o ^ic>@sVddlmZddlmZmZddlmZddlmZeddZed d Z ej ed d Z ej ee d dd Z ej ee ddsddZej edtddZeddZeddZeddZeddZeddZedd Zed!d"Zej eejd#d$d%d&dud(d)Zej eejd*d+d,dvd-d.Zed/d0Zed1d2Zed3d4Zed5d6Zej eejd7d$d%d&dud8d9Zej eejd:d+d,dvd;d<Zed=d>Z ed?d@Z!ej eejdAdBdCdwdxdEdAZ"edFdGZ#ej eedHdydIdJZ$edKdLZ%ej eedMdtdNdMZ&ej eej'dOdBdCdzdxdPdOZ(edQdRZ)ej ee'dSd{dTdSZ*ed|dWdXZ+ed}dZd[Z,ed~d^d_Z-edd`daZ.eddej/fddedfZ0edej/fddgdhZ1edddidjZ2edej/fddkdlZ3edmdnZ4ej eddodpZ5edqdrZ6dS)) annotations)jitconstexpr_function)core)mathcCs,d}|}|dkr|dL}|d7}|dks|S)Nrr)ilog2nr r S/var/www/html/RAG/RAG_venv/lib/python3.10/site-packages/triton/language/standard.py_log2 srcCs||d@dko |dkSNrrr )r r r r _is_power_of_twosrcCs||d|S)z Computes the ceiling division of :code:`x` by :code:`div` :param x: the input number :type x: Block :param div: the divisor :type div: Block rr )xdivr r r cdivs rsigmoidcCsddt| S)Nr)rexp)rr r r r,ssoftmaxNFcCsJ|durd}n|}|t|||d}t|}t|||d}t|||S)Nr keep_dims)maxrrsumfdiv)rdimr ieee_rounding_dimznumdenr r r r3s cCstj||jg|dS)zn Returns a contiguous flattened view of :code:`x`. :param x: the input tensor :type x: Block ) can_reorder)rreshapenumel)rr"r r r ravelAs r%c CsX|||}||}||}||}t|||}||}|||} ||} | | fS)a Transforms the indices of a row-major `size_i * size_j` matrix into the indices of a column-major matrix for each group of `size_g` rows. For example, for :code:`size_i = size_j = 4` and :code:`size_g = 2`, it will transform :: [[0 , 1 , 2 , 3 ], [4 , 5 , 6 , 7 ], [8 , 9 , 10, 11], [12, 13, 14, 15]] into :: [[0, 2, 4 , 6 ], [1, 3, 5 , 7 ], [8, 10, 12, 14], [9, 11, 13, 15]] rminimum) r jsize_isize_jsize_gijsize_gjgroup_idoff_inew_inew_jr r r swizzle2dMs  r2cCst|d|S)a' Returns a tensor filled with the scalar value 0 for the given :code:`shape` and :code:`dtype`. :param shape: Shape of the new array, e.g., (8, 16) or (8, ) :type shape: tuple of ints :param dtype: Data-type of the new array, e.g., :code:`tl.float16` :type dtype: DType r)rfull)shapedtyper r r zerosus r6cCst|j|jS)z Returns a tensor of zeros with the same shape and type as a given tensor. :param input: input tensor :type input: Tensor )r6r4r5)inputr r r zeros_likesr8c CsJ|r ||ko ||k}nd}||kp|}t|||}t|||}||fSNFrwhere) value1index1value2index2tie_break_lefttiegtv_reti_retr r r _argmax_combine rEcCt||||dSNTrEr<r=r>r?r r r _argmax_combine_tie_break_leftrKcCrGr9rIrJr r r _argmax_combine_tie_break_fastrLrMcC t||SN)rmaximumabr r r _elementwise_max rTrPreturn_indicesreturn_indices_tie_break_left)return_indices_arg tie_break_argTcCst|}|r|rtj||t|dStj||t|dSt|jjtdkrEt|jr6| tj }n|j s?Jd| tj }tj ||t|dSNr z"Expecting input to be integer type)r_promote_bfloat16_to_float32_reduce_with_indicesrKrM constexprr5primitive_bitwidth is_floatingtofloat32is_intint32reducerTr7axisrVrWrr r r rs  rz maximum indexr@)rYcCt||d||d\}}|SNT)rVrWr)rr7rgr@r_retr r r argmaxrmc CsJ|r ||ko ||k}nd}||kp|}t|||}t|||}||fSr9r:) r<r=r>r?r@rAlt value_ret index_retr r r _argmin_combinerFrrcCrGrHrrrJr r r _argmin_combine_tie_break_leftrLrtcCrGr9rsrJr r r _argmin_combine_tie_break_fastrLrucCrNrOr&rQr r r _elementwise_minrUrvr'cCst|}|r|rtj||t|dStj||t|dSt|jjdkrBt|jr3| tj }n|j srcumsumcCs8t|}t|j|}|dur||}t||t|SrO)rr\rr5raassociative_scanry)r7rgreverser5rr r r rIs   cCs||SrOr rQr r r _prod_combine[rzrcumprodcCst|}t||t|SrO)rr\rr)r7rgrr r r r`s n_dimsr(cCs:tdd}t|dg||ddgdg|}|S)Nrrr)raranger#)rr(arr r r _indicatorls *rr c Cszt|j}tj|jjdd}|j|dd}|t||d|dA}|j|jdd}t||}t ||k||Ak||} | S)NTbitwidthsignedbitcastr) rr$r get_int_dtyper5r_rarrr;) rflipr ridtypeixiyris_rightrlr r r _compare_and_swapss  rstageordercCsF|dkr tt|j|}n|}t|D] }t|||d|}q|S)zb order_type 0 == ascending order_type 1 == descending order_type 2 == alternating rr)rrr$r static_ranger)rrrrr r r r _bitonic_merge_hypercubes  rcCs6t|dgt|j}t|||}t||j}|S)Nr)rr#rr$rr4)rrrrhr r r _bitonic_merges rkr descendingc Cs.|dur t|jdn|}t|t|jdkdt|j|}|dur'|nt|}t|j}t|dg|}td|dD]} t|| | |krLdn|}qAt|d|dD]*} |rkt |t|jd|dn t |t|jd|d}t||| |krdn|}qZt||jddd|g}|S)ai Sorts a tensor along a specified dimension. :param x: The input tensor to be sorted. :type x: Tensor :param dim: The dimension along which to sort the tensor. If None, the tensor is sorted along the last dimension. Currently, only sorting along the last dimension is supported. :type dim: int, optional :param k: the number of top elements to select. If none, assume k = x.shape[dim] :type k: int, optional :param descending: If set to True, the tensor is sorted in descending order. If set to False, the tensor is sorted in ascending order. :type descending: bool, optional Nr+only minor dimension is currently supportedr)rg) lenr4rrrr$r#rrrrw) rrrrrlog_nlog_krrr r r r sort_impls 8 rcCst|||dS)N)rrr)rrrr r r sortsrcCst|||ddS)NT)rrrr)rrrr r r topkrLrcCsP|dur t|jdn|}t|t|jdkdt|jd}t||||S)Nrrr)rr4rrrr)rrrrrr r r bitonic_mergesrcCs,|dur t|d}|dkr|t|7}|Sr)r)rr4r r r _get_flip_dims   rcCstt|j |ko|t|jkt||j}tt|j|t|j|}tj|jj dd}t |j |dd|jd|dg||j|dd}t |D] }|t |||dA}qUt ||jj |jdd}|S)z Flips a tensor `x` along the dimension `dim`. :param x: the first input tensor :type x: Block :param dim: the dimension to flip along :type dim: int TrrNrr)rrrr4rrrrr5r_r#rarr)rrrstepsrrr r r r rs$ <rcCsDt||}t|jdkr|St||jddd|jdgS)a7 Interleaves the values of two tensors along their last dimension. The two tensors must have the same shape. Equivalent to `tl.join(a, b).reshape(a.shape[:-1] + [2 * a.shape[-1]])` :param a: The first input tensor. :type a: Tensor :param b: The second input tensor. :type b: Tensor rNr)rjoinrr4r#)rRrScr r r interleaves &r)NFF)F)NFTF)TF)NFN)r5rr9)rFN)rF)rrr(r)r r)rrrr)rrrrrr)rrrrrr)rrrrrO)rrrr)7 __future__r runtime.jitrrrrrr_tensor_member_fnr_add_math_1arg_docstrrrr%r2r6r8rErKrMrT_add_reduction_docstrrrmrrrtrurvrwrxryrrrrrr_add_scan_docstrrrrrrrr CONSTEXPR_0rrrrrrrr r r r s          '                       (