o I&ik< @s(ddlZddlZddlZddlZddlZddlZddlmZmZmZm Z m Z m Z m Z m Z ddlZddlZddlmZddlmZddlmZmZmZmZmZddlmZeeZedj Z!e d Z"Gd d d ej#Z$Gd d d ej#Z%Gdddej#Z&Gdddej#Z'ej(GdddZ)Gdddej*Z+GdddZ,Gdddej-Z.de/de eeej0gej1fffddZ2de ej0dfde/fd d!Z3d"d#de ej0dfde/fd$d%Z4d&d"d'd(edefde ej0dfd)e5de/fd*d+Z6d,d-de e e ej0e e ej0ffd.d/Z7d0d1Z8dS)2N)AnyCallableDictListOptionalSetTupleUnion)free_unbacked_symbols)index_prevent_reordering)get_dtype_size sympy_str sympy_subs sympy_symbol VarRanges)Vz indirect|tmp) MemoryDepStarDepWeakDepc@seZdZUeed<ejed<eejdfed<eejdfed<ddZ e de ejejffd d Z dejfd d Z d e eefddfddZddZddZdefddZdefddZdefddZdS)rnameindex. var_namessizecCsd|jd|jd|jdS)Nz MemoryDep(, ))rrrangesselfrGC:\wamp64\www\opt\env\Lib\site-packages\torch/_inductor/dependencies.py__repr__szMemoryDep.__repr__returncCstt|j|jS)z{c0: 128, c1: 512, ...})dictziprrrrrr r szMemoryDep.rangescCs\|r tj|j}|St|jj}t d}t |j |j D] \}}||vr+||}q|SNr ) is_indirectrgraph get_numelrsetr free_symbolssympyIntegerr$rr)rZnumelvarsvarrrrr r(%s zMemoryDep.get_numelrenamescCs*|j|vrt||j|j|j|jdS|S)N)rr)rrrrrrr/rrr rename0s zMemoryDep.renamecC$tjj|ttj|jSNrr'sizevarsZ size_hintr(r Z get_dtyperrrrr numbytes_hint7 zMemoryDep.numbytes_hintcCtt|dkSNrlenr r(rrrr has_unbacked_symbols<zMemoryDep.has_unbacked_symbolscCst|jtjo |j|jvSr3) isinstancerr+Symbolrrrrr is_contiguous?szMemoryDep.is_contiguouscCs6t|jtjr|j|jvo| St|jttjfSr3)r>rr+r?rr&intr,rrrr is_scalarBszMemoryDep.is_scalarcCstdd|jjDS)Ncss|]}t|jVqdSr3)r&r.0vrrr Hsz(MemoryDep.is_indirect..)anyrr*rrrr r&GzMemoryDep.is_indirectN)__name__ __module__ __qualname__str__annotations__r+Exprrr?r!propertyrrr(r1r6r<boolr@rBr&rrrr rs   rc@seZdZUeed<eddZdejfddZ de eefddfdd Z d d Z d d Z defddZdefddZdefddZdS)rrcCtd)NzStarDep does not have an indexNotImplementedErrorrrrr rOz StarDep.indexr"cCstj|jSr3)rr'r(rrrrr r(SszStarDep.get_numelr/cC|j|vr t||jS|Sr3)rrr0rrr r1V zStarDep.renamecCr2r3r4rrrr r6[r7zStarDep.numbytes_hintcCr8r9r:rrrr r<`r=zStarDep.has_unbacked_symbolscCdSNFrrrrr r@czStarDep.is_contiguouscCrWrXrrrrr rBfrYzStarDep.is_scalarcCrWrXrrrrr r&irYzStarDep.is_indirectN)rIrJrKrLrMrOrr+rNr(rr1r6r<rPr@rBr&rrrr rKs  rc@sjeZdZUeed<eddZdejfddZ de eefddfdd Z d d Z d d Z defddZdS)rrcCrQ)NzWeakDep does not have an indexrRrrrr rvrTz WeakDep.indexr"cCs tdSr%)r+r,rrrr r(zs zWeakDep.get_numelr/cCrUr3)rrr0rrr r1}rVzWeakDep.renamecCrWr%rrrrr r6rYzWeakDep.numbytes_hintcCrWrXrrrrr r<rYzWeakDep.has_unbacked_symbolscCrWrXrrrrr r@rYzWeakDep.is_contiguousN)rIrJrKrLrMrOrr+rNr(rr1r6r<rPr@rrrr rss  rc@s<eZdZUejed<eejdfed<eejdfed<dS) IndexExprDepr.rrN)rIrJrKr+rNrMrr?rrrr rZs rZc@seZdZUeeed<eeed<eeed<dZee e j ed<dZ ee ed<ejejdZejeed<d ejeefd dfd d Zd ed dfddZdddZede dfddZddZddZdS) ReadWritesreadswrites index_exprsN range_vars var_ranges)default_factory op_countsr/r"cs<tfdd|jDfdd|jD|j|j|j|jdS)Nch|]}|qSrr1rDdepr/rr z$ReadWrites.rename..crcrrdrergrr rhrirbr[r\r]r^r_r`rbr0rrgr r1szReadWrites.renamerfcCs<t|ttfs Jtt|j|h|j|j|j |j |j dSNrj) r>rrr[r)unionr\r]r^r_r`rb)rrfrrr with_readszReadWrites.with_readothercCs\t|j|j}t|j|j}t|j|j}t|j}||jt |||||dSrl) r)rmr\r]r^ collectionsCounterrbupdater[)rror\r]r^rbrrr merges   zReadWrites.merge read_writescCsntjdd|D}tjdd|D|}tjdd|D}t}|D]}||jq&t||||dS)NcSg|]}|jqSr)r]rDrwrrr z)ReadWrites.merge_list..cSrur)r\rvrrr rxrycSrur)r^rvrrr rxryrj)r)rmrprqrrrbr[)rtZ all_writesZ all_readsZall_index_exprsrbrwrrr merge_listszReadWrites.merge_listcCs$t|j||j|j|j|j|jdSrlrk)rZ rem_readsrrr remove_readsszReadWrites.remove_readscCst|j|jSr3) itertoolschainr\r]rrrr reads_and_writesszReadWrites.reads_and_writes)ror[)rIrJrKrDeprMrZr_rrr+rNr`r dataclassesfieldrprqrbtypingrLrr1rnrs staticmethodrzr{r~rrrr r[s     r[c seZdZdedeffdd Zdejdeejeej dfeejdfffdd Z d e dejde fd d Z d e de fd dZdd e dejde de fddZd e de fddZdejde fddZde dejdejdefddZZS)_RecordLoadStoreInnerr` normalizecs2tt|_t|_t|_||_||_dSr3)super__init__r)_reads_writes _index_exprs _var_ranges _normalize)rr`r __class__rr rs  z_RecordLoadStoreInner.__init__rr".c sB|js*dd|jD}tddt|j|D}tdd|D}|||fS|j}dd|jD}g|}t|}tj j ||t |g||\}}} t t\} tt||fdd|D} tt|| }g| } g|}|j}| r| d |vr| || r| d |vs|t| t|fS) NcSsg|] }tjj|qSrrr'r5simplifyrDxrrr rxsz6_RecordLoadStoreInner.canonicalize..css |] \}}|dkr|VqdSr NrrDkrErrr rFsz5_RecordLoadStoreInner.canonicalize..css|] }|dkr|VqdSrrrCrrr rFscSs i|] \}}|tjj|qSrrrrrr sz6_RecordLoadStoreInner.canonicalize..csg|]}|qSrrradd_varrr rxs)rrvaluestupler$keysr*itemsrr'r5Z_simplify_loopsr var_buildercanonicalization_prefixr#rr+expandpop) rrsizesrr*r`Z index_vars new_sizesreindexZpruneZnew_vars replacementrrr canonicalizes:       z"_RecordLoadStoreInner.canonicalizercCs4|jt|g||Rd|dt|dS)Nzload(rr)raddrrrrrrrrr load sz_RecordLoadStoreInner.loadcCs t|tsJ||t|Sr3)r>rArr+r,rrrr load_seedsz_RecordLoadStoreInner.load_seedNvaluec Cs@|jt|g||Rd|dt|d|d|d S)Nzstore(rr)rrrrr)rrrrmoderrr stores"z_RecordLoadStoreInner.storecCs|||d|dS)Nzstore_reduction(r)r)rrrrrrr store_reductionrHz%_RecordLoadStoreInner.store_reductioncCs,|jt||dt|d|dS)Nz index_expr(rr)rrrZrr)rrdtyperrr index_exprsz _RecordLoadStoreInner.index_expr offsets_name offsets_sizeindexing_dtyperightc Cs8|jt|d|d|dt|d|d|d S)Nz bucketize(rr)rrrr)rrrrrrrrr bucketizes(z_RecordLoadStoreInner.bucketizer3)rIrJrKrrPrr+rNrr?rrLrrArrrrtorchrr __classcell__rrrr rs*" -rcs(eZdZdZfddZddZZS) _OpCounterz,Shim to count how many times each op is usedcst||_t|_dSr3)rrparent_handlerrprq _op_counts)rinnerrrr r,s z_OpCounter.__init__cCs|j|d7<t|j|Sr%)rgetattrr)rrrrr __getattr__1s z_OpCounter.__getattr__)rIrJrK__doc__rrrrrrr r)s rcs&eZdZdedeffdd ZZS)RecordLoadStorer`rcs&t||d}t|}tj|ddS)N)r`r)r)rrrr)rr`rrrrr r7s zRecordLoadStore.__init__)rIrJrKrrPrrrrrr r6srprefixr"cs4ttdtjdtjffdd }|fS)Nlengthr"cs tt}||<|Sr3)rnext)rrEZcntrr`rr rCszvar_builder..add_var)r|countr#r+rNr?)rrrrr r?srargsizes.cGs6t|\}}g}|D] }|tt||q ||fSr3)rappendlistmap)rrr`rargsrrrr index_vars_no_squeezeKs rd)rc Gsbddlm}t|\}}g}g}|D]}||\}} |||| tt||q||fS)Nr ) SqueezeView)irrrZsqueezerrrr) rrrr`rrrrnew_sizerrrr index_vars_squeezeSs   rF)rrfnrc Gst|d|i\}}t||d}t| ||Wdn1s#wY|r-g}ngtj|}|jj}tt|j t|j |j |||jj S)Nr)r) rrrZset_ops_handlerr|r}rr[r)rrrr) rrrrrr`rwr_rrrr extract_read_writes`s"   r input_nodeztorch._inductor.ir.TensorBoxc Cs\ddlm}m}t|j|r"|}|}t|dkr ||fSdSt|jj|s+dS|}d}d}|durt|dkrt }g}|D]Q}t|t sLqD|j |vrRqD| |j t j|j } | durdqDt| |rt| dkr|dur~| }| }qD|| ks|| krdSqD|| qD||kr||fS|}|durt|dks=||fS)aX Returns the size and reduction size of all inputs, if the sizes and reduction_sizes (if exist) are all the same. It's possible that a node has multiple inputs, some are Reduction nodes and others are Pointwise nodes. In this case, reduction_sizes of the Reduction nodes need to be the same. Otherwise returns (None, None). r )ComputedBufferLoopsr)NNN)rrrr>dataget_sizeZget_reduction_sizer;Z get_readsr)rrrrr' get_bufferextend) rrrrZreduction_sizer\seenZ new_readsreadbufferrrr #extract_input_node_reduction_ranges{sP        rcCrW)Ncrrrrr rrYr)9rprr|loggingrerrrrrrrrr r+rZ%torch.fx.experimental.symbolic_shapesr Zcodegen.commonr utilsr rrrrZ virtualizedr getLoggerrIlogcompilesearchr&r NamedTuplerrrrZ dataclassr[Z MockHandlerrrZKernelFormatterHandlerrrLrNr?rrrrPrrrrrrr sZ(     4(AU (  "    @