U yʹjz @sddlZddlmZmZddlmZddlmZddZeddd ied d d ieejd d dZ eddd ied dd ieejd ddZ Gdddej j Z e jZdS)N) heuristicsjit)languagenext_power_of_2cCs|dkr dS|dkrdSdS)Nii )Nr r W/var/www/html/bwave/python/venv/lib/python3.8/site-packages/triton/ops/cross_entropy.py num_warpss rcCs t|dSNr rnargsr r r rBLOCKcCs t|dSrrrr r r rr)rc Cstd}td|}t||}||||}||||} ||||} tj|||ktd d} | tj} | t| d} tt t | d| } tj | | ||kdt t| } t ||| dSNrinf)maskother)r) tl program_idarangeloadfloattofloat32maxlogsumexpstoreZ debug_barrier) ZLOGITSPROBSIDXZLOSSr rrowcolsidxZ WRIT_PROBSZ READ_PROBSlogitsprobsr r r _forwards    r-cCs t|dSrrrr r r r)rcCs t|dSrrrr r r r*rc Cstd}td|}t||}||||}tj|||ktdd }t|tj}||k} t||} || | } tj|| |j j ||kddSr) rrrrrr$rr r%dtypeZ element_ty) r&r'ZDPROBSr rr(r)r*r,deltaZdoutZdinr r r _backward)s   r0c@s$eZdZeddZeddZdS)_cross_entropyc s~|jtjkstdjj}}jdtj|||d}tj||d}fdd}t||||||||S)Nz(Indices are expected to be of type long.)r.devicecsfSNZnumeloptr+n_colsr r rHrz(_cross_entropy.forward..) r.torchZint64AssertionErrorr3shapeZ empty_liker-Zsave_for_backward) clsctxr+indicesr3r.result neg_logprobsgridr r8r forward>s  z_cross_entropy.forwardcs<|j\}jdfdd}t|||dfS)aWe know d(-log(p[i])/dlogit[k] = -id_mat[i,k] + p[k] so we initialize the gradient as neg_logprobs, so we can just exponentiate to get p[k], which is most of what we need... neg_logprobs will be modified in place to become the gradient we want r2csfSr4r5r6r9rAr r rZrz)_cross_entropy.backward..N)Z saved_tensorsr<r0)r=r>Z dneg_logprobsr?rBr rDr backwardNs   z_cross_entropy.backwardN)__name__ __module__ __qualname__ classmethodrCrEr r r r r1=s r1)r:rrrrrrZ constexprr-r0ZautogradFunctionr1applyZ cross_entropyr r r r s  "