nVh ddlZddlZddlZddlZddlmZddlmZddl m Z ddl Z ddl Z ddl mZdZedk(reeZej%d d ej%d d edddej%ddeddej%ddeddej%ddeddej%ddedd ej%d!d"ed#d$ej%d%d&ed'd$ej+Zej.r ej.n ej0Zej2r ej2neeZeej4Zej4j9deej:ej<j?ej@eZ!ej<jEe!Z#e!jHjKe#e&e#ej0Z'ejPjSd(Z(e*e(dk(sJe+e,d)ejZjSd(Z-d*e efd+Z.d,ej^d-ej`Z1e.e-e1gzZ2d.Z3e4e-Dcic]&\}}d/|vs |fe3|jSd/d(c}}Z5e5jmDcic] \}}| || c}}Z5e4e-Dcic]\}}e'jn|e3|c}}Z8e8jmDcic] \}}| || c}}Z8e5jmD] \Z9Z:e:dk(s e:e8e'jne9d<"e4e-Dcic]'\}}e'jn||jSd/d)c}}Z-e8D]Z9d0e-e9< d1jwe8jyDcgc] }e| c}Z=e8jmDcgc] \}}|d2|c}}Z>e>d3ej^d4ej`gz Z>e5jyD]Z?e?d5vrJd6e?e5jmDcic]\}}|d7k(s |d8d7ggc}}Z@e jje'e8e-e@9ZCej^ej`d:ZDe jeCeD;ZFeFjjdkDreId<gZ7gZJgZKgZLe4e'jnD]\ZMZNeNe8vrKe7jeNeJje-eNeKjeNeLje-eNUe5jeMfddk(sle7jeNeJjd=dZQe4e-jyD]G\ZMZReQeeMz ZQe5jeMfddk(reQd>z ZQe5jeMfdd7k(sCeQd?z ZQId@jwee2eQgZSeFjdAZTeejeTdBdCZVidDeSdEej0dFe*eTdGdHjweWeVdddBeVdddBDcgc] \}}dI||c}}d*dHjweWeKeLDcgc]\}}e|dJ|c}}dKdHjweWe7eJDcgc]\}}e|dJ|c}}dLdHjweKDcgc]}dM| c}dNgzdOe*eKdzdPe>dQeFjjdRej^dSd@jwe=e1gdTe(ddUe(ddVe(dBdWdZYdXD]ZZee[j:dYz dZz d[eZz Z\ejd\e2d@eQd\eZjd]5Z_e_jee\jjd^ieYdddyycc}}wcc}}wcc}}wcc}}wcc}}wcc}wcc}}wcc}}wcc}}wcc}}wcc}}wcc}w#1swYxYw)_N)ArgumentParser)Path)List) ty_to_cppa Triton ahead-of-time compiler: This program compiles the kernel with name `kernel-name` in the file at the provided `path` into self-contained C source-code that embeds the `cubin` data along with utilities to load, unload and launch the kernel. signature is provided as a list of (optionally divisibility-hinted) types or constexpr values, e.g. `compile.py --kernel-name kernel --signature "*fp32:16, i32:16, 1024, i32" --out-name kernel /path/to/kernel.py` will compile triton.JITFunction of name `kernel` inside the file `/path/to/kernel.py`. Said kernel will be specialized such that argument 0, 1 are assumed to be multiple of 16, and argument 2 is assumed to be a compile-time constant of value 1024, i.e. it won't be part of the generated prototype. The resulting entry point will have signature CUresult kernel_{specialization_suffix}(CUstream stream, unsigned gX, unsigned gY, unsigned gZ, float* arg0, int32_t arg1, int32_t arg2) Different such specialized entry points can be combined using the `linker.py` script. NOTE: when resolving the scope of /path/to/kernel.py, the file will be executed from within its parent directory with the python interpreter used to run this `compile.py` script __main__) descriptionpathzTPath to Python source containing desired kernel in its scope. File will be executed.)helpz --kernel-namez-nzName of the kernel to compileT)typedefaultr requiredz --num-warpsz-wz$Number of warps to launch the kernel)r r r z --num-stagesz-nsz/Number of stages (meta-parameter of the kernel)z --out-namez-onz Out name for the compiled kernelz --out-pathz-oz Out filenamez --signaturez-szSignature of the kernel)r r rz--gridz-gzLaunch grid of the kernel,c$|jdS)N )strip)ss D/home/dcms/DCMS/lib/python3.12/site-packages/triton/tools/compile.pyrFs1773< signaturectj}|jdj|j |j ddS)Nr)hashlibsha256updatejoinencode hexdigest)rms rhash_signaturer#Hs? NN  )$++-.{{}Ra  rwarpsxstagescv t|}|S#t$rYnwxYw t|}|S#t$rYywxYw)N)int ValueErrorfloat)rrets r constexprr+PsO a&CJ    (CJ   s   , 88:r+x=z num_warps=z num_stages=)rz#Only 1 and 16 are valid hints, got r/ztt.divisibility)fn constexprsrattrs) num_warps num_stages)optionszMAOT compiling kernels with global scratch requirements is not yet implementedi32cd_cubin kernel_nametriton_kernel_namebin_sizebin_dataz, 0xrfull_signature arg_pointers&z&global_scratchnum_argskernel_docstringsharedr3 algo_infogridXgridYgridZ _placeholder)hr7extracudazcompile..w)cbinasciirimportlib.util importlibsysargparserpathlibrtypingrtritontriton.backendstriton.backends.nvidia.driverrdesc__name__parser add_argumentstrr' parse_argsargsout_namer=out_pathr arg_pathinsertparentutilspec_from_file_locationstemspecmodule_from_specmodloader exec_modulegetattrkernelgridsplitlenlistmaprr#r3r4meta_sigsig_hashr+ enumeratehintsitems arg_names constantskeyvaluervalues const_sig doc_stringrMr2compiler ASTSourcesrcoptscompileccinfometadataglobal_scratch_size RuntimeError arg_typesarg_names_not_1arg_types_not_1iarg_nameappendgetsuffixty func_nameasmhexlifyhex_ziprGparamsext__file__ template_path with_suffixopenfpwrite read_textformat) rrkvr-ynamerargs 000000000rrs # 34 z -F su CJi!%'  t#qGmn CNP  e#tJlm  dt.Y  t#HDIIHHHOOAs8??+, >> 1 1(-- JD .. ) )$ /CKKC S$** +F 99??3 D t9>>S/1E1Ec1JKLI!$s)! t~~&gdoo->?Hi8*45H I)2):;AQCq*;JZ/0K?P2QRRJ \\^GG|FB1#FF|G6;kkm OdaqBwQ#R() ) OE // # #v)y`e # fCt GD V^^C .F **Q.jkkIIOO !1!12$ 8 9 $   X &   Yx0 1  " "8 ,  " "9X#6 7 YYud #q (   X &   U #$F9++-.2#a& 99aUD !Q & cMF 99aUD !R ' cMF  (Hf56I **W C x$ %a +Dyd.. CH DIIs4!9d14a4j7QRtq!A3qc{RS  TYY#o_nJophdB9R=/4& 9pq  $))sS\^gOh$i84 " av%>$ij   "HQse9"HL]K^"^_ C(1, J &//(( T^^ SXXy(34 a a a !F$GX--7&@XcUCSS  ! !AhZq#"? @ E Ec J Gb HH;T-(224;;EfE F G GGql ] =TEX>; PHSp$i"H G Gsl `` ``" `  `$`+,``-`$ `*'`*(`0`6`<? a ?6aa