g<ddlZddlZddlZddlmZddlZddlZddlZ ddl mZm Z m Z m Z mZddlmZmZddlmZddl mZmZddlmZmZmZmZmZmZmZmZmZddl m!Z!dd l"m#Z#dd l$m%Z%d Z&Gd d e'Z(dZ)Gdde'Z*dZ+dZ,dZ-dZ.dZ/dS)N)add)typesirrewritesconfigir_utils) infer_globalAbstractTemplate) signature)utilstyping) get_call_table mk_unique_varcompile_to_numba_irreplace_arg_nodesguard find_callnamerequire find_constGuardException)NumbaValueError)OPERATORS_TO_BUILTINS) numpy_supportc|dkr||z S|SNr)dim_size index_consts R/granary/edev/quantum/lib/python3.11/site-packages/numba/stencils/stencilparfor.py_compute_last_indr sQ+%%cDeZdZdZdZdZdZdZdZdZ dZ d Z d S) StencilPassch||_||_||_||_||_||_||_dSN)func_irtypemap calltypesarray_analysis typingctx targetctxflags)selfr&r'r(r)r*r+r,s r__init__zStencilPass.__init__"s8  ","" r!c Pddlm}tjj\}}g}i}|D]4\}}|D],}t ||r|||||<-5|sdSjjD]\} } ttt| j D]Z\} t tj rt jtjrjjdkr|jjj|vrht'jj} fdt+t-jjD} jj}t1fd|D}|D]+}t |t2jrt7d,| d}|jjj}t;|j|| j| j | j!j"\}}}|j#d d}$| ||||j%||| }| j d| |z| j | d zdz| _ t tj rrt jtjrSjjdkrCtMtNjjd krtj(dj _\dS) zP Finds all calls to StencilFuncs in the IR and converts them to parfor. r) StencilFuncNcallc6i|]}|jj|Sr)valueargs).0istmts r z#StencilPass.run..Is6"Q"Q"QA!TZ_Q%7"Q"Q"Qr!c3<K|]}j|jVdSr%)r'name)r5r6r-s r z"StencilPass.run..Ls,'N'N QV(<'N'N'N'N'N'Nr!zITuple parameters not supported for stencil kernels in parallel=True mode.out index_offsets)stencilnumba))numba.stencils.stencilr0rr&blocksitems isinstanceappendreversedlist enumeratebodyrAssignr3Expropfuncr:dictkwsrangelenr4tupler BaseTuple ValueErrorgetget_stencil_irr*scopelocr'r(options_mk_stencil_parfortargetrrConst)r-r0 call_table_ stencil_calls stencil_dict call_varname call_listone_calllabelblockr6rO input_dictin_args arg_typemaparg_typeout_arrsf stencil_irrtarg_to_arr_dictr= gen_nodesr7s` @rrunzStencilPass.run,sZ 766666't|':;; A  '1'7'7'9'9 : : #L)% : :h 44:!((66619L.  :   F!L/5577' 7' 7LE5#D5:)>)>$?$?@@& 7& 74tRY//$7&tz27;;$7 JMV33 JO0MAAtz~..C"Q"Q"Q"Q49#djo:N:N4O4O"Q"Q"QJ"joG"''N'N'N'Ng'N'N'N"N"NK$/NN%h@@N",.M#N#NNN"ggennG&djo&:;B6DR NK!KJ L$.7:7:3JO%'JNN?D$I$IM $ 7 7w& t{B+!-!-I"'BQB)!;ej1>N!NEJJ ry117&tz27;;7 JMV33!-tzJJ';<<"$!TX!6!6DJM& 7' 7' 7r!c T|D]\}}|j}|j}g}|jD]} t | t jr|} t | t jr/t | j t j r| j j dksJ| t j| j j ||| t j ||| | ||_dS)z Find return statements in the IR and replace them with a SetItem call of the value "returned" by the kernel into the result array. Returns the block labels that contained return statements. castN)rCrWrXrIrDrReturnpoprJr3rKrLrEJump) r-rBexit_value_varparfor_body_exit_labelrdrerWrXnew_bodyr7 prev_stmts rreplace_return_with_setitemz'StencilPass.replace_return_with_setitemhs#LLNN " "LE5KE)CH * *dBI.. * ( I&y")<<:&y@@:%O.&8888OOBIio.C^UX$Y$YZZZOOBG,BC$H$HIIIIOOD))))!EJJ! " "r!c ,"STg} |j} tjdkr*td|||||| t j| |d} j| j} t j| j\}}t j | }t j | ||jj tjdkr#tdt j| t j | j j|jtjdkr#tdt j| j| jjS| j}| j}g}t%SD]S}t'j|t+d|}t,jj|j<||T|||||| \}}tjdkrCtdtd|td |t j| g}j|}|| }St;|ksJg}g}t%SD]}||||| ||}||| ||}|||||t@j!j"#||||dtI| %dz}t'j&||| |<t'j|t+d |} |j'j| j<g}!Sdkr |d}"nt'j|t+d |}"t,j()t,jSj|"j<t&j*+||}#t'j,|#|"|}$|!|$tjdkr#td t j| t'j&||}%ĉj| j} t jd}&t'j||&|}'t&j*-| d|}(t,j()t,j| jj|&<|%j./t'j,|(|'|gt jd})t'j||)|Td|j0vrq|j0d}*tbj22|*}+j34|+|j'stkd|'|*},n|'d},t'j6|,|}-|j'j|)<|%j./t'j,|-T|gt jd}.t'j||.|t@j7jj89|j'| j| j:jj<t'j|t+d|}/t,j;<tzj|/j<t'j>dtz|}0t'j,|0|/|}1|%j.|1tj@|j'jAjB}2|2dkrd}2t&j*-|/|2|}3t'j|t+d|}4t,jCD|j'j|4j<t'j,|3|4|}5|%j.|5t jEdtzjF|'|4gj3jj }6t'j|t+d|}7t'j,t'j6d ||7|}8|6|8t,jGj|7j<t'j|t+d|}9t'j,t'j6d||9|}:|6|:t,jj|9j<t'j|t+d|};j3Ht}<|<j|;j<t'j>dt|}=t'j,|=|;|}>|6|>j3J||6|>STfd"}Bt%| jD]F}C|Ag| jz}D|Ag| jz}E|B|<|C|||;|6||D|9d# |B|<|C|||;|6||E||Cd$ G|M||%j./|6nd|j0vr|j0d}*tbj22|*}+j34|+|j'sd}Ft|Ft'j|t+d%|}Aj3Ht}<|<j|Aj<t'j>dt|}=t'j,|=|A|}G|%j.|Gj3J||%j.|>t'j6|'|*|}Jt'j|t+d'|}K|j'j|Kj<t'j,|J|K|}L|%j.|Lt'jOtd d |I|K|}M|%j.|Mtt,jGjjj|Ijjjj'}?|?j |M<Q| | |tjdkr#td(t j| t'jR|"| |}Ntt,jGjjj|"jjjj'j |N<| |j./|!| |j.|Nt'jSd)d*}Ot'j|t+d'|O}Pt'j,t'j6d|O+|P|O}L| |j.|L| |j.t'jT|P|Ot jU| } | tI| %j.Vtjdkr#td,t j| d-||gf}Qt@j!j"W||%| ||"||QjX}R| |R| t'j,||| S).z> Converts a set of stencil kernel blocks to a parfor. r>rZrz#stencil_blocks after copy_propagatez'stencil_blocks after removing dead codez$parfor_index_varz-stencil_blocks after replace stencil accesseszstart_lengths:z end_lengths:z$parfor_exit_valuez$parfor_index_tuple_varz.stencil_blocks after creating parfor index varN in_arr_shapeshapezero_valcvalz-cval type does not match stencil return type.stencil_outputz $np_g_varnpboolbool_z $np_attr_attremptyz $none_varz$zero_index_varz$slice_func_varslicerrMr4rOrX$slicec j|tjfdzi} ||} t | t t jfsJt j|td|} tjj | j <t | t r*t j t j | || |} nt j | | |} | | t j|| r|| fn| |fd|}| j|<t j|td|}tjj |j <t j |||}| ||||<t j|td|}tjtjj |j <t j||}t j |||}| |t j||}t-tjj j j |j j j jj|<| |dS)Nrz $border_indrrrz$border_index_tuple_var)r*resolve_function_typerintprDintrVarrr'r:rJr\rErKr1r( slice2_type containersUniTuple build_tupleSetItemr nonedtype) slice_fn_tydimrWrXslice_func_varstmts border_indsborder_tuple_items other_arg other_firstsigsisi_var si_assignslice_callexprborder_slice_var slice_assignborder_ind_var tuple_call tuple_assign setitem_callndimsrjr-zero_vars r handle_borderz5StencilPass._mk_stencil_parfor..handle_borderAs~n::ZMA%!%!"sBFm44555}]'C'CSII,1J V[)b#&&; " "(2s*;*;VS I III " "fc : :I Y'''!#'0;T)V,,&)AT "."" 25~.#%6%x1H1H##N#N 6;6G -23!y9I3OO  \****:"3'!#}-0/0/03"5"5494D4M4M%u5.5. ^01W001CSII !y^SII  \***!z'>8SQQ /805 DL > >   0 0 0\&+. "1.$,OO!4^DD%    L N      !Q & & 7 8 8 8   0 0 0^T\-CZ!\ + + +  !Q & & ; < < <   0 0 0 V[). j u + +A}#(%(%&)++J,1JDL )   z * * * *%)%C%C g}l &&" {  !Q & & A B B B "M 2 2 2 .+ . . .   0 0 0 '55e<< $..v66,------  u 9 9A112B12E(3A 5#OOH33(5a(8)UCQQI   i ( ( (   X & & &   U]1::;q> )8Q88 9 9 9 9 "%^%8%8%:%:!;!;a!?13%1E1E-.}5I'J'JCPP,7,= ^() A::(^NNVE=),+,+,/11N050@0I0I E1#1#DL, -,,[#>>J9ZEEL  $ $\ 2 2 2  !Q & & B C C C   0 0 0XeS)) ?fk2J!/??Juj#66IGOOFGSAAM','7'@'@?I(P(PDL $ O " "BImY$L$L#M N N N .z::IveY44H---#+F3 -..t44~11';;LMMV$%TUUU#))$//#))!,,%--J&1&7DL # O " "BIj(C$H$H#I J J J,-=>>GfUGS11G).)9)B)H)H;F;L;E?;E;L*N*NDL & VE=+E+EsKKN05 0A0A"0E0EDL, -4S11J " *nc J J  O " "#4 5 5 5,5*0 2 226x 6))#* !#AQSV!W!W VE=+I+I3OON050K0KKL]0^0^DL, - " *'B'+~'+|'+~ 77Eve];%?%?EEH)BHT3$7$73GGK LL % % %*/*DL 'VE=9J+K+KSQQN " "(1c*:*:NC P P  LL* + + +05 DL, - VE=9J+K+KSQQN.;;EBBK0;DL, -i44G9WncBBL LL & & &.66{8= }q7H79;;C W\\~08(/C.0.1*33N.1DN> *umH&=&=sCCI+0+3GG &&|444"$+*;*;D*A*A3!G!G!#}]/K/K,/"1"14?4E ^01$&In.>3OO '0(- DL4N(, ^5H(I(, W\(B(H(*(*|$ -.3::;LMMM-.3::<HHHF0"55 umM&B&BINN Ibhqi&@&@&@-QZ[[-.3::;LMMM-.3:: ImY / /   ".~>>s>..001127;;===  !Q & & 7 8 8 8   0 0 0}k:;%,,Y N%(.)WdjZZ   7FC88999r!cN|}|dkrtj|td|}tj|j|j<t|tj r*tj tj ||||}ntj |||}| |tj|td|}tj|j|j<tj|td|} tjt} tj| } | |j| j<tjd| |} tj | | |} | | tj| ||gd|}| |jtjtjgi|j|<tj |||}| ||S)Nrstencil_const_varrcompute_last_ind_varr r)rrrrrr'r:rDnumbersNumberrJr\rEr@njitr r DispatcherrrKr1 get_call_typer*r()r-r end_lengthrorWrXrr const_assigng_var check_funcfunc_typg_objg_assign index_call index_assigns rrz!StencilPass._get_stencil_last_inds ??& 6I(J(JHKMMK-2ZDL) **gn55 G!y*c)B)B8CS J J  "y[#FF   \ * * *ve]:%>%>DDH*/*DL 'F5-0F"G"GMME$566J11*==H'/DL $I1:sCCEys33H   X & & &eh -Db#NNJ)1)?)?UZ 8"*>*>DN: &9Z3??L   \ * * *r!c t|trtt|dSd}t |i|j|jtjf|j |j }t|j dksJ|j d}t||g||jddz }|jdjj}|S)Nrc<tt|dSr)absmin)s_lengths r get_start_indz9StencilPass._get_stencil_start_ind..get_start_indss8Q''(( (r!r>)rDrr#r$rr*r+rrr'r(rQrBpopitemrrIr3) r- start_lengthrorWrXr&f_irreret_vars rrz"StencilPass._get_stencil_start_inds lC ( ( -s<++,, , ) ) )"="dn#'>EJ=$,#'>334;1$$$$ ##%%a(%,000UZ_$ *R.&,r!c \ j}|d}d|D} d|jvr=|jdD]} | vrtdfd|jdD} ng} |j| vrtdj|jj} |j} |j}|jdu}|r | dgz}| dgz}n"d|jD}d |jD}tj |}d }| D] \}}g}|j D]}t|tjr@t|jtjr!|jjd vr|jjj| vsBt|tjst|tjr|jj| vrtd t|tjrt|jtjr|jjd vr|jjj| vr|jjj| vr|jj}| dkr|g}n&t/|dr|j|vr ||j}tj|_fd|D}|r&|t7||| |}|rt|tjst;d|Drtdt7t=t>||}t7t=t@||}d}|t7||| |}| dkr |d}ntj| tCd|}tDj#$tDj%| j|j<tj&||}tj|||}|'|tQfd|Dr"j|jjjj)}nj|jjj}tj*|jj||}tW|j|jjjj|jj,|<||_|'|||_ |r|std||fS)z Convert relative indexing in the stencil kernel to standard indexing by adding the loop index variables to the corresponding dimensions of the array index tuples. rcg|] }|j Sr)r:r5xs r z9StencilPass._replace_stencil_accesses..+s0001000r!standard_indexingz[Standard indexing requested for an array name not present in the stencil kernel definition.c g|] }| Srr)r5r/rns rr0z9StencilPass._replace_stencil_accesses..2s0 P P Pq 2 P P Pr!zYThe first argument to a stencil kernel must use relative indexing, not standard indexing.Ncg|] }|d S)rrr.s rr0z9StencilPass._replace_stencil_accesses..IEEEaQqTEEEr!cg|] }|d S)r>rr.s rr0z9StencilPass._replace_stencil_accesses..Jr4r!F)setitemstatic_setitemz?Assignments to arrays passed to stencil kernels is not allowed.)static_getitemgetitemr>r:c<g|]}tj|Sr)_get_const_index_exprr&)r5vr-rls rr0z9StencilPass._replace_stencil_accesses..tsB"J"J"J9:#8"DL!#5#5"J"J"Jr!c:g|]}t|t SrrDrr5r<s rr0z9StencilPass._replace_stencil_accesses..s% L L LAZ3%7%7!7 L L Lr!z.s?MMM<=!L0EJ>MMMr!z=Stencil kernel with no accesses to relatively indexed arrays.)-rBrYrTr:r'rrWrX neighborhoodrget_tuple_tablerCrIrDrrJr3rKrLrrr[indexhasattrbuild_definitions _definitions_add_index_offsetsrGranymapr$rrrrrrrrEallrr9r r()r-rlrrgr=rrnrr in_arg_namesr/standard_indexedrrWrXneed_to_calc_kernelrr tuple_tablefound_relative_indexrdrerxr7 index_list index_varsind_varrrgetitem_return_typ getitem_calls`` ` rrz%StencilPass._replace_stencil_accesses#s $*00000 ,"6 6 6!)*=> N NO++$&MNNN, P P P P%1%9:M%N P P P   "  ;* * *<== = V[). j*74?  F!1#IM)KKEE<+DEEEMEE<+DEEEK.~>> $+0022W "W "LE5H S &S &bi00h&tz27;;h JM-JJJ J,1\AAbj11Bb&677B!K, <<$%fggg tRY//B.&tz27;;B. JM-JJJ J,1\AA J,19III!%!1Jzz&0\ ":v66F:?k;Y;Y)4Z_)EJ/7.H.X.XJ+"J"J"J"J"J>H"J"J"JJ$O%)%<%r?s rr0z2StencilPass._add_index_offsets..s"EEEq 1c""EEEr! old_index_var offset_varoffset_stencil_index)rQrJrGrIrrPrDrrrrrrr'r:rJr\rErr SliceType_add_offset_to_slicerKbinopoperatorr*rr(r)r-rPr=rxrWrX out_nodesrQr6rWrrX index_varrr s rrGzStencilPass._add_index_offsetss :#m"4"44444 EEJ},DEEE F F =C];;<< <  s:''1 )1 )A&qMM--- / "u -o > >!E!E 38: ]/0!y*Q-)E)E4A3 H H   ...&q)J*c** /VE -l ; ;SBB 05  Z_-!y-2BC)H)H0:C A A   ...=%00 !$,}/A"B$)J$8:: |JO4 BBBB 55mZ8A5#OO !!),,,:u-- !$,z"?$)J$8:: |M$675:EEEE 55j-8A5#OO !!),,,u)*@AA3HHI+0:DL (x|]0:CAAJ)-)M)M)1 EJ7OQS*U*UDN: &9ZC@@L   \ * * *   i ( ( ( ( """r!c ^t|trLd|j|j}i}t |i||d}|g}t jf} n'd}||g}|j|j } | t jf} |j j j j } t|| |j|j| |j|j} | j\} }t)|||jdjj}||jdd|S)NzRdef f(offset): return slice({} + offset, {} + offset) fcBt|j|z|j|zSr%)rstartstop) old_sliceoffsets rraz+StencilPass._add_offset_to_slice..fs Y_v5y~7NOOOr!r')rDrformatrcrdexecrrr'r:r&func_idrM __globals__rr*r+r(rBr(rrIr3r)r-rrXr^rWrXf_textrar4arg_typs slice_type_globalsr*r^re new_indexs rr[z StencilPass._add_offset_to_slices/ i ' ' 1y 77 C S ! ! !CAKK;&&((5%&&&JrN(. CRC)))r!N) r __module__ __qualname__r.rprzrZrrrrGr[rr!rr#r#!s:7:7:7x"""2ooob   D   K*K*K*ZBBBHr!r#c ddlm}ddlm} ddlm} ddlm} |j } tj | j } | | _ tj | j }d|vrtdddlm} | j}| ||5t#|||| }t$jd|j| |jj|jj|jj|jjd \|j_|j_|j_}| |jj|jj|jjd d |jj|jjt<j d d d n #1swxYwYtj | tj!} tE| #}tI| #}tj%&|t<j'd kr#tQd tj)| i}|jj*D]7\}}tWj,|t[||}|||<|||j.<8tj/| |t<j'd kr#tQdtj)| |jj*D] \}}|||< i}| 0D]}|j1D]}te|tVj3rte|j4tVj5r~t<j'd kr3tQd||j4j6|j4j.|j4j6|v||j4j6j.||j4j.<||j4j6|_4t<j'd kr3tQd|tQdtj)| tj7| | | _ | |8|d|fS)z'get typed IR from stencil bytecode r) CPUContext) cpu_target)type_annotations)type_inference_stager<z6Cannot use the reserved word 'out' in stencil kernels.zbefore-inferenceNr)r&r'r(lifted lifted_fromr4r html_outputr>zInitial stencil_blockszAfter replace_varsrfrnzAfter replace arg with arr)9numba.core.cpursnumba.core.registryrtnumba.core.annotationsrunumba.core.typed_passesrv kernel_ircopydeepcopyrBrrrTtarget_contextnested_context DummyPipelinerrewrite_registryapplystater*r+r&r4r'rr(TypeAnnotationrHTMLadd_offset_to_labels next_labelr$rr_the_max_labelupdaterrrrCrrrr: replace_varsvaluesrIrDrJr3ArgrC remove_delsget_return_type)rkr*r4rWrXrfr'r(rsrtrurvstencil_func_irrrr+tpr^ min_label max_labelvar_dictr<typnew_varr1call_typrnrer7s rrVrVs*)))))......777777<<<<<<l''))O]?#9::N+O01GHHN QRRR/.....)I  " "9i 8 8%% 9i G G!''(:BHEEEH\H\ H  2BH4D HM4I!I!E"(.0BA ''H$H$h(,  ( % % %%%%%%%%%%%%%%%%(2>8@8K8M8MOONN''))**IN''))**I ""9--- "" &'''^,,,H("((**$$3& a 0 0#66 #  .(333 "" "###^,,,(,2244##h" $O&&((::J : :D$ ** :z$*bf/M/M :)Q.., DJ4D#z 0@J0NPPP3=dj>N3O3T 0' (89  :"" 111 *+++^,,, (((+O B..t44Q7 HHs!C+FFFceZdZdZdS)rcddlm}||_||j_||j_||j_||j_d|j_d|j_d|j_ dS)Nr) StateDict) numba.core.compilerrrr*r+r4r&r'rr()r-r*r+r4r*rs rr.zDummyPipeline.__init__gsl111111Y[[ ( (  ! ! !% # r!N)rrprqr.rr!rrrfs# $ $ $ $ $r!rc<tt|||}||S|S)z infer index_var as constant if it is of a expression form like c-1 where c is a constant in the outer function. index_var is assumed to be inside stencil kernel )r_get_const_index_expr_inner)rlr&r_ const_vals rr;r;ss0 #Z)EEI r!c*tt|tjt t |||}||St j||}t t|||}||St t|||}||St)zWinner constant inference function that calls constant, unary and binary cases. ) rrDrrr_get_const_two_irsrget_definition_get_const_unary_expr_get_const_binary_exprr)rlr&r_ var_const index_defs rrrs Jy"& ) )***):w JJI' I>>Iz7I??I GY@@I r!cxtt||}||Stt||}||St)zWget constant in either of two IRs if available otherwise, throw GuardException )rrr)ir1ir2varrs rrrsDj#s++Ij#s++I r!ctt|tjo |jdk|j}t |||}t|j}td ||S)zQevaluate constant unary expr if possible otherwise, raise GuardException unaryz{}{}) rrDrrKrLr3rrfnevalrg)rlr&r inner_varrrLs rrrsj Jy"' * * Fy|w/FGGGI+JKKI y| ,B  b),, - --r!c*tt|tjo |jdkt |||j}t |||j}t|j }td |||S)zRevaluate constant binary expr if possible otherwise, raise GuardException r\z{}{}{}) rrDrrKrLrlhsrhsrrrrg)rlr&rarg1arg2rLs rrrsz Jy"' * * Fy|w/FGGG &z7IM J JD &z7IM J JD y| ,B b$// 0 00r!)0rrrpytypesr]rnumpyrnumba.parfors.parforr@ numba.corerrrrnumba.core.typing.templatesr r numba.core.typingr r r numba.core.ir_utilsrrrrrrrrrnumba.core.errorsrnumba.core.utilsrnumba.nprr objectr#rVrr;rrrrrr!rrsW  <<<<<<<<<<<<<<FFFFFFFF''''''%%%%%%%%PPPPPPPPPPPPPPPPPPPPPP.-----222222"""""" k k k k k &k k k ZVIVIVIp $ $ $ $ $F $ $ $   .    . . .11111r!