); +std::string trim(const std::string& field); + +// AFTER: zero-allocation read path +bool validate(std::string_view field); +std::string_view trim(std::string_view field); // return view into original buffer + +// Substring without allocation: +std::string_view line = "name,age,city"; +auto field = line.substr(5, 3); // "age" — no allocation, just pointer arithmetic + +// DANGER: string_view must not outlive the data it points to +std::string_view dangling() { + std::string temp = compute(); + return temp; // BUG: temp destroyed, view dangles +} +``` + +## What Didn't Work + +- **Replacing parameters that store the string**: Functions that save the string into a member variable or container still need `std::string` (ownership). Using `string_view` there requires an explicit conversion (`std::string(sv)`), which allocates anyway — but now the allocation is hidden and easy to miss. +- **string_view for null-terminated C APIs**: `string_view` is not null-terminated. Passing `.data()` to C functions expecting null termination is undefined behavior unless the view happens to point at a null-terminated source. + +## Environment + +C++17 (where `std::string_view` was introduced). Applies to all implementations. GCC/Clang/MSVC all optimize `string_view` to two registers. diff --git a/claude-plugin/skills/aae/knowledge-base/019-numpy-vectorization-over-loops.md b/claude-plugin/skills/aae/knowledge-base/019-numpy-vectorization-over-loops.md new file mode 100644 index 0000000..34708e8 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/019-numpy-vectorization-over-loops.md @@ -0,0 +1,53 @@ +# NumPy Vectorization Over Python Loops + +## Problem + +Python programs processing large numerical arrays element-by-element using `for` loops. The metric was computation time for operations like element-wise math, filtering, and reductions on arrays with 1M+ elements. Python's interpreter overhead (type checks, reference counting, dynamic dispatch per element) makes pure-Python loops 50-100x slower than equivalent C code. + +## What Worked + +Replacing Python loops with NumPy vectorized operations that execute the entire computation in compiled C/Fortran under the hood. The key insight: each NumPy operation processes the whole array in a single call, amortizing Python interpreter overhead across millions of elements. + +For computing Euclidean distances between 1M 3D point pairs, vectorized NumPy was 87x faster than the equivalent Python loop. + +## Experiment Data + +| Approach | Time (s) | Relative | +|----------|----------|----------| +| Python for loop | 12.4 | 1.0x | +| List comprehension | 8.7 | 1.4x | +| NumPy vectorized | 0.14 | 87x | +| NumPy + in-place ops | 0.11 | 113x | + +## What Didn't Work + +- **Chaining many small vectorized ops**: Each NumPy op allocates a temporary array for the result. A chain like `np.sqrt(np.sum((a - b)**2, axis=1))` creates 4 temporaries. For memory-bound workloads, this kills performance. Use `np.einsum` or numexpr to fuse operations, or use in-place operators (`np.subtract(a, b, out=temp)`). +- **Vectorizing inherently sequential algorithms**: Algorithms with loop-carried dependencies (e.g., cumulative state machines) can't be trivially vectorized. Consider `np.cumsum`, `np.cumprod` for simple cases, or Numba for complex ones. + +## Code Example + +```python +import numpy as np + +# BAD: Python loop — 12.4s for 1M points +distances = [] +for i in range(len(a)): + d = math.sqrt((a[i,0]-b[i,0])**2 + (a[i,1]-b[i,1])**2 + (a[i,2]-b[i,2])**2) + distances.append(d) + +# GOOD: NumPy vectorized — 0.14s for 1M points +distances = np.linalg.norm(a - b, axis=1) + +# BETTER: avoid temporary allocation with einsum +diff = a - b +distances = np.sqrt(np.einsum('ij,ij->i', diff, diff)) + +# In-place operations to reduce memory allocation +np.subtract(a, b, out=diff) +np.multiply(diff, diff, out=diff) +distances = np.sqrt(diff.sum(axis=1)) +``` + +## Environment + +Python 3.11, NumPy 1.25. NumPy uses OpenBLAS/MKL for linear algebra and SIMD-optimized ufuncs for element-wise operations. Performance gains scale with array size; for arrays <100 elements, the function call overhead makes loops competitive. diff --git a/claude-plugin/skills/aae/knowledge-base/020-slots-memory-reduction.md b/claude-plugin/skills/aae/knowledge-base/020-slots-memory-reduction.md new file mode 100644 index 0000000..5f70d86 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/020-slots-memory-reduction.md @@ -0,0 +1,56 @@ +# __slots__ for Memory Reduction in Python Classes + +## Problem + +Python application creating millions of small objects (e.g., tree nodes, graph vertices, data records) consuming excessive memory. The metric was peak RSS. Baseline used regular Python classes which allocate a `__dict__` (hash table) per instance to store attributes, consuming ~104 bytes overhead per object even for a class with just two `int` attributes. + +## What Worked + +Declaring `__slots__` on the class eliminates the per-instance `__dict__`, storing attributes in a fixed-size tuple-like structure at the C level. This reduces per-instance memory from ~232 bytes (dict-based) to ~72 bytes for a two-attribute object — a 69% reduction. + +For a graph with 10M nodes (`id: int, weight: float, visited: bool`), `__slots__` reduced memory from 2.9 GB to 0.84 GB. + +## Experiment Data + +| Class Style | Bytes/Instance | 10M Instances (GB) | +|-------------|---------------|-------------------| +| Regular class (with __dict__) | 296 | 2.90 | +| __slots__ class | 88 | 0.84 | +| NamedTuple | 80 | 0.76 | +| dataclass(slots=True) (3.10+) | 88 | 0.84 | + +## What Didn't Work + +- **`__slots__` on classes that need dynamic attributes**: `__slots__` removes `__dict__`, so you can't add attributes at runtime (`obj.new_attr = x` raises `AttributeError`). Adding `'__dict__'` to `__slots__` brings back the dict and defeats the purpose. +- **Forgetting `__slots__` on parent classes**: If any base class lacks `__slots__`, instances still get `__dict__` from the base. All classes in the hierarchy must declare `__slots__`. + +## Code Example + +```python +# Regular class: ~296 bytes per instance +class Node: + def __init__(self, id, weight, visited=False): + self.id = id + self.weight = weight + self.visited = visited + +# Slots class: ~88 bytes per instance (69% reduction) +class Node: + __slots__ = ('id', 'weight', 'visited') + def __init__(self, id, weight, visited=False): + self.id = id + self.weight = weight + self.visited = visited + +# Python 3.10+ dataclass with slots +from dataclasses import dataclass +@dataclass(slots=True) +class Node: + id: int + weight: float + visited: bool = False +``` + +## Environment + +CPython 3.10+ (for `dataclass(slots=True)`). `__slots__` itself works in all CPython versions. Memory measured with `tracemalloc` and `sys.getsizeof()`. Does not apply to PyPy (which uses a different object model that's already memory-efficient). diff --git a/claude-plugin/skills/aae/knowledge-base/021-generator-expressions-memory.md b/claude-plugin/skills/aae/knowledge-base/021-generator-expressions-memory.md new file mode 100644 index 0000000..8c7022c --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/021-generator-expressions-memory.md @@ -0,0 +1,52 @@ +# Generator Expressions for Memory-Efficient Pipelines + +## Problem + +Python programs that process large datasets by building intermediate lists at each pipeline stage. The metric was peak memory usage. A pipeline like `[transform(x) for x in data]` followed by `[x for x in result if predicate(x)]` materializes the full list at each step, requiring O(N) memory per stage even when only one element is needed at a time. + +## What Worked + +Replacing list comprehensions with generator expressions in pipeline stages that feed into the next stage or a final reduction. Generators produce one element at a time (lazy evaluation), keeping memory at O(1) regardless of input size. + +For a pipeline processing a 50M-row CSV — filter → transform → aggregate — switching intermediate stages to generators reduced peak memory from 4.8 GB to 0.3 GB while maintaining identical throughput. + +## Experiment Data + +| Pipeline Style | Peak Memory | Time (s) | +|---------------|-------------|----------| +| List comprehensions (3 stages) | 4.8 GB | 38.2 | +| Generator expressions (3 stages) | 0.3 GB | 37.8 | +| Generator + `itertools.chain` | 0.3 GB | 36.1 | + +## Code Example + +```python +# BAD: materializes entire list at each stage — 3x memory +filtered = [row for row in data if row['status'] == 'active'] # N items in memory +transformed = [compute(row) for row in filtered] # N items in memory +result = sum(t['value'] for t in transformed) # N items in memory + +# GOOD: generators — O(1) memory for intermediate stages +filtered = (row for row in data if row['status'] == 'active') # lazy +transformed = (compute(row) for row in filtered) # lazy +result = sum(t['value'] for t in transformed) # streams through + +# Built-in functions that accept iterators directly: +# sum(), min(), max(), any(), all(), ''.join(), collections.Counter() +# These consume generators without materializing a list. + +# File processing — never load entire file into memory +total = sum( + len(line) + for line in open('huge_file.txt') + if not line.startswith('#') +) +``` + +## What Didn't Work + +- **Generators when you need multiple passes**: A generator is exhausted after one iteration. If you need to iterate the data twice (e.g., compute mean then variance), you must either recreate the generator or materialize to a list. For multi-pass algorithms, `itertools.tee` helps but buffers elements internally, negating memory savings. + +## Environment + +Python 3.8+. Generators are a core language feature, no dependencies needed. Memory measured with `tracemalloc.get_traced_memory()`. diff --git a/claude-plugin/skills/aae/knowledge-base/022-numba-jit-numerical-computation.md b/claude-plugin/skills/aae/knowledge-base/022-numba-jit-numerical-computation.md new file mode 100644 index 0000000..612ac96 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/022-numba-jit-numerical-computation.md @@ -0,0 +1,78 @@ +# Numba JIT Compilation for Numerical Hot Loops + +## Problem + +Python numerical code with loops that can't be easily vectorized with NumPy (e.g., simulations with loop-carried dependencies, custom reduction logic, iterative algorithms). The metric was computation time. NumPy vectorization doesn't help when the algorithm is inherently iterative or requires conditional logic within the loop. + +## What Worked + +Using Numba's `@jit(nopython=True)` decorator to compile Python functions to machine code via LLVM. Numba translates a subset of Python and NumPy operations to native code at first call, achieving C-like performance without leaving Python. + +For a Monte Carlo simulation with 10M iterations of a loop body containing conditionals and array indexing, Numba achieved 120x speedup over pure Python and matched hand-written C performance within 10%. + +Key patterns that Numba handles well: +1. **Explicit loops over NumPy arrays**: Unlike NumPy, Numba makes loops fast — write naturally. +2. **Parallel loops**: `@jit(nopython=True, parallel=True)` with `prange` for embarrassingly parallel loops. +3. **CUDA GPU acceleration**: `@cuda.jit` for GPU kernels without writing CUDA C. + +## Experiment Data + +| Approach | Time (s) | Relative | +|----------|----------|----------| +| Pure Python loop | 48.2 | 1.0x | +| NumPy (partial vectorization) | 6.1 | 7.9x | +| Numba @jit | 0.40 | 120x | +| Numba @jit(parallel=True) | 0.11 | 438x | +| C (gcc -O3) | 0.37 | 130x | + +## What Didn't Work + +- **Numba with object mode**: If Numba falls back to "object mode" (when it can't infer types), performance is worse than plain Python due to JIT overhead. Always use `nopython=True` and fix type errors. +- **Numba for string operations or dict-heavy code**: Numba only supports a subset of Python — no dicts, no classes (unless `@jitclass`), limited string support. Use it for numerical array code only. +- **First-call JIT overhead**: The first call compiles the function (~0.5-2s). For short-lived scripts, use `@jit(cache=True)` to persist the compiled version to disk. + +## Code Example + +```python +from numba import jit, prange +import numpy as np + +# Pure Python: 48.2s +def monte_carlo_py(n, arr): + total = 0.0 + for i in range(n): + x = arr[i] + if x > 0.5: + total += x * x + else: + total += x + return total + +# Numba: 0.40s (120x faster) +@jit(nopython=True) +def monte_carlo_numba(n, arr): + total = 0.0 + for i in range(n): + x = arr[i] + if x > 0.5: + total += x * x + else: + total += x + return total + +# Parallel version: 0.11s (438x faster) +@jit(nopython=True, parallel=True) +def monte_carlo_parallel(n, arr): + total = 0.0 + for i in prange(n): # prange = parallel range + x = arr[i] + if x > 0.5: + total += x * x + else: + total += x + return total +``` + +## Environment + +Python 3.11, Numba 0.58, NumPy 1.25. Numba requires LLVM (bundled via `llvmlite`). AMD Ryzen 9 5900X (12 cores). Install Numba as described in its documentation. diff --git a/claude-plugin/skills/aae/knowledge-base/023-multiprocessing-cpu-bound.md b/claude-plugin/skills/aae/knowledge-base/023-multiprocessing-cpu-bound.md new file mode 100644 index 0000000..bbaa5db --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/023-multiprocessing-cpu-bound.md @@ -0,0 +1,61 @@ +# multiprocessing for CPU-Bound Parallel Work in Python + +## Problem + +CPU-bound Python workloads (compression, hashing, image processing, numerical computation) where `threading` provides no speedup due to the Global Interpreter Lock (GIL). The metric was wall-clock time for processing a batch of independent work items. With `threading`, all threads contend for the GIL, resulting in serial execution with context-switch overhead. + +## What Worked + +Using `multiprocessing.Pool` (or `concurrent.futures.ProcessPoolExecutor`) to distribute work across OS processes, each with its own Python interpreter and GIL. For CPU-bound work, this achieves true parallelism proportional to the number of cores. + +Key implementation details: +1. **`Pool.map` / `Pool.imap_unordered`**: For batch processing of independent items. `imap_unordered` returns results as they complete, improving responsiveness. +2. **Chunk size tuning**: Large chunks reduce IPC overhead. Default chunk size is often too small — set `chunksize=len(items)//num_workers` as a starting point. +3. **`shared_memory`** (Python 3.8+): For sharing large arrays between processes without serialization. Avoids the cost of pickling large inputs. + +For image resizing of 10K JPEG files, `ProcessPoolExecutor` with 8 workers achieved 7.1x speedup on 8 cores (vs 0.95x with `ThreadPoolExecutor`). + +## Experiment Data + +| Approach | Time (s) | Speedup | Notes | +|----------|----------|---------|-------| +| Sequential | 84.0 | 1.0x | Single core | +| ThreadPoolExecutor (8 workers) | 88.2 | 0.95x | GIL contention makes it worse | +| ProcessPoolExecutor (8 workers) | 11.8 | 7.1x | True parallelism | +| ProcessPoolExecutor (8, chunksize=100) | 10.2 | 8.2x | Reduced IPC overhead | + +## What Didn't Work + +- **Large objects as arguments**: Each process argument is pickled and sent over a pipe. Passing a 500MB NumPy array per call killed performance with serialization overhead. Solution: use `multiprocessing.shared_memory` or `mmap` to share large data. +- **Too many processes**: Spawning more processes than physical cores causes context-switch overhead and cache thrashing. Match process count to core count (`os.cpu_count()`). +- **Short tasks**: If each work item takes <1ms, the IPC overhead dominates. Batch small items into chunks. + +## Code Example + +```python +from concurrent.futures import ProcessPoolExecutor +import os + +def process_image(path): + img = load(path) + return resize(img, (224, 224)) + +# BAD: threading — GIL prevents parallelism +from concurrent.futures import ThreadPoolExecutor +with ThreadPoolExecutor(8) as pool: + results = list(pool.map(process_image, paths)) # ~0.95x speedup + +# GOOD: multiprocessing — true parallelism +with ProcessPoolExecutor(max_workers=os.cpu_count()) as pool: + results = list(pool.map(process_image, paths, chunksize=100)) # ~8x speedup + +# For shared large data (Python 3.8+): +from multiprocessing import shared_memory +shm = shared_memory.SharedMemory(create=True, size=big_array.nbytes) +shared_arr = np.ndarray(big_array.shape, dtype=big_array.dtype, buffer=shm.buf) +shared_arr[:] = big_array[:] # copy once, share across processes +``` + +## Environment + +CPython 3.8+ (for `shared_memory`). Note: Python 3.13 introduces a free-threaded mode (no-GIL) that may make `threading` viable for CPU-bound work in the future. Tested on Linux (fork-based spawning is faster than Windows spawn). diff --git a/claude-plugin/skills/aae/knowledge-base/024-dict-set-lookup-vs-list.md b/claude-plugin/skills/aae/knowledge-base/024-dict-set-lookup-vs-list.md new file mode 100644 index 0000000..02f5853 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/024-dict-set-lookup-vs-list.md @@ -0,0 +1,56 @@ +# dict/set O(1) Lookup vs List Linear Search + +## Problem + +Python code performing membership tests or lookups against collections using `if x in my_list` or `for item in my_list: if item.key == target`. The metric was lookup time for repeated queries against a collection of N items. List membership test is O(N); for large N and frequent lookups, this dominates runtime. + +## What Worked + +Replacing lists with `set` (for membership tests) or `dict` (for key→value lookups). Both use hash tables with amortized O(1) lookup, making them 100-1000x faster than list linear search for large collections. + +For 100K lookups against a collection of 1M strings, `set` lookup was 1,850x faster than `list`. + +## Experiment Data + +| Collection | N=1,000 | N=100,000 | N=1,000,000 | +|-----------|---------|-----------|-------------| +| list `in` | 0.12ms/query | 11.8ms/query | 118ms/query | +| set `in` | 0.06μs/query | 0.07μs/query | 0.08μs/query | +| Speedup | 2,000x | 169,000x | 1,475,000x | + +## Code Example + +```python +# BAD: O(N) per lookup — disastrous for large N +allowed_ids = [1, 2, 3, ..., 1_000_000] # list +if user_id in allowed_ids: # scans up to 1M elements + grant_access() + +# GOOD: O(1) per lookup +allowed_ids = {1, 2, 3, ..., 1_000_000} # set — same syntax, wildly different perf +if user_id in allowed_ids: # hash lookup, ~80ns regardless of size + grant_access() + +# For key→value: dict instead of list of tuples +# BAD: +records = [(id, data), ...] +result = next((d for i, d in records if i == target), None) # O(N) + +# GOOD: +records = {id: data, ...} +result = records.get(target) # O(1) + +# Converting a list to set for repeated lookups +if len(queries) > 1: # worth converting if more than one lookup + lookup_set = set(collection) + results = [x for x in queries if x in lookup_set] +``` + +## What Didn't Work + +- **Sets for small collections** (<20 elements): The hash computation overhead can make set lookup slower than list scan for tiny collections. For N<20, list `in` is often faster due to pointer locality and no hashing. +- **Unhashable elements**: Sets and dicts require hashable keys. For lists-of-lists or dicts-as-elements, convert to tuples or use `frozenset`. This adds conversion cost. + +## Environment + +CPython 3.11. dict/set in CPython use a compact hash table with open addressing. Performance is consistent across platforms. diff --git a/claude-plugin/skills/aae/knowledge-base/025-mmap-large-datasets-python.md b/claude-plugin/skills/aae/knowledge-base/025-mmap-large-datasets-python.md new file mode 100644 index 0000000..baa660e --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/025-mmap-large-datasets-python.md @@ -0,0 +1,60 @@ +# Memory-Mapped Files for Large Datasets in Python + +## Problem + +Python programs that load entire large files (1GB+) into memory using `open().read()` or `pandas.read_csv()`, causing high RSS and long startup times. The metric was peak memory usage and time-to-first-result. Loading a 4GB file required 4GB+ of RAM before any processing could begin. + +## What Worked + +Using Python's `mmap` module to map files directly into virtual address space, allowing file content to be accessed as a byte array without loading it all into physical memory. The OS pages in data on demand and can evict pages under memory pressure. + +For searching a 4GB log file for pattern matches, `mmap` reduced peak RSS from 4.2GB to 12MB and returned the first match in 0.1s instead of 18s (time to load the entire file). + +NumPy also supports memory-mapped arrays via `np.memmap` for numerical data, enabling computation on datasets larger than RAM. + +## Experiment Data + +| Approach | Peak RSS | Time to First Result | Total Time | +|----------|----------|---------------------|------------| +| `open().read()` | 4.2 GB | 18.2s | 24.1s | +| `mmap` | 12 MB | 0.1s | 22.8s | +| `np.memmap` (for .npy) | 8 MB | 0.01s | 19.4s | + +## Code Example + +```python +import mmap + +# Memory-mapped file search — constant memory regardless of file size +with open('huge_log.txt', 'r') as f: + mm = mmap.mmap(f.fileno(), 0, access=mmap.ACCESS_READ) + # Use like a bytes object + idx = mm.find(b'ERROR') + while idx != -1: + line_start = mm.rfind(b'\n', 0, idx) + 1 + line_end = mm.find(b'\n', idx) + print(mm[line_start:line_end].decode()) + idx = mm.find(b'ERROR', idx + 1) + mm.close() + +# NumPy memory-mapped array for large numerical datasets +import numpy as np + +# Create a memory-mapped array (writes to disk) +fp = np.memmap('large_matrix.dat', dtype='float64', mode='w+', shape=(100000, 1000)) +fp[:] = compute_data() # writes directly to disk +del fp + +# Read it back — no RAM needed for storage +data = np.memmap('large_matrix.dat', dtype='float64', mode='r', shape=(100000, 1000)) +result = data[5000:5100].mean(axis=0) # only loads 100 rows into RAM +``` + +## What Didn't Work + +- **mmap for sequential-only processing**: If you process a file strictly front-to-back and never revisit earlier data, simple buffered `read()` with small chunks is equally fast and simpler. mmap shines for random access and search patterns. +- **mmap on network filesystems (NFS)**: Page faults over the network are orders of magnitude slower than local disk. Copy the file locally first. + +## Environment + +Python 3.8+, `mmap` is a stdlib module. Works on Linux, macOS, Windows. `np.memmap` requires NumPy. Tested on Linux 6.1, NVMe SSD. diff --git a/claude-plugin/skills/aae/knowledge-base/026-local-variable-caching.md b/claude-plugin/skills/aae/knowledge-base/026-local-variable-caching.md new file mode 100644 index 0000000..f27adcd --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/026-local-variable-caching.md @@ -0,0 +1,60 @@ +# Local Variable Caching to Avoid Global/Attribute Lookups + +## Problem + +Python hot loops that repeatedly access global variables, module attributes, or object attributes (e.g., `self.config.threshold`, `math.sqrt`, `len`). The metric was loop execution time. CPython resolves global names via dictionary lookup (LOAD_GLOBAL) on every access — ~50ns each, adding up to significant overhead in tight loops with millions of iterations. + +## What Worked + +Caching frequently accessed names into local variables before the loop. CPython stores local variables in a fixed-size array indexed by position (LOAD_FAST), which is ~2-3x faster than dictionary-based global/attribute lookup. + +For a loop with 10M iterations calling `math.sqrt` and accessing `self.threshold`, localizing both references reduced loop time by 38%. + +## Experiment Data + +| Access Pattern | Time (s) | Relative | +|---------------|----------|----------| +| Global `math.sqrt` in loop | 3.82 | 1.00x | +| Local `sqrt = math.sqrt` | 2.36 | 1.62x | +| Global + `self.x` attribute | 5.14 | 1.00x | +| Local + cached attribute | 3.21 | 1.60x | + +## Code Example + +```python +import math + +# SLOW: two dict lookups per iteration (module + function) +def compute_slow(data): + result = [] + for x in data: # 10M iterations + result.append(math.sqrt(x * x + 1)) # LOAD_GLOBAL 'math' + LOAD_ATTR 'sqrt' + return result + +# FAST: local variable lookup (array index) +def compute_fast(data): + sqrt = math.sqrt # cache function reference + append = result.append # cache bound method + result = [] + for x in data: + append(sqrt(x * x + 1)) # LOAD_FAST — array index, no dict lookup + return result + +# Same pattern for instance attributes in methods +class Processor: + def process(self, data): + threshold = self.config.threshold # one lookup instead of N + transform = self.transform # cache method reference + for item in data: + if item > threshold: # LOAD_FAST, not LOAD_ATTR chain + transform(item) +``` + +## What Didn't Work + +- **Over-caching everything**: Caching a variable used only 2-3 times adds clutter with negligible benefit. Only worth it in loops with >10K iterations or when the attribute chain is deep (e.g., `self.config.model.params.lr`). +- **This technique in PyPy**: PyPy's JIT already optimizes attribute lookups; manual caching provides no benefit there. + +## Environment + +CPython 3.8-3.12. Verified with `dis.dis()` to confirm bytecode differences (LOAD_FAST vs LOAD_GLOBAL). Python 3.12's LOAD_GLOBAL is slightly faster due to inline caching, but local is still faster. diff --git a/claude-plugin/skills/aae/knowledge-base/027-itertools-lazy-pipelines.md b/claude-plugin/skills/aae/knowledge-base/027-itertools-lazy-pipelines.md new file mode 100644 index 0000000..6a7f5ca --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/027-itertools-lazy-pipelines.md @@ -0,0 +1,48 @@ +# itertools for Lazy Evaluation Pipelines + +## Problem + +Python data processing pipelines that create intermediate lists when chaining operations (filter, map, group, take). The metric was peak memory and throughput for streaming-style processing of large sequences. Each intermediate list doubles memory usage and requires a full pass before the next stage begins. + +## What Worked + +Using `itertools` functions to build lazy pipelines that process one element at a time through all stages. Key functions and their use cases: + +- **`itertools.chain`**: Concatenate multiple iterables without materializing. `chain(iter_a, iter_b)` instead of `list_a + list_b`. +- **`itertools.islice`**: Take first N items from any iterable without materializing. `islice(huge_iter, 100)` instead of `list(huge_iter)[:100]`. +- **`itertools.groupby`**: Group consecutive elements by key, lazily. Requires sorted input. +- **`itertools.compress`**: Filter with a boolean selector array. Faster than list comprehension for pre-computed masks. +- **`itertools.accumulate`**: Running reduction (cumulative sum, product, etc.) as a lazy iterator. +- **`itertools.product`/`combinations`/`permutations`**: Generate combinatorial sequences lazily instead of materializing all possibilities. + +For a pipeline reading 100M records, filtering to 1%, and taking top 1000, `itertools.islice(filter(...), 1000)` processed only ~100K records (early termination) vs the list approach materializing all 100M. + +## Code Example + +```python +from itertools import chain, islice, groupby, compress, accumulate + +# BAD: materializes everything, even though we only want 10 items +results = [transform(x) for x in data if predicate(x)][:10] # processes ALL of data + +# GOOD: stops after finding 10 matches +results = list(islice( + (transform(x) for x in data if predicate(x)), + 10 +)) # processes only enough data to find 10 matches + +# Concatenate large iterables without copying +all_data = chain(file1_iter, file2_iter, file3_iter) # zero memory + +# Cartesian product — lazy, doesn't build the full matrix +from itertools import product +for params in product(learning_rates, batch_sizes, seeds): + run_experiment(*params) + +# Running statistics +cumsum = list(accumulate(data)) # cumulative sum, O(1) extra memory +``` + +## Environment + +Python 3.8+. `itertools` is a C-extension stdlib module — its functions run at near-C speed. No external dependencies. diff --git a/claude-plugin/skills/aae/knowledge-base/028-struct-binary-data-packing.md b/claude-plugin/skills/aae/knowledge-base/028-struct-binary-data-packing.md new file mode 100644 index 0000000..8a7b1da --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/028-struct-binary-data-packing.md @@ -0,0 +1,53 @@ +# struct Module for Efficient Binary Data Packing + +## Problem + +Python programs serializing/deserializing large volumes of structured binary data (network packets, file formats, sensor readings) using Python objects. The metric was throughput (records per second). Converting between Python objects and binary formats using ad-hoc byte manipulation or JSON/pickle is slow and memory-heavy. + +## What Worked + +Using Python's `struct` module to pack/unpack fixed-format binary data directly. `struct.pack`/`unpack` translate between Python values and C-style binary representations in a single call, with minimal overhead. + +For parsing 10M binary records (each: uint32 timestamp, float64 value, uint8 flags = 13 bytes), `struct.unpack` was 8.5x faster than equivalent manual byte slicing and 45x faster than JSON deserialization of the same data. + +Key patterns: +1. **`struct.Struct` pre-compiled format**: Compile the format string once, reuse the `Struct` object for repeated pack/unpack — avoids re-parsing the format string. +2. **`struct.iter_unpack`** (Python 3.4+): Iterates over a buffer, unpacking repeated structures without Python loop overhead. + +## Experiment Data + +| Method | Records/sec (M) | Memory/record | +|--------|-----------------|---------------| +| JSON parse | 0.22 | ~400 bytes | +| Manual byte slicing | 1.18 | ~200 bytes | +| struct.unpack (per-record) | 4.80 | 13 bytes | +| struct.Struct.iter_unpack | 10.05 | 13 bytes | + +## Code Example + +```python +import struct + +# Define format once (uint32, float64, uint8) +record_fmt = struct.Struct('50% of runtime — a tight loop over numerical data with operations that can't be vectorized with NumPy (e.g., graph traversal, custom distance metrics, state-machine parsing). The metric was function execution time. Rewriting the entire application in C/C++ is impractical. + +## What Worked + +Rewriting only the hot function in Cython — a superset of Python that compiles to C extension modules. Adding C type declarations to variables enables Cython to generate tight C loops that bypass the Python interpreter entirely. + +Key steps for maximum speedup: +1. **Type declarations**: `cdef int i`, `cdef double x` — tells Cython to use C types instead of Python objects. +2. **Typed memoryviews**: `double[:] arr` — direct C-level access to NumPy arrays without bounds checking or Python API overhead. +3. **`@cython.boundscheck(False)`**: Disable array bounds checking in the hot loop (after verifying correctness). +4. **`@cython.wraparound(False)`**: Disable negative index handling. + +For a pairwise distance matrix computation (N=10K, custom metric), Cython with typed memoryviews achieved 95x speedup over pure Python and matched C within 5%. + +## Experiment Data + +| Approach | Time (s) | Relative | +|----------|----------|----------| +| Pure Python | 142.0 | 1.0x | +| NumPy (partial vectorization) | 18.4 | 7.7x | +| Cython (no type hints) | 38.2 | 3.7x | +| Cython (typed + memoryviews) | 1.49 | 95x | +| C (gcc -O3) | 1.42 | 100x | + +## Code Example + +```cython +# distance.pyx +import cython +import numpy as np +cimport numpy as cnp + +@cython.boundscheck(False) +@cython.wraparound(False) +def pairwise_distances(double[:, :] X, double[:, :] D): + """Compute pairwise Euclidean distance matrix.""" + cdef int N = X.shape[0] + cdef int M = X.shape[1] + cdef int i, j, k + cdef double diff, dist + + for i in range(N): + for j in range(i + 1, N): + dist = 0.0 + for k in range(M): + diff = X[i, k] - X[j, k] + dist += diff * diff + D[i, j] = dist ** 0.5 + D[j, i] = D[i, j] +``` + +```python +# setup.py +from setuptools import setup +from Cython.Build import cythonize +import numpy as np +setup(ext_modules=cythonize("distance.pyx"), include_dirs=[np.get_include()]) +``` + +## What Didn't Work + +- **Cython without type annotations**: Without `cdef` types, Cython generates Python API calls for every operation — only 2-4x faster than CPython. The type declarations are essential. +- **Cython for I/O-bound code**: Cython optimizes CPU work, not I/O. If the bottleneck is file/network I/O, Cython won't help. +- **Cython with complex Python objects**: Cython is fastest with C types and memoryviews. Operations on Python dicts, sets, or custom objects still go through the Python API. + +## Environment + +Python 3.8+, Cython 3.0+, NumPy. Requires a C compiler (GCC/Clang/MSVC). Install Cython as described in its documentation. Build with `python setup.py build_ext --inplace` or use `pyximport` for development. diff --git a/claude-plugin/skills/aae/knowledge-base/034-graph-clustering-lp-refinement.md b/claude-plugin/skills/aae/knowledge-base/034-graph-clustering-lp-refinement.md new file mode 100644 index 0000000..6a4c382 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/034-graph-clustering-lp-refinement.md @@ -0,0 +1,146 @@ +# Optimizing Label Propagation in Graph Clustering + +## Problem + +Optimizing the runtime of a signed graph correlation clustering solver ([ScalableCorrelationClustering](https://github.com/KaHIP/ScalableCorrelationClustering)) built on the KaHIP multilevel framework. The solver uses label propagation (LP) for both coarsening and refinement, plus FM-based refinement on coarse levels. The metric was geometric mean of execution times across multiple real-world graph instances, with a hard constraint that solution quality must stay within 0.001% of the baseline. The baseline geometric mean was 1.528s. + +## What Worked + +The combined effect was a **1.23x speedup (18.7%)** over 30 experiments. Below is each technique with enough detail to reproduce. + +### 1. Dense vectors replacing hash maps in LP inner loops (~7%) + +The LP inner loop accumulates edge weights per neighboring block to decide which block a node should move to. The original code used `std::unordered_map` — every edge traversal hashed the target block ID, probed the hash table, and potentially allocated a new bucket. Since this runs for every node on every LP sweep on every coarsening/refinement level, it dominated the profile. + +**Fix:** Replace with a dense `std::vector` of size `max_blocks`, indexed directly by block ID. Track which entries were touched in a small side vector, and reset only those entries after processing each node. This turns O(1)-amortized hash lookups into O(1)-worst-case array indexing and eliminates all hashing, bucket allocation, and cache-hostile pointer chasing. + +The same pattern applied to `maxNodeHeap`, which backed its key lookups with a hash map. Replacing it with a three-vector architecture (`m_elements`, `m_element_index[node] → position`, `m_heap[position] → key`) gives O(1) direct-indexed lookup instead of hash probing. + +### 2. Counting-sort contraction replacing boundary objects (~3%) + +Each coarsening level contracts the graph: fine nodes are merged into coarse super-nodes. The original code built a `complete_boundary` object (~16MB on large graphs), saved and restored the full partition map (~8MB), and used `vector>` to group nodes per block — all to support a generic contraction interface. + +**Fix:** A single counting-sort pass groups fine nodes by their coarse mapping in O(N) time: +1. Histogram: count how many fine nodes map to each coarse node. +2. Prefix sum: convert counts to start offsets. +3. Scatter: place each fine node at its offset position. + +Then iterate coarse nodes in order, processing contiguous runs of fine nodes. This replaces ~24MB of intermediate structures with three flat arrays totaling O(N) and eliminates the partition save/restore entirely. Memory access is sequential during the scatter and iteration phases, which is cache-friendly. + +### 3. LP sweep specialization and block-ID caching (~3%) + +LP processes each node in three sweeps: (1) accumulate edge weights per block, (2) find the best block, (3) reset the accumulator. In the original code, all three sweeps read the edge array independently, each time dereferencing `cluster_id[edges[e].target]` to look up the target's block. + +**Fix — cache block IDs for low-degree nodes:** For nodes with degree ≤ 32 (covering ~95% of nodes in real-world graphs), sweep 1 writes the block IDs into a stack-allocated `PartitionID blk_cache[32]`. Sweeps 2 and 3 iterate `blk_cache` instead of re-reading the edge array and re-dereferencing `cluster_id[]`. The 32-element cache fits in one or two L1 cache lines. + +**Fix — specialize sweep 2 for unconstrained path:** When no cluster size constraints are active (the common case in correlation clustering), sweep 2 only needs block IDs and accumulated weights — it doesn't need edge weights or node IDs. The specialized path iterates the `blk_cache` array in a tight loop with no edge-array access at all, cutting random memory reads in half. + +**Fix — cache partition IDs in constrained path:** When constraints are active and the graph is already partitioned, sweep 2 must also check `getPartitionIndex()` for each neighbor. Caching these in a `PartitionID part_cache[32]` alongside `blk_cache` avoids a second round of random lookups into the partition array. + +### 4. Pointer hoisting with `__restrict__` (~1.5%) + +The LP inner loop accesses edges via `G.getEdgeTarget(e)` which compiles to `graphref->m_edges[e].target` — a pointer-to-pointer indirection on every edge. With millions of edges per LP iteration, this adds up. + +**Fix:** Add `edge_array()` / `node_array()` accessors to `graph_access` that return raw pointers, and hoist them before the loop: +```cpp +const Edge* __restrict__ edges = G.edge_array(); +EdgeWeight* __restrict__ hmap = m_hash_map.data(); +``` +The `__restrict__` qualifier tells the compiler these pointers don't alias, enabling auto-vectorization and instruction reordering that wasn't possible through the accessor indirection. + +### 5. Persistent buffers as class members (~2%) + +LP coarsening and LP refinement each use several large buffers: the hash map vector, a permutation array, and two queue-membership vectors (`vector`). Originally these were local variables, allocated and freed on every call — once per coarsening level (typically 10-15 levels). + +**Fix:** Move them to class member variables (`m_hash_map`, `m_permutation`, `m_qc_a`, `m_qc_b`). On each call, resize if needed (capacity grows monotonically during coarsening since graphs shrink), then `assign()` to reset values. This converts O(N) allocations to O(N) memsets, which are much cheaper — memset is a single cache-line-streaming operation vs malloc's free-list search, mmap, and page-fault overhead. + +### 6. Stack allocation of framework objects (~1.5%) + +The multilevel loop allocates LP, contraction, and stop-rule objects at each level. Originally these were heap-allocated (`new`/`delete`), producing malloc pressure and heap fragmentation over 10+ levels. + +**Fix:** Stack-allocate them as local variables in the coarsening loop. Constructor/destructor run at scope entry/exit with zero allocator overhead. For refinement, the LP and k-way refinement objects are created once and reused across all uncoarsening levels via persistent smart pointers. + +### 7. tcmalloc_minimal (~4%) + +After eliminating the biggest allocation hotspots, the remaining malloc/free calls (from graph construction, edge arrays, STL containers) still added up. Linking Google's `tcmalloc_minimal` replaced glibc's allocator with one that uses per-thread free-list caches, avoiding lock contention and reducing fragmentation. + +**Integration:** Auto-detected via CMake `find_library(TCMALLOC_LIB tcmalloc_minimal)`, linked only on Linux. Falls back to the default allocator if not found. + +### 8. Smaller wins (~1.5% combined) + +- **`vector` over `vector`**: The queue-membership flags were `vector`, which uses bit-packing. Each access requires shift+mask operations. Switching to `vector` (one byte per entry) trades 8x memory for direct byte access — worthwhile because these vectors are small relative to the graph and accessed in the hot loop. +- **`MADV_HUGEPAGE` for LP arrays**: The hash map vector is randomly accessed by block ID. On graphs with 2M+ nodes, this causes TLB thrashing with 4KB pages. `madvise(MADV_HUGEPAGE)` hints the kernel to back it with 2MB pages, reducing TLB entries needed by 512x. Only applied to LP-local arrays — applying it to the main graph arrays caused THP overhead that was worse than the TLB savings. +- **Compiler flags**: `-fprefetch-loop-arrays` lets GCC insert prefetch instructions for streaming edge-array iteration. `-fno-plt` eliminates PLT indirection on shared library calls (minor, but free). + +## Experiment Data + +| # | Commit | Geo-mean (s) | Status | Hypothesis | +|---|--------|-------------|--------|------------| +| 0 | f683dcc | 1.528 | baseline | — | +| 1 | 79c85ce | 1.426 | keep | Dense vector replaces unordered_map | +| 2 | e35d815 | 1.427 | keep | Lazy reset for vertex_moved_hashtable | +| 3 | cf06cce | 1.390 | keep | Direct contraction via counting sort | +| 4 | 21c8fc2 | 1.362 | keep | Hoist max_blocks vector outside loop | +| 5 | f72cc86 | 1.379 | keep | Stack-allocate LP queues | +| 6 | cc13299 | 1.366 | keep | Stack-allocate coarsening objects | +| 7 | c373c2e | 1.352 | keep | Stack-allocate refinement objects | +| 8 | 9decfee | 1.340 | keep | vector\ over vector\ | +| 9 | b19650d | 1.350 | keep | Reserve FM vector capacity | +| 10 | 6a846db | 1.377 | discard | Pre-size maxNodeHeap — wasted time | +| 11 | 41e405d | 1.380 | discard | Software prefetching — overhead > benefit | +| 12 | 5a1af07 | 1.368 | discard | Reuse contraction buffers — no gain | +| 13 | aab4041 | 1.382 | discard | vector\ in contraction — no gain | +| 14 | 9c7ea1a | 1.370 | discard | Devirtualize FM queue — slight regression | +| 15 | d166e01 | 1.356 | discard | Eliminate PartitionConfig copy — noise | +| 16 | b02db57 | 1.369 | discard | Simplify relabeling loop — marginal regression | +| 17 | 8f1838e | 1.344 | keep | MADV_HUGEPAGE for LP arrays | +| 18 | 502ff05 | 1.395 | discard | Skip m_local_degrees init — regression | +| 19 | ceb9beb | 1.352 | discard | Cache block IDs (full) — overhead from max-degree scan | +| 20 | 0089f5b | 1.349 | keep | Cache block IDs for degree<=32 | +| 21 | 859d930 | 1.330 | keep | Specialize LP sweep 2 for unconstrained path | +| 22 | 0dcf027 | 1.316 | discard | blk_cache in FM — overhead on coarse levels | +| 23 | 97189b0 | 1.307 | keep | Cache partition IDs in constrained path | +| 24 | ec86900 | 1.258 | keep | tcmalloc_minimal via LD_PRELOAD | +| 25 | 7b5981a | 1.267 | keep | tcmalloc_minimal linked via CMake | +| 26 | 11135ac | 1.268 | discard | PGO — hurts non-profiled instances | +| 27 | b73363b | 1.289 | discard | MADV_HUGEPAGE on graph arrays — THP overhead | +| 28 | 92a38c7 | 1.244 | keep | Hoist edge/hash_map pointers | +| 29 | 8034afa | 1.242 | keep | Persistent LP refinement buffers | + +18 kept, 12 discarded (60% success rate). Final speedup: **1.23x**. + +## What Didn't Work + +- **Profile-guided optimization (PGO)**: Improved profiled instances but degraded others, hurting the geometric mean across the full benchmark suite. +- **Software prefetching (`__builtin_prefetch`)**: Manual prefetch of edge arrays added more overhead than it saved. The hardware prefetcher was already doing a good job on the sequential access patterns. +- **MADV_HUGEPAGE on graph arrays**: While it helped for LP-local arrays, applying it to the main graph adjacency arrays caused THP (Transparent Huge Pages) overhead that exceeded the TLB-miss savings. +- **Devirtualizing FM queue**: Replacing virtual dispatch with templates caused a slight regression, likely due to increased code size reducing instruction cache efficiency. +- **Reusing contraction buffers across levels**: The buffers change size each level, so reuse didn't save meaningful allocation work. + +## Code Example + +Core change: replacing hash map with dense vector in LP inner loop. + +```cpp +// Before: hash map lookups in hot loop +std::unordered_map block_weights; +forall_out_edges(G, e, node) { + PartitionID target_block = G.getPartitionIndex(G.getEdgeTarget(e)); + block_weights[target_block] += G.getEdgeWeight(e); +} endfor + +// After: dense vector indexed by block ID (cleared via touched-list) +std::vector block_weights(max_blocks, 0); // persistent, reused +std::vector touched; +forall_out_edges(G, e, node) { + PartitionID target_block = G.getPartitionIndex(G.getEdgeTarget(e)); + if (block_weights[target_block] == 0) touched.push_back(target_block); + block_weights[target_block] += G.getEdgeWeight(e); +} endfor +// ... use block_weights, then reset only touched entries +for (auto b : touched) block_weights[b] = 0; +``` + +## Environment + +C++17, GCC 11.4 with `-O3 -march=native`, Linux (Ubuntu 22.04), Intel Xeon. Graphs range from thousands to millions of nodes. Multilevel framework with label propagation coarsening/refinement and FM-based k-way refinement. diff --git a/claude-plugin/skills/aae/knowledge-base/035-streaming-hypergraph-multipass-eval.md b/claude-plugin/skills/aae/knowledge-base/035-streaming-hypergraph-multipass-eval.md new file mode 100644 index 0000000..c053eac --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/035-streaming-hypergraph-multipass-eval.md @@ -0,0 +1,123 @@ +# Eliminating Heap Allocation in Streaming Hypergraph Multi-Pass Evaluation + +## Problem + +Optimizing the multi-pass streaming hypergraph partitioning in [FREIGHT](https://github.com/KaHIP/FREIGHT) via its pybind11 binding. FREIGHT uses a streaming algorithm where nodes are processed one-by-one and assigned to partition blocks. Multi-pass restreaming re-evaluates the partition after each pass and keeps the best. The metric was geometric mean of wall-clock time across 96 configurations (4 ISPD98 instances, 3 k values, 4 pass counts, 2 objectives), with a hard constraint that results must remain bit-identical to the FREIGHT CLI. The baseline geometric mean was 82.9ms. + +The bottleneck was the inter-pass evaluation: after each pass, the code needed to compute how many hyperedges (nets) span multiple partition blocks. The original implementation built a reverse mapping (`vector>` net-to-nodes, O(total_pins) with many small vector allocations) and then created a `std::set` for each net to count distinct blocks. For large hypergraphs (200K+ nets, 800K+ pins), this meant millions of heap allocations per pass. + +## What Worked + +The combined effect was a **1.82x speedup** (82.9ms to 45.6ms) over 8 kept experiments. Below are the techniques in order of impact. + +### 1. Per-net bit vectors replacing reverse mapping + std::set (~1.60x) + +The original evaluation built `net_to_nodes[net] = {node0, node1, ...}` (a `vector>`) and then for each net created a `std::set` to count distinct blocks. Both data structures involve heap allocation per net. + +**Fix:** Replace with a flat `vector` of size `num_nets * ceil(k/64)`. Each net gets `ceil(k/64)` words as a bit vector. To evaluate, iterate nodes via the existing node-to-edge CSR, and for each node's nets, OR the assigned block's bit into the net's word. Then a single popcount scan over all nets gives the distinct block count. This eliminates the entire reverse mapping and all `std::set` allocations. + +For k <= 64 (the common case), each net uses exactly one `uint64_t`, and the bit-setting reduces to `net_block_bits[net_id] |= (1ULL << block)`. + +### 2. Objective-specific evaluation paths (~4%) + +For connectivity objective, the bit-vector evaluation is needed (distinct block count per net). But for cut-net objective, the existing `stream_edges_assign` array already tracks whether each net is cut: entries equal to `CUT_NET` indicate cut nets. A simple O(num_nets) sequential scan counting `CUT_NET` entries replaces the entire bit-vector machinery for half of all configurations. + +### 3. Incremental bit-setting during the main loop (~0.5%) + +Instead of a separate O(total_pins) evaluation scan after each pass, set bits during the main partitioning loop (one OR per pin right after `solve_node`). The evaluation reduces to just the O(num_nets) popcount scan. The edge data (CSR indices) is in L1 cache from the prior accumulation scan, making the bit-setting nearly free. + +### 4. Eliminate intermediate vectors (~4%) + +The original code maintained a `valid_neighboring_nets` vector, cleared and rebuilt per node via `push_back`. Replacing this with a direct re-iteration of the CSR edges for the per-net tracking update eliminates vector overhead (no clear, no push_back, no capacity management) while accessing the same data already in L1 cache. + +### 5. Avoid redundant copies in output path (~3%) + +The original code: (a) copied best assignment to a snapshot vector on each improvement, (b) restored the snapshot back to the working array after the loop, (c) element-wise copied the working array to the numpy output. Three O(n) passes over 210K+ elements. + +**Fix:** Track which pass was best. Skip the snapshot on the last pass (just read the working array directly). Use `memcpy` for intermediate snapshots instead of `vector::operator=`. Copy the correct source directly to the numpy output via `memcpy`, skipping the intermediate restore. + +### 6. Pre-allocate best partition vectors (~3%) + +Pre-size `best_nodes_assign` and `best_blocks_weight` to their final sizes at initialization. This avoids a dynamic allocation + copy when the first improvement is found. Small but measurable because the allocation is O(n) and triggers on the first pass of every multi-pass run. + +## Experiment Data + +| # | geo_mean (ms) | Status | Hypothesis | +|---|--------------|--------|------------| +| 0 | 82.92 | baseline | Original implementation | +| 1 | 51.73 | keep | Per-net bit vectors replace net_to_nodes + std::set | +| 2 | 51.41 | keep | Incremental bit-setting during main loop | +| 3 | 52.66 | discard | Fuse bit-setting and per-net tracking (hurts ILP) | +| 4 | 49.37 | keep | Eliminate valid_neighboring_nets vector | +| 5 | 50.44 | discard | Specialize post-solve for connectivity vs cut_net (icache pressure) | +| 6 | 49.62 | discard | Specialize bit-setting/popcount for k<=64 (within noise) | +| 7 | 47.18 | keep | Direct CUT_NET counting for cut_net eval | +| 8 | 45.73 | keep | Pre-allocate best partition vectors | +| 9 | 47.86 | discard | Software prefetch (avg node degree ~4, too short for prefetch) | +| 10 | 45.69 | keep | memcpy output directly from best source | +| 11 | 47.67 | discard | Hoist use_connectivity branch (icache pressure from duplication) | +| 12 | 45.47 | keep | Skip snapshot on last pass | +| 13 | 48.55 | discard | LTO for cross-TU inlining (increased code size, worse icache) | +| 14 | 49.75 | discard | Raw pointers instead of pybind11 unchecked accessors (pybind11 generated better code) | +| 15 | 45.58 | keep | null_buf replacing /dev/null file open | + +## What Didn't Work + +Several approaches that seemed promising failed or regressed: + +- **Loop fusion (bit-setting + per-net tracking in one loop):** The two operations have different memory access patterns (OR into bit vector vs read-modify-write of edge assignment). Fusing them into a single loop hurt instruction-level parallelism. Separate loops allow the CPU to pipeline memory operations independently. This was confirmed twice (iterations 3 and 5). + +- **Code path specialization (separate loops for connectivity vs cut-net):** Duplicating the inner loop body to eliminate a branch (connectivity: unconditional write; cut-net: read-modify-write) increased instruction cache pressure. The branch was already perfectly predicted since it's loop-invariant. The compiler hoists it automatically. + +- **k_words=1 specialization:** For k <= 64, `k_words=1` eliminates a multiply per edge. But the compiler already optimizes `x * 1` and the multiply is hidden by memory latency. Added complexity for no measurable gain. + +- **Software prefetching:** Average node degree was ~4 (820K pins / 210K nodes). With so few iterations per inner loop, prefetch instructions don't have enough lead time to hide latency, and the prefetch instruction itself adds overhead. + +- **LTO (link-time optimization):** Expected to devirtualize `compute_score` calls across translation units. Instead, increased code size and worsened instruction cache behavior, causing a 6% regression. + +- **Raw data pointers vs pybind11 unchecked accessors:** Replacing `vp(i)` / `ve(i)` (pybind11 proxy) with `vp[i]` / `ve[i]` (raw pointer dereference) was 9% slower. The pybind11 `unchecked<1>()` accessor apparently provides alignment or aliasing hints that help the compiler generate better code. A surprising result; do not assume raw pointers are faster than well-designed proxy objects. + +## Code Example + +Core of the bit-vector evaluation (connectivity mode): + +```cpp +// Allocate once: ceil(k/64) words per net +size_t k_words = (k + 63) / 64; +std::vector net_block_bits(num_nets * k_words, 0); + +// During main partitioning loop, after solve_node: +uint64_t bit = uint64_t(1) << (block & 63); +size_t word = block >> 6; +for (int64_t e = edge_begin; e < edge_end; e++) { + net_block_bits[ve(e) * k_words + word] |= bit; +} + +// After pass: popcount scan +for (int64_t net = 0; net < num_nets; net++) { + int distinct = 0; + for (size_t w = 0; w < k_words; w++) + distinct += __builtin_popcountll(net_block_bits[net * k_words + w]); + if (distinct > 1) + pass_connectivity += distinct - 1; +} +// Reset for next pass +std::fill(net_block_bits.begin(), net_block_bits.end(), 0); +``` + +Cut-net evaluation (no bit vectors needed): + +```cpp +// stream_edges_assign[net] == CUT_NET means the net is cut +double pass_cut = 0; +for (int64_t net = 0; net < num_nets; net++) + if (stream_edges_assign[net] == CUT_NET) pass_cut += 1; +``` + +## Environment + +- **Language:** C++17 with pybind11 bindings, compiled via scikit-build-core +- **Hardware:** x86-64 Linux (256-core server), also verified on macOS ARM +- **Compiler:** GCC with `-O2 -march=native` +- **Instances:** ISPD98 ibm01 (13K nodes), ibm05 (29K nodes), ibm18 (211K nodes) +- **Benchmark:** 96 configurations (4 instances, k in {4,8,16}, passes in {2,3,5,10}, both connectivity and cut-net objectives), 3 repetitions per config, median timing diff --git a/claude-plugin/skills/aae/knowledge-base/036-kahip-bit-identical-speedup.md b/claude-plugin/skills/aae/knowledge-base/036-kahip-bit-identical-speedup.md new file mode 100644 index 0000000..eb1a391 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/036-kahip-bit-identical-speedup.md @@ -0,0 +1,364 @@ +# Speeding up a mature multilevel graph partitioner with bit-identical or quality-preserving output (KaHIP kaffpa) + +## Problem + +Make KaHIP's `kaffpa` (multilevel graph partitioner, C++) faster in whole-process wall time while every partition file stays +byte-identical to the original's for the same seed. That rules out any change to the sequence of random numbers, to the order +of iterated containers, or to tie-breaking. Benchmark: 537 jobs (mesh, road and social graphs; k from 2 to 256; the six +preconfigurations fast, eco, strong, fsocial, esocial, ssocial), 3 repetitions each. The metric is the shifted geometric mean of +median wall-time ratios against the original. Correctness is the sha256 of every partition file plus a separate crosscheck of +864 runs. + +## What Worked + +1. **Look for algorithmic redundancy, not only data-structure tuning.** The biggest single gain (-7.6% geomean, up to 3.3x + on single jobs) came from 2-way FM. For every neighbor of a moved node, it recomputed the neighbor's internal and external + degree by scanning that neighbor's whole adjacency, which is quadratic around hub nodes in the coarse graphs of initial + partitioning. The fix caches the degrees per FM call (stamp array, reused across calls). Before each move, every neighbor + in the two blocks is made cached with its pre-move values, then the move's edge weight is added or subtracted per edge, so + parallel edges also add up exactly. The moved node itself is recomputed. The decision code then reads final values exactly + as before. callgrind's per-line view (40% of instructions on one partition-index load) pointed straight at it. +2. **Cheap whole-process overheads first.** A hand-written METIS reader (one `fread`, a manual integer scanner, the same + comment/empty-line semantics), a single buffered write of the partition, and computing the final cut once cut the + non-partitioning time from 102 s to 34 s in total (27% of a `fast` job had been I/O). +3. **Lookup-only hash maps can be replaced, iterated ones cannot.** Maps that are never iterated (FM "moved" table, heap + index) were moved to an open-addressing map. Maps whose iteration order feeds the algorithm (boundary sets, block pairs) + had to stay `std::unordered_map`. +4. **A global small-object allocator** (size classes up to 512 B, bump allocation plus free lists in one reserved range) was + about -3.5%. It cannot change the output because nothing in the code is keyed or ordered by pointer. Note: reserving 128 GiB + of address space fails under valgrind, so the allocator silently falls back to malloc there. Build a profiling copy with a + smaller reservation, or the profile shows a malloc cost that does not exist. +5. **Build the flow graph as CSR** in one pass, and map nodes to flow nodes with a hash map instead of O(n) vectors per call + (flow refinement -7 to -8%, lower peak memory). +6. **Remove per-call O(n) setup.** Flow refinement allocated and filled an O(n) BFS depth array per call; reusing one array + and resetting only the reached entries cut flow time by 8%. Sorting (rating, edge) pairs instead of edge ids with an + indirect rating comparator gives the same permutation (std::sort only depends on comparison outcomes) and cut matching by + 3 to 5%. +7. **Use a time-based sampler, not only callgrind.** perf was unavailable (perf_event_paranoid 4), so a 40-line + LD_PRELOAD library (ITIMER_PROF at 1 kHz, `backtrace()` in the handler, addr2line with `-i` for inline frames) gave real + time per source line and per caller. It showed different hot spots than callgrind (for example, a scan over hub + adjacencies at 23% of one web-graph job). Start the sampler after `taskset`, not before it: the interval timer survives + `exec` but the handler does not, so SIGPROF kills the child. +8. **Per-phase timers printed by the binary** (exclusive scopes: coarsening, initial partitioning, refinement, plus detail + scopes such as flow and multitry) make it possible to judge each hypothesis by the component it targets, when several are + measured in one full run. +9. **Software prefetch for whole-graph sweeps in random order.** Label propagation visits all nodes in a random + permutation, so every node, its edge list and its neighbors' cluster ids are cache misses. Prefetching the node entry three + nodes ahead, its edges two ahead and the neighbors' cluster ids (at most 32) one ahead cut coarsening by 14%. The same + prefetch in the k-way move loop was 2% slower: those neighborhoods are already cache resident. +10. **Small containers in the max-flow solver.** `std::vector` flags became `char`, and edges are reached through one + pointer per node; push-relabel time fell by 8% with the same capacity casts (a `NodeID` accessor truncated a 64-bit + capacity, and the cast had to be kept for identical flows). +11. **Check every hash map for lookup-only use, including first-occurrence numbering.** The ensemble of clusterings + numbered each distinct (cluster a, cluster b) pair at its first occurrence via an `unordered_map`; a stable bucket sort by + the first id plus a stamp array over the second gives the same numbers without hashing (coarsening of the strong social + configuration -16%, one job -20%). A distribution's output can also be replaced by its closed form when the library's + algorithm is known: libstdc++'s `uniform_int_distribution(0,1)` on `mt19937` returns the top bit of one draw. + +12. **Look for work whose result is discarded, not for constant factors.** After six micro-changes in a row were neutral + (0.997 to 1.031), a cost-center review asked, per configuration, which work is redundant (recomputed, rebuilt, or thrown + away). Instrumenting the flow refinement showed that 99.6% of max-flow solves end with a cut equal to the current one, after + which the caller undoes everything. The max flow is bounded by the capacity U of the edges between the two flow regions; when + U equals the current cut, push-relabel can stop the moment the sink has received U and skip the source-set search. Exact for + the configurations without most-balanced cuts (no random draws in that path): flow time -15% (eco) and -21% (esocial), + whole benchmark -1.4% geomean. Where equal cuts with better balance are accepted (most-balanced cuts), the same shortcut is + not exact. +13. **Reuse a pointer the previous lookup already produced.** The boundary update looked up the block pair twice per edge + (once inside the node-set update, once for the edge cut); keeping the pair entry from the first lookup cut quotient-graph + refinement by 3.6% and k-way FM by 5%. Fusing a counting pass with the construction pass of the flow network (one scan of + each region node instead of two) cut flow time by 4%. +14. **Watch for k*k structures built with a large k.** A flat block-pair index of size k*k sped up the boundary lookups, + but the same class is also constructed during contraction with k = number of coarse nodes, which made one job 37% slower + (quadratic memory fill). Allocating the index lazily on first use and only for k <= 1024 kept the gain (-2%). +15. **When bit-identity is relaxed to "same quality", drop work that only feeds random draws.** Later in the session the owner + allowed changes that alter random decisions if the geometric mean cut over three seeds stays within 0.1% of the original + (plus balance and no crashes). Three changes then paid: an array-plus-position-index boundary set instead of an iterated + `unordered_map` (quotient-graph refinement -12%, eco -20%); no gain computation for k-way neighbors that are already moved and + no longer queued (their gains were computed only so the tie-breaking draws stayed in sync; multitry -10%); and a per-round + incremental connectivity cache (block, weight) per node, updated on each move, for nodes of degree >= 32 (multitry -3% + overall, -50 to -75% on social graphs). Quality ratios stayed at 0.9998 to 0.9999. +16. **Per-instance state in an object created per local search costs O(n) per search.** The first version of that cache resized + an n-sized header array in each refinement core, and the core is constructed once per localized search; road graphs became + 7% slower with identical output. A static store with a global round stamp (never reset, so stale entries from other levels + can never match) removed it. +17. **Defer a side structure that the search never reads, and replay only the kept moves.** In both the k-way and the 2-way + FM search every tentative move updated the block-pair boundary sets, and every move rolled back after the best prefix + updated them again; the searches themselves only read block weights and counts. Undoing the tail by resetting partition and + weights only, then rolling the kept prefix back and replaying it in order through the same boundary updates, cut multitry + time by 36% (total -10%) and 2-way refinement on the eco/esocial configs by 43 to 50% (total -11%). A debug build checked, + every 20th to 50th search, that every cut-edge endpoint is in its pair boundary and every pair cut equals the real cut. The + boundary order changes, so this is a quality-rule change; the cut ratio went from 1.00049 to 0.99929. +18. **Visit only nodes whose neighborhood changed in label propagation.** A per-node active flag (cleared on visit, set for all + neighbors when a node changes cluster) plus a stop when an iteration changes nothing cut coarsening of the social configs by + 30% (quality rule, cut ratio 0.9998). Re-activating undecided or size-blocked nodes as well made it slower and worse. +19. **Move deterministic per-graph work out of repetition loops.** The clustering ensemble runs up to 18 label propagations on + the same graph and sorted the nodes by degree in every run; sorting once per ensemble and numbering clusters through an + array instead of an `unordered_map` cut ssocial coarsening by 15% with identical output. +20. **Per-search hash maps over node ids can be a pooled stamped array.** The moved-node table and the heap's node-to-slot index + were open-addressing maps created per local search (grown from 32 slots each time). A per-node array with a generation + stamp, taken from a pool of four stores (two FM heaps and a moved table can be alive at once; a hash map is the fallback when + the pool is empty), made creation O(1) and lookups direct: multitry -19%, quality-graph refinement -11%, identical output, + peak memory +70 MB on the 2.2M-node graph. Together with keeping the push-relabel discharge's distance and excess in locals + (stores through edge records forced reloads; flow -4%) this gave total -7%. + +21. **Measure where an algorithm's work goes, not just which lines are hot.** Counting push-relabel discharges after the last + change of the sink's excess showed that 35 to 67% of them (road and social stripes) only returned excess to the source: + KaHIP's single-phase FIFO lets nodes with label >= n creep upward one relabel at a time. A textbook two-phase solver (maximum + preflow on labels < n, then excess return with labels n + exact distance to the source) cut flow time by 12.5% and total + time by 5.6%. A line profile had only shown "discharge" and "global relabeling" as hot. +22. **The minimal source side is the same for every maximum flow; use that to change solvers exactly.** When the caller only + uses the set of nodes reachable from the source in the residual graph (KaHIP without most-balanced cuts), any correct + maximum-flow algorithm gives bit-identical output. Even the second phase can go: on the reversed network a maximum preflow + from the original sink already fixes the cut closest to its sink, and the nodes that can reach the original source there + are exactly that set. Where the caller reads the flow itself (most-balanced cuts from residual SCCs), a different maximum + flow is a quality-rule change; here it passed at cut ratio 0.99944. +23. **Check that the profiler records full stacks before trusting inclusive shares.** The LD_PRELOAD sampler had its backtrace + depth fixed at 8 frames, so inside deep recursion (w-cycles plus recursive bisection in initial partitioning) the outer + frames were missing: initial partitioning showed 12% of samples while its timer said 53% of the run. Leaf shares were + right; inclusive shares of outer frames were not. + +24. **Try a different max-flow algorithm in place, on the real flow problems.** Flow refinement solves many small, + shallow problems (a stripe around the cut between two big terminals). Boykov-Kolmogorov, with search trees grown from both + terminals and reused across augmentations, beat the tuned two-phase push-relabel by 17% on whole runs (d_flow 0.58x with + the next item). The source tree at termination is exactly the set reachable from the source in the residual graph, so callers + that use only the minimal source side stay bit-identical; most-balanced cuts read the flow and become a quality-rule change. +25. **Model an undirected edge as one edge pair with capacity w both ways.** Two directed edges with zero-capacity partners + (4 residual edges) become one pair (2 residual edges). For the residual relation used by most-balanced cuts, "f1 < w or + f2 > 0" over the old pair equals "net f < w" over the new one, so only edge multiplicities change. +26. **Filter before sorting.** The gpa matching permuted and sorted all 2m directed edges and then skipped the ones the grow loop + cannot use (back direction, rating 0, weight bound). Filtering first halves the sort and needs fewer random draws; the order + of the usable edges changes with the tie-break draws, so it is a quality-rule change (coarsening 0.59x on strong). +27. **Memoize a quadratic check while its inputs are frozen.** Applying a flow cut updated the boundary per stripe node and per + neighbor in another block by rescanning that neighbor's adjacency for a neighbor in the old block (deg x deg on hubs; 18% of + a web-graph run). The partition does not change during the apply loop, so the answer per (node, block) can be cached with a + stamp for the batch: identical output, one web graph 3.1x faster, total -1.9%. The shortcut "update only moved nodes" was not + identical: the calls for unmoved nodes also clean up boundary entries. + +28. **Put what one lookup reads into one cache line.** A stamped per-node map kept stamps and values in two arrays, and label + propagation kept the vote counter and the cluster size in two; each lookup touched two random lines. One array of + {stamp, value} and one of {votes, size} gave -1.1% total (coarsening 0.97x) with identical output. The binary heap without + its separate element array (entries {key, node}, a stamped map to the heap position, hole-based sifts with the same + comparisons, so equal keys end at the same positions) gave -1.8% total (multitry 0.956x). +29. **The same std::sort over a smaller, self-contained element type.** std::sort's steps depend only on comparison results, so + sorting (degree, node) entries with a comparator on the degree gives the same order as sorting node ids with a comparator + that reads two degrees from the graph: identical output, 1% on social configurations. The same idea was neutral 30 + experiments earlier; it paid once the label propagation around it had become faster. + +30. **Profile the configurations with the worst ratio, not the largest jobs.** With a geomean over jobs, the small jobs of + the fast configurations set the metric. Their profile showed costs outside the partitioner proper: the final cut, boundary + and communication-volume report in three graph passes (one pass: -1.2% geomean, identical stdout) and zero-filling of + memory that is overwritten anyway. +31. **Do not zero what is about to be written.** basicGraph value-initialized m edges and m edge ratings in start_construction, + and contraction sizes them by the finer graph's m and trims later. A default-init allocator for the two edge arrays, with + new_edge writing the zero weight and rating itself, keeps every created edge identical and never touches the trimmed tail + (fast 0.93 to 0.97 on small jobs); the same for the graph-file and partition-output buffers. Together -1.2% geomean. + +32. **Lay the data out in visiting order when the same sweep repeats over the same graph.** ssocial coarsening intersects up + to 18 label-propagation clusterings per level, each a 10-iteration sweep in the same degree order. A copy of the graph + stored in that order, built once per level and shared by all clusterings, makes the node and edge reads sequential. The + cluster labels become positions; they are only compared for equality and renumbered by first occurrence in node order, so + the output stays identical. 3-seed A/B 0.923 on ssocial jobs (web-Google 0.82), -3.5% total together with two + passengers. Five earlier prefetch variants on the scattered layout had only hidden the latency. +33. **Price a parameter reduction per bucket, then keep the cheap buckets (rule b).** Halving the number of ensemble + clusterings for every k passed the gate exactly at the threshold (Q 1.00100) and broke the per-configuration guard + (ssocial 1.0051). The same gate data per k bucket showed where the quality went: k=8 cost 0.9% for little time, k=2 cost + 0.17% for a 2x to 5x speedup (fewer clusterings give coarser intersections and far fewer levels). Keeping only k < 8 (Q + 1.00050) and then k >= 32 (Q 1.00078) took -8% total; a 6-seed check on the final binary (Q 1.00030, all configurations + within 0.35%) covered the bias of choosing buckets from gate data. Measure where the search ends before cutting a search + limit: 95 to 99% of the localized multitry calls stop after their first round, so the round limit is not a lever. +34. **Replicate std::sort when ties must keep their order, then make the replica faster.** std::sort's result depends only + on the outcomes of its comparisons, so a copy of libstdc++'s introsort (same pivot choice, depth limit, insertion-sort + threshold) gives the same arrangement, ties included. In the copy, the Hoare partition collects the positions of the + elements to swap blockwise without branches (BlockQuicksort style) and then swaps the same pairs in the same order: + identical output, 0.66x sort time on the edge-rating and degree-order sorts (2.5 to 6.6% of the profiled jobs). Together + with reusing the label-propagation setup of the coarsest graph across initial-partitioning repetitions: -2.0% geomean. + +## Experiment Data + +| exp | change | geomean vs original | verdict | +|-----|--------|---------------------|---------| +| 001 | baseline | 1.0000 | | +| 002-004 | timers, fast IO, open-addressing lookup maps | 0.8676 | keep | +| 005 | global small-object allocator | 0.8390 | keep | +| 006 | CSR flow graph + hashed flow node map | 0.8106 (with 007) | keep | +| 007 | raw-pointer label-propagation loop | (combined) | keep, prediction refuted (3% vs 10-25%) | +| 008 | pointer cache in front of the block-pair map | 0.8076 | discard (below 1% floor) | +| 009 | exact replica of `unordered_map` iteration order | A/B 0.3-1.5% slower | discard | +| 010 | incremental neighbor degrees in 2-way FM | **0.7489** (total 0.7868) | keep | +| 012 | reuse BFS depth array in flow refinement | 0.7322 (with 013) | keep | +| 013 | sort (rating, edge) pairs in GPA matching | **0.7322** (total 0.7659) | keep | +| 014 | reuse multitry start-node gains, replay tie draws | A/B 0.995 | discard | +| 015/019 | skip repeated no-op boundary updates of the moved node | 0.7309 (with 018) | discard (below floor) | +| 016 | k-way gain loop over raw arrays | A/B 1.007 (slower) | discard | +| 017 | witness hint for the "still adjacent to old block" scan | 0.7331 (with 018) | discard (+1-3% elsewhere) | +| 018 | residual graph reverse-edge lookup via smaller adjacency | 0.7309 (with 019) | discard (web graphs only) | +| 020 | prefetch in the label-propagation sweep | 0.7079 (total 0.7463) | keep | +| 021 | prefetch in the k-way move loop | A/B 1.019 (slower) | discard | +| 022 | push-relabel char flags, one edge pointer per node | **0.6986** (total 0.7317) | keep | +| 023 | push-relabel queues as arrays | A/B 0.998 | discard | +| 024 | `noexcept` hash functors (no cached hash codes) | A/B 0.994 | discard | +| 025 | `nextBool` as the top bit of one `mt19937` draw | 0.6924 (with 026, 027) | keep | +| 026 | ensemble numbering without a hash map | **0.6925** (total 0.7172, confirmed by the 030 run) | keep | +| 027 | SCC with vector stacks in most-balanced min cut | flow time 1.005 | discard | +| 028 | degree node ordering via packed keys | coarsening 0.992 | discard (below floor) | +| 029 | cluster-id remap via array | coarsening 0.994 | discard (below floor) | +| 030 | open-addressing index in the bucket queue | multitry 1.007 | discard (A/B gain not reproduced) | +| 031 | edge cut through the already looked-up block-pair entry | 0.6838 (total 0.6989, with 029) | keep | +| 032 | two-entry block-pair cache | A/B 0.997 | discard | +| 033 | one scan per flow-region node (edges and outer-boundary flag) | **0.6849** (total 0.6964) | keep | +| 034 | quotient graph via counting sort | A/B 1.001 | discard | +| 035 | open-addressing start-node dedup sets | A/B 1.004 | discard | +| 036 | label-propagation prefetch one stage deeper | A/B 1.031 (slower) | discard | +| 037 | reverse capacity stored on each flow edge | A/B 0.999 | discard | +| 038 | reuse start-node gains in the k-way queue setup (replay tie draws) | 0.6845 (total 0.6923; k-way FM 0.976) | discard (below floor) | +| 039 | 038 plus gain cache across searches (version stamps) | A/B 0.994 to 1.012 | discard | +| 040 | stop push-relabel when the sink holds the region-cut capacity equal to the current cut | **0.6752** (total 0.6879) | keep | +| 041 | lazy flat k*k block-pair index (k <= 1024) | 0.6723 (total 0.6817) | carried (below floor alone) | +| 042 | 041 + 038 (start-node gains reused) | **0.6694** (total 0.6748) | keep | +| 043-045 | residual back-edge count once per node, weighted degrees once, dense flow node map | **0.6643** (total 0.6613) | keep | +| 046 | (quality rule) boundary set as array + position index | **0.6327** (total 0.6047), cut ratio 0.99986 | keep | +| 047 | (quality rule) no gains for moved, unqueued k-way neighbors; no tie replay | **0.6279** (total 0.5844), cut ratio 0.99981 | keep | +| 048 | (quality rule) incremental gain cache for all nodes, per-instance arrays | 0.6261 (total 0.6017) | discard (roads 1.2 to 1.3x) | +| 049 | 048 with a degree >= 32 gate and a shared static store | **0.6245** (total 0.5783), cut ratio 0.99978 | keep | +| 050 | 049 without the gate | 0.6247 (total 0.5834) | discard (meshes gain, roads lose) | +| 051 | (quality rule) active-node label propagation | **0.5929** (total 0.5722), cut ratio 0.99980 | keep | +| 052 | 051 plus re-activation of undecided nodes | 0.6210 (total 0.5752), cut ratio 1.00079 | discard | +| 053 | (quality rule) k-way rounds: boundaries updated only for kept moves | **0.5739** (total 0.5140), cut ratio 1.00049 | keep | +| 054 | (quality rule) 2-way FM: boundaries updated only for kept moves | **0.5152** (total 0.4562), cut ratio 0.99929 | keep | +| 055 | degree order once per clustering ensemble, array cluster-id remap | **0.5060** (total 0.4447), identical | keep | +| 056 | degree sort on packed (degree, node) keys | A/B 1.004 | discard | +| 057-060 | pooled stamped per-node maps (moved table, heap index), push-relabel discharge with locals | **0.4819** (total 0.4124), identical | keep | + +| 061-063 | counting buckets for the cluster-contraction quotient graph, flow-graph buffers reused across flow problems | **0.4769** (total 0.4077), identical | keep | +| 062 | 24-byte flow edges with the partner's capacity (global relabeling without the partner read) | A/B 0.988, same as 061 alone | discard | +| 065 | gpa matching: grow paths over the sorted (rating, edge) pairs | A/B no gain | discard | +| 064-067 | two-phase push-relabel (quality rule for most-balanced cuts), node-state record, k-way neighbor prefetch | **0.4637** (total 0.3850), cut ratio 0.99944 | keep | +| 068-070 | no phase 2 without most-balanced cuts (reversed preflow), pooled solver buffers, stamped start-node sets | **0.4571** (total 0.3771), identical | keep | +| 071-073 | Boykov-Kolmogorov max flow, one edge pair per undirected stripe edge, no unused source set | **0.4219** (total 0.3222), cut ratio 0.99932 | keep | +| 074 | SCC stacks on std::vector instead of std::deque | A/B 1.004 | discard | +| 075-077 | gpa: filter usable edges before permute and sort (quality rule), prefetched random permutation, grow-loop prefetch | **0.4055** (total 0.3133), cut ratio 1.00045 (six seeds 0.99988) | keep | +| 078 | memoized neighbor-in-block test while applying a flow cut | **0.4025** (total 0.3075), identical | keep | +| 079-080 | more prefetches in the k-way move and label-propagation loops | A/B 1.007 and 1.014 | discard | +| 081-082 | BK: reverse residual read from the edge itself for symmetric stripe pairs; heap without element array | **0.3965** (total 0.3020), identical | keep | +| 083 | stripe membership bitmap in the flow-graph build | A/B 1.002 | discard | +| 084-085 | {votes, size} cluster slots in label propagation, {stamp, value} slots in stamped maps | **0.3927** (total 0.2986), identical | keep | +| 086, 090 | k-way and 2-way FM moves templated on the queue type (no virtual calls on the heap) | A/B 0.998 each | carried | +| 087 | label-propagation active flags as a bitset over visiting positions | A/B 0.998 | discard | +| 088 | degree sort over (degree, node) entries, same comparisons | A/B 0.989 | carried | +| 089 | two-stage prefetch in label propagation (repeat of a recorded negative) | A/B 1.039 | discard | +| 086-090 | full run of the three exact passengers | 0.3897 (total 0.2962), identical, below the 1% floor | carried | +| 091-092 | quotient graph of a clustering and extracted blocks built with their real edge counts (with 086-090) | **0.3842** (total 0.2939), identical | keep | +| 093 | matching contraction: coarse edges collected first, real edge count | A/B 1.020 | discard | +| 094 | final cut, boundary nodes and max communication volume in one pass | **0.3797** (total 0.2933), identical | keep | +| 095 | cache of degree orders keyed by the verified degree sequence | A/B 0.998 | discard | +| 096, 098 | graph edge arrays and IO buffers not zeroed before they are written | **0.3752** (total 0.2920), identical | keep | +| 097 | matching contraction without the precomputed coarse-target array | A/B 1.009 | discard | +| 099 | label-propagation prefetch only for the next node if active | A/B 1.001 | discard | +| 100 | cluster contraction builds the quotient graph from the coarse mapping directly | A/B 0.997 | carried | +| 101 | random matching built without zero fill (reserve + push_back) | A/B 1.000 | discard | +| 102 | one boundary insert per (node, target block) instead of per cut edge | A/B 1.003 | discard | +| 103 | exact replay of recorded gain events for hub nodes in the multitry start-node scan | A/B 1.038 | discard | +| 104 | most balanced minimum cut on arrays from the flow graph, vector-stack SCC, bucketed components | A/B 0.996 | carried | +| 105 | ensemble label propagation on a copy of the graph in visiting order (with 100, 104) | **0.3712** (total 0.2818), identical | keep | +| 106 | parameter reduction: ensemble clusterings halved for every k | gate Q 1.00100, ssocial 1.0051 over the guard | discard | +| 107 | parameter reduction: ssocial label-propagation iterations 10 to 5 | screen Q +0.6%, time 0.96 | discard | +| 108 | parameter reduction: 9 ensemble clusterings for k < 8 (were 18, 17) | **0.3657** (total 0.2738), Q 1.00050 | keep | +| 109 | parameter reduction: 2 ensemble clusterings for k >= 32 (were 3) | **0.3601** (total 0.2588), Q 1.00078, 6-seed 1.00030 | keep | +| 110 | parameter reduction: strong multitry search limit alpha 10 to 5 | screen Q +0.15%, time 0.95 to 0.97 | discard | +| 111 | single-clustering label propagation on the visiting-order copy | A/B 0.987 | carried | +| 112 | exact std::sort replica with a blockwise branchless partition (edge-rating and degree-order sorts) | A/B 0.992 (with 111) | carried | +| 113 | initial-partitioning repetitions reuse the coarsest graph's degree order and visiting-order copy | A/B 0.993 | carried | +| 114 | most balanced cut: topological sort on CSR arrays (with 111 to 113) | **0.3527** (total 0.2557), identical | keep | +| 115, 117, 118 | software prefetch in matching contraction, flow-network build, 2-way FM queue init and boundary updates | A/B 0.996, 0.999, 0.995; full run 0.3541 | discard | +| 116 | revert of 114 (its own A/B was 1.008) | A/B 1.003 | discard | +| 119 | prefetch in 2-way FM neighbor degree updates | A/B 1.000 | discard | +| 120 | two-stage prefetch in label propagation | A/B 1.017 | discard | +| 121 | label propagation gathers target blocks once per node into a buffer | A/B 1.018 | discard | +| 122 | uncached k-way gain loop with one slot reference and one weight load per edge | A/B 1.003 | discard | +Final result (converged after 122 experiments; the last 8 brought no full-run gain): the original and the final binary run +back to back give a geomean ratio of 0.3517 (2.84x faster) and a total of 0.2562 (6503 s to 1666 s, 3.90x), 0 incorrect. +Cut quality against the original: Q 1.00078 on the three-seed gate and 1.00030 on seeds 4 to 6. + +Every kept step: 0 incorrect of 537 jobs; bit-identical steps also crosscheck OK on 864 runs, quality-rule steps passed the +three-seed cut gate (A/A noise sd 0.0011). + +## What Didn't Work + +- **Trusting callgrind instruction shares for small hash maps.** The block-pair map took 7% of instructions and the boundary + set about 8%. Removing most of that work changed wall time by less than 1%, because these maps are tiny, stay in L1, and the + out-of-order core hides the hash and the modulo. Pre-filter such ideas with a quick A/B timing (two binaries, alternating, + 3 reps, a handful of targeted jobs) before spending a full benchmark run. +- **Replaying libstdc++'s `unordered_map` iteration order** with flat arrays (head insertion per bucket, rehash via the + library's own `std::__detail::_Prime_rehash_policy`, erase bucket fix-ups, `clear()` keeping the bucket count). It is exact + (partitions identical), so this is a usable trick when an iterated map must be replaced, but here it was slower than the + real map once allocation was cheap. +- **Fixes for one graph family inside the hottest function.** A hint that removed hub scans gave -10% on a web graph, but + d_multitry rose 1 to 3% on 17 of 20 graphs where the new branch was never taken (code size and layout in the k-way move + path). Four variants of the same idea all showed it; stop after two A/B rounds. +- **Below about 2%, component timers are noise.** Phase timers of untouched components moved by +-2% between runs, so a + component gain must be clearly larger than that, and the overall metric must still clear the floor. +- **Rewriting a latency-bound loop.** The k-way gain loop over raw arrays was 0.7% slower: the compiler already hoisted the + loads, and the time is in dependent partition loads and unpredictable tie branches. +- **Micro-optimising the label-propagation inner loop** (raw pointers, hoisted flags): only about 3%. The loop is bound by + random memory access, not by instructions. +- **Removing the cached hash code** (`noexcept` hash functors) keeps the order but gained only 0.6%; the cost of the iterated + maps is the chain walk and the modulo, not the node size. +- **Trusting a 7-job A/B for small gains.** An open-addressing index in the FM bucket queue showed 0.967 on one job in + the A/B and 1.017 on the same job in the full run; gains under about 3% in the A/B need the full run before they count. +- **Array-based queues in push-relabel** (instead of `std::queue`) were neutral: the queues are not where the time goes. +- **Deeper software prefetch.** Adding a second prefetch stage in the label-propagation sweep (cluster ids two nodes ahead, + the cluster weights one ahead) was 3% slower: the extra loops over up to 32 edges cost more than the misses they hide. +- **Caching gains across local searches.** A per-node version stamp (bumped for a moved node and all its neighbors) to reuse + gain computations across consecutive k-way searches almost never hit: the hubs that dominate the cost have a neighbor move + in nearly every search, and the stamping slowed the move loop. +- **Caching repeated flow problems.** Consecutive flow iterations on the same block pair never had identical regions (the + region search starts from randomly permuted boundary nodes), so there was nothing to cache; the waste was in the solve itself. + +- **Cache-everything for a per-node quantity that is read about once.** The incremental gain cache gave 0.84 to 0.86 multitry + time on FE meshes but 1.06 to 1.10 on road graphs (degree about 2.5): each node is queried about once per round, so the + build plus the per-move updates are pure overhead. A three-seed A/B on five jobs (two FE meshes, one road) predicted 0.963 + and the full run gave 1.005; pick A/B jobs per graph class in proportion to the benchmark. +- **Single-seed A/B under the quality rule.** Once random decisions may change, one seed of one job can be 1.2x slower or + faster from a different trajectory alone (144 strong 64: 1.21 on seed 1, 0.90 on seeds 2 and 3). Sum several seeds. +- **Keeping more nodes active in label propagation to buy back quality.** Re-flagging nodes whose vote was tied or blocked by + the size bound kept most nodes active (coarsening 1.21x) and the cut ratio got worse (ssocial 1.0047 to 1.0058), not better. +- **Packing the sort key next to the node id for a degree sort** gave nothing (A/B 1.004): one sort of n entries per level is + cheap next to the label propagation that follows; the same trick did pay for the edge-rating sort earlier (2m entries). +- **Reading per-config cut ratios from three seeds.** After 053 the eco ratio moved from 0.9998 to 1.0039; seeds 4 to 6 gave + 0.9938 and six seeds 0.9992. Per-config ratios over three seeds swing by about +-0.4%. +- **Compacting the residual edges to give global relabeling the partner's capacity locally** (24 instead of 32 bytes, no + partner read per scanned edge) gained nothing on top of the other changes (A/B 0.988 either way). +- **Small prefetches and carried-over sort keys in coarsening matching** (gpa grow loop over the sorted pairs, prefetch of + neighbor adjacencies in the k-way gain update) gave 0 to 0.5% each: below the 1% floor, only worth carrying as passengers. +- **More software prefetch in loops that already have some.** Prefetching partition entries, heap slots and stamps of the + first 32 neighbors in the k-way move, and cluster ids 16 edges ahead in label propagation for high-degree nodes, were both + slower (1-seed A/B 1.007 and 1.014 cumulative, 7 of 8 jobs worse): the hardware prefetcher or the existing per-node prefetch + already covers these accesses. +- **Plain std::vector instead of std::deque for SCC stacks** gave nothing (A/B 1.004); the SCC is a small part of the flow time. +- **Skipping inactive nodes word-wise in label propagation.** With 66% of scanned positions inactive, a bitset over visiting + positions (with a node-to-position map) skips them without touching the node, but finding the next three active positions for + the prefetches cost as much as the skipped tests (A/B 0.998). +- **Re-trying a recorded negative.** A two-stage prefetch in label propagation (cluster ids two nodes ahead, cluster slots one + ahead) was 3.9% slower, as this entry already said for the same idea; read the negatives list before starting an idea. +- **A stripe-membership bitmap in the flow-graph build** (n bits instead of the partition array) was neutral (A/B 1.002): the + cost on hub nodes is the edge stream itself, not the partition lookups. +- **Restructuring the matching contraction.** Collecting coarse edges first to build with the real edge count (A/B 1.020) + and reading coarse targets inside the visit loop instead of a precomputed m-sized array (A/B 1.009) were both slower: the + separate sequential pass hides latency that the dependent load in the loop exposes. Removing the zero fill around it paid. +- **Caching degree orders across calls** (keyed by the full degree sequence, verified on reuse) was neutral (A/B 0.998): the + repeated sorts are not on identical graphs often enough. +- **Replaying recorded gain events for hub nodes** in the multitry start-node scan (exact, A/B 1.038): one of a hub's many + neighbors moves in almost every round, so the recorded events are invalidated before reuse and recording is extra cost. +- **Removing per-job fixed costs below 0.5%** (zero fill of two n-sized arrays in the random matching, one boundary insert per + node and block): all at 1.000 +- 0.003; the total is set by the large jobs. +- **Cutting the label-propagation iterations or the multitry search limit** under rule (b): 4 to 5% faster on the affected + configurations for 0.15 to 0.6% of cut, far fewer seconds per unit of quality than the clustering count. +- **Summing sub-1% A/B results.** Three exact prefetch changes measured 0.996, 0.999 and 0.995 in targeted 3-seed A/Bs, but + the full run with all three was 0.4% slower than without them. Below 1% per change, a targeted A/B is noise; only a full + run decides. Late attempts on the hottest loops (another label-propagation prefetch, a target-block buffer, a hand-hoisted + gain loop) were 0 to 2% slower: those loops are bound by random block lookups, and the compiler already hoists the loads. + +## Environment + +C++11/GCC 13, Release with KaHIP's default flags (no flag changes allowed), single-threaded, 16-core x86 machine with one job +per physical core, graphs of up to a few million edges. diff --git a/claude-plugin/skills/aae/knowledge-base/037-cascading-reduction-rules.md b/claude-plugin/skills/aae/knowledge-base/037-cascading-reduction-rules.md new file mode 100644 index 0000000..1158b4b --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/037-cascading-reduction-rules.md @@ -0,0 +1,120 @@ +# Cascading data-reduction rules instead of batch reductions on a frozen graph + +## Problem + +Exact minimum cut on undirected weighted graphs (VieCut's Nagamochi-Ibaraki +capforest solver, with a Padberg-Rinaldi (PR) reduction pre-run). Metric: +wall-clock time to the exact cut over a 53-instance benchmark (33 generated +Chekuri-Goldberg-Karger-Levine-Stein families plus 20 real-world k-cores of +large web and social graphs), under a hard invariant that the computed cut +value never changes. On the dense, cycle-structured generated families the +solver was slower than a classical 1997 sequential implementation. Profiling +showed the capforest contracting roughly one vertex per round, i.e. quadratic +in the number of rounds, on graphs where a reduction rule should have collapsed +them in a single pass. + +## What Worked + +Reduction rules (Padberg-Rinaldi here, but the lesson generalizes to any data +reduction / kernelization rule set) are usually written as a **batch**: scan +the whole graph once, mark everything the rules fire on, contract, repeat. On +these instances the batch found almost nothing, because each PR test compares an +edge weight against vertex **degrees**, and a test only becomes true *after* a +neighbouring contraction has raised those degrees. A batch measures a frozen +graph, so the first pass fires on nothing and the phase stops. + +The fix is to **cascade**: contract on the spot and immediately re-read the +degrees the contraction just changed, pushing the merged supervertex back onto a +worklist so its neighbours are retested. One reduction enables the next, and +whole cycle families collapse in a fraction of a second where the capforest +needed a round per vertex. This is exactly how the original 1997 implementation +rode the chain (its internal passes remove about 87% of a million-vertex random +graph); a later batch reimplementation had quietly lost that property, and that +single lost property was the whole performance gap on those families. + +Two implementation choices make the cascade cheap enough to leave switched on: + +1. **Lazy adjacency over a union-find.** Hold each supervertex's neighbourhood + as a plain vector of `(target, weight)` and let entries go stale. Redirect a + target through the union-find only when it is read, and combine parallel + edges only when that vertex is tested, via a stamped (epoch-counter) + aggregation pass over its list. A merge splices the shorter list onto the + longer and updates one degree arithmetically (`deg = deg_u + deg_v - 2w`); it + touches neither the neighbours nor the longer list. An earlier version kept a + hash map per supervertex, and the maps *were* the cost: a million + supervertices meant a million allocations per call, and the run came out + slower than leaving the reductions off entirely. +2. **Run it only where it pays** (sparse rounds) and **retire it** once a pass + stops being productive, so it never taxes the rounds it cannot help. + +## Experiment Data + +Whole-campaign result. The cascade was one of about a dozen kept changes and the +specific one that closed the gap on the generated families. Exact algorithm, +geometric mean over the 53 instances, cut value identical on every instance: + +| configuration | before | after | speedup | +|---|---|---|---| +| 1 thread | 25.0 s | 2.80 s | 8.9x | +| 32 threads | 5.76 s | 0.65 s | 8.8x | + +I do not have an isolated single-change microbenchmark for the cascade alone, +but it is the change that turned the cycle families from "slower than the 1997 +sequential code" into sub-second, by replacing an O(rounds^2) capforest phase +with a near-linear reduction that removes most of the graph before the capforest +runs at all. + +## What Didn't Work / Gotchas + +- **Per-supervertex hash maps for the neighbourhood.** Correct, but slower than + running no reductions at all at scale. The lazy-vector plus stamped-aggregation + representation is what made the cascade a net win rather than a net loss. If a + reduction phase is slower than skipping it, suspect per-element allocation + before you suspect the rules. +- **Soundness under equality tests.** The PR tests fire at equality, and + equality is only sound if the degree of every vertex the cascade removes is + first recorded as a cut (an upper bound): contracting `u` into `v` because + twice the edge weight reaches `deg(u)` argues that `u` can cross the cut, and + that argument fails for the cut that isolates `u`. Maintaining that bound is + part of correctness, not an optimization, and it is easy to drop when you port + a batch rule to an incremental one. +- **Lazy aggregation combined with a min.** The test that sums + `min(w2, w3)` over the common neighbourhood can be fed stale entries that split + one true edge weight across several records; taking the per-entry minimum + overstates the sum (`min(3,5) + min(3,5) = 6` where the true `min(6,5) = 5`), + which contracts an edge the rule does not justify. Aggregate each + neighbourhood *before* taking any minimum. + +## Code Example + +Worklist loop, pseudocode: + +``` +worklist = all vertices +while worklist: + v = worklist.pop() + if find(v) != v: continue # already merged away + aggregate v's neighbourhood by current representative (one stamped scan) + for (x, w) in neighbourhood(v): + if pr_test_fires(w, deg[v], deg[x], shared_nbhd): # PR 1-4, at equality + bound = min(bound, deg[small]) # record removed degree as a cut FIRST + splice smaller adjacency list into larger + deg[big] = deg[v] + deg[x] - 2*w # arithmetic, no rescan + union(small, big) + push big back onto worklist # <-- let the change cascade + break +``` + +The one line that distinguishes this from a batch is `push big back onto +worklist`: the contraction that just happened is allowed to re-trigger the rules +on the vertex it changed. + +## Environment + +C++17, 64-core AMD EPYC. The technique is language-agnostic and applies to any +solver whose local reduction rules have preconditions that depend on the graph +state the reductions themselves change: minimum cut, vertex-cover and other +branch-and-reduce kernelizations, and constraint propagation. Developed with +Agentic Algorithm Engineering while speeding up VieCut's exact minimum cut; the +study is described in the AAE paper (13th DIMACS Implementation Challenge, 2027) +and the released code is KaHIP/VieCut. diff --git a/claude-plugin/skills/aae/knowledge-base/038-kahip-quality-at-fixed-time-budget.md b/claude-plugin/skills/aae/knowledge-base/038-kahip-quality-at-fixed-time-budget.md new file mode 100644 index 0000000..8e90521 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/038-kahip-quality-at-fixed-time-budget.md @@ -0,0 +1,155 @@ +# Improving the cut quality of a mature multilevel graph partitioner at a fixed time budget + +## Problem + +Improving the solution quality (edge cut) of [KaHIP](https://github.com/KaHIP/KaHIP), a mature multilevel graph partitioner, by adding a new preconfiguration `ultra` on top of the `strong` configuration. A previous AAE speed campaign had made KaHIP v3.26 about 2x faster than v3.25; the goal was to spend that saved time on better cuts. Constraints: imbalance 3% as in strong, every existing preconfiguration bit-identical to v3.26, and kaffpaE / KaBaPE (the memetic algorithms that call kaffpa as their operator) should gain too. + +- **Metric Q** = geometric mean over jobs of (mean cut of the candidate over the seeds + 1) / (mean cut of v3.26 strong over the same seeds + 1). Lower is better. For kaffpaE the same ratio at a fixed `--time_limit` is called QE. +- **Benchmark:** 73 graph x k jobs (k from 2 to 256) on 19 graphs (meshes, road networks, social and web graphs, 10k to 2.2M nodes). Tuning set: 53 jobs on 15 graphs, seeds 1-5. Held out until the end: 4 graphs (20 jobs) and seeds 7-9 on all jobs. A/A noise of Q measured with strong seeds 11-15 against 1-5. +- **Quality gate:** a change is kept only if Q improves against the current ultra base, measured in the same batch, with a consistent sign across seed batches or k, and no more partitions above the balance bound than strong. +- **Time gate:** per job, ultra must stay within the v3.25 strong time: median wall time <= 1.05 b + 0.05 s for every job (b = v3.25 strong time), and geometric mean and total ratio <= 1. +- **Identity gate:** every kept change passes a bit-identity check of all old configurations (2907 runs over all preconfigurations, k and eps) and of 28 CLI paths against v3.26. + +## What Worked + +The final ultra reaches **Q 0.9777 (2.2% smaller cuts than v3.26 strong)** on the tuning set and the same 0.9777 on held-out graphs, at 2.0x the time of v3.26 strong and below the time of v3.25 strong on every job. Below are the techniques, in order of impact, with enough detail to reproduce. + +### 1. Jet negative-gain label propagation before FM (~1.4%, 6.5% on social graphs) + +Jet (Gilbert et al., SIAM J. Sci. Comput. 2024) is a refinement built for GPUs, but it transfers to a sequential multilevel partitioner as a step on every uncoarsening level, run before the existing k-way FM. Per iteration: + +1. **Candidate moves:** every unlocked boundary node picks its best adjacent block, and is a candidate if the gain is non-negative or the loss is small: -gain < c * (connection to its own block), with c = 0.25 on unit node weights (finest level) and 0.75 on weighted (coarse) levels. Negative-gain moves are allowed; this is what lets it leave the local minima that FM converges to. +2. **Afterburner:** recompute each candidate's gain assuming that the neighbors with a higher priority (better gain, ties by node id) have already moved, and keep only candidates whose recomputed gain is non-negative. This filters out conflicting moves without sequential dependencies. All kept moves are applied at once, and moved nodes are locked for the next iteration. +3. **Rebalancing:** when a block is over the bound, the iteration rebalances instead of moving: it moves the nodes with the smallest loss per weight out of every overloaded block. +4. Keep the best balanced partition seen; stop after 12 iterations without improvement by more than a factor phi = 0.999. + +Then FM and the pairwise flow refinement run as before, on a better starting point. + +The decisive detail: **the rebalancer must consider interior nodes, not only boundary nodes.** A version with a boundary-only rebalancer (to save time) lost nearly the whole gain (Q 0.9977 instead of 0.9852), because when the boundary nodes run out, the overloaded block cannot shed weight and Jet's negative-gain moves are rolled back. The final version keeps the incremental cut and boundary bookkeeping and replays the best partition onto KaHIP's boundary structure node by node. + +Jet is weakest on 2D meshes and web graphs (worst jobs 1% worse), as the paper reports, and strongest on social graphs. + +### 2. The unlocked flow model F_H in pairwise flow refinement (~0.3%) + +KaHIP's flow refinement builds a region around the cut of two blocks and computes a min cut. In the classic model, the nodes on the region boundary are fixed to their blocks. The unlocked model of Heuer, Sanders, Schlag (ACM JEA 2019) instead connects each border node to the source or the sink with the weight of its edges that leave the region. The min cut then finds the same cut or a better one, and it may move border nodes too. The change is small (one network construction) and helps most at large k (Q 0.9939 at k = 64). It costs 8% time and gave Q 0.9974 over 10 seeds, with the same sign in both seed batches. + +### 3. Spend saved time on global work, not local work (~0.7%) + +Before building new components, single parameter levers on v3.26 strong showed where time pays: + +- **Global work pays:** 4 V-cycles Q 0.9967; 8 V-cycles 0.9938 (but 10 jobs over budget); 256 initial partitioning repetitions 0.9958. +- **Local work does not pay:** 30 FM multitry rounds Q 0.9999; FM adaptive stopping limit x30: 1.0004; flow regions 16x larger: 1.0001. + +The kept setting is 3 V-cycles for k = 2 (4 went over budget on one job) and **best of 2 independent runs for k > 2**. Later, more V-cycles of one run were tested instead of best of 2 at the same time: 0.5% worse for k > 2. So two independent runs beat a longer single run here. The likely reason is that the multilevel hierarchies differ, which gives diversity that more refinement on one hierarchy does not. + +### 4. Memetic algorithm (kaffpaE): rank overloaded individuals behind balanced ones, and keep the input individual + +These are two small changes in the evolutionary layer. Both are needed once a stronger operator is plugged in: + +- **Imbalance penalty:** kaffpaE's fitness was the cut alone. Combine operators partition at a relaxed imbalance (up to 25%) and then refine at the real bound without enforcing it, so an overloaded individual with a small cut could win selection and the final choice. The fix sets the fitness of an overloaded individual to total edge weight + 1 + overload, so every balanced individual ranks before it. The final partition is rebalanced with the existing negative-cycle model if it is still overloaded. On the 8 eps 1 instances with invalid runs, the share of runs over the bound fell from 16 of 24 (v3.26) to 1 of 24, with equal cuts. +- **Keep the input:** when kaffpaE starts from an input partition, the input only seeded the individuals built from it, which are re-partitioned. So at eps 0 the result was worse than the input in 9 to 12 of 124 runs. Inserting the input unchanged as one more individual makes the result never worse than the input (0 of 20 reruns of the worse cases). +- **One partition per operator:** best of 2 is switched off inside kaffpaE (repetitions = 1), since the population already provides diversity and every second counts at a fixed time limit. + +With ultra inside, kaffpaE gains 0.5% to 1.6% over v3.26 strong at the same time limits, while the v3.26 speed alone gave kaffpaE 1.9x to 2.6x the rounds and almost no cut gain. + +### 5. Guard-field pattern plus identity audits (process, enabled everything above) + +Every new behavior is a field in the configuration struct with a default that reproduces the old code path. Only the new preconfiguration function sets it, and the new code lives in its own files. Each kept change then passed two audits: + +- **idcheck:** 2907 runs of all old preconfigurations, compared by partition hash against stored v3.26 hashes. +- **xpath:** 28 CLI paths (all programs and modes), 3469 runs, comparing output, return code, metric and balance. + +The audits found what review missed. They separated the paths that are nondeterministic in v3.26 itself (global multisection, wall-time limits) from real regressions. They also exposed a pre-existing flow refinement bug (an empty "no change" result read as "move everything", which kept a stale cut value). The bug was fixed in ultra, and later in strong and eco as a separate, measured change. + +**Pitfall:** configurations compose. Most modes call `strong()` before the chosen preconfiguration, so a field set in `strong()` must be reset explicitly in configurations that are built on top of it but must not inherit it. + +### 6. Held-out graphs and seeds + +All tuning used 53 jobs and seeds 1-5. Four graphs (one per class plus a random geometric graph) and seeds 7-9 were only run once, at the end. The held-out graphs gave the same Q (0.9777) and the new seeds a slightly better one (0.9729), so the candidates did not overfit the tuning set. Measuring every candidate against the current base **in the same batch** mattered as much: machine noise between batches was of the same size as most component effects (0.1% to 0.3%). + +## Experiment Data + +| Step | Q vs v3.26 strong (tuning set) | Time (geomean of the v3.25 strong budget) | Verdict | +|---|---|---|---| +| v3.26 strong | 1.0000 | 0.32 | baseline | +| + 4 V-cycles (k = 2), best of 2 (k > 2) | 0.9932 | 0.59 | kept | +| + unlocked flow model F_H | 0.9912 (0.9974 vs previous, 10 seeds) | 0.64 | kept | +| k = 2 back to 3 V-cycles (budget) | 0.9914 | 0.61 | kept | +| + Jet refinement | 0.9777 | 0.65 | kept | +| kaffpaE imbalance penalty, keep input | (kaffpaE only) | | kept | + +Final ultra: + +- **Quality:** Q 0.9777 on the tuning set (social 0.930, mesh 0.986, road 0.989; k = 2 0.954); 0.9777 on held-out graphs; 0.9729 on held-out seeds 7-9 (all 73 jobs). +- **Time:** 2.0x v3.26 strong (total 2.09x, max 2.9x). Against v3.25 strong: geomean 0.64, max 0.95, 0 jobs over budget. +- **Balance:** invalid runs ultra vs strong: 0 / 265 vs 3 / 265 (tuning), 1 / 100 vs 2 / 100 (held-out graphs), 1 / 219 vs 4 / 219 (new seeds). +- **Contributions:** best of 2 / V-cycles 0.7%, F_H 0.3%, Jet 1.4% (6.5% on social graphs). +- **kaffpaE / KaBaPE** at fixed time limits (1x, 3x and 10x the strong kaffpa time; 10 instances, seeds 1-5), QE of ultra vs v3.26 strong inside the same binary: + +| Mode, limit | QE ultra vs v3.26 strong | QE ultra vs v3.25 | Evolutionary rounds vs v3.26 strong | +|---|---|---|---| +| kaffpaE 1x | 0.9907 | 0.9932 | 0.96 | +| kaffpaE 3x | 0.9912 | 0.9845 | 0.87 | +| kaffpaE 10x | 0.9946 | 0.9896 | 0.79 | +| KaBaPE (eps 0) 3x | 0.9870 | 0.9890 | 1.02 | +| KaBaPE (eps 0) 10x | 0.9838 | 0.9876 | 1.05 | + +- **Experiments:** 45 measured candidates, 6 kept (the commits in the table), 39 discarded; one more (a combine operator) was dropped before it was built. The campaign used a written log instead of results.tsv, so the count comes from that log. + +## What Didn't Work + +- **Relaxed-balance k-way FM** (move a balance-blocked node to its best block with room, or overload temporarily and accept only balanced best states): v1 Q 1.028 to 1.061 (worse, up to 3.8x time); v2 with positive gains only: 1.0001, because the pairwise flows with most-balanced minimum cuts already move those nodes. +- **Randomized top-l move selection in FM** (diversity): l = 2 / 4 / 8 gave Q 1.0004 / 1.0036 / 1.0070, adaptive 1.0014; neutral to worse in kaffpaE; about 0.5% better only in KaBaPE at eps 0. +- **Larger flow regions:** 16x gave 0.13% at +40% time (11 jobs over budget), 32x the same 0.13% at +160% time (43 jobs over budget). +- **More F-cycle arms** (an arm on every level instead of every second): 0.26% for +34% time; 6 jobs over budget. +- **Jet parameter changes:** c = 0.5 Q 1.0006, patience 24 0.9995, phi 0.9999 1.0000, a second Jet pass after FM and flows 1.0000. The paper's settings were already the optimum; FM and flows absorb small extra Jet gains. +- **Weighted sparse tabu search** on the coarsest level (MITS-like, edge-work bounded): 0.6% to 1.1% better coarsest cuts at k = 256, but over 10 seeds Q 0.99966, inside the noise. Uncoarsening absorbs coarse-level gains. +- **KaBaPE relax-and-rebalance rounds** (local search at a relaxed bound, then negative-cycle rebalancing): mean cut 0.10% to 0.13% worse on 194 eps 0 instances; they cost 40% of the evolutionary rounds. +- **Two runs plus a combine** instead of best of 2: dropped before it was built, since recombination is the memetic algorithm's job and the single run should get better local search. +- **ILP region improvement** after the evolutionary search (Gurobi, KaHIP ilp_improve): regions up to 0.65M nonzeros are solved to optimality in 57 s; 1M hits the 600 s limit at a 0.075% gap; 2.5M to 5M leave 0.33% to 1.2% gaps. The region size has to be capped at about 1M nonzeros, and on near-optimal partitions 1 h per instance improved the cut by at most 1 edge. + +## Code Example + +The guard-field pattern that kept 2907 of 2907 old-configuration runs bit-identical while new code was added: + +```cpp +// partition_config.h: the defaults reproduce the old behavior +bool flow_unlocked_model = false; +bool jet_refinement = false; +bool mh_penalty_for_imbalance = false; +bool mh_keep_input = false; + +// app/configuration.h: only the new preconfiguration turns them on +inline void configuration::ultra( PartitionConfig & partition_config ) { + strong(partition_config); + partition_config.flow_unlocked_model = true; + partition_config.jet_refinement = true; + partition_config.mh_penalty_for_imbalance = true; + partition_config.mh_keep_input = true; + // k = 2 has the least time headroom: one more V-cycle; otherwise best of two full runs + if( partition_config.k == 2 ) { + partition_config.global_cycle_iterations = 3; + } else { + partition_config.repetitions = 2; + } +} + +// call site (simplified): the new code lives in its own files and is entered only behind its field +if(config.jet_refinement) { + jet_refinement jet; + jet.perform_refinement(config, G, boundary); +} +// ... the unchanged k-way FM and flow refinement follow +``` + +## Environment + +C++ (KaHIP v3.26), GCC 13.3 (CMake Release build, -O3), Ubuntu 24.04, Intel Xeon w5-3435X (16 cores). Runs were single-threaded, 16 at a time on fixed cores, with nice. kaffpaE and KaBaPE used MPI with 1 process per run. The graphs have 10k to 2.2M nodes and up to 2.7M edges; k ranges from 2 to 256 with 3% imbalance, and eps 0 for KaBaPE. + +## Links + +- KaHIP PR #175 (flow refinement fix): https://github.com/KaHIP/KaHIP/pull/175 +- KaHIP PR #177 (ultra preconfiguration, merged into master; review history in #176): https://github.com/KaHIP/KaHIP/pull/177 +- ultra manual: https://github.com/KaHIP/KaHIP/blob/master/manual/ultra.pdf +- Jet: M. S. Gilbert, K. Madduri, E. G. Boman, S. Rajamanickam, SIAM J. Sci. Comput. 46(5), 2024, https://doi.org/10.1137/23M1559129 +- F_H: T. Heuer, P. Sanders, S. Schlag, ACM JEA 24(1), 2019, https://doi.org/10.1145/3329872 diff --git a/claude-plugin/skills/aae/knowledge-base/039-pseudoflow-solve-direction.md b/claude-plugin/skills/aae/knowledge-base/039-pseudoflow-solve-direction.md new file mode 100644 index 0000000..565de53 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/039-pseudoflow-solve-direction.md @@ -0,0 +1,116 @@ +# Choosing the solve direction in pseudoflow / push-relabel maximum flow + +## Problem + +Speeding up an established maximum-flow code: the highest-label FIFO variant of +Hochbaum's pseudoflow (version 3.23), kept single-threaded under a fixed build +(`gcc -O4`), with the flow value required to stay correct. Metric: solve time +(user CPU from the end of input to the end of flow recovery) over 164 DIMACS +max-flow instances (rmf-wide/long, wash-rlg, polygon, road-graph punch, computer +-vision stereo/segmentation, graph bisection), reported as the shifted geometric +mean (10 ms shift) of per-instance median time ratios. Every kept change had to +pass three checks on all instances: flow value equals HIPR, the program's own +check, and an independent checker of capacities, conservation, and the absence +of an augmenting path. + +## What Worked + +Three transferable ideas, headlined by the direction rule. + +1. **Pick the solve direction.** Maximum flow by pseudoflow (and push-relabel + generally) is *not symmetric* in s-t orientation: solving the reverse graph + can be dramatically cheaper, because the algorithm's work depends on where the + excess has to travel. Before building the adjacency lists, cheaply estimate + the excess reachable from the source against the deficit at the sink, and + solve the **reversed** graph when the source side carries clearly more excess + than the sink side has deficit. The decision costs microseconds and is made on + the solve clock. It fired on 71 of the 164 instances; on the segmentation + (ALE) instances it cut solve time from 3-39 s to 0.01-0.02 s. Measured in + isolation (the original run on reversed input files), the orientation alone + explains about three quarters of the gain on the instances where it fires. It + was the single largest change of the campaign (chained factor 0.66). + +2. **Flow recovery is a real phase; profile it, do not assume the push loop + dominates.** The biggest single cost here was *not* in the main loop: the + original recovered the flow by sorting each node's arcs by flow with a + quicksort that goes quadratic on equal keys, decomposing paths, and scanning + all arcs. Replacing it with cycle cancellation and excess return in + topological order, over lists of positive-flow arcs only, was a 0.82x factor + overall and down to 0.26x on one vision family. "Sort by key" hides a + quadratic whenever keys tie, and post-processing/recovery phases are routinely + left unprofiled. + +3. **Verify a textbook optimization is not already satisfied before importing + it.** Exact initial distance labels, global relabeling, and lifting trees that + cannot reach a deficit all *failed*. An instrumented run showed why: in the + highest-label code the labels already equal the exact tree distances, so none + of these can save a single relabel. Instrument the quantity an optimization + targets before spending days implementing it; the standard trick may be a + no-op for your variant. + +A correctness note that is also a general trap: the original stored node excess +in a 32-bit `int`, which overflows on two large rmf-wide instances, silently +violating conservation. A 64-bit excess fixes it. On large instances, audit the +width of every accumulator. + +## Experiment Data + +164-instance DIMACS max-flow benchmark, single thread, median of three +repetitions, flow value verified on every instance: + +- Total solve time 2283 s -> 761 s (3.00x); shifted geometric mean of the + per-instance ratios 3.18x (2.80x excluding the four ALE instances the direction + rule makes trivial). No instance got slower. Peak memory 23.1 -> 14.2 GB. +- 132 hypotheses tested, 29 kept (one of them the correctness fix). Final code + ~1800 lines vs ~1300 original. +- Kept-change factors, chained geometric mean over the respective control: + +| kind | factor | main effect | +|---|---|---| +| direction rule (excess vs deficit) | 0.66 | stereo 0.42, polygon 0.47, bisection 0.46-0.52 | +| memory layout + huge pages | 0.77 | all large instances; rmf-wide 0.82 | +| flow recovery (sort, cycles, arc lists) | 0.82 | vision-surface 0.26, ALE 0.31 | +| hot path inlined; split roots to front | 0.83 | all groups ~0.93 | +| prefetch in bucket and child walk | 0.95 | punch 0.92 | +| path segments for long pushes | 0.98 | rmf-wide 0.85/step | +| 64-bit excess (correctness) | 1.01 | two rmf-wide instances become correct | + +## What Didn't Work + +- **LIFO buckets:** help vision by up to 30% but slow rmf by 1.4-2.4x, and no + per-instance signal separated the two regimes, so it could not be adopted. +- **Merge along the largest-capacity candidate arc:** halves rmf-long pushes + (242M -> 81M) but is 1.6-6x slower on rmf-wide; again no cheap regime signal. +- **Exact initial labels / global relabeling / lifting trees:** no-ops in + highest-label pseudoflow (labels already exact), as above. +- **Compiler-level tricks** (PGO, `-march=native`) were deliberately out of scope + (source-only goal), so they are not part of the result; worth remembering that + a fair "engineering" comparison usually holds the build fixed. +- The last ~10 hypotheses cleared nothing above noise; profiles put the remaining + time in the relabel scans, bound by memory latency on large instances and + branch mispredictions on small ones. A useful stopping signal. + +## Code Example + +Direction rule, pseudocode (run before adjacency lists are built): + +``` +src_excess = sum of capacities of arcs leaving the source +sink_deficit = sum of capacities of arcs entering the sink +if (src_excess > 1.5 * sink_deficit) // threshold fitted in-sample + solve with s and t swapped (arcs reversed) +else + solve as given +``` + +## Environment + +C, `gcc -O4`, single-threaded, Intel Xeon w5-3435X. Target: Hochbaum pseudoflow +3.23, highest-label FIFO. The direction idea transfers to any +augmenting/preflow max-flow or min-cut solver whose cost is asymmetric in s-t +orientation; the "profile the recovery phase" and "instrument before importing a +textbook optimization" lessons are general. **Caveat:** all tuning was in-sample +and the thresholds (1.5 excess ratio, 1% terminal-node rule for split roots, +segment length 32) were fitted on this instance set, so re-tune per benchmark. +Developed with Agentic Algorithm Engineering; the study is described in the AAE +paper (13th DIMACS Implementation Challenge, 2027). diff --git a/claude-plugin/skills/aae/knowledge-base/040-parhip-bit-identical-speedup.md b/claude-plugin/skills/aae/knowledge-base/040-parhip-bit-identical-speedup.md new file mode 100644 index 0000000..1166590 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/040-parhip-bit-identical-speedup.md @@ -0,0 +1,561 @@ +# Speeding up a distributed (MPI) multilevel graph partitioner with bit-identical output + +## Problem + +Make ParHIP, the MPI graph partitioner of KaHIP (multilevel: label propagation (LP) coarsening, +distributed contraction, evolutionary initial partitioning on the coarsest graph, projection and LP +refinement), faster while every partition file stays byte-identical to the original for the same +graph, k, preset, seed and number of processes. Benchmark: 180 jobs on 20 graphs (social, web and mesh, 0.3M to +23M nodes), k = 8 and 64, 1 to 32 processes, fast, ultrafast and eco presets. Metric up to experiment +029: shifted geometric mean of per-job ratios of the whole mpirun wall time (reading the graph and +writing the partition included). From experiment 030 on (owner's decision): ParHIP's own "total +partitioning time" (part, shift 0.01 s) as primary and its IO time as second headline; the wall time +only as a check. + +## What Worked + +After 29 benchmark rounds (49 experiments, 29 kept, 19 discarded): **part time about 0.62 of the original +(1.61x), IO time 0.565 (1.77x)**, every partition identical (seed 1 in every +timed run, seed 2 checked separately after each keep, from experiment 030 on also clang + libstdc++ +builds). Up to round 14 the decisions used the wall time (chained 0.609); the trajectory was later +recomputed on part directly against the stored master medians. + +1. **Per-phase timers first, then one hypothesis per phase per run.** A small RAII timer per phase + (io, coarsening LP, contraction, initial partitioning, projection, refinement LP, ghost exchange, + write), reduced to max and mean over ranks and printed by rank 0, costs nothing measurable and + lets several independent changes (one per phase, one commit each) share one 30-minute benchmark + run. Each change is judged by its own phase. Cross-cutting changes (shared data structures, + allocator) are run alone. +2. **Output writing was 12% of total wall time.** The text partition was written with one + `std::endl` (flush, so one write system call) per node, process after process behind a barrier + chain. Formatting each process's labels into one buffer and writing it once, with the same barrier + chain, took the write phase from 109 s to 0.7 s (25% of total time at 32 processes). +3. **Evaluate side-effect-free conditions lazily in the LP edge loop.** The loop computed a size + constraint (`getBlockSize`, a virtual call plus a hash lookup during coarsening) and other tests + for every edge, although they only matter if the candidate block is an improvement. Testing + `improvement` first and `continue` otherwise gave LP 16 to 19% faster and 8% overall. The + random tie-breaking draw must stay exactly where it was (it is consumed whenever the weight ties + the current maximum), otherwise the output changes. +4. **Replace std::unordered_map where order is never observed.** Coarsening block weights per + cluster id, the LP neighbor-weight map, the label mapping and coarse node weights in contraction, + and the global-to-local ghost id map: all are pure lookup tables (or iterated where the order is + irrelevant: request lists whose replies are index-aligned, additive weight messages). Open + addressing with a power-of-two table, multiplicative hash and a vector of used slots gives + identical values. A missing key must behave like the original (weight 0). +5. **Keep the maps whose iteration order defines the output.** The contraction's hashed edge map is + iterated to build the coarse graph, so its iteration order fixes the coarse edge order and with it + every later LP decision. It was left unchanged (no `reserve()`, no hash change). +6. **Parse text input by hand.** mmap, memchr for line ends and a hand-written strtol-equivalent + into flat arrays (io phase 0.76); every process still scans the file to its first line. +7. **Projection grouping by sort.** Sorting (coarse node, fine node) pairs and grouping replaces an + unordered_map of vectors and keeps the message and update order (project 0.91). +8. **Stamp arrays instead of reset scans.** Two places scanned a node's edges a second time only to + reset per-PE "already sent" flags: `addLabel` (called on every label change of an interface node, + in both LP phases and in projection) and the contraction's ghost mapping. A per-PE mark that holds + a per-call counter (or node + 1 where each node is visited once) removes the second scan and sends + the same entries in the same order. `addLabel` alone: coarsen LP 0.97, ghost exchange 0.95, + projection 0.88, refinement 0.97, overall 0.985. Use a counter, not the node id, where the same + node can be processed twice. +9. **Skip work that provably does nothing.** The global ghost update after each projection scanned + all local nodes; only interface nodes can have ghost neighbors (projection 0.93 total, 0.79 + geomean). One lookup per LP edge instead of read, add and re-read (coarsen LP 0.96). Reserving the + local edge array at the average share 2m/p before reading (io 0.94). +10. **Prefetch in the LP edge loop.** `__builtin_prefetch` of the target's node data eight edges + ahead: coarsening LP 0.95 total, up to 11% faster overall on the largest graphs, up to 4% slower on + small ones (net geomean 0.998). A prefetch never changes a value. +11. **Devirtualize both implementations or neither.** Block weights are accessed through a virtual + base with one implementation for coarsening and one for refinement. Keeping a typed pointer to + each and calling `ptr->Concrete::getBlockSize(b)` (qualified, so non-virtual and inlinable) behind + one null test gave coarsen LP 0.99, refinement LP 0.98, ghost exchange 0.99, overall 0.991. A + first version that devirtualized only the coarsening path (typed pointer or virtual fallback) made + refinement 8% slower: the extra branch plus the remaining virtual call cost more than it saved. +12. **A pool allocator for one container type, not for the program.** The contraction's hashed edge + map (`std::unordered_map`, iteration order defines the output) allocates one node per coarse edge. + A stateless allocator for this one map (single nodes from a free list in 64 KiB chunks, bucket + arrays from operator new) gave contraction 0.979 total and 0.944 geomean, overall 0.992, peak memory + slightly lower. Iteration order of `std::unordered_map` depends on hash and insertion order, not on + node addresses, so the output is unchanged. The global new/delete replacement of round 4 had + reached a similar contraction gain but was rejected as too invasive for a 4% gain in one phase; + the type-local form keeps the gain with one header and one typedef. +13. **More provably-empty work.** A per-call stamp also in the global ghost update (projection 0.90 + total, 0.94 geomean); returning early from the ghost block balance update when the label did not + change (ghost exchange 0.96). + +14. **Reproduce the iteration order of `std::unordered_map` instead of using it.** The coarse edge + order is the iteration order of the contraction's hashed edge map. libstdc++ keeps one singly + linked list: a new key goes to the front of its bucket's run if the bucket is occupied, else to the + front of the list, and a rehash re-inserts the list in order with the same rule. So for an + insertion sequence into B buckets the order is: bucket runs by first-use time descending, each run + by insertion time descending, computed epoch by epoch with the container's own + `std::__detail::_Prime_rehash_policy` (initial bucket count 1). An open-addressing array map plus a + counting sort per rehash epoch gives the identical order: contraction 0.718 total (0.842 geomean), + overall wall 0.963 / total 0.938, part 0.956 / total 0.929. Conditions for keeping such an emulation of library + internals: a randomized equivalence test against the real container as a test target, a + compile-time guard that enables it only for library releases on which the test ran (here + libstdc++ 11, 13, 15 with g++ and clang++) and falls back to the container everywhere else, the + identity check with a second compiler, and a gain of at least about 2% overall. +15. **Dense array for the locally owned part of a sparse key range.** Coarsening block ids are global + node ids; most blocks seen by a process are its own nodes' ids. A dense array for the local id + range in front of the hash map for the rest: coarsening LP 0.911 total (0.904 geomean). +16. **Find the first line of a process without scanning the prefix.** Each process counts the node + lines starting in its own byte chunk of the file, one Allgather of the counts, and each process + starts at the chunk holding its first node: IO 0.919 total (0.899 geomean). +17. **Small things that still count:** one reused receive buffer and an `MPI_Request` vector with one + `MPI_Waitall` instead of a heap request per Isend (ghost exchange 0.988), and a linear scan over a + 16-entry inline array instead of the hash table for nodes of degree at most 16 in the LP + (coarsening LP 0.987; the weight per label and every tie break stay the same). +18. **Prefetch the hash map slot of the lookups that integrate a received message.** The ghost + exchange, which runs inside both LPs every few hundred nodes, looks up every received global id in + an open-addressing ghost id map with a multiplicative hash, so each update of a large message + waits for a cache miss. Prefetching the slot eight updates ahead: ghost exchange 0.929 total + (0.966 geomean), and with it coarsening LP 0.969 and refinement LP 0.989, part 0.984 (0.980 at + p = 16). The same trick in the contraction's label mapping: contraction 0.986. The exchange had + been read as "waiting for slower processes" from a profile; part of it was this integration work. + A prefetch changes no result, so it is the cheapest kind of change under a bit-identity + constraint. + +### Profiling a latency-bound loop + +perf was not available; an `LD_PRELOAD` time sampler (ITIMER_PROF at 1 kHz, `backtrace`, symbolized +afterwards with `addr2line -i`) gave per-function and per-line shares with each job weighted equally. +It attributed 39% of the samples to the LP's hash table frames, yet replacing the table for small +degrees gained only 1.3%, and moving the labels into their own dense array gained nothing: in a loop +bound by a random load, samples land on the instruction that consumes the load (the hash of the +loaded label), not on the load. Before building on an inline-attributed share, test it with a cheap +change of the presumed cause. + +### Code layout as a confound in per-phase judgments + +A contraction change (038) gained 2% in its own phase and lost 3.7% in the refinement LP, whose +code it did not touch. The coarse graphs were bit-identical (hash of every level), the slowdown was +reproducible to 0.5% over 20 timed runs, and it came from a 16-byte move of the refinement +function: code linked earlier in the binary had changed size. A timed padding sweep (a dummy +function of 0, 16, 32 or 48 bytes in front of it) showed that the refinement LP is 1.18 to 1.33x +slower on single-process k = 64 jobs (1.02 to 1.06x for k = 8) when the function starts 48 bytes +past a 64-byte boundary, and equally fast at the other three offsets. So roughly one change in four +draws a large false penalty (or a false gain, when it leaves the slow offset) in a phase it does not +touch. Remedy: `__attribute__((aligned(64)))` on that one hot function, behind a macro that is +empty on other compilers; with it all four paddings measured the same, and the change itself was +neutral. Procedure when a phase whose code did not change moves beyond its band: compare the hot +functions' addresses modulo 64 in both binaries (`nm`), bisect the commit, and run a padding sweep +before judging. Untimed runs on a loaded machine are useless for this (they suggested a random, +stack-address effect; under exclusive the effect was strictly per binary). Retesting two earlier +refinement-side discards with the alignment fixed confirmed one of them as a real loss. + +### Leaving fresh memory uninitialized, and checking it + +The largest io gain of the later rounds came from not zero-filling the edge array before the graph +is built: an allocator whose default `construct` does nothing (`std::vector>`) makes `resize` leave new trivial elements uninitialized, and the +builder writes every entry once. io 0.848 total (0.871 geomean), peak memory 1445 to 912 MB. The +gain is a first-touch effect: the fill touched hundreds of MB of fresh pages ahead of the parser, +which then wrote them a second time. The same trick on short-lived, reused receive buffers gave +nothing measurable (glibc hands back pages that are already mapped), and on the much smaller node +array it lost more in the builder (a per-node branch) than it saved. + +Two pitfalls. (1) `resize` keeps old entries, and a caller can depend on that: ParHIP rebuilds a +graph in place on top of an old one and later reads fields that the rebuild does not write. Zeroing +"fresh" entries in the builder must therefore zero only the entries beyond the previous size; the +first version zeroed all of them and changed the output (caught at once by the quick hash check). +(2) Bit-identical output on one machine cannot prove that no byte is read before it is written; a +latent read can pass here and differ elsewhere. Check with valgrind memcheck +(`--track-origins=yes`) on a build without `-march=native` (valgrind cannot decode AVX-512), at +p = 1 and p = 4, on social and mesh graphs, fast and deterministic eco presets. Run the p = 4 cases +a second time over TCP (`--mca pml ob1 --mca btl tcp,self`): data that crosses process boundaries +through shared memory is not checked, but bytes passed to `writev`/`sendto` are. Expect noise from +the MPI library itself (uninitialised bytes in Open MPI and PMIx headers, origin in libmpi, +libopen-pal or libpmix) and classify every error by its origin stack; here all 214 reported errors +in 72 rank logs were of that kind, with 0 invalid accesses and 0 uses of uninitialised values +originating in ParHIP allocations. + +### Relaxing identity to a quality budget: measure the cut noise first + +After 52 experiments (30 kept) the owner relaxed the identity rule: output may change if the average +cut moves by at most 0.1% (Q = geomean over jobs of (mean cut + 1) / (master mean cut + 1) over the +same seeds, cumulative against the original, Q <= 1.001; balance, determinism of the deterministic +build and the interface stay fixed). Freeze the bit-identical best first (tag plus patch series) as +its own deliverable. Then measure the cut noise before any experiment: on this benchmark, original +seeds 1 to 5 against seeds 6 to 10 gave Q = 0.9926, and the sd of log Q over random 5/5 splits was +0.44%, so a 0.1% budget is far below what a few seeds can resolve (uniform seeds would need about 40 +per side for SE 0.1% and 150 for 0.05%). The variance sits in a few jobs (one social graph at k = 8 +with sd 53% of the log cut at p = 1; the median job 1.3%, mesh graphs below 1%), and those jobs are +cheap, so a Neyman allocation of seeds per job (S_j proportional to cv_j / sqrt(cost_j)) cuts the +cost about tenfold: SE 0.05% for about 2.6 CPU-hours per candidate on 16 cores. The delta-method SE +from 5 seeds underestimates (0.27% against 0.44% observed) because the cut distribution has heavy +tails; validate the protocol with an A/A split of a large master reference before trusting it. Even +the original goes over the imbalance bound on 2% of the runs (social graphs, k = 64, 16 to 32 +processes), so a balance rule must compare against the original job by job. + +### Keep rule for multi-component runs + +Judge each component commit by its own phase, not by the overall geomean against a fixed threshold: +keep it if (1) its phase improves beyond that phase's run-to-run noise, (2) the overall time is not +worse beyond noise, (3) the output is identical, and (4) the code is not disproportionately complex +for the gain. A phase that is 5% of the run can improve by 20% and still move the overall geomean by +only 0.3%; with a fixed overall threshold such clean wins (here: the interface-only ghost update) are +lost. Measure the per-phase noise as the largest deviation from 1 when the same binary is measured in +two separate runs (instrumented builds; we used four such pairs of full runs). Worst A/A deviations +(phase total / geomean of per-job phase ratios), full set of 180 jobs: + +| wall | coarsen LP | ghost exchange | contraction | io | initial part. | projection | refinement LP | +|---|---|---|---|---|---|---|---| +| 0.2% / 0.8% | 0.2% / 0.8% | 0.5% / 1.6% | 0.2% / 0.7% | 0.1% / 0.4% | 0.9% / 1.0% | 0.2% / 0.4% | 0.2% / 1.2% | + +On the 29-job quick set the bands are 2 to 4 times wider (0.4% to 1.4% total, 0.9% to 3.3% +geomean), so a screen only selects candidates and a full run decides. When two commits of one batch +can touch the same phase, run three binaries (control, first commit, both) in one alternating run. + +### Phase 2 (0.1% quality budget): output-changing speedups + +Once the output may change within a cut budget, the largest single gain came from balancing the +work per process at the finest level, not from local code changes. The per-round block-weight +Allreduce of label propagation was about 2/3 waiting at 16 to 32 processes; a per-process counter +showed equal node counts but up to 7.9 times the mean edge count on one process (social graphs +stored with hubs first). Splitting the input so that each process gets about the same +`adjacency entries + node lines * average degree` (counted per 1/1024 of each process's byte +chunk, one Allgather) gave part 0.945 geomean (social graphs 0.87 to 0.92 at p >= 4, meshes +neutral), ghost-exchange time 0.74, IO 0.90, and a slightly better cut (Q 0.998). Lessons: + +- Bytes are a biased proxy for edges: later lines hold longer node IDs, so a byte split gave the + first processes of a geometrically ordered mesh more work (meshes 1.04 slower). Count entries. +- Half node, half edge balance was as fast as pure edge balance (edge max/mean 1.9 instead of 1.2) + and avoids processes with almost no nodes. +- A counting pass in the reader must be branch-free (vectorized) and skipped with one process: the + first, branchy version cost 29% of IO at p = 1. +- Coarsening order: visiting nodes of a degree class in node order (counting sort) is cache + friendly (coarsening 0.88) but sweeps mesh geometry and costs mesh cut (+0.3%); shuffling chunks + of 128 consecutive nodes keeps most of the speed (0.91) with neutral mesh cut. About 2/3 of that + gain depends on the given node numbering (randomly permuted inputs: part 0.986 instead of 0.96). +- Quality gate that held up: Neyman-allocated seeds per job (cv of the cut over sqrt of cost), a + master reference with twice the seeds, Q = geomean of (cut + 1) ratios on paired seeds, A/A SE + 0.05%; screen on a quarter of the seeds, discard above Q 1.003; balance as one-sided Fisher tests + of over-bound rates per job with Bonferroni, plus a pooled test. + +### Phase 2: coarse-level rebalancing, fresh seeds, and measuring on a second host + +The finest-level split left the coarse levels with equal node ranges per process, so hub-heavy +graphs got up to 2.9 times the mean edge count per process again after contraction. Four attempts: + +- Rebalancing every coarse level by edge cost: part 0.995 overall, youtube 0.84, but meshes 1.012 + (p = 32 1.028). Moving coarse ranges costs locality where the split was already even. +- Rebalancing only when the equal split puts max > 2 x mean cost, with an extra pass over the + coarse edges to count: on held-out graphs 0.85 to 0.97, but one graph 1.078 with identical + output (the pass itself). +- Cost from fine degrees summed per bucket of coarse IDs during contraction: no extra pass, but a + 256 KB Allreduce per level and a 64-bit division per node: output-identical p = 32 jobs 1.011. +- Two-stage reduction: each process sums its local buckets per equal range, an Allreduce of p + values decides, and the full histogram is reduced only on levels that fire; power-of-two + buckets (shift instead of division). Output-identical p = 32 jobs 1.003, part 0.989 / 0.982, + unseen hub-heavy graphs 0.956 with a road-network control at 1.003. Kept. + +Lessons: measure the overhead on output-identical jobs with repetitions before judging gains; +decide from a small summary and reduce the full data only when needed; graphs looked at while +tuning stop being held-out (confirm on unseen graphs plus a control that cannot fire). + +Selection bias check: after four kept output-changing changes, the cumulative best and master +were both run on seeds no earlier run had used (beyond the 2x master reference): Q 0.9939 against +0.9953 to 0.9958 on the selection seeds, so the noise picked while selecting did not flatter the +result. One balance flag with 3 seeds (1/3 vs 0/3) disappeared with 200 more seeds per side. + +A second machine (Xeon Silver 4216, same core numbering, about 1.6x slower) inflates speedups: the +same pair of binaries gave 0.956 there and 0.989 on the main host, and the direct master ratio +0.49 instead of 0.55. Keep one host per trajectory, report the other host's figure with its label, +and build both sides of a timed pair on the same host (`-march=native` binaries from the other +host are a different code generation). + +A byte-counting loop that gcc vectorizes is still slow with 64-bit accumulators (every flag is +widened); 8-bit counters over blocks of 255 bytes run 2.6x faster (1.96 to 5.1 GB/s per core) with +identical counts, and IO improved 2 to 3.5% at p >= 4. + +Process lessons from running unattended chains: gate every step on the command's exit status, not +on grepping a log (a failed cmake left a chain waiting 3.5 hours), and give every wait a timeout +that raises an alert; kill only by a looked-up process ID, never by pattern (a `pkill -f` matched +the shell that issued it); record which MPI build each run used, and compare against master only +from a direct rerun with the same MPI on both sides (a ratio chained through an intermediate +binary was 0.6205 against 0.6155 direct); do no heavy work on a shared host while another user's +exclusive job runs there, including checksums and report builds. + +### Phase 3: completing nonblocking sends, rebasing, and the input graph build + +An audit of ParHIP's nonblocking sends (MPI 4.1: every request must be completed, and a send buffer +must not be touched before completion) found requests that were never waited on. A PMPI wrapper +that counts outstanding requests at `MPI_Finalize` is the reliable detector: MPICH's +"pending references" message misses requests on duplicated communicators. Counts at p = 4 grew +linearly with the number of partitioner calls (164 to 820 per rank for 1 to 5 calls) before the +upstream fix, and one remaining case (the evolutionary partitioner sending a population size that +it then changes) needed `MPI_Waitall` before the buffer is reused. An apparent hang was the eco +preset's evolutionary time limit (2048 / p seconds), not a deadlock; use a deterministic build for +leak tests. After each upstream merge, rebase, check partition identity of new master vs old master +and of the rebased branch vs the branch before, keep a map of old to new commit ids next to the +results table (its ids stay as recorded), and run one timed neutrality check. + +Input graph build: per-phase timers (12 phases) plus per-process counts showed that the build, not +the parse, was the largest phase on meshes and social graphs, and that its time followed the number +of new ghost nodes per process (delaunay_n22 p = 16: 1.25M ghosts on one process, 0.253 s, against +89k and 0.065 s elsewhere). The cost was growth: 18 rehashes of the ghost map and doubling of three +arrays. An exact distinct-ghost count from a bitmap over global ids (only when the bitmap is not +larger than the local target array) and one reservation cut io by 14% in total on meshes at +p >= 16 (delaunay_n22 0.73, social p = 32 0.90) with identical output. Two alternatives lost: the +ghost-edge count as a bound over-reserves 4 to 14x on social graphs, and key/value slots in the +open-addressing map (one cache line per probe) made the build 5 to 7% slower. + +Code layout also moves IO: on the second host, five binaries with the same reader code path at +p = 1 gave stable, binary-specific reader times differing by up to 8% (eu-2005 0.920 to 1.000 s, +spread within a binary about 1%); two of them differed only in communication code. The run order +inside a pair matters too (the second run's IO is 3 to 6% faster), but alternating the order cancels +it; layout does not cancel. Use p = 1 jobs, where a parallel-only change does not run, as the +layout calibrator: report IO claims below about 5% net of the p = 1 band, and confirm them on the +main host. + +Number parsing in the reader: once the build was cheap, the parse was about 12 ns per adjacency +entry, dominated by the per-digit branch of the strtol-like loop (number lengths vary, so it +mispredicts about once per number). Reading the 8 bytes at the number start (`memcpy` into a +64-bit word), finding the run length with a branch-free byte mask and converting up to 7 digits +with three multiplications, with everything else falling through to the old routine, made the +parse 1.7 to 1.9x faster single-core and io 0.902 in total (social 0.89 to 0.93, p = 1 on large +files 0.82 to 0.86) with identical output; a fuzz test against the old routine (also under +address and undefined-behaviour sanitizers, inputs in exact-size heap buffers) guards the +equivalence. The in-run gain at p = 16 was about half of the single-core one (shared memory +bandwidth, page faults of the fresh target array). An earlier sentinel parser that kept the +per-digit loop had shown nothing (round 12): the digit loop, not the scan, was the cost. + +### Phase 3: splitting LP time into local work and waiting + +After the reader work, io is about 5% of the total and the part phases are the lever (coarsening LP +197 s, refinement LP 102 s, contraction 90 s, ghost exchange 83 s inside the LP phases, over the +180 jobs). Two contraction experiments failed on wrong premises, and a per-rank split of the LP +time changed the target. + +- Contraction: per-rank sub-step timers showed that the order-preserving edge map (local build, + ordered iteration, the second map on the receiver) is 60 to 80% of contraction on the slowest + rank. Putting the key and the weight into the probe slot (one cache line per lookup instead of up + to four) made contraction 2.3% slower: ids are handed out in insertion order and lookups come in + runs of the same coarse source, so the id-indexed side arrays were already cached; the 4x larger + slot array and a gather pass cost more. An exact multiply-based modulo in the order computation + saves only 6 to 16% of that step: the order computation is bound by its random scatter. +- Coarse rebalancing from the exact edge copies each owner receives (instead of fine degrees): + a firing table (one diagnostic line per contraction, every job, seed 1) showed it fires on fewer + levels than the fine-degree rule (24 vs 41 social jobs) and not at all on the motivating job. + The excess of the slowest rank was on the sending side (its hubs create 4.5x more local coarse + edges), which no choice of coarse ranges moves. Lesson: check the firing table, which costs + minutes, before redesigning a decision rule. +- LP split (per rank: total, block-weight allreduce, probe wait, receive and integrate, Isend + Waitall, final barrier): refinement is imbalance-bound. On enwiki-2013 (16 PEs) 9.6 of 15.9 s + are the per-round block-weight allreduce, while the total compute differs by only 10% between + ranks: a round is a fixed number of nodes in random order, so its edge work varies with the + hubs it draws, and every round costs the maximum over ranks. Social and web graphs: 50 to 74% of + refinement is this wait; meshes 20 to 25%. Coarsening is compute-bound (local LP plus 38% for + integrating ghost updates). Meshes show a separate, large compute imbalance in coarsening + (one rank 3.2 s, most others near 0) that is not yet explained. +- Rounds of equal work instead of equal node counts: a round ends once the visited nodes' + degree + 1 add up to (local edges + local nodes) / rounds. In both LP phases this cut the + refinement wait as intended (enwiki 9.6 to 2.5 s), but cost 0.23% cut (fast presets 0.6%, one + job 20%): in coarsening, a process with many low-degree nodes now visits far more nodes between + exchanges, so the labels it sees are staler while clusters form. Applied to refinement only + (moves between k blocks under a bound are insensitive to one round of staleness), the cut is + unchanged (0.9998, held-out 0.997, unseen 0.998) and part time drops 4.3% total (refinement + 0.79; social graphs at 16 and 32 processes 0.90). Round boundaries are part of the coarsening + algorithm, not only a communication schedule. +- A branch on a loop-invariant flag in the LP node loop cost about 1% of coarsening (first + V-cycle, identical work, three alternating repetitions); one template instantiation per value + removed it. Measure such costs on the first V-cycle: later cycles see a different partition. +- Count what the ghost exchange receives before tuning how it is processed. A third of the + coarsening updates and 96% of the refinement updates (enwiki, 16 processes) changed no label: + every interface label set outside label propagation (projection, V-cycle reset, initial + assignment) was buffered and sent in the next LP's first round, although those callers already + leave ghosts consistent. Recording only while the LP runs removed them with identical output; + part time 0.942 total (16 and 32 processes 0.95 gm, projection 0.76). +- On one process the timed gain of dropping the work counter was 2.4% of refinement, against 11% + in an untimed single-core screen; in the same run, unchanged p1 coarsening moved by 3.7%. +- An allocation that lives across phases can cost an untouched phase as much as code layout: a + per-node list of adjacent processes, built at a graph's first label propagation and kept until + the graph is freed, sped up coarsening (0.94 at 16 and 32 processes) but made refinement 3.5 to + 9% slower (timed) with identical output. Untimed variants (3 repetitions, spread under 0.4%) + separated the causes: moving the new code out of line did not help, while building the lists + without ever reading them reproduced the refinement loss (up to 11% more refinement work). The + mechanism is unproven (likely the address shift of everything allocated after the lists). Treat + any longer-lived or larger allocation, such as a bigger hash table reserve, as a footprint change + and measure the other phases too, not only the one it targets. + Follow-up: freeing the lists when the coarsening label propagation ends brought refinement back + on the social graphs (enwiki 1.047 to 1.010, youtube 1.052 to 1.028), and a reserve of 2 or 4 + times the needed hash table size also slowed refinement by 1 to 8%. On the mesh (hugetrace) the + loss stayed, and a run-time switch that never built the lists, in the same binary, still showed + 7% more refinement work than the base, while two diag binaries that differ by one getenv call + differed by 3%. So part of such a loss is a binary effect, not memory: separate the two with a + run-time switch in one binary before concluding footprint, and let the timed net on the full set + decide. +- Split a hot spot before optimizing it: of the ghost update integration in coarsening, 96% was + the block weight hash map (random keys: the labels of the updates), not the label stores. A + software prefetch ring over the incoming message (ghost id slot 8 updates ahead, ghost lookup, + node data and the new label's weight slot 4 ahead, the old label's slot 2 ahead) cut that loop + from 3.7 to 2.8 s on the critical rank, with identical output; part time 0.976 total on the + full set (16 processes 0.964, coarsening LP 0.946, refinement unchanged: no new allocation). + Reserving the map to its final size instead lost (doubling the zeroing, refinement up to 8%). +- An optimization that crashed once in a deterministic build and then never again was parked, not + called a flake: per-node lists of adjacent processes (about 1% net) segfaulted on one 4-process + job and hung another in one packed, oversubscribed identity batch; 6 plain repeats, sanitizer + (address, undefined, libstdc++ assertions, an explicit range check on the list path), valgrind + and three reruns of the batch all passed, and so did the base. Keep the failing batch, core dumps + and per-rank stderr when a run fails; a rerun loop without them only shows that it is rare. +- UndefinedBehaviorSanitizer stops at an upstream line (uncoarsening.cpp binds a reference to a + null complete_boundary) before reaching ParHIP's own code; suppress that line to get further. +- Split the label propagation node loop by degree class before guessing: per edge about 90 to 140 + cycles, and the short linear-search path (degree up to 16) cost more per edge than the hash + path, so the cost was not the hash table but the first misses of every node (its record, first + edges, own data and first target labels; the in-node prefetch 8 edges ahead never covers short + nodes). Prefetching these a few nodes ahead in the visiting order (record 4, first edge line 2, + data and up to 8 first targets 1 ahead) cut coarsening LP by 11% (178 to 158 s) and refinement + LP by 7% on the full set, part 0.938 total, identical output. The untimed diag gain (13% over 8 + jobs) carried over; trust such a diag over a guess. +- Sizing the hash table prefix per node (pow2 >= 2 * degree instead of 2 * max degree, 8 MiB for + enwiki) gave 2.7% of coarsening LP but only 0.7% of part time: inside the 1% band, not kept. +- Collecting a moved node's adjacent processes during the scan (a bitmask) to skip the send + rescan lost 3.6%: every ghost edge of every interface node pays the process lookup, but only + moved nodes save the rescan, and in refinement few nodes move (refinement LP +18%). The diag had + only measured the coarsening phase; when a change touches code shared by both LP phases, measure + both. +- Projection was mostly the global ghost label update (enwiki 3.0 of 3.1 s), and its time was + work, not waiting: the first message arrived at once, but each received (global id, label) pair + is a random lookup in the ghost hash map. The same prefetch ring as in the ghost exchange (hash + slot 8 pairs ahead, lookup and node data 4 ahead) cut projection from 25.1 to 16.5 s on the full + set, part 0.979 total, identical output. +- A prefetch ring must be larger than its lookahead: a first diagnostic version with 4 slots at + distance 4 overwrote each slot before reading it, gave wrong ghost labels and crashed later in + refinement. Name the distances and the ring size as constants and `static_assert(lookahead < + ring_size)`. +- Run the sanitizer build with libstdc++ assertions on prefetch code too: `&v[i]` for a prefetch + of the first edge of a node a few ahead hit `i == size` (a trailing node without edges). It is + harmless at run time (a prefetch never faults) but out of range; form the address as + `v.data() + i`. Only one graph (enwiki) and both process counts triggered it, the mesh did not. + +## Experiment Data + +| Round | Change (phase) | Phase ratio | Full geomean | Verdict | +|---|---|---|---|---| +| 1 | buffered writer (write); block weight map (coarsen LP); mmap reader (io) | 0.006; 0.76; 0.76 | 0.738 | keep all | +| 2 | LP neighbor map; projection sort grouping; contraction maps | 0.97; 0.91; 0.94 | 0.972 | keep all | +| 3 | ghost id open-addressing map plus binary search for the owner PE (cross-cutting) | io 0.92, comm 0.93 | 0.983 | keep | +| 4 | small-block allocator for operator new (cross-cutting) | contract 0.96 | 0.995 (quick) | discard | +| 5 | LP lazy size constraint; contraction flat adjacency | 0.84/0.81; 0.99 | 0.921 | keep; discard | +| 6 | LP one map access per edge; refinement touched-entry reset | 0.96; 1.035 | 0.986 (first alone) | keep; discard | +| 7 | packed-key degree sort; interface-only global ghost update | refine 1.14; project 0.93 | 0.998 (second alone) | discard; keep (per-component rule) | +| 8 | reused receive buffer; reader reserve; contraction stamp | comm 0.99 (gm 1.00); io 0.94; contract 0.98 | 0.994 | discard; keep; keep | +| 9 | addLabel stamp; LP prefetch | see 8. and 10. above | 0.985; 0.998 | keep; keep | +| 10 | std::fill to clear the refinement weight array | refine 0.994 (gm 0.987) | 1.000 (quick) | discard (inside quick noise) | +| 11 | global ghost update stamp; skip unchanged ghost balance | project 0.90; comm 0.96 | 0.993 (both) | keep; keep | +| 12 | newline-sentinel number parser in the reader | io 0.99 | 1.001 (quick) | discard (inside quick noise) | +| 13 | coarsening-only devirtualization; both balance types devirtualized | refine 1.08; see 11. above | 1.007; 0.991 | discard; keep | +| 14 | node pool allocator for the hashed coarse graph | contract 0.979 (gm 0.944) | 0.992 | keep | +| 15 | dense local block weights (coarsen LP); per-chunk line counts (io) | 0.911; 0.919 | 0.960 (both) | keep; keep | +| 16 | array map with reproduced unordered_map order (contract); degree-sized LP table | 0.718; coarsen 0.988 (gm 1.004) | 0.963 (part 0.956); 1.006 | keep; discard | +| 17 | node labels in their own array (cross-cutting) | coarsen 1.010, refine 1.024 | part 1.002 | discard | +| 18 | refinement touched-entry reset; Iallreduce overlapped with the ghost receive | refine 1.140; comm 1.054 | part 1.017; 1.002 | discard; discard | +| 19 | no allocation per ghost message; small-degree linear scan in LP | comm 0.988; coarsen 0.987 | part 0.998; 0.995 | keep; keep | +| 20 | MADV_POPULATE_READ on the reader chunk; contraction adjacency as one array plus message swaps | io 0.991; contract 0.982 but refine 1.037 | part 1.001; 1.005 | discard; discard (layout, see above) | +| 21 | contraction adjacency as one array without the swaps | contract 0.994 (gm 0.986) | part 1.001 | keep | +| 22 | 64-byte alignment of the LP function; refinement touched-entry reset retested | all phases in band; refine 1.140 | part 1.000; 1.028 | keep (hygiene); discard | +| 23 | retest of the separate label array with the fixed alignment | coarsen gm 0.986, project 1.033 | part 1.010 total | discard | +| 24 | edge array not zero-filled before the build (io) | io 0.848 (gm 0.871), peak RSS 1445 to 912 MB | part 0.999 | keep | +| 25 | contraction receive buffers without fill, swap (contract); projection receive buffers without fill (project) | contract 1.002; project 0.984 (gm 0.996) | part 1.000 | discard; keep | +| 26 | node data array not zero-filled (io, contract) | io 0.989 (noise), contract 1.006 | part 1.001 | discard | +| 27 | degree sort on packed keys, retested with the fixed alignment | coarsen 1.004 (gm 1.013), refine in band | part 1.002 | discard (the earlier refinement loss was layout, the verdict stands) | +| 28 | prefetch in the ghost exchange (comm); prefetch in the contraction's label mapping (contract) | comm 0.929 (gm 0.966); contract 0.986 | part 0.984; 0.997 | keep; keep | +| 29 | early stop of refinement LP when moves plateau (output-changing) | cut +0.2 to 0.3% | about 0.2% part per 0.01% cut | discard (late sweeps carry the cut) | +| 30 | degree order by counting sort, then random 128-node chunks per degree class (coarsen) | coarsen 0.865 | part 0.956 | keep (Q 0.998, mesh 1.000) | +| 31 | Iallreduce of block weights with one round lag (refine) | wait moves into MPI_Wait | part 1.00 | discard (load imbalance, not latency) | +| 32 | input split balancing entries + node lines (io, all LP phases) | comm 0.743, io 0.896 | part 0.945 | keep (Q 0.998 vs previous) | +| 33 | coarse PE ranges rebalanced by edge cost at every level (contract, all coarse LP) | youtube 0.84, mesh p32 1.028 | part 0.995 | discard (locality) | +| 34 | the same, guarded by max/mean > 2 with a counting pass; then cost from fine degrees per coarse-ID bucket | identical-output jobs 1.078 (one graph); 1.011 at p32 | part 0.995 / 0.983 | discard (overhead) | +| 35 | guarded rebalancing with a two-stage reduction and power-of-two buckets | identical-output p32 1.003; unseen 0.956 | part 0.989 | keep (Q 0.9958; fresh seeds 0.9939) | +| 36 | 8-bit block counters in the reader's counting pass (io) | io 0.978 (p >= 4: 0.95 to 0.97 total) | part 0.998 | keep (identical output; second host) | +| 37 | ghost map and arrays reserved to the exact ghost count (io build) | io 0.949 (mesh p16/p32 0.86 total; net of p = 1 band about 0.92) | part 1.003 | keep (identical output; second host) | +| 38 | adjacency numbers of up to 7 digits from one 8-byte load (io parse) | io 0.930 (gm) / 0.902 (total) | part 1.001 | keep (identical output; second host) | +| 39 | contraction edge map: key and weight in a 32-byte probe slot (contract) | contract 1.023 (untimed screen, 6 jobs) | part 1.006 | discard (no timed run) | +| 40 | coarse PE ranges decided from the exact edge copies per owner (contract) | fires on 24 vs 41 social jobs | n/a | discard (falsified by the firing table) | +| 41 | LP rounds of equal edge work in coarsening and refinement | refine 0.56 (enwiki), cut 1.0023 | n/a | discard (quality: stale labels in coarsening) | +| 42 | the same in refinement only, node loop instantiated per round rule | refine 0.788 total; social p16/p32 part 0.90 | part 0.957 (gm 0.974); unseen 0.938 | keep (Q 0.9998) | +| 43 | one process: refinement LP without the work counter | p1 refine 0.976 (gm) | p1 part 0.972 (gm, about 3% placement) | keep (identical output; timed with 44) | +| 44 | ghost labels recorded only while LP runs (no buffered no-op updates) | project 0.76, comm 0.91, refine 0.94, coarsen 0.95 | part 0.942 (gm 0.953); unseen 0.956 | keep (identical output) | +| 45 | per-node list of adjacent processes for the ghost label sends (lists freed after coarsening) | coarsen 0.96, refine 1.03 (lists kept) | not timed as the freed variant | parked (one unexplained crash and hang in an identity batch) | +| 46 | coarsening block weight map: prefetch ring over incoming ghost updates | coarsen 0.946, comm 0.855, refine 0.997 | part 0.976 (gm 0.984); unseen 0.986 | keep (identical output) | +| 47 | label propagation: prefetch across nodes in visiting order | coarsen 0.888, refine 0.928 | part 0.938 (gm 0.944); unseen 0.974 | keep (identical output) | +| 47 | LP hash table prefix sized per node degree (on 47) | coarsen 0.973 | part 0.993; unseen 0.987 | not kept (inside the 1% band) | +| 47 | moved node's adjacent processes collected in the scan (on the above) | coarsen 1.006, refine 1.183 | part 1.036; unseen 1.020 | discard | +| 48 | global ghost label update: prefetch ring over the received pairs (project) | project 0.660 | part 0.979 (gm 0.987); unseen 0.988 | keep (identical output) | + +From round 17 on the full-geomean column is the part time. Cross-run A/A of the same binary: part +0.9% geomean / 0.2% total, IO 0.5% / 0.4%; within one alternating run, IO moved by up to 1.6% +between two binaries with identical reader code, so IO differences below about 2% are noise. + +Noise between two passes of the original: geomean drift 0.4%, per-job sd of log ratio 0.027. +Per-phase noise: see the keep rule above. + +## What Didn't Work + +- Overlapping the per-round block-weight Allreduce (Iallreduce, one round lag): the wait just + moved into MPI_Wait, because it is load imbalance between processes, not latency. +- Stopping refinement label propagation early when moves plateau: the late sweeps are worth 0.2 to + 0.3% of the cut, so the trade (0.2% time per 0.01% cut) is poor. +- A small-block allocator replacing global new/delete (worked for sequential KaHIP) gave only 0.5% + here: after the hash maps were replaced, ParHIP's allocations are mostly large vectors. +- Turning the contraction's vector-of-vectors adjacency into a flat array changed contraction by + only 1%; the order-defining hash maps dominate that phase. +- Sorting nodes by degree on packed (degree << 32 | node) keys gives the same permutation (same + comparison results) but was slower: the per-call key vector and two copy passes cost more than the + cheaper comparisons (refinement 1.14x slower). +- Resetting only the touched entries of the refinement LP's k-entry weight array (k = 64) was slower + (1.035) than clearing all k entries. +- Reusing one receive buffer in the ghost exchange: no gain beyond noise; the small, short-lived + vectors are cheap for glibc malloc. +- `std::fill` instead of a hand loop to clear the k-entry refinement array, and a sentinel-based + number parser in the reader: both inside the noise of their phase (the compiler already vectorizes + the loop; the reader is bound by scanning, not by digit parsing). +- Devirtualizing only one of two implementations behind a branch (see 11.). +- Sizing the LP neighbor table per node by its degree (power-of-two prefix of the table): coarsening + LP total 0.988 but geomean 1.004; the per-node sizing costs the small mesh jobs more than it saves. +- Moving the node labels out of the 32-byte node struct into a dense array (hot/cold split): no gain + (coarsening 1.010, refinement 1.024). The prefetch already hides the label loads; the edge array is + the larger stream. +- Resetting only the touched entries of the refinement weight array, tried a second time with a + first-use test per access, and a third time with the function alignment fixed: 1.14x slower each + time. A full reset of 8 to 64 entries is a few sequential + stores. +- `MPI_Iallreduce` started before the ghost receive and completed after it: ghost exchange 1.054. + The reduction was never the cost; the waiting is imbalance between processes, and the + non-blocking collective adds progress overhead. +- The wall time includes a fixed launch cost (trivial MPI program: 0.11 s at p = 1 to 0.21 s at + p = 32), which caps the wall-time ratio of jobs that run under one second (floor 0.39 for those + 39 jobs); a program-internal timer avoids this. +- `madvise(MADV_POPULATE_READ)` over the reader's own byte chunk of the memory-mapped file: io + 0.991, inside the spread of the io timer between two binaries with identical reader code. +- Sorting nodes by degree on packed (degree, node) keys, retested after the layout fix: the + refinement loss of the first test was the layout draw, but coarsening is still 1.3% slower + (geomean): the key vector and two copy passes cost more than the cheaper comparisons save. +- Not zero-filling the node data array (much smaller than the edge array): io in noise, contraction + 1.006 (the per-node branch in the builder costs more than the fill). +- Not zero-filling the contraction's receive buffers: contract 1.002; short-lived buffers of + moderate size come from already mapped pages. +- Inlining the key into the probe slots of an insertion-ordered hash map whose lookups come in + runs: the side arrays it replaces were cache-resident, so the 4x larger slots only cost. +- valgrind/callgrind failed (illegal instruction) on the `-march=native` build; a separate + profiling-only build without native flags was needed. + +## Code Example + +```cpp +// LP edge loop: draw the tie-break exactly as before, then test cheap conditions first +bool improvement = cur_value > max_value; +improvement |= cur_value == max_value && random_functions::nextBool(); +if (!improvement) continue; // nothing below has side effects +bool cycle = !config.vcycle || G.getSecondPartitionIndex(node) == G.getSecondPartitionIndex(target); +bool balancing = own_block_balanced || cur_block != old_block; +if (!(cycle && balancing)) continue; +if (cur_block == old_block || G.getBlockSize(cur_block) + node_weight <= cluster_upperbound) { + max_value = cur_value; max_block = cur_block; +} +``` + +## Environment + +C++ (GCC 13, -O3 -march=native), Open MPI 5, one process per core, Intel Xeon w5-3435X +(16 cores / 32 threads). Graphs from SNAP, DIMACS 10 and the Florida collection (METIS text). diff --git a/claude-plugin/skills/aae/knowledge-base/CONTRIBUTING.md b/claude-plugin/skills/aae/knowledge-base/CONTRIBUTING.md new file mode 100644 index 0000000..fed8004 --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/CONTRIBUTING.md @@ -0,0 +1,57 @@ +# Contributing to the Knowledge Base + +If your AAE session produced insights that could help others, we'd love for you to share them. + +## How to Submit + +Open a pull request that adds one entry file and a corresponding row in `INDEX.md`. One entry per PR. + +If you are an agent running AAE, ask the user once during setup whether you may submit a PR. Only proceed with explicit permission. With permission, open the PR after 10 experiments and update the same entry and PR every 10 experiments after that (see "Contributing Back" in `AAE.md`). Use `gh` or the GitHub MCP server to create the PR. + +## Naming Convention + +Name your file `-.md`, using the next available number from INDEX.md and a short kebab-case description. Example: `003-loop-unrolling-matrix-multiply.md`. + +## Entry Template + +Your entry should include the following sections. Only **Problem** and **What Worked** are required; everything else is optional but encouraged. + +```markdown +# + +## Problem + +What were you optimizing? Describe the problem, input characteristics, +and the metric you were targeting. + +## What Worked + +The key insight or technique that produced improvement. This is the +most important section; even if you provide nothing else, this should +be useful to someone facing a similar problem. + +## Experiment Data (optional) + +Results from your AAE run. Can be a pasted results.tsv, a summary +table, or just the key numbers (baseline vs. best). + +## What Didn't Work (optional) + +Approaches you tried that failed or regressed. Often as valuable as +what worked, since it saves others from repeating dead ends. + +## Code Example (optional) + +A diff, snippet, or pseudocode showing the core change. + +## Environment (optional) + +Language, hardware, dataset size, or other context that might affect +whether this technique transfers to another setting. +``` + +## Quality Bar + +- The insight should be transferable: useful to someone facing a similar problem, not just a project-specific tweak. +- "What Worked" should be specific enough to act on, not just "I made it faster." +- Experiment data is encouraged but not required. diff --git a/claude-plugin/skills/aae/knowledge-base/INDEX.md b/claude-plugin/skills/aae/knowledge-base/INDEX.md new file mode 100644 index 0000000..1eea1cb --- /dev/null +++ b/claude-plugin/skills/aae/knowledge-base/INDEX.md @@ -0,0 +1,46 @@ +# Knowledge Base Index + +Optimization techniques, experiment results, and lessons learned from AAE sessions contributed by the community. Agents should read this index to find entries relevant to their current problem, then read the linked files for details. + +| # | Title | Problem Domain | Key Technique | File | +|---|-------|---------------|---------------|------| +| 001 | Cache-friendly blocked recursive sorting | Sorting, large arrays | Blocked partitioning with cache-line-sized buffers | [001-blocked-recursive-sorting.md](001-blocked-recursive-sorting.md) | +| 002 | Gradient accumulation as batch size proxy | Transformer training, limited VRAM | Gradient accumulation to simulate large batches | [002-batch-size-tuning-transformer.md](002-batch-size-tuning-transformer.md) | +| 003 | Arena allocation for JSON parsing | Parsing, memory allocation | Arena allocator to eliminate per-node heap allocation | [003-arena-allocation-json-parsing.md](003-arena-allocation-json-parsing.md) | +| 004 | SoA vs AoS for cache efficiency | Data layout, particle simulation | Structure of Arrays to maximize cache line utilization | [004-soa-vs-aos-cache-efficiency.md](004-soa-vs-aos-cache-efficiency.md) | +| 005 | Branchless programming | Array processing, partitioning | Arithmetic substitution for conditional branches | [005-branchless-programming.md](005-branchless-programming.md) | +| 006 | Small buffer optimization | String/container allocation | Inline buffer to avoid heap allocation for small objects | [006-small-buffer-optimization.md](006-small-buffer-optimization.md) | +| 007 | Move semantics avoiding deep copies | Container transfers, pipelines | RVO/NRVO and std::move for O(1) ownership transfer | [007-move-semantics-avoiding-copies.md](007-move-semantics-avoiding-copies.md) | +| 008 | constexpr compile-time computation | Lookup tables, constants | Compile-time evaluation to eliminate runtime initialization | [008-constexpr-compile-time-computation.md](008-constexpr-compile-time-computation.md) | +| 009 | False sharing avoidance | Multithreaded counters, parallel scaling | Cache line padding to prevent cross-core invalidation | [009-false-sharing-avoidance.md](009-false-sharing-avoidance.md) | +| 010 | PGO + LTO compiler optimizations | Whole-program optimization | Profile-guided and link-time optimization for 20%+ gains | [010-pgo-lto-compiler-optimizations.md](010-pgo-lto-compiler-optimizations.md) | +| 011 | Memory-mapped I/O for large files | File processing, large datasets | mmap for zero-copy lazy file access | [011-mmap-large-file-processing.md](011-mmap-large-file-processing.md) | +| 012 | Open-addressing hash maps | Lookup-heavy workloads | Flat hash maps (absl/robin_hood) vs std::unordered_map | [012-open-addressing-hash-maps.md](012-open-addressing-hash-maps.md) | +| 013 | SIMD vectorization for batch operations | Array math, distance computation | Explicit AVX2/SSE intrinsics for 4-16x element parallelism | [013-simd-vectorization-batch-ops.md](013-simd-vectorization-batch-ops.md) | +| 014 | Loop tiling for matrix operations | Matrix multiply, cache thrashing | Blocking iteration into cache-resident tiles | [014-loop-tiling-matrix-operations.md](014-loop-tiling-matrix-operations.md) | +| 015 | Compiler intrinsics for bit operations | Popcount, clz/ctz, Hamming distance | Hardware bit instructions via builtins/C++20 <bit> | [015-builtin-intrinsics-bit-operations.md](015-builtin-intrinsics-bit-operations.md) | +| 016 | Reserve and preallocate containers | STL containers, bulk insertion | reserve() to eliminate reallocations and copies | [016-reserve-preallocate-containers.md](016-reserve-preallocate-containers.md) | +| 017 | Hot-cold data splitting | Routing tables, large structs | Separate hot fields into dense array for cache density | [017-hot-cold-data-splitting.md](017-hot-cold-data-splitting.md) | +| 018 | std::string_view avoiding copies | Text parsing, function parameters | Non-owning string references to eliminate allocations | [018-string-view-avoiding-copies.md](018-string-view-avoiding-copies.md) | +| 019 | NumPy vectorization over Python loops | Array math, distance computation | Vectorized NumPy ops to bypass interpreter overhead | [019-numpy-vectorization-over-loops.md](019-numpy-vectorization-over-loops.md) | +| 020 | __slots__ for memory reduction | Object-heavy programs, graphs | Eliminate per-instance __dict__ for 69% memory savings | [020-slots-memory-reduction.md](020-slots-memory-reduction.md) | +| 021 | Generator expressions for memory efficiency | Data pipelines, streaming | Lazy evaluation with O(1) memory per pipeline stage | [021-generator-expressions-memory.md](021-generator-expressions-memory.md) | +| 022 | Numba JIT for numerical computation | Monte Carlo, simulations | LLVM JIT compilation for C-speed Python loops | [022-numba-jit-numerical-computation.md](022-numba-jit-numerical-computation.md) | +| 023 | multiprocessing for CPU-bound work | Image processing, batch computation | OS processes to bypass GIL for true parallelism | [023-multiprocessing-cpu-bound.md](023-multiprocessing-cpu-bound.md) | +| 024 | dict/set O(1) lookup vs list search | Membership testing, filtering | Hash-based containers for constant-time lookups | [024-dict-set-lookup-vs-list.md](024-dict-set-lookup-vs-list.md) | +| 025 | Memory-mapped files in Python | Large file processing, log search | mmap/np.memmap for lazy file access without full load | [025-mmap-large-datasets-python.md](025-mmap-large-datasets-python.md) | +| 026 | Local variable caching | Hot loops, attribute access | Cache globals/attributes as locals for faster bytecode | [026-local-variable-caching.md](026-local-variable-caching.md) | +| 027 | itertools for lazy pipelines | Data processing, combinatorics | C-speed lazy iterators for memory-efficient pipelines | [027-itertools-lazy-pipelines.md](027-itertools-lazy-pipelines.md) | +| 028 | struct module for binary data | Network protocols, file formats | Pack/unpack binary records without object overhead | [028-struct-binary-data-packing.md](028-struct-binary-data-packing.md) | +| 029 | Preallocation patterns | List/array construction | Preallocate at final size to avoid O(N²) resizing | [029-preallocation-patterns.md](029-preallocation-patterns.md) | +| 030 | String join vs + concatenation | String building, log formatting | join() for O(N) string assembly vs O(N²) += | [030-string-join-vs-concatenation.md](030-string-join-vs-concatenation.md) | +| 031 | deque vs list for queue operations | BFS, FIFO queues | collections.deque for O(1) popleft vs list O(N) | [031-deque-vs-list-queue-ops.md](031-deque-vs-list-queue-ops.md) | +| 032 | array module for typed numerical data | Compact storage, binary I/O | array.array for 72% memory reduction vs list | [032-array-module-typed-data.md](032-array-module-typed-data.md) | +| 033 | Cython for C-speed hot loops | Custom metrics, graph traversal | Typed Cython with memoryviews for 95x speedup | [033-cython-c-speed-hot-loops.md](033-cython-c-speed-hot-loops.md) | +| 034 | Optimizing label propagation in graph clustering | Multilevel graph clustering, LP refinement | Dense vectors, counting-sort contraction, sweep specialization, allocation elimination | [034-graph-clustering-lp-refinement.md](034-graph-clustering-lp-refinement.md) | +| 035 | Eliminating heap allocation in streaming hypergraph multi-pass evaluation | Streaming hypergraph partitioning, multi-pass evaluation | Per-net bit vectors with popcount, objective-specific eval paths, incremental bit-setting | [035-streaming-hypergraph-multipass-eval.md](035-streaming-hypergraph-multipass-eval.md) | +| 036 | Speeding up a graph partitioner with bit-identical or quality-preserving output | Multilevel graph partitioning (KaHIP), FM refinement | Incremental FM neighbor degrees, fast METIS IO, lookup-only maps to open addressing, small-object allocator, CSR flow graph, O(n) per-call setup removal, time-based LD_PRELOAD sampler, early exit of discarded max-flow solves, quality-preserving relaxation (dense boundary set, incremental gain cache, deferred boundary updates with replay of kept moves, active-node label propagation), pooled stamped per-node maps, two-phase push-relabel, exact solver changes via the invariant minimal source side, Boykov-Kolmogorov max flow, one-cache-line lookup records, same std::sort over self-contained entries, no zero fill of memory about to be written, graph copy in visiting order for repeated sweeps, per-bucket pricing of parameter reductions, exact std::sort replica with a branchless partition | [036-kahip-bit-identical-speedup.md](036-kahip-bit-identical-speedup.md) | +| 037 | Cascading data-reduction rules instead of batch reductions | Graph reductions / kernelization, minimum cut | Contract on the spot and re-read degrees so reductions trigger each other; lazy union-find adjacency with stamped aggregation | [037-cascading-reduction-rules.md](037-cascading-reduction-rules.md) | +| 038 | Improving cut quality of a multilevel graph partitioner at a fixed time budget | Multilevel graph partitioning, memetic partitioning | Jet negative-gain LP before FM, unlocked flow model, saved time on global work (best of 2), balance-aware memetic fitness, guard fields with identity audits | [038-kahip-quality-at-fixed-time-budget.md](038-kahip-quality-at-fixed-time-budget.md) | +| 039 | Choosing the solve direction in pseudoflow max flow | Maximum flow / min cut, pseudoflow & push-relabel | Reverse s-t from a cheap excess-vs-deficit estimate; cheap flow recovery; instrument before importing textbook relabel tricks | [039-pseudoflow-solve-direction.md](039-pseudoflow-solve-direction.md) | +| 040 | Speeding up a distributed (MPI) multilevel graph partitioner with bit-identical output | Distributed graph partitioning (ParHIP), MPI label propagation | Per-phase timers with one change per phase per run and a per-component keep rule with phase noise bands, buffered output, lazy side-effect-free conditions in LP, open addressing only where order is unobserved, stamp arrays instead of reset scans, LP prefetch, devirtualizing both implementations, a type-local node pool for the order-defining hash map, reproducing the libstdc++ unordered_map iteration order with arrays (guarded and tested), dense arrays for the local part of a sparse key range | [040-parhip-bit-identical-speedup.md](040-parhip-bit-identical-speedup.md) | diff --git a/claude-plugin/skills/aae/report-template.md b/claude-plugin/skills/aae/report-template.md new file mode 100644 index 0000000..8bc0c5e --- /dev/null +++ b/claude-plugin/skills/aae/report-template.md @@ -0,0 +1,450 @@ +# AAE report template +The report is a LaTeX document, `report/report.tex`, that inputs files from `report/generated/`. +You write both yourself: copy the skeleton below once during setup, and rewrite the generated +files from `report/session.json` and `report/experiments/NNN/{meta.json,change.patch}` after +every experiment. Then compile in `report/` with `latexmk -pdf -interaction=nonstopmode report.tex` +(or run `pdflatex` twice). This edition of AAE contains no report generator program; the +example below shows exactly what each generated file looks like for a session with three +experiments, so follow its structure and macros. + +Rules for the generated files: + +- Escape LaTeX special characters in all text from `meta.json` (`\ & % $ # _ { } ~ ^`); + text in backticks becomes `\texttt{...}`, text in `$...$` is passed through as math. +- `metric` defaults to the geometric mean of `measurements`. Compare every experiment with + the baseline (experiment 001) and with the best value accepted before it, respecting + `direction` (`lower` or `higher`). +- `trajectory.tex` plots `generated/traj_*.dat`: one data file per series (best so far as a + step line, baseline, kept with the patch number as label, discarded); each file has a header + line and one row per point. +- `experiments.tex` has one page per experiment, kept, discarded and crashed alike. The code + change is read from `generated/diffs/NNN.diff` (the experiment's `change.patch` without the + commit header); shorten very long diffs so that the page stays one page. +- `accepted.tex` lists the kept changes in acceptance order with their patch files. + +## Skeleton: `report/report.tex` + +```latex +% AAE experiment report. Static skeleton; build_report.py writes generated/*.tex next to it. +\documentclass[11pt,a4paper]{article} +\usepackage[T1]{fontenc} +\usepackage[utf8]{inputenc} +\usepackage{libertinus} +\usepackage[libertinus]{newtxmath} +\usepackage[a4paper,margin=2.2cm,top=2.3cm,bottom=2.4cm]{geometry} +\usepackage{microtype} +\usepackage{amsmath} +\usepackage{booktabs,tabularx,array} +\usepackage{xcolor} +\usepackage{pgfplots} +\pgfplotsset{compat=1.18} +\usepackage[most]{tcolorbox} +\usepackage{enumitem} +\usepackage{fancyhdr} +\usepackage{listings} +\usepackage[hidelinks]{hyperref} + +\emergencystretch=1.5em +\setlist{itemsep=2pt,topsep=3pt,parsep=0pt} +\setlength{\parindent}{0pt} +\setlength{\parskip}{4pt} + +% colours +\definecolor{ink}{HTML}{1F2937} +\definecolor{accent}{HTML}{1D4ED8} +\definecolor{safe}{HTML}{047857} +\definecolor{unsafe}{HTML}{B91C1C} +\definecolor{neutral}{HTML}{6B7280} +\definecolor{soft}{HTML}{F3F4F6} + +\input{generated/session.tex} + +\pagestyle{fancy} +\fancyhf{} +\renewcommand{\headrulewidth}{0.3pt} +\fancyhead[L]{\small\color{ink!70}\ReportTitle} +\fancyhead[R]{\small\color{ink!70}\nouppercase{\leftmark}} +\fancyfoot[C]{\small\thepage} +\renewcommand{\sectionmark}[1]{\markboth{#1}{}} + +\makeatletter +\renewcommand\section{\@startsection{section}{1}{0pt}{-1ex}{1.2ex}{\Large\bfseries\color{accent}}} +\renewcommand\subsection{\@startsection{subsection}{2}{0pt}{-1ex}{0.6ex}{\large\bfseries\color{ink}}} +\makeatother + +% status badges +\newcommand{\badge}[2]{\tcbox[on line,colback=#1!10,colframe=#1!70,boxrule=0.4pt,arc=2pt,left=3pt,right=3pt,top=1pt,bottom=1pt]{\scriptsize\bfseries\color{#1!80!black}#2}} +\newcommand{\keep}{\badge{safe}{KEEP}} +\newcommand{\discard}{\badge{neutral}{DISCARD}} +\newcommand{\crash}{\badge{unsafe}{CRASH}} +\newcommand{\baseline}{\badge{accent}{BASELINE}} + +% boxes +\newtcolorbox{hypbox}{enhanced,colback=soft,colframe=ink!25,boxrule=0.4pt,arc=2pt,left=6pt,right=6pt,top=4pt,bottom=4pt} +\newtcolorbox{verdictbox}[1]{enhanced,colback=#1!4,colframe=#1!60,boxrule=0.5pt,arc=2pt,left=6pt,right=6pt,top=4pt,bottom=4pt} +\newtcolorbox{summarybox}{enhanced,colback=soft,colframe=ink!30,boxrule=0.5pt,arc=2pt,left=8pt,right=8pt} + +% diffs +\lstdefinelanguage{diff}{ + morecomment=[f][\color{safe!85!black}]{+}, + morecomment=[f][\color{unsafe}]{-}, + morecomment=[f][\color{accent}]{@@}, +} +\lstset{language=diff,basicstyle=\ttfamily\scriptsize,columns=fixed, + breaklines=true,breakatwhitespace=false,postbreak=\mbox{\textcolor{neutral}{$\hookrightarrow$}\space}, + tabsize=4,showstringspaces=false,keepspaces=true, + frame=single,rulecolor=\color{ink!20},backgroundcolor=\color{soft!60}, + xleftmargin=3pt,xrightmargin=3pt,framexleftmargin=3pt,framexrightmargin=3pt} + +\hypersetup{pdftitle={\ReportTitle},pdfauthor={}} + +\begin{document} + +%====================================================================== +\thispagestyle{empty} +\begin{center} +{\color{accent}\rule{\linewidth}{1.2pt}}\par\vspace{6pt} +{\LARGE\bfseries \ReportTitle}\par\vspace{5pt} +{\large \ReportSubtitle}\par\vspace{4pt} +{\color{accent}\rule{\linewidth}{1.2pt}} +\end{center} +{\renewcommand{\thefootnote}{}\footnotetext{\raggedright\hspace{-1.8em}This report was created using the Agentic Algorithm Engineering skill: \url{https://github.com/CHSZLab/AgenticAlgorithmEngineering}}} + +\input{generated/summary.tex} + +%====================================================================== +\clearpage +\section{Trajectory}\label{sec:trajectory} +\input{generated/trajectory.tex} + +%====================================================================== +\clearpage +\section{Accepted changes}\label{sec:accepted} +\input{generated/accepted.tex} + +%====================================================================== +\input{generated/experiments.tex} + +\end{document} +``` + +## Example session data + +`report/session.json`: + +```json +{ + "title": "AAE session: example solver", + "subtitle": "Running time of the solver on 2 inputs, branch aae/example", + "target": "src/solver.cpp", + "metric": "time", + "unit": "s", + "direction": "lower", + "resource": "peak memory [GB]", + "machine": "8 cores, 32 GB RAM", + "branch": "aae/example", + "date": "2026-11-21", + "summary": "Two hypotheses tested: the bucket queue was kept, the cache-aligned layout was discarded." +} +``` + +`report/experiments/001/meta.json`: + +```json +{ + "exp": 1, + "title": "Baseline", + "commit": "a1b2c3d", + "hypothesis": "Baseline, unmodified code.", + "intuition": "Reference point.", + "prediction": "-", + "status": "baseline", + "metric": 2.572, + "resource": 1.2, + "measurements": [ + { + "input": "rgg_n_2_20", + "value": 4.1 + }, + { + "input": "road_eu", + "value": 4.7 + } + ], + "verdict": "Baseline established.", + "reflection": "Profile shows 60% of the time in the priority queue." +} +``` + +`report/experiments/002/meta.json`: + +```json +{ + "exp": 2, + "title": "Bucket queue in refinement", + "commit": "b2c3d4e", + "hypothesis": "Replacing the binary heap by a bucket queue reduces time by ~15% because keys are small integers.", + "intuition": "Most time is spent in heap operations with integer gains.", + "prediction": "-10% to -20%; could regress if gains are spread widely.", + "status": "keep", + "metric": 2.181, + "resource": 1.2, + "measurements": [ + { + "input": "rgg_n_2_20", + "value": 3.483 + }, + { + "input": "road_eu", + "value": 3.976 + } + ], + "verdict": "Confirmed: -15.2% geometric mean, matching the prediction.", + "reflection": "The queue is no longer the bottleneck; next look at memory layout.", + "patch": "patches/0001-bucket-queue-refinement.patch" +} +``` + +`report/experiments/002/change.patch`: + +```diff +diff --git a/src/solver.cpp b/src/solver.cpp +--- a/src/solver.cpp ++++ b/src/solver.cpp +@@ -10,3 +10,3 @@ +- BinaryHeap q; ++ BucketQueue q(max_gain); +``` + +`report/experiments/003/meta.json`: + +```json +{ + "exp": 3, + "title": "Cache-aligned node array", + "commit": "c3d4e5f", + "hypothesis": "Aligning node records to 64 bytes reduces time by ~5%.", + "intuition": "Records straddle cache lines.", + "prediction": "-3% to -7%.", + "status": "discard", + "metric": 2.203, + "resource": 1.4, + "measurements": [ + { + "input": "rgg_n_2_20", + "value": 3.51 + }, + { + "input": "road_eu", + "value": 4.02 + } + ], + "verdict": "Refuted: +1.0%, more memory.", + "reflection": "Padding increases the working set more than it saves." +} +``` + +`report/experiments/003/change.patch`: + +```diff +diff --git a/src/node.h b/src/node.h +--- a/src/node.h ++++ b/src/node.h +@@ -1,3 +1,3 @@ +-struct Node { ++struct alignas(64) Node { +``` + +## Generated files for this example + +`report/generated/session.tex`: + +```latex +\newcommand{\ReportTitle}{AAE session: example solver} +\newcommand{\ReportSubtitle}{Running time of the solver on 2 inputs, branch aae/example} +``` + +`report/generated/summary.tex`: + +```latex +\section*{Summary} +\begin{summarybox} +\textbf{time} (lower is better): baseline 2.572 s $\rightarrow$ best 2.181 s, an improvement by a factor of \textbf{1.179}.\par\medskip +2 hypotheses tested: 1 kept, 1 discarded, 0 crashed. Accepted changes are stored as 1 numbered patches in \texttt{patches/} (Section~\ref{sec:accepted}). +\par\medskip Two hypotheses tested: the bucket queue was kept, the cache-aligned layout was discarded. +\end{summarybox} +{\small\begin{tabularx}{\linewidth}{@{}l X@{}} +Target: & src/solver.cpp\\ +Branch: & aae/example\\ +Machine: & 8 cores, 32 GB RAM\\ +Date: & 2026-11-21\\ +\end{tabularx}} +\subsection*{Contents (one page per hypothesis)} +{\small\begin{tabularx}{\linewidth}{@{}l X l r r@{}} +\toprule Exp & Hypothesis & Status & time & Page\\\midrule +\hyperref[exp:1]{E001} & Baseline & \baseline & 2.572 & \pageref{exp:1}\\ +\hyperref[exp:2]{E002} & Bucket queue in refinement & \keep & 2.181 & \pageref{exp:2}\\ +\hyperref[exp:3]{E003} & Cache-aligned node array & \discard & 2.203 & \pageref{exp:3}\\ +\bottomrule\end{tabularx}} +``` + +`report/generated/trajectory.tex`: + +```latex +\begin{center}\begin{tikzpicture} +\begin{axis}[width=0.95\linewidth,height=8.5cm,axis lines=left,xlabel={experiment},ylabel={time [s]}, + tick label style={font=\small\color{ink!70}},label style={font=\small\color{ink}}, + axis line style={ink!50},major tick style={ink!50},ymajorgrids,grid style={ink!10}, + legend style={draw=none,fill=none,font=\small,at={(0.5,1.02)},anchor=south,legend columns=-1,/tikz/every even column/.append style={column sep=10pt}},clip=false, + xmin=0.5,enlarge y limits=0.08,xtick distance=1] +\addplot[const plot mark left,accent,line width=1.2pt] table[x=x,y=y] {generated/traj_best.dat}; +\addlegendentry{best so far} +\addplot[only marks,mark=square*,mark size=3pt,accent,mark options={draw=white,line width=0.6pt}] table[x=x,y=y] {generated/traj_baseline.dat}; +\addlegendentry{baseline} +\addplot[only marks,mark=*,mark size=3pt,safe,mark options={draw=white,line width=0.6pt},nodes near coords,point meta=explicit symbolic,every node near coord/.append style={font=\scriptsize\bfseries\color{ink},anchor=south,yshift=1pt}] table[x=x,y=y,meta=lab] {generated/traj_keep.dat}; +\addlegendentry{kept (patch number)} +\addplot[only marks,mark=o,mark size=2.6pt,neutral,line width=0.7pt] table[x=x,y=y] {generated/traj_discard.dat}; +\addlegendentry{discarded} +\end{axis}\end{tikzpicture}\end{center} +{\small\color{ink!80}Every tested hypothesis is one point; the step line is the best result accepted so far. Labels P$i$ give the number of the patch in \texttt{patches/} that a kept change produced.} +\subsection*{Course of the session} +{\small\begin{tabularx}{\linewidth}{@{}l l X r r@{}} +\toprule Exp & Status & Hypothesis & time & best\\\midrule +E001 & \baseline & Baseline & 2.572 & 2.572\\ +E002 & \keep & Bucket queue in refinement & 2.181 & 2.181\\ +E003 & \discard & Cache-aligned node array & 2.203 & 2.181\\ +\bottomrule\end{tabularx}} +``` + +`report/generated/traj_best.dat`: + +```text +x y +1 2.572 +2 2.181 +3 2.181 +``` + +`report/generated/traj_baseline.dat`: + +```text +x y lab +1 2.572 - +``` + +`report/generated/traj_keep.dat`: + +```text +x y lab +2 2.181 P1 +``` + +`report/generated/traj_discard.dat`: + +```text +x y lab +3 2.203 - +``` + +`report/generated/accepted.tex`: + +```latex +Applying the patches in order to the baseline commit reproduces the best version:\par\smallskip\qquad\texttt{git am patches/*.patch}\par\smallskip +{\small\begin{tabularx}{\linewidth}{@{}l l X r r r@{}} +\toprule Patch & Exp & Change & time & step & total\\\midrule +\texttt{0001} & \hyperref[exp:2]{E002} & Bucket queue in refinement & 2.181 & 1.179$\times$ & 1.179$\times$\\ +\bottomrule\end{tabularx}} +\par{\small\color{ink!80}\emph{step}: improvement factor over the previous best; \emph{total}: over the baseline.} +``` + +`report/generated/experiments.tex`: + +```latex +\clearpage +\section{Experiments}\label{sec:experiments} +\subsection*{E001\ \ Baseline}\label{exp:1} +\addcontentsline{toc}{subsection}{E001} +\baseline\hfill{\small\color{ink!70}commit \texttt{a1b2c3d}}\par +\begin{hypbox}\small\textbf{Hypothesis.} Baseline, unmodified code.\end{hypbox} +{\small\textbf{Intuition.} Reference point.}\par +{\small\textbf{Prediction.} -}\par +\par\smallskip{\bfseries\color{ink}Measurements}\par +{\small\begin{tabularx}{\linewidth}{@{}X r r r r@{}} +\toprule Input & baseline & previous best & this & vs.\ best\\\midrule +rgg\_n\_2\_20 & -- & -- & 4.1 & --\\ +road\_eu & -- & -- & 4.7 & --\\ +\midrule +\textbf{geometric mean} & \textbf{--} & \textbf{--} & \textbf{2.572} & --\\ +\bottomrule\end{tabularx}} +\par{\small Peak peak memory [GB]: 1.2} +\begin{verdictbox}{accent}\small +\textbf{Verdict.} Baseline established. +\par\smallskip\textbf{Reflection.} Profile shows 60\% of the time in the priority queue. +\end{verdictbox} +\clearpage +\subsection*{E002\ \ Bucket queue in refinement}\label{exp:2} +\addcontentsline{toc}{subsection}{E002} +\keep\hfill{\small\color{ink!70}commit \texttt{b2c3d4e} \textperiodcentered\ patch \texttt{0001-bucket-queue-refinement.patch}}\par +\begin{hypbox}\small\textbf{Hypothesis.} Replacing the binary heap by a bucket queue reduces time by \textasciitilde{}15\% because keys are small integers.\end{hypbox} +{\small\textbf{Intuition.} Most time is spent in heap operations with integer gains.}\par +{\small\textbf{Prediction.} -10\% to -20\%; could regress if gains are spread widely.}\par +\par\smallskip{\bfseries\color{ink}Measurements}\par +{\small\begin{tabularx}{\linewidth}{@{}X r r r r@{}} +\toprule Input & baseline & previous best & this & vs.\ best\\\midrule +rgg\_n\_2\_20 & 4.1 & 4.1 & 3.483 & \textcolor{safe}{-15.0\,\%}\\ +road\_eu & 4.7 & 4.7 & 3.976 & \textcolor{safe}{-15.4\,\%}\\ +\midrule +\textbf{geometric mean} & \textbf{2.572} & \textbf{2.572} & \textbf{2.181} & \textcolor{safe}{-15.2\,\%}\\ +\bottomrule\end{tabularx}} +\par{\small Peak peak memory [GB]: 1.2} +\begin{verdictbox}{safe}\small +\textbf{Verdict.} Confirmed: -15.2\% geometric mean, matching the prediction. +\par\smallskip\textbf{Reflection.} The queue is no longer the bottleneck; next look at memory layout. +\end{verdictbox} +\par\smallskip{\bfseries\color{ink}Code changes}\hfill{\small\color{ink!70}1 file, \textcolor{safe}{+1} \textcolor{unsafe}{-1}}\par +\lstinputlisting{generated/diffs/002.diff} +\clearpage +\subsection*{E003\ \ Cache-aligned node array}\label{exp:3} +\addcontentsline{toc}{subsection}{E003} +\discard\hfill{\small\color{ink!70}commit \texttt{c3d4e5f}}\par +\begin{hypbox}\small\textbf{Hypothesis.} Aligning node records to 64 bytes reduces time by \textasciitilde{}5\%.\end{hypbox} +{\small\textbf{Intuition.} Records straddle cache lines.}\par +{\small\textbf{Prediction.} -3\% to -7\%.}\par +\par\smallskip{\bfseries\color{ink}Measurements}\par +{\small\begin{tabularx}{\linewidth}{@{}X r r r r@{}} +\toprule Input & baseline & previous best & this & vs.\ best\\\midrule +rgg\_n\_2\_20 & 4.1 & 3.483 & 3.51 & \textcolor{unsafe}{+0.8\,\%}\\ +road\_eu & 4.7 & 3.976 & 4.02 & \textcolor{unsafe}{+1.1\,\%}\\ +\midrule +\textbf{geometric mean} & \textbf{2.572} & \textbf{2.181} & \textbf{2.203} & \textcolor{unsafe}{+1.0\,\%}\\ +\bottomrule\end{tabularx}} +\par{\small Peak peak memory [GB]: 1.4} +\begin{verdictbox}{neutral}\small +\textbf{Verdict.} Refuted: +1.0\%, more memory. +\par\smallskip\textbf{Reflection.} Padding increases the working set more than it saves. +\end{verdictbox} +\par\smallskip{\bfseries\color{ink}Code changes}\hfill{\small\color{ink!70}1 file, \textcolor{safe}{+1} \textcolor{unsafe}{-1}}\par +\lstinputlisting{generated/diffs/003.diff} +``` + +`report/generated/diffs/002.diff`: + +```diff +diff --git a/src/solver.cpp b/src/solver.cpp +--- a/src/solver.cpp ++++ b/src/solver.cpp +@@ -10,3 +10,3 @@ +- BinaryHeap q; ++ BucketQueue q(max_gain); +``` + +`report/generated/diffs/003.diff`: + +```diff +diff --git a/src/node.h b/src/node.h +--- a/src/node.h ++++ b/src/node.h +@@ -1,3 +1,3 @@ +-struct Node { ++struct alignas(64) Node { +``` diff --git a/img/icon.png b/img/icon.png new file mode 100644 index 0000000000000000000000000000000000000000..b4b67a04f0ccffb65c2dddcca048b8099bfc9895 GIT binary patch literal 98293 zcmc$_1x#FF*Dg9(DNvxeOK~m5tw@V|ahD>+-F1-SPK&!!+#LpYcX!v~I?SE6-?=xr z$$#>nlarhY$)04`?|x;iXFbnaYlnW3mq0}(L<Rr=s8W)mN&o;n^ea5zH7xY<8`wGn zfB^ucL_evxEuE}7d+FXhApYL>+-dGBODH%CQ=wO)JFET*$7MidbZInT@TM;)!ukWg z;0iynR@MZrbb7PQ5{De&xHJDH7#v%k$IWe?vPZ++lzDf|IY%tCpW%6Sl~z&J-kbXx z`Zxe!D!^Cf?-$=cvw(hL{#^i=|E|J+W(EC7{<{E>{$0tTbNut6`gi#+ZvLPD`oA^w zfA7KnYeWApt?9ov^#9g&{a+jUzxUw(;*kG~oBvPP&;M`e|C6TqpP~Q9nx=vdoc6aE z0K(`uS14ZHj|_Esd+4E!gi!N_9%WV{BfZM|bexaJJ->UtD&~$@3K&ua`OE!tfC(7_ zDnUn`?;&g?kKa4191DYj?xI=6RZ(+_S#`+%^FJk`h9}SF*&;wTad(2p3(A~*d8E5x zW$gH3BGIu}q-(btF&he{P>SygSbz-;sYZCfgd^u;Z;JT4{<3~K?g6i*gM->hMcXg; zE^YRcT31nEw)g7j@e<F3$=qQ6o*U`0TSph+eVjDa1C?$m(*Y>FK(|YHX+iD%eq(Og zsI%5hjr^NTXTy8pfJV2n!;*8k*+{pun-5K+Db5=Bj9*ngr-@Q*iRNLd+UsR22g=8% zsoE9K&<Fj5F5kJ-_JU3a0dUYXmVw#2S=ymU^~U5fr>@VIJ7vQDivp|mb2faI0M9-m z&vLFZeW5_w>7=6lYrw__H>_6xf#uuF+>&MX+Dx|EOb0NppELiP@VVp!Tedz9C*q$| z4g^BQy97oE5y<c<4j)K1RC}bImr81s<UO5`DWUuK2So(t*Wjwhq<B(7<go|Hugr~d zegyTxe$}cUthFmMSN`Qk`CfJsej`cEG-4NPH=IN4W0Itbqh9eeX}Yv?eX#QDOn$&0 z=s5Ufh|E|L_iCray3dyii7HG7TM`p^fFr4xJS$o#8<EDwI<<k-u4(&t<f*L5nSxT7 z&^i`1?ntm3=F|nz!caNz;y2aiT<ZIU<}D5*OeoBq1(?{z*)O3MyFdaMQOOv`G)@YC zgX5Oe!IZ*A^IR-(oO<Eb@m(AUzE$2MCKLQ>JYzKjH6OaL;!pIO49T=NKkPmbl1~`N z?n!be@4yi>u!q@S<{JM4;=BUcLN0T&h!9af<4_`FHljh}jp`oiSok>$;{H&2i)YzH zIWm4pUQ=m1Cwp0JHxq$A>|^=Gh|I-6lUZDGZv;Y?DRa!xgnNU&)rezJJ$YgQ6c}M! zKL%mIL`~Ca%5p)noDOEA-A{}ZIy(z$euIceCN*(fMj-*(O2d+`lUmbD>924`16y?7 zUz{z~rJLnQVu?{~eI4UIw#pSJa#9P7g8>|DPN32Ou0Os~BvUMnAR{K>rkmqdr=$p0 zJ=gfMhmPd&L@^>9!n|mvRXbOkKjIsmHJrwDd2EuQ6s$6Ao;x;YSt}(0X9)P?k}<O2 z%bynkhDL~D64&ae!H9usHSe7dPQB<D=#D1KrZ@@~a-92y_&D(7o5`e}nL-Qqst!f- z`K%|*a>rZ_@MG}*1~8{rEn-AMbX+!eR#s9>-PsNGNt0QreAtl^IJlwCA8eL2-!W@h zYSrQzg;7u1ns^&3G&X<DVJ5VmS@cLRo``?@*I0P8+zF0CDJlEWqfrluj85IE7-Uuq zPX;^<RgF3aSOYDK<>idR$`%YG7|f_921!|u!~~er$~SA3j7m6{vU3kV{2j(jS?cS_ z{pKW=xaiZ4g~T0`(aP6DN!X51ug<a1B}&$MuNwPFzUd_reC5*VR>PScQX&yg-m*!v zeDPFXWjknihGMBN4eU1Ms~&k`g7l9cQ%-02TS;2P@c6I|AG-a^UH1q$BkAkS(tp)r zM#4u)y}F>4YsE{GOfm})FX5=iJq?2S6S(u31B#-_2>E^7oTe&^R#%WdW%T9^lN2+N zhSA`oW-0%imkggv!7b;Ub)S^Xq6|WLwt+jBGtI-Wo-?nLE^(d<k^P4%d~Xp1ZN515 zEgaZ)=N8CEp7O1r@f1|safD{$?|WU5cIVU;<fqE|y(UB=4*UTRE^MnER~>t>8B;s7 zs*{c%fCv1U&h?P5sT<IwZ(-o4Vr-P^oUMYz*W)iz{slF{jf@l@RC8ySBT(j3)U3Oc z3{M@Jdd6ejDuhNI<YiPI)kKm3fA-c_FB{m-=4fSnbRZceqi_@j-*>vYiq@E(5=l>k zW!OHqYO&Fe^ifVe!LdEE##U7;o4pq6FQQCb8Mn+3Z0E4|e}5u10Sfl#bf5adVSJiy z-ej0{`))gzj90;)pqkrc2a8U4?Q^KElw;SfhL>RGRW_z#3=8-ZIF!;>bL7ad(phPs zz&3T`$xY-OFcXW@&5Ens%s#cGO%1g%oyNJ90}01Xvg-pu;b`5MY2<Y8l1pno$_U!O zp@qDTl)8uZP3Tw#aq<Ta2IqA~?SQ=0ND@uf#Z%v9sr}lo4XtKn?u*}ycv8+H;dOd4 z+3dZVLW-iMT?wCFjRyWZSbwHMm!+KvdMpY}`L<<raZKluUq-RqSHVhjwPEyDy=s3> zgp)raC8L0K{cYblCj|`8h)sIC(xNM2?-%laK`uNLR>;Q9>2~LLQe0BZ=j_+68t(mQ z{dkyeM`EdKS-(#`wsY$CNS>z4V>_H;*QqFBB*{k}Fe8~K76hX>{#nN5u4!>08D0jb z;8yJsy_*|X&3l)kGuaOmMOsICw{WOiTXh8$?cA!|(&i3@x=crSS>{fuE{E6geF?>M zZ~vZ5419H-0s&64054k(C~m0Voa-P6O)L^4P|QexJZwpghDFQr*(G<7>}Yp15cuv` zd4i&u<Z~L!<N_W;i7*PxUmQu#0cNmyIvDRQRL-uz9T_<_l2qnYeaY+ZQ7D!_cEQIj zV`pABj<}g(?Irzi(2LJ}YIZVZ?gyPkUY78`rATC;^(f#I2$))!xe4VXlGG10mIc2- zoTK;0`2WJrT{&I{3qB_b>|!q7R;{R!e$jVW7=T&icc|CeGl_#|1~+TuU+*^Gkh=aR zG)?21;)ugQf;|*CjVPnxGcoM3*<8LQz^=Pn`OdW!^1^=)HgQVnQTSn&I<`+=D{V@w zj`=r?TUP5HxlezlRS41WaJUOkktZPv7shT}gL_)hsU5psXo=FB(>5dhJZHqEl9er9 zy!q@H&oZNdkMPe>Q;1l}J<fjq-q}>l-bX|#p4)`?{h8rXeWjL^`7$x)sp}de3E-1K z%VjiamfQ<h6;v?cl!W`^q{1m!2!htK_i6N-ab0W5-lFpaiW1TvdNStrM_O_A3I*-1 zhxpm?MuGhC$p=oM@;~kgBgdx~a7oPnqL!o(*AJh*i%oWg*CssIdM*|uXJh)<JU3}~ zxsG_+`pLG(Wp}3UyWbk5E0CwqO^L=P9(rZ;wNFcm`Te^e@qMbxX2izHHbsFR-w4CL z?0zepb25G`UCbO7W{$dG!>F$DX|F4NF@RcC$j=7*eizd_f6^=d89Mzx{Kq!_aN>Om z!AEN)jk8mIZ#mj*ksEi>*g`i;E&1!~;9K`$=C&uS5i!veFL`l1$!DurU>OP--OeY% zzaH+P%CoD+l8%$c`<^1df1)@=-k~28@NRbexXLvh%3(b2`b}xhgqfC9jr+cZdnI&$ zrWS%3ME-MgYnRpT)Q`{8ed0Y%kDBVL;v~fzKZlTf3#J1qTGXsqZy%%4Z-C9zs64tk zniLSN?f*<#VOB_?>11aEiIRc|90apPV!cZKIt@Q}#*#SO>|~QW?eHxaOOaFy;EDy8 zTfniw7zz9i^}j!mN=_YEH7t)LPfAEhAw8UF2gYB5k@?*JgdrU#86ohy*b@9+B*n|C zsWgD3>9c7*jH#qk^#GpPvq=0y%&WhMs>0<Ud>yCRSw)TfKCn18TSS3JaicilGls&4 zt~2k+Ht)<d!Az1@Fp{wx22rAgA<r`Z@6+?LU>=50O&=t1DX)bjr}^%`V~I(p!}zs* z+C1CW^(@(`eLT5TM4*eG@@1#yQ{B<GxQ><n&zhwN>`T<sNt3u3NRA&HKAoGrN#^VJ zoAu~v(YI5b%U=y3^!6<<ChSm%J&=hP6#uuw&9U9Zx;_qew>y<on47c@cEx=>7ic=z zei%k1w&E+HlaSN#Mkv{8_O@fQ$2;5oX-vf@vsEa4oj)X`q6ru${=+RN@<Bb-KAqPS z?$aXQ;230TgAV!5H5xMFB*T`{gY`67*1%qT;t<iF(M8pWc`#04Js_0_iT_L+4;l$Q z9qy|&KPBq+&zhwcI9KWe6*r{s0y#ZP@;vqIJYmh3FmZ2y9`IO`i<s#YM`T4-8GYD! zFn@Ffxl}<^m<}(0Oy;jmrCLUq1GBK1$ZkJtmc(=KNt`!@wn98+Fe?E`1T8uRKg3>| zyVM<=*5Lny9=_QDn90mE0w#J<S|N!a05g&kc|GiwLXzK)t3il*_Tq--B$B(x`0!c5 zgBS<%;!EYn5Jf0V=lq3OyAdd!D2$^KbAJqzv7bQ#oEz{Ph*Q1JbG0TZxoJVYYuizp zaf;4#zyROLioidX#Ij7?FC)J~{fn(6CNVcM^2ixN%v9T{OPj>Pw7+sk0Zwhl{z*96 z5$K>p(Q5^Kgi>0pkZ++*{#l@k>IXCYzwQ(@7s0w|xd+ciVs_{$t5TMWB#h?*Uu%|r zSmjER5@i-@yrtRrMj$P@iqV))Y^#FaD?Z8#?T=yCimVgwCgME=T`SttN2Qm$R(@`p zQMDibgF!;Ijx%XA6gk8?{(4u6Y>$yJsDGWmBX(i*F}@H{J{24Hx8tJJOq5F!mD8YG z$3N)l?st+TUBkP353$4-wq^K^@$aV1O`2`fjJNGuKYeE-aAVR#9Sa=S{Jw`_gi@Ff zp4i>Qk6Sz9X<iU;CWZ?Nd#-m!nk5WG<J~_Me^8)~eD`mML@5zPZTVOxXc3)#;yzDH z9ZZBY=X7u~YjzGPlP&EtH*8FY{s`WvOsxKJOPQQWF=>gc9|`4XP?MO_te0tf8R<Zx zoa{AqzydS{APFl)1!D<B9)NHTV2cug?GZkfS!K;HnfRx)UJHNkiTE_bl7iD}ptV4r zJ0{}riy;?H*I4T0IbhPs<3x8q$C<|KJ<{$bW}e?ySOMdlKtdqOzb(p(R)ot1HBTgN zpXIz4WPhbIS$Su@rRAY#y*Bwq#vqNgkDuj}GRN(2qA}>v-Xl}BZ=~&S<lCiq&56>6 zKfC$qDIOA`^8UeUU%k<9Qx!tXLY7{6!?$J*Z|Dgj>Yv!=Prcp=R5Yzkx^#$AA(Na- z(8*Kvk-E*+`6R-u`_GX-DsZcba}RnI4`fDFy~0KSI9X$&0d$R*l754L+UxO?LKrvc z&pcRiP2Dyo{%Q5<?cJYR>0HL1Ad(TQV1J@X;PRzC($o~4-^;hHLS#PsnPB(>ik#{6 z@Df0bl}+j+Fwv7GgMf{o`5N!N=>qfFOXAyDa!qGu;!^F{jCp+nowG0Qw6q09{f~xh z_+Np~(VkC&1p43WBXP`2Lxy9bjC1TMA~DTVgVbOEx|cN}GJqIkuQB^hAf(=BYC;3= ze8G%yjO68*EY7c!8(AqOiZ|bt;}kQ{NN%C^(`J<ETW|vk6fGS^;5`Ng#xZ$C6am6< z+3*1JP-M2*mWqosADx~Fl2Pkdvfi_1#}JWb7R``jzZB}jWB@)YiF?HWKD}0dZmd?c z+rH!3C7~$f)S-ER71OR6kvc*=oI^f`gzzq#4nS`HyA(lkOg;P}-1HU1*%Ca_zRKRJ z{_P@uu2jk%s|Hb*g^)e<#d<JJ#bN5HynVm0AF=B?%B5*$a}vOC+`F7<uRfPu5qR50 z4w%SGn$(wb@+%6bLk38qJ+r|$wL_7;g5~07{CF(ZyL*jUGz|?HNHcJ|@)cCPd}G*o z(2^Etiiss^RA9RPboe{lObb!OU2$V*5)@imif0St>&6n6Ts})nz9Me@9r$f>u$L{J z`~ui-hTfkJGcouD;dK@}0TuXY8m>X-WVrHp)sss4-j~{_bW-qxO>-|No8^Rp?AT@x zK893ZX^5-mOhfKne<F!BqU=f0GR%2d+?fa3BxG{%8wX-V-i95TI6E~PruDp}DHNyq z&bcuX0a-N>g!te)+oSPRTukL{*J6)hp^z<<iDWawuCM8eduSj|5<!x(+yGP(YyOSX z-Aj`3W(+VH^UO<dkG^n@Bh^b~D7rY2y|4W9H|X(gED(r-A8Wwe8AO#S$_KR4NMx1Y z<UcIEs%=ZM8j=eB43Hl&JCKxv70;`0`HRXzZ3oB_7$c;0lM&G(WLXo)M7hZn4XjHA zu|YWiU!x#>dS4@G;CFwX3xd+6!X-b}PW+p58&3jV)EOHfgIpit1XN}<dUyLl6=&-? zwK`SMzJ){<!Uy6_zrKY-Oz>!r1RIGu+y;JjA9Y97#=ZiIE-ycM{#xpA+SdkS^kzM_ zI<20<Ucw`nHkFob0_Hub9||(Fl1B#o8<Z?H{{w_we~|nPW}FW<PWNg4xmoV9Jvz3| zmfI3>{cC9b%LKz;QkdwtT(|uGE($*@Z*P;}6mB`>C6rF1q1j?u#QY}qkk4|FE=L*D zSL|38;@00esg8sRAdgEMgK=GXXKL~--M;V4^96scrf}NTzMv=g_4PRZtWhbz7wr=f zz&9$1gX41lSRimwYl!q~PYq|u4q;yp7aKgpui~ej>Z6xM+K(y`5eHWB$vJ-5l!mV7 z9?y`M=-Fs$-5!)xz3iR#rxT<8x3R23LU+$s$y5(&!1Y7VK?4ruJY;#q-?MPgG((Kl zD^1YD$Mrf#|DYMtRaNEISK2}~mc~TG_+{RN!)tPd3eg0{YkuTEFw6<3v2aL7bgjHM zvG_MRcUWIJ_Kf%04Yv6|<t2k{+HQi+qnP+<6~;d9{@jQnsVbKDStV!ZQ=}rg809;> zM}vi`RLQt4*j*N(n+SSEkDf{KkJ?agl#%_d&JyLul}a42{ewUoVsIL!VW<D-<sUb| z&qbV6t~~q*uJE$)bo4`A4>#f#&LV(<Ki+q2uI1l_nV8lW`pCTO?E7hsI}vZz?@F4c z1}Ui$UUSH2p%Mnh!2Ka|2Rae=z3d{Owvpq?JZXtcanI0}^#`cqYqQg`4HBdDsMpS7 zvtU%LzNpL=1p7Oeu*sAgety<g%dJ!BUTc`zcf^KYtHNDQLnA_qz!CswgLA3Xyg2y+ zyM8GZ!lecBghrIU-tj2d;y-ti&q%w+V0e8ynd>knNUNiUOV4<`0&XT4eft12#-K>@ zIxAgKZhcF!5gA}T>}1xUyGb&~qN?}2>`<^5d0)o))U$`>%&gLI)eA>-W@%m4Wsh!* z+T6Z-y&BGfrULvyLzYixrybm9F~R#F(v^JxA+zPjD`Vd*iKSxO4}`^GWiV*{Vh|Xy zk9<%=G0RIMNBGcnMMcWnU7jig9M^v-o|$)jWcA9fR##P0#QsA?Ht^FwP&f>W=~)iJ zdj@WkjW9IT*EMbSp?W6hBv+bO7>|&;>^Xc2V@XHnE?urNa7pn>f!)51O4z*r47;Cq z2$ftRFD=_+w0cxus)XU-1%=5~+N3$Rg`Q?+D)jmB)QE^=$h<1gekkyRKinQ%N8^`> zp#dgnRA6Au;+YPGgvMQO2tJ8ew0dG2dk&UHL(#L=d1OyV`xfaYzoITJ-}0TJi)IZ? zg9bIr(K5z`wL<XtPuE(K4;a7r!WmAoxpmg9MX%RUpL-a1mh)<CId7|maDUaRLQjSw zO*UdA<gQs2vcA5T!MaVC@$H-A-mk)xM{M^*mVIec+;Zs2VE;LpFC?#_<f>zG$hqWB zcR!hjBY`hi<s%r?a?W02Onbz=nttAOV>qFvEAp5fQI9(0;-X}cv(;uEIJNtWmROhH zKt9hxc2a|mV&yLG_~}&Ki(lzek_NS@wo|=|X{N=AOnm)8>1We-hHP3GPRvh9tkieZ z<z9`)<t?=a6MV*!T}^ZEM5^~kkgS$gnn>UWBhty^e%#Ysqy6EySzwwsp=-c4+Wz#< zaAW&u5NO7hajB`2+DH~EL>h51M1L^v_cR{Hu~Oo$;9XeB2cxSS=m;vsMvW!sR3!~j zu)f1I3LEl*iVwV%)I0p&r;>7thN@}S+g43udzL4T@h-<B!%{R3A9{QTIiG(4hcz{s z3+I6^z3u&^w)aOR9et5e_S;m!xw~v}-Lc+#$(`2A{B0(`h5cw=*GlCA1c86#gG2;c zMKV_7g$^jqP!8=W^SA8q(UW!Mw<v{sC83$dcP01z{>GuHHZEu-&_H?R=Rj?gP#);9 zE-`{{t63P1WzMIpf~y>87a@jIU?DqAurDgQmSs2Zc(#gj@T<K)1fyibs%;F{%ql~g z-)U|@irnD?rkyjNYNq}(#qssDw!AtKQN!A^qX02Y&Op02=Xsm$z*rR;Ny-Kzy;L%7 z=80}y@iZc`9a{>$V!Hfp(Z9Xrv}tiUugfsc&HZKfp}Hw+JWqo8OwljJmi^W9!0LGy znU7{}-K_SDUi%Su-E!s}b~Ll8#qcb((JQ0RS$i$_lIqc{pXztz!-2T?D>^6$$7)>Q zWTD55v}|tf6Ze`s?;$?g+|3Qf&VnqV;@6w_%Gl7w*(x4wqauTVOS*ai2cSZwP?8O+ z+%a5>6Xn2qd$mF=qaCWKs^I9@{5<0ceNqD@|2t|=Qm=tKFghH-QOx}Q@wj7-c`Ou5 ztfbBaZ3CaZz~fA3ioa<^l_G^n=~Zqd;o*HF3{>^^5b%;ecA#eDSTW<E(U=~vJZAom zSQJ=F@oufS3@_axBO(b%`4)JjeyZmb3p%pYy*Mlu(rEd(wRm#msW%*{!QRK60L<$B zXg}Q@NyPR_hi2$v7YX<7;7hj$=bQ$|W@qWAc><<Dtr08>jp<>*198v-w$XQ^&!zMS zwr$#3|G3%^pVOyB8M6pg$NJ=&Ucx6)6Mh~PyK2_zcWfsyUm-6IRasfPPUn|6&^1=> zUJtCb)n$5ZrEv_#{J???QBsP$r9Qvh9po8<ftADjm7>fU1`9~3wsC?(78b=z$9|st z1;dKqLul1Bjncnn>VWq2PR~|@cEBaqn;hdX47mJ!DBZ|rQ+D!>8!zI)Xt?YFjcMWP zG+Il3oY2X8vk@`M?L+g{`b+O%m4T)L#ETMnxmrT6Ww_F(@T+aS*OhzwbTkT?ec6!5 z#o7L99qAjz<)>YCL9Hx%STdRzn0JOx*e^S@V4tT{`0>auW~&Hz;zPsB9-Eg(j8PZc z+tW3|2EF$EwCBFgIj$<V*<tAT0BS(YEfWoRjLdV4=RH$iN3_535CNNKw(*dSpd#14 zR9{2^aQ@Vzvjy@afidVveX%;4TVtDllWQqFw>_P?<^Bz4x$B+4`NTJEZxM3B(ems_ z&-KG>GMfMejn=j6KD>I`^#)*DU$u8;C)V3gy6InXd`b{`s1R?-7&iO;?oj~+#;}D~ zJEum0bE$xg3Ad8+i&eUNhi@mk(8vzF>h{pfOLU(t$lCL!k6OlI>?3>m`xt`{-|tS4 zPM6~IjJ}CdA{-4Lc&Z@6QbBbWc{&ar9q_??JBxx-EJ{K>AjhiVo4l2^A2f)mf5r)= zOiIa5hATHP^rDTQ8@Z_C&oHmzQ8$V#%Wlhg_4I`K|DmPUWn7%u5$QPb9Y9C_&~ ztXEx@jlYTWaccq<OcPiq0IZ<N0H028vjcreGQKK{vAN?D$4Ku&B}!TpN^bigw+p-O zmm3!4bwMOjQ6lcs=BG1C{rxYJ4m`wI+nLTf^T(ijieD0XbQ4seb4S>p&xN7TN{*WD zqjPXe=1xLFg2yHGtQyW14x}2o{qA_|(pDs1j0MSJ5R|*&y;O*V#0{e<)<2ct_R2-O zHE6NZ5?F(md6Mk5A9cf7=5X|nfGLE9WhR%j>zWyE>n~>%%R#F_9+{G5&mB+xT?o@6 zISU@wkXA-wvFpq;*q(aF+q%B?ISpphSAA>~p>HuTc6D7e*Gv*=-1OguTQ^;pS{NC5 za$yR5iHB_8BZtM!syz%b4hs4p<`PhxPW(dS1>`Y)93q#~wJ(P3T`84UjMKy)zsxK# z)i|0FG0vTP(0@%`J?Z35N$b{@XVJEIB)WMblGa~-xZ%~R6^s|^PwG;$;(P&H_s8Fv ztY66XA<C52Wl#c(_)I;EF@qt;u0&@(YdE2si?4;usBWA3&Z4V`x=#oK%O)i^KU!A= zy<@3ZudtEgv4l8|2f8NKJMzbRKNa<w{7f4w=S5+JGUaEd*zMb(>+j+nF+L!@>K9<* zLlhCI+-?9d&o+tnSwzA#-0b>^o|iolqZV5qD&K>$;l-qNk<ZOo#h7f2Vxu@((rvMo zU?~|or<Qirt<8A*o$~Et5n<WlS7@@g(IK3*z=cW!^un~xtGs0i162cF_i}~)T>f@c z;yD-of>L=^6v@*kE5QN}-#zz65~f9#8!B2hx^kChE=tsp6Yg215~A=I;|pt4atpBw za_9h4Z;An*JO(Q*fnd$;5SU+jxdJnl?;SUyHN$z<3{JO8o^>K;YU^__%j1|IaPN{D z53BQqk<<f;xO*g~VpD>uo3u*x)M1;qQ06MiS-efw!}LWzrYilewj1L<XKk$+c)V{J zLs(C}QyorO>&&z+(adDKtfaDoR=sj<c~w#E0mRATwV5?Id{4634)Sb@^)a|&M~kNE zs&3skNbRX%%{P0QjLy!d>p$Po(ry+MuLBiO;G0h(9$#~)<#7^<#lQ*kecGRk5r2h4 z@2Y=lzj!?(G*bjLT^-x630@%<3P5ttwONOv-?GMC$Kz9aps8eEY2zV6?%DNEU=icX z{TCW7{hHH3Y!G=1V?MO_gImGQdCy;{oiR1k?ZJ`^RT<a8E>VJAn{@OM%+rf(=~k?Q zsyXHqEN5-*zG=yekh=zIpOe;G3%ldF+$F7H8)0<8Sg)7&f)$41=mZY|tV;p3fIQTn z0O95&HU@`rPSWsSsg_=Blg9SZY`nq9{));0J=&+fbYB`Mav{_?Cn2dewRljsW7vc% zD;u~eDt|?Sa8c%mFi1hf?iH_pouzOp9B{wLc7=7mxc|P|dfD`~tKRIhK~lQC*Mjke zD2Hp+O`yWb_dB578^egxtrN>Xpz&R~KzL5iRV4oT!RM(6em02G#AMkAOf(f`gWcN> zYVQ4LApBe~G|Z^)RduW|I$#zgTF18@ObnQUlLnBJ{4P;I*W;mFh?FpJKpQ7>4qg<{ z;@W?CYvwSjURRlY?QGj?iiD;Wu2K$aSVtPKlcl&6e`$BRpte3K_qiJ;0JGE?%ZH?9 zC_Zlowb?>RL%s)7e&snA0Z7ASskJ{TPxm10v^=H{J*w>?{$><`%S$S&Mg_xD+sL-W ztEi(e+$l{mK@UTzSs^9a`SB3xo7qyi(Auz3Rqj$sTI^;9SK03M@3Ylae01=NaDd87 zwZ9^4)}06v;MxHy_>QH|>9HtRZ%Np{;YPK9%(_Hq`VAE8Np4e;k;FZ#O3vNKo|OY` zEN9HDSDG=_T`c!RbD3YLlPBkdwpIymPOXeU80QHiFo_u9Koc96$Wj(>P7|SuHq|M9 zeQn4~`S$Y7-tG0iBhzI~Ehqn*b0*Pibg2#vlB#oOE`?l%F`Wo#MM{!>$*MckN4l2! zzFMX2$!0dymh|-16Zlc>X`za_P#yGC4qSb-`8e$d8~OIoGt!U_^ZOb&<TH!#M7RlW zosi!00*m@_Ih-AC`l{*Hx@f?0h>U~oueHnQBj)Hn+7I<FBBm_ga_DWEF^d=Y&PB`X z;uo;x+;|)}W=9K)CYy^#qi3KXIV<?fE`nLp_Svh>y<boy@pQ&`Vse$%Df#nO4Yi=^ z&K~K#OViHy-F_O-XZWz{=!nxaBR{xX;%@eM*>pvMtd3YLso${V+n7q*s?cvq=0e1r zQbJz9wH_Ueub!pZ`hvdCT+;DbJq^g6@}<J8OCO<1HyxghX33#I@>K)q=y-3sE=eM$ zbnaw(3)Ea4z{>jN%9?2T&0N_CUAs;Ulc=bU<-*fm2K79=Wn$q>e{IQUpKZotqjSCl z8@74R-trYrzGuJk9ol*|(<>FV{=w0KRmU^I`B2(h2ndoZ>OLH`Y>R%r;#^BaeBGfx z0*6?icJ^?)oN}j>bN7`o_ZO!1#6URlpBm9Mc|U6O^95n_;(|5e1`}Ti)G#EL%V*+v z*UV0j%aU9*i1k4XUo8z}`jYxC(QA_m>&ol}ic4T5I?ewG#zZ}J&~bjf%$sJV4_)m# zCO%sN{mc5xF6sjzz3=4AB2G7<zE?IYNj`+<A*|fGIWEZ9gc&ie`ZOtlGY+z=)gI<~ z*b(Gj|7oZ8@B@N&x{^#`YE7O(Gx<Jpo*eOy#1QBr=tl&4S1DjOpZ;Rj4ht&;y`w5> zP$n&UnbWWP5<W*@-HWtdl&nh`dfd5WG79U?iOb}ABfw%KMI$M`41flZC7w8F$Swla z8lI+GUlY#sG92}Aw^zP^N%n`zXcUBdPLU6P#V*|OJt<L>6+UC1v6*$GyEYceS@Y^f zi#M#fO;4)5u<q=mLhe8aj{Xjl6Hz4jz!yj9^cGPLu^htm3C_2FG9AwIX`3Jmml$VD z=4Jm$SM3_KlG^SYjfWxq_=@IP<(F8HMYicWV*3!wNcq?5{L0-UXnOLcxOY_Kwg_1G zC*Mb^Zw5uahFs3ybJ!1NTqj%~pQ=t5qfhx6-;j`Ypk#3?hMev|#oG{16M;$o!hDzD zw~-FdGn#jwlnH30%CQu#D>QRCIBJ0bP&X?!na>mzO|yNPEvqUnR(bM;&l#kkPqURk z`;#|U`q6O@(cMfJXiBwgk}OCQ`~m6w-o4ats_o;rQsTj>G~IYEAk(*ykpBFEMw7jN z{@kbOq326?O6eA3@>_lVjuhT1o4aA_2-=BLl}OxjcJLKRzvsFJ<1v*}1NsC-%D4D5 z=DCGC!odif{E@6Ko%vDkr=;+3tT9ewlU*G<K`gQLqx0pOx;RAA0Uq;jU7DO^J}NhT zMD$M&5txHgUS;Cn&pyFGtFY3n1CIjpZZzoO&QGz8Px>cJY13ZP3pKx>#~JFQt4f2P z7I>j+DFJR7Qh$M_B5&Db!`+k2YE3fVLP%^REH}=!P?5=)ScRSou2P-9&D_H9P@q@M z@@Z!&{nU6zsq89|>OrT%4fHX)d*9z^pEH|6?ifq8zcYAtI1QJ|4Gq24Ru)i$*wb~e z@6wdk1=s68e~uPce#j-D`vUY+JI>q2d37B%k9P11+XJRW69jqe+w%YEhO$<DHJW0r zK)PV3zhdZb%$V{!re$5sewygERk7;!y)G*CeKYdq1baKX##8S;hL-X2KD(h`+pnw8 zpX%Xc@2i+}OqEowYmg=|FfXv+eS$>cls5ROmaM?Qr{xtjiu(Zc=DRh<r+p)7Z|~=3 zHOmXKvmwF#v^ljEk8Qcg^*&HnWh7y;_nhY0FJ(FK6ILwl8*xU)TpXP2si~nvJ)5S) z@+uKiV5A~E6ui8k;6=)^oG<96Y_v|yEEYv;IpJ(y=i%?n2s^vZk!d)Z!eQ6PUw?v? zsxX6X`%tWHS^x6mSb9s&Q2=*k|IE`Tyhk`Lf?y4tS<rQOUL2G8IPr4QAGI1^9ToNh zAj!vC?q5_VT+aECP6yJS29SnxzSld^uIXd_R=eXst<%S6rSX!VIJmf|_2DhQT_u41 zxRgEk<%%pS6ia*Aq1SO-z3oc=ik_fGo}uadr5*H~?|qH_g!$$bFZ=Wg2xPR<xN<c< zpBvjzAg2o?VIz>kHjfV?26%jg@{VyXa{Z=rzUb-V!m)R#X%G6^V_(z1ah!R3t`BLW zfs^@9=2kF3U;SC%OD!~PB~}?MX499q<L$g@bDy8z?}>l)Ch4&{H%qhZUHR;}`*ueV zVy%YJUB>+!+e3Qb+<60$y*8{<vtb`g?0ph;k<|S8oHmMW<IC$s89L%MWril5zOF7V z+$(PgxS#}$k>wQ9vW~0cY3+$d*=Pa<>+{#=YpXZvSsBxCtS*32|36mWfX_?rF<z(l zV?_C=Ncx*M_B=uiAvDwt!k06~(KoDLK%!*O8_Q>`yh|;vD1+PMp#)=;)Z@n>>(3Au zZ*P}M8<VSDRA@w$*IwXgRWJH#-MiHC`DWeqoZgm+vlU$D+-(DQCgr;$sF&muDx2Hf zq7-8?5c3J>-ofW|P-R8c=zg`Axmv{aaz2V+)SYpk2&L1vwx4bu>sLKWY<GvsL9p-S z(R(!L9RLSa&>DtEF?}wDtL`CZ-WoaIZR-!&v2Gtcgu#N6Pl_(s<8F*LYD^mScP*CC z6!SX@L&Lp}lE?JlD$ukRmC;=1`arYl*YUo+5;N9la(vmb&Um7iTHwujb3*_GN++V* zx7}Q`%bGE)y<p<X{T3Psm-7_cC(^xx7;bK6&)Z9aDQ{VG5kw@|fa@uosTCl-t0#0h z0=s<@>sl;H$q!JY$voApXHrw1?o}85#vzy2mR1bq&EwxO0SB@2$@nk32zm)|0@Wty zFPfbT_cWkp!nahGI_OD}CXRPNh^H<p2iUywhK$QpY$c3Sn@p(k!Y{!yHbT>)wwy@8 zaaV1d2&FVXp6yKB*M>FjFmXs_sw^L-K+yXUe9N6&nBk#fN#X{+5dc0|UeWX7>3ng3 zc$YDc8&z%GlH~MS9W5sgUW}(f<H|8hTSR|zAxfjEk<~jwei&$N@x5K&MQ7?5s?&Hy zx?|(yoI<#}NJeDymO^8M6+N-fp0uHLyxc(;$ncbm@U|U?+xvd;*u+uqie5F9w0!N- zOj;K*yZolY+g#)TJqB4E<=HUQ)s}DrU&yE96x0xfZ{I@=6-no<PZ=9~qocHhmsfqj zXZx;AzZ*NZbu5LMR!IBN4Mwa17oMK0?k1;tKXdC{b}*y9dwz>NggVf>{N3L`J?M1% za|6}XS>!sC{dyl-YD^&(@wulifwcSFF)|0ZQ7~lO1_?lR=D2WdG}P<2@Dzdmo1Vd$ zZSq|E+?h(v_n5$&E$pWIV^;#b4BJ&zzvdUfFB7I|7OpS$4dU6!H42QadxXmelgy^} zVi!yg<&WSyqV~t)-IdPh7iMTS5X(w&x*ojOfQo`!yD}LgLHh0@*dllm`~(sLzu8)y zBRz2r^!A4$0ya`2z$1n%^9W?DRut}PQHyhz{M^WCskGuGvj-3TGo4I+Qd`5Ehf@G& zsA2g&u2!7rJMv$^rzZ?0<cO$3Aa^ytULC3Bb;YmOD|j5TKP9kzhfdav1aD5v+`oe* z&JAM2fLa834u6D_98&H@0q<t(nWBPv7%jRJ-Kv~N5X9O#>BIg-KXueg=n<aREZd{) zlH%w!0q=hLQ@3(+5#7^Cjiu0IQA;@_+StaBLmiqx7kDKN$m(xAlU;98?$&Kp{$XS` zD0Lp%H^P+=>c(}h<<?%{ma0JPbXk}%0<Wtt(W>pr{88v@VmesCt5efS?TS9*5`8!U zT<gj5RF3>`!!p@KD~#UxKo$Ew!99}{Gd#*0A~3wsSvFdh%56KpdUvpSPYawoEJXLY zwhd-}I5g$AX({s@-^Gg8>z@do5ZRCq%-HO6Y<c&XIO+~(I%duPNK5g<vZsUKJfynj zq32Q+w5dha*){eR&a~Z<$1CqcK%G8>o~2I=(u~Vjx#?%qx(Cfqfm)BAN3jS(USK@l z^r;=5)W7!pDE@S#BBTM`pgIg_x6d4bX@t)e>Q%>9<Dp-f_I3fjLv;(-X)Nv_1^YOG zFpIMY7XDZCx<mH)&(PTNb^EbOeW?N$a8VL_A2gk?R35lbQoM|HVO^o&sxbLI7>ntI zkke|c`V08(rC6-$OFktqA$Mb@(Y`aJ$fdt#v>)ME`pGh%8&vJpvL=duikW3v05j}V z|8b}=#K~{uO8Lg~Bq=5<Nd*O7H1%{MGvQN!IOw9Uyyd(cvNcSWk>mZ;ms$0Wnt^T% zw@LO|Cn98X6Nw&m8z5`}3Bv%LBA<`gPukqu;IDwAJmZi~<>F8iKIfKYHf+_1C_~E) zO}KUY@(OO>s`EP%9CK>y>IYeq>JHcjTn##3W<iukM~PCMA3WXL!O+GL6IHexpl?RT ztCMYL#lgSl6_Sp8TTJaiA$QW_b!ckiaTI?Wqk6OaM?Y{ya)rSM<8+r4vl{mVX3G>M zw^PpgZhjKv^A**nLE2w-N^tN)7-o^;#Lzs`2WZqpE-ZuvfhE?qgCtr^&ZC}xB%AS_ zAokU;s(^w*`yyTvsdai+&vP^kY-?D(e|=j9Q5lKf#-v7D%>OgA((qhh`gCH+XvIj) zLw6Cu>j}?;h|y`YWE!1!SLXCwdYRrTUW}rU#~bIXeJn-0r`vZx;7rU9N4p?$Y#qls z@9P8WhBfI3lT2@ym9XOVM817aU8)*C2)j6+13+@|EebTLT0IFNE5DQ_#D(g@0n$r$ zXeDzR)P%eEv^pDqe11c6Kc&vPu71%Z;F@A)*Pt^D-BHXHvRkXi&AU>Ms!k<imxhKq z@6~e1&x+={a0M67CMU>nIg7fc;XyBK7I)ZN?fe>4LR8#=Si*G&s0I34mWQUgZN#fT zAJ1bcmKGP^){LLziW8t9qm6Rz?VFA#8<BkM;d7FN#))&#daYp?9k!XTFyoiMtUHR~ z5%gp+{{3B*&Bvg$cicowVu-*WT!Isgcc&sXGm?=NTaKoiR+AeRO*xe3c?R7nd$V`m z@mKK%O^ci#H!p+XE!r8O3|+)L5zTaZ11sZk;&!E99;#s?5hq$W-)d*YWiA-_?VLr) zB-7#QEmOd_fg>dsJA(^Jk!QW2tqC=Y648Px`FgdBOmK*OWxcA-`{%avxsyMemmXl_ zLux-$lFZUW`9}u^QpCD_zU$P4O~?!CoX=iFRSMyT)*mSy+_%G^{ZiBT-kYO$zvQ_~ z)hURMe+xJZKd5O}T;bqH2+m$Vt~7N7pCxkin!2FkQCY4qV2tm1m)MjoU1wpIpH%-) z<$N><-L0tLbu}_z@vZG7SRP&c97!B;W0dB+{$wtI&CvS7M?<29({!1i7{9c%r~Gl< z8*fc}FM}^|uq;YY57LWQN9Mgk(ScFc9n>Vx^UiaO_~~FWZkdFBUBV;LCGzKRk!Y65 z01Ww!0Wm<bia+k&TIf=qU@@P?fyzj2F6&swdyI1ly<J18OEu@@Efop|eKuHIFgeCf z?1$rhmL?f8vjc!43}M*iJ&$taT)iZE#8}s2!U&S6Y@~{WR}00OweI@z_>pyxsZg~# zXAaI=;YY%rduO)udG9hG@wUfe@8`xn<=dCj$ss&_tEN16gNw?3GVYu6OasQ0`rFp4 z?)?1GLd{?_g2x4;zMP!yH4nEo&U0Ypc_^(G8dHplCFi4dM797R4;$)>7Ge-kdk?p_ zY@#>>TE|9Dx(%a*U=Jo;bmNKCvu<uI)99kkp1JXDHe1wkoh<I0xhncsBy^e#3zmN! zKfYjZ#J#Y-Q)j+ZXNIO3!X91wBgqr2R8hB}&6sPNW+<W)3wY#ZoNu3`p$fPYmY|uE z4`+nt?z;O?$|!=`Em0wli%SjoyFK4H23j&`)Py>$I_M7_PcBU7yVi%q)l_e}nzAkS zM~w<zjo2M#NV>ZxrhjV}et<2Tn&77UBaeL{dzkAZ_PR^!z}t<?o8)`5VLskqZz<zc zNQb&~Bg5v3hCv14FMI5?Ifn!*`VaiUo<l;ivByMFRov~hf{zI`#V{-{FdGIv-j(8S z73b$)BPX8G-(N(h&Y!a4;4+XZ7)%$ygWiT|aoc8)hV#+N8*X*p;c1aqg4G)={5rCG zPqzC|fk*59RkNEY)WMX;5$q}z+?#rFdC|hW1Am&yg=HFa4pKR@WpaV0lbYT%;2Hs$ zp*~rK2gI_MAkK0;(Qr;nhhfb>sk7?G5<J(ZUEP-H-@*GjmM$qg{p08W;ung51&j2u z=|y2b9o_ry8)2V&()!8fKfbe9$g~+4yr|G!o++B=vY4r)ed$$YZnP%Cc$hD}snFif z;1jFW2r0T&0A7~qt$N+WLfk(&v_!|;y!Q8ebfQ6Ci)U;(F@8e&`L2?aFQX%dut{hM zAj|+QzOJWjx6~>X(D4V4X)+MhEHyo!iQ&DnNT{t6hG6MKlV$OCb1$5riuD^MS6yvj zlE7V+^`sVx4LgA6vuwx1ASOtb!V)j&S9gZbkbw14tP+CrldaG6@0)o!Afbapkl;&_ zl0N8WA7s|9F_*?0K<3^EZdrR8pp0i}TZvb{cZ_VyBS-W-3x<+upZWAd{-+*0yz^Rv zk=6P%^Hcq`4EFr5=PmZ$Fr54JkbziknR@KnwA}$A&Xq>iqxIFBd6!wPZn6y5b2{!V zE;wHoV?gQ8nhMTptef*<!cVCBKx-nwu9*f~?S5)xD3iamUBO=kWroxmR6c^yTV)b^ zIrm2)7tvPT!Tn|0S1(W6wlD78Ib!~U#`N^&g2T>Rp;akM8wt$1E`BobhCK~ge^XyM z5a`QlEk2;Y?Xw>d$-yrwU8)(EcXejDooW5w;Z=9U7QYVSPdDDBbJZV~Jg>~|mwrzN zYqR&wo_e)2DokwSz3|RoJSN^I`S-O0?63`c+UFO)rHhB?Dq?rw)ZI=ijP~~0Pr7@d zkxM1le0xwU3sY6c!mwYfX3uO5f6s<~K*nv~4_U=6D+*Cln*&}1nz1hTM_Ok5HkQrA zU8>_R$I8wleE;R=G}=rbJQX0O{=E>d-Q>s@^)p4b-2;m+_1JsS%~umWP!`jImvK%H z$*Qul5n0snT#NXw8t`V$C$96f;<33d@>q-R%T;>Ixx2rag`fZ--Rjk8yjE%Uu)O1@ z)-;%<$whsru^*6SbQI!U#>##1qI-#!{SwYX@pQK=5s7VGKbL^<CTQe`f{jM4q_vf7 zqwiJYAhQerHkB*5-q!Z9!Ss-=)5?Kk90tDAxhquF7Bp(4%Y8p$77Lr^LitA;!~~6{ z61E;)8>pXE8JuK3%6l+<xhm^aaw<BP=TvI5oD2-+JDJDHEPOVScwrj+x$1zDFmlmM z_-3}k#u`tKB%t{*t$U%<@H;g7G$n?|m<`r-j6xOM?3_<O2U<q{x;KaMHOj##n_DDJ z7Z0C7xOolLGwS2u3ImR%AXFnm|EfVw5NHp4*EXEkFc%AA<(?NX5`T7Vf%q14uYbHQ zbX{ay9>Cc9CW8E$hB8_l%~Cea%Sg;(=7DtQPruC65cC}#eZ7P8Apt0z^bm*0B$+wk zjzvAXE?&u7r0CTh#&&GBUE49AK%=21hXe<YwoQKqKl9cy>CLH`wM12i9<|OXQ<_|l z7tWP*qRubI_V+xW#vb-dn1DXLs(Jx<5Ws{vnq9p{KCC#xAPe$F*w({Y%Tbdf<<pRq z>Kk6(h$?7>-0z0yXwSQTyMCFlU2g#!j3S=v-Syo}Ozdj5+w*Qu->1od$I8i(w|msN z&E4x<AN1}#E<6?evlTt5#?9@rj%v$gR}`Kbbs_jK*5!pN_D=FwTq5^BPdH+aL<KU0 zkZs0VS6<We9>Q7PUdREF<8HoOD~LYLuVh2p0ZI;v*IhSc)F+d3tfJS6vC<F<5snfX zE|i2$7U@C??6&*78EtvDHl4###5cz)8illN6gi(00K&B4nA*iPTJn?#?`c6Mm858h z+_uo(B+%Bw{+$q57Fq<D^}bEEoh>ei`jl!smLYI*(Jz`h+G0U{_?Z3t$)yF_($)vj zh#;H4A8n96?TphK-5>tAi!E}C-rJhfyY=8Ab=~22OPpcAg$~>^pFIyhS@mQp9K=th zo3cqSj)fCW6gd_?u;@h?T2=SVH)_9G@n#|Zu-og$zK6_gbp0)&e(O`A*EfeqvztV5 zB&sx0l#|t(9YS5HFS4DYSHAQtgd_pW3+H~Qz;X>wbvzw~^yy%k5pIZp<FFpw*?WLw zhAoGSw(LT2VNa*=2w$jZ)kcb?IOrw;nokj#z7u*%+Y4uLGO7OfJNk|zlR|5|ABdeE zbt?{05Ze9ta{n$`@F<0C7k~US)nnwy6MIw(N;5uG>dEYy6hUi}xwB5C=q*sT4Q&O+ zXPq>lciI`h{0W@VLQK~h*&Yz;HL$J=XtLX03bFO<(zM9qJ6YT~awIjns8d9kJQ)yL ztU34AMRDG~vHFb!?QMTFj29CrbO6M7y)z6~MaDwKYh3USSYp#;c$m&iDSpJSu67s` zNIXsfj%ra%<tiHP@NIr?{9}a0>tqX+ejz(2Cq50&Y&Kuk`g+4f_+9F0A6;eBB<0#L z(;dQGZ;lrHijBg$@kwYWrQ`j-_w$g$_^u6@Op~@&v39Q^O#(BefQ1l*-D>*&SYTMg z7Q3$8O5TF@MRLvW+PS`GHVtIL+t$gYc4#70x@bz%jjzs)Bm7k&wA-n#x=B}gurObB z`OV$O^YWrZJ4+a0Jn?laT2n?2E?II@?!Zz9aq-D<pT}+tskr0sSP=*aa+HA&^X0=I zAB=e!Uy{%9xyx_26>5cr(E|w+y)5bJS@(8HQnW8J@z7m9EJYnafbJn_&PTu2a2Sm^ zIr|;>#B)L^m}lw2{)7x+7Negxh_+TtBYbmR+Ywu2cggp3Brno>Y$h$p`y4;pJDa8) zaLW6=;!&XF5(7&WH;x&>qLFW<F}qG6NFWuTReLNHX^wM0Rk!=&li<NtIDztlW9EYX z{fu9^4<JkXK9pG{6ay8vC7Z2^H1f9Tu%Iy%%CtCfVBbR3QPj%S81-hQ2-cfdA~y11 zJpH$dr3D3hZFf1Us`asyq9vSKii)=l63c7m>-sNjPo?*pYY4flS2NQeWfB>jo%<cc zam9^9AH}I7H-Di?0<z{;OkqL$gy%8s>AwqF$G=M$%RuX0H!b?+CBZ(uctx69XRa7n zm?}RTcI9u#G_u4kw6z7kw9FWs8o5Nu#w;wzP#GmaI{e?aBthkJOQvhtnoXBC0$??S z9oqGCLl}7ychQBm&dt*JSlqbgBYobYf8&w*LQ+zUX?>ee=EqD;{o;H~vAc_S^}S5H zWPc3DSftSj8c(JucRpIqFZD;U<KXnkse7+2MGLAO9Y<fJrA0`?Kgv@^{XLHJ`{aAj zaNCby=P`W73~(mjrF@($?<nmYG(tOo+pH^!#4RY+PS4u+8`65=7c=;C?ay0uqU-CM zD-7E?9djnRvUZtd`grtr@htM1QQq%DU#I}JZ&|(0T0#fpeHS;>FEKXasI)|Zbr=^? zaA-<V_~rFrG71rCK+V8mWHTML(8L(vG8Da8aZPn4-qCEn4uN*;+sl=ta*wl~RT7hH zDONqu&O6Tuck#mFAlJfLH$yq*HLMUnt=-Uf$wF!@t)3bPmgm}=2OzI-cLw(z5V)L; za4E4)g7VWmtYNoV5%Qi{>wG|nx|6$equMyaoXN8xQrDAlj~}|!(F^@#Z$_WZ3azA* zVf@!si2ANR*XgPv!S;X{X^y_RBdb*LTroE1*()grzO)~-WVhefK|+=GMm72<XX}-$ z`zwyAcXtY+4nh3YrnYANQMV_71aBxa+*6JpJ8wIK_W{u64lZ3s>lF?7{vt{8WHT&> zKkvW@EImy7CypHLZ2C0e-u?uxsGx0>-CM0+l3DoUMtPEZ4nVtnc-O)nkQ{x#bn_|J zLz`~v(@ZLF*fo&7;xikX^#%6Ip1W4}nG^+igp<ner9t4>3J1MvmhFVL%k=SPY4ogI z1%RYtT@S7$helTE`%v8gewQ76AI&Q*;R1jE6m&r=!K6ebb=oA^C?qT`!Lj3gpihtP zYw>Mpf3OQF&}*+$bF#R2quzSd5g<Uk<aDa=)5!8Ab>fR3q0>^hs{k~pdlW&_)u2Xl z`fe~+Jlu!AA$3U9fkDIPhupIfQU`X7(CThYo@L+$BeaO+pbMO(2=|Pb6s+83s`2EQ zH|lK0DLYkH4}ezhpyq{=BZYLz0^XbY;_V<2J<rot*YUWZBf&$n6$R_7f=hW@CM_-= z&|)yMVK$?pVNAsOJb}$i*)8ZHyV^RArh5L40@8;XIIlO{)18r7@a#V{uV?AGqTgvo z{4z3)h(ZAW&S-B$LrxCbUm?OPylQ$i6BI<1zRD9j<izLu*A=4KOZ89C)!o16f!{!R zNQIsk-Qzg)#SCaw<7sn@izi*o^nn2+nFQy?xVL!qZD<cZGW3*JnX;WkQNJX=+oD!* zrgH8CzuNPncd{en|6%DW7@~Z>?n+84-5@C~T~g9Wh;(;%cZmo{ry?EF-Ho)QEC@>@ zNOw2vJO2Lf2Y_dGX70?r=bm$>z_3@h%K0d&vVU*%W9KR)hw)Fe!*89I0mcYAhMN|G z*=EkY`YF0W97ON5&!4$0N1xiyUQ1@bR(#ofyW70kqnKK@lDW{{E$)Br#2M#(C^*<> zDUBL!p_skgPA(oNmFQ&~ZCxuz#$(&pOO@yzcfm#(RtCvrA1M(DeA_WLDV<MdiPc)^ zylAAWtkt9OJ&HqKVEl{uD5?Lh|J)>_iEWMF34M2K%90<}7bm_)w@O<4Zs#S)#2J1< zV0uDr5_zXC3J&GX<lD*B&}&!elAG=P;V(wd*-|^3?k7C}C2d?^<4X*M?CTSbWa~Pt zpi2@!&qdqZxs8RlRX9H0SGIh5lGjF4jv@%i<iV8P<`YyMEx9U{Ga6ns0_(pylq#R2 zS(PTtPdc$<`SCFdC=bRb>QF_ZXpQLT)(39Ewxk}~S~n1EGak=M4*njMHv<L6R_(Jc z=}ogleLpC#$8a@cmT~Rkqek-lyau<5E!!>C&00`-Z#ZKH*?f)J6fcZ&*S;P@eb?Pe zF=EZ~!}up=@rj*l0CZs=z*0j0x~kmB#@_2BGPP~ImZ)w0JbtRdrk;PYbkDD`>2f9b z1?_!Q11kXR$VN6@*(UqHa1dJpanD^_jOC&{WFnubFV(t_7|Se!zQXD5XQ%7;Tu!qA zciSuW(v<*yldkN<DZx`Qjsj(n{CUbekQkLdH*x+K`6c^<Maar&v(pmXFq*8>xc!zX zy-VmE{#e^x#zB?NWNq1_cJ*u0#VzKHsna=TQvI3ozoa9ZxUjZ;sY{r`Kk}<4{a+b` z;N{>NHX)V``wwzMWSyYX<Yo=4j3(ipSHphB&bIWJzLHo-AMt493l(y=-j#wYog#5@ z%}2!kYX=vUgSZ7J=mbF*(}xV5`@fw@R8hBH10Ob&1jWs6P%kr>YQIyPV1pc}vFmmr zRZzUIErqi?)y;II0TaNi9er)^zm+Y%EKMYbKHVD_xlL4%ZH-a^y%OllkbQX!)84kX zA|t87jLOku6s(UST}~^pE064&6mUsuEAwdo6S36pP&Kw&Cob(AU1f`H{HGu>Kqrl_ z`p2U9de2t|nserxLEKSQ$%=E9Th+IdTIq#uBAZM>D_PuZF4)xHKyHBjn_a6X#ib6R z;cX1S)}$()DvrNsB}JdEQrIPN8DF`bzdX9iBwaoMB3^uIP#1L4Y%m_YASx>nJic7; z6)+_eQh&u*EIn6k+?|>pwCuXbE=X&({+2@rC8fafe8s$7ZV$@z%$<@4;nVnPV+s7) z!t4|6PJ;y~3+d}b6UThQ3-|R-&Ok~fhIE_w-TQW7rhEuZ3q}yMlSqYX_oeBYIReqD z-EJfemtSyyPP_H`H1z-#C_4ry1E*?y)6u-1L(YA-<nG@`GF~YqytEw4I(Y3^TazRH ziZuW3N|u=!l!yH6hKsH`g2I057#v?w35#6MSNaBbVqe;c|HDG5rGF+q{#1(TM>c>S zd;J-`TI)I2HvTo^qZVN|_Q7WgbP*cAAr^N}I#pSxAwI1QIA+S`r=lM9GM1RDC3ELZ zZl5Z#!GFBYE6XEHd3-``5KVZ^#?@3PPR?Rw73X_Huhtb<LiNDUwB@e#k;(cQ=3o{> zKKe8AwiSa-3bj1<?pVym$1t;_sIEvXM*}Q!&9~uYF?Y*NmP%aOcMLy$kweRV3Fdxk z01V-?9sM}KAa=eIQHd{?Qt!Ik2pDl%^TqvTcV>7qzQwiMf$R7JB1biy!8UDEFBj_= zV}^C(lW*u>gDMS|S~veKn`2STiVjk)1p^*#@5{*5%DFG#zsG)<0tXz8M(5{VS<BT@ ztCjH7XcrkhxTb3`kR6bjh2w0LF#!%Dmshr1S>ye|k56KZ7y6?5rh(^dR*KyMMv!>V zH+_U9$t>@(9et5N=T4^MJh8-S1zQSU$)tA5-u3z_FA6>I{%<+AvKOUSJ0K8N+o-|F zeL}<$poHrXXM51NjU9>?tqFIu#c$n}NESTa%|pZC&t3!2f$PHq;C%I6HBsYyz5OQQ zWZBcAAGLpTs2em5bD5%7izO|QU)FIR&wl(V>LfUDgAnY_ueYu>+#Yo&%1uZR<G<=- zrV541zyksw&)TCn66^4y1@psu{{uN!;6NtQ_6MIbvWEBPlRv~cND!fMi>b$3wYa&T zNm)hcuB)?mJy~OU=AKt=#}n=<H{pI=ZG{ztt@>BVLDrrz_J(y`bC&d4P@U&1m(9<E zJb1shyQ!-C9%I4s8VDD~Adoh}h3`}A=CQaFT(?vglWzxlK74cw)oZkW9P(hxhws;o zlb-BzAEt3j*{yl;E2cW+%WN1rOE_=E0xiLiXw(T!#XF{8lx~nMv=Y+>uBg107Stu$ zFM~&Nn4fozE?u>M5^=J?ZjPSq=K31|8r-z9yV>D;Spihqi}`vZPco$407~p4D4uxT zyPrk02<Wvga0rB_WOLsyh9SaSIgrP07v}UkiUX{iZqy^c!5!Y%QC#}+)t4?}=fQ)D zu*Wep1~njG`2<+lN9;<wkv{<oTel$LfBCOgtF5Q=?QA9SNt5wt2MJ;&HH~Ux+g;W# zkFZFkCp;_ghogm-=f65}M#oJUWIsq4660lpA$1z4uEZ8($tmSK-%XKNOm82f0<$7f z(-j*Y1!^s>d>s%twBYQG|FWQYP8C5l;yNLTvuWfaeRtJE$7FVUKJRnJlg&{=G}iqx z3d*dG^Z{D}QEEHFB{Yv{4AZkawpJl9QDnW4>QHNl`RAo%oDOLFwKer!v3HEu9d-2m zvg?i)xweuQNtjgF;a9O#=gsM*0ys;1-(h4e?-2RhPmwf91^J&JQ5St=FLkUaJghBh z@K;>?i1h1iXfjz{7hg|20c@50)SU^+=3Hp}J<(jH=+;Hr`F1Y47~9Tg{&#KXF9VG) zgKTTRe9h=xD+I03w?syx7uJc&=1c51KoXN15s{7x$>Ak6d#f@0Y&LkkJwu(04B$gn zV_9rp6%QgG0e_=&yGYEzFMW0CyCVJ?pPba@<HH^i>$}~(qyXqZM7<_04l@F1(90cC z&>r7aI*pD@v>6+LrkxH?INGNf!fIf#aR^tB;w4wYK8}RW!}dvitQCHD*<JNu#M#oy zS^$S30KcnaeRlv#nx0wGOq&cDhOMy3_gY&9EzAI&MheY6*cRJrcZ)h|Eo(UF1I;Fp z6Z4&ppn3SB8v)0CkqMM$$g~Div^}+?l0Wdyp_b#TIFE9H6VK=Wdh+qF#YBXi?~kov z`z(aBc*TzcI<MMn%qx+)@ITm(WKL>d8dVO)(x4W9^?HdKfqZpf=xul^^x(f|H&RC> zERZz<fgCi(0O$s<ui&*1D2PL_$ZC~^V}@Mqpk3B=z(T7<QvOXw>&;oyqI(@ca3oon z8yYS!2>Wd9xGcx*vqsf_WiB9p85u2>GF~Z_B7Q*(E=#Tq0j|>|!H~+ueB8>@L-=o1 z=UQ`TMtyNEOC9b3zEp}O$r(LP6~2w(&Hbe)cJ()RNiOT8MU2aC{>@044~P)*2CnhT zZ60}QPBc!3`aGIypQw^*>oY7d2NKXMUR4QP%i`l*ZEZX1@LCaj9Lh&C{D|CSUlT)W z^SCn6<NMP6PYi$SrmF?Fsg#}oF2tGx6A9%YCA&oQvJ53@5$hmE%FObl9>z-YS5NPU z+M~%_ZT54DrJA8l1bY3nXvEl;`ypBxdtSM1DnPMG_Vbzo#qr>@+H(zNMB<9qPRZ7B z^jO-o;%Fufn(1aMVzUcUpt$nbGBPIPofL7<z3IwUxa4#jr&9lj1=5Fk5*;9^o$saI z3u?*boXYhYS#BY8WWB2z;uJ6u%aSs8#pa!h>bMgVE*;w~6(f<|!$S8G*;{V>6}*4~ z?9zB}^(v;;*-2>r#4loSjEV|N{1UT(Uk`?889DeFeD4$m`-eU}G~&ZT;|_Eiw4l6{ z((-Qo!BQsr&PLaqAh1m{eN-p{hn&#J5zw`g=eTb@cLsFDT=UU-^JqJIRANh)xJ%7Z z+DgyACY09Cg5dRSdD#CJ!W}`Z#)344;3rhW#9{RUrlX*+hGRB1Z{W4Elr?>~=&q(# zB>tg0-VZ9Rm>T(0#KE@N-pP7vynwPL1o||D(9HdnHR`z{f-Ilkd+Xt!*39Jc{ojPK z?VGzAN=3oth%)Z2W&gcFoM=|#1^h5FS{Fy?4i&P^gVpy)Y~j~diZrjT{@c}bgQcq8 z%3yCIH*w1675Oh=`}WnuyCW~ftc?B8BgP3VE*TVnLObuPB)#4!+PYTT^0sXhQEj|0 zoVncx*F7cA9_`%w-xsKj7FGUt-0VTbD@>4JWkU0P&>VDy2nq{}K6%L+osL^pS_(fM zxwL0HsvXgdX^I{&iF=shGMB5Zn&wF;ME?&}B}Yt>=(=N$%P@7S7r9#oMWJkKqIa@K zzS(h3k{EBBjQhq5U~zGH6US2c2Ll-rk8fgn)jxZvJobdH$?trxioh_);4D%<Y9*Y~ z&IxO_e9EeclV8)QDqPaQxuP=V0`twk<S8dV&?}8&aEYM5^0Hkg>@QqsdtF7=pxNqN z`VLTP0zs~{71OT7w?(fn`tIxiFd&##OFy~+TF{U@)vRlwht7G>)tRFdxQ>R#pB%=a zEi@=E`~jR9K0-o(J29O72nlD98NTXgWB8^dyPWzG+wT+&>geRq88kK@qh&}ir?res zG~6>=fif=O3bB6pBu#5<J?5hpYL!W>d}#=DV}`FEVn#!>^tc9wQV|$_Mn~q^Cyn_4 z{$L#{{P#X7q3vQZK$Tfa6Lfz)s*w9I$`6dDrd#Aem#!AAjh*`1Dlqcka+pvti1?(9 z6u6u8bIQ#Ew+R6U>5!fX?|Bm4;5KyH4Zgs$q^`dHRp6xh-H_=?g$~)$%1jQYB>W~W zTBlAoeF1c_^qe|%a6Yx}zp>zYAL-Eh^B<-kI}DPZ6~jZk#G7s74>d<F@L`HaGKt!T zn)o43I=ep_iJZdfT|$++U6t2>u~uhEWmS#Opq(OL*(3O0XF_?BX}>i7@t8a>2ohYG z&wZIGBPo>{VBhsF^zrlu4wk2zYsAkcTY~3dR|hQ3yQR(9cnj6J6)_NM!A!}=M{6=D zJOpOA;IU;;;N>AUFDb2C=8EHs^g-%bvp9lhlqW33kg3EA6%SQ0U})*8=kh!HCm_7T zO6!<N4KQ6~RXRB|bPuUAH7v86zoqBA9c%-`U@v-gX&l{1zL2Bk5)%;RcwOSL{fy=_ zMX}KF?VbyU9w>_My6*}*1J(+Kfll~t1>R~kuGo&F#PV>4==XSQ+@wz*2@F%X@@SsX zMaR9ueN}%xXh#Yhy68QK9+bUjrneUJ4P)$3`9!qqE>2S2&<7vr|ItM8L^h~W+3sw( zVuPGg-0!##-y!{5xU2X``1lW5uWe5I-6^a4`)ht#AS0bI@>ZSd8XOIP0m|nhLMdzy zF!Smi{LX(o*{e30$(76m4byq{b7$`@1Fz|zM`^Nflb=J8)4%C1m;)-HZq|7yk*AWu zy`$u0%AV)hKu_@z8pNtS&`=EJjL>DPHGh6|`H`D>W|;)zkcIe6))Y0{J!IEuICUkt zU>ja(_59z-#TYi*t9ly~5=Rv0YtU_KBM3^(;_n4?^>#$h9Ul5Q!bMK1n#sjoh3b^W zGL^+1A{PFWnz}}T{jXW~Wq-dLM})4;;6J)v;&r<4+kzwBP8nHgY5~l7$PbeYzJp~u zGOQs*ObfSV&_Q^b*K#9ev~e%;wz6{m{M}L|(VkaF`Nu1ufR_s8c3dhQ!{JTtoS9FR zyeP{Ld2iKS5-q>_$>aNq{=mg}P-!Ym&P(*KOb`gyrC9Dp{ro2g8&n&#Iwe&c|M4U) z?O>o14B=n4zHK6Ws@`>0nUP{cCU|Nc$NXNi@#dE7gI-f(W3b(D^7CIJb|mUFR+gl- zb)-b>3cDJlEED{s(i8u@1I6teyLuKlRNn5gpUNx~!X&JK7=AeQK{1M#=b^+d_g(Ki z$apQWo8QEjhZCM1!s|`PC{W?XriaHyplIwD^XymvVtS=miSH|oLhpWX%D}x|@12HA zEduL=k|e})g5u+sZjixiG_VKLRnQz42>JhL{ate+C1l6PlyU0{R7_o{n*-WoK;e9$ ztCQx<XZBBI8RE4>uZ#28I`;_&=V)$9eLtQKMFtUBxBSJEU`sC=FySw_2g3s0<R?)d zC;~5a=k%SoRoG^#H<ILANh9>^fBgPJfh>k!C*!{W*c_5tl2UzAYT$;q7o75Bd(E%i z);?0HqIQ0SL34`=kz+&H;wQUK&i8q3HI~MGoe-X`qT12PYn4&ys~6xy*RljOhs^<t z%`=47`t{a(zk;Ci`k>DF6HZkdM}o(1fPpk)<Z4q;&RX(in;>jV5*rfU$<>TB(PuLi z=&3xJACzP!ceb4!@p^-*Df-dv)6(SYTNnX+`KZ$A^P_ZVnfc~oA=k%@Zhp_X3&D3B z|BXpW$V_Em_FMjrP2FMLuQ$S;<b*526o}xO*9_{qJtuj%wR4byYdE#f$((2V?3S1< z`?OgA_$F_R?>e{O3g10!^o+z82}n1coAo$Ob(G{+qJ~ITN<LIduNBUoS#HMNKi~wp zRoM6c%74t1qDylHlw<RujNd$is?Oh_hLgPi*-8BiodVYJSuxZ7;#-s@z-K-&x!s@p zSKp<+1$aHh#k^f%V8ta%$CXoDs_3zJkB(yMXSg#gKIr~oq8ia9@9yfT6aH(cQxq3k z!S&^-Ammvt7KA5_SJc1zyplL~B9Rvga@3n-xxWl6A;{s6JlePTDnWHO6UG|pR5rHP zcGVXgX*VmqG=U6K_rg6`dx^fI(=k0q+2hvdwEtGmMyO_kIfjraKR`!(Uz~u`Rr%X* z|2(%zI$Wfc;gs;g^^BlLaL*-V#P$itNO>;%SX~GoF=W`Wvr|6qbCWP;O0k}BPLE}G zT@0k>5fB(S2V_JmN|8;_KFc9{M@>_t)V6AG1lx})TWw~l6ty}s@?&&~iTFx=m3^BC zntwbnEx|6|k(KmEw)8UqFr0W*?<;to2MPbXCManJW{phGt!RHa871pIu6F^7fKK?; z{@hj}S=}$8I+qhu(@~Fmw9SzIii%3#fCa~n2A{?W{c)cy?B>7NP)obN5Rg~b2OS5N z7Y@|9>nvGihh6Q<?_Q+QLTxjhq_N{K-yuNU2jkiwTpIw!U_b0~)zy51FHzfv@csD_ zA<eGvPWELxm5f^4bz@4Pn;<NuP%LF4u+Hw|5NGktZtlPZI4Kj|>p^$Dv50Fo0FW{v zzL;UN3Nf*GLzA5^n*cFDpE1LD^p}&Us;kgJzU+c*Tq^fsMFpT_1f5;LQ28Bf1>aVG zMP0rt5*6#xc||<GESy9)<jB9dA;_pS+;#6dm$|Cb8SO6O1G7pNH?mRejY(XdK%rjB zeEs5epmX_%CsDD%c*FIZ;2xKt#tp?W)|P=K_@pGBM}L#m5U&ze5DprkFzIPU@iOeZ z3pMfFa%2CO$A~D?*4*WfE*$P2>N&O!bp<}C9`tm19(<xQB3bTe+hE<Bcwx5eE4<a7 zPj!%0)6T+9b+Kz#(LlJ3PZC`j@4iT9gG~>yVe7423h;0ZfA~Det9Pg#dQt7E{=s^D z+bdVZiNOpmj9+<Va5?qvVfVVoc1Fr?d2~|^yS&RN#{E%Xmow~P)yGRGB|05AI548) z>J^<Z(U8omV*fC!x#b{gjujh^Pnyb#0o19(#+7>+93<WG;kNtJ7eRgn2M0(^4po77 z2fJQdx3H?I4=+Z3gp6&Gy1MBzW@+=o-6&%icGfWQtvh6}h)G#vOvr=lJ>ecPW(uZ& z3+&YubxV0=zA&Qh;Jf{98or65=(m&*G@L6oZN0RKBkO+q+&1IT58B{a^xW$yL_Cf! z{=^S+aiUXPtNLsB>(dB8?&nK|=0hj%;QD|<!)6tUEj!e4WjQWd5i<=S(Gkj^c9ZT8 zn|J^u!MeLj0*%(Q$mNX@ej~4d)5_5&%1zQ~pUJ5Wa*?P5z(V>T{(c8L;$K@@8x_;w z+iGYG2ISVyyiH9Rj$_}F#cpxNihdsj&KY$SPsJS{ISB(pz*TJE+p<F%nc^xWVXqZq zRaxwj9U!CMghjPnd)P^dt_9Rr!cJYiZKl#A9tVF1WBebH?`An2?7IL(u<_ya%j)E0 z^A2DsuIx^?5E+1=FDZVhX+W{q$}6umFdiY`uqyZOjuYDXG%Pmuk#wfz&S>;wAAvm* zH0L7=(AbAd!{-J8()F488;+7{z6XGtHs9f`j-O7x-n&oHTzmn6bV%}A-FXNp+aI+# zg3tU@9?0g~=F6Vl(fW<NUh&tv87^yspBH^XHY({v%G@L{^ji8OOG$W8ae0Xc0%R|S zf4coD&<>-|)>bj#bGhUvo^0zwSk>+rf0CEQr^x#Ev{dL%CUHw4-%A`z?%lLeCgHWB zNmO>+^4etO29&niEa_*UE+r4ELxxWJhdkH?8Ug*?xJaoF%cx4gpC1QLAQK~sL^n#V zTX7-uR+iYedq(<dh4nRx4_}N9TesrA9X(xg`2TzDFGFS5`XnIz54ExSxwmo;>yqU| zd^w{L<71UGi4T{-9R&mh^m^!LGRrgb2mqT*jn<Ua$MNIy@ICKxPdX{V*>4tsKzy^E zcv4mF+s6;*BwPgmTcfc?d<PLny#rd=bgig3v6JC;x?aElK|~UgTsz~TgV>-h1s7&w zwt4o>J*o>N0`aD%)1w}$rg~2bu%KCLL9HuM4_pU+m2~*Pe%x2j!cM5|YUDUe<onf9 zwzFnIyZ;0ImqzYe#=PhxWm`e0`n&l35!W;+l0Fx|0<vUCL8g(>8aGMeod&<)Crhl` zNM5_dA;F9YAa4EL#JxX<dvb8BIjTiF9iCb~e;-jPeJXPw8+qX1UM9T<fTF^WmtJd6 zv}Pg!JLkKJxfHN>i>?Wh*$JdQf|YQH*@@ls`P`ofZe_##O_%O(+Kz7W2{9YpXo0RH z-e<l9E%5Y>{83r&k8LZIM;LM7ld1iMU_OY=Mtx+ik^a{{82p1h2l>?&pma1eLSQsl z7DqGoeq4mtCiQNi4&u=3E*>hxW?s<Wy3@R;A!Oa_GsFHkTcKk9n}>!31+x91-v_r{ zuN=qr8Voqk_4y+XXd2Ltxch@*(xXj7f9MatBP9Vl7|5Odk4=d5QyI<brDQ9f(LpB2 z=Gf-lnx^%lhkPHLJY#l)*3MA$;FqlX71!myXDdBfRHb}EQRB=YX%200>e)zfwg4h> zyGD_@mcjC3_pQ3agAJ5LN(jUz59jX{iO-hjr9Bf9$;0@9g2O?h(F1d{gP$-AmvuAb zYB4F`at>|FR&-^UgxjZ_Wqv<Q;DLD<p%^s^^j@>me_dsY3T70<kR607Saxub)Gc(O z!)Q?X9veIdr^cM*>}Kcy2We(B((=lBE}7!w;He+q9W#;luvh_|`uqDmyK#$j=mw}U zd`bGx7h*COu_t*wJ_jANUTWRdU{bzbcM9UhVRdu}5y264|1+)Cw*QTFdbSJypLa!$ z*Up!s*MV<E-s)-@N&dPXx8~US@fpZ6FIih^^(u$<jC5BX=Rf%YD=Dl@rc<hsrI~Bq zt^2d6qjt$^ZC$+_$K`g4uDdRF(b6s{8(LZEociE6L4!rtTrnSL5D+#A6F%T2<o~H( zBRP78*LMAZ)zIqAIS!!kiB8X&Hu!lOoNkGIxxabj%YzR%b?)4PKy&1Yf9{0>2$Z~* zu^3p)?-QE>yjIR|l|irps|=RK=8!!opJvn6`1ZnPj&Y`=R@nJ<(0>!4>!)d=cfr1t z4+zGySC3>ci+67Nsw41+-!tishp8Be-P<&11|JHM&;2LwNX#EvxkKfb!sbNy92E{O zi+0}{f$kZFA`iXZ<F=T2ivEHvI`qi&=<!!OLAFOX@=5qywvi_jbFC31xW174QZAbE z)#*KZ$!(U$W(HVh+Id|ze@$Dvrj<*erAt}zRvM4L)?(Jk<#l!nCoaBfTT9NF`qrmK z4qAaoJK?U1mC~1NT`ho~5uO)BGdV5Z|82pGU4V9fZg$=ieQ;@oAkI%Y$Y*-)50Q%* z@UZ<lYh=dyb-ii%e|hAea7tU++bedX&cmLlR(_y2N!Zp!L<hz<;y7QnG=HTfE}BT0 zyE&eT?OnCcH6L}MeZk}{pKJMp2|A{cO=e8Pr0xc_96#S?&+#2e5XeFx0>!Cy;iigj z=P#84ZleSL#<Y)c6lv<NxoEJ#uhi13+6+soRKhU0baPG+Tmg(Uk@Rr?rYm{<%Tnw7 z>+U}uZI4p}J^EV1sY~w~b|lu&C>zInxRZQ~NBBs@v`zaBo<JaDK(kx2jv1EDxmN&n z>aw5^ZErwMZoOsWUsiuyPh#?rs7z`3nScm-%~>{U@jt(Y$dii~I4>gR?PK~iJ;r0c z3@LURe6Cdvgzobczv{sqni1cg&sLQ{888^q2Atz{Z<Vc#iZe41n~PMaZmWXim96YB zt~=p@5B)@bU)bI@lqK|UE7qgP0QN@0%j4`UJh|n|D2CS=yxX@g>WD~LkyYV`g^`N( zM>)Eve4@ly$pcE}c|+^O-=g!eWM&^Ygw4<$y@6BqK0(;9P(#1|-=-B5J_m)@DpSlx z{w`C~FmR#0KeN@RPKt{E?{{pr+jO@_n}c@PRwigF+0rDkg@MlF^McR|+pp9ksXdqW zU`Hd8QCbTca9O(FLS4tH{Fij5gGH+#kQ@zCvp8nf#Yhb!ft8asxLQYN^H3?Q?O;al z>hVBTVbzvTCXdDZzgU)aS@``{MOkyfNTL>S%9}RT%-xoAg8=$fk+~4*-9+!J>-p%7 z?8IJOu@d&F3j%RQYRC|oE3xQLI6Ykox;Hyx3W*`(*Z0-wg<+_{Z58oKKE6QxM{^nE zb^B0QKz|4v(&E{qE5^pQy=sO8+eH>s-=whh1c|IIOt;9k*`}{nR6&o4AdvYDzs#|t z-_Jknb=}wAF0}ydH5E`_V2Rzj6tuS&>BsDuD_dhlgX8a`g!HidUmE%OW@~3*49lhl zB)q7i4%;DOuRbN6I>tez$XhR)ngQk4^T0Hm)=3Br@f!$*wvNK;aR<!myJf$Cm7uP_ zaL0wSFJ1A47ddaCMmb4R|4j$#pRVQ~U)n3|Y6cFnf1#^$IGT~nz2|Y>3^EA#<>aT< z(i=p^->=8ng@?nNSAA#bDs9?4^aS$9I&g>BPkm<{+RbO+;jz^2BXVM^&ya*x5#mPy z-5KxK*jBPsLC@wBO#E%IUD{Q%Z@{Z-H6oN18M!yZ-TjAIXlkFFwA84K=>!c8PYJQG z)Babcr}kQ&G=N<S%DTse#$}h?9p~(Ruz`S(Wsxe6o>Gq{RWUa=r=HKrXQcp*edan< z_+MVi&_#kfbpG<QefZCPa>wdR2aD&Hu9!ol-YL`zk#>D_Y@3{dpUhM43#{*r%F~k8 zQi16TFzYh0qF4s<b5WhorJxv?keJR$1D2IcT&gwVEb6EEq7EEAw}f#$8r!^%0t;TC z1<Sr{1H?+;=3mLp5*KE8ceZrcPsB)0*^J0{lyKxfGj>u848+Z-oddHIaHO!sfxq0{ zyYG$sOAF0N9Wo}3LR%lrrQx+L;yjiN85(U5B@{;E*waBPHEeDMB}*rYJ55^dpnvwX zrmOQ1m|;L5gK}2;M>Wchm=ls%c-E{_E)>34>Q8n>4qJzTj(X-xo%7-Y!`6iOSmH;w z25rgSEEt&Ea{1^6ul9=KATPiV8`gT52W-x@%$c2eNhg6Ta`EXP@T>=t2@lG3b|(Vc z#Ib9|mmjmm9euz(yZ4R>Y_K6t9$=~RZLMtX^lN2<P^b1qYM0+t(9BhHP})-d!m@E+ zU5ovV5@?|+i~X*Gebsc}UJZmmg2!tOt+I>n%j@I5-lFqp{CPV;+N>pxChkQ2D)4IB zFkMh5EIA+!-UQ_ne>T>-i!oDSYbMXDtJSrIHGl+Mjt6uDOlZ<v6)!_V<}F$$(?nAF zu0Mk9#^fyqHbACros60Eu-sktm&fQ%^+@jf@yq*(=U;M&1Z@22IB0bcqh`;PYq!R& zZ@p%6$Ra38$dm1WoOSVLT1OQ=&X-!dp-d&@m)o||pXps3<TvvC^I9qj0PQ~wK$KjO zgpfUCeUXzP=lA5XCvBCd+Db#dtXM$84GMJcwx6E7{?v=js}Xi33A?wGFl4Cp7`ZH~ z%e!%@{5tE{eh~)kHbpJOLGKJI4&GlzA%S4T)J#2nrS5@m=EJ+_X##TNKfMXgq7+%f zJ$Els#<UVnn!e-*I4Hy5#HLzp_~P$zj~2=beRz=p21f*n!h-gws`GsOTvV)pz4JmT zY2O#Ix>ZXs$?G~pVHUcrMo45(a<m5;fMy*&+!J@LuN$8>W&>(VGn34nxK_5?d{>z9 z(&vNhZ1tu+F^BV0CQ_gr3~GfZ6Va<Fqhi0`I-l5!{rV}L1OiFi6XPV5yNhy;RpbM$ z94kqJoGgb+m(?`0eED`UOl7~r^X5pDw-R7C#hS9{Ur=2$Yd7^|x#hhS*3%naCr+M& zPg+;A?k>2fqu$KGW}IfyUAV#FQ`=8mHZ$2l6E8)_)%axT$%VPM&3sZ<NM4{1mk}IK zbt!ghTFzrp>g}eK-gtffl6sEiP~#2v!Z<l-rU4bowpWka!OH$J^G;@!>ID@9()v?@ zomJ`Xvv|?l*P-v80!Y5KcYo-4lSoVN>5eYYVREw-ZDBr0%QhCGfPLS3zQCf(KoHKc z>9mES5ku+Qo4ZO6yujUaiOfdRm!Va1?wq>7Z#1})G4ht45xLD+7I=<1kW?+4S(3Zd zSIu1G_%}E_o(}}Tl?1fVv^2`^%>1aj&+G{rxExVH5Ggzv$cp`bC;gRORa6pt!c0w| z{>gUkxDe-S3-x-I%RY~_R*|EE-XbpOpYua_bIgVL+k`pZ#ngaypdKyT-j9Crh3udo z@B*d!8X~;lMbCd(>K)=s*cyg}AL#{}fqsU_V1q;U`MMuAP5UD%=ox5Amj|n6^905X zM{p6dNxtyMdAR=H5#@rJ_n0lbN1j?a(hj;gntVikN{gaFLY%6N>tw89wM<%E8$VP1 z=f1L~ftF6-NQ^!Fz97PD#00IFvkvfhGo-Ef9_E0g5HRv>ew~2Vmo~esW_+X<F~P~m zW}g8c>&x7sfUFQ$d<R-jkaXg1DPcXXkFp;`u-Z*vh5dc>XVtTwy#w}RM)VmQyMBS7 zDB0C@U$`%6ovcKG#Ep6QUx@euY$*7UCv~Tbzwm-Uy9yeLzMFLQns}mRi!<U@TE6Td zoAZw>IKlGl@#xxKE+EY@DDX+tMxwxK>@~F{K{z`)R07B3?s)>hdGVBi+j)^b0|29V zlipz1k$QjvDXN;86tkw_f9c(1rU-9x!LS(0LNFkxaCJ}Y=v6M<wBwdqn?0e|FZsx_ zP{-u-&imW})e$`2AcxGsWMA-gh9G48%XL0FZZIXvcRrD5)tO(}lnRpr7&0rF(c4Bb zzcrl#%Ig|0ubL=2sY@Thw-v<W#PN`pQJ0>8Xe<!&vlWo9f;-GHQrbo@w1Bo@Wm)Ew z>1vzOK!b2hGZq9w?NR=>2bVWMC7aWeeE<pM4MTt=j#0!|b{X{e)lbO9?%SG1PYNLN z0t;VTuh$Dq3(?H`b8-9s{qb~rapA>;F=Iq>_?uu#WCw_E;#pBxd0~Nn_u>Objd}Fm z@h-?P|IS^76khZ8FL)Zj%}y@g*yU5DfP<fU<6}fe`LQ0Rjqfc2v;|ir|Bnn9K<1d~ z<gZLn*}`vVgH~K=U_G|=$}zuLpt`*8i=;}Y<*@KDk<AWc^d__T>i~c+FY-`|a`Mnm zxCr7P{#)u!Pf4pNt3jEjBEgLx^K_mtHOP8f#?$A!Ves1Ta%cxmW>2hEGpenkj(%u? ze~qTavT953FOH@GES5fww0$0L1c{DlKd=TT_x1oP&?K*0xbLq@lV%C9FeV&*vyu#1 zS<=(E&^a1-+-IZRED85kHy48U+5H78kF-_p=QsqR4nL~7Q+4YYSXOs|HE@qb-eYD~ zz|kGHKS}Wo`(m@`?#g$WQjd!1Nff1%mfH+Ge&{}_VcVHcUcv8ox;4`RUI|!D;G!wu z65~Y8PjqKP+Gd=fNul{jo`Wh*`zLGl!;VB!E^Up+;e^pEiN>u=kRdI07x^C^dO>Tb zqP^gQ&sQ3oBWu6EDqEj7&uGh=<fh6y-~fh6kfVW@Un?t2=<~Rt6wg5Huv+E^ZlY0? zJaDAB8a%LHn+TJb-2p7Fd-g8&A>Ou8G=X}K*b@V99dwaeqAENL>mQH-6ioedmds|6 z@KQaY{C_c+Ts2itJq@1F0sT?MXE;s*BNv}0`N6*c3`!Zq?`&r=<r4hvhb=JQu$HKa zOOSB-MpyARYN<@Q@xKzn2m^)fpIOeeEwAD;Dv5e?LZj#WP^}A%k$GW?9=j&7X;<6v zc(T80v1VtBvYyiY2H=r2k?k+(9XSO~BG+@c{5j76Ih{P8DL*eQHnv^<C#id;wez47 zC_1#QD6}6AwR9VqQx#j%5Ucjf4eVzfv^)Dh`hnNUd{YfF0`kA_w4{>=qXj6^04<jL z<2m-_L$-#i$6~Va((t>7(vKJ;sd?D@d5&9dCrQnO#X9p$jFq<MeL>Lwen<ffpHLY_ zwCYBu{PJH6QSg-;*}qejCB?HKG-A@b?NcfT8C6dhz1m^5;0T}vzyP)hrH8Z`YW8T1 z74f8w0UoaBa7)l}MMLTHBk-0^j_fJkol~o($(RYxlT0sgF<=Wj*`gZ3t#DeH903bw z0T^xe8U4-`v-;4TDje{m9)_CTNSzP={<~AzAKcyx7d^RR9=zVhe~pAT`p6!7^o(-< z+W^LA>ZJ;>t3TT~_-GLKca?e~j#BlAEJ7Gt#0Q9f8D{L^UTebQjaKJf=M&kMuwnq; zagf7do`+?QgJ-&<Q|CvZh8rMw>OSj6wP0p_hXK|hZG1M!)W-L)7io9!zxpZPYy5Zo zM6Mx{Zs(&`4=Rwvq<OC<?g7WMW1Dsp6~8w0>4tXzn@RUn?YPL!bY0UMS*Iua-(<GO z=xe7`R)nbjyHsPp%@t?!F-?XLRLl+P)iHntUARFw=~KxWO|)ApM6yl&NXB1Zg8=|! zsX3qRU*9NH_CEh>V6+elUgE(7L}DAez@?$rUjB-nO|LHtcohF{fGz_rFY*@?szAM; zJsAGGW}r2bgD&<F&|b%GT=49?Uc&}1u^OKQFp7!ZmmgsaicI|Wim$xBs%q%~ejs4P zR2((xzlWPakmh(5GhQa7#;&GH@#sk*fKksY;a|_G^-yR-*NNwA8M8r_=E3G-ww9U_ zMzhS;Is9v3R{F47FCy}TJuq~As1|&CM*&X;#Tk%57G*T^zihurA>ap_^PGW-`FC-V z*-L*ep)*MY$N|5HdVyJJ&<5rma!s$OTBWqE-QynX>wF8n)kY)u+d_S>6%6c$A*B4W zBjh>Bz!@FJ$9bX|XZ$N*MRv}NZgy0<BfwW33LW}ypB}h(N5`H&GJ(hs7>`PyeRo+& zSfSGF5A@9gV!8j;t#fIBy%rT79U0F5Ys2><HOO;V2ShN3R@%u9xNeT>w9R;Zg|e=9 zXi$*<jqg=~ioJN}d<cPC_(LLEI4o;RcdU^SkKNuguF*1^n7<~bu}6w7B6W0)WsTt$ z+y6Pn{!TJp&=yPCIEf?p!P^1u1BNOiZ-Qw;229_*0{uQIEAo|ud4M_o+78lLlmcW6 zy{K`w@Fh~X{rmfh1_H-K-fhObZVLVy%LTkl*O>;x)sboM2T-td%qV0YyaoS@dxGjE z;7O(T7psshC5-d9p{%_8>C$0!`pp5Z&|KZk_e00qcd-vpo%QTk6s7w508?R53Dg8n zXDW8V=<GtUs^Cw+I{RtJ57tq#y`%i^$2tgDm`WYKOS%CxkpU%D+J@F|0y)1(T)As{ z(bt<PD?rGTEr3zW=V41`?<Ch%UZ5KV#3g|JnDb?!FYLfZ=V@@^4}j;fS!GQmY<B5{ zRwgO7xZu&NjZaJwT?w!1B)NH??c{Mz0hfyAFp)dacRd;o`*{{otHnW0AV&kg*2zB- z1_#$3p`q&J$(3WP9L+Zb+9$yu!7G*<R-Wd4$G}dNA*-7XWZ~f%jX5CY9;-|Uf8!%j zOWsHfvuk{4i#ZTRSB@K@`R5%n1^y!t)%WoJH-9#{$#n$2^BWLsnWPDPVl)Vy`|Wxx z58UIT<A?FV!VIn=P?3`O&uf8l*>>z(2Ji81^Gzor22H1f0|POd5M~8OK!@7+?HKa* zJt}w|<>vyr&5qYh-2}csig84vx(Vm5nv6%mO;t(1v0=sL3UZMky0V>Mhv%J|1BLb` zlHPc7bL2&vp>2Ne#ATW7Gop(AJ&-oY(V`0Im7nO?T*~;<{nd$vt)>{+g3_*3z{9oK zY%HjZPptN0Y6w`dH_og#Ipp4Igv@_;V2$Tq-*QoN?>B1O3sV-{1Co5#o;tAOCP&)P z)u#57Nz96J2*HII!M)Y&91e>?9=WV~T$)i~gxUMO*YcNbaOA0OifNJ2tEt-_1>^EF z1WoUnY*!8NBqxO-90bGs$+J66rM5az`E^XC5Vg?%UL|#J12R>h)Q)qZZ6ySFM&-iA zhe{;HOH)<PdMma%JE_Z@Jz(fC+oq%y>f(kW1Ml9aT=EB4vfr17`i(@2WAM;btu0Bm zC`bl!xs;3E?r1!9G_O}FLQlR23(cT4hJ8M=mq93@uLZ2m?7TX2KH(7C2hf|%JlA?I zRR>&wJck-&?tT<^smIJvwF^e40b77jwDA7ztN^$JYr5b5ppGG}$U{ixF+M2i>hWn( z2fe#^%08Qz)JqMjh=JE1YW3>%*iSJs?>+ti&qS~wpS_!gHvL1tA4jmflHh8SF~@2j zP82l6(&~S+;X@k;Sofq~fH&F+2CuHy{@~T9Y-7OxHdMKO(bQ!5GwLrV{X!V47e-!^ z9O6qy$!O1+Y&an8cVZ`p@IF8f0=NdQ=ab(wIL|0<iUdwo(D=)N#a!RmImLC}0k-7A z-OsXwRQp$d^FbNnt|@IFHlQL`R9F*TlSY$WM)PFb(i&;;0X*2Le8~<J3RkQxu6Or4 zzzM5`(%}~~a01EW<aRU#$CA*7w;_D;^Dhbr<c~<0_qT?qOnTgZ@gwJjh+*17V!%Tw z*Vh!V@;Te-VZRSt5bil-*&rb>UMe*ci#{c_*<o-8NjF6NPX1bSY`3qCBt|wUM-DQ% zDF2t%pk0=}9++BX>WOBmcWkKgzF&AG0QbrQ6YB9i(SP&dJQ0*qFT6oQqW~Iv1)MVe z1GCmm`?^uQnhw6+BSm;Duq8Z<R0@}y|H6H6uyX}`bI$T#`2`WM=0LM+I2YPDp-Fox zDDL%!_Rk@Yqw}cEblVA}neZQia}aCAeQT7+dtGD+`#F>PF9j*uKN+iHy<U708%7=R z;pJ0B^Jq>9<7iG?Vlzc!sSnQE9Jmpj46i(o;eppp)c41`vRl4t$mZi7G;2bPYf8L{ zf-u47E^7rzk6mKvo0RudFkX#m9s`)W(v$~QHn#X+d;UbTdno}eZYK4_8;b1LxWV=A zlg)Cr<WHSu?>qq=hXT0{B#5Ix&!T^IZ(HBkdEQ{QqST>s?T841sJ@g*h)7BvUrOSc zl&Wdo?75RfJd!H@*B{zennZ-)S)r;n?`<&(|5)eBozlvk)=Hair8eIx%UTO<l;G1m zJs4uan)~!7`IT`BMYberQk~BbK~XIyiJ~fIBE|C^dr0f1rRvnJ&QJ01ST7T~E|MY| zdDrtX+?%m|;m$ZREPu~P&M)s}WUU0Dg=#XA#4|eiXJ9OJHCA!a-NZ>!Eiw3SV7y8& z4<#tC!L){fh6Zn8>F2_dW)_I3rgbuB%(z$ckIcS~2z0QZ=}h_gD1rnl;bCO3qlZJ? z{0=&+S&pTE@r_p9wzC$q75kUNo!XI|@E1$3<WuQYQ_VwJN$7M4>2-{4=C2k0eOUhU zFE;@f@D+OSnLOriq-q+-1fQvw?ALn&4LB6i2NBKeOXafHqGFM#mqy8nls9@oJ0zv* zUSoP7oR!2(c*4s1tx-8oc)qnecE{@<DRerC7V=yXiQUvArnc?BVIG-yTl!l<TYWxf zv#ljQ5Hl&dS2DlPoo3oI#V7wlOl#>z!WmJ7U>fv=Hl=wlzm*p2Dyo`(?Cj0eN-vtk z`otr&!bAgC2R6O9!R~9-wq|cllTYc56q+aXROir6P>XQB<lJzaR~8lh#@`;PWsORd zNHv#`Go|f>PiL!o7)<dROlgb@x;z-~>5t>^ZHeqtWn9<r)tdDcGf5iveHBMlLz0O` zQ+y^RkJ}%5fG4GjOQ*V<kxdYSh7uYXF*=5h&1;~R^0k!7R!uX9xs<fG;%Ub~lTcem zGn>P>SFpHGU)ej1r)c80UwPRlTi9;iFDAwe_)>j(2+oS(?6&41!L$s8(hYr5E+6WZ zhw&(hz3v37J5?WQ(pBgQLkz`(8#Reb6*Qc^K?D0DkNPM4EvA$NKC-H#?Zo10=yPOM zo^LVdlFB~>wSH@}l157R)@UOUgy+brMFf-_pKb66ZS2qlq5dmxCI?5=ci=@)?QPFO zgCu_WOMR}j=OKw^2<@hOoB58fs+dzrgMhvi>?dZhpIZmw&a*nLI@tf{!_*Db2=e55 z+@2%JKcm-4l$KAV3#A*XUF;r+i{=cq=lLm9D;Oy^jZaS+DbowpR+n6boQr4M?c1jn zsW|f~<$1+7f<p~2Y|hq2g7m_-#-d6e2-)1iv)=F#V|@yBy6rbBp3dL<mN%fTp^@^n zkXcD&`Ohh!9u?YtEA68hV{lbXrjt*l(@D0_;fbh~NTl;j_CTalO=Yc}Gj(cxtg)se zjp{Qsp(pH538(CSYu_?lRH)>h+Hd{ruke5PM?r`hLji0Lj?*c95F}?4jfT<-H#{SY zrYK(M6@>9fYY=Eii1AU}J^fl~SM!xn_%%zsdFacbTDqu2|5qZ<3Or}B@sXhBjo(H! zX(n4L`|@hE)s%8yZV?APeGg+KO=teD>UUK_b0DP!5~@jv7*oR?d`Ev8YweUx#(y{7 zbwzR<OC||w&6_q_Yrij^POhPrY|TR&^-@7SQkrvEgR2%nNrOj2u$&o(Hq_ph`FjG^ zfG2hq^((MN$YLyO+XS~h<Hkt6fm**LRI1w4o}f*3X*Af9oGe^uXQBzxc|oxP8W<f5 z)qx71Xk!dT1KLLwt^eRKNj^i$A2Q|AiD9Vw5nJgSS7@ieiW?<OKu;Pab4=Kh{+|DH zKTVA`ot&YNwy#+-7Q>t|7isjtPDCW9<*TolD0r%$D(#?h0=QT*v@Z+>QPE4*xKpq= z3r2*IS@NMY6}?wDXyQkn_s2Y?pG%ADjR{cT9}Z)Q1;MsT8nWln!IGVtRo`kYmCCP) zz>Q&6z(<@ZTn`7Vk5%h)<C;=1!}dJH(I}E_zP7kXQ2UxRHoozeyuLiBG2I8#C)<H# zJf&O7X?#tq_#stOC*LIzCfbfpp*FuE#zQh{Pr)|ke~tvTt|eBoOoC31Q{D=0m6AsF z=P9s~s?jL0eipB`vMJ@lM??LFD>i5R>L<gTHTYbjwe{aA^Jo%Gnb2<QCadjM!kACA z3C29QGLqQaS)-{n96YVB9dr>C6Xo8we7;4HgZP#O`Y6P*#3;0ER{huu+O|ar9kQ<d zs+f9t#Yz(SDKDWKH>xjL?(Exr#G@b2fSiVDDQ^Bmx0uBo9hst%pvgi>{+vB`x`R}~ zOK72Dj?bQXF#B<X!WV^PhN_8ZAWHg;@CISRFH?XfW*1IgOw^Q0a-1tkWbHfW#HAOF z;wEf#2~_{!L#xI-gfeGK?muvNDW7^rBtVV%zEv>fODE3T%CL^e=<v-S6zu8Jk;clr zfyzDYN0>lH0LD7k1@$Cv+z@O5D7W1ot5SsY=kL+P9U~o)*&8)v$NXFa3nd|#8Pu7F zH1=a?@kI9W9E7)l{P=lfo;jCZj7(j(wPXHRFm(iB-I!_Pb5&MCG~|ce>TCg5S-HF_ z9zJ=+FMjs5avXn3qLFg6hFxtCPp38xHPvr@W6-FC=pu*AdG#D*^1q7&T(Xqn#*o`D zMve=;h%GGX4(2C5#^nr!nsezyGrV2M7v=Y$S$DROp~-drNfq<andAk@jNR{_Lz$FS z{S?s(#4BOXyW=6}qG~eVRS72h_n`MgLfNe8teWXV_QGqB4)7}yoasnEOT7^?yp(J^ z*5)CP`ZT1+5s^4#&Y4z@l3&t5r)K;{EiJgvZuK2C`6&zf0NE$y{@E|gT~f+lrZTA~ z*<V*AfG;o>sG;VZY31E;xRa{Jk7}mRt;K(*y~kw+TMJD(KUebM4T5`cWl3%h2mbl< z1gJIN*HMOija00P#q7MX&HA~F&$t-j539!FQ_uUwx3r6EqRg`xs-_sw?j78bJ|g9O zsY$;uc`&u%qMPZobr}@pfrpv(N2HbWMa?<B`Z|9YFFTzKCIS6cwT?>2K+W!n7;i+L zdCmM7p`TqZ&Pc69D)+%<v^Fa_HgeA@M3(l603W5P9k1vMk#7h@=*dd$gaDcjUS97@ zu;Hm4wHi6Tr4ovD<4PMPEbFxvdP5Lo0PD#We@Px8-Md|jm#8$Q-c<$Na(G2NMi@0_ zUqer>_>I$&M;E*k*qcIthAO!#&O_e!fJOn{;9w|FQv35)>J}-}9&9M3$s%}mL=R!P z1OT31yhW*-c5IxL%B_%18o#in!^cLZN%)f+kD-Oriejc1DcxQZ0aJr{nz@Id5k@VU z*4TX`MZfFK4KD3-ezIy^p;T1(15f%f^3#?e7=4fZ@4@7P5x1Zn88bzRGa~`?47a*| z%eod!6t#t@e$#|^FQ3U=Ma7fSBpB1-nk&9acd(2avsS!6l*FfR8*^xC$**yqIKm1? zl+#&#nH3)$P;r7WMQcbl+8pBzX$S^;DKXm9T3sQ%<~3o3^ilRi;HYWCf_-Hps@a<b z1-S!EO`I_Q+<28|(oy}6d<%MZ?DjlgW6~_8V~`QE-n|M|w#m&TEZ#>@RS9oT;N(*+ zkS<0Bi|h53ht9wd?uXFRYGD0{ao|0S@nL$mcAnh9eY?cQR)G9zIKsBUb6+HbDTF|k zA4drm`E#b=e8~IyM^-ZxlU*xzDPfN15wa!3PC>rflrqaOF7tiR?gS0N=%cPUIFIBJ zUa=bVe37<O`j{weo%B+HuWiJ>rfKGf)|M|FIcC^Bh<;xB9LD#_#cMIXA|5hz)25AS zqcf>qZ7(CRWj~D_qKDr<=|ZK@mRLx&wND0bW`K?&SD!Z~P7obLWa-C;uhYrE(IaXm z_YtRFfsfdMuW88KhOHc?&im~ABZ8=M#zZ}lufTvdEAD6MH|r*vWmC3p_4{J##^8t6 zYm+C)z4~LPu3nG(sV~*ey+d*5UAi_6Df_VpF)lHhxC*~xbs~e6?lg6(=J0jj3+$~$ z`^LF_hlZ0FCTPwI9y;=~Gs&uZY$U`i;~!>x3l5$|tYmtH=$3V3_Kni{tKvWKB${$3 z-h7-TjZsmEyZmS-r5jGe`ETUc5<BnVTT#>74|r=jte#~{^lvK~F(aiubO$q%#yop& zO1Zt%V-h5=&rhmEn7&|NTd)7)zdKes5*Z=O`DYvp$`~7Tw&0xjnK6DVt`x!q5P1g7 z2D>vo(Y_Yhh5wL<M&nDLGIgt;$?qD!a29xW?x(KNz44(QH%2B8iWn&0IKM;~m7?lj zO*?|pBs{qK^C?P4cQ8)H$K0VS(l9~(qVDv>Hz8b+uzdL!xWQR=it>tzab(_ShAqpC zYj~)t$>pA3zQ^!&w$9+sYzz7NhMs>f{oGVDa70kSS2Vq8Em&OrQUP1GaO@BpIUpoh z7IqvU??r0H;$b4=i;oaZX#IlA2+RbmSz>20&+bJlT*WUM4Xk<pu64z2Z7DOD2;$S+ zE4N_Do9Y$VvtxeUa_q9PwCa@VvCZ1~jP9IYM-JjP;@|X2q0#5w(|ii+7#_*M$zi)@ zplI+Zq*UfDKd5Q!G}A{Mz0fM$cd25+LZ283tjkWx1g8pYfkXYz)mJvpvN0QbsD9S- zI-N{#ZyecSqI5Ky+Snz@7%GqRc*n*PMhM7`0EA8CeWB#YFS)2U3&5Eu<mvo=0aMb@ zuCqm>zuO%Cj8tBC@Jd<VxB2TeRqP4o0c|?+h>wW1-R5;Ix}INIp5{%ZSTqO*EZ@q} zMX;oPavw;wxj*NLSSpYHjLa<U!U-mLyV)dxM{tUs?BrIAVkohDXIy!<r3ftDt7lJn zOucD?^S}4e#9(RUy_~aWujB0$jG-GMN%Mhf&n{>enJxyOpJmDv70K@NDHOcF4+irP zIe;Y}b*Q8sqiN5#_E(uDk6JK;#4@U5gLT~J<^Gl5o(Dxr$`c{wDfOzR@gt{tVQw}{ zJ$;*CsPXtX6No!eqx?!*1qIr+=80lE$J{HT;@Icc;SQci>OJSUAuYrm-}Agl(8PzI z5mRx$Vp|?7U$BWNDTXJp6LY5bIy4C4d7k3Mgo(s7kosOCZPT?@`jFXu2#f<W!9&S6 zgKjoz>hYDeF?fXxlgC*-RgR3M<&`kj?R#Rn_=K}<_<N%PpBH}-r2&OG1mObj+ko^Y z)r(tMg!uOWv`fRZIeBdu3xn^3xH&cm9bh-TWbk%;X*&wGM2C5>`MYWD0E6#C<{kpv z1-;?ZcwhfGH?wKBfo+6-J2A;XQLcHj&$EpoM<b-XZjvW~K=rcv2?9T!YSgF|)Bln6 z*HKly&l@nj>5vZT?hueh8l+PJX+&u$>23t+lJ1s}7NkMCn?rYZANoAo&-eGP^*-xa z>-=?g-8pm3Tr+dO>6ZaN(pw~ll9O=>h93esP#$7sRqc9qCgO4J-L%z+ECnE*0}oA> zRuWu14b7XLOYv_9jz_EKGmO*>*<W`4K3J~bWAAP(6xi{MbP9MK_D6#L^^G;13|5jY zgKxfGkp_v}WCS90Vx(%!y}14j>Eowq*#5}dcaBXp%8b^5S22S3j3|4b!6iy?wU|lr zLFqAttj1<|kRviUNX$Mn{07>ZJo-<j95djO6A3CrqIJvkb_X?Fb&@@N-zJYZzo$Y2 z4^;T>sqjuu#5+0fS<Swed=w*nB=M_}Al{kR0j??FEZyDhl3mO}aJG=CYo5ShS21A2 zm--8?y>Hi=I<g6(^lNf{n(}_#k%Q|C<rP)w)~cg!oILfHV@!M|z0&n-aXf!b?t*Oa z%y{c_24YxQay%xZo2=^Jl_x@mo$KkmG&l+9d#%FiXNhr3*CuW=mS@ArkrQs+@+iz1 z5b|2jJd{t#9}B(C=xCInr;NN+Fa{Gp@AY?pUkt#a^93Fr7=>42s?%P#hh(PxZF-OB z+yTlR8P^qOTfLxP&p7Qob83s4mrElt2torst?du#T_^nnn`FrSy1L&wW5-T^NmEI- z9y?js<oEU!o^Wy{3-Td7(K@V$sw&0YH%OZn)Hy;q{vLvQq{O`2zfK+)hpC;KaHTAr z7`zMj<K8pCMs9)oNXHbtM@8Ke@A?w`UOLymE9+CWP@3Ds5HTn(%o7T$bCC;OTZg|y zy=Q@2qb)8<YKC_R_)Z}`?xzDkK6r&Xq4+Oe>;0jwvbUw@4<HAYZS?Z3prP)E2)a%u ze!LscFI+{O5{jEkyJgtOt*YA51I}~b;qsU*%!rVK#IlQ{^R*=GtKVBdoLb3N`%yz1 zR&8CWGd<*Kah0Z%%%WubE#rWKqz(9<3}iZB+CBJ!bV<B57(8+b3~_Zmt2|7DXj5U` zJA?tAh$<Rn2W%rBny2DOAX-0X4!F{Wvw9bFcq3oTyg&pG9QkWJ`|TWgS7p7eJ4L>G zkGb@M$^>it0?d8Af(0Q`&tRNpQ$E3Lx({J10YYF8tOJQ)>+=1JI{An0gYZ{e>p9}2 zSdB~b4ivYFc^QdXI2i7&E1u3c(=W<rZ`TTL1pj(y3>l`%8!HR8#Ef##alglI*Y|mA z?cyv6x@CGW4zpj<0p-h99Tv+|VeuJpaS%s$P8CQ?R+GG-zY^lO^nTN<`bRVVMe>UB zKfRqF(x;4h1*WMaTH7dL(&ZC`j=RW@2K0C&3M#nbxF*Xu=Y25E-zgWFitc>h5UIB# zS@xgM{PQM@M7hNdDMvoL2xg>?EWY4J3q{E4)r&kv8h;Vo6=2D`<-DCT5=w~FUqd6@ z5ChR5!paT&h+UtJ$O*~pdzt+L<~72}uXSX#@{^jc$=fm&V3)a#0f%SH?5C^KDW~>w zo!ke_A8pq<k!F9KRCaDVeOwHo?qAL~sB#1*Tw1GhSwGhc8O{fubvfx92QFrsaO;zj zeL-yvzR@{!=9d;3jNz<c^}7rgWtkr!qr=+|)qdv7aCLf14K5?)G-sg6j_D0>Tw*$( z5|_$WPrT0dGJ=;^bDwmj47ie7|4RKb;t!7_W=L0>eX0t-bVYSrg`$?pA4Sm##wc_0 zYM$|;aM8C?8q1WqUf=}3h>N>v=iSwhqWJ_0J;xBJGy))R2oCM~uxY3`RA@IqB7bc0 z`|}f==W7)WYirj3q{vK5S}612eB(Dz+@U`CTuYE$b>oWD#eaX>OTQ6$eY>uYwJn&j zN7hm-1s@w2QYc9zv655OfQ|`xQi4dsrn*l>lMmcFdSNnPs@auiWE?1%qk@-68b_L+ z%;^6Z2!VF%lukt1bnWf;#Qm0RiO5TEXW=<W3YOFQ*@u5{JJjqtcLHa%XKuvvo$IL@ z(6$OH1JOw8@1RC>0IRWpwlo=;_4o={?zsDNT-t$SgD_JCO>ECbEWNbP6I{h+>Nvuj zASbR>j)ZHkt+?jVJ_D|w_lY7s&c5X{D4O{&6%y{ng+bu`le5vZe|VJWTUXoB8;DO< z6V4j|PF^YF5EE1DRWm%O8s>#$FfH+7<4s2+kVu)&0^k$5263@u9F^sR+%{<Ar0O)g z{*y8m>YwzkN?N`M>j6+=5F+l6Mjsw(Xv{R-E2?OH^#LYznfD@kXu%%5uQh_Be>XbF za>YCQ_OGNtp61v1oaFB|)SFLw`**So8?6_={S4qcz$_sn&>$BYk9lvwsv3l7>{dt+ zzE-~vC&(1aHk27hM)cGN_CC!K%$$;VTZ&iqYo03N8+HUnN^D%*?pJe1XY4fia`*?o zn-vkUii;+g--++(QGRNixv4^u=}<lU5Zg)zTc`gBp7Lk-g4kf31&MU?JD+s>_T}`V zCq~-@UaLua1^8F=4jnA2uZ_dekTK?Z=W1Rp#xr~xh%b-h&7Zdb-*}2`%ab3+uFEqi z@3g{c6t;UXYB+j1+`4-yvYWB{;j_W1U+f(zQ9cbl9X&MxB<^QE)Ofq;2#yDsH$4x$ zJJsZ?>o|iM0Me?(p`S3U(oN343hsq^e*^$oloz}skX4h~=15tk3)h4J*;wbVIYczE zaJ2($=pl#q()dX$>N^qQ#dHN&&&70nzjx74`7F7fQQ@g!TvcgZ+%WuN(n^V^>=w7@ z*GucHg)EHMC8eS8Pa;$Zr&NIdYivvctUFWPss)1qvp}I)|7H<5Kh7WMDTYQL9O#Nx zm7>eKe+ZEw6wnng2tDCn)h4hqN@({}<`mZS)`Z`QKUVJl+-F7ge|%QJ53Y5WOoLW_ z32%pyVk<qu16*QZ?#>;M{4Xr)^cW*T5kv)AxPW=q6Z*~U?Oo+UMp}^q)=}62a|NBG z+(qz;c;E$$8n{T3DZCWt|Fx`yj(3&%_>G!H+kr%;=!aE2;Q+)<B}t|V<aF`r6ap*T zmELLe)-uPt5+uhXTKenm>|VSjrOr0Qp`byc!Oabh&U+KS{~Z!nphn|D6R4hmd<g+x zA1@^jSRVMV!ydoo=_IrH<fz6Y`oZ{Af`(ZDfbm;VY1^|Mi%Y*hJ+p-ZiHs!HTsc3c zlz4LW=BA~83n>*9#&$B!h6};eJ54z>im-e2&OEX;A0+N~62KN)1iA!ZaNth}To9zK zlXGB$*@k3DG$+t&i8BrW(70UT9XUOG+5I-W*&i26o2BckarXQSAf8lRo7($V%jc%@ zm9#Ool&%LtE5bHR64+uw{=im0<|lASI}B_wuv<F|idb!n;h*>WufEt827o^_xc<oA zqG9gzS_o@0(HPr3{;6+e_Pp@;`6*>H&-<RP5WE-%v@i{4BmsY*Un%Lw2y7kJ2H-~G zjn8`QJuOBkR!wh+fv|_*;*<J0$fJ=VbycbN6Gb6U^1l>X&Ve&0BFQ~f)M?I=Bpdec z%kJsOQn-LCfCU2m`slRJC<m}DDhx0W_}Rq}8|mTm1T``Ej(z+nMCs1wEy7N&U_zl~ z-fs92R(L5t&LV67@`r)^+{dl}(Xevl8|4d9Rf&5}RDY3|z@_-W#d~2?|CY<=h>R27 zQ*q)wCHkBtCIHZ-i~Ek;_XR)!XP7ahs-39XT7i_B%?cREeXK4Z&8+1^UoFq|{SJeJ z1K^EcHlmVr+#!y9l~HxOh!+i7X{YY)6a)ZMqJWq}D}cO95@9O{`L;e7iY+h{b+?Bf ze6cHyHNP19`~sfZ{&iJav1OLrPZ^K$MHHEw%Yg^C$n)o*r!_F1k86t5-@H?>R$g6h zHK{6AY->=hmk3<HsRQcIWD+ua(m24s#XA@e{~Z>v4r>X>OQ8)FZrmA*g<`*W0uQV= z!NkFengCR>6*^yh=HG`@w<lE1RrO`Wk?DK8&049)0hkzI8(5r1{NTPezQ6-6-~*{C zII!J8Vt*Jt>@70p+SjBF#Od$pDi!;4;@@@H6@X!#sL_lItZA_Gv19CCM_`oq4g%e< zVKL$mVS_?JS$`-tAn|Moz%!ALzQ;g_`OM##X9A=C?BG*m?Yt9>DuqevTxtMkWcZSq zGlT>%W_<Jut^Gm&o|KmSCI~8C!bEY%T6QIP?Ajc*Y0{~~z6ZUzx;H^5&h3Z<OqCFf z1wZ|MBUi6ZfBe<+vCY~z16<88Z7Dy+!^_QYPq~4{@BpHOk;Vs+vF4xRkxYo{>$PDO zhgikwMX6g+)t1b*nX1O@HuJq+C#|(B3DqAI>h=qZv3YN6w%`&*M`C<D)Lbsg1U#P` z{e-`=5aC6!H(w65w|6DB3_w?=)Y>OsEv4Yznqm8G1}tn*+|DNZ-g--Oy!=sB>3hWn zg_xci9L2Y)iPiUJ)ybTAZ|)bb9wJ#KNN-egDX7aPxR}0BslCaI?~veRmlTZPwVH!6 zTw&^EqtNK~<vxmoYq6d>w{h1CEhnb3^9;srQcF}@{@UZkSYOvC@fJ&_-nyeUPH)!) z>qiQLEbR_dBx6KsqF^bpzycjtXD#Ds9^GG~@}@u}PXw>!Tnxk2G&ux#dh-PyE>_7{ z$F1oHp%i=vA5M5JngthGRMOU~ylxZ(u3=Zvxr<*7YFoFoLUY?h9jKS9s1kM^T<l!q zYziz-eWvx*8EBC_&E$|ODy0;$Wc&4$Hw~3khybepodg-G|9nt^W%_dJM%BQie>)Mo zlh4$ADVr_+3GU^{1K`dDA8T?2*RK7cU`vI+If6cyYwyU~KYPH*J$4TP6*hLg2&c?t z3&#ap&HRM}&#P~&6W0~Rfaap5h5RTg59#2C^lX(@6+)McH?2dX`3%u$nimZZ3htYc zxc)v0z{lz%<ya#;Q))X11z>;6SR*#ta)v@(x)Q#)dMNcM%llbkwRKc}`6(C5Vuu6V zaR57wikNTDzG9$@j15|Axe1C6nE5#uU}CuQVIs>iUVo+2<v51xYiRS>nl6kD#JO+{ zZlQ3xf0laHx5^-T)c=Y#u-$A}{dp@Y=ay_$jIu7F%Zp91YG_DW0G=_Zs1(kxvi6U& zsfS~H&XEiQoFDdE)=>`5vOl;}rwqF~yqG^=C~#aTI$W#P{=Q2HnWnB>`r$ggu`LYb zr0S+<>^$?uWJ5(!-YgS8xjy<<Djp5W$yJz+p-S0r3Wt*ALI8p9>@VWZBpg4fNAu~h z%S>3}fHQNX%|)Kuyc><Ij(abGv<xcZQ(YbIRU`o1mdLy#X(IzJnC}8(Bqzp7I-?!Z z5(l326lTqxgfe^hlZQrW#a<EyTe6d0K|O0JIMizRlc%g2=8yT)Q}F}EK8{mee3R-- z;aZeU5tWYPS2OeYHf1i5pEI_5cycnVkOXh#n56Z`FeoYJUaB*;p_c4|rbtrcWsy~q z(jdT}dK^#xomQD)gu7jn3n9dgVJ8InPfi{eJSaG++a2Z!9NXmjIxsqhgErsBb!CdU za16a?o<m7W?WP;~y<c5<#4=W{cTARbD^3<7ewO_4exTMudZ4Yw+Tp49`cQkub9j@& z=*Cdwy!Ws#&1dR*Ol7;QVWr;b>HUNIy<YVF8xBgZ&Fk9MR5iYdH-Nxr#HsLsN`zQ8 z>A3oyhNBfH{a<WPgC3`i`IQv+Kk0;5gbs}wFK-s@N?D4yQ(Mc{uchTwp;AWOn79}> zaRe%u7Jvz{x!aE`sEZYAu=6mtc=Btk+z}^91aDwQ9MqoT-_+rMcD=|nl?CmR{`KDT zcV;D4dd3|QIkbAJRNqEj9w?W!t=mWKew5iAS{#08pj?1$JP4!-D>Xrm7Tv~JA>A6* zKjyV+_n*_6?gg@hzm_YdKtpI%z|y-{<MrW@AJ#?6ZoD1t(y*bY;vy-nUpGhn=4V4G zn|w+<yla!~=b)w|7Z4+nuki8t_FH$~;C(C03ds9H6p7Y|KXE!fJT%jg7`sm3+*e0B z5bH(lUN^s$*pvPaOU2E>*t<gP@KyuVHK&OX;kfcshF#NC!xmAIA2bQ!@g4W47e1V1 z25SfhHtuBkG$${qmpDx}?&Qjyv}ZxHy-OFVT8mf2Rxx`T1;j<NpyC_pUi>2W-&)yh z1)e27v$=`hd_!x2;U2koQTo_uGZ56a<7#T5A>Ai5aPeAmFLrX*uJWS#sQn$?f!^LL z5F%`2!W(0qEjV<Y!V)$sv#OukVf-+t*HWm8vbH$Qz@=P^(J{iQpE(74Ag})Wl^@kQ zIs|yesRpFWZJ)<}_}F+oMVpP5z~0ylxt){_ZTSy+^Ac}uM;bXloiuCD_21bt*A9u1 zsF;3igl6$okkx+s<$JnvXc=j$@V@#wpw;8kjnH@M?#=42iVu2614twO^HQ=P%t;z! zJE9CVt|sxy3qmU@Ean>)mheRB51HK>X__kaAP+B=cM;}|aXO<a4M_ERO>RgHnBy4A zxrTa~9+1-%Fu$i40g_VD6^9|)W-9bj1toK%g{@D4ZR4>Czk<DkaJS37HfI-{vqcts zH};yD7MpzFt=yVK6Rlls40Y2Sxk_8=mmcH%6bV)3Ek=sWN{Z3>Da0Oo2X^9qyDL0w zT)VB-7IXBB@I!13Pp!lz3%7<*6F$je5ZXMg9cx6rTBe!Q{axVulyt3S6&O@D@3C7m zyUQ6|bnslc)asQeI<+4yP<O?8(_?Kn-_*!qX5MSi;03ujLD-c(D@%K`eg=#<(hUHD zA&cY;ZEben5e88(&kWHXlw>Oh+v~qmJA-BYB!FQvF29ts(6HWVQlAeA2lZu?#TyQm zS+p~onoBZiGbxb60_SUK>{OxQm%=iy^F4w6012wVIatXZtn&p*cAGzrRDl-+JUD|* z#Q{o+^>{T-NHfAxX}F;%ioEK1Hvcj63frT%?S%iM+ah86#dMc-0)_mbuDJzSpoWs; zE2XHXbxul;yUZP(!RrQ!)pjb;n?h0WtcMH9N(fcHeDDtgl248Y43Lm;Bj00f6jf*J zRJz%{Lran6s_O|kyGqq}1Un4r)`lZ}J{n=+bj#QJFN?#78h39ivxJuT-J90t>9jd` z%a<PZx&}1JcNQqvmXcQvzX2^|B;%-9SB>FAvQjXs@YniBuYaPap`fBZwSIr%%jfLp zv3T%7>ikc23ew`y{e8XRs67b3PoAInh9E}TNqA<Gwr5j)8S8g{`Hn)HW-@jXkY18a zt57`i6e$a@e(i@$APt?c@SW4t`(R{fwBz8Vm|1>0M^^l_&+)Om`B&)2bnX<@nj-?L z9E1B(0|Ob`#MZ{wDW3d=(CJPDMLv?F{Sn{wi55Q1y7TRKr+(e<_Qib9EJw6&G#L&V z2Myc1T?cowTRzQ&Ums2d`rXcUJ;c-%W~iQLi(fgq%<u9?J!6$j-g}OPt#GIV0+S{L zETouv>@xX*gEmCLf+DjQ0mOgXUVZXYvF^DmfLjT;`iKL3L*N7}{*O~Z3v3F|Mvm1* zRb9AT`#04^utv<Z!$@FOpaO{e<VL_$CZ>S*c>*WJO!o0dN_%t9>awHjMytD|1anVi z;UAV~5%&N!;gH7X0DF!VkHW_Fx6&p`a+dBt49@woF9?%zy?3MCy16MvA>O5C7FvcE za4Z(T7q8m<Cfl)RTfH6huCQ0<=h8l5N$!a)A1&4iSrv8ZYT42?Kjr{-!$XnCB(ohD z+I$^c!*-lE*KR>(7XdlZ|2iXk=#dBK$?L3z0;;W?syNM-TB_3{HIi!DudUr;KT$N@ z;<)Lt!!=ngY=7?mL@@`<5%VV$<MKPj(~6J~)UU<GDdQNEo6DPgk9!Ftlt!j}ywRVd zpX(+Zf^l|Bjhwu$(-v_z9-7}Yy;<OFX!y9lmubTGUS&$X^G4~}Au#y!N|F({7qR@= ze$`YHN&PM0{Kfk#C+5h0KibSD&p0`I^@!A#cJRZxX`DE+_%||of<2`Qbj5d(6Q(Lk zxX9)az7!Aj!OtPHwjOKlRH1u*IDmw#CgvdtwHNYanzGvhZ;Uhxg|eV4^U1$0D)i<> zz$WA#jpD=RkSK9<dV=`?t`xZQr=K!<it~Qi3~ZDhT$COR4!-yLz?K}mpyu{Ac{W-* zG@gy4@)bv=VFMveH&lSH#p!6S@#K*ZUYGX=`9V~)TXr}OMh}i6$zKawS1cE0{<r{_ zw8;w9Fc~DXjM~DH$rQtF{(2$gL^S)WkwVtEs#S9~%HKcdv%br@7$R(Lg(2Sxqk+3< z%x*@XBG--P5rm*z45t%Rw2w4-=bJx%@h2Q`ycdm*gAp`H|0XUfLT~aO7d3~Em<ij= zPxLdDQSR=lvK)=<YD)g4ukle^SL^t-+6kG!T52c<g^*+Q#q*n!CNIaH-vXy)t#_jp z*q64S77U0LaS_dyS_yn3QPU>#!%XL$UH;C9Ju<C&x;uUpSX4fv&v<#R(`{K7ttr^y z@b|Mzu(Vi}a@JRszGum|ew}+qM~2e1A@?-TC}3-Sq+Zh!(4`4$@RJgcfEz${=6GBY ztT1`B$wcmUBmJm0nk&CxDesW~S5n*^#`8p8_iFXJqitk+iJ+ZH?MXAd{S!=WVkPuM zSD^8C$M81Uji|2HQ`s&SZz%ep`T)O%yV;vlj`;1O>mm~&W@GVJTo^AxA5|fNaj#_a z8E|9zYfyU!@GJJ_+!(HyuF9?ap)^lKh(N3`9tnQ5!U)d~?HtdH(!N~xaNc4=2CivL zk`5tRh>QL^hl`hz-<9wX$%fg>Ev68#Z8@mM1~!8XGbvPd^}jJRpBg6eZ}~`12!*$= zcf{zaE!ADtk6bGmCY%=<9^R~YZLj*eyeSygmX4zro}lUzJPW5)megdP!lR~Fh~L7j z=i>^dM+?NU(B15RS+=}=co^6j{g|uJ)y9@9UU;1fX=at8l6=HQ1&oWbhtPe0W4A9P zKPP-#i<f?#0@wcvIbW#oti2*s+LvnxLmXjf_H<dOw`9AOXpBF@`b+Dis6j}Ap+Qv5 z+7mYZ;r-rtEfocS;;V|XUxsaeTuUsZ7&>NbR*%};|LCU_hvX^q<A(*NLg^b{zZMMd zP+NHn=i*<<#=ifYkbtB(%;|rb4aQx-g}kQce9=KN43E7(b5{u0#+hqBevi@qK&D9$ z5)4}A5?NnIP-<5iWgspQ^KmT<ao;6QPL;bpl<OvZpfNh}gwF5A6mTZ;31Vb4ZPt+~ zg}Df1AK8g{Y*Bw!;c(M+QF$sg<C{q>LJA)$Ju+?3PAv|p@d#J>$@b&?RZmmKzKnOC zhE69j%WFSPSr*;F&Y~c3(K&c0Tb{7e{;;>xmUWRf)jCJswCD+dhfGk&la7?%+c8U# zXJM{KY_RYFItnPF=sgtEDr9KkxQ3L);cU0f>ZesYFXdt^beNO3=8i-;bkm!sj~!C; zv`6bPwepERw?&(Ol)D_D`|iTA_{VGPByD=~nvqN(B7nJ(pFq3Ot{TjcA~`rts4sov zw(#}s@FH1SIB<5|tu9j|q{5LLW)l$iM8b$ivW8I<7oS6~02-lhY0;Ag*rKh8nesW` z@zkNV`l|<p1kHm9ZxXr2^#!+X=a%x29FCPpaX4Ks&u_%v#<qykmLim<T<MGdAcvM* zBrP{oI0+kaT%N@R7E7y9St?-%;yjL=QoZ`9UHghb{OvnaGfJ}X(v?Oht;NpQ9#}=; zBk5&RSj5Lof`v8mQua!2&IzI{oB@uN`a2?bMyDwUQlnAWqOQgIXLnC+D;1+0_G<Q( zW7vEWwZRoupbNpw);NL!QG)U!DV&7sO4pH*w?lz22+dy?PLl{2hL_Dxq?T=^)zZkT zrcN<?sotxZU6+^?g(f8Ms@WDY_?nNi_$NTnb5y}fas7-db(7f4uWg~|XQonL$3-p1 z+pH9{4f<CTp`wmsiWV0YMF)o1A~h?C^(Ixna-F`!ZB~zl%ucToX*Icwy)ru@j;g;) zBIU7g)zX})SoD5aK@o#aYiIqjs7CW;(xsz(!HDt(57A5>9UcB6RGYx2T?C&Abb8hz z4FdpZcBQ@#h0(18lb!6f>$ANrG?aziOBypEL=X1f@9?5{9gOqYaA<b5(aR#!=V&Uy ztXPvmc2Cr&^eIV{MQP!EmM!)%J{I|#6>jbem9#<aqXjRuG_nyl{(8CH=3Hh7nDza> zwy^r#z`F7?tG?04Y!-n9w?>dBNzJd0-aP#cS<y&z=B@Yw#s8@5`PiS1c`Qm3P?tFK zYt?n}zLfWx)-98?Nt~K(?@jpIK-Y(#r97#}VeSN-_*^h|-JYlvUwe*SI9gP`PWTRL zFu6s7E3F(+OwUa6L%~}S_|&hpTM#3wgB|NH@0XPyCKuapcN;Dza-<eUHb3m<>Ie%L z{`Br~yZHr$h%|L2;7xq4Xnd-!T#YsrBgMzD=C)vc!|6uhRqOE~!hbX)cbM>6Kzlmo zVb{2e?9^FVlwNLM)Z=z9p7RLA<%v57BT(^bk4O7~6)1N0g;bG*t3r1dV?tuy9t~$O zb(E}<2qp5+7R>|UaB!}E7#I@(B~<sVMGgkwI+RVg7vsf4q^m9<>~HEjYr$fKqZar0 z7OE~^?kpFA|CCIN_V9iF7A*j6cmJDi^*M86*wst+lu9PuNXWfKDWwm@v0}bO7I^E( zU<K+i3cI#yW1&LmZ2$t*l`_rGfOf~NLV;$QXYoa2;+JR-Q<PzE{si0&j+q^~kAB$q zM45$w_Ben9-<PTp0yBGu=fO2(<5}N&O67juS`kJ@<9`0z54QfiSyHy=kd@|msX6xY z%kyz9xYy29YbMNA?-0h?ohK`Y?WNI<<)BozRTh6uE|2X+oYwxhkz#q2ANTV-to{)p zf$Mb3bU)wURlWMV=ckGzBj!P397ruRifTTf#vAOLL5BHNOIq6n2l(HxM1JPgPY~}l z!L1r7$G7&{o~JTea6a`BwvOm4cb-{wKDh3Z<NQX@o%eFJezM?)rrmkk_GmA(_#pE| zYcclo)>3s3S;I<`$;fS%L)GEcdBe)%&sRSQ0&@vIUglEy!ls)t8|>Rl7@yW!pi6Pd z*gn&u$Hv|av#v+;RD7>i8=d;X7tp<W?_Qqv)b_22%<g^5g=OHy$l26yDFf-Clb8K2 zEPOo1EV66-i%2>TAC<)YX?dWQCM1XuFgDL&?$t8jzj!zG+sIxxSYij8-Ws&y?2?Df zq~EnyLKf#U4;O?B26pO%ZF|5FE;BmZu-LrYCjB=H!cVUYxiDqYL=v?KRnwOv^Te_n zn)7)Ctj}xeYr?scmrZNZb;b6ecRpf%m7K^y&pcZ`#&&Ak{b>Md?(;K;Oi~Be-NUog zo|;uFzY?j0++G>&EH)45D+Y3Ks^%jwLnd%n-9^Z;?mF24fPnf#^N=O+FE~;)kQbv$ zVaWc3d#@y0HW)IsQjpu8NR95p4|j;rq5L%WHG`)>lR&Ne6B0S7Kz_Vs--hdVV8lSk zrmw`sx%a}g9_uCV_|LBmUPc@>?j7HwQ%maE;sw3o{((|S-?l8Mqt}u1NrvNhyK!!y zw63<Er2juZHjwibQQ|^Q9)<+ubcFS3PzJH1lyQt<VxHNeoo0t1TdL#I&j<Z`kw3M3 z+Gkm@FHK2npqc>G$LaCjsX%1KxOtUu)HA+f;Y?37Ft9u4lde(xgdzOXb%-1+5KNkG zraW~p9n4=!4s!Qa(M3B5F3*1l`RUD#8=fmdp$8qvVy$@xf?xRCk#he*N#YS?eE_Mj zI%aBww{RmQ|9QYQW&tWzR7TLo1tcUT-myzYk6I7aFte$jkb-IDt>o}BlLQJux;yCF z&PP;lZ4}|2`uM;uf!pgF+z?5;c(3s-6#y8gTxl8y;o5LF@OHQ{C}%B={^z9y{<T#G zsA7SaSTixk2V1f(_gUnQu8X*X(J6uv^Pw$Rg|TWkC0z2J3>5&FQMt&A;}2l{;Xj$Q zTEVt5f8|w4+@AO;pJx7=JR?C!7F7H0c%*BfpB1!R03@k}vvy{wI}`B51oUxg&!`E2 zQ`{!$jfqtr+;bV}!Nh+k6(INy7)O?z3W3SM0KoD+KK6cY3v7TU!lN=Jq$l(rV+Q1? zkGo`k(pWLgiv|AWve75-BLE&9PRR|LHhx8s#)P@pn#9llzWNPLhDT5}()))Q8s5Jb zPLA{PH41V^<I*bx=<YCcmfa31HdXpb%6viuUH9O6{0_(~!emet@UZ#8$BC5|=m8lX z<AlZ`UnDOEt^f110TSyH#cf@0;lL*V3FJw1dUf6x{Harq?>kpWqwm1wh_m4*0bW{< zy}?SL?+3sP{eFpi3;>A-aLURansXqOM$OcG?S8jSMxO!zj?6&(Y+bUM!c-0_I13WU zyee7*VdTE%lf0Pxsh76@x6d8iuLHo*gTHvTLhh1f5~TrK2pGe^eMgK75dUQfk2Av^ z<}QARd!O~)6l25h0t*!m-rqp7xEI!eXAItoPhD6H*FoRm4YYtp7@)3GkXWr0@f8WY z6p!2U!|c7hJebu5Ug6rMjz{z_Dv1TJ09<fjI>f@BA`3m-yFuXJD%r^Z<UPGgSpN|< z0AvPGoTMX|g-7r=fV&_C9B&S3-<ty+)EeDw_Y;*%AG;WA*Y!W%uU}$Ha*Yre8?x1} ze3*H-U+>M~E(&dG5IZkiWQ)%%-%GXdb*@6H(ej_i2l9CUM1_N<Xh}pNkjDf38{q>I z22NjTX>>4lU_3mZ@VTwjT5B}w+W8%;aG`K9$@^+<5UYa4ud@IET$?F8Yo>!SY#ur< z0LQ2jo8<~ow@Q}X4uXy)DUJ6_K7lzBV8j2r52s|`It`r%1S1_uwmw&Ji)EcqlpWgA zBu3z?E^-+laT7pq4D_QBM;D?4DR6)}^5lmT_*1Hs{DLT`iJ&`3q5d0?akUp+8mnKp zz~KY9_~86E5$6yA@FN`=a9!_L^jU+e!MhA%>cR)C=)gxzUw}pD{g*5cr&FylXW{za zl(&}GEZK7YPQT4r-90@xy6H3b%lgLKIK;Zdkv!CUPLsX5I6rZWZ+9dCI>-$VTpNd( zNh6V@#q8be)z!_fj55p!&HmjfBP~i(={5+JN(29~OI|kTagv%QLJw)tuxE__J#WBY zylmdktV~XI`ORC>@25CHxXLnV!GPhr9f$c0*IFFR-aHym-b0v_?*a`}!JDaCDs#q& zu0p+Y)wz(bHHxFMk~qFeW2yrHxcHpp0b^oRe*@9_j3KwxW;dTip(ku_wXSu9P_h;l zJL3KM>uxIh>w+U<k;~!ssE@~4ubMBF{v)&$+Ac2CvhMmTse|QrUww4H$hs)lxa-9B zea!DNMBmq|K0L`hOm3Pa;tY57xxLS&EB0dArl_eOY`Q)asHphGWub$pfK{I@cE50F z!9yNV_sef9=EN|~YMZ|F%b38k<JdULX|wqAeq6<&drP^+^@&e2Bq@4;;d{N`*S+z$ z%jZLn;c~C-(M9{_k|Mz=mAB2LjAl5qt&WGk8kRNnMeGz2sBG*#ANu1Gma7lvXEc4M zMwj&=nu|PozpuB}Ax8FXCx)qm+V-usyZ41)&BNmO{I=+bzLowtiH-twjTc4h#2~zq zP^B)v!K;D?OeumUq~fn)p_L`oI;b|>GJ??3sNd11+Dd4=nJ3lD_x-!4x(CMGLQUps z$3=vr#`+NguQy~Yr2-9FntIcY<i3{AuDG*Mo6QoPM7i1Nr{r<mve~ye9HiyY1moEK z)T#Bgx>ygcsNI#`b`!myvmJ_vB3UBrZOXCpX=R%8ED?D9j^ht#v8ajN&-v&d5GCoP z4%=aO&JTSNC@-UbSa^d{ps5ySSoLM|{jSS$;oOXdMa5fW`<}Y5%S|#YNpUSvyV8VE zk^BDWo_b{?4inDGkImE5s|!0sbq+|=vvO^><D)WIn}PzbO>x@F3Wd3&$5a-ym#*h_ zZ|ropxK-<j*?Z<ie;>}iqG?sp?zK`4j<q^p&h|Q*B4a&Xa{u6>IbEs09$EK0auM{u zj##vv-FexSIy>U0q!>p=*DPQj4*rNv8b#fDTEc0Xup!fZE2BY}6%+6d_$RoBaZhKA z83$+}GE>)rsQ&Dz0D8l#6?dm!K2l7{Y$av<&RO;(>aqm>{@TBko?iA@QDxg<nO=<d z$0s(N>isx?HMCyd%ZW;xG?5@mL~_o%LrUYrK<do3qzf{=ls#|HXnUP~zeFco8ks<H zYI^hn!p<4yT4$t$UNjI5bm&4RN774&@hL*Ml3$hmt>mpYyZAakXc*&jGAje&l<p>k zOlu_=s%T4@P`FLisQKP?rY}v?d!d^d{;8v1aL#2)fZFM~t`bqU%y$_;%`DkVesK~m z=r#4pGRKYbVf&msdp?+=<dgK}>7~Q|7W$^3{J~W=C*vo-=PV2VCxaY;0tg_XDjx<+ zx^T&VsPYPD(pa1Py;DDPGI1-*Fg-cmSxiw<Qy)087bcuM{YFDT25cn?Q&Z`}tLusZ zgQrrjlhwUZ0=I#r<$C+mH;^C7@0y`;50Lp9iXMe|(3^7B%re%S`TM4N(e~O4YFCz0 zzB#I8<70dp{u<iB(Pj%>z(A4XrosdvR@q-CCAa0~3596ektnaz)vdH6q*Uu0ktXYE zt6@VB-Ik4ttRS6}aP}LtKb#Qqqt>+I#%CuJ7l6vo&c6>I*xI4suOJ1ARL%xVHc$C@ zUG51-4Qm~CD_-sI13!w}2A|(m+FLSfqq*`LtXm(%_XM??H>TU}<Mg^mYPnd^jZNH4 zQRFHZ>DHC)f7Y;0?(i)Aa*(7|*ca}k)o`pgRRjCMX%?CKYsKR<`l4Eub#--NVRB_X zDlV%u?VyLR(8|rKqSbzN^m#YN84VR=PTqZNfY{OETC=H+QLP2Fr9|RLg|9)?KI`b6 z*uS}cE(u9GiM3%XTl7B6G<iw5Gk8#Di&?CJVWQvY1ucD@<yCbZ)DZLkPj2*32tbZk zez!by+!bu5ysRv#E&kon9EwHNJ}oT5s|K30rwa{Uwu*sp+P}@LxA%i=72M>SKCF6K zs5F`W@KHIXdRTMT|LzPaUu<zvQW+dLI)pCG-VWOfe^xNeGoPx_oI***qVnA|z5LGb zf7}eyb<K#2KEx}nGTjbN*$AwI^l#-{JpV3axC3!jO~{dtJ<O@H4)e~UIyyu<>-6Hp z%jTnPc!ccypa}a#l)FRj&6Mp#)BTpuk<#JNTy(vreM+77tk>itEV_4eAeg$p%_-DK zL@@s8ZlW5JdF^((*vxYM)nH^N>ZDh@Nlp!98kam0EFvM^&mEU2YB8pTS)$SMix;5H z@2oP&3#ZlQ3k?s+J=)T_WE<nB3sz>MeBMNPKTOj=<+F3Q**4&r_n0V<1;$UTD;?Pi zdsH<~J)GftmA*Hy8r8KkQ45W?FYq1O;Eke%lEY?rLPPb~+;935UtqF9#qPhhit6?r z>8!V18J(WZRNpwN4ZhpD5x}MtrBZD2WSG64$h9Cm8F-MhOThN@^7{N;+VU!Jt|ltI zY;wu3?J3IQhKlTm^3UUGHruFT`zBbmSHj4)Uc>L3%oB~yb^Q4HnrRlq=Z(1JRV&pK ze8sBc+vlgfYR==5B3aJ<!<cvUQq(|^;st`nKV(V`LN^XH!1k&Ysj$3OZY(T$)kf*M zj}X?F-)CRa+RE!eXDv`mPa3ro?ZRJ&e5e7hbQEE&G+)m)fbL(X1u2Sxq3Dw~F|EVo zI4M0(I_6C|30a<dS5M@Yk|}i(H^!Qhrkm$a;!05JKabpBrI0ip9wUUR7S9yClS;;F zckc(S5IEXyE&CM%pLKPmo4}FnoBU8!=VD5=vn-#(r5#SYn&Py%ZpOPso1XhCvUPpz zQ}Or(y>QrLZgSa>Sh6#7X<gfjQ=Xf9LB`hHG!$>!sBO;UK7WPR36ZnK75JTo^ij>N zQb6NLWoqNPY%;GjZx@@K_~u#8?l>QUrIjd}L6223c)wjA0u4PLt2Q`5av=Pzb+glJ zynu9_^~PPIgfU38wROdtSzg#p!PoeJv^*CJXuQh}Km(|m;HiDq&>7)s)-zl)?fTV? z--X?ITfKfJuvWdMGg*#Go`SzS)?!;`@c;^bebLfQgROdcazD3Mo*(b1BWH6?U)?y^ z>qWXd=r;OnP3k}|%=W8cPgJevu5OGmMth)d#a(SdR(qtBTjNG~_@kBOMRY4Z!xm=Y zGRi<||2zEJlkJG9Jipt~-dT((EGg)|`NzSR7FySqbuAsm!edL5UfK6$m7?e46?{q= zHWv(F@luzj#;)|n`?iuieo=1rh1pmuZYJl#YSA=Sp(Og$A4Iox_$h}}w>z-=L42=< zit;*!kFk|^(*rMT4cDI3LhJV5Q*?rr;I<G!9O|En0}zLF8r)okd9;#bvdOF=1^+># zvXg?|nXeX&46M{!Tm4EuCMJnw8cF<ZbuRui5+#~IY;bCCO>Xgxy$3RJAe}X->DMsr zyhKK7pqW$jM$~e}t0~v3#fh~ws<@dyxkwp)Zm7)0M!$YG2i<Uggx(#@v&P5C>-7qV zWz@Dx!-#u&mHImY=jh@x-7tSdW}Y>J{Q~=>9I?6O-In6mK~-|)f?nGfHYx9_+UBy_ z)?bDBt+m3uCx!`($vnrz-fGR_mq>TbJ{N-pEoc>gmU^=&-uziR-P?sMOgT{-%vjr1 z8aW9|dHZS0Y0Euq-`ucKw5~m<eS|CE$VH7&`5ypPgtIh31c=Sj`m1rW=R4p3xH_g6 zl7(~T=fAdbWq8=Q>G@q~)2B{=V^v!5O3CHY`&Cqe?nD^WCF(GYHFt$N=KY814I4`Q zFe!AY{-djc2bg4~0B;*nNpMRmjb;|Nm9cA+;B+PhTt{E82A)Ypm|`{8A#Y&`0>euR zi*L)XJx7jM8>&a7Euvin3@|E|OsRI!g}h{<TJy}eUk1mLw>aL+E%uS>fdT2fg)i{v zBi-!7aW)ePTudCjhE4m%w?qF&V5Rjn=WJXofIk;~T(Hj?UBZ3QnTN<UU%Q_D?g;=8 zN&aWB2p3=)d=d=c)xmIYV&3%lA)gxDRd~gpb!o?;Lb>HO!|q1e%DpF&UYb;V2U{s} zO3<6o!U;w*xUuB*as5m`oJl`l=kG?}_c#%DA`^CxDCEMN#anG!AFo0vGio!a%%~b3 z#iYyM8i}j@aF*?TxIU0MYr_*o$u@hlv(BSLWn|rc0%7*LzV|AFoE}&SI-d2HRBWyL zB!MO9j}IxIYF!eVDF?TB6D$RNS)BL#<02BF5EXNWpoMhk%!>D}=BzLB`;PM;GcS!d z2mr@xeFS_!f@54`*(cB+>9}@1lXcK}$)LxtBv*hhei@7vV6Ip)`k<e=u|6Og6=lT0 zrkz>OFu+Hl`|6~L)?&-jWE#0AZLWM?=R>dP`P+|lQ#@wHs6z~kG_&4&`yYMg2GlNw z4EEE&%{bY0m@(j*xq?)vO3m^d@+9uFU#FI})iCmp$GYooGvmJ<Q6-_(0xVf0)Bk53 z3jZvPX9bMwUr`af1>N6lGZAb?2B)8KrRp#4r_cQwp$6%Gj?dyk{416QAMt2{5q-PU zKM>qCHOB~P-xmbL_V_1|CF+EWWY$JEvN6ng?=h{y&I2}8L?;F){u+6$u8dGRUh!{P zkubErjN7mmJj*|Qy4}X+d|)h5G60FIW~?Om^cq0kN3%#&9uDY#dG;0FpB&BqM*tfh zqJ%HAzxv@mR@)m72WD-R%u7q{enagfJ)gGb^GW(AzwQNF7Ct!*sv2{r``gEU>bAV@ z&n2C^-?@bv!!|tRXZaZCeE;GUkZ{|CC=u)W9Nn)*-|;L5jC<XmRjN|)d%T2z5cHS+ z2mMc3d(4`5I~0kS4=9JxtdDcIWMX>}tO-X<oyzON)fF*Vq7M_FH0C#&78bT6tdl(% zBhfd$)Z4p$6j{QIlh6Q|-ZwR%$0hvFQvLy`zhC`=1OI~wea%r?u8z@cy?Fcg$&}&s z49h6P>dYKzM!B}i{7~jm*f62Fe1dSRKL*tWW^?m|I#*f}wJ|4WL!7o}pOAo68FOu+ zZNxp3FmtAKbMNCjlEQLta+^1|Q$h1OIn+D0)@St~3yd1aSX^A*(p2v|Gv^YAE7foC zZj+q#djVKUfp_R$mD2tGn%#X<vAp@t)D7g8P;jjs8;44*542yEmAIPWAoQ4Nxz;Sc z2%1X>rPiQKEQRky_|HR9jOZWjZ1d+)JQ&!<td8c<&ig$O0UqwCUptzr$CGs`w)`@! zW%pA>c_kO$$H5#x-pkCxGt<vcTg~QeOr2N008w$qVy&IZ{6H5435C|AVS9y$Ej;zy z%z@gxhYhdliEuXF<v;GqaiwBln>+56fS7<}TE@xyhQ12tOlwNTVy(~573O~0PF=#^ z>o|@4c27jDlzgiBFUMQab5;HH_}(S-4X76D;aNK+H;Hm+Tn*0G=RO+K>>OI&xbS9+ z1nTV8L^VT?Dil-<`R4^ef${~OdjDMsona^|=(zxK;SLUQNJ%(%{0164bv8lAe*Ic0 zpF(t+^>MVOc>Jkd3tn0{#Jpm;dSUruCB?z3-exncuCPa4P3~duLZ|w~iDE+Qngu=Q zwrJ6fS1kEcdD^RHGw9A_T~u4qRKnLUPAgBrD@iANj9!yb>L)Hk8BBc(!OnKxd)30r zfn%r8YgZRn4=+zIpXXfkOq=<8FV_kdx8?}B!HY|k$M&ZKZ)QUW#6{Isk5b>eECSI2 zR63W>-ztiX+GaC&*~he|A%8eBNNf2?^xF<6<m_wCpN%S4AyYHQ$JaMkSd}|wr7dRA zUx!Po$<|we|Hs{Fdj+=NJ>_p7eyJxcHrV0Sz(v;wm}@YQ;rz+r6gq%dw)lL1)c8P8 z*So;+`SAtu39Jb=%;U}yiiFZXPXVoL83VrjR{E|`I&6CdeLdxqH_Nd{_<a?;yoK2j ztgGk!`EH#2h3>~igU8sSO~O96+3k*xnii!n@p{ckbtB1W7cC2E>w_ruLVf7c`=l)3 zQ|XeoQdBBWGi&J=h;5tCDMGf_0h`)V?Hun>XF|o`#YTZlBaU{n*KUonHATYCl{2c4 zV3T?V-jQWXJSl0<%u$QeZ{ZQVf?wuC2aY@z92}Lpzjt<)p<-uq_<oObqw9ws2x^;4 zrxU#;|3?LuHg~7lnB=t@TUC|h9C|wP|0xYP2dK3>e$oS5=rVD{=KVAsuQO;_kXZG$ zz+`e;%X3?1*Qi}sL-q-(%G1vZ7i^LBQX!X$Au6r8$<l>ww$kt^Ywx_ZZ*F6SJ1<{Q z^sU(ZTGdWCc;mROPx&+$KRcT=%U}7=J}XV0`4NrAy}KAg`6fN9v>a9H{!fWdmvl@O zkrz^T6YJb;&_j&Nb`D+do9K3kk(Zj*MT*zV2~%6m)oq*r*@Bl8ycJuSe*1Af7+yB< z;0P&!z6n8|CiI)~vel<jHFOzj+j*AzmW)MiOKE7A>cX+=0c*Q1Ip=pgZH%V34o+Kn z?r%1=u7T%)EI)J$e2JzMo1NZ{IWMpEoWG|KhI}Y!KK{j|h6y<%@>I+I(r%`vrI=T5 zu_*lzj1U?f!yzI2KMrzia9V#;m~1is%qu-mPU^MJ<e!4P4iLHd|MO&qu!r@hT-YJN z#R{OWIoG;H3?<CN?5&R?m**TB!}%6jXCq-y%c=Ezlh1NL#v7$Wll}Ieo36?~RQxeF za~K(Ph<PdEceYl7fmReK&X~`8{E6dZV~O_}X`cDZTpg_P(yJ^HZ|4D`8}O{b;*pv3 zQ@_ubh~oIC51|RCw9fl`)oh7g?;)mZto(jAWa+pY`I3^KR9vG6zZN{#+Gi@E3F0+H zw?6uAn+U%zQepNwKE{Tuh&>~nYe%;^c=7_u=b1i-Pkb>a@|gp#AxqD55PQ((k3##- zm&#)pyX0~-vV-%uRSc-SW)W4y>tqCC>8KG=0Y^r_l;w`@pDW#4@|=&~*__f4qn_7Q zZ9+6DXl0E)gf*xbyWN!~LHcJfDHgeVgOg%@X6GeJmbkmv<wO(5O|&2Tteu4u-$niv zrTEvipe&XpRKM+hCHtZsF6P!gzu&NBHj)=)4(83Gwa~GKc89h8xsJyZ0qhocPu6vV ziQX<0u>+O#d=DpB1XD}uN%;SZ#eBi)uZkA0`nr%l2(*cD0Hh%Geab&x3DG9T;f<U? z_tqcKVeU}~JpL+f=4o{9oSo9RhQ*D0TxpNBy>_UQJ6lp3qiEqeajzP8*h<ZAbI!Bo zmXgBaZ#}<K!_J_zPlX3#$R|P^63ChP-QWQl0&C%m;b#F3wcg#GX4RJ4D<>vyO_3-U za*v(jc5nO1b)L>5NY=a`dw#>O;52NphaqSaqXsjd=k^?@pJUZ0n>%fW>BmKjY_ajR zOw&fKc{k9U=FGh6+Jg}&WWkNc6%|D1g5?oaS}_^>1@MLj?+Ul!g2RnO!z35>4BU7K z!$^G#C&2WLQY)!fQ54=<dV8V=_I5{&aqkxA5}{dl$<f!%r}Mu#d>ojC=6ny%a`She zAVL-6i)V}3{4F4x-sbt##P?f?XF4&)m7l-G?P%`KXx-E3cD*<opK_)7m#1>I_sZe* zc70tFp0)7JQA}FW=~APKjO{nk&AmU+jd>G$O6x0heCE0x?JQfbcE^#MOnr!!naJSU zafaW+-bBI(K8ZFOU;D+>X{J&2W&N*f7fL5tg_(KpgX8=@CcS$VFcRR5X*c@S%Af05 z3hTi>!#MUim-}0Nd*A7oh`TW497}#H012dhs6&kjXReLEJxb;)oKh|2YtT#e&$;(J z_gYmAWb~T5kzgdn#rU*pe=>V$3Ha5rJJ82pKw$f}4VvY9pBPs5C4N-D?QBbw<<s_r zQ~UkE^G>b@!sJ#(1-qMv&5q^IzSHA_y*9~@vA;c*PTm&}naNEUgNY-qu4>f<7*?zw zKYY}-ql?4C4EbDb?B^y28AdLw{r^&__N9pq^0Oz-O*XF0Bjye1@^~~dXs*A(aJ)0s zYxVNlrb!clRApEEzEDpvk>9?OlXqqYT|QyN%55Rs)-`!ZmKJ>A+@-;}&h0+;U@S^1 zR#QQtr@zl#1P|-3gB{$5#>mv|Yu1Cvt$xAkEp4Q`42RXSTtiDTcrrnsuIRcP!vr;v z2Jm6+b!|n0w&@M~2B*&X^@%$6L*G`CUQD$~&}<!}dI+Z#JFyrNMY|cMsIa}rJxNlB zI>+RpcCMTl+q>~|^!pXgiy!ARSF78zPcI^g=i5Bfrv4D()ZO?e3WTXqsK*dHYA&{7 zCNPI+1vpV^hYjDpj(dzV$1qNK^mm@M{}`wD7Vkf-Puz~=uS;`$&WsAWN&rNOPdTK- zxwJm=+%=6HHJ?a`Eh_3uN|_HV>DjuzR*~!Y{(&93o*JZ4maY0UkxcT#8S{fAp+Xv5 zETO`bCf5sK>r51tv7ZWWh!rPGGG3;Bb~DiQjUSRSFA*+M7B3c3B)*`eMr65!RYYp9 zX1!yzVc0pkRKY;)m&;kL{gS5d_QEb`kA6T3p=DZEf{|Z5fTRfrHftT#pD=OriN$$Q z&)^4UGR>AV2fBaogkTzqaxJnDGQ4pSm2*+)E9jDuWK?_TlFNmTH4ZxxPz&Ex!xA$x zy`}&FHCP~seBl_U>pI{2#1n#aW3yEzV*FDmlpYNvLa}VPH^K3A&g=AP-0dc_ck4=_ z1BR0T`Eb~Tiy_WG5>NslK(^X%PhCKN;qAy{=`R#oRN^@4nS{~x8TErlx&snURP^Zz z3K{mX_yXLCI53Pw95Utn9^Rjw)(`<CP{CZ9nCRT@i^0QFS)2ehyO%=-#??YaC$|&( zHW=&WorwIYR&dm}pGhE3wr?Jn+8-|ikf{)12pu+^4zd_z6fbd+z;rbKr5E-(sO(fY zLxycAuFCZzrajkx7oDlP*~UmFU;e>G6yq94Fi(9a|Ibz!bWBctORvBho)_F=c$~cG z`DvLENHHTHo!r8q*5LxvCO3Y8(}jYkG{n?is~QJZJaBrtXXe9!n;JpWlu0s+%S-8u z!RXo$tgzzh6AfanKG{iF9x+A*`Z>R56zor67%Sj)KA`xg-tGsWnNv{o-KEvA_{f-8 z!P<^%y5MKi+0MLewpNKl95P63K(yOI@_xh5opM$Y1(g5a0@ROZF@`nU3rTgs`@?5Y zku%8PsVS9wb|pco9RM~kSn_GSX|X_(7}OS$c@-DEZ9hqMSKE$D<B-U9#7m35Pg;Hw zPBT}ChX)5(y+4*fMk9aoV_v^CX*WolqWZ&2<SVq@_p*7Fe#k!!I3`sJV%YVcj3`#N zB~#(ALIbI5Fo0IgL!Obk8uPyi_EsVJTRelV7}o>cw<X2ATk=?){eNis3ZSUI_wP$D zuyifmDacZSh;)aPbS>RTcgNC=bT^2ANJvR{2qGdVNQa<wNxj$a@Bhv?jx*jp^~5<( zeV)LVb6--E^ZO#rkI=y0g_3>nA1S2J8V9bTm@tEXLL-$@Z6tCRG0J)o)@>w-w)xC| zoCN^0AehVD&;X5Mm|Eth4iSkP7%cy9*A_x?TSEp<GxWEqzIDp6V>dd8lO#ubnb2R` z@f)N`fX74%=}1bDcl@mIoDS_jFD6a|=j(@}49pKc@4?&hTNemvr%7Y7na0486yYuj z{(y$YmL+D^z)?yOtbGLEKpdo3)?16ui#4{e4;HH}i&0Lp^nVGqjX7a~X19i2Hh0Zl z&oSYD@l#TCt#@kPi%q;-nVHevZ!z#v+F}BYb+%HI1*hN%D3hpQguS<Ec=K0L98+v5 zVDmp>=zb0nAZ}>D|80m%KCKRi#>+~N8BJy~je=pKVpp)_DNRDYObR?=a4ir|C~GFq z<nr;x`TW}3bHt9=eDHBhnkY7q15%k3!fjh#J=@|)DyMA3493X`fm8CD$qLHs{6eM< za|5dZez+H1xH$*866G<!(Qi0Nw%#s<1uvDQqY#!zsY$MXbAZ*euLWE)N5B64Htx{q zE$$uaE9gfY%0xL0fJQMe=^+x4!=fGiV18}}I<+(y>;%VnIrDNnqFjt{sTc_Q;G~J@ zX>Iy-_@EL4PO0ArSzk#IWA`6P{CAj<4#_X3zSMTZ$UB^!y(wsbCTS+R*JSpa%>Qcs z@W3<b4IHv+B|;ZE(adBON0|k=549MvYWt0Lq_MX!G4M4Ocaa{LnuS-TiVDK#K%<LT zz-eZ`_r^9`t23EPKp%E4cx0L)OCSe;RU|seB)m3k?TzgZijjA{?847NJQDUUq)0Zw zyhn<m!BJ7LI>V?g$(3*k0)$?J*gw7RW>N5^RAZP3Y{oe2%7$Cl#Z4B37`ac;9GJ@4 zP~!+SNTgVaWJS5Vp{PK%6sVqOoDxocp|ACX#P3J7?TJ@^$XAGz3D$GF9))MmB4S6R z<G3TN!63h|1}TW{q3~0o1D}~z05JDth=fC#MX380H9lK~*6-)Sw4*}PAY6eOTWHEm z#cZe@Tw%)~M?MWRd*ZluB`Wt<qI>E(&j+lKC5U4h0eh3riZ%^7c;H?6_ws{uA@DS5 zTQnInvN&$6?LYGP<CQflv{D34p)b&vYi~53CYzhDK`ix}fJ-Kgq54S5Bq@SV3IOI# zFY4sS1oDPV1v4)~{PXq~6lzeayE1uY2T;&#hog6*V7VJmo6$rx01=FU8*(~$E>r7h zml44tbFm?19<^qDoM_HM4Ey1vpQK)Yk(FGFaA2xFLCGPgjSc0JWrzSJ9d1DEg=>39 z1<r8z{m;?^IVGlA5d{b+7z_R8R3;;&`$n6TB)2x)&E^inKdbW>FE@)cgmd<lN%zIC zx@PQQF$2?cj(9?8#oY%By;Y0*w^RxnlTYxUkx-!S;bYuTkiGVqg)p+P`P281J4CS_ z;$YLT*VPCmm4V5cLXE1^Ke-{P!v|22b@ar%<&4Dw_dqqxs~L@hnHh<2E_!JQ)xW&% z3&%kX3qNad<5&l)pqo~T!;ttG+?mIICD09`vRQt9MdZeZWmB=$%W!KO%7nU-Xu7o0 z{N3z|ORB^<p@9|>EZNySAAG@&U)TTEPB<iahqz5vqpCIHM*kE*(L4&jG(&ugS3W9l z7B(O^-)f+AKZN2#w_IcB29(Bk%eB}kJx%nYQu+yo@M~g=#|b!M89tkf;i5~TFt^8( zzTs1UEZ;&EMBwDpeT8oqD#i#q@M#~HZwjLvMhF^i9K-cNa|+uZTWGy&I(Nspv|ZWa zs3=!2#q!#UiNyYI?uZA40z-M>n5Q#-+$pp@#37nNCfr=g9BxR7q&~+_rGuVD7}$(< z=lS2NlfOhkrsfxH9_(R{XgRsYArz@uP;uxbQ(U2_#gf@WXFn4-<j8-beUHDSAHujI z+LrPd)e@v?)|W5JwQ)>k7-ZCOdD3k(zDjx0HNSY;#fLS8?}K|@TfRB>iOk5tdu0f3 zNzsL9#FTp5JiCMgUCrUOpzl}GlmMxe%r9OcWvRubk`&o|D68XHBj`w?gdD_Yk|dp{ zC2Zjlix@cc8YC!D1XBHc;41)`v=M8Kd-}UJgDL)S?z=JX$3&G?>uK>>#gwgtGerhM z<o7t`+q>mP-DxKxl8B#sai789sNc{cRrfG?)0sw*MwEb->CnPDZu=T1v0O)SLIKn4 zGmGvJrd9B}M-;LUfGb<CFzEo_$9*ltfV?b~L2S}S0}wg6#laqzOi&Zim5m~(y)A}A zsslBY&>7cY=)E-Z)!vSnYy`ZjkKYj-EU(`-SZ1R#`oD)FOJ*dCCQQVMKa9=PEm|bY z$sqC#q&}CE+&U5HTYSN>xsu{JrM=#I&=xwIA&uo=t+(lD4dw2_oiHS&vx$G+egfi> z_Jx@qt=Z@TB0KBW8!R$;%N#sw`Z_Pgt08@B<K+?dBaDh7m|)ec;-<l^I=@AZ<Lb&( z($9~{_Hc3G365GCY^=0+o9rpM(%uQMzJXKKF5fh}fwEVy@3E5d5pC!W7FbReQwn>( zSlGo%h0Q@YB{J|9!JQ*~8BWAv@19zjQCxYn^`PV#0>KN~FvL}bquB{^zRcP_CBxAv z1og8h{)WJNb#5xsirGmg=s+Be3#g$T!qPrb_Tn!!!a4hdG2db^M6UUQ6|6KLIF!Ux zEvK7bjp*~`XaBscW<l`n*@L(n0w`wMmD<-X4#D<*KuM{dx`k-M#?-HUl7aBBmGv}t z8fyPQ67)S+-@p_o{Jn2N#V4M|Y$9pJ=YP6SrgrUyaKOq1gQzlM#`5D&US9XIUx=gy z`&O)ZJ0`5EOO!AfHB%dKQYixO^|tnJi|z7PY+8TwQ|rR(B4$dPeA%dc9}{+0mwbV) zHgg-*P3s<~TO>(A7_aB+FJ2wz*bc;hSTD+=@B{Tx=gh*fmHs%*e5B@g?p71|dSuCS z(lNU@JyOxZ!Oq!TBJrVi^`=S4ppE)eCtV~VRg9;^3iT;dB;m&~opZEjQH0Wl>o<Cx zrlVFkB=+B|`JMTxGT|?Yl;Dc-<4t;L^=wR7%(DU;3T!rQWTAD%`yF#@j4`v91(h{n zVdB2tQ_Zhzxe6~Ys|W99W<dM6AW;9oF^!6WW$Z%ug-rU?<J?+ItiLqVvj)t$Hat%h z1ZjYc$!x+?*Rz!5XO)48(Kcc-A!$aWwC00P@^<nbYr9X=%4n&mCiSd-DE;$iXUGjz zhtI!Z?(V~K%9-2`Rk!c)<W6ma!GUjg{D@es0Tqg^&5c#M%v@RbBbGSG+L?)5{fM)T z`iU6H3xARHXaVO|G6|}W7wmVlYZzT{f|2Zz!!A=(D%etLojG!Zfo!d|Hd#~rz*KK+ zL-+By-G>~DNz(U&b}L(irst+NFT_0ldVMuS8Ku+BORrqCdQJ%2oflMu=h@rE2Yl>d zWmQX%mWcpH9bcj`XMUpOp=)Ru_8arBBcgKgCXy4U!pojg(ved0HXy}7Q7)gFdH)Vc ziQlbPfy=ou@*&z9^Z6^y?e6TgH`cgxo5MNI>`?XBB8Hto?872Og45y#hWEwh`MhbB zRpZLDLOqr&X!^=i!0;&9nN?cEF$*@MY0BJXnbMrFiprq{dmP`XK1HD3;k`r10hcsn zO)mQf)2`qPe>?Zx(cJadq(qG6Du+;tn<_U4jOU);ss{sF_KWmhZo*zXm2X`;Fe&8! zGIw1t_$_>G(H?^#zOioBs)_Ygtb9Zoy2{ifYR00#r{|CUWr?Oduw?2?zQs*)(MQ(Q zZbhifc)wR+p((b{G>0-9Imr#(e|$Vzk%#T-Fs2wpoujVQJv2%-fcBj)(ZBgiOTiG) z=dO1%DzjH|rn@t-tnxSMk8jfYy7eba4tir{YY$`Jt3R&xO`-Po+S*kKJL<&#C^tVv z+BWqkcwCyX|Li5F5d1cUE=7w!)IQvVKW7t>pvuLR@u9XdmT*8Xz1x-bIPKE|7q78S zYJ`<A8JUE=HCBOB^4ra0*u&dMF2-{;v&%e44#$l;c_w!sk#sflm%9&UF%>GfHdc76 zVX5|ECGo<6@9Pd8`&&f>{#dw7cdgTrfs`<fyXt8kZ$wlL=X8n{W==cqt;raG2Ngib z;{1r~1?`ZcEHet^K8F?sk+Xv;nFMw@0a0%#fJa<Ah5PO8_?)?012%Ra4bTS~dnL`> zl&%#ed6L%T>b1<w+Ug_d(~VTowFSg-X$Yf8KT%gvXvxsMAJA{6Wn>vC+&{f{^2{)N zPPBJgJE%^sPXl3;orh;|GO}Hd3_8E~6mv<Rco{TDEmoG8|M))e)rd~Tfsq`I41Pmt zRP&qgj05vL4(~OVw1{*H0Xo8pmj#js?tXvhFYe`nS9$#d5Ah#GX6Juc@;MLJ_!~j| zo{}-LTJ1>$J5jgjt+l}OrojgH*PjS0Hj3NV;`I#WSg$X9o<1*I=^?$IZdr*?(k4N9 zdK_qG`{w06EyV~!BxJ5kZrJAr^AxA)8;<0~8Gq8Vd`_JQG9Bd1gF%2#3bapu`|`Oo zsw8=ED<V9fiR`;omh)p%KP$%|Cu^bR)q~P~LqxH<LYXPGk_LVi7NhUm?M#+K-QR>- zT}HX<<ZQ|Lqie6cydJDn6XNug6VSWiR>}c6u!%Bqo6Y^7dtjYTwB}FR-RIfFqKrN< zV^zIkP)DkJ?MhZ(1<(Fbi{V^7RZ8R~EWW52KOXJriOFI0rNn0xUirXJc*(hWl-hWy zo+!*YKyD!80C5i}r3z^rhH19}$$i&Xqka`fZ5`cMTkpbkpAnthJ6(OjPj@!@WjB9j zgfhCKf#EKqq7<&|@46QjW(upy=HZwK_X|l@4g{#EK%kQWUY?>(3VP7+uy}0ta4VS^ z<;t<8q%h}{I`CHJU8Qk{Ac+<5Ox|4`yh#FE+79Pc`3wY;k`vq%w6!5m7UMFZ&;7TU z@9qw5CR4stFcSrPo=5Xh_P#Eo?6nV3?j~a6L=!P!M@+6TogP8D1Irr@6^BRpO+Io! zLt=|*${u^VC&?nwe>Ou`zoyc{|F$yiH=+z>{rzjIGrTv}5y->^H3&s4@S_(@=>knc ziX|ZYPx`{GG(@gH7%Wo|dCv|g@7z6?#+%tgW0Y(U<hSrJ0vMpU2a7REI17KGzUIpj zAH(J5<Og;$9#2q#nqZAV48`Ldp*jt?C>?7YM-8-BK}y169@^dFUgcujh~(&;kCzn+ ztHJ*8<I@_eq6Zf+c6;q!uaQN;){)b%Gf@)Kg_tROwJgG+t1Ja6X(&z>+~1LwDTA`B z3%4J#W{J;&g<i}+7)@x^7!{R|$XM4Lkfb}4l)ZG*!96JS?^bYc(IW;s^kV!{OTd7$ zB7xfH<J9<kmq);2zYDL8!Q*p3ve*%3M9v~9u<d7xs;JPB2I)<p10`3E%Y##Q<HtuV z8qi1NF&jdUg$M79M7=|oYEr`PlaG+ovC=nX7hV=;MZ0ov_+1dijP_kooG8hRR^d4X z-`VI@haP2t%n|JEBV=LcuN=iXuW(I!&@rIvj|rbnqF*d>XZAW!#B9vYKuO@?qkI(2 zI-TQc6+RJDvyLquR2GXs3J|8~kq*BYPEJQr{gR_lo34uJUQ?*99d0g(b?|TpJ~e|p z2SV*GBUEFBC?b8~-v~+`-s3($Q~`!=X9k*%O3_F*lYC9K0#laIiLN$n-c4fQiN=Gb zE6ve(Q+C&^g<E4jQ9}CN0=pD1jRcx6hLBbVySLSk&nLdD!ZZP;iWKrYgfT5_0~)TW z<2`xV@E54`*#5KiCmS|E6)RTN56Cn(B@a`^%Lz!?lSJb}uEN!-1Ij8J;1P)}U`9Iz z4A->$7#);58$HK4o%?175RBVA0t`a8RBco=;}Gexinj%KSr}u4`Od*V%ko2^2oUt; zKzKu>r`*Z(j7Zhe?xCtxnbq6w>*9f?DAmy=M{R!Btm(@%V|LO}9Pl;g+66_p3toCM z5IIsVoR<HR-zF?Ce~O-O$k$|&V3zF~#3I1Aem_pD<&W9Od|LGRmDM*8PqE()6P_p_ zG;sV>G0sGz>K7%uEugYt1tha+CPdQBqCewOj2xWb--`K+ZcX=b(k|^rd|4N$#2bJe zTx66M#c>f_1!zT5!KpR2Q5AtuVoC7bLs_!)R@+0MA7{W5b%9gMk|WVlF!QUu7U({! zY8%7HZ9d>GwO<NtVD~dngPO~|5Vi|FxCi6~LQ}=q(6Z?YjrEL(`>%$C6`n7La^UYW z7YYj*r(+G3&;F+59q7}L$ttlL1?YlF;z+CkYI#p-WlXqQlId_LH%8^b`&1!jN8D-) zM<f)%#2IVhiAD01I5inl_S%2K*CMGLx*b*yxwmgcg$Et>jPAXl+gtGT>Rmvy4mptl zK7axp4!TiN7Ff15ZV8a!@|C2)<(CO1eWg^cg2=9x&9k|w{^th#hxYXk>Bp33PwrVY ziX9Sd!wTD-QQ((pl{FBc@9q2h``ItaF_G#`uF|rW9s2}9yobutCQ;2*Jo)*=$cE8a zQ1hgY2D^m|h4ML&2>*%#Pkv71X`0ld8fDK~a^>;V80Bzb8BBOV21<VOfSqD8m39NQ zPb4oNf7w^M#ML5Da1D;`HJdIQ@UFVJxHYobQF{#mAxTDb2)g9Gzda_hVfv7phGtI5 z!qv`f16EMtnJbH)G>rZ8iyMgN@t<?jL;Ss867z%wqFC4jO+*sSYQHL#w1T<97D&!) zeHn$jk@<R1Y;qLUw8jG>@`TtwA||%JUzXa3Z(!b?g_DP%#9dI%{D)AeTC*F805Jz| zWgSX_IrAchrXQ*T&8fzjwAfuvN|*9dua`h*x;H}pQ~cn5BKchT-eZ^Y6pj#^!KJxZ z7_0InlrD9b^hZ@y>?rSRX5zSZF7bXT>RJ5n8BO}J61wJR^IbLUxu_CDpgXpa$dhWk zm(2U$+f0O-X2##gu$+X2t!!fSyF#(>QfMjIh0uH%crFFoFSb4~vTjNbh<HG9*%Iqi zqAEyR)o;4rwk#nRaA2S&aWC9&Kv8vSdzP6n_?^WTV2%`g&ic+=jHv(H$M=nYqlQ$} zxofVyV(r(CE<kd<qX4yTeshF~;R3|6F}5g0KXJ@Z_y-~F@u*^P2)QB32lK<9e8V^N zyRTV5ZOu+-O!Q*<v@Mi)UxCmG6O|zZqxZ9z;sfo;6+84GC+&bP6Q_25HvrK&{CQ^P zeI3eejv#=y<YXwKwE(=haIsILqx^R9f>2xho+iBf&AYW%WU~W|_m&(M&dJH#WMXVn z3#ys}j=-B0xAhn?%$<$QR2mj%Nks!ReXXtC8J#tJWa$VQBQtI$BQGz$9_kBIR<(}| zi9sZwR$~9nFxKCD<VWGoj3Fv;oh#zgGX<yFN%s|E!E9Q<>BS3p@J(}~KDn2pOGsCk zvdRNHwT=^k5Z)l@+E&dxM<&Im@ft50P(I2y6KRjH*l`YWENm%%^Zb7o^S<IFTWL1r zDj+=zeVDInbkAIW6#Jvw@<GW+BDE6jH@dcPTlYv+NyQo_t{SZgS4wsy-`lA(v{LMW zjiG6xPXVgAyf#c@=3B+F{=>}Kh1{*|LkzWVEtD&!v&YR~_}{!;hgp4Xu#d~FebG&D z=o1ou@V2KE`r@r-Kvy?uwWb{#za-q;Gqyfe{zLVZb!D{ojgp%<$Zd#PcwOI`>hZl- zdBkN=7HYA@J#MHSub?0Zt;ed_=88!T7s@EA0A+<me6hE%rY1haa{T{X;FBGg5X`@y zg(OqIBj90fE-~2a@~#9EY{`MiJN|5Pqt!h2bhVF*h>&D}{e8mnuF~alqM3<?x5Tr* zR&;sW6A3TYG*w{y&bJ?9_(MFT_Ubp^)=>6>-VLS0;pm7~`~3}awzr+xu<vHRaa_@d z>#7XC7R_m#nlj0es}t||l7tfHRH|p8s~IIeYW@OGXdX)&6nVmiZIow4ZWQgeelA9{ zOO~%DW37iBfEtbris_-5N-a~f(nH_%a7zkA!O>>W@bFL>%s@soRzAX(zV_(;JDy#K z`EF9E(I@@|@yzaf5ISg15I~wg35Xqo?mzh<$zG+akv<N5tW7Ek5Y<c!{H-b?LOgA1 z&--9tFvcuji8=C|G$3#Oyx`=->X4q1RK;vp?^1c)RgB<cf6J=T6HD4K3ZgT%vfsSQ zkIbqKT`A(3UrP);Vn#)%0tw!Xq#^<t<f?TXrdRJR_6;p&9&G-&XLdKvbbGT*34gvR z*sVrWv)oj~)*Ljv%rziD+CeX&1cWo2yG(tST9GZdj1SMoM}KDpjab+A$tS7S33UW( z7+h!A%SOnB9GrL8*mRSHj#;#lE*=bxkx78ii&p`1d?>?X!tgX^yrk=(WYN!CcJ$3& zx}j^|xtYnnXE1r_hRi9BxJFuvCNa}Y49(n=)}xQ`si`agmoDD77eT5KoEHXPf!5OX z&8PnH6)`NMh(>zy?R+soJ}NVD@Cs8NdrsSS#bz+@f|QX2^mH9l25VSQ1*(cbFRH*l zTHTvT`l{>PDv%7f5;I<UF{W&KnlIQQ@4G()1{jW((A<4oNT$JE038!?IJe)F3;3R% z+y!qCh>>G~rtIq{pVXb`G<7JM<UR#3uSo30v#~EJchJqrE5rCYf4uJe0bx8ZaMZhJ zs#oSzBBdn9&dm5m;E#r@f>lLv#Cp-R^ZQq5xgu<@MC%0)PT%3hSx;_mSKeD)p#IQ) zuVdNN310an2>IcmD>!hGV)`cF<%W2<OyA-GAOIM!Jv){VfPE8~$DLRW()B|Pc=8K$ z@wH;Yj;iX2kc41Q>wiXyUG7+#;Zt!Qv}~pE=TjOt?%&SK^)yHVs{anTeiFVZ%}FlC zZZO@yGL(cRHH|gTG_c)H?muU;_6%-GO1X^JYoN`DF;_4+#nGx%e(-nva-;J~#DKco z_d$7`_3$<C1=46P-!@SKwmugxo|M!CUYxf0v&U?Id%iCF1PXG+y3VM^T{hqxUUkqN z{tEC?-77q3L3J%Sri*?fxO}tlCG6gD^=_vD;0V2<2TMDfuqFX3uTk_7k_i{EC@nk5 z5;B7IApDB<NkK249BpTTnIT=8)eUu33Be>&6U*Pg=OFYfgM7#Wh%wrse0NV*KESYh z9uhtFdZXb*21;E*UVw6<E)nfdnfd9+hUXx4TMQQ552SNt`9s*~o)g6vtq#2rHMFC| zPolRi`I${~7yl+!Z%QtJr;?0&qqn7++cYMM0LjE|r6xTpn>e^u{kD|<MU6(c(61mC zqK|JzTMH^EJ~v17${BQV#Zp+TQ7|35aXC5IhMSman<P^hd`;rUVv~)|ycvp_wf-_p z?|+HY*U%E}C~|80#*0TGfFn)H@MT)P@@<seRn)f+Pdg2qF%Vq~M1NfUKmUs5tmR+r z#C#L9C6M5J{|VqAA_Nw1D)jyphc`uO>|KO7>~BJoF=hObr~%vKN9{?j!9Os}ROwAx zyFKVS`VPWLA_okVD=LxWaim8BS_NS7B35`g4G86_VNZ~PQlvMxijwweg=X#Jx%#_n z@t^w-yAoTj?9JZZT^MjFawYCOS5%jS^-hDj*{3qeFhMDp)<I>(+S_7wTS_&ZySoJz zPPr;gmfu{DZ4o(e02q!o7twPw$i3uM3J1Y-eINYro!|80Ff5TM17RSJ>qtD?HEYd& zh?>Xoa+hq8Yt+Ur(mEAyqF7+koqBI6w8w~0Yus~H#B}p0yF`!>04V9GIu<Q_L{yHy zkmX|VtzDk)HXi*l2IF@>Z;}ZlF<r^SXhQ4y7Bcn(D>lCUz&*(M9AJ?z7)r!swC^Ux zU(kbUVxeocW7m7wvS>0~E+`Hi03(QL7SesDqhR4EVs|XGk*O-rI??{S<vUlA%?QwT zM6IlLU-?-^93>PAg7H-W`dXFRgSWhK5^GRko}xSeExT4aB>OH)6xq*M#1i|Yli81m zK1m?W$`1i(fuU7NbSd(u&+QFlL_f=f1%s7sml6sq6QtHWk*vLvV61Oc0`$tg%q*6! zj!I?(+1kraJ{vkxvx3qgtQdgic?e*hx1OqHu`$Ilkv|nf7w1yrjPoYTik21+OAd<# ziV@~d)2IF@<KAcrv92DzwkA3d&_mGZM1fG}FbEjhyjHL<!02~VeFR_>GxNg)Ozo<` zVbuW&^Qft=5If~kunXbnUL}W(6?4*toE+=12nGH%5$Yl)fPsm0dXm+LQ2Ne2X3=}7 zW~7Ne>Vu-u`~ff~Fv0PXF#=tR5-5$OpvfvoyLalg$f~1|8)q-7B?M(;nl?%@3zsia z!K>yJH)~rq?RTr({4HM=Rk<ixtfynuNDPpU#}}3UmSeRymw=R@)8k_N?G^XH|EeQ! zc&M4=pNii>iZ1!_e1hryK@(okjG*k~@shHrbVrgPnKqG@P-JuSx<y*2q9unhdX`&~ zkRyDCJx4aag2S&VyVMTkI+jg!mQn-@kbMbL(|upT!E>OnaIVNgm>l(Qntlygc;@w? zT~tT$Tac`@fm+ugs`F><9DpZn4OQiy;#_KNr94Se&TKEdq;zlwC*92>*_WF|CA_<u z#jgcHGvIt-ktu{39@L0{1Cx4iXxsOAgU8)V$))3cl11E7GY|QZoQ5wUZJe0qcMaTV z`+(|qk6|!!wza#{o`wP6JyyYCQ`DCp^ux{~Rn_EE%Ts-vs&9a*CZUkZQqt29uWrH$ zFb=v8r=^{2ih(itZ9LMc56a`i{tu}1m*-v)mjtK`KYtqcy_xyCh+zX%E^|&8$9bKl z#@Xi<7TE*fNj>48OjdhUtXJ(RGxy1SzSgl1r9qI5IFx~-$`ORqOp--Sp@agbXiyop z_yMctO1wXIVDB<?5LY(a4qsxDNS=Zx|FW%y1=BMsdK(iZ#f&90iSuI~yFCy=*HdYk zn<R)=V@N&`SwkLSA&|<*!SpmbcBH>_v}z000%pViT|a`4bT^9yr4AkNphB)WwOWL$ zhAd1pp)Wq10!yug?7My;dBk7>DgZ&rElf__2vjpzl6N963l*g0WDa|d`gjzCSoZ%K zkB&P}w+ndC%4G)LnH>AF1R+9;^XoC5bm9)I5}^Pd#L$1#h<@)0nkvZ(fp^QRcc0Mo zZOsQCA{kH^Za^qF!2(2^4U9zKtvzJwi+?Z3M6}vM<@SmyJPYX?)>2nPxT>~>;3X17 z77t1>$H|jCTy>M}MMNXKj*sSKz~3Bzx2;0i;^%X^ZhszqSHi=*8RFxV?4_yDAO;Ya z0LZDS)o*AziRRo56y+2^?x{4dMS9|?Je^Chp^LG8k;vwz6y^e_FXyJ;r@+Wt9&6%! z8jk*gl<YO6^eO3?Jy)|3u=xyK%Fy|7rjXTD4sNUl_;R_<6<%$Q<^S{$PLtdJCJ@M_ zMwb-#QqJ8JK=$p5e)f#C9iyeM9d9tteX{v?M=6dEc`SsRK*7it^6|`84I6saR5$D3 zL+m@A9%c<4V%<8{7a_;~l|lmr{IHSm_Z6dvrd3O(OAs?N`Zk;p>(LbZxAm<k*F1}2 zv?b6J7z}DsMlp{_!Ag5k@kEYdDTR<muAX{}zmYoEwgl$?g%%kGhfG%-ykH5Ok0w@O zU{fnRc@;r1!ej+F6wL@lvaopn@9K%EY~|$(F`~rC+m8CMlI;RO-w@J<4c2=BCrPr) z{jp+KoZ>yHAWI=%26>|eRBm(Y`U#?r==e5F3fM~HMg@w|QyYmz_G0ICf3s^v%Aln| z<Q4(MMTtSg{4d4&EurwzagAxmkKw5_&PMtWADn-H&A}!5i57mxA9_^J{c#wCGU)e2 zk=rXNwJ%mJUgtUPhuy;;2<1NZ^vpSVWhZ=Xme6F`@XW-qU?M(wq@joo*zDh_X^?lO zTSF-ysN#_bsboK>QCBmejfIy^YFRgY14Ow_9@V&)sQ6@~j5dROb3(9vSSp5hkVAaC zj==zI|4!FiU}g+qqH6kPVp#yq<bw}w9m}#%;Y)g=0Fx{*q!G6qW*KW4E(he;b+keH z>-{Y`P`@PoL5(kv>>rea-}cqxM?KSMhI5`!g<7ON#Qs<I=Ca@Cit8a<&jV7h|Jf;p zoDOO*@06=<UIbIlOI~yV%D0%)q=}g3&ozN;uxn!8qSJst)XhW_h>p*&ZF?^>qpGRH z_@kOMQVfI>^P6XmcW#=<?AyyC18<v4e2taRVU3b!Uq4ACOW+lhl=D-}ZAF;!eY)uy zy}b0MOeBBb!tqv~Fo^{0&t=yv+mFH_fC{unY$cGtiD0RI7osJ^3T`hvxa17o77s}h z<yuwxU}*<}B>SOD&^$!gQ|;l-b%Ph1WB}G^*zRbV9&_wWeO{9;RIupnyAPngXACvh zXCm4n;Y%NK5j2=n#!F){*>sI(qQ}1`lSB1Dk$hq{Pip<1Y)<gmg?es%V#Y4UVUGs7 zq?^>>gI!$ppnpj_sLd%_+orX;(Gq-q3~{>3S$CnWqwat)4wZ3dg-=ph9E6_b3d~`$ zVixa2MW!<@^pI=DL)KzogFv*#6+%Fjo`xMW#F4TfpfFiZ8m$&{?LG<}`93LE`h(`| z0_t1PNb6~&bIhM~QB^QwgaGq{N-ocqsUwp#MXGCwp=U$8Df7Qnsb3GmClQT}`LXJ< zma~LbOsKtyQiSDWP8-necenwJkt_yKtiRUT&X2@k3V8CQk*93v4EO9=>aaK6=yfU2 zcsVuAOd!H>oH*ZSvG7do>}si!6-ebOImy*cNC^huxi*RNkK|_Km<F5xK$R}8{J&;i zrS{$jy*;(#B`Fd#C-xdaZy5OWoXb>)Ktil->!o78uJiygx;6=yo`sIZy_gDNK2>>C z_o^tuP!a=zG)JW;iOe3e6398Qv<60UM5bBdV(kRc9ivft2jNg#{_m8?3GQ|w9&+5_ zQgd|hSZ*d+^%t7Jb9uT{HGysCN(gwI`ozIoyOYhD5Gy}x9IS@oEG`@*1pUKG%k%{^ zVicynGNIYh?dQT1!pxyf-123N;zVc*7=Tu=q$od3jV(Q~G~3O4h9uUGg}>Ml{>yBp zk#z!><z6B3s)**FH(zeUu4GA@;W0xUMEB@L!ebn4aiSr3etGpfeh|VZNq{oZCNhOP z9>E8Ikei5R(YPO&Wl^J>-|Raw6=fsB#0!%dX)6tIX(!fA%KJ<qK2+CN&T0^;_#J!@ z^Y5MhAx8#|Zr~K^?@@c>sM#861)bNfU>gDS1hzM`c9fmQL{b_c>pQiA?K*tL(3x{3 z!#eIoT>dpgeUxPVc(>l(L5wbTXjK>$dPzC1NHP4^tfNwqQ6CM=cgSyjCGtX+xONEs z5K=G5%58|S+Y6V+!VJo?V<C)3#{*}uVjY1xuZx(!`=(~|h(3A9LW+YNofMxW>rx8B zSxb`~MRFX2FLN(RoZne5=`JxC<_(qF0kOYF%)fUZ>%`-%y|`X-2T=m!xMlst$m2$& z^Jks-mpI~HWTC)7ae~SV@D5tcrZCB<$sojKQk%ydY$-Q<UGlkAj`cZGW5VBtLQ#9Q z%@}-e{;B5eC}4N@5g`i0%K$-SzZ}OBMgmdvu^kinnn?PrllATy4m%7<De^u{K;|6- z;v>JCAAMvB3-&v^l${NY6|7N6-;$k&eIF$B(abtH+Pz25a(qqhi`CQn7hW{#+&sbs z(LZ@oP@n_ojE^C1xi)Gx&nBq&pDJM@`9|Xg*)3vaUq*ajK-52TMjxY`#00bHYQXo5 z2}P8wAUW8$<j59P9*w{Mb^YU%XF5YMxv)@j@F;frbQ1RtC~AL1xxT;UCj@%EQdk(_ ztE@`nR!wg50?5zLL0REcNgdK+Q5GwTIgZW7$E8E@Q2nzGHeeRY&iUfiq-`8h&Psf& z&jMh>d#&IZ^;{{>$Znlyet?_6cKY6eN(5NssJwHvLYEc^-7n^F=h%#wMU{U)W!G5# zmf;CR#sfNK;QS&?{U}g%bp!IbKlJwE*5E6c69&tqe&eJtBgjLPJZQnF&q$rklF^cn z;Kp!&U)qCV7>F6_1=tLZujxD8X}yGabi$BhT;DMsg`=cPFL)_!S{^snJF>{xa>s=; zCY0KN%&tZrUhb);O-jkc8rt=17&?%{?cKR-4L4Lwqxy8Cr^8B)i-{UknjMNG(Ywi~ zvmy<4yFiiDd+j>M2;0v=L0}@}(L)*LRzhF?;)u*kY%sL)dqDNI?aiR6))yp=N9una zrJUp*G3SUN260I3nqi6LI-10R8OaJ{Wc<@%20;<T8GUx!A_7t`&iLn3*&x{BbR>}? zFuZJjk-m04($Q`lG#-eBxD+tS{-M<og9#8!A}4?CPCroF+xvz7PTH7o9ftJ4BoPk2 z6_iQ?gG5fPDovv!Wn%OZ8aN2~JHCAI)fLgFsyR6rAWer04+7sOB4gy^M#VjphBWBC zx5PkJXa9Fy^3kc`{WoA1pm0LHBiKiTD47u59-jLYWEy@wjS@~|D<AnAnG9cXBC;!y zNFAUZZ2+Dr;EpLE<3WN|!daUiqfnAYmO=)_b>Yg0?y1DN*R^v4&B}?HRCLN1F1hV+ zjidRB+Av09rsijG+Lt=9^5Lnjq2?mfh;0$(B$AwE{^4GJ%&AJHG=>vwDb;EmSu7L- zfB$sB0e7flvAuU!O90la;Om_YFnk#&R$c=bgssNGj10nl)R7)Ya&co}qas%NT1?_y z^mPf_;!3W@<uXaKif6;P&)A^pEK+HauA`rtFe7Fs;mCv)vZS>L<=(iI((%&XcnfJl z-xv;z%9Y=)nQO4=Uic4R5)I~EMT%?L@MN68!1B3p?65jS-8T&p;YVk}Qt=T$b~0*l z3j!UGKGvDvCZS;Gh{OW473h_S;jqd4KGnGmc|<z892IBTZ0UXn3vPskBr(R(WT;zb zR{b1w0*;B?vWqKF)k|dw+pz?p6Qys%@<U1TJ3qPxhl;U)Hf?s+Og9SC@Ef*zLKj^j zH;ans-@8%XSnfxKGwa8$VsvcSaRRw`BwrDIa3|O2U<r$$30Y($#|cwWXqP!+bf_O2 zDFo*p;YTL$*-#d~ic4E9i}gkSdW7nZqt{gG`Z;9FHLJdt2?So)<^(13>emn`LdSm{ zK;8F3^W{@2Y}tM^LGW4tFB!q-v%Ne?M3~b~tCOQhD>>rG4;)8n9FkLe{b*&BndW%h zmc<<YT1_N-X&f;Y(wih{-2E?!;b<`OI-dNzt)PF%rU<$>*R6&9L;!d*ye`A0BlZua z>2!y&JxZf_4Pr|p;oxc$!nUC3e6y#CO!nild1~H~DGY3~VOgwFX|jZ9r)^~nxfYzc zv@{1rU8AItkt|1JDn4PC1e1{=HlClqHublgl@R;~;f*`D7#-U*t$}gUp;lMTU(^Bp zjXw$d>slx&lAxGo0Ye)w>G7RF{DMm>E_fNFT4|<OOJ&G{aqs!OfdZ)o&bSggWF_Ul z4D~`QM~4o!xNsI0Sw@ujDU|n7=6!$r+YO#By@@rq^8~qjf%PlV<#7vGA%NQc^II;} zT+PI)B+C2n($P5Lq>+J$V>)g}L_Nh}m^n<1PjHCAN)BS5%cYdYeVWFhoCd@eUK1px z3qIdG(<Y7IxFR(sG$0;~<X~aof{sSitGeRql1=}{Aq3&_5ciIME2qH&tj*bQB!?L` zNOhvUk9!>3#c~3{TNN2=!bEPsaywmy*^@}PC*E;KPlBHJ>L%{EG#14s29?H1Inz+{ zm{R>X8?3wg{mcVZnYovm>llR;QIVBW^1La$u7CV#s&ICK%~8PM>?!@W13eMNOUshY z^AQ(AR8(M;UQkVmg(+56c7bT15jw$_kXp({Mk7s3YF{i8fz4RXY?3LfL~afCg*Vl- zf{^Dj@>T7VPHB+J%yFv;q!Sz{iHCE?M>xtVHd;`mWWgqc$)~4k0~KR$if2LH<^iX5 z<1u+BB!uDKqgA8L)jE5}Xy?z+6C4YpGByL%HLRZlSz(W?d>UsHqx2<wkWd~R_x-3y z^&>&K*E(R?gMkjKoGz(ccuL7$#0I*O7P?X@i<}lI_zx+y{MppSfti;Eq$~z0ESlsP zv{ItHDLpe6T;nf|MJ}li?L+0UP&5c#)xh9BKq&n#FO$66n|O(K)aYy$RsyCeGK{9C z5XV`tZMdw9yx4%2!XDcf5U6Qe%V}8~Xj;o<lIiG@DMf!fIToUn`^=(5!lIOVRy-$B z!EkRSXrgVUl!0zQYGn`^e%DLW-|Xrkwaoj|gPaOPtCi~YMUxdQ$;AWr9b2*Wwe1rz zc1#ZDHfC*YHF)<~M$kudjcGuKTuGlsRU6Z=!*XRj$CpCD*6wk298IO8wM=7ibzB(w zwwTm3)AaYrY>IJUHkk27t%!*psR13`MwS&e=4Dtr#GJ!|CLSbjg+>+IQ1P-9K#GIJ zrJbFcz}<3j<8F0{J=bRj^)f}OwPxnfV9G|^tl-29x10d!)Jo_lWnvbk6zBkJKz&Dj zO>M1xZEa0QZB1^eeQjk;Zf)(rL$*vGS4R>goB={S;+D^V9o~DqmMo%_f}TWV@HDlS zvxx(L+e$?#>^X$-=sxUeT#dVm)bgu-+>a=bQ0ac(JqHwQAkyLuMNYC=KN(prUH11~ z=00f%o<emxsTwh<(i|z6NS{Gp3qzX2+?La1YbzaOn51Tf)O2)A7E<f&6&HkP_c|&a zWh^Yt=tXnA=}?i4eJZlBet6(+IF<xLAM>H`!+rM;SRB)%IF?6gT#w?IIOFQ<j}%e1 z$zay2KhC4hBz2<hRc8J$061fRN?`N!Y4OfcC(KnfvdhwQ?T=>8dIP2BUMgv^Dk=I! zXv-Y$lgg#D$Z4wTPRp$h;$e;1<yMwS*@JnL_4^KK0TSrakda2lOFDC5n!qB~0f%1I z@?8K3X;hMxdI`?mR9-E!rWFTuj6*x6pX;B3At|Qs;y&t{5dB9Ra%APSYaaxN)Ns1v z-)K+fLnk@y71QgCV3a8<@(NNUND)-}%p#Y@qDCn2_ki|d8(Vg>sN2(Io>VT=N>10> zer7qkB&e8>(QY8z3&U$9cI9;D0Nj~6*pv+N>J4wcqRx;5X|0%X0I%S44GjdWvL>#a zi#bdtA`he;c3LlXI7q}n@NY<dF(f~GdG@ldY%jMGB)(n5B3^E3^*BN#y|R{zy9Wb^ zFzpL(5g8+$zH?||_-j=1kGTrJsK=<bsdMud8}_`zn*g5`fHauBqeJGEk1F}CQ4AK> z$gY^i5WRkYjryqvf}*CaqcblkrxP9MucITRLpf;_dQJ!+h*v95)G)jT0v}k>39!KO zBkYAAIvS{w+#azN_?5OmE*-!|jHRge5+hN&gKoeTWIdV@IF$O<Oz`Rzlo2{|RH6=1 zfFajlPtHQh9$plh&j_k1EGdq|KYOXorHbbY6f@H9jXeX2w;y3J95FUp8%J2@@~>6_ zOZgY%tFb;R@aq9ARndonkE9~~vJH%kW0Ga#k=upNnKQt3<3w`drCd1756D)}=6c0h zeWMdqP4g$xyGYIz-sfa!mXU*s8%xBcQcnC#@Hu!00JL8u3q*;LT6XUbz+d{O0W7<| zncB6{aF%0$?s?FIZlOB=9rt(Ubs?sm4_HEl_kAhiG)m{EVfDo*KJNbpfhP?$FPdny zaYr@ONJDbzz)>O?t;8PO!j=Uv|CP=FKqw(U&ys;%TmA=cDsUPgG+{?V-rszFW;H8g z!kgs3aWqC}E^4Kz?-vAf4m$xg5Pw|KD$-W;24}yJk91idSB(H1h?IX@?imia9>*O* z?sGg;05<4bzp52R(t=ACj{*_^060|bZlsqCbzNUQI93(!NdbV2E4|sH=9h>_OR$`= zcn|cwhV~D>Y;T!yF~Bj?UJthO2ki>N>`H20W6lrep=3I>sZn3Cz@c#PK6iw2)^dHg z6#v&-4)_c&bbi!q^c7pI37a1r2E>^i&}u0I$I(Xr55P5n5*l-)Y8z6r>=iiwUnLVb z2+p|Hut+WL?|Q#OjQ_r}>iwqgswoK>U^4nu^1tQuHJwePb~*m<zmbl6k3WFRC2k>W zA|b~FQAq(D)ZzBXR$rXo^%btd|E+~T!oFioIsj$G*Y`V@T@d&UVx^4Ps;2!d@Jh_> zH?&zC0w9cW_;4VMCF4AgU8VQYz;8~eBW%g+T}hWrWWWx0z^bT5AIzhr0sKjqhfuRP z_yxSE&-Jvl0K@m^zut4e2l4H>w-~+)TCoLYFLB=fqrs#ebZZ}b*f>-ef3JrD2;CoO zzD!y3{ck18w7K-)C)q{#6xQ`ihM|`{QT5XQ&GZ)m5h&;Ye^gJJL*n&%=D*{+2C>=6 z^EYvhhAj)&i_bw?rT4p*&{Axob{&uMc!{C}fpx(^wjLs5_T{^F;Fa-~jC2UDCm03k z-87V&2f66+AbZpzp-B%^6#NieJY4V95Tps*&sRUxf#SP=BMy6ZVAw~&i0PL<sPa)4 z|2qcINgm!pN&e#{xExN4bwjlOhJ0B_fj#Tm{sx?fJ@yD+p8^B}0J7QjvwkL$ZNh!m zm+Z?Q&7eqv9TNx6X1Dsg+=7*I^fC9t|82#CZn!1#ogUe0;(iDD;Vsy#6g>X%nLEbq zCF8gG)+IR)o!2C9@>zv1&S%)a?hy)s;VEF!nLWOdd(kAj?eUq(=W-6uu&$`aC6!*B z>ziFdt6+m`ntd(U#p~D9Awf4)E)Dw4y>{1L`-ye;Zz*BnS-UI1*}tg~NfySwxO&FL zuX(~L2!3hN`~`E8vJv5dNas#@+3TmCeMN>Z1J+7-Ltah%X2oQd-u*M4`;3cx`CCc9 zk**yS;~(<?+3EP?&VoA5oRa}r(ghB6^nbw}tYY?UzhLXN@mcH)ka$?9Ei$$6V0ra$ zQEBw*AXHQ>s;wdNAI|`ja+5-B8@^%qmP4Y+GFTuAk}B>-)T8W-h4^rt?7E(h8`_6| z?rwPsp|i@4`G0;k82MN#BsEm0#W~OX6SUBF3~L_Y4|=5A-2>}pAWu0pr<CVTZ|{{N z$aAlwr;0ut&17Bj7$zl~Z5(`ZS#qe{rGAQn{8xJSzs1Y!iZ*>+F==cghz|(^OLZgm z$|kbww!iD#H3$i}1zr$5SWQjI{}R>C_eL&jG_4L#Gx^cPE=B#{?BOL&wAS1+7GHCj z*!-uZ=|@8p;ZpZ9M`_2mBk#OaPB#6;TdFS(pZrmcKmQa{yWitre>;)>%4fJ=IIv?T zNyo`ixW0OZ<lj!ju$Z4ttgDkA{2u9i!K8+C0Wn%e@wW-`Hr`O|4|0F{=H#6fWP9pN zX?ZttV&ijDME9^vS>)3(fQ`xKf9k&9nK*;@-#8;pXD{-ta86f*bmDvpL&g7%LnSLM zC}<knU@B%r(fG~q%qPz~xVD45NY=`1L-OI@P;1kzP6f}w7u0d!Gs-wAb#N%)guJ7= z7U2;>Bz^pb-BT{|za{-D?mt#JejQIJK9?f|9`g6ZXCIpbc~p#h+B`>?eu@VjR8f4| zHR#5q9f|O%_W_G~J^x7#gL)r`1E@}Zu<RHI>->vyh6>tfnT*7x#})xyAN{wdW6VSJ zd)D$CUr`#*?~8VrWrhT<eQA5qou}bWe9QjWby!z7Z`v%Z@ZZ8V89{9h85Sd{rw-HW zn0i6ILvU&m#K$Kw9y$0dS>{Xv0x$>@WRYa8KdPNgm_(U$tvs}X9O@65>@ixPo}vF+ zPtOm|vN`>CoD0DZ&)Z(Qm~<nSi78}ABB+H-4*w*#ct)*=&7XS@v@CW6sS#P~G5!sC z9QGgl`Zi(Mb82@Jat~epAa?Az#U0V3$N$@6$PCf;k0TLh@0~xS!VRt==K^mYDYNC8 z+x>!w-~6iZ`E)+ng;7f1L%T16`5f#Nr2m^05%m;(3wPuzw|u*SrRBSpUG`@ew;NNP z<hgGbmfPKmAL3ti-|EqS{H-==>{9O_M3gG>8&~Fk)Iql#Vk@)N`IGd}zC3Y%;IZ#q zxlc;rzWBj}noF0*+hh-$Akx;Q)mVEFvGFRUUP?V%28@9Huw4D4cKi9{fNpM#>OS!9 z>*CZ|0m}!DEHRI4MahEy6vIWIy~puXNDNVXT+RK-%HL1FO`lXW4EkNyKX@so*l-9l zoE*>kjs(n*Y?G&;!7m$WdPaFUC%&=(Vdf8xQ&VPU-HUzh3lFI(>jcvUw<Aabw$U>a z2aj*C3oEmFCL$pu&Ps6lY1`rdc49;8xct2R^#1T;g?sIag!C@OE8{}GLo@9`-yNl0 zOi--zW$<L%lLNr;T$6jVK#`w?oaV9B#clGoM@aDcTWr)|c;v_Si_C}jE{?X`UE5J1 zSuMs>m^ksuBq|f&o2NwJAJ@Ex7-@}=^Ni2-?bpWLFAO-2xb%D`A!2PQk9pM8X!+pZ zBK-GuKP!4=0%q7g>Q_Hi^bB{fJ{TAZkpuvN$m@vsMPgvYU$o&7z?Qz(`C0Sifc=lj zZ%^h(hu_n$oPS*RBtJa%A%`rtCjH(Wk;4}K$M^6@h>X4tPalv<SeNqxq^Vt7XVyQH zgFWi69bAt#P_vG;pOH)W8mXO775ofYt1HFF>%D7;UT`t%UhV7gf1W!Un~Xp3;6Xv7 z3a2OuGMCY|)x|#-u3Cv5+vu20M7#6pX?jS?3gHX8R7ZUK>GH$vb}kvu{~;pKYrZm( z71!MsC9xu+?d|^rXOjG)qg!NXYj^cgE$zOeHgd<XL;B~{>^mYbEzEO@;?-{))_@0v zPo6~|zCF7-RR68-`EVhJ|LEp*P`k25l1UIf671s(2*hpO;VOQTCa+IczjfTGdH223 zGL^UI^j9v${hh1rcXYF#8eOYDB0`RA_ggEi3A9D-PmK;Mj%d8$!3G#+Tgd;Ucv8%G z&tW)qpyKJJ(+l3~J6_&GwfKV_*h!80TbdxAxN!ngG2yF2HLpJnAvRM-C0q+UJkx*4 z9O7W$J?W1M-jni-18$a?nL~FyRX;n)lam4y6xV+i`nO%^N%;Ns_X_^lvz4?_6kPZ9 z&*asDVdrVh-}9$J(nr&HRA}a4xCDd+7gbLfYT8W7FSPK}(L5RXD7k)SZMGzC*3mX? zPpgLJ`{#W7qa-ThOB09d3z9sDg`G`k@4K5;YqK|lqAL+=U~sON8}T$4Te~U6j{0@Q zs}N5pBa7pbI&Xo&{hOlOsTv=)>eC3qj+VNnt@j~U;FUc5oslUP<l5cj=^bw;Ndb5X zEg|HeMw+DQNIEz#>c<wXvgK^0PLsH}+OU82vSX@frQLVaGFteAzCnmF(vJ9HM`Fog zvwE|@aopJ;_+cD7=DujIcXy8l8G~Di%-sEJJty&{$4sodYxMkVXpREy<9;=M?)9Ik zwjlfKCF~0#o%3h5D<Quf@@jViYeEO(l<O<v9S@`K!9@U*KxLd-a|`!^t-1DG2<_Li za0#;jS3I>O{Egs;sP_>K#!9o(E!WSJ;x-E=;?=m-tlmAbJNN}!FC_w#CZHqP#Zr7R zGJY{9^k&1d!EE8C$I*OhDEgz=yUS!k3&{HSJQwGyNfI>ziR+hlfAein86v|2p47^d z|0gXej*y6fuB)$<B1CFyHFRBmaxIpd8|*Gz-#>I&ZEhTj{}T7Bd9}GAasNv-LH!sH zLX-ZwZhrggCc5lH1OV8y+{q4k{yKr?kB(6VzjHhc`FW10nH9T!#q;iI@Q3s1Q-b?G z<oXX?betdEN(!Q9|K1P`Rj*6}?yA6o$}j<pGEUqTOW0pf=k<jMkRC@XwfBUV-|vSM zsH{Alo?vbX#iWGb)@HV}S=FKaCkpQqugfN6jYs!5>FR1619u7NY7N|$hoX6-F~#w* zM#e)y{&V_Sy4Bt({ce(M`2)@W_W2Ank2tiCC-e8U0vMbpIx6aW?$_m$7<*jx-j0~l zEZ&2(_QN8!e@5avP94CICL|%1PICMhc3*TViQk)#r7d(Ct5Tv~H#O8Zp_)nWyd3mu zJ}uVA5$A7_v*=CH)XuFmDx;O<x1Pnu4FYWy;O^jSfOrx*yQ!L!sH}@YpZCS=3g8Li z82}77_~~QLwM+b#RRKE(Ex#Q<vwu4Lq5rP@M{BRpuR9-K<_EtI_H~y%HnETVCa#mQ zn!L<slYb)D9{fONnZ!!i5oarnt)suV-2W;RxE&kY{qRSBbIix|&CiEd4}lj~#USBv zV*Bao1MbYGtG-~tHP6(2Rti*DBQgIh*3F{-j|dZl%|Tn@P;&wA{)eZFd$`AzZF*UT zX<1nJeepZPncuGteycdMvH4BB`gtevNUvfr@YvC|_*p3UN<mds_-aYf<l8H!dqaJB zzoOoH$Z5f5xN%f8O=orm!kOF*%~5XP^xUsd^OR&TM$f;vM2q);RhQ-B^*s}o3;_=T zXo~OtBcAqtYLik&?cusah2+o!)c@rB$%IQEZTjdBOhCQe|2nMokE?DySKze}l>g6m zba~6aH}@&<nSMJ<$BstePKKxD=`?r-6;DRcZv-*!F1s)KLj;Qot~7MWzuWjIzZFC> zge?2>`FLy4kTZ?X-7)U|vr{B>;oSPRh990*RdKX>xJfB}|Muhm(e%|}QGU<cScn3G zgmj36(%rBMB1lSicS|=bAG*6sy1R2h326jrmhOh7w|)=b_j<33fA!4H%$d38o_Xew zRaQ7Bkz@|QQ#$^9zhP5XWGS&}fB!*bE?MlVEID?ijg{@u+Ja_gMh#{OIQ~Bfga6f& zJv>;SBe1~%k29$9G>c+BGx)4;BrkxlYtDhKVpNnd?+_%6UeC?;l>3%^NF710%52;; zoo%Zc<e;x$2vhDQy?hBsW0ZkkOfWLF%ziJoHh%-D=Oai(T`%BGPkEsacA3D4t7Age z&24sW#0r%h_%x<y^LTPSy7T_i;3bF3+>Fe~Yd{Q4V2}s&U9(fZg@YyZenXc-n%?Q? zdiML;?5g-p^v!Bo)D2|P&abwZ-QKiqD8W`_kBF@G#v;4tr`&@zreO)C`N?=jS`G~( zsJ-cmH*B@l+8f>{91g2p*2WjPbWZzar(G^ZOwOj#J5t+d1l;KZ1q()>2uRKAeFIl) z5?c_)7?YJFZMXO9U&V*;5QccoEfJP4bEKMX$UzcCzx%TQmiVs<@7Q^DSKCkj)<t3l zeO~Xt70Uhi52sS`wVgE0|L(?GNZkx{aGu%OJL0OTAAGTpC+v?Sb)?nMWA`)h9b8J= z@6h!o|Jm#dlYoIuABcmXf1Uugqk1}C7Wj1pdX&W;n=wu4TR_;pojW%Y?1}n))1@DX zc8CpfKpwjoid};e$!?{w{4AWHp>XON-lYLlS?{U{PjOh4F0&euWa@)19B`^#`7~8h zr&UR{E+)iV#O*hs0-DA4FTV8GAkXG+{GoC9$zQ3pr>|%_J+Id~H0zzS<C|OJm_Mba zDFJxTH}Hfktfq2yh;>xce{$*S&b$IV`>z&0$suBz$V_XLmtNwr0u!KcCzWB|e6?!t ziFO@Nc`PGY%Fd@}HAqZa_aH>^9yOVs6bmSDy^V?qIW6rV%Q7E}pZ{S)O%-fN+tq&4 zf8DcAsHv_{^F34BJD*a~Fboaf6&Jqrm)2lc3BQw22;~f`0A8xu;z=tvPM}lsDF&B5 z3P8v1G;lQQXKbNv)Y<&*s-L%bGG?&x#a(iDOv|WP6HB-RD86>CgL>{W|3BA!xt0Vg zmVU9AHl1jvtAD?Rydp#Y6N$638jAZzLXw1+dH<?&=1mxu2BD;?XFb8BYgYxR89g=& zLF^CN%^*KxJ)-~@gd6V>!N&9}ORJx=jPkhRT!Fkpj~Ykk$@H_<pX;7bW?qOqO8u`J z>>!Au8QXRCvroZe*8zqY4IZ$uJb@2VyxCLRLCM&{|7y--ek?uGBiH5tU2qEFN#7*Z zsvkCW($Y3iV{d*=@LvM-kmx=$F#7pVxls_19L39pgGvLwn@G)!W}T=;+TcO2CI=Fp zss6(mP|{%QDr>#;fF3(ZjYGG?&DqNZJ_i3}eh(b_-7ji8Gm4nP#V^GXcn9ojuze>) zxbg^0?T|6*WF`d*zOCr`OAXc=|I(Ce=BkgIg>0&hm}J?ZiF+yvGHyre*?Lq@Z1ATU zv~Eswteo$goCw1TUw;+dJ%yd`zmE1`tMuAjSskgP{`L6AE+ukwk_9;QM|hq+@eV`F z_C0IvV%=U!GNIH#J#U@E{?S}rn4#E<UJ#5Jc<ExrM9f>U^ummRLhxNIyZ&8l%aL<^ z$cyW*?=K?lQ~ej5QB@}~y=+TzB&fq{=*V$_kkVDm2C@ZwkSBm*@hG1B;B0i7vSCQs zEA>qHfwGEwomID}DfT08xws>xO<u{*X4}o9!&>`XdJNfMg{v<6d1*D+fXmTGB+Y2W zro-x&>w&W(wE^SEh=<ti5tF!3-3_|C#TFTm>vvw2D%}YVniLV!Y=K<nrmFJ%pwa3? z5Sq8)YN^;22lzP5=IH`QVx&{i<y&0^3Jz`Gi-PawF`p!IBZTqaAdKJIw<Jeq6VmKb z;8l@v1k<v7URsgJz(GC+mbBb(g6HM?ZjY#8BP<TW@buLOsPh!wXSPpq3aL9@1M@yd znxe_^?;NvbY1D2<u9lqbah;66no$&XK9mMzwKFSn>CV({uzCF^780Q;e$<}810c7i zIAo4w6dYH*sabAuIp;v%$5d6wx;SM!Hwyfn#^`!DTtHV-x#~1)V&f5dKSdwiZr_q^ z%rw6nUCmNqczZ*WB_h--4%g)<v|rdLE3Kb*wvPjFVYoT*Bo}b<53{Ik3Za__y<Btc zU?j|gO=-z<4|zSO<ZfVtS7>i(WdeKYF|hfM!&%gMS0(6MeEQvBiNoE1q;0)_+4|1h z&oiLHHhN`hbyC=+UD1jj8HujvcZh!kiSyjpW9h&{){7Bj+zK8>@<eGJy|2o86!`SC z)tV8z5Ez5Tc$od>oIA=0AVl7g%ubs{tj^5Pz^}pnLHNQhLXxd$g%R!v8N@aTKRx(= zKOc4Ap2}w1D+kc$Le99tn848vramy7sRlyi%GxQ?;WwNNeVG}XIoo-5Ig!c8XL<{n zvnT7Q8@9=OWV=VdzS^N~YPClSoQ<KFc~n8&;|y9xHceC-boOL)O3hqC;IXQKXz{t7 z-NB?Xt>h25Yf9oB&7Jc}&C1iRU_p`BkoMxwIx9gjkAM|MFo2r_dKnFFjS$g0cU2oR zh@p({^&h$1yWG{wzJf%dUpFnwPKGm6bxS^*JZj~JO{AD79onHo3V_}vJpRLim#51h zpx<;a(v=E#$B#i%&edZeI!t!ALg`0UrS7EjNv`W-EXy~JhDptt?<>GputaV0J;}O? z!MsZ*;J4=(`L)D{VJ`?TZaPw_xKRO3%SOnR1hfWqJXb}+1e3xZ>vX^QZK&O5e3Z+p z5AK3(lJSIT?YxAxV?JE4k8e$DI&zymo&DZCkwfK9K$Du!H*MF!BJ*jT+j<qC0%uac zQ1U%AY)`@T=59zX`dBQ!Qm>pY$mIC+(;x;gvY2X-LcA0|4M-5)eFC;e)Uv#}`?pR4 zKK=6v5|=jG0=zgUzI<~_MI6bEy4q`sVToUpU!?E$<}y@$=?8fJPfjlNQ?tm_v7T^Q zsxMzPruk80JT&Dx!$187%F#p09}U4)WAh);iK+JnxB9lt?0l~4Ei$3cTs{H>EY`_p z4r?_o;_7Ba>+TSJvwEDUZ@Knd_tf(g#qm#Tkj>`0t6|V`g)yLmKhj)rj{>lIFy^7Q zb@d|m6DFXr|Le|%I5UKtN21$UStQkYbNuf%=G${|tHeLa&b=6^2wBNo(nwut<LZjg zr)T#&T)=B^D+{;9b#BGJfAGC-W4QhkMoWbS0Ml@H*k7qNKG1}v{=Yp%O50I?GM`&= zqb#3h_aJZV<AG2r0G$8w`J#q#MVAMrqU&n&0r>8q`5Ajbz76``aF+%Rf!OZ9)VISE zg$)<s)X-D$`DYn4Ieu-6?&0c4zgvwp!$RGSb122?e>xiYR^-`wg}X-){KU8^^#}1w zspFxS?0+PhEG-l*^U`S?pvb@~55ajSSjA&HbkE`QR|ajOwqM@9$g>E@+h-QH@P7kH zlNRb9+q<p>$5Q;(CjNJkp&H`nODIW*tye4{e)Jg&Hz+vp)IbP;GH=bOqjcA}tMm_K zS0v(Z&#}7<|1|o}1BpJ@ujzxXCWRe>gYP8p<l$Q6PWpYyo2tsd1^!EYY};3>u(RuD z*i5BW?*2a==fLO<HL%hE<PFy{Xc-YsmPR&dEnDeKkx^|@jwuFIzed!7f4}T17}8pq zRx54QJE8Bdj=#i6V5uji>300q^a*{r9leWu7o?^D*-a^LjHVAbUFhVrH#MX^zoO|9 zAW-F1Am}0?&;<QBQU3N3>szle7%5q}-zL(X$R`S3T0wUo_Nr{Q=J7=o`<)(zD?sk< zJi{SwQq_s@?px_A3V;7Y`O&Fi#lxNzHsM#vpMX@qQwLWPo<&giZiyZx(x07*dqqM= zqwMbPEj4(EoS+;zwy;4jA^T0OQ$rwx_+9^g5!9D7NhI!k=e9=`j5&%yWRQqW@GD-+ zAS*KPRJRe4+(EOrWxyH00wsU7e6xr>!Jn%>Sf=9(RBn<UVq5%a_!1S_Co7+i8YePs z8%p04sF-m#_AG`1CWGPYOG{Ny8`%-_e6I$f5%-yg&>>xU%HU0{0b!7JrLKnF4!1(* zM9r!l@)&@CZcTY|HShmY@4fw>AapyM3TCt!P*ZQ5;#7Cg*`bK|_QpUohl(3gI+~^$ zbp^@YqiCW2o4eY0CEy`$lqdaM5oLau4J7UHAKJG?#OrJ<KYai>MnRQw9u4Ce=|ZXD zBXLG&YxaD}Cx=SnS+-q@w5a5qKe6aT-)44ooMvo4lAk_Aw46`Ag&AQf`vrXVb9>W; z#>VsWYMK`76{q?;`J@KDgVq<U+DkWmL-yb8_VEo!9YU?7KtaptJMk-nWQjePNz=nh zSjL=PlpXM{J^BYlU~c{z<59hIiEF4r`LwA0E#m^w2WTdx;pMgppL-B1Ks@ATT2bwV z{>Pz%nDuIR(yB#0`-XevPsIu`0?2+m4nD_>+O1D9JXGj=<<T5{qp3azc3v;IN1{)G zNOgX#ZPncEfhM_T)5$rnVIQxZ-OeE!33AY7QuF%xV2RPh8~t<Vk&Pvbi>PjV^zSFT z0}iXTon~|Q7vYel+w-1z;4344?>!S;4?*2IGHapyugZIG0SfciulsQ?I(D-|D?_iF zt0=k0hS4FXo_8eR%aipW?h`_LpZ#Bixu^{tO^X(h_b*$JH!z5OZkb9RDS4}9!`_Gg zsDb3kIhPk%e#~+`j_@+{uR&gJ<Ge<L5j{3BZ<$ahZTx7@CGa^&+lN>-PqBL!=Q8fU zXA(bNlE_cL2MgATimYFCmV0j2b<Gq2EAZP<4aFgFJcy?dFJ~_aq|x^{K;(w(gYc%| zGj*X!_wZjYu*ZBUbISfuJ|?jlZk@Gl#ULyI6yQQUtaqRV)5c-4*gnE#>jsRVYBnrW z5x}6GA?j4;@gGM!RmfG-ez!I~JMtWFB{~`;#L3{QPG;q)sQG!19#Dlu5Kg~34^!%V zgLC6R0Lra;ih=*R#Ga1meLDJX3~vJb-kOlD_G<Ik2j%_yI=mWxaIST^W4Hp;QHY|b z5a6hPlf8GCW5oSdcaOf8gPo68zn@r(7HGN|y}B0&p&M+USf>}$CBfg^4*I#Qmn+`Z z)%eT*NKc8DyBbCn6W+3-mcy@nb~6FyT^a4c#lO4tnCcmv{Q40{6xc3JpgP~YaWhMW zzMIbMT*K!3;O59&-{U!j-f(JlDV&>OC?42b`FIOYk+6?E6&^V<cimU8`~2R;o-8}B zKZ4+QQH2(JDv<0S7Qc?`WFLxyo?!bcj26Z8glwNA<R@)-K8<1f(9(r4fP1dK>@S>) z@kk=hWEM;5Wdl|lUrncSSR9sbSvl@)jF49P0#A%qe*0sTq3H7@oH+SC1zunny*T&| z!J$ho<(3}+1IP@DTtpY2K+@bu{8pE`9Q^JUhxOIZbipB2GpqKO6JG(?A+(9)u>jN( zU&%+qzGIhpy23U?A5Va#W+!loNx-I`cizn)bC5~nc7;h`n?fW_crgdQ&3SKWDXcTs z6u<-dm;CgS5FdjU<0TdMz#-9Fw(l0(R$O<&PB5TD2%kE6jyZQX{3rM$8YDo82H@C( zhrferF8&JGmSR^`Nx+CP_E4YI!?|V5G5hnfe5_EgQqNYv3Q*Axp*7?*;~)_Dg>-ZD zl<3(%ls6lDwbwK`*TH~V)=VmMtpkuxK*=qqbAE0iwMj>;?%LzCA>kz<1~3kQFCpI& zYp7D~Bwwv`sd?Cf9+?#p;&#D(N|l#uAPU=`!|Oaewu=p-S2zv}iS-gOwCixy)piSi zYtHQf|1Wq#UtHkH9Y)0Od_bT>zR3rpd7UD`F<VDGF75CZj)-$-djD+W8W;_pUAYU% zlIr6CtqfkLf6t@m(uZTtoegPi3$A2GN!rhJgJ`RX0dPbcg#XL+vj=>Gr9zp#;IU!n z=Cq-K6QTI{ho|ayUGF%MroW$?YQBt=*#N2|lu^HM2&ww3FA*B9$f}cTvf9{<+CTmL zC)}eC1;W;0WE8jH95`poqBIO<>1cRy^?dGVN!!JxYtZgz`@PA3ybei<o5?^OQznc3 zR7b}Fz4qT6%^QoI|A3FEvyerM{XZ)JtNkdfj>C^H4##$xq5&F9r$<Oc%(LXpW7^W; z?m0Y!htG(o$XU6{D49UpNP<8+ZZX(il#l0HIbzhgF532Zz-SN1(6s1IFyKPUn1j5r z4W?uNp^<+SXn4K7tn<_0nVPL=p#S8A^P5x6ZkM0k1H41X2UIj)AD<pXGTYqy&=C<k zQR=od3VOYZF&$w(hWOnKkEXp}ag*4y52}9r0~nFWtGVkBc9=2wb7I_*%I=F?nb6Q3 z$WA0ycUwJtZqTvFf9@ohfRI<vzeu42KAYVh_$es)n@KiaoBhLvPW#9I=_~$5$vGxd zT#2kCb4Ua7=RBKZRnNd~vDv#fRKaguYiB;EAG>lDKCBU<sBHic`4=YB*E87H$9kPs z0qPn}C#a;l>xQasVB2^EtKG$yl4R^J{lzg^HO>WnzP$shAJ6a}nyHsKQekb2B!=42 zVU=vJky{a&IUV*j$vN*k)@UaD&rV5dAk^#J9XtwvGsL+1Et9{J-;o9M-sG<7A<A84 zSqgpnuw<Avw-Y=7yrYpc#5cD2FKG#N7u&3Z!z=N}Cf`wC*S)+Ag#VSssGf8`7XcQT zLfo2JMQG_+beT%_Gqq@>3P<;b2@-6%tPl5y>6U+|lvAY&T{*-rq^uSyu)gTo-*qKp zMnBlzsfU4Vm%&)sS|jE<5s0WJNPdD`8O7V+K1&1iSl?kVZhdzvuUFCfCj$q^Hlj;~ zFQSxfH}MnfW#BfX^T(}*yhrU^RG$HXdsuzYQnu82V<*<sCG+FRJ=-OHeh0J200!WV z&CY$B+Cdtn6_{andsXo$f4V8Tc6oLAUbp>BOEpVNruOMagpFfT-p~x+2OeS^sd@R) zj$LnO%c9BP)!}u;J+_Cr2fDrPpK1r+?24_$<b^iZeZ)m(4QWO|{J!4{ZB<^!p<PFg zO*g~*z4u3xWvn(hR*u*~t~ZkAjFI8_jx}1BCQJR+v#(T+7|Z@yw=O?>P_ogr9_eD; zwesRd`i`+3<IaP0`F&Z<*?Hkg%WD2(Bpp*c!fwkX@>mH6jZR<O>#?mC$3_>6s-HXR z)^|5=sYw6q>@(vku2kwQ{N-L=28x{VPxyM76>aE8Ba97rL#w#Fv}1m^4pTS<{9ZV9 z7kq*|k{v8;pXE`oq+qJ6JPp}+wz}eaIDZVjT3<5$nza_$lB|w&pRmk2`eqUs$Op9b zPaV5Bn`bAJzEyoG`#03&`dhdc5WQg&S#u}0m=Szu6VcaFWw7RCrp2E|e|Bv9<}BYC zu=`pHH5h^Xa*BG#W~qzJuM}dfjL)nj_)bE`U-P}h3my1R<4xD)_@K5x!msC#EpH_E zK!Ct$M6LNMD4G1`4a{zl)^SZPfs@Baj&2@$uW(tLU#Z*p?8{EOJ^NVp*Z2FxlUMv= zB5e%3XUEg03TOTB@|14R-R>=^K70vci$m@XSkGW;u{FbtP13Nwo(swU0(!cioN`Q4 z?Bwh-ZJ{)JTeM4k!!P@4_OV(|&KTp&gJsMVVXra{FP~Tt3gvZ*5^((dBC6=muP=lb z{nsDgL>E>J1J3J4cHsXQaNe(yAtvMsUyD>(%C1`|JP;!(B8e4OC8AOI_q}a|v7KIL zcn7oUX`wj{`gMQMuKqxm3EiP2Uf-9pF6$3-ny%2)9fWU?q*VD|H!Ibrvs-w@%l&OM zUD+S_T>d6d?PP=9!!!fa52%Wb7<;e=7bm8gFIcHnrmtEQJt(xw{groG43KovMT5?l z@;CD!29`&kRH!sg29iGNtN(S3)sNq*>{9rZYMA_3^C#rtww3j|UPmT7%#12HqYY2m z(L=-n#+#Ke;>043?yHNN@tluXe)d<z@gM%{pUnr`2#T+89a8n9yLNlcuSff#)FAKD z+*nJ*G`htbhSze;ST}k%2!IRJQJ*g1;41}#Ct<|>|6kL~9c-?D7gtT~h}#6OhvfX6 zJ}tB^XP=8{=-cQUXkVVRQA>>>8fhY4wJS69Oa{ewp%!}*gHAh0;kWNENiYBs$(%DN z?RupzqWml46C#(84rW_!R2duPJZ!jTh0>tu#$`l<jq_EJE}6{Bx0i|&Gj*%?0E>l! z5#F@nrccK|FxtE9xf8b~4{au03Xhl;n^S-i5HUcRJJRK_NtRgzYAKI4xf(P7&o#2T zl4ZudrgyF7tWOl3)`x$9Mjgvt2t7kQWW2(?{OpH(fE6tUj7K5KJyXtbUCVaKy8Ty* zAAH|Wp0O1EW_bam*Jb?H!f&WE`ou@#7haJ<oA(=+(l82rO?R4mb7@}DdUMmS3*>WF zGtE@k%#m6rSA`#c@+sa_cz`aA#v;yT!*XCZQ1ESk6!^a7*yF(x=j<2BF|+Vj5d8SU zBXU_jilAVJ-`k4+g91WcG+b|X?^!Fa-SpW#Lc~vAt=88k(eZTth&W#E8woO3d%U}d z_^+`22=j@YqHC<6hJlGcvTtTz4^mz6#L_=6^7S51&u8cM{C4U}40&L_uzX%Yj<F4P zDt@;Y^9);B+`nZ#5G01`X!HX*G@{O@@2n}jq|O=hjlZ%AXlEULu6TXJu+h?aDGyIO z$i?|2>T<|LL)oA;j|Fk)y7iGZ7}g(PpCNIj@}wXk-g_8w&<<LQpgg2zAcaL`J4rQd ziWqh;7>{+X&3)9m-GTUVs*qb<EH_r(&bpcFYNWhbYI9olxcmo#_!aki{MMxIMtJAL zi7v_3gdrb*X9%lDk`f*PWzq_ybawo^vbOi-Sl~koY2I(~T^<yhgo9H@rPGqKg(qv! zh#co7d2Ec;-pZUmlu{miME>DmTMQ4X%2)l(>wY7BaK&4wBxFjS64DxFyZ_6axE+{Q z`lvx0-9MfwE<IVCE@};q14I6HShp?N41a5K`JP@Ic^uGa{;OK<BJ0O{TWsh~qZcNQ zqy`~uAXGrBK;le4rkOfG{K*dhX!ihe5=0;zCT85&CsuxQbo$vZdyP&bSAdBzff<&& zQpj$MB*YH8ghaN1=LkqF9ZkG0r3o>UEg<1*v>zwo#OG*JU~m2Mpn+c8&s{)~@`S@W zCl4Hn6A$>{yZk*=TAxC*HIkKV=h1rtnhyE-e@Q)FHHB802UDqB)$h4zydm$#mArao z)uUczc~S3kK7RHfSRWwRJ;8adz71)O_ylcNxR$QqcgJlN%HNDG@>F@?j<dg#jtET| zqA674q@9@WgymctcS&$rJ1X>)8dU-iS)&@lk_;)DwZ{f@ZzU!kVPLcx@8#PJW`sKZ zms5qk#&GqgY@B!|<|{iPzdE}fS*xF9@JwH!J!Fx>zMDL2u|j(8v88|6i?@6HK!&Bw zT%I4sAe<DO1Pcj%AN<vi<CGm93|LK08V}$_UZ^YKtr=Q)Py+y<sCiVgpJjMc8c%G= z*FD6(DACO1PNH~8`is||5TGCdj6)p@!z+vs!Ui8QfoA#ntvVF?80z6lF_#WPi16%D z(aQYjis&j6kgK;FVRoozSaSM$ba_;4mj(6o01urMA6X=h71v}}t}1LWj<D6Ukq|G{ z&va-pys0Cp{ecl?V7Y?~jFBgxl~Q?%m9RP?BtA%vV-ezUf^)*-a_U07Vg4+9^G$O) z<t_zgNfVXjNT|5WlX6@j2rmdg(;rnIMvNg+XG0|PO=4szoPrfgkRWzy@ZMSXm!)sF zCCT2Ky&W`PRZHpo=U$(k3ZEV59xJ%G+H<Ys<E+Q&Te;-#08S9eYF^fE#>CgF$Iivy zZ~ZGqJyXzs(VFpn#Bv?G?eZ_|=)lyBVy;*7j@E9xR3DoH$_RgA8Ij+7$eB9^@!igA zDAHwB*&8(Jww!?D@tr-~QRI?~^?hdHA18O+n}qC6-3-JQ=EVMz%#ls6E}t9yNAJu# zwd*AI{p?B=33buZ>-<r-+@OebioTdqSEI)G@gc!-rEqwO(oy<qY0hT!_ln~YbF83L zk3Gwj8}raA`XtCBBYcg11B<OLC1>5bkLgirl|M~6u6&JDDDO{B{qCs3JZL5sU+Q2j zPsiFs%Sr)3_&fi_{l-(z-J&oK+?}8!+0g4XuLx=ABaAd0rS}Sdv+I+FTg85{u}q_{ zU=?(VuRmh7Eju7j&{?&@C0rOs@Yt$O7r)s(1_#IQhaT6V#rzdG6AdZY6PQ$tYSd|O zTp#^sO*ju#i@UNM(;oAphaBCpi7*Yn(Xl0)M_uP=bg=o$vZj^3C}fnaxmKYpe(?k& zEXw3_dl8;o;!D#`!)OV_c846zlELK|blb4Cz)GR4Zgpe7;q$oZUkm5aSzhCyxccQF zu&332Yp^4=BS?Erpg)kt!BfP@`qTEKrx^IyCW#ECb3p`Zb*opJaw+K39g$$dQG8#E zpO%LxF(kD|XEJF;3fx(>sH)2qV>;*&MwGPNzx(7|k`IZdovWcOTj(uY;3YN2qP=aE zx{vf`34~i7l_SHv<>P!fPbFXHEV{lD<ls-AOJ$JGn$@hrht{)JHP3Kcbtn>q{rh}Z z)7|t*yjkwJb_C5ak#&w)oHRTQb?JpV5hMd&1AIJ^S(NYYBvNDy-&O>-9_-{+ci88T z=AF9gLA}<9rD+uj&s@0cTHdQJ^ZjY{D+OU-%wefp-*3lqZKR2-sSHU$5VIjZ0y|;B zuL*Fsi_0CgQ{3~czYD5>bIRliXC0hfU&*IErTfHo_uE5&!gaHx3m5_z7m#OrWM$Mk zqH#m=vgtt#R%J>K>qA(QnS!}{t@b%w!9!HJ%_2V&I-<z|`o7onu~3{pTxH>8IiG`w z16i_t2r^V_<9p|5t`Z^t{>WYEzq;eZ-;qyq672H4w5(k6UINc`NEC)}+iwSQDPHHA ztD#OrG9$Q~5-STd!kTNwk0`&pUrSpJ#>x|Fr01p;M#n`;y#MKFkTT07xWH>1PHz1v z#QidELsPu7{U&gyj1qL{p~Eh&)Jl7RH(E;^`l;vpJA}12?1=y&bNdACdkKSO=T*y; zB>Bc)V9i|Apejue2N>`+7KV{o^67LL={+1we?#!PWvXI`68oEN!5^id>Mu657+PDN zIi(j4%4Suu=l&xug#LF`xG(qBrWcSf8bv?XG6CmHx;Ka;$GIxz2n1{L_AT0fe~Myo zk#X@N!F3EI+So{^q=e;#_Pv21z?Q{o4tiF5zj7tV{ng&4adkfWK|%N3(+WJ9Y9F=2 z-iHHAJ;qx})fN1tn}L`v!}E7V&9`pvznsZoyAf-%(+^P*Q%-%$^}A9@5iAHfN-L|F zAp9+TL3D0Op<SdlwHVscf>A4@996y1JI{pRIxvHHnIwLBn{%(puTk85Ip0BdwfmxO z8NS=Wl)-2>Pwcr<<sIttE2c~%^fWt>O7eGiiUEAWu${MmldK!@rDrjQ9C*LYh&}wi z@@*A&T4~I{K$FQ3#S*B2@14L#$)gzLY+WFvLPWsfPGN|Q<^*9E>@Vx4%SHgoGW0f{ z{I^3IW8U@*ck!c6e&fvNIvL!j7?G9d97#;9=q#;FGPpkI>PEnlln98_*Nd?lnU6os zlFoj}#s4+=uehj8Q@FBg2aAl6F@te*|Aqfe8KI`n(VSS}0Q*w#K-oC_@IZ8XUp$og zLBx}<z%@BCQ+%5Z9a9O)n!mbh8fXm5^_J+Bes#Jutp`X5UAO};T>tnY8vS>OE~JgA zV=X(zqG673#MyMlS;@G)up##x^IIf_Q9S-1!mAM)l3_v<$YHLp+wpLIp%XUv&Uq#_ zR}5U={Gh5p{?l@kz#n9}R*F57narw(I&6L^e?WwmH8`z)dC5{C@gCy_;JsY*_xADy z(T>5>mL@zyf9#r=M~^W2o<%TCX(W~iMv{zQEKljwUCuThoo&bb0M7+zt{#%-B0SYq z4mv!<Sc&D~oPAM7J{BvsY?^w#Y0dp!b$?V7*ct_cG?<tPFjze|f<&S9(qwtO*Bu6$ zHi*Tp%|S6hQ;E;VT@B!ts&n7B3H)>t=n#hzkGia_ZD!PApYOyDmsp`s#;>-Ah@_B* z=<IB7cJzmrUiPzc9QZL(Y(N7UpJUHR=AZ3#lEuWykF`WUsL7Y_s?Gy@!G6?pU8h~% z|NaR%m`7_l-{wdY($|M&0>iuhzNp)q3IEGu!sn}2wv=b>0`!*O3Su<grI4CM2rQDD z`V&vB{Bkw;6DIQfT<<x?8&0ns9k+8RZ~y4f3Z$+=l*LQF_^;tn6*=^tDJaS8u-o78 z?z1lbi}cLYA_E)l%+Pm(81{_Pewo{{9d64Oj+BF#tR%U+2v0S@{<Py*dIpP|u<gRp z6N}i(M$7jza#y)gJSVP;E!|xuHt!{_12vmZ49sPsTjL{lx|deG*JT`iatHe_d?ABZ zPL7t*MQBuK!e7at$fj_gn%O0UOI!&ku6!^K14FCUDzErbT<YE?VstDphQ8`@NIE69 zdp$F>YZY afWTn4IxG*G$CZt;<rxWkyQgBipU+kHv3qhzReLs3u^)n`)j>b(KnU z+G^5AU|7WG7b_Tu!Cex-R|Os`CsvF>|8y|<A$NKO_V0bO!KHdWma_@r&1V${Jb^3c zfi2f90w8^`zdVlQbnH&*g!=v+PXeZR3q#t^&r*+n1b?PglGQ_+TdVw~q?R-&_jVod zlaTQYK`m(%#!|nXT%_<qRO>M#u2$`}bZw#*dOH?l?%CK%WjG_xFflCB3%fv8|3VrY z79zG^cokGBvlIE9|3UdH1<~V3WUPP(B?&BkRii!3wgu)#agF1DpLbV74BJ!k#G`|( zGrXJYEUIPkfc@}B0r5qwGbx~+Fk%-U(t)Q7F-_V9ske`hx(libGOy6c=qu+Tro<<= zfk&F?!*><GTewwXz<EDbI$5hGHgev}q)Cap6cp@ZFcXZj194IN%j^6lunsTh2**bx zEBuBDPwAh~Zgr(^lU04;xJk%@3*7I<j3i3;<P~P7UCLj%i<uWLx<5MO07x|6T^A*7 z{`kiov=84#_C%co`jzWHAyCT4@4lzQooa3==hK$ztIPU1rFX7Y7des&fT>)blN$=o z6Wv)#{Xv{XR)B#p|8Ck->aaE&+1&0|DhuHC9my(_wjc<l>Cs8r@p*M^irCx2S?=%k zHSk4{6mP@LwSI^5ASCe3lxg+jA{in)^F2Ln!OQfqO8KaRrx>u;@1~imQT6|{$e9MG zr)+PY4{S|%9T%j;pp;gcG}!;1F2}hX7Z^0a(~+stD0G=$)1Zw!p}o6)**Q3O8S!t` zF=Bu#)Ytv6h~!0VBj`b(k*ZY+8DAvnaio;2LPVSh_uvf*@B7g%Xc-Jgk$?`4TbsX) za{4%~UVSt;Z({i1Y#!h~`kNyCBPo;lp3)4SuYV>7VfH03HW1IKNnd?m|1kw06uNbd zsWbyl5X)aZxSOqsL7F(uG%2Ci*UbwwQP4wE;r~$g$p=F%h*8~COXQEXpM&1#&&0Q+ z+#V)Ra<M}6U%+Ej^GZSO@8>88=r>hmA&bsM-leO9AtwNgyb%UbiS&Q_7pZ&V!&zgt zmAhK4CC2_%D)aXnkoljlpCvo^=LM2sL<B*Pdwf5>;AOb6i8zf3bpUNriqCz;^4io| z`VA6vtu|g;ycldI*(vfDPDc(ccE?r>Y%Ou&w)u=!hwGqM@1WtpJzNk|%fZzyzpfRh zU?u{v|DE)*<xICn%GuxJG4|qjh|x|c>T8VxXynBAblZum2i9`DoUN;t5_dUvp}yHY zZ$uGMBT@K%H{!ouX@?hOW~Is)_!Fe*1~f`g|AD(bmD3im3&N7a;PO^t?oIht{)7Vy zW!IuA+=gi(40OA%%BGp5gAQt$=^}T&F8K1$39m8J^qG;iEzk}BofG3D(Hq~G_f9JZ z0*7`1LB~*+?&p50!|K(+ER@b0;mdpKrWf6sYPjO+Q*Sq3uk{qSJzx;vepJ2{V9<(x z#VHD}=w?ki+`OZm(4$^0_qX%Oleo8@m~hd={52g&?poj2r*wh_Dxh57do#bx{%u_U z-IhVit}7+1Izz8HM{B9*|4IpdV$m)F!=#erZ5`YkvP+KS)Eh%-W#;rzw0+TEZ+i$Y z_TY}SqG&4=6Rd}aD7l7g73B73_s+%C?UvoY+hJrixn?C>xAU;4V)wa%i%L}IH={Po zuFiHTm=j+=ofJMsA8v;7u48Hudv*nqsg&WyVN~BRC95>~V1__K#<b+PCw&}%ao*5V zhf-3Qk@@8ll<f%_pZNU28C3;8@K?k*>@(Jod-!X?X=+I`SooPT#76lk{*w&63VQp$ zHUdkC?u<IZC(}eti*awdcT8zI?@B@rzVct`nMUVQfo}k)bQwO(7Zj(~cJXI?Lafn@ zebu3Z1%?UgU?@x}P<ixG)TmGxIm|I;e27MkdDS4{*BnrFy`&;4Yb2q>OP(erQLVcR zycP?F3dTDimdeYK!)#haHER;0hn{Js<@vcDmW1zAGoWI>9IlFZ_Rn}fO@3+Ov~5Di z8TKdwe`v-4?9N>g_r(-0o98L#J)Xuu_xG?0cipxd`nM84VwP<BH|PM?N-f?wF&OfN zqjHjBWgfNGlp`ULv5>C%{tc(JxA0Cmr%-yIevTykNxpv+-jg3fzhR$qx>a<$b!nBy zX1EY*krkkps&DVZ5WK1dM62Dx1<n23MpvKp-hN{6vXVpXF%Y~~jVtS#6cprEH1hg1 zrG&P~R=KvO$;*Yf4(-4Avg!B&$77+~@a<C$PXMK`7LxwhWzh8LyxZ?IyCys~<i7Fi zE>#h-VXr07H-QYg{aq&MzJNpRiSqc1EMo2$-fi_t*39XS6_6bCP5t%&*~CSAEh{Sj z3)zZlS<J)~JrQD2$hh97hfiN7R6i2NmpFd>Lu)mtj~O}@x>Yt(RvWLz1>=aVS+3)E zxp3=~u#l7*4qxmxP6Z)UQw45|I%Egf<Q3FJI~mZk+xxjo8n-=Fa%KriB65ZfEhj?D zM9~n2J@Q69?|`$`T%a?m<q0BvtFbt{`(b?RSpy=l_rf=9v)Md`%;LMHhQHT;y+>DC z$d$1&ZIGb{e5bL}L%5niA|^xl39FB-&UI5V#AD}=OEljmKm0lyk4;Yg)hd*b>D95L zV-D=XZ04ns_N(KIh*9@{>=<EW+6T_|{eco?4zFU@tFIOeJ1abyZDl|I2Vg83gT3q9 zcf2)oNu-I5ubnHclQ;eb21+9kMRNL2#up&z@GMD^T!yza<!^up^^E+<?c<G;PfAXV zd6r&?p}XFdt06KWW?(alln!$DB@4=W;a>V;<YnWF5yHfpSGtE)fkAPEeo#q+iR)1Y zqPJ*O4;0)IMBdVyTiLe#_rHrXs^cylw6Cgt(MlR^e8%#48=I0r&uvN}CXSBN94%F! zT|5T<woj{8QI+SnoNnh=Gzdb8K8YPmaEBs$`ax&3Y`T?K-g?izVhY7-Z^-6tbaI*M z!XLmgXdjW6g#N*Zl*kEZanV7bn69lo%o<_h$MwSI#u6Jz)I}CGtzTbe5w$rdxH-s8 zcvWpjjGev<%A(;-Umm9&^jrUfa;gxE(?nEcHh)=`(3Y55O{(^6{E3k3A3@n~7Z1K< zVZIz-cFBFWmRmuGyPuR(&l=GGwz2J7F9xk_VcCf5HXhct3kY(0T*2*LIY!>qM`Nbz zeMA7%;Zc<7dou|*ffs-JbDTypJ3>=;n2Psm61DT@j{G-T1|NTNa7W|km{zI#NU#?o zo&(ZjVNp^mLuORdK3TZ^w?sSGwJ_UA_klsTDBX&R>z$UZJlsX{>W}JAxgFG;PIZvL z`HemJqiB;gE1${NqLtgGDdG>#+fJ&QKQc21&pJP@u?=HcZ*5e6irD@(h+*{RFI{l+ z22aFgG7%NY!f#MV6DP;<beqqr=<Ib-F0~iTa|!mGq9S9=yzPZeupmNL=mT>|F$X_x zjIRI5aL?kDS(S?D-xGGeOdy;W&Tx$=d3ojlx#dl5Q^9{gZqef9`T=FDo$tD{ZlHML zDbc%|X%o!bz%G=XV>D9cTqHp+ybR@64n$4r$I7F~-vU+WG)tn22HBKBx6Z`5OUm!J zX{F~$W-rEmKklG#!@;l!<}Fguyo>wBKZu7?cm6Bjo|;>I$`s^#KE4qIu|W6&;01^s z`-?VyZzl2@^tZjd3)m$lP(^p)0HB)5;X>Mrq9#R{L>if$fNFjD^h?+_>eAsiP(__y zkNfY?NW<P5f~lBE0l)8;vT}QAr+Aw1Zz1{-)?1n@cRk)rhlxMQFX(RYld+!GH0zP2 zp1brJHiK#x@!LRqaT72*)3Ir{-J2%9>sBh@*{KR%A!aCreYA^Kd8f?Yhar<>Zl_qX zn#UMxUnc2-ixK`(0~u|{=So;Z+|0SC&@o%tsDY<<bvDzRA`xG{?CE5q98nj0t;xtw z*IG^cJLc9lUO|N}wzARcV>?dapR{8@%1fGlGj`<8OIiM*BA%VIjW@_0!%jW7@@fup zaD2m+smGl4Kdv+uwJ(ZfvX6G+D(F0Jo6Y~u*XzAsPKoz$v6&^8wCQH-9SmXD9aG;y zrDhz`6+dSPLRejmt?G8a-H<-@XtDS4nb>0xV6g10cS}_+CT-cNep1uDu)@3hOwI?9 z5nFB+4J|dr-~?WweyNVq(4ozNz0)?#K?HJCcBe$4HEYzs+b0BVYBE?}S280UhB2a! z4(<^L0E85E(5QJ0RQfC+c<G?1Wni(MzGar?sx%ewsAO%c=?`<c5CP5cJLBb4)!W5p z_&fhg3!iD=@DL6TB~CT5+hiwf4x9`P2x?I9gYMVCA3I3c<4}?r9E{>cZu}uW1`&q# zEKk8SX@?2)&PhbLIC3AGWL0#9|8Ge}BrfFcdLWFZDU;Ihej(n<_2|<f&CWA`{cB@7 zXdsBVole|2Bb9XxcRaq8?+btcNTTHw%(c;i*?DrD?p1cd)nHWoCmB*LbZEIr_rX%8 z;JfCQ3M=q$zm?ixaDK@77I8VVcmWT-=<^iJdD^taWT-EqY(gXPC(lLSo^kMV?sW!K z(tTB7`@=d75P>&Xp#ePET&T!pF_NsY&~MW~_Xy){P*}%U)@M?goQ6*pG^f6->aUIY z;nb;{*GLZGz<}`XTTf%Db_p?ub-6J&iy9__m#M`_ZOqjKUgg|iC)0c~j4=69S-DbQ ztqeVCrZRu5w+7(SHXqNWj29*FCLv_v&0oi(#v%G&^~+cw_Ofv2XcdAYcGPPx`I(pj z59J3LdP?Z2o>b})@Q(NXOK?2#{O;%Med`PwS-CnH!EMe9?@lY3MS5OYe3~z`fYd<n zr5^!O<0~>crS3Wt<ds^6ZFv3@z-SF282%?E^6%M_RlF$q62wu}L9)c&V)dA-!%h)> zU%wTWyHi|_m$UDxu$Z`elpKJ-u`lM5aA?_J%^ur#+^g;X8Q10K_FMB>)sq+-3s}+p zP<m@r;4qe(woQZe+s*;n3~-O8*j2xh2Csxu)OkvWoa<|Yq>-n;faHW~w(Ae{VxUmr zE{r;h7^8d+x}cBCTKh0&vTU#(_v8u}!&72tT62H+l%!0hpbnWuG=uiFoYp^VaMX=o znY8A?n=lhqkqlK}AW1co_7tTz+S_`1_nf3U@pughgHkZ7T!F_Zg%^QeQe3@W9k8c( zctT}YtG^OD3!=rq2%G;{M9<dsDUiQVeZV5t!*UnxKTY)-bgvI>%QX-d8uEZ>)FVWa zZRz~g$^_mV67VEvh&DNRnOvN7Zq%e!sYTKQ@n;ly+9k6$rUWym2u5lQY)V%>(Q5p1 zkNA^6yKA`{l7&WOx_#C8@b)tRw%XWrtMJwNG8gtCL8AJrhFRHhIfYp0_bl4xb#U^U zn~-%g`&EF2mv(+0JctoG?k45<v373RuEK{Z`r-Rw&aXHh3&+HI=@n#1u1J4=s{q>$ zS{D9?1@Wc1aToXXT%}*tFBj<UbZg<PgiU-ld%PZ!6vaWpAp%{U-pqlQoK~vruuqs1 zt2osJN5q~FW8m{N%H%OGQ+J7jnA`S34Rw%vvyUIa_7xI}%cin|*ea%ggb1&iKB4!@ zZ$8~AD$9^jMO3qLXO$l|2r^P^Zk-3S^Wiz@XMZ4yYIAfEMgQt8-sCHOPu#RRWmDmm zOt<Ix0`LpQ*hijp_6INa)$^XJt9e|mk55=IzjGBh`2Nb#$0wGFcVYVHI}s3x_>&&D z4^kU&Si}V0b+bsw=G*y~<c%j6ddx%wue8oxUm9_(y<nKos$yHBDiQ|;#ce<P3h^;9 z?V}cXSJ!*pr2SQAzn{^?N;r`iQ*wF#z8l~+M(8(l7qLjarSjI-7REn*ugBFn>W=Z> zO-as;xxIM4fc8cGVxcjJ);L^Ubf{X8!`d->3QQ`g+GG$|yC}(2;OrOhQLj}ni~vkE zS@$x<OX$_fN`?C&{i+|Jd;pWv`Tsk0@;e=kcgz(8XsuAiD>o+DZLb3^$bTkZYT4eM zlFXcQKceR_@?98O`44A(vbJaa6ZDBzPd;?lvQNMxHFZZmWh%P3^q2hcHDCHMqjb%J z*cF?IST9hS+E?&ML$pZqG{PbpTtmm*&lAOMPLDhcC{&Klp62)-{Hr^#5`*L6cxYD` z(244xfok7KhpT6XTH_raqfLGtGIB?}V!}t)`xTFk)*Gcin>5{CtsV9W94MoiVh$4N zoc3EY!<!I+95IW7`unb+IR3#iAZl5^oT=~Qx?|arcr8_0P&|J7!CsLWp!uF(kt?I| zkMZ*EAF*-*uERbL{E-8uDA?zlYAPBytS*jw*-|wwFQ=l$myYZ$bN|eX=hK0im6=+i zZ2Znf_0l;|(H)BE2VB~UB&*~NN5qAeWGJ6b71*by$OdoSN#q%z{OOoS@4w1%lq6x| zE?U{4mw)cdg8|qon!c83Yzzc<*vJ7-qVINh)-Uuz>pLJF0$q;#>qv2T#4+kPW(V+i z)rf&h-7#fbu)Yib%1hCtpU#?F&L(-KtFm>THqUDD&w>&xEB~e(zbR8SX&5w4Mfd8a z%I`BO#WPrEP}w$khc0I$RmXl$uk8M`w$b(v!m+#C&%htiG^EP##V$Fx7jK<}3T(xw zym%VY!K&_qMe_Ly$c`91QI86%{Ib+2WnzXA81^yinl%P(LVVNjRVn%0#P3lQiIsY# zGZ42yOZs>VZ1KNyZo3^mZ7){DRufe@9clw~=>4nf2B^az^rE~}GBX`4&R!8%@G7F% z09y=LnqZxbu0Lc{i+rT5D=TLWSRk=Quue3W`t_-<M)7X3O?d`pSHOp^m#zQI#uYVo z3qO-GHCNKs%^lh7Hz{>LbNpxEb|0cs$>;MT)bhbo)&^u2;a>5yarv<W1ILfhJgFSk zeF=_~Vr%A1o#wWE2|#t0!5kZxxtGx9xU|IKB}wRi183p4>EDDYLFV=5cAfey17ME_ za*fZDEL3~rzJDj(41Bah70<7$+pHU3m+^YY1adQV6m&N0mD^vH6wY`r>t25VENQk^ z1iRdvsku8<E;hFJ2_~4j@<!y(rZ8yoWj?QT;Vc~z@icF}cDsZFcpIaE=!xsA&m5Sv zL6iqJ*@XZt;bne!67~fAT=awuZkT4avZ~)EW%wbS(_*b8Rk3fA;>Ta(A2NF*3g>T% zr`c;J4?Z(IO(S_KThMHm%I)*yoyPbK7XbR3iZ*vo{JahIrE?oZ9e*vgYW3Mw%Cn6t zmUaufLMsuJ6Ernk&~`dD(NcWQ%aDkqO+LmB7W@&&qG>sGH8?T<Y=a0`R5faY^HFdm zH=YQtj|LE$A+rJbd4UgJrwR`*=U);xlYNRMns4EB^wp^Gq2I9ld7-j--MEtxh1XhC zGK5d@=OBA3+3kC;+)x@{@^hBng$v^J99n{}H~Qrib|M(OlYM__KHC2p(J^c3gUn-@ zqsTkNt_8@N=Bja`7wP8^61v_+mCarreN`F*-#`GXQ7Sy=oZ|ozT$<m;%3J0?_YEdG z)~6#M5Mv?sYK8Fo)m+ien}hAI-SCT9C`$qzPNz@F=D~>O;jZ#f!^^S~UXi#gT1rZb z&NL7yZTBQyo6V4h+qu4kl!+*%QlBw~P4I2eO(vW*2g$U^C+p-N5|Z(=i<N-;a5zEp z4Ow|+C}qx2&PJ>&*<w9+VDX@%ECWgM?mhq`Hu6C!LksSs0Pk^y4=C<48z!f=m19q8 zB4s8tnTd2|{`@EXlwB_e-yC{#ACK|w#DjUxx4_Q4FSZ-;S45v3-pPW9(^#+mgL^O| zQ4x3!jBBU)Bz-I&ZPV6u&XviQ?VpD2)@KDB)&0gWBj`ORPwYRzu0sGd=+4PHP;Mkf z5%V9Id8{Ojo%`BHnkP_QrRh5ua++)?7bsdd-I5F2DI?{M$AHUU@{JhaHa@zB-&aLV zAC5$N%Zen+`0kxZtI**hy6pH?ktZX0;nLprvO~Tnv)QhNtDBY-wy2Vm<xYEtZ+TCX z0gf!K^o<TaCdMeV>3!0N1_2vcj|ewyB89=Fb*U`epQ{^X)_3&MmUBN~P*tiSZ!g8+ zAgYw0$KLAk#(vte_J@<TF26d)w%scf89oe?C$h%u4gqyvE@+;#3lhKU<+CZWvJzPu z4cuc|My{T<+V9`zQ&VO0?;2V}xwZk|^&94|gIrm<fe7wpbB4NxNcaKbC&jm;G;yz1 z@5WccR(0%UkeT(wt7Urd&6!NJ%W+jM+xyj=bj=hat6E&4zyME?XYUd*0WF7#>CkKL zILxLjv&@<ga6@(*J&}3(0=vOYX+r@#x#IzQc=gik*OwVkxm0@D&!e9+W&g>X>g2K! zO#)R>_^gons%9HKmlS%4*Ya%&e1e3%)3sG7qY6B(fexvGydoyp*(MB%%HmFXJgSN^ z%Qm0;qR0w#I$!d>sBu`*FQ?$SxUKfM>7HBQ9%^A1J=-q2$2u(lUQk?|LwnAM`<pER zT|0D>!w7-fR@*BCD_qZX3~V6!x}+<_(_qm?C8wK?obpTd(=dl=!%7Xanp6hOKKl-E zV<4ncw}2Ez*+*mKh}!8s6Tbrx1RA~^!N&j$ptKgvp*BIewg(Zu74dSx;QL)&o@m(? zf!hhR&;4?8!26Z!*0`!`K1k6sH-eoHKsUlme(gfG`ug@_+Dp07ADd156kDOm{6&bb z!mmFGV3C`18j5nJyJ74WsE?1g;_U_gAk<?m8GEPuFLp355=7C0_JyWA4R`$oN-R!g zVgT29T1P^?`}v8$v9C+x6t@A-CvALOk>~{f#`DMmJ*QKyy1iQcp<A_3w&=aL0j0~= zW_A&;QBI6Ul5;KEKBJo@AD!S5&nMi8E%&Ez6N}hq%IE=JH8~>qHN{OYzDIj^jY&9a zs6=VDNhklt!^y5k=Srn#Hpb=wwe6H*!XHl!I2bB&GZ9N$lYENtxw8=)gi)w+ukN?r zUzwgN)L*m=Yx;TYw};ULm*5agT^=q`dkR8anM2=Nq(Qy<iShuVY!jvz<T@PB`NPTk znIb*Tv$Ud)ELLUpS_S<^$wFQg81q=60hKZ7kC}ZC#jLO9a1aQwkrAD&O1_|;lldK9 z9#zb~^kYoq<Xs8V?M@wk#k+`Xf`y-h&gX4$(?J>klazo8YlvJJVg0p|JA%-Fn#V1h z<_KwjynrEKtg+9wFT~-{Ca@vxlW~aJzYHx%{T6hzgj&nimMiwhA$pss!t>&-6UCFL zphr1Vc82cU+juW3U(xqtrg6oyOY7I_H<?As#D9*z^Wc)A#H2OaYl&%zHc*K#K_3%l z4m%SRcbhj!beig?C^EGD<0UmQ6*$-UV_K)@83WT%76QUZ&vJp5;_T#WUMtDP_v5|` zCM^n6_Y>br@m67_YWT(<C92Ggc71B1<j5|UaF?jr(!M{9dAH_B-dy^bz<B=sZl-F( zBGvDvlmJihca<s!Y%;)n1z`rktA!6Kl5Ee?E~uhY7&U<gma<tj@gLWJo!NTBJC~;` zSxjV0QmQ2ho&JU9X8!CEbcSsump-|=Ttl5kgut3=LAe$;w=cFeP~NSW-dw`k8y$AB z!ROG@yoZqD;LqhnXTP_RRR8l61-wg+wD-9PD6{Ba5ZLV6+rba^ED}6gaR4AK_}_!f zcMj!V^B1ut;=X%lhN$m%8x!o=uFswvM(xvC$ooePf8JDe+ua1Wdl|4^T9WC@!35J~ zIJbU2d8?-OW9Y>6aT%RUHqbduuS<G8MABnOB#!f=xRgJ!plC4X|7!c{sHncT-x<2Q z5s+>e1w@c?aFA{grKLedx_dwxq(ebEL_)f|q+5{g?uKFR!Qc1ZKkob9yMFgw?>+0R z8P*#1K6^js*-w6+y)zS$7s4NAn%D$93HYJGk}VRACHn!m;@mi~aDVJ>F%Gj#z+Nk^ z(5=VRS3^j921HDzG{W1J@BsbN@yC{(aa5a>u_HZ=Y(>A>@x}`vWm_WVQ|Ehfd%P(j zRamG!VSGetym*}Kmo`y?JX$uy#5sX~**++s+5YoYmxt<%ONu!G2mqI(HlD0LX-88( zEnZXb$L8{zi%@F=5f_uZ&g8nFP&yJ{hCfIl=1pqmUWR}Yg*nKEfm~+lV|J?JKchCt z`h^G23{0%|>F*YyF5Y!)B8>J=O<mBvF9CF(&2{GWtC~9HKeY3V@RzWA1K3V}*2Mga zf#C=Fe)oQfTE{uzzE`l@P1<gPT^m|S{DBMY4>i2J(3Ee=d0qheD6-{p$%fZCH68yw zb?R$bPBLU|nY%%!g{EwYbK=J3<A@6Whoz#A#DH2CCgmP2%3akBTjIi)m5&l7v}ATv zQbEsaowR~<s6>vCfGgy6To5~zi?Db1KS<lgLaRL%YxJeokhsqeUCtd{S8<&8Y+xT) zx>AxNfZdRN-J}Gc@@!YTCuWn1Zb$TM$rX51?}uKF0M;a8^R~<mSBvQg5P00=SWK1C z1rpUKw?^<)lEi+5A-6swtx(Iu=S(+SIix2?yQN&o)sQ)jlz}ifN@^^@a=a*yOq>*# zif#dClWtnmn}=h3<B|f)fFRksv?(H3h*e3y?d$%r#4u5T_cdP#FfyFDuo;2SF7nVE zF=WuNd5)3w``SW!G9}R~06)JN;A<mqBD{K*J+FCIs|HBj*xG5`SEiuXwu%hGst1}r z7pR{M2z`uiRr)^6Bz&f|(;$&4Wp?OYA1@L>h*K;iJwjznyMd~i`LB)%zGMoT9@Dcu z^VrzJUhahY$i&#VdAH=}pfI+U;_{;b=n(6lac<R?q5CnLCzcuST9J*c)7UFP;1`bT z_Q+~)ucFGSM<Y1epL~!Z&LKJv<FT4f{;V*(f>jf;(4Ok>i8EBVWOe-|uvYDy#B1q1 z$-bUm7PYyN&uNkFJBwA}4K{Pv_4<*#wvC%O7;UfauC94Gk9*b0&A5cp`3tkLoia;o z-mR7zaMB1&wE}M=*^78|FOO4m2@$+Z1`Di^w^a+u<M|Ye1%JYhfg`;;3~lS?lO$&a zfMx%ZCPl<tG3$M%2Eu%5%*mdihEbBIsjq;Pw|pgki*gfZ<i?Nz^>3F7p~B?5Uc)i3 zpY*uJBvL?>g3xCnfN^>)GTMNAv>>xCOXkn;pS=Oy$Mz&F>I}3q0A7#6WA#G;vU0!5 zuH6U7+B+fw$>lgt4Hp=(tjrzWOn!ZRMTwz=pr>Ta)=9lw{du;M_F2+lZTXup6j~?j zB6jJr_MKmM05AwWMKl^@sbm%GdHJzecJhs6joPum!F+D(qGeCGArH<tg<R8uou+@W zR$GVe$FZ!AK)2p<Tc@Wt9<0xXNW9;w^tlo4R+%{K9s)DZu|EWC<_nI<L?I{A(BnLI z8K;-#=@u#u`nB_y<qGVC&XqEH?7BJF>Flm54s#3iUI&Q%s<YlL_pLkxIA-lvcNiDI z>)0Ao%|){{Oji95k{=8qPqql#H&}lse7->@*g}4hMUiWhZJlmlxNi*n0X71LLEq$5 zL)tac`na+yx})Nr6XE}w)r}KL#iBeupMsT%;^rKmYNf<@kWqgJqL!T}Jw6X?vW-D- zqJ));CZDsLp)X~k%bSm1)~NubFZ~FFCQ-uL6H{p>j9Vq1oT@wM4|Hfa&V7lwT8xRY z!)^X(iuCQCTXDWE%aG?GDteNw??K+8dMavijd1I<3Ac(VJHX#A3TXZzPT3UZ^@l+* z<`*_287zJ~rb)FzvmcUoT`%G-!CJZk^OutMuV<Rlh#m|H3~!Gpeh$&qM1KYh6PvL@ zYIR?fG%%=H>Me)%`dlVogLF)EJr={&C){*=U!vo|-bet!6Pm6|I&@NeZW5V`Dtmz) z<_Cg-!r(=<FCn`$GnJoXLpHXla4N6u6oNd{QZiIfFeLl#RtgB`E#KcZ>+W;WRGK%j zF?wnc$D(^xn!1I2=no@^m%Rlt^1$GQcheK2My0X2x|Pf#&vKUO%Xj4yD{8}<4ET+G z<!06k{r8S5)|ipsU^s6w*FDQ8G$JY3S7js2mOE|(<9d9w$1CvO8zelr^G>3ESsPUo zJ*0h+=v#@;jtl4&9>L63<MxwJpPX&+wkLp7$%*TCqZ|Mr=+*ZjAV$C5%x%c~<tgVS z_bdJ=T_$6+1>G`6&w+G9Q6UvwWLC=iI7#u|^fW=woyT&zCA;={*;(n!hj9eCT49z- zM|0{X(@7rwq604NU|-dEHYL7zD)}T>^r(e7pCDckzg|qPoT>uPd2RDE^f?OMGlNNs z_GAE30m3C+IcMF!8g}lswl%Qcz3_e?x?nTTB$ugaejX{e?y#rTB9MCU;Nk@`)^P0k z*eCJ56^gs|(aHkjtH6{`yf-s6NW1B4rXLcoi^)_^zfH|8J@Br|sIzjtW35?4uMn1; zOHEh4L5b0ht*CuDI;~O`jJD`*DWhAb5TgTW+AD@!WAF_fnlp_lM3||Y6i5|52dfn4 zOiAn>+mJ1n*@?e4OlUy9xS6F-wauT_u8#Wn{0p0o-T-4b*tc_irgTy2XJm)-`~W!W zX5`lX>*|<9D<14MqAAvkSY7F&hR(-2p9B{TNJJ%<#o~D)f%Er72_l9=S^4LhxIMie zrDQhOJI?*EnkILbj+9T$8I9W{nPSiLX?(W^pry&byRq~Qi-n^l8O3zz#x1=`V^AC? zf<AH<p2wD2RY)V==8|9@*?{HCitX%u|2jDdFC=?CBd^u^L~D?(>`4HHylR_<T1sp* z`4hYrN+Zd-aRH0!&!3f%5eJ=*X?cK(DBoI^{U-{#9*F7QFD$u@PB6T3R&k5d7(lKu zoH+t&W{JC84EfyU59Zun7Ae|<E@t!jQT@u~?N*(xaE@t^E#q=rJM-~)T}d3zMftmq zLzBDN_RVO6JziOdA3cODJFY<#PsPZ>(jBNKt_`nWmpx&&KO^SWqhrr2C4ZCMmtaw( zMmb;5qsKgH3Sl}uPfmOBL(uaHs1kZRdNbf)FQ(UH@!f0U&#MRxRKT{Uc)fRVX?FJh zd825yM%TN%vxL@WUR@D)E)X{J?_`aSxy=`EX70o)_Xp?**YhiZI5!bgh6F8y$~@vx z1d%;5W2#=?gsrb%vk>XoC+<((I8Y<Rn@VIqE*SA{(+Wo`gwL*$@hqWYH-a-c@5W<! z*>TiSHt%#TNAwlh<=AP1YRgk`M<V7RN|znOha<dRn%QZk%?z-Sc|UiFr>)^L6{<0s z97A<-8jgu#P6e{@P}~lViR2@6=*4>)Nm^e;rl)Fugvm~fLC>o@-H)!K%LzBoDR==^ zf<+LNUJI+#ah>d@a$hDnV%lS%8cLqepH$C1t5b<GV0z|$U;-!?%6_j}u00^D4^1&h z|58i=?Bj6zkg;RvC(`k{UHDJR2f02RA0s4#%?Dxu#g4fNAb(?NQ$)ROm?`%=DW;ZL z0LmZiqgz|hy>_V~ea7<XeH`a};|XJyVn$+*%uAZ)0&!X&_V{B9n%m@{CEY|E%=a<p z-as;}NAN|4&+yy&I+*?3=*z&+?A}+T1!D~+?4CLB$8zc!f@P}N1DTc>tRmcR+?nuD zAHs{cg`PNfIiB*0B2bB-`G1{csh0fm+44xO4i1wGm&D?mtr^|CpX*@*PEkIG>n!i5 z)<SwW_L1s9+{F8|QBtqzsk>a@d?|oDLpOGV{qVM(o|ja7Oux9XQ{|awKXAFA_9X72 z?cjR7ode|;L!2Xi@G}gS_g*N8l;qy6CM?##NYNk9IZZiU7&r<xy|k3c9wcI44R#Vi z3G~Z^ke}_|Y%LVkyKP-oKk`_xy&JF*DiiSqRtF&4pHhN0QP8BY7{mLxOAfH8DvD%? z1LnSNAJ8AWSDCMlOxak(u~Pxr`zhTLMKV+|mp_9KwBU|2J}$S5V`r3Iyblg*B=`3& zJ=NA~?x5;^r3))5`2tb$EZ$31U*F0Q(SO;sDo$l!*J%Nu`AxHAQ=y(J#xk<sAChH! zr4Q})@C$Sm>2t*HlVFclW*iWM;j++t=2Ou4_=%@dJzLJA1E=CH%XAh3GW12v?Ns`B zZpeqfLB8vGbMhD<H((D?{UI@WJsnRFx^lZE$8yc<dlHa)642;k%J&(oc#O0}^*jG4 zYtuZ<l7;<CbAygqWJ#rg+wnLiakcrmyEa=6g=Kd$+DFWQk}A(S+@MTAiB?QL2EbB{ z<g{J*dTek2F;z=APj&3JJ{Pi{y)0$#pT{v)nc0$_)+5*H_^`(nQG-UIxMSFSYbPUi z)IuMcX$kJcvXW7Jgz|>Nq(h#o3cHmeZ=@Y0di_=9_WZKUTH~EFoue*KVj7`{;!|7@ zGXw)<q}lGPT38+dgaiBZciuvb4_k{Te<mUi5#+S=+(prl&-`5gjN!N54Znp%=;fAA zQ7x-y20NMO@hWEQ?$mA=gdtg7aQy(Ge)Ka>c<-{cPzmXhvkfFg+NTy4=Esd_Q>OA# zDIv^R=)2ke=a-$0=4vG{QyLQd-_dH0O?TOxc-RUI8L;D0o77y(AXP!w$V$6M=pmUB zBeCge1RkntU#Bgd?ac)D#GOvW$VHO%Zo~lzJtqME2kK`OX9Y}M?@HH-o&3h(mIr;S z&Ubqw7v9cWtIa=VD>#J&^fTz?u7-|p$-iIC8XeeL4;oaZFneF%!s6GkiRTv)Rp+yn za42y+>kBFh5l9s`vYK0)6so@Z3rJBuf{rCer|XH8MJtQ4lX@g(yCr55L-UE)x2uB{ zDbx<4Yg~fQO-dJUA8O`paWQ0~lR&XDI6|*&!jSGN7o4FoY?r?;YYHzMy^te?^Hk_p z?<_BW={B`;bGU&3r99PL;}uf!z^;hjshBJLxeMU1A#h)wlS1Kw`UytE3fb&8^>eK| z70;TRfiza|vWG>a;{3H?JQ*YRa=e;x7phz5RZ6dadX&2;Uoz!83~>Xgn=ulERS2*{ zzzIsd?r&Rie7rD>qBbx$U5i79E6b?a4i|l9APIDYtTWn5zmiN^7R!tZRnlweZT~sB zBjL2Qp_w=Z*|U3<#VL&Knkk{izfbpbzAT6o*W@&;(@x`^4)Ww@>f`wB!!@K^Bj$;U zW)Zew&+%<oR?vxex#z?xty@{G^T;?VSGF2l5XaWIi~XpjRAJnfl;HOZRPCh}2;GZz zyL~wlT)~`rOTXZ^Z~@U?S(5r;SH+3*u8ZNl`yJX_!f7b>5OPCwRkP)Mmcv~`YF3>w z>|D$hD<!}Cmy&&C`c(57^wde3JbgSgh78#$$AVeid!5Wh5DTdX4ZOS??37VM7seog z5X_2S(b{sMU0tXFPt+ur{c0J75(>f&aye2qM681>;@7cZSed?ozAOIceK#X}m`3(J zC}<z)93OHt&klyc{7m(~JQxKSel4M*=y1*O=5nrUsdwH5d6S|GRD5*tu3}&9-q`A; zqP27fr4>FiG}|AVVqxiH&Qmw$DCA+>vdQ62U}a0E-YmnUu)(1VUCv%M%A83zJB>4c z_KLi&@2A&vub(O1cQtzw#wFKZ4BxOu<z>pwN;s$8c<tC>xaOl4QA}dpOduVQ|A(-% zw({p5!l5We_O7TPNZ>lS&n8kc#C@RZ2e;eO14AKdoU<u;=Ik{Q>ut~SMynY9TyD)M zN^LklvNW&JI^Sout@>FCbCW3*DUI4VU=2;teJ{;l25WPg#6R{~Pr~GGWBk;zPV`$3 zB~@c|vfDM<-qE>dR68HHDk{X>Ei%(w(1pL#=)s{UfTRUV{Xm(+GnU)L_;`eSdJOWa z153uYHz@;4BGMvh>7Kj%jH}(>g@IiDA^+XtZ1)m(iBK*O5O3IT75k$j&X9)^T=va? zeB4Lfg2WJ!bd;pVCanCq(jb{O0V8n(D_IUCH6l2JJfW;ZfD|cnCBR9_9b&tW8xEV6 zp(lDy6gmk>x5=zIOML)zM_DlHppswWiy=kRZ3K4H<y_8A8)cn8*m-=s_Ic#ZLE%_F z`B>X;Jq|c+UVkGibSIt2JNJicXj{pWC;7*e{yd+k$@{`}Uye*2WrqqjtAkJBb96V2 zw(`BEOip%pcW3lS<jwW~aWmCX<MR-{qvE35!!;n)sCzo90m!;vv^d1jyU`iQ6&w}P zbPU(=>kG55X;s%G-&F-ler_3s$GR`cI!*GsJlh?LvevRP9DG(iYly*$gcxAd{>@)@ zZfin68f1we+ixLiEkYI;?(^2RzqG9!CeqDW-Rth!Hd)e^s?1jJc77=>^`#d+Yi1v% z^r<_q$_eSR!=(cw+do|ZU^o}il=SyS#KWfye1^Lxi*DzFSnri#PTaqiWZobFFVU1U zKK93>a=R8dQCc^|*NtVfO6PmpMJg?EH5kxl!Nv;k4OTkT<w3BI3<*DR=euT9`ir#n zx-o>f88$|iqA}s+f#hWhkdc?2{g?OG&0B$ad%+NT*lC))oIK!y)5fNPXkQDP!k~T@ z;3uvB!xEvA_}lce5Tj&oIKRjn4g08Z9p|qTmgYeSx2FW)mWCG6`8g-^ntdU@MHwql zKOR7kS<BeL!;OGQrM2Q;4LVIYuFds(h<%HDU4#oaKQL|oC~2tkI#U79@Gh)yjJWwe z2Cx(m?)7f7neS`cR?~AT5X3<+i+7GyjwcCv68&}hsZm-k_6*XFf7JnRs*&FDG?U|A zdGuTL!Pf*|7Uo3EP11dTT`87O4`f))kICVJi1rELPy7c?&ye4@O)u@^Wgk4!+6+^2 zLb%TRtO)Bpqt_Wk>%^*DBMfz#*0An&nC&O=8zl=k_xW;?#5bKvKuS>X3H+gE%by99 zM9y4-vO2*FkILGD-vM6KxNNK&;&67vD0?Xeke_NH9V-+VBKsR!uD6o3VBdidXp-|^ z!xynjri)}IZc)5A?e!u0RO8!czXkJXueU^?_75Hqhk}eIm#q@UymeS@_RBsPc5<0H zc#8oAbU<P(h`;%r5V9<T3kvL*LYznC(Rd~wHWzyG3|Dy8xB-^DF2owMlOVyR$)dEQ zE7q-c8cuJkCYbv<RNBCy9Pvxnv*%MLD5<iZ8+eFwNc}*NBwgH8zt_)1pVwt-bmvdg ziH=tbE|`mvotZgHCV?Vp$lgHGr$5SlH_!<bQ=T+V`m`N)=l!{62mzVbDPVNczpD9W zTXQ$CH1lfs8Z${B5MV@sK?=bW-XG;A$Hl8lP(H{HVi~&D)bAO+T3uYFUHjff+4RTl zEEHd0YOrgLg4s7O;B&r{0s|1l76`307o($U&)QrGaw%<{cl9NFyNV}8M#IH{;>&Th zswJiWM~|-M_*MGVdKqftzRJy4z`bh3&c*watgre_4$JTkh}hV1S#W|{07X}QXI2Ma ziyA0B^N~IQo8U?586gN1&gPfurbezYfxrobkiV4ZySS$saz@RZZ`}u|ajCn5E~VuG zrFBW-i|p`*aXpvbSxURLlDHR}2p^!B!_kLuaAg+_<lF#Gdh?tU52&RSAUVU=^or~U zPI!a4;z))$gs?Z%-^?x+<e?@|#9>IC8Zkk9sVZ(~kL$FR*|2FU$`(ALSAY~rx=qn% zTo}^(?oyTo5-G!Tg&1%dHja`6Hb<~L8wL$77DelCk(`<EoH=1lSYh|k=L8DR+C}DC zJf)+G+G<%>92T<t;oke_Re%PF=qv#{6!Z{d69mlX`OI&#KTV-pvL-$t1pU-hRFg0O zv%@b?)Y(a2yGkrE=&DVvtmjn+Ddn6spQ?ot54KggelQ>NBMxO`fjbbJJfL)}q;qmg z%fON`26be=2Dm+#T5jswQWR#~1}!IFQ<^O&Jo7LR5B5Vrx&5^?h=`U+|GcR3t~Vka zwF}~_bD}%bpO^0Jk|wjm@X&l19qVB(5Mgs9_(fA6;7<NsyF=kmm~j)06<*8!?6Oc$ z1&MwpoY;uVm)0igcsuI*NW~Y(6J?&^1WBn!<yY$N=4EIwj<j}KUn*nN4=>w21}6zG zqZx9*zCdi(FUt#j6|bz;%xg4muYUA}VaT_;QCxsutC*%*PgdY61&+^2ZSdsO`K5sJ zY&iDrhc#9_>u^PDp0T)jg#l**?@)?pYO%pQtf4ZrsLZS(lfo*kW)jB~RWgvnFu`+T zP(QhxNBYvV^mbOIrqMoowa?<>&Om!Yh=dn;0Qdt01e}6`4QNwKd1rOCo<;}MUa6x# zDt&OC)#h#bF23yvPwoXu679WaPvl4oR-$n8dd5zQNjhP=f6ed=<*sqy>}m!M{oy_P zvhl-)a1=vFZ(W&DiQ(vcls(l;Kyy~Eo?Ghck%^yF<2(o}C&e#X*s_!`wH0eYhwx(D zYq!n_f}f)lvDp)35bN679%8|^z9575wX@dMIoz56=(uu?UvPqc6M7L(%k+Mi<ieP! z^Y^6BIh)!3zzBY+g<w9Y9hB*3AzfTNR#i78txSsX=H3bNZ8nJbkls+GwD=J-meh6R zOU~wxa=g2R_`)e&8odmu>b^&dPROcry{JEaq>j=NbK4Xd_>I02IsJ{W<VMDs_k+~y zWSJb1_!$wRsf3gmv_+NtHv<u}hlE7z|2VXIOK9N^`ein55-M5z$0){^C(kp;z33&b z=G4gFCFb!)wWe8S!29#1B`I-;8qkvPpKtg!Zb=Q#M`VJ6xs$N)9_|k_YI_wa8;qIP zd(}#!9&A!{{Fn@yM9QwDPm(S03UzI^k`CFC3B-bYnaD;#a}S>^I#<yjooZ*b-|xr- zN;>Yea$#?B?oa$G)=<lot(7UY&fm54+#!Yt`cIgPs9t=qq8(dh49!c+++e$?{-&rL z@zfX<kk<E9vWl&^pccK+?>nZm33L5Sg#MWSpu7hxkCj2K-zoi8-&s*NQ%t!0oL%XW zC2}ZyXm-_7qW?NbjqKjiGFs<|d}Ux6qrX*Z%W)BBN&+AoL6C!O2=jP?B-P5KB%^>2 zs#N}G<l(Y8jelg*(mE=m2Y)CEU`&Jh3A$FM=tN5zew!|+uJ_43_v_EQ*<Bt)N@KIs z{rUvk9v7K@He@&6k1?hpxb%p&B@R9t5Sb8#kXGWrxpz|_H*jeJ1T8x%#t$zO>MqQo z)SDu&qexKh|BO<IJVQxKUHWn%D6NAAnzFgR3DSSxE#@WU%8w1KJW!^Bppk;-UXrAi z?SdXP#8U0yk$1X&ku8WS<W2RdnxCR@B4%s3z)emT2L_!lTPucGapX!`8m?(BDlh1M z28yXQE*x8(gr3tfH`k*N+k9&XKol#qAemeRoofzewV^%Ze%A*O7^9Phsb0^)zUa=j zsM2qcJnG-+gXKJ$YZ`|DB%O{S&G>iKhjvp0&TgE}Bb6-F$;yE1{Dvj{MtaGuEedb@ z>jyL`*U(&u!`|l4RcP9L*kIL)XHVEDND7#TLkLq`Mb4|y(H-9Yk4600R%9HpPwRU5 zn#iP30vV0BnC(Wckt}FNoT16R@vspzaEmN8$^4&J8}CHg&>=Z=_k0wb1@Y~nsn1rL zON-&@u6wJ^DV^NWspR|_A4F6*<`_>nf+8nYX&?ekqSn{MW6bEBKiv?q^#O{W&~_@9 z?J@4hVMmQ|*wE0VKRLX_H{?{z_jZjWu6OqBZWv(HZ2fdvBUD%P9MJg$_&*V;I<`I$ z<$Guw)hp0{wPnC4ZOANyY5!R_sOqRKjE)UfWpi+j)fcn{v4o(fjOvpObj*FW6WHyN zRugRI$U|M3j_%gn+ZJoR8d<t`&OrU#f_<Dr!o4<c^!rY}_MNzu5t%*46rS#Ugdr89 z^NsV5lA5C)J)MfHH=Bl~23(bFiBb*JQU+ETVtqBmDY4kqiR%l@V*Nc85mte+p_-nx zT`}KBUi~&)syI0wY+l^v8~C_RE??6#16?FM{1hc+>#32qT{!AGRKjHl=CWaBdI`R# z11>AxA?A4ZHW_8gSDVmTGZ8G*x0<{3NdF-)mx3;UJ)}Q)*;_8~08>!_uDiGb6d6v< zdCkvx1rc7^4FG#TfqVmljxL2>7RFo9>XX;*=hS|A5&k%2pUydh5Cj<Ruh!>Dklnow z4}{9aq{te5QJ+3pv~x)Z?(R5)zpF9>u<_Eb$IaHJF;wIdfE1r>-izcHI|Md{JtiFS zwbdLyCx@a;zS;@Xhc08ocx&uho?k=*a$!nS;gYCW<9>0K>J0Pb<-d=es=EFetIv&d zOD)F{JKfyQH!BP(^rr*z5GPx~Zjt?#aVV9(?%*RahSI9q`k`yyA2niNz=myMf)UHl z%=5oLAl%j(j7Ohu!%NFm<0*Ikq@Tp;%c0eJ?lyUi%%&s~FSjuTiy=3rqw|_4=bVDh zbUqrR+_NOGpWsrow9BHHN70~G@<wrg=QtP(#AZn_ma<%*)AiHsH(Hu8|HHK0Ue7_) zj_WV`^SEjIxMDl)tk{2vAl9{~kzVvFt#+Anz@B%A(CU%cV%L7@WS|g+uX#mIJ-}8c zN53t^7$8cIPAU)#A5+A>ZZrfW_Ad9B;(*33YhT&mbD)E$B9`4oe3cBGwBG$ZgtBF8 z4VseVz@pC4H7#f;Mh!bo#CBXr+XU3{VMFyBFM(KHFCN!pnClb4X64w0S;5)E;_rJa z(H<X(2Tk!nQ_W>&t8p1*z{!|Jio1M$0JOc7i(1}f`vqb$TbzM!IGsQSh#?Clgpk^= zxZRZ;z2uEn6`)yeLr1_ZR*-~vR}Ty{D)TN#C_gTvpxwWxk6ErzK>oOx74tu%=10qg zP7ddI9PFYXv5Dn#<T}+&YGq@eFtkxHc4AqOfJ!r52jhlL8f-TU^DZ(MM*X=-LAfX< zfFQSyV28>aqnF`?)Qe5&Ck$|9b0o+?&I6HT354$Ajg-Jm9U_yW4g@N&vuB=u;28Qy z1%f>~nvP&wr8`e5Y{<3dTcoL#1%q-Ob=zqH%_OHA$^^tVU0}uhW4@Fq#8!k~Vhy+S z-uYCj#xn^e02Q9$9CxDOYGk}$)W(02@@<CX$p)Pzr;0Zw-X=WJ^91<<Yw0a-c3d5| zy?D-ak)Ho374K0F^Kug`)!{3}jzdn@<B%rbx67Ma0=#j9JO*X~5?loKTkr;L72Ei3 z%}&h5nvTD6x}P;I6QGfbT@XKeJmq5!BB<=~!G!=l@-08Q%xUjOWBn)OnEIL=eo2`Y z>XNmpMg$wE>mrjjA5PD;(ckUW`~*O!#=s`4{WA5T8#(X2Myn*0IS$wJ+(REDsDnCZ zS(9<fEP-N5EZW1fiR)`?tIyCBJvN~PTyi-FnnHVcKZjHf74$6~=nzi$yRHFo*5t>X z)P)N&KY58`3otqtzGQ8OM$!`JARe0)#Zo%|gHmdPThPSAFH&g$FUW=GpoF*+II9R$ zNQq^&_^n8n+4y79d8PTe8~Px6HVr_3wP8A%lRU86^B^W_O3V-6?<>3B;0@lv_pH)_ zD_;Lx1>*FKyuoXm#Gw=et`ACuXJW0<=cJqb`sVN4Cv9M?;J=*=(R#uOvsv9k1sWAb zMM6FT0DxcR_}ipsPY*OM$=`NdvjRQYwZ$Z#4W{$G;b0u^&qY1?ycv#4aP^?6wQs)Z zI{G_<5X!w5+u|RFX88+1*#d-e&=5&>&A3aRk)~Wdqbd>kAteQ{K!=g?Bl(A!o5|rm zE7D+Z?QXEc^C*{ZtpKc{+KDlh9ds`XVMN$OB(f7gkkj>KRHxF@58)J2;>og4-S#>S zER*?}N<OF83|cF}CFxP&OA>6ZZX%LEIMCLzAMNs6a(q`v82#N>vL(AN@V$=Uk?~`R zaFGk;=jY9T#UI|xaYbk=@FDwsE#pXvMtG*xQn(D=p_mKm^aX091X4A2YPP=)<D>oG zwk_;X>slAN6c;(`hp&Zw*r{@%@3w8vm);GMy!jH%D-;(Ji(@Gx_)PbU)r(e)hB0^Z zp_8J8$t_{w8;p{VVBta5`}u&9YI0Z;T~E{-6h?vIJUq?fv+u{A53c(V<>hynDA14$ zOYL>CEHskPw2`Na1xlD*fu0fy-xp?A8R~wOgTILWRr!#wY>BFiLyE^{G`G{e8E%g? zrQ9pojgeMg>MhYxm;yUgd6f7_IW^2L`?`ZU%*|JP1H>r9Z;6+rBPG&vgpA(f;$0B6 zEh&QU1+6bFxq6@1YYd-+k;Jr)uodCN{A16oYaP15R~3&Qf90vQ$8ZP*cm~p4X1e=y zF?xu&wBq{5dv6<E9t#}QzvBtRRrw5L0VMl2`yQi(+X=ssI=O?v-qR{;gma;)yhw4h zvQHK)NTN^%K>VR8+bJt9GBw{mnV^DoKP`;zS62Zj3}x_N@&t1#^BkH{A*L}5bN#eB zVGN9j?=M;B0-g#zLO##x#I990Nq%AYagJW1LiNd>9hz9T!<Wi)p#>?gICC^$|0H~H zD4@fIW*B1h_de_IF~Z@RKc%cQC50HHa|yRK9nQ?9AiHgD$*5u*40YQKb@Ei7$5nX% zfh(9ij^^aM@jCQvZmC@4;e(-a_y*PxDU>4=y2EhK&!CP0x`8F)r=n)~UxBhDp2NIN zRI3m_O|O=3Yv_vM1pXW7%7lZRdT&ty5j~};6sE;ojthu@B_$7NQV1=AXUG9=(k@9} zsL|*p#9{K6ZbUAYKC*wcBs=^Nj~udzQsC~BA`zXr!IHVT_sYmA^Kb=i!=>h#O8qkx z+|`iLRr!K4I59VpCf6>#iDt34N5DJH1aRvO^C6i=V56pX;H(E7M<@-f<s0Csg9G@I zve#J3lT~F$tw)>PNCF>xa)rJ~;dyPJ%u*0dA<rrE!qquDc1Z9^TOdZ{-Iw&X`@`Yv zZ~>`gbToGvr0;J`X|;XDLa*9~ilIN8k4;t4$})%Fh3aqI01nn5V{Ba_xcT_0@kkuV z0*>pjDfi^$Zw+9{jG)L^VNH9erTrD+M)UKSdmsGflv3(ncB0Mz*gS}JfZWX1IVA~R zy;}BBq=gr1B5-qNXr^ca?;=dlN&{IX8xOq6zl7&$a9xbOxGx<W=ad#Tlg-1~ihKSw zM(Y<_FF%}2>V%UB>p!Zf9dCtvW#+jJRri05!T9a0MAshk+`5Wx1RnW#0rZWQY~BNG zY^W$J|Lv+XS1BVX$wz0Q^!r{uaQ`nACHutqeka*~^`VxbV_;ep)_*>ND=q(8_B$3^ zA-c08IHlX4Kx5e0Ua*;eQsrr0l4@~Q;A6gE(-IKG@R9Woa1NlR12&~pja<TRia0{i z{Vc?_7mNbF-{|$2WKa<^PpGCV(<!3Aqbr&UdW&)|hXEs!mk6sC!xQ0@Hm{&JtQ1BT zE~|`%U562H{(+1T)UT0*3H*~$A=>eag0k?p@W^#(4*b6?b>H|XQik~c`|pan|Nh%0 zu+s1U2CjQ<zQ2IQe}75+{RG`xO@A*yP~Kk;|9<|5#{c(U|4Uu}dk_A9>RRfb!T)!5 z{a+&2zc=dt4_*Iz5B@*A<o~>}?>|KS&wu?l`R9LY{NKr%|8)J&oP1#`NKB_N2}=Df RyyN?k%fXai6hn>t{|9>E-&z0w literal 0 HcmV?d00001 diff --git a/knowledge-base/022-numba-jit-numerical-computation.md b/knowledge-base/022-numba-jit-numerical-computation.md index 31118df..612ac96 100644 --- a/knowledge-base/022-numba-jit-numerical-computation.md +++ b/knowledge-base/022-numba-jit-numerical-computation.md @@ -75,4 +75,4 @@ def monte_carlo_parallel(n, arr): ## Environment -Python 3.11, Numba 0.58, NumPy 1.25. Numba requires LLVM (bundled via `llvmlite`). AMD Ryzen 9 5900X (12 cores). `pip install numba`. +Python 3.11, Numba 0.58, NumPy 1.25. Numba requires LLVM (bundled via `llvmlite`). AMD Ryzen 9 5900X (12 cores). Install Numba as described in its documentation. diff --git a/knowledge-base/033-cython-c-speed-hot-loops.md b/knowledge-base/033-cython-c-speed-hot-loops.md index 98f8c2e..19c1b5b 100644 --- a/knowledge-base/033-cython-c-speed-hot-loops.md +++ b/knowledge-base/033-cython-c-speed-hot-loops.md @@ -69,4 +69,4 @@ setup(ext_modules=cythonize("distance.pyx"), include_dirs=[np.get_include()]) ## Environment -Python 3.8+, Cython 3.0+, NumPy. Requires a C compiler (GCC/Clang/MSVC). `pip install cython`. Build with `python setup.py build_ext --inplace` or use `pyximport` for development. +Python 3.8+, Cython 3.0+, NumPy. Requires a C compiler (GCC/Clang/MSVC). Install Cython as described in its documentation. Build with `python setup.py build_ext --inplace` or use `pyximport` for development. diff --git a/scripts/build_claude_plugin.py b/scripts/build_claude_plugin.py new file mode 100644 index 0000000..9d6e16c --- /dev/null +++ b/scripts/build_claude_plugin.py @@ -0,0 +1,271 @@ +#!/usr/bin/env python3 +"""Build claude-plugin/, the edition of AAE for Anthropic's plugin directory. + +The directory accepts only Markdown, text, JSON, YAML and images from plugins of +organizations that are not Verified, no symbolic links, and no commands that +download code. This script derives such an edition from the sources of truth: + + AAE.md -> claude-plugin/skills/aae/SKILL.md (adapted, see EDITS) + templates/report/* -> claude-plugin/skills/aae/report-template.md + (the LaTeX skeleton plus the generated files of an + example session, produced by build_report.py) + knowledge-base/*.md -> claude-plugin/skills/aae/knowledge-base/ + img/icon.png -> claude-plugin/.claude-plugin/icon.png + +Run it after changing any of these and commit the result: + + python3 scripts/build_claude_plugin.py # rewrite claude-plugin/ + python3 scripts/build_claude_plugin.py --check # exit 1 if claude-plugin/ is out of date + +Every edit asserts that its anchor text exists, so a change in AAE.md that would +silently break the edition makes this script fail instead. +""" +import json +import re +import shutil +import subprocess +import sys +import tempfile +from pathlib import Path + +ROOT = Path(__file__).resolve().parent.parent +OUT = ROOT / "claude-plugin" +SKILL = OUT / "skills" / "aae" + +PLUGIN_JSON = { + "name": "aae", + "displayName": "Agentic Algorithm Engineering", + "version": "1.12.0", + "description": "Agentic Algorithm Engineering: autonomous performance tuning through falsifiable " + "hypotheses, benchmarks, numbered patches, and a PDF report.", + "author": {"name": "Christian Schulz"}, + "homepage": "https://chszlab.github.io/AgenticAlgorithmEngineering/", + "repository": "https://github.com/CHSZLab/AgenticAlgorithmEngineering", + "license": "MIT", + "icon": ".claude-plugin/icon.png", + "keywords": ["performance", "optimization", "algorithm engineering", "benchmarking", "profiling"], +} + +REPORT_INTRO = """\ +The report is a LaTeX document, `report/report.tex`, that inputs files from `report/generated/`. +You write both yourself: copy the skeleton below once during setup, and rewrite the generated +files from `report/session.json` and `report/experiments/NNN/{meta.json,change.patch}` after +every experiment. Then compile in `report/` with `latexmk -pdf -interaction=nonstopmode report.tex` +(or run `pdflatex` twice). This edition of AAE contains no report generator program; the +example below shows exactly what each generated file looks like for a session with three +experiments, so follow its structure and macros. + +Rules for the generated files: + +- Escape LaTeX special characters in all text from `meta.json` (`\\ & % $ # _ { } ~ ^`); + text in backticks becomes `\\texttt{...}`, text in `$...$` is passed through as math. +- `metric` defaults to the geometric mean of `measurements`. Compare every experiment with + the baseline (experiment 001) and with the best value accepted before it, respecting + `direction` (`lower` or `higher`). +- `trajectory.tex` plots `generated/traj_*.dat`: one data file per series (best so far as a + step line, baseline, kept with the patch number as label, discarded); each file has a header + line and one row per point. +- `experiments.tex` has one page per experiment, kept, discarded and crashed alike. The code + change is read from `generated/diffs/NNN.diff` (the experiment's `change.patch` without the + commit header); shorten very long diffs so that the page stays one page. +- `accepted.tex` lists the kept changes in acceptance order with their patch files. +""" + + +def edit(text, old, new, count=1): + if text.count(old) != count: + raise SystemExit(f"build_claude_plugin: anchor found {text.count(old)} times, expected {count}:\n{old[:200]}") + return text.replace(old, new) + + +def cut_section(text, heading, next_heading): + """Remove from `heading` up to (not including) `next_heading`.""" + a = text.find(heading) + b = text.find(next_heading, a + 1) + if a < 0 or b < 0: + raise SystemExit(f"build_claude_plugin: section {heading!r} not found") + return text[:a] + text[b:] + + +def skill_md(): + t = (ROOT / "AAE.md").read_text() + + t = edit(t, "or when the user invokes $aae in Codex or /aae in Claude Code.", + "or when the user invokes /aae:aae in Claude Code.") + t = cut_section(t, "## Codex sessions\n", "## Setup\n") + + t = edit(t, "by opening a pull request to the AAE repository and updating it every 10 experiments (see [Contributing Back](#contributing-back))", + "by writing a knowledge-base entry to `report/kb-entry.md` and updating it every 10 experiments, which the user can submit to the AAE repository (see [Contributing Back](#contributing-back))") + + old = t[t.index("**Template.**"):t.index("**Session data.**")] + t = t.replace(old, "**Template.** This skill ships the report template in `report-template.md`, next to this file. " + "During setup, create `report/experiments/` and copy the LaTeX skeleton from `report-template.md` " + "into `report/report.tex`.\n\n") + + old = t[t.index("**Build.** After every experiment"):] + old = old[:old.index("```\n", old.index("```bash") + 7) + 4] + t = t.replace(old, "**Build.** After every experiment (after logging), rewrite the files in `report/generated/` " + "as described in `report-template.md` and compile the report:\n\n" + "```bash\n(cd report && latexmk -pdf -interaction=nonstopmode report.tex > /dev/null 2>&1) || true\n```\n") + t = edit(t, "This writes `report/report.pdf`.", "This writes `report/report.pdf`. If `latexmk` is not available, run `pdflatex` twice instead, or keep the sources and tell the user that the PDF could not be built.") + t = edit(t, "then rebuild the report with `python3 report/build_report.py report`.", + "then rebuild the report (see [Experiment Report](#experiment-report)).") + t = edit(t, "if `metric` is omitted, the generator uses their geometric mean", + "if `metric` is omitted, use their geometric mean") + t = edit(t, "The generator compares every experiment against the baseline", + "The report compares every experiment against the baseline") + + t = edit(t, "open or update the session's knowledge-base PR (see [Contributing Back](#contributing-back))", + "write or update `report/kb-entry.md` (see [Contributing Back](#contributing-back))") + + old = t[t.index("This methodology is supported by a community-contributed knowledge base"):t.index("### Contributing Back")] + t = t.replace(old, """This methodology is supported by a community-contributed knowledge base of optimization techniques, experiment results, and lessons learned from AAE sessions. A copy is bundled with this skill in `knowledge-base/`, next to this file; the latest version is at https://github.com/CHSZLab/AgenticAlgorithmEngineering/tree/main/knowledge-base + +Before or during a tuning session, consult the knowledge base for techniques relevant to your problem: + +1. Read `knowledge-base/INDEX.md` and scan titles, problem domains, and key techniques. +2. Read the individual entries that seem relevant. +3. Use the insights to inform hypothesis generation in the AE cycle. + +""") + old = t[t.index("### Contributing Back"):t.index("## Philosophical Foundation")] + t = t.replace(old, """### Contributing Back + +Permission is asked once, during [Setup](#setup) (step 8), and stored as `kb_contributions` in `report/session.json`. Never write a contribution unless it is `true`. + +If permission was given, contribute every 10 experiments (loop step 12). Each session has **one** entry, which grows over the session: + +1. **First contribution** (experiment 010): write `report/kb-entry.md` following the entry template in `knowledge-base/CONTRIBUTING.md`, and record `kb_entry` in `report/session.json`. +2. **Later contributions** (experiments 020, 030, ...): update the same file with what the last 10 experiments taught (new techniques that worked, dead ends worth recording, updated baseline vs. best numbers). +3. **Nothing new**: if the last 10 experiments produced no transferable insight, skip the contribution and note that in the report summary. + +At the end of the session, tell the user where the entry is and that they can submit it as a pull request to https://github.com/CHSZLab/AgenticAlgorithmEngineering (new file in `knowledge-base/` plus a row in `INDEX.md`). Only share transferable insights, and never include proprietary code, data, or paths from the user's project beyond what the entry needs. + +""") + for banned in ("curl ", "build_report.py", "gh repo fork", "gh pr create", "$aae"): + if banned in t: + raise SystemExit(f"build_claude_plugin: {banned!r} is still in SKILL.md") + return t + + +def report_template_md(): + skeleton = (ROOT / "templates" / "report" / "report.tex").read_text() + with tempfile.TemporaryDirectory() as d: + d = Path(d) + (d / "experiments").mkdir() + (d / "session.json").write_text(json.dumps(EXAMPLE_SESSION, indent=1)) + for m, diff in EXAMPLE_EXPERIMENTS: + e = d / "experiments" / f"{m['exp']:03d}" + e.mkdir() + (e / "meta.json").write_text(json.dumps(m, indent=1)) + (e / "change.patch").write_text(diff) + shutil.copy(ROOT / "templates" / "report" / "report.tex", d / "report.tex") + # run the generator without LaTeX: only the generated sources are needed here + env_path = "/nonexistent" + subprocess.run([sys.executable, str(ROOT / "templates" / "report" / "build_report.py"), str(d)], + check=False, capture_output=True, env={"PATH": env_path}) + gen = d / "generated" + parts = ["# AAE report template\n", REPORT_INTRO, + "\n## Skeleton: `report/report.tex`\n\n```latex\n" + skeleton.rstrip() + "\n```\n", + "\n## Example session data\n\n`report/session.json`:\n\n```json\n" + json.dumps(EXAMPLE_SESSION, indent=1) + "\n```\n"] + for m, diff in EXAMPLE_EXPERIMENTS: + parts.append(f"\n`report/experiments/{m['exp']:03d}/meta.json`:\n\n```json\n{json.dumps(m, indent=1)}\n```\n") + if diff: + parts.append(f"\n`report/experiments/{m['exp']:03d}/change.patch`:\n\n```diff\n{diff.rstrip()}\n```\n") + parts.append("\n## Generated files for this example\n") + for f in ["session.tex", "summary.tex", "trajectory.tex", "traj_best.dat", "traj_baseline.dat", + "traj_keep.dat", "traj_discard.dat", "accepted.tex", "experiments.tex"]: + lang = "latex" if f.endswith(".tex") else "text" + parts.append(f"\n`report/generated/{f}`:\n\n```{lang}\n{(gen / f).read_text().rstrip()}\n```\n") + for f in sorted((gen / "diffs").glob("*.diff")): + parts.append(f"\n`report/generated/diffs/{f.name}`:\n\n```diff\n{f.read_text().rstrip()}\n```\n") + return "".join(parts) + + +EXAMPLE_SESSION = { + "title": "AAE session: example solver", + "subtitle": "Running time of the solver on 2 inputs, branch aae/example", + "target": "src/solver.cpp", "metric": "time", "unit": "s", "direction": "lower", + "resource": "peak memory [GB]", "machine": "8 cores, 32 GB RAM", "branch": "aae/example", + "date": "2026-11-21", + "summary": "Two hypotheses tested: the bucket queue was kept, the cache-aligned layout was discarded.", +} +EXAMPLE_EXPERIMENTS = [ + ({"exp": 1, "title": "Baseline", "commit": "a1b2c3d", "hypothesis": "Baseline, unmodified code.", + "intuition": "Reference point.", "prediction": "-", "status": "baseline", "metric": 2.572, "resource": 1.2, + "measurements": [{"input": "rgg_n_2_20", "value": 4.1}, {"input": "road_eu", "value": 4.7}], + "verdict": "Baseline established.", "reflection": "Profile shows 60% of the time in the priority queue."}, ""), + ({"exp": 2, "title": "Bucket queue in refinement", "commit": "b2c3d4e", + "hypothesis": "Replacing the binary heap by a bucket queue reduces time by ~15% because keys are small integers.", + "intuition": "Most time is spent in heap operations with integer gains.", + "prediction": "-10% to -20%; could regress if gains are spread widely.", "status": "keep", "metric": 2.181, + "resource": 1.2, "measurements": [{"input": "rgg_n_2_20", "value": 3.483}, {"input": "road_eu", "value": 3.976}], + "verdict": "Confirmed: -15.2% geometric mean, matching the prediction.", + "reflection": "The queue is no longer the bottleneck; next look at memory layout.", + "patch": "patches/0001-bucket-queue-refinement.patch"}, + "diff --git a/src/solver.cpp b/src/solver.cpp\n--- a/src/solver.cpp\n+++ b/src/solver.cpp\n" + "@@ -10,3 +10,3 @@\n- BinaryHeap q;\n+ BucketQueue q(max_gain);\n"), + ({"exp": 3, "title": "Cache-aligned node array", "commit": "c3d4e5f", + "hypothesis": "Aligning node records to 64 bytes reduces time by ~5%.", "intuition": "Records straddle cache lines.", + "prediction": "-3% to -7%.", "status": "discard", "metric": 2.203, "resource": 1.4, + "measurements": [{"input": "rgg_n_2_20", "value": 3.51}, {"input": "road_eu", "value": 4.02}], + "verdict": "Refuted: +1.0%, more memory.", "reflection": "Padding increases the working set more than it saves."}, + "diff --git a/src/node.h b/src/node.h\n--- a/src/node.h\n+++ b/src/node.h\n" + "@@ -1,3 +1,3 @@\n-struct Node {\n+struct alignas(64) Node {\n"), +] + +README = """# AAE for the Claude plugin directory + +This folder is the edition of [Agentic Algorithm Engineering](https://github.com/CHSZLab/AgenticAlgorithmEngineering) +for Anthropic's plugin directory. It is generated by `scripts/build_claude_plugin.py` from `AAE.md`, +the knowledge base and the report template; do not edit it by hand. + +Compared with the plugin at the repository root, this edition ships no programs: the agent writes +the LaTeX report itself from the template in `skills/aae/report-template.md`, and knowledge-base +contributions are prepared as a file for the user to submit instead of being opened as pull requests. + +Homepage: https://chszlab.github.io/AgenticAlgorithmEngineering/ +""" + + +def build(dest): + if dest.exists(): + shutil.rmtree(dest) + skill = dest / "skills" / "aae" + (dest / ".claude-plugin").mkdir(parents=True) + (skill / "knowledge-base").mkdir(parents=True) + (dest / ".claude-plugin" / "plugin.json").write_text(json.dumps(PLUGIN_JSON, indent=2, ensure_ascii=False) + "\n") + shutil.copy(ROOT / "img" / "icon.png", dest / ".claude-plugin" / "icon.png") + (dest / "README.md").write_text(README) + shutil.copy(ROOT / "LICENSE", dest / "LICENSE") + (skill / "SKILL.md").write_text(skill_md()) + (skill / "report-template.md").write_text(report_template_md()) + for f in sorted((ROOT / "knowledge-base").glob("*.md")): + shutil.copy(f, skill / "knowledge-base" / f.name) + # the directory holds or refuses plugins whose text downloads or installs code + banned = re.compile(r"curl |wget |pip3? install|apt(-get)? install|brew install|npm (i|install) |npx |" + r"git clone|gh (repo|pr|auth|api) |conda install|cargo install|go install|^#!", re.M) + for f in dest.rglob("*.md"): + hit = banned.search(f.read_text()) + if hit: + raise SystemExit(f"build_claude_plugin: {f.relative_to(dest)} contains {hit.group(0)!r}; " + "point to the tool's own install instructions instead") + + +def main(): + if "--check" in sys.argv: + with tempfile.TemporaryDirectory() as d: + tmp = Path(d) / "claude-plugin" + build(tmp) + diff = subprocess.run(["diff", "-r", str(tmp), str(OUT)], capture_output=True, text=True) + if diff.returncode != 0: + print("claude-plugin/ is out of date; run python3 scripts/build_claude_plugin.py\n" + diff.stdout[:3000]) + sys.exit(1) + print("claude-plugin/ is up to date") + return + build(OUT) + print(f"wrote {OUT}") + + +if __name__ == "__main__": + main()