jR NddlZddlZddlmZmZmZmZmZmZm Z ddl Z ddl m cm cmcmZddlm cm cmZddlm Z ddlmZmZddlmZddlmZddlmZddlmZdd lm Z m!Z!e j"jZ#Gd d ej$Z%d ed ededee&ee fdee%e%ff dZ'd ed edee&ee fdeee e j(e)fe e j(e*fffdZ+d ed edefdZ,d ed ede*fdZ-d edee*fdZ.d ed ede&fdZ/de!de&de!fdZ0de!ddfdZ1dZ2e2de j(de j(de j(fdZ3e2de j(de j(de j(fdZ4e2de j(de j(de j(fd Z5d ede6fd!Z7d ed ed"e*defd#Z8dS)$N)CallableDictListOptionalSetTupleUnion)FakeQuantizeBase ObserverBase)_is_activation_post_process)getattr_from_fqn) GraphModule)Node)NSNodeTargetType NSResultsTypeceZdZejZejZejZejZejZ dS)NodeInputOrOutputTypeN) __name__ __module__ __qualname__enumautoFP32INT8FP16UNKNOWN FP32_OR_INT8EC:\Projects\Historical\venv\Lib\site-packages\torch/ao/ns/fx/utils.pyrrsQ 49;;D 49;;D 49;;DdikkG 49;;LLLr rnodegm logger_clsnode_type_to_io_type_mapreturncn|d}|d}|d}|d}|d}|d} |d} |d} |jd kr|j|vrtjtjfS|j|vrtjtjfS|j|vrtjtjfS|j|vrAt ||d } t| tsJt| |||\} }||fStj tj fS|jd kr;|jd ksJt|jtsJt||jtfd | D}t|ttfs|rAt ||d } t| tsJt| |||\} }||fStfd |D}tfd| D}|rtjtjfS|rtjtjfStj tj fS|jdkr:|jdkrKt ||d }t|tsJt||||\} }|tjfS|jdkrwt ||d }t|tsJt||||\} }t ||d}|t jus J|d|tjfS|j| vrAt ||d } t| tsJt| |||\} }||fStj tj fStj tj fS)Nfuns_io_type_fp32funs_io_type_fp16funs_io_type_int8funs_io_type_fp32_or_int8mods_io_type_fp32mods_io_type_int8mods_io_type_fp32_or_int8meths_io_type_fp32_or_int8 call_functionr call_modulec38K|]}t|VdSN isinstance.0 target_typemods r! z7get_node_first_input_and_output_type..Ns>1 1 -8JsK ( (1 1 1 1 1 1 r c38K|]}t|VdSr3r4r6s r!r:z7get_node_first_input_and_output_type..`>) ) -8JsK ( () ) ) ) ) ) r c38K|]}t|VdSr3r4r6s r!r:z7get_node_first_input_and_output_type..cr<r call_method dequantizetorz handling needs to be added)optargetrrrrget_normalized_nth_inputr5r$get_node_first_input_and_output_typerstrr anyr r torchfloat16)r"r#r$r%FUNS_IO_TYPE_FP32FUNS_IO_TYPE_FP16FUNS_IO_TYPE_INT8FUNS_IO_TYPE_FP32_OR_INT8MODS_IO_TYPE_FP32MODS_IO_TYPE_INT8MODS_IO_TYPE_FP32_OR_INT8METHS_IO_TYPE_FP32_OR_INT8 first_arg_prev_node_input_typeprev_node_output_type"is_known_fp32_or_int8_input_moduleis_known_fp32_input_moduleis_known_int8_input_module prev_nodecur_node_dtype_targetr9s @r!rDrD&s11DE01DE01DE 89T U01DE01DE 89T U!9:V!W w/!! ;+ + +).0E0JK K ;+ + +).0E0JK K [- - -).0E0JK K [5 5 50r1==Ii.. . ..52z+C %%*+@A A)13H3PQ Q M ! !w-''''$+s+++++r4;//-01 1 1 1 ._get_scale_zp_from_function_argss-dB FF *4Z@@*d++R :;Lc0R0RRRR'4((LZ-L-LLLL$R):;; !"gn556""r r0rr1c38K|]}t|VdSr3r4)r7r8 module_objs r!r:z)get_node_input_qparams..s>1 1 4?Jz; / /1 1 1 1 1 1 r )+rCr5rrArBrGquantize_per_tensortoqaddadd_relumulmul_relurEr nnqLinearConv1dConv2dnniq ConvReLU2dConv3d BatchNorm2d BatchNorm3dConvTranspose1dConvTranspose2dELU GroupNormInstanceNorm1dInstanceNorm2dInstanceNorm3d LayerNorm Hardswish LeakyReLUReLU6BNReLU2dBNReLU3d ConvReLU1d ConvReLU3d LinearReLUscale zero_pointrFget_node_input_qparams)r"r#r%rWrOrarTres @r!rrs)r155I i & &t 89T U###|&&  u8 8 833Ir1aHH H  #'3<#,!O O O33Ir1aHH Ht  & &)*C00000%b)*:;;      ## """      1    =:$j&;< <-01 1 1 1 C\1 1 1 . . * . S))R9QRR R 4r c|jdkrt||j}t|rt |jdksJt |jdtsJ|jd}t |jtsJt||j}t|rIt |jdksJt |jdtsJ|jd}|S)a If node is not an observer, returns it. If node is an observer, navigates up the graph and returns the first parent which is not an observer. For example, graph: (node_non_obs), node = node_non_obs : returns node_non_obs graph: (node_non_obs -> obs0), node = obs0 : returns node_non_obs graph: (node_non_obs -> obs0 -> fq0), node = fq0 : returns node_non_obs r1rr) rAr rBr lenargsr5rrEr"r#node_objs r!return_first_non_observer_noders w-#B 44 &x 0 0 $ty>>Q&&&&dilD11 1 119Qz4get_arg_indices_of_inputs_to_log..;s1DDDdil););t)C)C!)C)C)Cr rb) rrrArBrGrhops quantizedoperatorrjrange)r"results` r! get_arg_indices_of_inputs_to_logr(s 49~~  w/!!  59#6#:HLIII ;59ei&9&=x|L L LDDDDU1XXDDD 3Jr cd}|jdvrtj|j}nP|jdkrEt |jt sJt ||j}tj|}|S)z Returns a string representation of the type of the function or module pointed to by this node, or '' for other node types. )r0r>r1)rArGtypenamerBr5rEr )r"r#r8 target_mods r!get_target_type_strr@su K w222nT[11 M ! !$+s+++++%b$+66 nZ00 r results model_namec i}|D]k\}}d}|D]B}|D]+\}}||kr t|sJ|dd}+,C||||<f|||<l|S)a  Rekeys the layer name of a results dictionary to use node names from `model_name`. For example, transforms {'base_op_1_0': {'node_output': {'model_a': [{'ref_node_name': 'linear1', ...}]}}} into {'linear1': {'node_output': {'model_a': [{'ref_node_name': 'linear1', ...}]}}} Note: we cannot use these node names directly because they are not guaranteed to be consistent across models. This is why we extract the results first and rekey afterwards. Nr ref_node_name)itemsvaluesr) rr new_resultsold_layer_nameresult_type_to_resultsnew_layer_namemodel_name_to_resultscur_model_namelist_of_resultss r!'rekey_logger_info_on_node_name_of_modelrOs,K29--// A A..%;%B%B%D%D   !3H3N3N3P3P  /!Z///////%4Q%7%HNN    %*@K ' '*@K ' ' r c d}|D]X}|D]A}|D]*\}}t|dkr|dd|}n+|r|D]|}|D]c}||}|D]D\}}||kr tt|D]}||d}|||d<Ed{dSdS)ay If `fqn` entries are filled in for one of the models in `results`, copies them over to any models which do not have them filled out. A common use case benefitting from this is comparing a model prepared by quantization to a quantized model. In this case, the model prepared by quantization would have `fqn` entries, and the quantized model would not. Nrfqn)rrrr) rmodel_name_with_fqnsrrr model_resultsref_model_resultsrrs r!maybe_add_missing_fqnsrvsq ").."2"2%;%B%B%D%D   !-B-H-H-J-J  ) M}%%))$Q'.:/9,   6&-nn&6&6 6 6 ")?)F)F)H)H 6 6%$9:N$O!1F1L1L1N1N66-J !%999 "3}#5#56666/25925 a(//66 6 6 6 6 6r cfdS)Ncx|^}}}t|trt|ts*t|trQt|tr.innersSB r5 ! ! %jU&;&; % r4  %%/D%9%9 %GBKK ; ;S/w//uuh9&99::::N EL ) ) %jU\.J.J % %]]__ %]]__ 8u{ " "bh%+&=&=4%W%%q(%f%%%r r)rrs`@r!8maybe_dequantize_first_two_tensor_args_and_handle_tuplesrs)&&&&&&2 Lr xyctj|}tj||z }dtj||z zS)z Computes the SQNR between `x` and `y`. Args: x: Tensor or tuple of tensors y: Tensor or tuple of tensors Return: float or tuple of floats )rGnormlog10)rrPsPns r! compute_sqnrrs= AB AE  B  BG$$ $$r ctj||z dz|dzz S)z Computes the normalized L2 error between `x` and `y`. Args: x: Tensor or tuple of tensors y: Tensor or tuple of tensors Return: float or tuple of floats rb)rGsqrtsumrrs r!compute_normalized_l2_errorrs9 :A!|((**adZZ\\9 : ::r c|dd}|dd}tjj||S)z Computes the cosine similarity between `x` and `y`. Args: x: Tensor or tuple of tensors y: Tensor or tuple of tensors Return: float or tuple of floats r)reshaperGr functionalcosine_similarityrs r!compute_cosine_similarityrsE !RA !RA 8  0 0A 6 66r c|jdkrM|jtjtjt jt jtjtjfvrdSdS)Nr0FT)rArBrGrhrjrcatstack)r"s r!op_type_supports_shadowingrsM w/!! ; I I L L I K    5 4r idxc ||d}|l|\}}t|t|z|ksJ|t|kr||St||St|jt|jz|ksJ|t|jkr |j|S|t|jz}t|j|S#t $rt|jt|jz|ksJ|t|jkr|j|cYS|t|jz}t|j|cYSwxYw)zu Given a node, gets the n'th input to that node, normalizing args and kwargs to the best of its ability. T)normalize_to_only_use_kwargs)normalized_argumentsrrrrr RuntimeError)r"r#rnorm_args_and_kwargs norm_args norm_kwargs kwargs_idxs r!rCrCs :#88 T 9   +%9 "I{y>>C $4$44s::::S^^## ~%K..0011#66ty>>C $4$44s::::S^^##y~% 3ty>>1 DK..0011*==  : : :49~~DK 0 0036666 TY  9S> ! ! !s49~~-J **,,--j9 9 9 9 :s.AD &DADADAG>AGG)9rrtypingrrrrrrr rGtorch.ao.nn.intrinsic.quantizedaor intrinsicrrptorch.ao.nn.quantizedrltorch.nntorch.ao.quantizationr r torch.ao.quantization.observerr torch.ao.quantization.utilsr torch.fxrtorch.fx.graphrns_typesrrrrgEnumrrErDrrintrrrrrrrrrrrboolrrCrr r!rsp DDDDDDDDDDDDDDDDDD ...............############@@@@@@@@FFFFFF888888 55555555 i     DI   xN xNxNxN#3,<(=#=> xN  "7 78 xNxNxNxNvM MM#3,<(=#=>MeE%,-.elC6G0HHIJ MMMM`  :        44DI0 d      $ $$$$$$N6M6d6666D:: %EL %U\ %el % % %:9 % : ;5< ;EL ;U\ ; ; ;:9 ;:77%,75<777:97& T d    !:4!:[!:s!:t!:!:!:!:!:!:r