a kh?@sddlmZddlmZddlmZddlmZddd d Zddd d Zej ed dZ ej ee dddZ ej ee dduddZ ej edvddZeddZeddZeddZeddZedd Zed!d"Zed#d$Zej eejd%d&d'd(dwd*d+Zej eejd,d-d.dxd/d0Zed1d2Zed3d4Zed5d6Zed7d8Zej eejd9d&d'd(dyd:d;Zej eejdZed?d@ZdddAdBdCZ ej eejdDdEdFd{ddGdHdDZ!edIdJZ"ej eedKd|dLdMZ#edNdOZ$ej eedPd}dQdRZ%ej eej&dSdEdFd~ddGdTdSZ'edUdVZ(ej ee&dWddXdWZ)edddYdZd[Z*eddd\d]Z+eddd^d_d`Z,eddddadbdcZ-eddej.fddddddedfZ/edej.fdddgdhdiZ0eddddjdkdlZ1edej.fdddgdmdnZ2dodpZ3ej eddqdrZ4edsdtZ5dS)) annotations)jit)core)mathzcore.constexpr)icCs4d}t|j}|dkr*|dL}|d7}qt|S)Nrr)r constexprvalue)rlog2nr F/var/www/auris/lib/python3.9/site-packages/triton/language/standard.py_log2 s   rcCs$|j}t||d@dko |dkSNrr)r rr )rr r r r_is_power_of_twosrcCs||d|S)z Computes the ceiling division of :code:`x` by :code:`div` :param x: the input number :type x: Block :param div: the divisor :type div: Block rr )xdivr r rcdivs rsigmoidcCsddt| S)Nr)rexp)rr r rr+ssoftmaxNFcCsJ|durd}n|}|t|||d}t|}t|||d}t|||S)Nr keep_dims)maxrrsumZfdiv)rdimrZ ieee_rounding_dimznumZdenr r rr2s cCstj||jg|dS)zn Returns a contiguous flattened view of :code:`x`. :param x: the input tensor :type x: Block ) can_reorder)rreshapenumel)rr r r rravel@s r#c CsX|||}||}||}||}t|||}||}|||} ||} | | fS)a Transforms the indices of a row-major `size_i * size_j` matrix into the indices of a column-major matrix for each group of `size_g` rows. For example, for :code:`size_i = size_j = 4` and :code:`size_g = 2`, it will transform :: [[0 , 1 , 2 , 3 ], [4 , 5 , 6 , 7 ], [8 , 9 , 10, 11], [12, 13, 14, 15]] into :: [[0, 2, 4 , 6 ], [1, 3, 5 , 7 ], [8, 10, 12, 14], [9, 11, 13, 15]] rminimum) rjZsize_iZsize_jZsize_gZijZsize_gjZgroup_idZoff_iZnew_iZnew_jr r r swizzle2dLs  r'cCst|d|S)a' Returns a tensor filled with the scalar value 0 for the given :code:`shape` and :code:`dtype`. :param shape: Shape of the new array, e.g., (8, 16) or (8, ) :type shape: tuple of ints :param dtype: Data-type of the new array, e.g., :code:`tl.float16` :type dtype: DType r)rfull)shapedtyper r rzerosts r+cCst|j|jS)z Returns a tensor of zeros with the same shape and type as a given tensor. :param input: input tensor :type input: Tensor )r+r)r*)inputr r r zeros_likesr-c CsJ|r||ko||k}nd}||kp$|}t|||}t|||}||fSNFrwhere) value1index1value2index2tie_break_lefttiegtZv_retZi_retr r r_argmax_combines r8cCst||||dSNTr8r1r2r3r4r r r_argmax_combine_tie_break_leftsr<cCst||||dSr.r:r;r r r_argmax_combine_tie_break_fastsr=cCs t||SN)rmaximumabr r r_elementwise_maxsrCr?return_indicesreturn_indices_tie_break_left)Zreturn_indices_arg tie_break_argTcCst|}|r8|r$tj||t|dStj||t|dSnft|jjtdkrt|jrn| tj }n|j sJd| tj }tj ||t|dSdSNr z"Expecting input to be integer type)r_promote_bfloat16_to_float32_reduce_with_indicesr<r=r r*primitive_bitwidth is_floatingtofloat32is_intint32reducerCr,axisrDrErr r rrs  rz maximum indexr5)rFcCst||d||d\}}|SNT)rDrEr)rr,rSr5r_retr r rargmaxsrXc CsJ|r||ko||k}nd}||kp$|}t|||}t|||}||fSr.r/) r1r2r3r4r5r6ltZ value_retZ index_retr r r_argmin_combines rZcCst||||dSr9rZr;r r r_argmin_combine_tie_break_leftsr\cCst||||dSr.r[r;r r r_argmin_combine_tie_break_fastsr]cCs t||Sr>r$r@r r r_elementwise_minsr^r%cCst|}|r8|r$tj||t|dStj||t|dSn`t|jjdkrt|jrh| tj }n|j szJd| tj }tj ||t|dSdSrG)rrIrJr\r]r r*rKrLrMrNrOrPrQr^rRr r rmins  r_z minimum indexcCst||d||d\}}|SrT)r_rUr r rargminsr`cCs||Sr>r r@r r r _sum_combinesra)in_dtyper*cCsXt|}|dur|Sd}|r8|jdkr2tjnd}n|rT|jdkrPtjnd}|S)NrH)r_unwrap_if_constexprZ is_int_signedZ int_bitwidthrPZis_int_unsignedZuint32)rbr* out_dtyper r r_pick_sum_dtype s rerr*)Z dtype_arg)r*cCs0t|j|}|dur||}tj||t|dS)Nr)rer*rMrrQra)r,rSrr*rdr r rrs  cCs||ASr>r r@r r r _xor_combine%srfzxor sumcCs&t|jjdtj||t|dS)Nz#xor_sum only supported for integersr)r static_asserttypescalarrOrQrfr,rSrr r rxor_sum-srkcCs||BSr>r )ryr r r _or_combine8srmZ reduce_ofcCs&t|jjdtj||t|dS)Nz%reduce_of only supported for integersr)rrgrhrirOrQrmrjr r r reduce_or=srncumsumcCs8t|}t|j|}|dur(||}t||t|Sr>)rrIrer*rMassociative_scanra)r,rSreverser*rdr r rroHs    cCs||Sr>r r@r r r _prod_combineZsrrcumprodcCst|}t||t|Sr>)rrIrprr)r,rSrqr r rrs_s )n_dimsr&cCs:tdd}t|dg||ddgdg|}|S)Nrrr)rZaranger!)rtr&arr r r _indicatorks *rvc Cszt|j}tj|jjdd}|j|dd}|t||d|dA}|j|jdd}t||}t ||k||Ak||} | S)NTZbitwidthsignedZbitcastr) rr"r get_int_dtyper*rKrMrkrvr0) rfliprrtidtypeixZiyrlZis_rightrWr r r_compare_and_swaprs  r~)stageordercCsF|dkrtt|j|}n|}t|D]}t|||d|}q(|S)zb order_type 0 == ascending order_type 1 == descending order_type 2 == alternating rr)rvrr"r static_ranger~)rrrr{rr r r_bitonic_merge_hypercubes r)rrrtcCs6t|dgt|j}t|||}t||j}|S)Nr)rr!rr"rr))rrrrthr r r_bitonic_merges rkr descendingc Cs0|durt|jdn|}t|t|jdkdt|j|}|durN|nt|}t|j}t|dg|}td|dD]} t|| | |krdn|}qt|d|dD]V} |rt |t|jd|dnt |t|jd|d}t||| |krdn|}qt||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)rS) lenr)rrgrr"r!rrrr_) rrrrrZlog_nZlog_krtrrr r r sort_impls 8 rrrcCst|||dS)Nrr)rrrr r rsortsr)rrcCst|||ddS)NTrr)rrrr r rtopksrcCsP|durt|jdn|}t|t|jdkdt|jd}t||||S)Nrrr)rr)rrgrr)rrrrrtr r r bitonic_mergesrcCsFt|}t|}|dur(t|d}|dkr<|t|7}t|Sr)rrcrr )rr)r 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}qt ||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 TrwryNrr)rrgrr)rrrrzr*rKr!rMrrk)rrrZstepsr|rlrr r rr{s $ <r{cCsHt||}t|jdkr|St||jddd|jdgSdS)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)rjoinrr)r!)rArBcr r r interleaves r)NFF)F)NFTF)TF)NFTF)TF)NFN)NF)F)rFN)rF)N)N)6 __future__rZ runtime.jitrrrrrZ_tensor_member_fnrZ_add_math_1arg_docstrrrr#r'r+r-r8r<r=rCZ_add_reduction_docstrrrXrZr\r]r^r_r`rarerrfrkrmrnZ_add_scan_docstrrorrrsrvr~rrZ CONSTEXPR_0rrrrrr{rr r r rs          '                    (