o ei/@sdddlmZddlZddlmZdZdZdZGdddejj Z dddedfd d Z d d d Z dS))OptionalN)_NnapiSerializercseZdZUdZeejjje d<e ej e d<e ej e d<dej j dej de ej de ede ed ed effd d Zejjd e ej fddZd e ej de ej fddZZS) NnapiModulezTorch Module that wraps an NNAPI Compilation. This module handles preparing the weights, initializing the NNAPI TorchBind object, and adjusting the memory formats of all inputs and outputs. compweights out_templatesshape_compute_module ser_model inp_mem_fmts out_mem_fmtscompilation_preferencerelax_f32_to_f16csDt||_||_||_||_||_g|_d|_||_ ||_ dSN) super__init__r r rr r r rrr)selfr r rr r rr __class__g/var/www/addictedbytheproject.nl/epg/venv/lib/python3.10/site-packages/torch/backends/_nnapi/prepare.pyrs  zNnapiModule.__init__argscCsd|jdur td|j|j||_dd|jD|_tjj }| |j|j|j |j ||_dS)Nz'comp must be None before initializationcSsg|]}|qSr) contiguous).0wrrr 5sz$NnapiModule.init..)rAssertionErrorr preparer r rtorchclasses_nnapi Compilationinit2rr)rrrrrrinit0s   zNnapiModule.initreturnc Csd|jdur |||j}|durtddd|jD}t|t|jkr5tdt|dt|jg}tt|D],}|j|}|dkrR|||q=|dkrf||| ddd dq=t d | ||t|t|j krtd t|d t|j tt|jD] }|j |}|d vrq|dkr|| dd dd||<qt d |S)Nzcomp must not be NonecSsg|]}t|qSr)r empty_like)routrrrrFsz'NnapiModule.forward..z args length z != inp_mem_fmts length rrrzInvalid mem_fmtz outs length z != out_mem_fmts length )rr) rr$rr lenr rangeappendrpermute ValueErrorrunr )rrrouts fixed_argsidxfmtrrrforward@s>      zNnapiModule.forward)__name__ __module__ __qualname____doc__rrr r!r"__annotations__listTensornnModuleintboolrjitexportr$r3 __classcell__rrrrrs. &rFc Cst|||||\}}} } } } t||| | | ||} Gdddtjj}|| }tj|}dddtt |D}| dkrAd}n dd dt| D}| d |d |d |d |S)Ncs eZdZdZfddZZS)z5convert_model_to_nnapi..NnapiInterfaceWrappera0NNAPI list-ifying and de-list-ifying wrapper. NNAPI always expects a list of inputs and provides a list of outputs. This module allows us to accept inputs as separate arguments. It returns results as either a single tensor or tuple, matching the original module. cst||_dSr)rrmod)rrBrrrrs  z>convert_model_to_nnapi..NnapiInterfaceWrapper.__init__)r4r5r6r7rrArrrrNnapiInterfaceWrappersrCz, css|]}d|VqdS)arg_Nrrr1rrr sz)convert_model_to_nnapi..rz retvals[0]css|] }d|dVqdS)zretvals[z], NrrErrrrFszdef forward(self, z): retvals = self.mod([z]) return  ) process_for_nnapirrr;r<r?scriptjoinr*r)define)modelinputs serializer return_shapesuse_int16_for_qint16rrr ser_model_tensor used_weightsr r retval_count nnapi_modelrCwrapper_model_py wrapper_modelarg_listret_exprrrrconvert_model_to_nnapiisD   rZcCstj|}t|tjr|g}|ptd|d}||||\}}}}} } tj|tjd} Gdddtj j } tj | } dgdd| D}| d || | |||| fS) N)configrQ)dtypec@seZdZdZdS)z-process_for_nnapi..ShapeComputeModulezCode-gen-ed module for tensor shape computation. module.prepare will mutate ser_model according to the computed operand shapes, based on the shapes of args. Returns a list of output templates. N)r4r5r6r7rrrrShapeComputeModulesr]z\def prepare(self, ser_model: torch.Tensor, args: List[torch.Tensor]) -> List[torch.Tensor]: cSsg|]}d|dqS)z rHr)rlinerrrrsz%process_for_nnapi..rG)rr?freeze isinstancer:rserialize_modeltensorint32r;r<rJrLrK)rMrNrOrPrQr rSr r shape_compute_linesrTrRr]r real_shape_compute_linesrrrrIs:    rI)NNF) typingrr torch.backends._nnapi.serializerr ANEURALNETWORKS_PREFER_LOW_POWER)ANEURALNETWORKS_PREFER_FAST_SINGLE_ANSWER&ANEURALNETWORKS_PREFER_SUSTAINED_SPEEDr;r<rrZrIrrrrs  ^ <