o ,h0@sddlZddlZddlmZmZmZddlZddlmZ dej fddZ de ddfdd Z dej fd d Z   d#d ededeededeedeedefddZGdddZGdddZ d$deeefdeeedeeeedfffddZ d%deded ede fd!d"ZdS)&N)AnyOptionalUnion)_get_device_indexreturncCtjdkr tdStdS)Nwin32z nvcuda.dllz libcuda.so.1sysplatformctypesCDLLrrL/var/www/html/scripts/venv/lib/python3.10/site-packages/torch/cuda/_utils.py_get_cuda_library s   rresultcCsR|dkrdSt}t}||t||jdur |jnd}td|)NrUnknown CUDA error CUDA error: )r c_char_prcuGetErrorStringbyrefvaluedecode RuntimeError)rerr_strlibcuda error_messagerrr _check_cudasrcCr)Nrznvrtc64_120_0.dllz libnvrtc.sor rrrr_get_nvrtc_library s   r kernel_source kernel_namecompute_capability header_codecuda_include_dirs nvcc_optionsc sddl}tddtddffdd }|ds!d|}|r*|d |}n|}|d } |durF|j|j} | j | j }g} | d ||rc|D] } | d | qV|rr|D] } | | d qgdd l m }dd|D}| dd|Dt| }tj|| }t}|t|| |dddd|||}|krt}|t|t|j}||td|jt}||t|t|j}|||t||jS)a Compiles a CUDA kernel using NVRTC and returns the PTX code. Args: kernel_source (str): The CUDA kernel source code as a string kernel_name (str): The name of the kernel function to compile compute_capability (str, None): The compute capability to target (e.g., "86"). If None, will detect from current device. header_code (str, optional): Additional header code to prepend to the kernel source cuda_include_dirs (list, None): List of directories containing CUDA headers nvcc_options (list, None): Additional options to pass to NVRTC Returns: str: The compiled PTX code rNrrcsL|kr$t}|t||jdur|jnd}td|dS)Nrr)r rnvrtcGetErrorStringrrrr)rrr NVRTC_SUCCESSlibnvrtcrr check_nvrtcJs  z#_nvrtc_compile..check_nvrtcz extern "C"z extern "C"  utf-8z--gpu-architecture=sm_z-I)COMMON_NVCC_FLAGScSsg|]}|dkr|qS)z--expt-relaxed-constexprr.0flagrrr ysz"_nvrtc_compile..cSsg|]}|dqS)r,)encoder.rrrr1|sz.cuzKernel compilation failed: ) torch.cudarintstrip startswithr2cudaget_device_propertiescurrent_devicemajorminorappendtorch.utils.cpp_extensionr-extendlenr rc_void_pnvrtcCreateProgramrnvrtcCompileProgramc_size_tnvrtcGetProgramLogSizecreate_string_bufferrnvrtcGetProgramLogrrnvrtcGetPTXSize nvrtcGetPTXnvrtcDestroyProgram)r r!r"r#r$r%torchr* full_source source_bytespropsoptions directoryoptionr-nvrtc_compatible_flags num_options options_arrayprogreslog_sizelogptx_sizeptxrr'r_nvrtc_compile)sh        rZc@s2eZdZdejddfddZdeddfdd ZdS) _CudaModulemodulerNcCs||_i|_dSN)_module_kernels)selfr\rrr__init__ z_CudaModule.__init__name _CudaKernelc Cs||jvr |j|Sddlm}|}t}zt|t||j| dt ||j}||j|<|WSt yJ}z t d|d|d}~ww)Nr)rr,zNo kernel named 'z' in this module) r_torch.cuda._utilsrr r@rcuModuleGetFunctionrr^r2rdrAttributeError)r`rcrrfunckernelerrrrr __getattr__s$     z_CudaModule.__getattr__)__name__ __module__ __qualname__r r@rastrrkrrrrr[sr[c @steZdZdZdejdejddfddZ    dd eeeefd eeeefd e e d ede e ddf ddZ dS)rdzT Represents a compiled CUDA kernel that can be called with PyTorch tensors. rhr\rNcCs||_||_dSr])rhr\)r`rhr\rrrrarbz_CudaKernel.__init__rqrqrgridblockargs shared_memstreamcCslddl}|jj}|sg}g}g} |D]Y} t| |jr?| js*| jr&| s*t dt | } | | | t | qt| trRt | } | t | qt| tret | } | t | qtdt| t j t| }t| D] \}} t | t j ||<qz|durddl}|j}t||j|d|d|d|d|d|d||j|d dS)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: rq)rJr7_utilsr isinstanceTensoris_cudais_cpu is_pinned ValueErrorr r@data_ptrr<rr4c_intfloatc_float TypeErrortyper? enumeratecastr3current_streamrcuLaunchKernelrh_as_parameter_)r`rrrsrtrurvrJrprocessed_argsc_argsargptrrr c_args_arrayirrr__call__sV        z_CudaKernel.__call__)rprpNrN) rlrmrn__doc__r r@ratupler4rlistrrrrrrrds*  rdrY kernel_namesc Csddl}t}t|tr|d}t}|j}|t | t ||Wdn1s2wY|s=t |Si}|D]}t}t | t |||dt||||<qA|S)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,)r3rryror2r r@r7rrcuModuleLoadDatarr[rfrd) rYrrJrr\rvkernelsrcrhrrr_cuda_load_modules*   rFdeviceoptional allow_cpucCst|tr|St|trt|}t|tjr2|r&|jdvr%td|n |jdkr2td|tjsAt|tj jrA|j St |||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``. )r7cpuz(Expected a cuda or cpu device, but got: r7z!Expected a cuda device, but got: ) ryr4rorJrrr~jit is_scriptingr7idx_torch_get_device_index)rrrrrrrNs        r)NrNNr])FF)r r typingrrrrJ torch._utilsrrr rr4rrrorbytesrZr[rddictrboolrrrrs^    |]   1