o ZhC@s<ddlZddlZddlZddlZddlZddlZddlZddlmZm Z ddl Z e dZ GdddZ GdddZGdd d ZGd d d ZGd d d ejZd+ddZddZddZGdddeZGdddejZGdddeZddZddZdd Zd!d"Zd#d$ZGd%d&d&Zddd'd(d)d*ZdS),N) NamedTupleOptionalZnnapi_serializec@s@eZdZdZdZdZdZdZdZdZ dZ d Z d Z d Z d Zd ZdS)NNAPI_OperandCoder N)__name__ __module__ __qualname__FLOAT32INT32ZUINT32TENSOR_FLOAT32 TENSOR_INT32TENSOR_QUANT8_ASYMMBOOLTENSOR_QUANT16_SYMMZTENSOR_FLOAT16Z TENSOR_BOOL8ZFLOAT16ZTENSOR_QUANT8_SYMM_PER_CHANNELTENSOR_QUANT16_ASYMMrrO/var/www/auris/lib/python3.10/site-packages/torch/backends/_nnapi/serializer.pyrsrc@seZdZdZdZdZdZdZdZdZ dZ d Z d Z d Z d Zd ZdZdZdZdZdZdZdZdZdZdZdZdZdZdZdZdZdZ dZ!d Z"d!Z#d"Z$d#Z%d$Z&d%Z'd&Z(d'Z)d(Z*d)Z+d*Z,d+Z-d,Z.d-Z/d.Z0d/Z1d0Z2d1Z3d2Z4d3Z5d4Z6d5Z7d6Z8d7Z9d8Z:d9Z;d:Zd=Z?d>Z@d?ZAd@ZBdAZCdBZDdCZEdDZFdEZGdFZHdGZIdHZJdIZKdJZLdKZMdLZNdMZOdNZPdOZQdPZRdQZSdRZTdSZUdTZVdUZWdVZXdWZYdXZZdYZ[dZZ\d[Z]d\Z^d]Z_d^Z`d_Zad`S)aNNAPI_OperationCoderrrrrr r r r r rrr  !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^N)brrrADDAVERAGE_POOL_2D CONCATENATIONCONV_2DDEPTHWISE_CONV_2DZDEPTH_TO_SPACE DEQUANTIZEZEMBEDDING_LOOKUPZFLOORFULLY_CONNECTEDZHASHTABLE_LOOKUPZL2_NORMALIZATIONZ L2_POOL_2DZLOCAL_RESPONSE_NORMALIZATIONLOGISTICZLSH_PROJECTIONZLSTM MAX_POOL_2DMULRELURELU1RELU6RESHAPEZRESIZE_BILINEARZRNNSOFTMAXZSPACE_TO_DEPTHZSVDFZTANHZBATCH_TO_SPACE_NDDIVMEANZPADZSPACE_TO_BATCH_NDZSQUEEZE STRIDED_SLICESUB TRANSPOSEZABSZARGMAXZARGMINZAXIS_ALIGNED_BBOX_TRANSFORMZBIDIRECTIONAL_SEQUENCE_LSTMZBIDIRECTIONAL_SEQUENCE_RNNZBOX_WITH_NMS_LIMITZCASTZCHANNEL_SHUFFLEZDETECTION_POSTPROCESSINGEQUALZEXP EXPAND_DIMSZGATHERZGENERATE_PROPOSALSGREATERZ GREATER_EQUALZGROUPED_CONV_2DZHEATMAP_MAX_KEYPOINTZINSTANCE_NORMALIZATIONLESSZ LESS_EQUALLOGZ LOGICAL_ANDZ LOGICAL_NOTZ LOGICAL_OR LOG_SOFTMAXZMAXIMUMZMINIMUMZNEGZ NOT_EQUALZPAD_V2ZPOWPRELUQUANTIZEZQUANTIZED_16BIT_LSTMZRANDOM_MULTINOMIALZ REDUCE_ALLZ REDUCE_ANYZ REDUCE_MAXZ REDUCE_MINZ REDUCE_PRODZ REDUCE_SUMZ ROI_ALIGNZ ROI_POOLINGZRSQRTZSELECTZSINZSLICEZSPLITZSQRTZTILEZTOPK_V2TRANSPOSE_CONV_2DZUNIDIRECTIONAL_SEQUENCE_LSTMZUNIDIRECTIONAL_SEQUENCE_RNNRESIZE_NEAREST_NEIGHBORrrrrr%src@eZdZdZdZdZdZdS)NNAPI_FuseCoderrrrN)rrr FUSED_NONE FUSED_RELUZ FUSED_RELU1Z FUSED_RELU6rrrrr rc@seZdZdZdZdZdS)OperandValueSourceTyperrrN)rrr IMMEDIATENUMBERED_BUFFERZNUMBERED_MEMORYrrrrrsrc@seZdZdZdS)TorchScalarTypesrN)rrrQUINT8rrrrrsrư>cCst|||t||kSN)absmin)lhsrhsZ tolerancerrr approx_equalsrc Cs@tjdtjdtjdtjdtjdi}||}|D]}||9}q|S)Nrrr)rrrrrr)op_typedimsZ ITEM_SIZESsizedrrr tensor_sizes rcCst|}|||<t|Sr)listtuple)tupindexvalueZlsrrrchange_elementsrc@sjeZdZUdZeed<eed<eed<eed<eed<eed<eed<eed <eed <eed <eed <d S)ConvPoolArgs2dz*Configuration arguments for a convolution.kernel_hkernel_wstride_hstride_wpad_tpad_bpad_lpad_r dilation_h dilation_wgroupN)rrr__doc__int__annotations__rrrrrs  rc@r)DimOrderrrriN)rrrPRESUMED_CONTIGUOUS CHANNELS_LASTSCALAR_OR_VECTORUNKNOWN_CONSTANTrrrrrrrc@sJeZdZUdZeed<eedfed<eed<eed<eed<dd Z d S) Operandz"Represenation of an NNAPI operand.r.shape dim_orderscale zero_pointcCs(|jtjurdS|jtjurdStd)NTFzUnknown dim order)rrrr Exceptionselfrrruse_nchws  zOperand.use_nchwN) rrrrrrrrfloatrrrrrrs  rcCst|dksJt|dksJt|}t|}t|t|kr$tdt|t|kr0tdg}t||D]+\}}|dkrE||q7|dkrO||q7||krY||q7td|d|t|S)Nrz.Non-equal-rank broadcast is not supported yet.rzCannot broadcast shapes: z and )lenrrzipappendr)Zshape1Zshape2s1s2retd1Zd2rrrbroadcast_shapess0   rc Cs|\}}}}|jdks|jdkrtd|r7|d|j|j|j|j}|d|j|j|j |j } n ||j|j|j|jd}||j|j |j |jd} |dkr]d}|dkrcd} |||| f} | S)NrzDilation not supported yet.r) rrrrrrrrrrr) Z image_shapeargsZout_ch transposebatchZ_in_cin_hin_wout_hout_w out_shaperrrget_conv_pool_shape s  "   rcCs|tjur|S|tjur t|dgt|dd|dgS|tjur5t|dks3t|dks3J|S|tjur<|Std|d)NrrrzBad dim_order: .) rrrrrrrrrrrrrr fix_shape$s  (  rcCs.|tjtjfvr |S|tjusJgd|S)Nrrrr)rrrr)rrrrrreverse_map_dim5s rcCsd|d|S)NZs__r)op_iddimrrr flex_name@src@seZdZdddZddZddZdd Zd d Zd d Ze j fddZ ddZ ddZ ddZddZddZddZddZddZe jfd d!Zd"d#Zd$d%Zd&d'Zd(d)Zdd+d,Zdd-d.Zd/d0Zd1d2Zd3d4Zd5d6Zd7d8Z d9d:Z! *dd;d<Z"d=d>Z#dd?d@Z$dAdBZ%e&dCdDZ'idEdFdGdHdIdGdJdKdGdLdMdGdNdOdGdPdQdGdRdSdGdTdUdGdVdWdGdXdYdGdZd[dGd\d]dGd^d_dGd`dadGdbdcdGdddedGdfdgdGidhdidGdjdkdGdldmdGdndodGdpdqdGdrdsdGdtdudGdvdwdGdxdydGdzd{dGd|d}dGd~ddGdddGdddGdddGdddGdddGddGddGddGddGddGddGdZ(ddZ)ddZ*ddZ+ddZ,ddZ-ddZ.ddZ/ddZ0ddZ1ddZ2ddZ3ddZ4ddZ5ddZ6ddZ7ddZ8ddZ9d*dddZ:ddZ;ddZddZ?ddZ@ddÄZAddńZBddDŽZCddɄZDdd˄ZEdd̈́ZFddτZGddфZHdddӄZIddՄZJddׄZKddلZLdddۄZMdd݄ZNdd߄ZOd*S)_NnapiSerializerFcCstg|_g|_g|_g|_g|_g|_g|_g|_i|_i|_ i|_ i|_ i|_ g|_ d|_||_|dur8i}dSdSNr)operandsvalues operations value_dataoperation_argsinputsoutputs flexible_shape_computation_linesmodules constantstensor_sequencesjitval_operand_mapcached_immediates used_weightsZ weight_offsetuse_int16_for_qint16)rconfigrrrr__init__Gs&z_NnapiSerializer.__init__cCs t|jSr)rrrrrrget_next_operand_id] z$_NnapiSerializer.get_next_operand_idcCsHt|tsJ||jvrtd||}|j|||j|<|S)NzDuplicate tensor: ) isinstancerrrrrr)rjitvaloper operand_idrrradd_tensor_operandcs   z#_NnapiSerializer.add_tensor_operandcCs&t|tsJ|}|j||Sr)rrrrr)rrrrrradd_anonymous_tensor_operandps z-_NnapiSerializer.add_anonymous_tensor_operandc Cst|jdd}d}d}|dkrtj}nc|dkrtj}n[|dkr-tj}|}|}nK|dkrCtj}|}|}|dksBJn5|d kro|j rkt |d d}tj tj f}||vrc|}|j }|j}ntd |d td td|jdtt|j||||dS)Nztorch.rZfloat32Zint32quint8qint32Zint16 nnapi_dtypez `nnapi_type` needs to be one of z for `int16`y`int16` isn't supported. If you're trying to represent NNAPI qint16 with Pytorch int16, set `use_int16_for_qint16 = True`zCan't handle input with dtype '')rrrrr)strdtypereplacerrrrq_scale q_zero_pointrgetattrrrZ nnapi_scaleZnnapi_zero_pointrrrr) rtensorrrrrrrZop_codesrrrtorch_tensor_to_operandvsT    z(_NnapiSerializer.torch_tensor_to_operandc Csxt|ddr tjntj}|||}|||}|j|t|j D]\}}|dkr9| ||d|d|dq#|S)NZ nnapi_nhwcFrzargs[z].shape[]) r rrrr rrr enumeratercompute_operand_shape) rarg_idxrr rtoperrrrrrradd_tensor_operand_for_inputs    z-_NnapiSerializer.add_tensor_operand_for_inputcCs|||}t|j}|j|t|j|j}|j|tj ft|j }d}|j t d||||tjkrA|dddd}|j ||S)Nriiirrr)r rrrrrrrrrrrstructpackrrpermute)rr rrrZtsizeZbuf_numoffsetrrradd_tensor_operand_for_weights      z._NnapiSerializer.add_tensor_operand_for_weightcCsvt|tsJ||f}||jvr6t|j}|jt||tjdd|j |t j f|j |||j|<|j|S)Nrr) rrrrrrrrrrrrr)rcoderr cache_keyrrrradd_immediate_operands     z&_NnapiSerializer.add_immediate_operandcC|tjtd|dS)Nir)rrrrrrrrrradd_immediate_int_scalarz)_NnapiSerializer.add_immediate_int_scalarcCr)Nfr)rrrrrrrrradd_immediate_float_scalarr z+_NnapiSerializer.add_immediate_float_scalarcCs|tj|r ddSddS)Nr)rrrrrrradd_immediate_bool_scalars  z*_NnapiSerializer.add_immediate_bool_scalarcCs"|tjtd|t|fSNr)rrrarraytobytesrrrrradd_immediate_int_vectors z)_NnapiSerializer.add_immediate_int_vectorcCs ||jvSr)r)rrrrrhas_operand_for_jitvalrz'_NnapiSerializer.has_operand_for_jitvalcCs|j|}||j|fSr)rr)rrrrrrget_tensor_operand_by_jitvals z-_NnapiSerializer.get_tensor_operand_by_jitvalcCsF||\}}|jD]}|dkrtd|dkrtd|q ||fS)Nrz0Flexible size is not supported for this operand.z!Operand %s has runtime flex shape)r+rrrwarning)rrrrsrrr'get_tensor_operand_by_jitval_fixed_sizes  z8_NnapiSerializer.get_tensor_operand_by_jitval_fixed_sizecCs>|j|}|dur||d\}}|||}||j|fSN TensorType)rgetget_constant_valuerr)rrrrrrrrrget_tensor_operand_or_constants  z/_NnapiSerializer.get_tensor_operand_or_constantcCs(||d\}}||}||j|fSr/)r2rr)rrrrrrrrget_tensor_operand_for_weights z._NnapiSerializer.get_tensor_operand_for_weightcCs.|j|t|t|f|j||dSr)rrrrextend)ropcoderrrrr add_operation sz_NnapiSerializer.add_operationcCs||jvsJ||j|<dSr)r)rrrrrradd_tensor_sequencesz$_NnapiSerializer.add_tensor_sequencecCs ||jvsJ||f|j|<dSr)rrrctyperrrradd_constant_valuesz#_NnapiSerializer.add_constant_valueNcCsd|j|}|durtd|d|\}}|dur0||kr0td|d|d|d|S)Nz#Could not find constant value for 'z'.z Expected constant value of type z , but got z for value 'r)rr1rkind)rrZtypekindrecordr:rrrrr2s  z#_NnapiSerializer.get_constant_valuecCs>|dur|j}n t|t|jksJdg}t|D]0\}}|dkr*|t|n|dkr7|t||n|dkrA|dntd|dq|dd |}|jt j krad |d S|jt j krmd |d S|jt j krd |j d|jd|dS|jt jt jfvr|jrd |dStdtd|j)zHReturn a TorchScript expression to build a template for a given operand.N(r0z-Unknown dim value, dimensions should be >= -1,)rz torch.zeros(z, dtype=torch.float32)z, dtype=torch.int32)z0torch.quantize_per_tensor(torch.zeros(1), scale=z , zero_point=z, dtype=torch.quint8).expand(z).contiguous()z, dtype=torch.int16)rz!Unsupported output operand type: )rrrrrrrjoinrrrrrrrrrr)rrrrZ shape_partsrr-Z shape_coderrroperand_to_template_torchscript%sT           z0_NnapiSerializer.operand_to_template_torchscriptcCs|||t||dSr)rr)rZ out_op_idZout_dimZin_op_idZin_dimrrrforward_operand_shapeXsz&_NnapiSerializer.forward_operand_shapecCs |jt||d|dS)Nz = )rrr)rrrexprrrrr[sz&_NnapiSerializer.compute_operand_shapecCs|jdddkr td|jtjd}dgd}||d<|gd|d<dgd}|||d<|tj |||d|fS)Nr)rrz1Automatic transpose only supported for H,W == 1,1)rrrr) rr_replacerrr)rr7rr)rin_idrout_operrrrrrtranspose_to_nhwc`s   z"_NnapiSerializer.transpose_to_nhwccCs|j|jkr ||||fS|j|jf}|tjtjfkr$|||||fS|tjtjfkr6||f|||Std|jd|j)Nz2Automatic transpose not supported for dim_orders: z, )rrrrrJr)rin0_idin0_operin1_idin1_operZordersrrrtranspose_for_broadcastts   z(_NnapiSerializer.transpose_for_broadcastcCsH||\}}|dkr|dksJ|Std|d|d)NListTypeIntTypezCan't handle size arg of type 'z' for 'r)r2r<getElementTyperr9rrr get_size_args z_NnapiSerializer.get_size_argc Csdd|D}|ddksJ|d|dg}|d|dg}|d|d g}|d |d g}|d }t|d ks;J|ddgksCJ||||||S)NcSsg|]}|qSr)item.0rrrr szD_NnapiSerializer.get_conv_pool_args_2d_from_pack..rrrrrr r r r r r)rget_conv_pool_args_2d_common) r kernel_size packed_configZpcstridespaddings dilationsZoutput_padding group_numrrrget_conv_pool_args_2d_from_packs z0_NnapiSerializer.get_conv_pool_args_2d_from_packc Cs`||}||}|durddg}n||}|dur%||d\} } nd} |||||| S)NrrQ)rSr2rX) rrYstridepaddingdilationrr[r\r]rr^rrrget_conv_pool_args_2d_from_jits     z/_NnapiSerializer.get_conv_pool_args_2d_from_jitc Csvt|}t|dks Jt|dksJt|dksJt|dks$J|\}}||||g} t||| ||gSNr)rrr) rrYr[r\r]r^ZkernelsphpwZ real_paddingsrrrrXs z-_NnapiSerializer.get_conv_pool_args_2d_commonc Cs|d|dg}g}t|j}||||ttt|jdd|D]\}\}} | ||| } | |j | j j q.t|jD]\} } td| | || qM|j} | dkskJ| dkssJ| d}dg}|dkr|g}d}n|d kr|j|}t|}n td ||durt|t|ksJt|D]0\}}|j|} |j | | |j | j j |r||nd}| || |j | |d q| d g}d}td |t|j t|jt|j t|jt|j}| ||!\}}|"dd|j D|"||"dd|j Dd#|g}t|d}|ddksEJt$|d}t|j D]G\} \}}}}}t%||}t|D]"\}}|dkrt&||}|j' d|dt(| ||d7}qbt)dd|D}| |*|qP|"|| |*|j+| |*|j| |*|j|j'"|t,,dd#||j-|||j'|fS)NFTrzProcessing node #%d: %rrzreturn [r0r?Z TupleTypezUnsupported return type: rAr Ziiiiiicss0|]\}}}}}td|t|||VqdS)ZiifiN)rrr)rVtr_mr-zrrr s  z3_NnapiSerializer.serialize_model..css"|] }tjdg|RVqdS)rN)rr)rVxrrrrjs rz ser_model[z] = cs |] }|dkr |ndVqdS)r?rNrrVrrrrrjr).r%nextgraphrr;typerrrrrrrrnodesrdebugadd_nodeZ return_node inputsSize outputsSizeinputsAtr<rrrrrrDrrrrserialize_valuesr5rCrrrrrrserialize_intsrr'r) rmodelr return_shapesZinp_dim_ordersZout_dim_ordersZ self_jitvalrZ input_valueZ input_tensorridxnodeZretnZ retn_inputZtemplate_return_linesZ return_valuesZ retval_countrvrversionheaderserialized_valuesserialized_value_dataZ model_offsetrrrrr-Zpt_drrrserialize_models                       z _NnapiSerializer.serialize_modelc Csg}g}t|jt|jksJt|j|jD]*\\}}}t|}|ddBd}|d||}|td|||||q||fS)Nrrr$r)rrrrrrr) rrrZop_indexZ source_typedataZ source_lengthZphysical_lengthZ padded_datarrrry/s z!_NnapiSerializer.serialize_valuescCstd|Sr&)r'r()ZintsrrrrzAsz_NnapiSerializer.serialize_intsz prim::GetAttrcC ||Sr) add_getattrrr~rrrF z_NnapiSerializer.zprim::ConstantcCrr)add_constant_noderrrrrGrzprim::ListConstructcCrr)add_list_constructrrrrrHrzprim::TupleConstructcCrr)add_tuple_constructrrrrrIrzaten::unsqueezecCrr) add_unsqueezerrrrrJrzaten::tocCrr)add_torrrrrKrz aten::detachcCrr _identityrrrrrLrz aten::reshapecCrr) add_reshaperrrrrMrz aten::flattencCrr) add_flattenrrrrrNrz aten::slicecCrr) add_slicerrrrrOrz aten::sizecCrr)add_sizerrrrrPrz aten::catcCrr)add_catrrrrrQrz aten::meancCrr)add_meanrrrrrRrzaten::quantize_per_tensorcCrr) add_quantizerrrrrSrzaten::dequantizecCrr)add_dequantizerrrrrTrz aten::addcC||tjtjSr)add_add_sub_oprrqrrrrrrrU z aten::subcCrr)rrrrrrrrrrXrz aten::mulcCrr)(add_pointwise_simple_binary_broadcast_oprrzrrrrrrr[rz aten::divcCrr)rrrrrrrrrr^rz aten::relucC||tjSr)add_pointwise_simple_unary_oprr{rrrrraz aten::sigmoidcCrr)rrrxrrrrrdrz aten::softmaxcCrr) add_softmaxrrrrrgrzaten::hardtanhcCrr) add_hardtanhrrrrrhrzaten::avg_pool2dcCrr)add_avg_pool2drrrrrirzaten::max_pool2dcCrr)add_pool2d_noderryrrrrrjrzaten::adaptive_avg_pool2dcCrr)add_adaptive_avg_pool2drrrrrmzaten::upsample_nearest2dcCrr)add_upsample_nearest2drrrrrprz aten::prelucCrr) add_prelu_oprrrrrsrz aten::addmmcCrr) add_addmmrrrrrtrz aten::linearcCrr) add_linearrrrrrurzaten::_convolutioncCrr)add_conv_underscorerrrrrvrz aten::conv2dcCrr) add_conv2drrrrrwrzaten::log_softmaxcCrr)add_log_softmaxrrrrrxrzquantized::linearcCrr) add_qlinearrrrrryrcCrr add_qconv2drrrrrrrzrcCrr)rrrrrrrr}rcCs|j|tjddS)NT)rrrrrrrscCrr)add_qaddrrqrrrrrrrrcCrr)rrrqrrrrrrrrcCrr)rrrzrrrrrrrr)zquantized::conv2dzquantized::conv2d_reluzquantized::conv_transpose2dzquantized::addzquantized::add_reluzquantized::mulcCs:|j|}|std|d||||dS)NzUnsupported node kind (z ) in node ) ADDER_MAPr1r<r)rr~Zadderrrrrus z_NnapiSerializer.add_nodecCs,||d\}}|d}||j|<dSr)r+rx outputsAtr)rr~rHZ_in_operrrrrrs z_NnapiSerializer._identitycCs~|dksJ|dksJ||d\}}t|ds#J|d}t||}|d}| }| |||dS)Nrrz __torch__.name) rvrwr2rxr startswithr-r rrrr;)rr~Z obj_ctypeobjrroutputr:rrrrs   z_NnapiSerializer.add_getattrcCsL|dksJ|dksJ|d}|}|}||||dS)Nrr)rvrwrrrZtoIValuer;)rr~rr:rrrrrs  z"_NnapiSerializer.add_constant_nodec Cs|dksJ|d}|}g}g}|D].}|dur1||jvr1||\}}||nd}|durE|dkrE||qd}q|durS|||||dur]| |||durl|durnt d|dSdS)Nrrr0zMUnable to handle ListConstruct node. Neither all constants nor all tensors. ) rwrrrrrr2rr<r;r8r) rr~rr:Z const_valstensorsinprvalrrrrs,     z#_NnapiSerializer.add_list_constructcCs6|dksJ|d}t|}|||dS)Nrr)rwrrrr8)rr~rrrrrrs  z$_NnapiSerializer.add_tuple_constructc Cs|dksJ|dksJ||d\}}||dd\}}|jtjks-J|dkr3|n|t|j d}t |j }| |dt |}|j |d} dgd} || d<||| d<dgd} ||d| | d<|tj| | dS)NrrrrQr)rvrwr.rxr2rrrrrrinsertrrGrrrr7rr) rr~rHin_operrrZreal_dimZout_shape_listrrIrrrrrrs      z_NnapiSerializer.add_unsqueezecCs||dSrrrrrrrsz_NnapiSerializer.add_toc Cs|dksJ|dksJ||d\}}||d\}}|dks,J|dks6Jt|dkoA|ddk}|jt j krN|sNt dt d|j|j}|j|t j d}dgd} || d<||| d<dgd} ||d|| d<|tj| | dS) NrrrrPrQr?zSCurrently, reshape is only supported on NHWC tensors if the target size is [X, -1].r)rvrwr.rxr2r<rRrrrrrtorchzerosexpandrZreshaperGr)rrr7rr~) rr~rHrZ shape_ctyperZis_trivial_reshaperrIrrrrrrs*  z_NnapiSerializer.add_reshapec Cs$|dksJ|dksJ||d\}}||dd\}}||dd\}}t|jdkoK|jddkpK|jddkoK|jddk}|jtj krX|sXt d|dkrc|t|j7}|dkrn|t|j7}|jd|t t j|j||df|j|dd} tdd |j||dDrt d |jd||j|dd} | ddkrt d |j| tj d } ||d| } t| D]\} }|dkr|| | ||jdqtd d | D}dgd}||d<|||d<dgd}| |d<|tj||dS)NrrrrQrrzGCurrently, flatten is not supported on NHWC tensors unless C=1 or H=W=1css|]}|dkVqdS)rNrrVrrrrrj(sz/_NnapiSerializer.add_flatten..z-Flattening flexible dims is not supported yetzOnly 1 dim can be flexiblercsrm)rr?Nrrrrrrj9ro)rvrwr+rxr2rrrrrr functoolsreduceoperatormulanycountrGrrrrErrr)r7rr~)rr~rHrZ _start_ctypeZ start_dimZ _end_ctypeZend_dimZis_trivial_flattenrZnon_flattened_dimsrIout_idr}rZinputs_1rrrrrr sX(     z_NnapiSerializer.add_flattenc s|dksJ|dksJ||d\}}||d\}||d\}||d\}||d\}durHddurOtjdkr[|j7ntjkrbddkrrtjkrr||dS|jdkr}t ddkr|j7n tjkr|jkrt dt fd d t |jD}| | d|j|d }d}t |D]\}} | dkr||||||d|>O}qdgd } || d<|fd dtt|jD| d<|fddt |jD| d<|fddtt|jD| d<|d| d<||| d<|d| d<dgd} || d<|tj| | dS)Nr rrrrrz#Unable to slice with flexible shapez0Slice start value should be less than stop valuec3s$|] \}}|kr n|VqdSrrrVrr) dim_valueout_lenrrrjis z-_NnapiSerializer.add_slice..rr cg|] }|kr ndqS)rrrU)r start_valuerrrWzz._NnapiSerializer.add_slice..cs g|] \}}|kr n|qSrrr)r stop_valuerrrW}scr)rrrU)r step_valuerrrWrr )rvrwr+rxr2sysmaxsizerrrrrrrrGrEr)rangerrr7rr) rr~rHrrrrZend_maskr}rrrr)rrrrrrrCsx         z_NnapiSerializer.add_slicecCsr|dksJ|dksJ||d\}}|j|d\}}|j|}|d}||||dS)Nrrr) rvrwr.rxrrrr;rr)rr~rrrresrrrrrs  z_NnapiSerializer.add_sizecs|dksJ|dksJ|j|d}||dd\}t|dks+Jg}d}d}|D]F}||\}} |durMt| jd} | j | d}| j |j ksUJ| j |j ks]Jt| jdt|jdksmJ| ||| j7}q3|dusJ|j t|j|d}| j t jkrt|jdksJgd} n} ||d|} t|jD]*\} }|dkr| krd fd d |D}|| | |q|| | |d| q||| g}dgd}| |d<|tj||dS) NrrrrQr?rrrrrr + c3s|]}t|VqdSr)r)rVZip_idrrrrjz+_NnapiSerializer.add_cat..)rvrwrrxr2rr+rrrGrrrrrrrrrCrrErr7rrs)rr~rrZin_idsrIZ out_dim_sizerrHrr nnapi_dimrr}rrrrrrrrsP      z_NnapiSerializer.add_catcCs|dksJ|dksJ||d\}}||d\}}|dks,J|dks6J||dd\}}||dd |jtj krat |j dksYJd d |D}n|}t } |D]} | dkru| t |j 7} | | qh|jtj kr|s| ddhsJtj} n|j} g} t|j D]\} }| | vr| |q|r| dq|j| | d }dgd}||d<|||d<|||d<dgd}||d||d<|tj||dS) NrrrrPrQrZBoolTyperNoneTypecSsg|]}gd|qS)rrrnrrrrWsz-_NnapiSerializer.add_mean..r)rvrwr.rxr2r<rRrrrrrsetadd issupersetrrrrGr)rrrr7rr)rr~rHrZ dim_ctyperrZkeep_dimrZcollapsed_dimsrZ out_dim_orderrrr-rIrrrrrrsH      z_NnapiSerializer.add_meanc Cs|dksJ|dksJ||d\}}|jtjkr$td||dd\}}||dd\}}||dd\}}|t j j krOtd t j }|j|||d } dgd} || d<dgd} ||d| | d<|tj| | dS) NrrrzqMost hardware backends prefer NHWC quantized tensors. Try setting `t.nnapi_nhwc = True` on your tensor inputs. FloatTyperrQrzKPyTorch NNAPI export only supports quantized tensors with the quint8 dtype.rrr)rvrwr.rxrrrrr2rrrrrrGrrr7rr) rr~rHrrrrZ scalar_typerrIrrrrrrs2    z_NnapiSerializer.add_quantizecCs|dksJ|dksJ||d\}}|jtjddd}dgd}||d<dgd}||d||d<| t j ||dS)Nrrrr) rvrwr.rxrGrrrrr7rrv)rr~rHrrIrrrrrr s  z_NnapiSerializer.add_dequantizec Cs|dksJ|dksJ||d\}}|}|tjkr.|jtjkr.|j ddd}| | d|}t |j D]\}}|dkrL|||||q.rrrcsrr)rrrrrrrjrz#Size and scale cannot both be None.rz(Flexible batch or channels not supported)rrzint(z * rB)rvrwrr2r<r+rrrrRrrallrrr"rrrrGrrr%r7rr)rr~rZsize_jitZ scale_jitZ scale_h_jitZ scale_w_jitrr scale_ctypeZ scale_argZ scale_h_ctypeZ scale_h_argZ scale_w_ctypeZ _scale_w_argrrrrZarg_hZarg_wrrrrrrrrrr~s           z'_NnapiSerializer.add_upsample_nearest2dc Cs|dksJ|dksJ|\}}}}}||fD]}||\}} |dvs.J| dkr6tdq||d|||dS)Nr r)rQrz6NNAPI Fully-Connected does not support alpha and beta.T)rvrwrr2r<radd_addmm_or_linear) rr~jit_bias jit_input jit_weightZjit_betaZ jit_alpharrZ scale_valuerrrrs z_NnapiSerializer.add_addmmcCsD|dksJ|dksJ|\}}}||d|||dS)NrrF)rvrwrr)rr~rrrrrrrsz_NnapiSerializer.add_linearcCs4||\}}||\}} t|jdksJt| jdks J||d\} } t| jdks1J|r:| } n| } || } |j| }|jd|jdf}| | d|j |d}|jddkrn| |d|ddgd}||d<| |d<||d<| tj|d<dgd}||d<|tj||dS)Nrrr0rrrr)r+r4rrr2rg contiguousrrrrrGrErrrr7rrw)rr~Ztranspose_weightrrrinput_id input_operbias_id bias_operr weight_tensornnapi_weight_tensor weight_id weight_operrrrrrrrrs2    z$_NnapiSerializer.add_addmm_or_linearcCs`|dksJ|dksJ|\}}}}||\}}t|jdks(J||d\}} ||d\}} ||\} } | dksGJ| d\} }|dusUJt| jdks^Jt|jdksgJ|jd| jdkssJ| jd|jdksJ| t j ksJ| j t j kr| }n | j t jksJt j| dt j| | dd }|}|j|}t ||dt j}||}|j|| }|dksJ|dkrtd |}||}|j|}|jd|jdf}|j|| | d }dgd}||d<||d<||d<|tj |d <dgd}|!|"d||d<|#t$j%||dS) NrrrrrQZLinearPackedParamsBaserrQuantized convolution multiplier is greater than 1. This is supported by NNAPI, but not by most hardware backends. Try training a model without quantization-aware training. rrrr)&rvrwrr.rrr2r __getstate__qschemerper_tensor_affinerrqint8!_make_per_tensor_quantized_tensorint_reprrtouint8rr rquantize_per_tensorrrrrrrGrrrrrr7rrw)rr~rjit_packed_weight jit_scalejit_zero_pointrrr out_scaleout_zero_point weight_ctype packed_weight raw_weightraw_biasunsigned_weight weight_scale bias_scaleint_biasr multiplierrrrrrIrrrrrrsp          z_NnapiSerializer.add_qlinearc Csd||\}}|dkr-|rdnd}tj|||jd}||}|j|} || fS||S)Nrrr)r) r2r<rrrrrrr4) rrrrr:_valueZbias_idxZnnapi_bias_tensorrrrrrget_optional_biases     z"_NnapiSerializer.get_optional_biasc Cs|dksJ|dksJ|\}}}}}}}||d\} } ||| \} } || jdd||||} ||ddd|| | | dt j S) Nr rr0rrrrF rvrwrr2rrcradd_conv2d_commonrrr)rr~ jit_imagerr jit_stridejit_pad jit_dilation jit_groupsrrr _bias_operrrrrrrs6 z_NnapiSerializer.add_conv2dc Cs|dksJ|dksJ|\ }}}}}}}} } } } } } ||d\} } ||\} } ||| | \} }|| jdd|||| }||ddd|| | || t j S)Nrrr0rrrrr)rr~rrrrrr Z jit_transposerr!rrrr"rrrrrsDz$_NnapiSerializer.add_conv_underscorec Cs|dksJ|dksJ|\}}}||\}}||d\}}|j} dgd} || d<|d| d<||| d<dgd} || d|j | d| d<| t j | | dS)NrrrQrrr)rvrwrr.r2rr"rrrrGr7rr) rr~rZjit_dimZ_jit_half_to_floatrrrrrrrrrrrs  z _NnapiSerializer.add_log_softmaxc Cs|dksJ|dksJ|\}}}}||d\}} ||d\}} ||\} } | dks7J| d\} }}| dksFJ|\}}|\}|dusSJ||jdd|}|t j ksgJ|j t j krp|}n |j t j ksxJt j|d t j||d d }|}||\}}|j|}t ||dt j}||}|j|| }|dksJ|dkrtd ||d| | |||||| S) NrrrrQZConv2dPackedParamsBaser2rrrr)rvrwrr2rrr_rrrrrrrrrrrr rr r+rr rrrrr)rr~rrrr r r rrrrrZ pack_versionrZ opt_tensorsrZrrrrrrrrrrrrrrsn        z_NnapiSerializer.add_qconv2dc !Cs ||\} } | jd} |jdkrd} |rd}nd}n|j| kr%d} d}ntd|j|}||}|j|}|j|}| jt j krV|jt j ksMJ|jt j ksUJn2| jt j kr|jt j ksdJ|jt j kslJt | j|j|jsxJ|jdksJntd| jt| jd ksJt|jd ksJt|jdksJ| r|j\}}}}|dksJ|| dksJ|| }|dksJ|| ksJn |j\}}}}|| ksJ||jdksJ| }| rd }tj}n d }|rtj}ntj}dg|}| |d<||d<||d <||j|d <||j|d <||j|d<||j|d<||j|d<||j|d<| rR|d|d<|| |d<|||d <n|| |d<|||d<dgd}t| j|||}| j|||d}| ||} |!| |||| |d<|"|||dS)NrF)rrrrrTz$Group convolution not supported yet.rz#Unsupported input type for conv2d: rrrrrr r r r r rr)#r+rrrrrrrrrrrrrrrrrrrurrtrrrrrrrr%rrGrrr7)!rZjit_outrrrrrrrrrrZin_cZ depthwiseZweight_permutationrrrroneZ_kern_hZ_kern_wZout_cZchannel_multiplierZkern_drnum_argsr6rrrrIrrrrrs               z"_NnapiSerializer.add_conv2d_commonc Csd||\}}|j\}}} } |dkr||d|d|dkr"td|rj| dkrE||ddt|dd|jd|jd|jd|j | dkrh||ddt|dd|j d|j d|j d|j dSdS| dkr||ddt|dd|jd|jd|j d |jd | dkr||ddt|dd|j d|j d|j d |j d dSdS) Nrz Input channels can't be flexiblerr>z - 1) * rz - rz) // z + 1)r+rrErrrrrrrrrrr) rrrrrrrrZin_chrrrrrrsB..00z1_NnapiSerializer._handle_conv_pool_flexible_input)Fr)NN)Prrrrrrrr rrrrrrr"r%r)r*r+r.rr3r4r7r8r;r2rDrErrJrOrSr_rcrXrry staticmethodrzrrurrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrFs2  1      3   p      "#$%(+./012345I 8K 51"7 &+6(b& I  ' A orF)rr|rcCst|||||S)aConvert to NNAPI and serialize torchscript module. Parameters: module: Torchscript module to convert inputs: Tensors used to specify input details for NNAPI config (optional): Optional config to attach to module return_shapes (optional): Specify shape of outputs if your module uses runtime flexible shapes to set output buffer size for NNAPI use_int16_for_qint16 (optional): Use Pytorch int16 to represent NNAPI qint16 values )rr)modulerrr|rrrrrs r)r) r'enumrloggingrrrtypingrrr getLoggerrrrrrEnumrrrrrrrrrrrrrrrrrrsV b  ! m