9jRddlZddlZddlmZddlZ ddlmZdZddl m Z dejfdZdejfd Zdejfd Zd eddfd Zd edefd ZdejfdZdejfdZdejfdZdeefdZ d$dedededzdedzdedzdedeeeffdZGddZGddZ d%deezdeedzdee edfzfdZ! d&d ed!ed"edefd#Z y#e $rdZdZYwxYw)'N)Any)runtimeTF)_get_device_indexreturnc ddl}tjt|j dd}|j|_ |j|_|j |_|j$|_|j(|_|S#t t f$r`tjdk(r5tjdtjjdd}ntjd}YwxYw)Nramdhip64win32 amdhip64_.dllzlibamdhip64.so)rocm_sdkctypesCDLLstrfind_libraries ImportError IndexErrorsysplatformtorchversionhiphipGetErrorStringcuGetErrorStringhipModuleLoadDatacuModuleLoadDatahipModuleGetFunctioncuModuleGetFunctionhipModuleLaunchKernelcuLaunchKernelhipFuncSetAttributecuFuncSetAttribute)r libs Q/media/conek/DATA/Code/OCR/venv/lib/python3.12/site-packages/torch/cuda/_utils.py_get_hip_runtime_libraryr$s 0kk#h55jA!DEF00C00C!66C22C 44C J  $0 <<7 "++ %--*;*;A*>)?tDEC++./C 0s4B A,C<;C<c|tjdk(rtjdStjdS)Nr z nvcuda.dllz libcuda.so.1)rrr rr#_get_cuda_libraryr(-s, ||w{{<(({{>**r'c^tjjr tSt SN)rrrr$r(r&r'r#_get_gpu_runtime_libraryr+5s! }}')) ""r'resultc|dk(rytj}t}|j|tj||j |j j nd}td|)NrUnknown CUDA error CUDA error: )r c_char_pr+rbyrefvaluedecode RuntimeError)r,err_strlibcuda error_messages r# _check_cudar8=sn {ooG&(G VV\\'%:;")--"; AU m_5 66r'cBts td|^}}|tjjk7rJtj |\}}t |tr|j}td|d|dt|dk(ryt|dk(r|dS|S)aCheck a cuda.bindings (cuda-python) call result for errors. All cuda.bindings runtime calls return ``(error, *outputs)``. This helper unpacks the tuple, raises on non-success, and returns the outputs (``None`` for zero outputs, scalar for one, tuple otherwise). zcuda.bindings is not availabler/z ()rN) _HAS_CUDA_BINDINGSr4_cuda_bindings_runtime cudaError_t cudaSuccesscudaGetErrorString isinstancebytesr3len)r,errout_r5s r#_check_cuda_bindingsrGIs ;<<IC# ! - - 9 9 : # 5 5  7 gu %nn&G\#b ;<< 3x1} 3x1}1v Jr'c  ddl}tjt|j dd}|j|_|j|_|j"|_|j&|_|j*|_|j.|_|j2|_|j6|_|j:|_|j>|_ |S#t t f$rtjdk(redjdtjjddtjjdg}tjd|d}ntjd}YKwxYw) Nrhiprtcr 0r z libhiprtc.so)!r r rrrrrrrjoinrrrhiprtcGetErrorStringnvrtcGetErrorStringhiprtcCreateProgramnvrtcCreateProgramhiprtcDestroyProgramnvrtcDestroyProgramhiprtcCompileProgramnvrtcCompileProgramhiprtcGetCodeSizenvrtcGetCUBINSize hiprtcGetCode nvrtcGetCUBINhiprtcGetProgramLogSizenvrtcGetProgramLogSizehiprtcGetProgramLognvrtcGetProgramLoghiprtcAddNameExpressionnvrtcAddNameExpressionhiprtcGetLoweredNamenvrtcGetLoweredName)r r" version_strs r#_get_hiprtc_libraryrcfsB .kk#h55h?BCD"66C 44C!66C!66C11C))C!$!.check_nvrtcso ] "oo'G  ( (g1F G==, $$&)  m_=> > #r'utf-8z--offload-arch=z--gpu-architecture=sm_) include_pathsrhz-Iz12.8zPCH requires CUDA 12.8+, got z--pchz.cuzKernel compilation failed: rJ)' torch.cudarprgencoderhget_device_propertiescurrent_devicerr gcnArchNamemajorminorappendrtrrAssertionErrorryrurCr r0c_void_prQr1r_rUc_size_tr[create_string_bufferr2r]r4r3rWrYrarSraw)rzr{r|r}r~rrr source_bytespropsoptionsrcuda_include_paths cuda_path directoryoptionnvrtc_compatible_flagsrw num_options options_arrayprog c_kernel_namereslog_sizelog binary_sizebinaryc_mangled_name mangled_namerrs @@r#_nvrtc_compilers0$%HM ?C ?D ?!''0L! 001J1J1LM ==  $)$5$5#6 $)KK= !> G }});(<=DDFG/0B/CDKKMN8&v.'2 I;'..012* 6I NNR {+224 5 6 u}}!! "V + #@ASAS@T!UV V  LG$" 3F NN6==1 2 3;< NN5KLTDKK(LMg,K__{2W=M ?? D## LL  m3  & & (      &&w/M//mDE  & &t[- HC m??$''fll8.DE))(..9##D#.9#)):J:J:L9MNOO//#K**4k1JKL  ( ():): ;F&&tV45__&N$$T=&,,~:VW'%++224     d!34 ::| ##oMsQ7c@eZdZdejddfdZdeddfdZy) _CudaModulemodulerNc ||_i|_yr*)_module_kernels)selfrs r#__init__z_CudaModule.__init__Js 02 r'name _CudaKernelc ||jvr|j|Sddlm}|}tj} t |j tj||j|jdt||j}||j|<|S#t$r}td|d|d}~wwxYw)Nr)r+rzNo kernel named 'z' in this module) rtorch.cuda._utilsr+r rr8rr1rrrr4AttributeError)rrr+r6funckernelrDs r# __getattr__z_CudaModule.__getattr__Ns 4== ==& & ?*,  V ++LL& dkk'6J  !t||4F"(DMM$ M V #4TF:J!KLRU U VsA.B// C 8CC )__name__ __module__ __qualname__r rrrrr&r'r#rrIs/3v343VV Vr'rc eZdZdZdej dej ddfdZ ddeeeefdeeeefd e dzd ed e dzddf d Z d eddfdZ y)rzT Represents a compiled CUDA kernel that can be called with PyTorch tensors. rrrNc.||_||_d|_y)Nr)rr_max_shared_mem_bytes)rrrs r#rz_CudaKernel.__init__ls  %&"r'gridblockargs shared_memstreamcBddl}|jjj}|sg}g}g} |D]O} t | |j r| j s'| jr| js tdtj| j} |j| | jtj| t | tr:tj | } | jtj| t | t"r;tj$| } | jtj| ;t'dt)| tjt+| z}t-| D],\}} tj.| tj||<.|ddl}|jj3}|dk\rQ|j4dk(s||j4kDr3|j4dk(rdnd|j4d}t7d |d |d t9|j;|j<|d|d |d |d|d |d ||j>|d y)a Call the compiled CUDA kernel Args: grid (tuple): Grid dimensions (grid_x, grid_y, grid_z) block (tuple): Block dimensions (block_x, block_y, block_z) args (list): List of arguments to pass to the kernel. PyTorch tensor arguments will be automatically converted to pointers. shared_mem (int): Shared memory size in bytes stream (torch.cuda.Stream): CUDA stream to use. If None, uses current stream. rNz?All tensor arguments must be CUDA tensors or pinned CPU tensorszUnsupported argument type: znot configuredzonly z bytes configuredzKernel requires z' bytes of shared memory (>= 48KB), but ze. Call kernel.set_shared_memory_config(shared_mem) after compilation and before launching the kernel.r;rL) rrh_utilsr+rATensoris_cudais_cpu is_pinned ValueErrorr rdata_ptrrr1rgc_intfloatc_double TypeErrortyperC enumeratecastrcurrent_streamrr4r8rr_as_parameter_)rrrrrrrr6processed_argsc_argsargptrrr c_args_arrayiconfigured_msgs r#__call__z_CudaKernel.__call__qs_& **##<<>D13 KC#u||,{{CJJ3==?$Yooclln5%%c* fll3/0C% S) fll512C'!??3/ fll845"=d3i[ IJJ+ K0#f+58 ' @FAs$kk#v?LO @ > ZZ..0F  "  & &! +zD!!/ 01GG 783  & & ;   &6"r')r;r;r;rNrN) rrr__doc__r rrtuplerglistrrrr&r'r#rrgs'V__'foo'$'&/&/ ! _ CcM"_ S#s]#_ Tk _  _ d _  _ B(6(6(6r'rptx kernel_namesc 8ddl}t}t|tr|j d}t j }|jj}|5t|jt j||ddd|s t|Si}|D]c}t j }t|jt j|||j dt||||<e|S#1swYxYw)a, Loads a CUDA module from PTX code and returns a module object that can access kernels. Args: ptx (bytes or str): The PTX code to load kernel_names (list, optional): List of kernel names to extract from the module. If None, will return a module object with __getattr__. Returns: object: If kernel_names is None, returns a module object with __getattr__ to access kernels. If kernel_names is provided, returns a dict mapping kernel names to _CudaKernel objects. rNr)rr+rArrr rrhrr8rr1rrr) rrrr6rrkernelsrrs r#_cuda_load_modulers '(G#sjj!__ F ZZ & & (F IG,,V\\&-A3GHI 6""G2   ' ' T"FDKK,@  $D&1 2 N!IIs /DDdeviceoptional allow_cpuct|tr|St|trtj|}t|tjr;|r|j dvr+t d||j dk7rt d|tjjs0t|tjjr |jSt|||S)aGet the device index from :attr:`device`, which can be a torch.device object, a Python integer, or ``None``. If :attr:`device` is a torch.device object, returns the device index if it is a CUDA device. Note that for a CUDA device without a specified index, i.e., ``torch.device('cuda')``, this will return the current default CUDA device if :attr:`optional` is ``True``. If :attr:`allow_cpu` is ``True``, CPU devices will be accepted and ``-1`` will be returned in this case. If :attr:`device` is a Python integer, it is returned as is. If :attr:`device` is ``None``, this will return the current default CUDA device if :attr:`optional` is ``True``. )rhcpuz(Expected a cuda or cpu device, but got: rhz!Expected a cuda device, but got: ) rArgrrrrrjit is_scriptingrhidx_torch_get_device_index)rrrs r#rr-s &# &#f%&%,,' {{/1 #KF8!TUU [[F "@IJ J 99 ! ! # fejj// 0::  "68Y ??r')NNNFr*)FF)"r rtypingrr cuda.bindingsrr=r<r torch._utilsrrrr$r(r+rgr8rGrcrnrprrryboolrrBrrrdictrr&r'r#rs   F&++.+6;;+#&++# 7 7 7:V[[:6FKK6&$fkk$tCy6&*%) $ O$O$O$d O$d{ O$ + O$  O$ 5#:O$dVV