=Vh9[dZddlmZddlmZddlmZddlmZddlm Z ddlm Z ddl m Z dd l m Z dd l mZdd l mZdd l mZdd l mZddl mZddl mZddlmZeZeZeZej8ZdZeej<Zeej>Zeej@Z eejBZ!eejDZ"eejFZ#eejHZ$eejJZ%eejLZ&eejNZ'eejPZ(eejRZ)eejTZ*eejVZ+eejXZ,eejZZ-eej\Z.eej^Z/eej`Z0eejbZ1eejdZ2eejfZ3eejhZ4eejjZ5eejlZ6eejnZ7eejpZ8dZ9e jte jve jxe jze j|e j~e je jiZBe jve jte jze jxe j~e j|e je jiZCd8dZDd8dZEe9ejZFe9ejZGe9ejZHe9ejZIe9ejZKe9ejZe9ejZe9ejZNe9ejZOe9ejZPe9ejZQe9ejZRe9ejZTe9ejZVe9ejZXe9ejZZe9ejZ\e9ejZ^e9ejZ_e9e jZae9eDZbe9eEZce9ejZde9ejZee9ejZfe9ejZge9ejZhe9ejZidZjeje jZkeje jZme jZod8dZpd8dZqe jZs d9dZtejZvd8dZwejZxejZy d:dZzdZ{d8dZ|ejZ~ejZejZejZd8dZd8d Zd!ZejZejZejZe jd" d;d#Zej Zej$Zej(Zej,Zdd4Z d>d5Zd6Zd7Zy)?a-Experimental library that exposes XLA operations directly in TensorFlow. It is sometimes useful to be able to build HLO programs directly from TensorFlow. This file provides Tensorflow operators that mirror the semantics of HLO operators as closely as possible. Note: Most of the operators defined in this module are used by the jax2tf converter (see go/jax2tf for details) and are used in SavedModel produced by jax2tf. Hence, we need to maintain backwards compatibility for these operators. Please reach out to the JAX team if you want to make changes. ) gen_xla_ops) xla_data_pb2)attr_value_pb2) constant_op)dtypes)ops) array_ops) bitwise_ops) gen_math_ops)gen_random_ops)math_ops) random_ops)random_ops_util)special_math_ops)np_utilscdfd }|S):Wrapper that restricts `fn` to have the correct signature.c||SNname)xrfns U/home/dcms/DCMS/lib/python3.12/site-packages/tensorflow/compiler/tf2xla/python/xla.pyunary_op_wrapperz#_unary_op..unary_op_wrapperIs ad Nr)rrs` r _unary_oprFs rcdfd }|S)zGWraps a binary Tensorflow operator and performs XLA-style broadcasting.c|xsg}tj|tj}t j |||\}}|||S)zInner wrapper function.r)rconvert_to_tensorrint64rxla_broadcast_helper)rybroadcast_dimsrrs rbroadcasting_binary_op_wrapperz?_broadcasting_binary_op..broadcasting_binary_op_wrapper{sL#)rN**>6<<HN  + +Aq. ADAq a rNNr)rr's` r_broadcasting_binary_opr)xs ('rNc2|j|jk(sJ|j}|tv}|r5t|}tj||}tj||}t j |||}|rtj||}|S)zCPerforms an integer right logical shift irrespective of input type.r)dtype_SIGNED_TO_UNSIGNED_TABLEr castr right_shift)rr%rr+signedunsigned_dtypeoutputs r_shift_right_logical_helperr2s AGG   ''% - -& .u5N a(A a(A  " "1ad 3& ]]65 )F -rc2|j|jk(sJ|j}|tv}|r5t|}tj||}tj||}t j |||}|rtj||}|S)zFPerforms an integer right arithmetic shift irrespective of input type.r)r+_UNSIGNED_TO_SIGNED_TABLEr r-r r.)rr%rr+unsigned signed_dtyper1s r_shift_right_arithmetic_helperr7s AGG   ''% / /( ,U3L a&A a&A  " "1ad 3& ]]65 )F -rcdfd }|S)rc|||Srr)rr%rrs rbinary_op_wrapperz%_binary_op..binary_op_wrappers a rrr)rr:s` r _binary_opr;s rctj|}tjt j |tj |gd}tj|||S)Nr)axisr)rr"r concatrconstantshape broadcast_to)rdimsrr@s r broadcastrCsW A!   D!9??1#56Q %   5t 44rc6tt|||||Sr)minmax)arbrs rclamprIs SAD !14 00rFc d} |r|j} | xs |j|jk7xs| dkD}| *tj|j|j} |s| r/t j |||||||| |j| | |  St j ||||||||j| |  S)aWraps the XLA ConvGeneralDilated operator. ConvGeneralDilated is the most general form of XLA convolution and is documented at https://www.tensorflow.org/performance/xla/operation_semantics#conv_convolution Args: lhs: the input tensor rhs: the kernel tensor window_strides: the inter-window strides padding: the padding to apply at the start and end of each input dimensions lhs_dilation: dilation to apply between input elements rhs_dilation: dilation to apply between kernel elements dimension_numbers: a `ConvolutionDimensionNumbers` proto. feature_group_count: number of feature groups for grouped convolution. precision_config: a `xla.PrecisionConfig` proto. preferred_element_type: the result `dtype`. name: an optional name for the operator. use_v2: an optional request to use the XlaConvV2 op even if not necessary. batch_group_count: number of batch groups or grouped filters. Returns: A tensor representing the output of the convolution. ) window_stridespadding lhs_dilation rhs_dilationfeature_group_countbatch_group_countdimension_numbersprecision_configpreferred_element_typer)rMrNrOrPrQrSrTr)SerializeToStringr+r result_typer xla_conv_v2xla_conv)lhsrhsrMrNrOrPrSrQrTrUruse_v2rRprecision_config_protoneeds_v2s rconvr_sN-??A ))syy  Q  #%11#))SYYG   " "  %!!/++==?/5        #-);;=-   rc4tj||d|S)NrL)axesr)r tensordot)rZr[rs rdotrcEs   C14 88rcfd}|r|j}|xs|j|jk7}|*tj|j|j}|s|r)t j |||j|||St j |||j||S)NrK)rSrTrUr)rSrTr)rVr+rrWr xla_dot_v2xla_dot) rZr[rSrTrUrr\r]r^s r dot_generalrgMs-??A # ? SYY(>(#%11#))SYYG   ! !  +==?/5        );;=-   rc2tj||||Sr)rxla_self_adjoint_eig)rGlowermax_iterepsilons rself_adjoint_eigrmns  ) )!UHg FFrcZd}|r|j}tj||||S)NrK)rVrxla_svd)rGrkrlrTr]s rsvdrprs2-??A   Q'3I JJrcttj|}tj||||j|S)N)meanstddevr+r)rr"r random_normalr+)musigmarBrs rrtrts4 R "  ! ! E rcttj|}tj||||j|S)N)r+r)rr"rrandom_uniformr+)minvalmaxvalrBrs rrxrxs4   (&  " " FF&,,T rc^tj|}tj||||S)aDStateless PRNG bit generator. Wraps the XLA RngBitGenerator operator, documented at https://www.tensorflow.org/performance/xla/operation_semantics#rngbitgenerator. Args: algorithm: The PRNG algorithm to use, one of tf.random.Algorithm.{PHILOX, THREEFRY, AUTO_SELECT}. initial_state: Initial state for the PRNG algorithm. For THREEFRY, it should be a u64[2] and for PHILOX a u64[3]. shape: The output shape of the generated data. dtype: The type of the tensor. Returns: a tuple with a new state and generated data of the given shape. r+)rconvert_alg_to_intrxla_rng_bit_generator) algorithm initial_stater@r+alg_ints rrng_bit_generatorrs0"  . .y 9'  * * }e5 rXlaVariadicReducec |xsdgt|z}|xsdgt|z}|xsdgt|z}|xsdgt|z}tj||||||||| S)aNWraps the XLA ReduceWindow operator. ReduceWindow is documented at https://www.tensorflow.org/performance/xla/operation_semantics#reducewindow . Args: operand: the input tensor init: a scalar tensor representing the initial value for the reduction reducer: a reduction function that combines a pair of scalars. window_dimensions: shape of the window, as a list of integers window_strides: inter-window strides, as a list of integers. Optional; if omitted, defaults to strides of 1. padding: padding to apply to 'operand'. List of (low, high) pairs of integers that specify the padding to apply before and after each dimension. Optional; if omitted, defaults to no padding. name: the operator name, or None. Returns: A tensor that represents the output of the reduce_window operator. rL)rr) input init_valuewindow_dimensionsrMbase_dilationswindow_dilationsrN computationr)lenrxla_reduce_window) operandinitreducerrrMrrrNrs r reduce_windowrs>"AaS3/@+A%A.!AaS3/@+A%A.%E!s3D/E)E  8x#&7"88'  & & )##'   rcf|tj||}tj|||}|Sr)r transposereshape)r new_sizes dimensionsrs rrrs3Az*A940! (rc2tj||||Sr)r where) conditionrr%rs rselectrs Aq$ //rc t|||Dcgc]\}}}t|||}}}}|t|Scc}}}wr)zip_slicetuple)r start_dims limit_dimsstridesstartlimitstridespecs rslicer sS%( J$H   5% UE6" $  5;  s< XlaShardingc|jd}tj|||jd}|jj dt j ||gS)zGradient for XlaSharding op.shardingunspecified_dims)rr _XlaSharding)s)get_attrr xla_shardingop _set_attrr AttrValue)rgrad sharding_attr grad_shardings r_sharding_gradrse++j)-** {{#56- n..? rXlaSpmdFullToShardShapec tj||jd|jdjj |jd|jd}|gS)Nmanual_shardingrdimr)r full_shaperr)rxla_spmd_shard_to_full_shaperinputsr@as_list)rrs2fs r_spmd_full_to_shard_shape_gradr*s^00 kk"341##++- ++e {{#56  # ,rXlaSpmdShardToFullShapectj||jd|jd|jd}|gS)Nrrr)rrr)rxla_spmd_full_to_shard_shaper)rrf2ss r_spmd_shard_to_full_shape_gradr6sD00 kk"34 ++e {{#56  # ,rc tj|||dn||dn|td|Dtd|D|S)aEmits an HLO `CustomCall` operation with multiple outputs. See `CustomCall` specification at https://tensorflow.org/xla/operation_semantics#customcall, and `mhlo.custom_call` specification at https://tensorflow.org/mlir/hlo_ops#mhlocustom_call_mlirmhlocustomcallop. Args: call_target_name: Name of the user function. The function signature must conform to version 3 of the API, see `API_VERSION_STATUS_RETURNING_UNIFIED`. All operands and results assumed to be in the default layout. operands: A sequence of tensors with possibly different types. result_specs: A sequence of tensor specs for all results. backend_config: A string that encodes a metadata for the backend. Empty string by default. has_side_effect: Indicates whether the custom call has side effects. `False` by default. name: Optional name of the operation. Returns: A tuple of output tensors. rKFc34K|]}|jywrr|.0rs r z!custom_call_v2..m>$**>c34K|]}|jywr)r@rs rrz!custom_call_v2..nrr)operandscall_target_namebackend_confighas_side_effect result_dtypes result_shapesr)rxla_custom_call_v2r)rr result_specsrrrs rcustom_call_v2rIsP>  ' '')1R~.6eO>>>>>>  rr)version platforms function_listhas_token_input_outputdisabled_checksc |tj|||d|||||| } t| tjrd} | S)zSee documentation for the XlaCallModule op. https://github.com/search?q=repo%3Atensorflow%2Ftensorflow+path%3Axla_ops.cc+xlacallmodule&type=code r) rmodule dim_args_specToutSoutrrrr)rxla_call_module isinstancer Operation) argsrrrrrrrrress r call_modulerusL ##    !3% # S]]# C *rcy)zMaximum version of XlaCallModule op supported. See versioning details documentation for the XlaCallModule op at: https://github.com/search?q=repo%3Atensorflow%2Ftensorflow+path%3Axla_call_module+%22int+kVersionMaximumSupported%22&type=code rrrr%call_module_maximum_supported_versionrs rcy)Nplatformrrrr"call_module_disable_check_platformrs rcTtj||||j||S)N) slice_sizesrSindices_are_sortedr)r xla_gatherrV)r start_indicesrSrrrs rgatherrs2    );;=+   rc Vtj|||||j||S)N)update_computationrSrr)r xla_scatterrV)rscatter_indicesupdatesrrSrrs rscatterrs5    +);;=+  rc,tj|Sr)rxla_optimization_barrier)rs roptimization_barrierrs  - -d 33rc0tj|||Sr)rxla_reduce_precision)r exponent_bits mantissa_bitss rreduce_precisionrs  ) )'=- PPrr)rLNNNFrL)NNNF)NNNNNr()NNN)FN)__doc__tensorflow.compiler.tf2xla.opsrtensorflow.compiler.xlartensorflow.core.frameworkrtensorflow.python.frameworkrrrtensorflow.python.opsr r r r r rrrtensorflow.python.ops.numpy_opsrrF_maxrE_minrrr?rabsconjcosceildigammaerferfcerfinvndtriexpexpm1floorimag is_finitelgammaloglog1p logical_notnegrealroundsinsigntantanh bessel_i0e bessel_i1er)int8uint8int16uint16int32uint32r#uint64r,r4r2r7addsubmuldivmodremmaximumminimumatan2complex logical_and logical_or logical_xorequaleq not_equalne greater_equalgegreatergt less_equallelessltpow left_shift shift_leftshift_right_logicalshift_right_arithmeticigamma igamma_grad_arandom_gamma_gradigammac polygammazetar;rreverserevbitcastbitcast_convert_typerCrIr> concatenater_r-convert_element_typercDotDimensionNumbersPrecisionConfigrgrmrpxla_dynamic_slice dynamic_slicexla_dynamic_update_slicedynamic_update_slice xla_einsumeinsumxla_padpadrtrxrxla_recvrecv xla_reducereducexla_variadic_reduce_v2variadic_reduce no_gradientrxla_replica_id replica_id xla_set_bound set_boundxla_set_dynamic_dimension_sizeset_dynamic_dimension_size!xla_remove_dynamic_dimension_sizeremove_dynamic_dimension_sizerrxla_select_and_scatterselect_and_scatterxla_sendsendrrRegisterGradientrrspmd_full_to_shard_shaperspmd_shard_to_full_shaperrxla_sortsortxla_key_value_sortkey_value_sortxla_variadic_sort variadic_sort xla_while while_loopxla_dequantize dequantizexla_custom_call custom_callrrrrrrrrrrrrs 7043.++-.0*,124         H$$ %  8?? #(..! (..!(..! h(( ) 8?? # (..!,,-   (..!  '22 3 '22 3  ($ KK LL&-- LL&-- LL&--  LL&++ MM6<< MM6<< MM6<<   hll+hll+hll+hll+l../h../h..// !("2"2 3%h&:&:; $X%8%8 9 %h&:&:; X^^,X//0X334X--.X001X]]+hll+ $[%;%; < -.IJ01OP  1' (B(BC +N,L,LM !("2"2 3 #H$6$6 7 x}}- y** + ""# ((51  K\ }}9#66.. BGK-- ";;   .   44#$  -` ' '   % % )GG !, M M 0!77  # #m$ % 'CC&CC/01/01//--  " "  ' ' ))  '^  " J  0 (4Qr