gt4$dZddlmZddlZdZdZdZGddeZ Gdde Z Gd d e Z d Z d Zd ZdZddZedZdZejddfdZejddfdZdZdZdZd dZdZd!dZdZdZdZdS)"zz The Device Array API is not implemented in the simulator. This module provides stubs to allow tests to import correctly. )contextmanagerNzArray contains non-contiguous buffer and cannot be transferred as a single memory region. Please ensure contiguous buffer with numpy .ascontiguousarray()c"eZdZdZfdZxZS) FakeShapez The FakeShape class is used to provide a shape which does not allow negative indexing, similar to the shape in CUDA Python. (Numpy shape arrays allow negative indexing) ct|tr|dkrtdtt||S)Nrztuple index out of range) isinstanceint IndexErrorsuperr __getitem__)selfk __class__s ^/granary/edev/quantum/lib/python3.11/site-packages/numba/cuda/simulator/cudadrv/devicearray.pyr zFakeShape.__getitem__sJ a   9!a%%788 8Y%%11!444)__name__ __module__ __qualname____doc__r __classcell__)rs@rrrsB 555555555rrcBeZdZdZdZdZdZdZdZdZ dZ d Z d S) FakeWithinKernelCUDAArrayaG Created to emulate the behavior of arrays within kernels, where either array.item or array['item'] is valid (that is, give all structured arrays `numpy.recarray`-like semantics). This behaviour does not follow the semantics of Python and NumPy with non-jitted code, and will be deprecated and removed. cHt|tsJ||jd<dS)N_item)r FakeCUDAArray__dict__r items r__init__z"FakeWithinKernelCUDAArray.__init__*s*$ .....!% grcNt|trt|S|SN)rrrrs r__wrap_if_fakez(FakeWithinKernelCUDAArray.__wrap_if_fake.s' dM * * ,T22 2Krc|t|jjvr-|t |jj|S||j|Sr )dirr_ary(_FakeWithinKernelCUDAArray__wrap_if_fakegetattrr )r attrnames r __getattr__z%FakeWithinKernelCUDAArray.__getattr__4s_ s4:?++ + +&&wtz'I'IJJ J&&tz'='=h'G'GHH Hrc<|j||dSr r __setitem__)r nmvals r __setattr__z%FakeWithinKernelCUDAArray.__setattr__:s  r3'''''rc\||j|Sr )r%rr )r idxs rr z%FakeWithinKernelCUDAArray.__getitem__=s&""4:#9#9##>#>???rc<|j||dSr r*r r0r-s rr+z%FakeWithinKernelCUDAArray.__setitem__@s  sC(((((rc*t|jSr )lenrr s r__len__z!FakeWithinKernelCUDAArray.__len__Cs4:rct||}d|d}|rtfd|D|d<tfd|D}||i|S)NcHt|tr |jj}|Sr )rrrr$objs r convert_fakesz@FakeWithinKernelCUDAArray.__array_ufunc__..convert_fakesPs##899 %inJroutc3.K|]}|VdSr ).0or;s r z.Xs-!@!@q--"2"2!@!@!@!@!@!@rc3.K|]}|VdSr r>)r?ar;s rrAz.Ys-44!]]1%%444444r)r&gettuple)r ufuncmethodargskwargscallr<r;s @r__array_ufunc__z)FakeWithinKernelCUDAArray.__array_ufunc__Fsuf%%    jj  A!!@!@!@!@C!@!@!@@@F5M4444t44444tT$V$$$rN) rrrrrr%r(r.r r+r6rKr>rrrr!s&&& III (((@@@)))%%%%%rrc.eZdZdZdZd'dZedZedZdZ d'dZ ed Z d(d Z d Z d Zd)dZd'dZedZdZdZdZdZdZdZdZdZdZdZdZdZdZdZdZ d Z!d!Z"d"Z#d#Z$d$Z%d%Z&d'd&Z'd S)*rzj Implements the interface of a DeviceArray/DeviceRecord, but mostly just wraps a NumPy array. Trc"||_||_dSr )r$streamr aryrNs rrzFakeCUDAArray.__init__es  rc|jjSr r$nbytesr5s r alloc_sizezFakeCUDAArray.alloc_sizeis yrc|jjSr rRr5s rrSzFakeCUDAArray.nbytesmsyrc t|j|}|S#t$r}d|z}t||d}~wwxYw)Nz#Wrapped array has no attribute '%s')r&r$AttributeError)r r'attremsgs rr(zFakeCUDAArray.__getattr__rsV -49h//DK - - -7(BC %%1 , -s =8=c,t|j|Sr rr$)r rNs rbindzFakeCUDAArray.bindzsTY///rc*|Sr ) transposer5s rTzFakeCUDAArray.T}s~~rNcRttj|j|S)N)axes)rnpr_r$)r rbs rr_zFakeCUDAArray.transposes!R\$)$???@@@rc|j|}t|tjtjfvr|St ||jS)NrN)r$r typercndarrayvoidrrN)r r0rets rr zFakeCUDAArray.__getitem__sKi##C(( 99RZ1 1 1J T[999 9rc8|j||Sr )r$r+r2s rr+zFakeCUDAArray.__setitem__sy$$S#...rc|tj|j}nt||tj||j|Sr )rc empty_liker$check_array_compatibilitycopytorOs r copy_to_hostzFakeCUDAArray.copy_to_hostsC ;- **CC %dC 0 0 0 #ty!!! rclt|t|t|}}t|tr t|t ||n7t j||jdrdnddd}t ||t j|j |dS)a Copy from the provided array into this array. This may be less forgiving than the CUDA Python implementation, which will copy data up to the length of the smallest of the two arrays, whereas this expects the size of the arrays to be equal. C_CONTIGUOUSCFTF)ordersubokcopyN) sentry_contiguous array_corerrrmrcarrayflagsrnr$)r rPrN self_coreary_cores rcopy_to_devicezFakeCUDAArray.copy_to_devices $(.. 38 c= ) ) ; c " " " %i : : : :x&_^<Ecc# H &i : : : ).(+++++rc4t|jjSr )rr$shaper5s rrzFakeCUDAArray.shapes)))rc@t|jj|i|Sr )rr$ravelr rHrIs rrzFakeCUDAArray.ravels#_TY_d=f==>>>rc@t|jj|i|Sr )rr$reshapers rrzFakeCUDAArray.reshapes%.TY.???@@@rc@t|jj|i|Sr )rr$viewrs rrzFakeCUDAArray.views#^TY^T)r)r?rCs r z'FakeCUDAArray.split..s0    !     r)rcsplitr$ranger4)r sectionrNs rrzFakeCUDAArray.splitsC  XdiwD 7)K)KLL    rrr )Nr)(rrrr__cuda_ndarray__rpropertyrTrSr(r]r`r_r r+ror}rrrrrrrrr6rrrrrrrrrrrrrrr>rrrr]sy   X   X ---0000  X AAAA:::///,,,,,**X*???AAA>>>,,,,,, 111111000111000111000000000111000000111      rrc|jr|js|Sg}|jD],}||dkrdntd-|t |S)aG Extract the repeated core of a broadcast array. Broadcast arrays are by definition non-contiguous due to repeated dimensions, i.e., dimensions with stride 0. In order to ascertain memory contiguity and copy the underlying data from such arrays, we must create a view without the repeated dimensions. rN)stridessizeappendslicerE)rP core_indexstrides rrxrxsn ;ch J+==v{{!!d <<<< uZ  !!rc|jj}tt|jt|jD]\}}|dkr|dkr||krdS||z} dS)z Returns True iff `ary` is C-style contiguous while ignoring broadcasted and 1-sized dimensions. As opposed to array_core(), it does not call require_context(), which can be quite expensive. rFT)dtypeitemsizezipreversedrr)rPrrrs r is_contiguousr so 9 DXci00(3;2G2GHH v 1991v~~uu EMD 4rct|}|jds!|jdsttdSdS)Nrq F_CONTIGUOUS)rxrz ValueErrorerrmsg_contiguous_buffer)rPcores rrwrwsM c??D :n %3dj.H312223333rcp||}}|j|jkrtd|jd|j|j|jkrt d|jd|j|j|jkrt d|jd|jdS)Nzincompatible dtype: z vs. zincompatible shape: zincompatible strides: )squeezer TypeErrorrrr)ary1ary2ary1sqary2sqs rrmrm s\\^^T\\^^FF zTZiTZZ122 2 |v|##j***djj233 3 ~''j,,, 677 7('rTc tj|dd}t|||jjdvr tjn|j}t tjtjt| ||j|j |j  t|S|||dS)NFTrvruMm)bufferrrr)rfre)rcryrwrcharint64rrgrvrxrrrrfr})rPrNrvto buffer_dtypes r to_devicer-s (3U$ / / /Cc z#&9>T#9#9rxxsy  Jwz#//44\BBii     dS d""     #f-----rc#KdVdSr r>)args rpinnedr>s EEEEErcVdD]}||vr||t|i|S)N)portablewc)pop device_array)rHrI unused_args r mapped_arrayrCsB(##    JJz " " "  ( ( ((rrrc2tj||||SN)rrrrtrcrgrrrrts r pinned_arrayrJ :E7%u M M MMrc2tj||||Srrrs r managed_arrayrNrrcxd|vr|dnd}ttj|i||S)NrNrre)rrrcrg)rHrIrNs rrrRsD%-%7%7VZZ ! ! !QF T4V44V D D DDrcZ|jds|jds |jdkrdSdt|jD}|ddgt |jz}|jj}|D]\}}|||<||j|z}t|S) z[ Given an array, compute strides for a new contiguous array of the same shape. rqrrNcg|]}|Sr>r>)r?xs rrz2_contiguous_strides_like_array..hs6661666rc|dS)Nrr>)rs rz0_contiguous_strides_like_array..is !A$r)keyr) rzndim enumeratersortr4rrrrE)rP stridepermrri_perm_s r_contiguous_strides_like_arrayrWs y CIn$=Qt76i 44666JOOO'''cC $$$G Y F$$  #)F## >>rc>|jdr|jdsdSdS)Nrrqrsrr)rz)rPs r_order_like_arrayrts) y >)Bssrcxt|}t|}t|j|j||SNr)rrrrr)rPrNrrts rdevice_array_liker{@,S11G c " "E cisy'# % % %%rcxt|}t|}t|j|j||Sr)rrrrr)rPrrts rpinned_array_liker rrct|tr|dfSt|tjstj|dd}t |||dfS)NFTr)rrrcrhryr)rPrNrvs r auto_devicer sj#}%%Ez c27 # #h  S&$ ' ' --rc$t|ddS)z$Check if an object is a CUDA ndarrayrF)r&r9s ris_cuda_ndarrayr s 3*E 2 22rctfd}|dt|dt|dtj|dtdS)z,Verify the CUDA ndarray interface for an objct|st|tt||st|d|dS)Nz must be of type )hasattrrWrr&)rXtypr:s r requires_attrz4verify_cuda_ndarray_interface..requires_attrsesD!! ' && &'#t,,c22 H DDD##!FGG G H HrrrrrN)require_cuda_ndarrayrErcrr)r:rs` rverify_cuda_ndarray_interfacersHHHHH M'5!!!M)U###M'28$$$M&#rcBt|stddS)z9Raises ValueError is is_cuda_ndarray(obj) evaluates Falsezrequire an cuda ndarray objectN)r rr9s rrrs+ 3  ;9:::;;r)rTNr)rT) r contextlibrnumpyrc DeviceRecordfrom_record_likerrErobjectrrrxrrwrmrrrfloat_rrrrrrr r r rrr>rrrs*&%%%%% 3 5 5 5 5 5 5 5 59%9%9%9%9%9%9%9%xX X X X X FX X X v"""$    333 7 7 7...."   )))!iSNNNN "y$cNNNNEEE :%%%%%%% . . . .333    ;;;;;r