U ʹj=@s<ddlZddlZddlZddlZddlZddlZddlmZmZm Z m Z ddl Z e dZ GdddZGdddZGdd d ZGd d d ZGd d d ejZd+ddZddZddZGdddeZGdddejZGdddeZddZddZdd Zd!d"Zd#d$ZGd%d&d&Zddd'd(d)d*Z dS),N)List NamedTupleOptionalTupleZnnapi_serializec@s@eZdZdZdZdZdZdZdZdZ dZ d Z d Z d Z d Zd ZdS)NNAPI_OperandCoder N)__name__ __module__ __qualname__FLOAT32INT32ZUINT32TENSOR_FLOAT32 TENSOR_INT32TENSOR_QUANT8_ASYMMBOOLTENSOR_QUANT16_SYMMZTENSOR_FLOAT16Z TENSOR_BOOL8ZFLOAT16ZTENSOR_QUANT8_SYMM_PER_CHANNELTENSOR_QUANT16_ASYMMrr_/var/www/html/bwave/python/venv/lib/python3.8/site-packages/torch/backends/_nnapi/serializer.pyrsrc@seZdZdZdZdZdZdZdZdZ dZ d Z d Z d Z d Zd ZdZdZdZdZdZdZdZdZdZdZdZdZdZdZdZdZdZ dZ!d Z"d!Z#d"Z$d#Z%d$Z&d%Z'd&Z(d'Z)d(Z*d)Z+d*Z,d+Z-d,Z.d-Z/d.Z0d/Z1d0Z2d1Z3d2Z4d3Z5d4Z6d5Z7d6Z8d7Z9d8Z:d9Z;d:Zd=Z?d>Z@d?ZAd@ZBdAZCdBZDdCZEdDZFdEZGdFZHdGZIdHZJdIZKdJZLdKZMdLZNdMZOdNZPdOZQdPZRdQZSdRZTdSZUdTZVdUZWdVZXdWZYdXZZdYZ[dZZ\d[Z]d\Z^d]Z_d^Z`d_Zad`S)aNNAPI_OperationCoderrrr r r r r rrrrr  !"#$%&'()*+,-./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 #sr c@seZdZdZdZdZdZdS)NNAPI_FuseCoderrrr N)rrr FUSED_NONE FUSED_RELUZ FUSED_RELU1Z FUSED_RELU6rrrrrsrc@seZdZdZdZdZdS)OperandValueSourceTyperrr N)rrr IMMEDIATENUMBERED_BUFFERZNUMBERED_MEMORYrrrrrsrc@seZdZdZdS)TorchScalarTypesr!N)rrrQUINT8rrrrrsrư>cCst|||t||kSN)absmin)lhsrhsZ tolerancerrr approx_equalsrc Cs@tjdtjdtjdtjdtjdi}||}|D] }||9}q.|S)Nr rr)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@seZdZdZdZdZdZdS)DimOrderrrriN)rrrPRESUMED_CONTIGUOUS CHANNELS_LASTSCALAR_OR_VECTORUNKNOWN_CONSTANTrrrrrsrc@sJeZdZUdZeed<eedfed<eed<eed<eed<dd Z d S) Operandz"Represenation of an NNAPI operand.r.shape dim_orderscale zero_pointcCs,|jtjkrdS|jtjkr dStddS)NTFzUnknown dim order)rrrr Exceptionselfrrruse_nchws   zOperand.use_nchwN) rrrrrrrrfloatrrrrrrs rcCst|dkstt|dks tt|}t|}t|t|krHtdt|t|kr`tdg}t||D]X\}}|dkr||qn|dkr||qn||kr||qntd|d|qnt|S)Nrz.Non-equal-rank broadcast is not supported yet.rzCannot broadcast shapes: z and )lenAssertionErrorrrzipappendr)Zshape1Zshape2s1s2retd1Zd2rrrbroadcast_shapess$   rc Cs|\}}}}|jdks |jdkr(td|rn|d|j|j|j|j}|d|j|j|j |j } n@||j|j|j|jd}||j|j |j |jd} |dkrd}|dkrd} |||| f} | S)NrzDilation not supported yet.r) rrrrrrrrrrr) Z image_shapeargsZout_ch transposebatchin_cin_hin_wout_hout_w out_shaperrrget_conv_pool_shapes  "   rcCs|tjkr|S|tjkr@t|dgt|dd|dgS|tjkrjt|dksft|dksft|S|tjkrx|St d|ddS)NrrrzBad dim_order: .) rrrrrrrrrrrrrrr fix_shapes  (  rcCs2|tjtjfkr|S|tjks"tddddg|S)Nrrr r)rrrrr)rrrrrreverse_map_dim-srcCsd|d|S)NZs__r)op_iddimrrr flex_name8src)@sfeZdZdddZddZddZdd 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'dEdFdGdFdHdFdIdFdJdFdKdFdLdFdMdFdNdFdOdFdPdFdQdFdRdFdSdFdTdFdUdFdVdFdWdFdXdFdYdFdZdFd[dFd\dFd]dFd^dFd_dFd`dFdadFdbdFdcdFdddFdedFdfdFdgdFdhdFdidFdjdFdkdFdldFdmdFdn(Z(dodpZ)dqdrZ*dsdtZ+dudvZ,dwdxZ-dydzZ.d{d|Z/d}d~Z0ddZ1ddZ2ddZ3ddZ4ddZ5ddZ6ddZ7ddZ8ddZ9d*dddZ:ddZ;ddZddZ?ddZ@ddZAddZBddZCddZDddZEddZFddZGddZHdddZIddZJddZKddZLdddZMddZNddZOd*S)_NnapiSerializerFcCspg|_g|_g|_g|_g|_g|_g|_g|_i|_i|_ i|_ i|_ i|_ g|_ d|_||_|dkrli}dSNr)operandsvalues operations value_dataoperation_argsinputsoutputs flexible_shape_computation_linesmodules constantstensor_sequencesjitval_operand_mapcached_immediates used_weightsZ weight_offsetuse_int16_for_qint16)rconfigrrrr__init__?s$z_NnapiSerializer.__init__cCs t|jSr)rrrrrrget_next_operand_idUsz$_NnapiSerializer.get_next_operand_idcCsHt|tst||jkr&td||}|j|||j|<|S)NzDuplicate tensor: ) isinstancerrrrrrr)rjitvaloper operand_idrrradd_tensor_operand[s   z#_NnapiSerializer.add_tensor_operandcCs&t|tst|}|j||Sr)rrrrrr)rrrrrradd_anonymous_tensor_operandhs z-_NnapiSerializer.add_anonymous_tensor_operandc Cs t|jdd}d}d}|dkr*tj}n|dkr:tj}n|dkrZtj}|}|}n|dkrtj}|}|}|dkst nn|d kr|j rt |d d}tj tj f}||kr|}|j}|j}qtd |d qtd ntd|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_pointrrgetattrrrZ nnapi_scaleZnnapi_zero_pointrrrr) rtensorrrrrrrZop_codesrrrtorch_tensor_to_operandnsP   z(_NnapiSerializer.torch_tensor_to_operandc Csxt|ddrtjntj}|||}|||}|j|t|j D],\}}|dkrF| ||d|d|dqF|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_inputc Cs|||}t|j}|j|t|j|j}|ddBd}|j|tj ft|j }d}|j t d||||tjkr|dddd}|j ||S)Nrr riiir)r rrrrrrrrrrrstructpackrrpermute) rr rrrZtsizepsizeZbuf_numoffsetrrradd_tensor_operand_for_weights      z._NnapiSerializer.add_tensor_operand_for_weightcCsvt|tst||f}||jkrlt|j}|jt||tj dd|j |t j f|j |||j|<|j|S)Nrr)rrrrrrrrrrrrrr)rcoderr cache_keyrrrradd_immediate_operands    z&_NnapiSerializer.add_immediate_operandcCs|tjtd|dS)Nir)rrrrrrrrrradd_immediate_int_scalars  z)_NnapiSerializer.add_immediate_int_scalarcCs|tjtd|dS)Nfr)rrrrrrrrradd_immediate_float_scalars  z+_NnapiSerializer.add_immediate_float_scalarcCs|tj|rdnddS)Nr)rrrrrrradd_immediate_bool_scalars  z*_NnapiSerializer.add_immediate_bool_scalarcCs"|tjtd|t|fSNr)rrrarraytobytesrrrrradd_immediate_int_vectors z)_NnapiSerializer.add_immediate_int_vectorcCs ||jkSr)r)rrrrrhas_operand_for_jitvalsz'_NnapiSerializer.has_operand_for_jitvalcCs|j|}||j|fSr)rr)rrrrrrget_tensor_operand_by_jitvals z-_NnapiSerializer.get_tensor_operand_by_jitvalcCsF||\}}|jD](}|dkr(td|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|}|dkr0||d\}}|||}||j|fSN TensorType)rgetget_constant_valuerr)rrrrrrrrrget_tensor_operand_or_constants   z/_NnapiSerializer.get_tensor_operand_or_constantcCs(||d\}}||}||j|fSr.)r1rr)rrrrrrrrget_tensor_operand_for_weights z._NnapiSerializer.get_tensor_operand_for_weightcCs.|j|t|t|f|j||dSr)rrrrextend)ropcoderrrrr add_operationsz_NnapiSerializer.add_operationcCs||jkst||j|<dSr)rr)rrrrrradd_tensor_sequencesz$_NnapiSerializer.add_tensor_sequencecCs ||jkst||f|j|<dSr)rrrrctyperrrradd_constant_value sz#_NnapiSerializer.add_constant_valueNcCsd|j|}|dkr$td|d|\}}|dk r`||kr`td|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)rr0rkind)rrZtypekindrecordr9rrrrr1 s z#_NnapiSerializer.get_constant_valuecCsH|dkr|j}nt|t|jks&tdg}t|D]`\}}|dkrT|t|n6|dkrn|t||n|dkr|dntd|dq4|dd |}|j t j krd |d S|j t j krd |d S|j t j krd |jd|jd|dS|j t jt jfkr4|jr,d |dStdtd|j dS)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: )rrrrrrrrjoinrrrrrrrrrr)rrrrZ shape_partsrr,Z shape_coderrroperand_to_template_torchscripts@         z0_NnapiSerializer.operand_to_template_torchscriptcCs|||t||dSr)rr)rZ out_op_idZout_dimZin_op_idZin_dimrrrforward_operand_shapeGsz&_NnapiSerializer.forward_operand_shapecCs |jt||d|dS)Nz = )rrr)rrrexprrrrrJsz&_NnapiSerializer.compute_operand_shapecCs|jdddkrtd|jtjd}dgd}||d<|ddddg|d<dgd}|||d<|tj |||d|fS)Nr)rrz1Automatic transpose only supported for H,W == 1,1)rrr r) rr_replacerrr(rr6r r)rin_idrout_operrrrrrtranspose_to_nhwcOs  z"_NnapiSerializer.transpose_to_nhwccCs|j|jkr||||fS|j|jf}|tjtjfkrH|||||fS|tjtjfkrl||f|||Std|jd|jdS)Nz2Automatic transpose not supported for dim_orders: z, )rrrrrIr)rin0_idin0_operin1_idin1_operZordersrrrtranspose_for_broadcastas   z(_NnapiSerializer.transpose_for_broadcastcCsL||\}}|dkr2|dks.t|Std|d|ddS)NListTypeIntTypezCan't handle size arg of type 'z' for 'r)r1r;getElementTyperrr8rrr get_size_argps  z_NnapiSerializer.get_size_argc Csdd|D}|ddkst|d|dg}|d|dg}|d|d g}|d |d g}|d }t|d ksvt|ddgkst||||||S)NcSsg|] }|qSr)item.0rrrr xszD_NnapiSerializer.get_conv_pool_args_2d_from_pack..rrrr r r r r rrr)rrget_conv_pool_args_2d_common) r kernel_size packed_configZpcstridespaddings dilationsZoutput_padding group_numrrrget_conv_pool_args_2d_from_packws z0_NnapiSerializer.get_conv_pool_args_2d_from_packc Cs`||}||}|dkr&ddg}n ||}|dk rJ||d\} } nd} |||||| S)NrrP)rRr1rW) rrXstridepaddingdilationrrZr[r\rr]rrrget_conv_pool_args_2d_from_jits    z/_NnapiSerializer.get_conv_pool_args_2d_from_jitc Csvt|}t|dkstt|dks(tt|dks8tt|dksHt|\}}||||g} t||| ||gSNr)rrrr) rrXrZr[r\r]ZkernelsphpwZ real_paddingsrrrrWs z-_NnapiSerializer.get_conv_pool_args_2d_commonc Cs|d|dg}g}t|j}||||ttt|jdd|D].\}\}} | ||| } | |j | j j q\t|jD] \} } td| | || q|j} | dkst| dkst| d}dg}|dkr|g}d}n8|d kr:|j|}t|}ntd ||dk rlt|t|ksltt|D]d\}}|j|} |j | | |j | j j |r||nd}| || |j | |d qt| d g}d}td |t|j t|j t|j!t|jt|j}| ||"\}}|#dd|j D|#||#dd|j!Dd$|g}t|d}|ddkstt%|d}t|j D]\} \}}}}}t&||}t|D]D\}}|dkr t'||}|j( d|dt)| ||d7}qt*dd|D}| |+|q|#|| |+|j,| |+|j| |+|j|j(#|t--dd$||j.|||j(|fS)NFTrzProcessing node #%d: %rrzreturn [r/r>Z TupleTypezUnsupported return type: r@r Ziiiiiicss.|]&\}}}}}td|t|||VqdS)ZiifiN)rrr)rUtr_mr,zrrr s z3_NnapiSerializer.serialize_model..css|]}tjd|VqdS)rN)r)rr)rUxrrrrisr z ser_model[z] = css|]}|dkr|ndVqdS)r>rNrrUrrrrrisr)/r$nextgraphrr:typerrrrrrrrZnodesrdebugadd_nodeZ return_node inputsSizer outputsSizeinputsAtr;rrrrrrCrrrrserialize_valuesr4rBrrrrrrserialize_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|jks tt|j|jD]T\\}}}t|}|ddBd}|d||}|td|||||q.||fS)Nrr r#r)rrrrrrrr) rr~rZop_indexZ source_typedataZ source_lengthZphysical_lengthZ padded_datarrrrus z!_NnapiSerializer.serialize_valuescCstd|Sr%)r&r')intsrrrrv*sz_NnapiSerializer.serialize_intscCs ||Sr) add_getattrrrzrrr/rkz_NnapiSerializer.cCs ||Sr)add_constant_noderrrrr0rkcCs ||Sr)add_list_constructrrrrr1rkcCs ||Sr)add_tuple_constructrrrrr2rkcCs ||Sr) add_unsqueezerrrrr3rkcCs ||Sr)add_torrrrr4rkcCs ||Sr _identityrrrrr5rkcCs ||Sr) add_reshaperrrrr6rkcCs ||Sr) add_flattenrrrrr7rkcCs ||Sr) add_slicerrrrr8rkcCs ||Sr)add_sizerrrrr9rkcCs ||Sr)add_catrrrrr:rkcCs ||Sr)add_meanrrrrr;rkcCs ||Sr) add_quantizerrrrr<rkcCs ||Sr)add_dequantizerrrrr=rkcCs||tjtjSr)add_add_sub_opr rsrrrrrrr>scCs||tjtjSr)rr rrrrrrrrAscCs||tjtjSr)(add_pointwise_simple_binary_broadcast_opr r|rrrrrrrDscCs||tjtjSr)rr rrrrrrrrGscCs||tjSr)add_pointwise_simple_unary_opr r}rrrrrJscCs||tjSr)rr rzrrrrrMscCs ||Sr) add_softmaxrrrrrPrkcCs ||Sr) add_hardtanhrrrrrQrkcCs ||Sr)add_avg_pool2drrrrrRrkcCs||tjSr)add_pool2d_noder r{rrrrrSscCs ||Sr)add_adaptive_avg_pool2drrrrrVscCs ||Sr)add_upsample_nearest2drrrrrYscCs ||Sr) add_prelu_oprrrrr\rkcCs ||Sr) add_addmmrrrrr]rkcCs ||Sr) add_linearrrrrr^rkcCs ||Sr)add_conv_underscorerrrrr_rkcCs ||Sr) add_conv2drrrrr`rkcCs ||Sr)add_log_softmaxrrrrrarkcCs ||Sr) add_qlinearrrrrrbrkcCs||tjSr add_qconv2drrrrrrrcscCs||tjSr)rrrrrrrrfscCs|j|tjddS)NT)rrrrrrriscCs||tjtjSr)add_qaddr rsrrrrrrrlscCs||tjtjSr)rr rsrrrrrrroscCs||tjtjSr)rr r|rrrrrrrrs)(z prim::GetAttrzprim::Constantzprim::ListConstructzprim::TupleConstructzaten::unsqueezezaten::toz aten::detachz aten::reshapez aten::flattenz aten::slicez aten::sizez aten::catz aten::meanzaten::quantize_per_tensorzaten::dequantizez aten::addz aten::subz aten::mulz aten::divz aten::reluz aten::sigmoidz aten::softmaxzaten::hardtanhzaten::avg_pool2dzaten::max_pool2dzaten::adaptive_avg_pool2dzaten::upsample_nearest2dz aten::preluz aten::addmmz aten::linearzaten::_convolutionz aten::conv2dzaten::log_softmaxzquantized::linearzquantized::conv2dzquantized::conv2d_reluzquantized::conv_transpose2dzquantized::addzquantized::add_reluzquantized::mulcCs:|j|}|s,td|d||||dS)NzUnsupported node kind (z ) in node ) ADDER_MAPr0r;r)rrzZadderrrrrqwsz_NnapiSerializer.add_nodecCs,||d\}}|d}||j|<dSr)r*rt outputsAtr)rrzrGin_operrrrrr}s z_NnapiSerializer._identitycCs~|dkst|dks t||d\}}t|dsFt|d}t||}| d}| }| |||dS)Nrrz __torch__.name) rrrrsr1rtr startswithr,r rror:)rrzZ obj_ctypeobjrroutputr9rrrrs   z_NnapiSerializer.add_getattrcCsL|dkst|dks t|d}|}|}||||dS)Nrr)rrrrsrroZtoIValuer:)rrzrr9rrrrrs  z"_NnapiSerializer.add_constant_nodec Cs|dkst|d}|}g}g}|D]\}|dk rb||jkrb||\}}||nd}|dk r|dkr||q2d}q2|dk r| ||||dk r| |||dkr|dkrt d|dS)Nrrr/zMUnable to handle ListConstruct node. Neither all constants nor all tensors. ) rsrrrorrr1rr;r:r7r) rrzrr9Z const_valstensorsinprvalrrrrs*     z#_NnapiSerializer.add_list_constructcCsF|dkst|d}g}|D]}||q&|||dS)Nrr)rsrrrrr7)rrzrrrrrrrs    z$_NnapiSerializer.add_tuple_constructc Cs|dkst|dks t||d\}}||dd\}}|jtjksZt|dkrf|n|t |j d}t |j }| |dt |}|j|d} dgd} || d<||| d<dgd} ||d| | d<|tj| | dS)NrrrrPr)rrrrsr-rtr1rrrrrrinsertrrFrrrr6r r) rrzrGrrrZreal_dimZout_shape_listrrHrrrrrrs      z_NnapiSerializer.add_unsqueezecCs||dSrrrrrrrsz_NnapiSerializer.add_toc Cs|dkst|dks t||d\}}||d\}}|dksXt|dksltt|dko|ddk}|j t j kr|st dt d|j|j}|j|t j d}dgd} || d<||| d<dgd} ||d|| d<|tj| | dS) NrrrrOrPr>zSCurrently, reshape is only supported on NHWC tensors if the target size is [X, -1].r)rrrrsr-rtr1r;rQrrrrrtorchzerosexpandrZreshaperFr(rrr6r r) rrzrGrZ shape_ctyperZis_trivial_reshaperrHrrrrrrs,  z_NnapiSerializer.add_reshapec Cs.|dkst|dks t||d\}}||dd\}}||dd\}}t|jdko|jddkp|jddko|jddk}|jt j kr|st d|dkr|t|j7}|dkr|t|j7}|jd|t dd |j||df|j|dd} td d |j||dDrFt d |jd||j|dd} | ddkr~t d |j| t j d} ||d| } t| D],\} }|dkr|| | ||jdqtdd | D}dgd}||d<|||d<dgd}| |d<|tj||dS)Nr rrrPrr zGCurrently, flatten is not supported on NHWC tensors unless C=1 or H=W=1cSs||Srr)rjyrrrrrkz._NnapiSerializer.add_flatten..css|]}|dkVqdS)rNrrUrrrrrisz/_NnapiSerializer.add_flatten..z-Flattening flexible dims is not supported yetzOnly 1 dim can be flexiblercss|]}|dkr|ndVqdS)rr>Nrrrrrri$s)rrrrsr*rtr1rrrrrr functoolsreduceanycountrFrrrrDrrr(r6r r)rrzrGrZ start_ctypeZ start_dimZ end_ctypeZend_dimZis_trivial_flattenrZnon_flattened_dimsrHout_idryrZinputs_1rrrrrrs\(  "    z_NnapiSerializer.add_flattenc s|dkst|dks t||d\}}||d\}||d\}||d\}||d\}dkrddkrtjdkr|j7ntjkrddkrtjkr| |dS|jdkrt ddkr|j7ntjkr*|jkrO}qdgd } || d<|fd dtt|jD| d<|fddt |jD| d<|fddtt|jD| d<|d| d<||| d<|d| d<dgd} || d<|tj| | dS)Nr rrrr r z#Unable to slice with flexible shapez0Slice start value should be less than stop valuec3s"|]\}}|krn|VqdSrrrUrr) dim_valueout_lenrrriRsz-_NnapiSerializer.add_slice..rr csg|]}|krndqS)rrrT)r start_valuerrrVcsz._NnapiSerializer.add_slice..cs g|]\}}|krn|qSrrr)r stop_valuerrrVfscsg|]}|krndqS)rrrT)r step_valuerrrVlsr )rrrrsr*rtr1sysmaxsizerrrrrrrrFrDr(rangerrr6r r) rrzrGrrrrZend_maskryrrrr)rrrrrrr.st            z_NnapiSerializer.add_slicecCsr|dkst|dks t||d\}}|j|d\}}|j|}|d}||| |dS)Nrrr) rrrrsr-rtrrrr:ro)rrzrrrresrrrrrws  z_NnapiSerializer.add_sizecs|dkst|dks t|j|d}||dd\}t|dksVtg}d}d}|D]}||\}} |dkrt| j d} | j | d}| j |j kst| j |j kstt| j dt|j dkst| ||| j 7}qf|dk st|j t|j |d}| j tjkrLt|j dks:tddddg} n} ||d|} t|j D]Z\} }|dkrl| krd fd d |D}|| | |n|| | |d| ql||| g}dgd}| |d<|tj||dS) NrrrrPr>rr r  + c3s|]}t|VqdSr)r)rUZip_idrrrrisz+_NnapiSerializer.add_cat..)rrrrsrrtr1rr*rrrFrrrrrrrrrBrrDrr6r ru)rrzrrZin_idsrHZ out_dim_sizerrGrr nnapi_dimrryrrrrrrrrsR       z_NnapiSerializer.add_catcCs|dkst|dks t||d\}}||d\}}|dksXt|dkslt||dd\}}||dd |jt j krt |j dkstd d |D}n|}t } |D]$} | dkr| t |j 7} | | q|jt j kr&|s&| ddhstt j} n|j} g} t|j D]0\} }| | krX| |n|r:| dq:|j| | d }dgd}||d<|||d<|||d<dgd}||d||d<|tj||dS) Nr rrrOrPrZBoolTyper NoneTypecSsg|]}ddddg|qS)rr rrrrlrrrrVsz-_NnapiSerializer.add_mean..r)rrrrsr-rtr1r;rQrrrrrsetadd issupersetrrrrFr(rrrr6r r)rrzrGrZ dim_ctyperrZkeep_dimrZcollapsed_dimsrZ out_dim_orderrrr,rHrrrrrrsF      z_NnapiSerializer.add_meanc Cs|dkst|dks t||d\}}|jtjkrHtd| |dd\}}| |dd\}}| |dd\}}|t j j krtd t j}|j|||d } dgd} || d<dgd} ||d| | d<|tj| | dS) Nr rrzqMost hardware backends prefer NHWC quantized tensors. Try setting `t.nnapi_nhwc = True` on your tensor inputs. FloatTyperrPr zKPyTorch NNAPI export only supports quantized tensors with the quint8 dtype.rrr)rrrrsr-rtrrrrr1rrrrrrFrrr6r r) rrzrGrrrrZ scalar_typerrHrrrrrrs2    z_NnapiSerializer.add_quantizecCs|dkst|dks t||d\}}|jtjddd}dgd}||d<dgd}|| d||d<| t j ||dS)Nrrrr) rrrrsr-rtrFrrrrr6r rx)rrzrGrrHrrrrrr s  z_NnapiSerializer.add_dequantizec Cs|dkst|dks t||d\}}|}|tjkr\|jtj kr\|j ddd}| | d|}t |jD] \}}|dkrx|||||qxdgd} || d<dgd} || d<||| | dS)Nrrgp?)rr)rrrrsr*rtr rzrrrrFrrrrrDr6) rrzr5rGrrHrryrrrrrrrs     z._NnapiSerializer.add_pointwise_simple_unary_opqparamsc CsF|dkst|ddks*t|ddksDt||dr||d\}}||d|j\}}nN||dr||d\}}||d|j\}}nt d|d|j |j kst| ||||\}}}}t |j |j } |j| d} |dk r2|\} } | j| | d} ||d| } tt|j |j D]\}\}}|dkr|dkr|| |||nn|dkr|dkr|| |||nH|dkrV|dkrV|jd t||d t|||| |||qVdgd }||d<||d<|||d <dgd}| |d<||||dS) zEHelper for pointwise binary broadcast ops with superfluous extra argsrrr/zCan't do a NNAPI binary op: z on two constantsrNrrzassert z == r r)rsrrtror;r)r*r2rrrrNrrrFrrrrrDrrrrr6)rrzr5 fuse_coderrJrKrLrMrrHrZzprryZd0rrrrrr_do_add_binary8s\       z_NnapiSerializer._do_add_binarycCs"|dkst||||dSrc)rrrr)rrzr5rrrrrmsz9_NnapiSerializer.add_pointwise_simple_binary_broadcast_opcCsH|dkst||dd\}}|dkr6td||||dS)Nr rrPrz*NNAPI does not support add/sub with alpha.)rrrr1rtrr)rrzr5rralpharrrrqs z_NnapiSerializer.add_add_sub_opcCsV|dkst||dd\}}||dd\}}|j|||||fddS)Nr rrr rPr)rrrr1rtr)rrzr5rrrrrrrrzsz_NnapiSerializer.add_qaddc Cs|dkst||d\}}||dd\}}||d|}t|jD] \}}|dkrV| ||||qVdgd} || d<| d| d<| || d<dgd} || d<| t j| | dS)Nr rrrPg?r)rrrr*rtr1rrrrrDr!rr6r r) rrzrGrrZ softmax_dimrrrrrrrrrs   z_NnapiSerializer.add_softmaxc Cs|dkst|dks t||d\}}||dd\}}||dd\}}tjtjd}| ||f}|dkrt ddgd} || d<dgd} | | d|| d<| || | dS)Nr rrrr))r>r)rr z9NNAPI only supports hardtanh with args (-1, 1) or (0, 6).)rrrrsr-rtr1r r~rr0rrrr6) rrzrGrrZmin_valZmax_valZop_mapr5rrrrrrs   z_NnapiSerializer.add_hardtanhc CsP|dkst|dks t|ddks:t|ddksTt||d\}}||d\}}t|j dkst|j ddkst|j ddkr| rt d| | d|}t|j D]4\}}|dkrq|dkrt dq|||||qdgd} || d<|| d<dgd} || d<|tj| | dS)Nrrrr/z8Per-channel PReLU only supports channels_last right now.z.PReLU requires fixed size for dim 0 and dim 1.)rrrrsrtror;r*r3rrrrrrrrDr6r r) rrzrGrZw_idZw_operrrrrrrrrrs4   z_NnapiSerializer.add_prelu_opcCs|dkst|dks t|\}}}}}}|p:|}||||||} | jdksf| jdkrntd| |\} } t | j dkstt | j | | j dd} | } dgd}| |d<|| j|d<|| j|d<|| j|d <|| j|d<|| j|d <|| j|d<|| j|d <|| j|d <|tj|d <|| |d<dgd}||d| j| d|d<||||dS)Nr rz'NNAPI does not support dilated pooling.r Frrrr r r rrrr)rrrrsrrbrRrrrr-rrrrrrrrrrrrrrrr$rrrFr6)rrzr5imagekernelr_r`ra ceil_moderimage_id image_operrrrrrrrrsN   z _NnapiSerializer.add_pool2d_nodecCs|dkst|dks t|\}}}}}}}||\} } ||\} } | rZ| rbtd|||||} ||\} }t |j dkstt |j | |j dd}| }dgd}| |d<| | j|d<| | j|d<| | j|d <| | j|d<| | j|d <| | j|d <| | j|d<| | j|d <| tj|d <|||d<dgd}||d|j|d}|||| d||d<|tj||dS)Nr rzANNAPI doesn't support count_include_pad=False or divisor_overrider Frrrr r r rrrr) rrrrsrr1rrbrRr*rrrrrrrrrrrrrrrr$rrrF _handle_conv_pool_flexible_inputr6r rt)rrzrrr_r`rZcount_include_padZdivisor_overriderZcount_include_pad_valueZdivisor_override_valuerrrrrrrrrrrrsf    z_NnapiSerializer.add_avg_pool2dc Cs|dkst|dks t||d\}}t|jdksFt||d\}}|dksjt| dks~t|ddgkrt d|jddt |}| }dgd}||d<| d|d<| d|d<| d|d <| d|d<| d|d <| d|d <| |jd |d <| |jd|d <| tj|d<|||d<dgd} ||d|j|d| d<|tj|| dS)Nrrrr rOrPz@NNAPI only supports adaptive_avg_pool2d with output size (1, 1).rr r r r rrrr)rrrrsr-rtrrr1r;rQrrrrrrr$rrrFr6r rt) rrzrr size_ctypesize_argrrrrrrrr7sB    z(_NnapiSerializer.add_adaptive_avg_pool2dc Cs$|dks|dkst|dks,t|dkrH|\}}}n|\}}}}||\}}|dkr||\} } nD||\} } ||\} }| dkst| dkst| } | } ||\}}t|jdkst|dkr| dkrt dn|dkr|dks.t| dksDt| dksVt| dksdtt |t stt|s~tt dd |Dstt|dkr|d }t|d kst|d }|d}||}||}n| dkr| dkst| d kst|dks.t|dks.rrrcss|]}t|tVqdSr)rrrrrrrisz#Size and scale cannot both be None.rz(Flexible batch or channels not supported)rr zint(z * rA)rrrrsrr1r;r*rrrrQrrallrrr!rrrrFrrr$r6r r)rrzrZsize_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|dkst|dks t|\}}}}}||fD]2}||\}} |dks\t| dkr:tdq:||d|||dS)Nr r)rPrz6NNAPI Fully-Connected does not support alpha and beta.T)rrrrsrr1r;radd_addmm_or_linear) rrzjit_bias jit_input jit_weightZjit_betaZ jit_alpharrZ scale_valuerrrrs z_NnapiSerializer.add_addmmcCsD|dkst|dks t|\}}}||d|||dS)Nr rF)rrrrsrr)rrzrrrrrrrsz_NnapiSerializer.add_linearcCs4||\}}||\}} t|jdks.tt| jdks@t||d\} } t| jdksbt|rt| } n| } || } |j | }|jd|jdf}| | d|j |d}|jddkr| |d|ddgd}||d<| |d<||d<|tj|d<dgd}||d<|tj||dS)Nrrr/rrr r )r*r3rrrr1rf contiguousrrrrrFrDrrrr6r ry)rrzZtranspose_weightrrrinput_id input_operbias_id bias_operr weight_tensornnapi_weight_tensor weight_id weight_operrrrrrrrrs4     z$_NnapiSerializer.add_addmm_or_linearcCsj|dkst|dks t|\}}}}||\}}t|jdksPt||d\}} ||d\}} ||\} } | dkst| d\} }|dk stt| jdkstt|jdkst|jd| jdkst| jd|jdkst| t j kst| j t jkr&| }nB| j t jks8tt j| dt j| | dd }|}|j|}t ||dt j}||}|j|| }|dkst|dkrtd |}||}|j|}|jd|jdf}|j|| | d }dgd}||d<||d<||d<|t j!|d <dgd}|"|#d||d<|$t%j&||dS) Nr rrrrPZLinearPackedParamsBaserrQuantized 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 )'rrrrsrr-rrr1r __getstate__qschemerper_tensor_affinerrqint8!_make_per_tensor_quantized_tensorint_reprrtouint8rr rquantize_per_tensorrrrrrrFrrrrrr6r ry)rrzrjit_packed_weight jit_scalejit_zero_pointrrr out_scaleout_zero_point weight_ctype packed_weight raw_weightraw_biasunsigned_weight weight_scale bias_scaleint_biasr multiplierrrrrrHrrrrrrsp         z_NnapiSerializer.add_qlinearc Csh||\}}|dkrZ|r"dnd}tj|||jd}||}|j|} || fS||SdS)Nrrr)r) r1r;rrrrrrr3) rrrrr9rZbias_idxZnnapi_bias_tensorrrrrrget_optional_biasDs     z"_NnapiSerializer.get_optional_biasc Cs|dkst|dks t|\}}}}}}}||d\} } ||| \} } || jdd||||} || ddd|| | | dt j S) Nr rr/rr rrF rrrrsrr1r rbradd_conv2d_commonrrr)rrz jit_imagerr jit_stridejit_pad jit_dilation jit_groupsrrrrrrrrrQs>  z_NnapiSerializer.add_conv2dc Cs|dkst|dks t|\ }}}}}}}} } } } } } ||d\} } ||\} } ||| | \} }|| jdd|||| }|| ddd|| | || t j S)Nr!rr/rr rrr )rrzr rrr rrZ jit_transposerrrrrrrrrrrqsL z$_NnapiSerializer.add_conv_underscorec Cs|dkst|dks t|\}}}||\}}||d\}}|j} dgd} || d<|d| d<||| d<dgd} | | d|j | d| d<| t j| | dS)Nr rrPrrr)rrrrsrr-r1rr!rrrrFr6r r) rrzrZjit_dimZjit_half_to_floatrrrrrrrrrrrs    z _NnapiSerializer.add_log_softmaxc Cs|dkst|dks t|\}}}}||d\}} ||d\}} ||\} } | dksnt| d\} }}| dkst|\}}|\}|dk st||jdd|}| t j kst|j t j kr|}n@|j t jkstt j|d t j||d d }|}||\}}|j|}t ||dt j}||}|j|| }|dksxt|dkrtd ||d| | |||||| S) Nr rrrPZConv2dPackedParamsBaser2rrrr)rrrrsrr1rrr^rrrrrrrrrrrrrr r*rrrrrr r)rrzrrr rrrrrrrrZ pack_versionrZ opt_tensorsrYrrrrrrrrrrrrrrsp         z_NnapiSerializer.add_qconv2dc !Cs<||\} } | jd} |jdkr6d} |r0d}qRd}n|j| krJd} d}ntd|j|}||}|j|}|j|}| jt j kr|jt j kst |jt j kst nh| jt j kr|jt j kst |jt j kst t| j|j|jst |jdkst ntd| jt| jd ks(t t|jd ks.. 0 0z1_NnapiSerializer._handle_conv_pool_flexible_input)F)N)N)NN)N)F)F)Prrrrrrrr rrrrrrr!r$r(r)r*r-rr2r3r6r7r:r1rCrDrrIrNrRr^rbrWrru staticmethodrvrrqrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrr rrrrr rrrrrr>s  /     /  n I :I 51"5 $+6(`&I  ' AmrF)rrxrcCst|||||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)modulerrrxrrrrrs  r)r)!r&enumrloggingrrtypingrrrrr getLoggerrrr rrEnumrrrrrrrrrrrrrrrrrrsV b    R