WARNING: BEFORE STARTING ON A STRUCTURED KERNEL PORT, CHECK OPEN PULL REQUESTS TO SEE IF SOMEONE IS WORKING ON IT ALREADY. https://github.com/pytorch/pytorch/pulls?q=is%3Apr+is%3Aopen+structured+in%3Atitle

RFC at: https://github.com/pytorch/rfcs/blob/rfc-0005/RFC-0005-structured-kernel-definitions.md

Here is the list of all functions which are immediately eligible to be ported to structured (as in, they have both a functional and out version with exactly the same signatures):

• abs (Anjali)
• absolute
• angle
• sgn
• conj
• acos Port all unary float functions to structured #56082
• arccos
• add.Tensor
• _add_relu.Tensor
• addmv (addmv: port to structured kernels, improve error checks #55746)
• addr
• all.dim
• all.dimname
• any.dim
• any.dimname
• argmax
• argmin
• acosh (acosh: port to structured kernel #55540)
• arccosh
• asinh Port all unary float functions to structured #56082
• arcsinh
• atanh
• arctanh
• asin
• arcsin
• atan Port all unary float functions to structured #56082
• arctan
• baddbmm [Structured Kernels] Port for baddbmm and bmm #64805
• bernoulli
• binary_cross_entropy
• binary_cross_entropy_backward
• bitwise_not
• copysign.Tensor copysign: port to structured kernel #55040
• logical_not
• logical_xor
• logical_and
• logical_or
• bmm [Structured Kernels] Port for baddbmm and bmm #64805
• _bmm
• cat
• cat.names
• ceil (ceil: port to structured #57589)
• clamp (Add OptionalRef; clamp: port to structured kernel #61361)
• clamp_max
• clamp_min
• clip
• complex
• polar
• cos (cos: port to structured kernel #55564)
• cosh (cosh: port to structured kernel #55563)
• cummax
• cummax.dimname
• cummin
• cummin.dimname
• cumprod
• cumprod.dimname
• cumsum
• cumsum.dimname
• diff
• div.Tensor
• div.Tensor_mode
• divide.Tensor
• divide.Tensor_mode
• true_divide.Tensor
• dot
• vdot
• row_stack
• erf Port all unary float functions to structured #56082
• erfc Port all unary float functions to structured #56082
• exp Port all unary float functions to structured #56082
• exp2 Port all unary float functions to structured #56082
• expm1 Port all unary float functions to structured #56082
• floor (floor: port to structured #57587)
• floor_divide
• frac
• gcd (gcd: port to structured #57624)
• lcm (lcm: port to structured #57628)
• _fft_r2c
• _fft_c2r
• _fft_c2c ([WIP] Migrate _fft_c2c_mkl to structured kernel #55730)
• inverse
• kron
• kthvalue
• kthvalue.dimname
• nan_to_num
• ldexp.Tensor
• log Port all unary float functions to structured #56082
• log10 Port all unary float functions to structured #56082
• log1p Port all unary float functions to structured #56082
• log2 Port all unary float functions to structured #56082
• logaddexp (logaddexp and logaddexp2: port to structured #57629)
• logaddexp2 (logaddexp and logaddexp2: port to structured #57629)
• xlogy.Tensor (xlogy: Port to structured #60814)
• xlogy.Scalar_Self (xlogy: Port to structured #60814)
• xlogy.Scalar_Other (xlogy: Port to structured #60814)
• _logcumsumexp
• logcumsumexp
• logcumsumexp.dimname
• _log_softmax (Port log_softmax to structured kernel #57374)
• _log_softmax_backward_data @SplitInfinity
• logsumexp
• logsumexp.names
• matmul
• matrix_power
• _compute_linear_combination
• max.dim
• max.names_dim
• amax Port amax to structured kernel #72124
• mean.dim
• mean.names_dim
• median.dim
• median.names_dim
• nanmedian.dim
• nanmedian.names_dim
• min.dim
• min.names_dim
• amin
• mm (port mm to structure kernel  #57755)
• mode
• mode.dimname
• mul.Tensor
• multiply.Tensor
• mv
• narrow_copy
• native_batch_norm
• batch_norm_elemt
• rad2deg
• deg2rad
• reciprocal Port all unary float functions to structured #56082
• neg neg: port to structure kernel #57212
• negative
• round
• rsqrt
• silu (silu: port to structured #58050)
• sigmoid Port all unary float functions to structured #56082
• logit
• sin
• sinc
• sinh (sinh: port to structured kernel #55538)
• sspaddmm
• stack
• _stack
• hstack
• vstack
• dstack
• sum.dim_IntList
• sum.dim_DimnameList
• nansum.dim_IntList
• sqrt Port all unary float functions to structured #56082
• square (port square to structured #58266)
• std.dim
• std.names_dim
• prod.dim_int
• prod.dim_Dimname
• tan Port all unary float functions to structured #56082
• tanh Port all unary float functions to structured #56082
• tensordot
• threshold
• trunc (trunc: port to structured #57350)
• fix
• var.dim
• var.names_dim
• norm.ScalarOpt_dim_dtype
• norm.ScalarOpt_dim
• norm.names_ScalarOpt_dim_dtype
• norm.names_ScalarOpt_dim
• frexp.Tensor
• frobenius_norm.dim
• nuclear_norm
• nuclear_norm.dim
• sub.Tensor
• subtract.Tensor
• heaviside
• addmm (addmm: port to structured kernel #57417)
• hspmm
• eq.Scalar (Port eq kernel to structured kernels. #60177)
• eq.Tensor (Port eq kernel to structured kernels. #60177)
• bitwise_and.Tensor (bitwise_and: Port to structured #60813)
• bitwise_and.Scalar (bitwise_and: Port to structured #60813)
• bitwise_or.Tensor
• bitwise_or.Scalar
• bitwise_xor.Tensor (bitwise_xor: Port to structured #60812)
• bitwise_xor.Scalar (bitwise_xor: Port to structured #60812)
• atan2 atan2: port to structured kernel #55130
• tril
• triu
• digamma Port all unary float functions to structured #56082
• renorm
• lerp.Scalar
• lerp.Tensor
• fmod.Scalar (fmod: Port to structured #60809)
• fmod.Tensor (fmod: Port to structured #60809)
• remainder.Scalar
• remainder.Tensor
• addbmm (Port addbmm to structured kernels #60647)
• addcdiv
• diag
• cross
• ne.Scalar
• ne.Tensor
• not_equal.Scalar
• not_equal.Tensor
• ge.Scalar
• ge.Tensor
• greater_equal.Scalar
• greater_equal.Tensor
• le.Scalar
• le.Tensor
• less_equal.Scalar
• less_equal.Tensor
• gt.Scalar
• gt.Tensor
• greater.Scalar
• greater.Tensor
• lt.Scalar
• lt.Tensor
• less.Scalar
• less.Tensor
• take
• take_along_dim
• index_select
• index_select.dimname
• masked_select
• nonzero (@krshrimali)
• gather
• gather.dimname
• addcmul
• lstsq
• triangular_solve Port triangular_solve to structured kernel #61857
• symeig
• eig
• svd
• cholesky
• cholesky_solve
• solve
• cholesky_inverse
• qr
• geqrf
• orgqr
• ormqr
• lu_solve
• multinomial
• lgamma Port all unary float functions to structured #56082
• polygamma
• erfinv Port all unary float functions to structured #56082
• i0
• sign (sign: port to structured #57588)
• signbit (Port signbit to structured kernel #57936)
• histc
• hypot (hypot: port to structured #57627)
• igamma (igamma and igammac: port to structured #57626)
• igammac (igamma and igammac: port to structured #57626)
• nextafter (nextafter: port to structured #57625)
• fmin
• fmax
• maximum (maximum and minimum: port to structured #57630)
• max.other
• minimum (maximum and minimum: port to structured #57630)
• min.other
• quantile.scalar
• quantile
• nanquantile.scalar
• nanquantile
• sort
• sort.stable
• sort.dimname
• sort.dimname_stable
• msort
• topk
• pow.Tensor_Tensor
• pow.Scalar
• pow.Tensor_Scalar
• float_power.Tensor_Tensor - [Port float_power kernel to structured kernels #60855]
• float_power.Scalar - [Port float_power kernel to structured kernels #60855]
• float_power.Tensor_Scalar = [Port float_power kernel to structured kernels #60855]
• normal.Tensor_float
• normal.float_Tensor
• normal.Tensor_Tensor
• _cumsum
• _cumprod
• _cat
• _mode
• bucketize.Tensor
• searchsorted.Tensor
• mse_loss
• mse_loss_backward
• l1_loss
• l1_loss_backward
• multi_margin_loss
• multi_margin_loss_backward
• multilabel_margin_loss
• multilabel_margin_loss_forward
• multilabel_margin_loss_backward
• nll_loss
• nll_loss_forward
• nll_loss_backward
• nll_loss2d
• nll_loss2d_forward
• nll_loss2d_backward
• smooth_l1_loss (Port smooth_l1_loss to structured kernels #67404)
• smooth_l1_loss_backward
• huber_loss
• huber_loss_backward
• soft_margin_loss
• soft_margin_loss_backward
• elu (elu: port to structured #57619)
• glu
• glu_backward
• hardsigmoid (hardsigmoid: port to structured #57622)
• hardtanh
• hardtanh_backward
• hardswish (this one is complicated, see Converting hardswish to strucutred kernels with metatensor support #66899 )
• leaky_relu (leaky_relu: port to structured #57621)
• log_sigmoid
• log_sigmoid_forward
• log_sigmoid_backward
• rrelu_with_noise
• softplus (softplus: port to structured #57620)
• softplus_backward (Port softplus_backward to structured #58482)
• softshrink (softshrink: port to structured #57623)
• softshrink_backward
• (XLA) adaptive_avg_pool2d (this one is a little tricky because there is specialized derivative handling; see ngimel)
• adaptive_avg_pool3d
• adaptive_avg_pool3d_backward
• adaptive_max_pool2d (adaptive_max_pool2d: port to structured kernel #56317)
• adaptive_max_pool2d_backward (adaptive_max_pool2d_backward: port to structured kernel #56799)
• adaptive_max_pool3d (adaptive_max_pool3d: port to structured kernel #56320)
• adaptive_max_pool3d_backward (Port adaptive_max_pool3d_backward to structured kernel #56800)
• avg_pool2d (avg_pool2d: port to structured #58987)
• avg_pool2d_backward
• avg_pool3d
• avg_pool3d_backward (avg_pool3d_backward: Port to structured #59084)
• fractional_max_pool2d
• fractional_max_pool2d_backward
• fractional_max_pool3d
• fractional_max_pool3d_backward
• max_pool2d_with_indices
• max_pool2d_with_indices_backward
• max_pool3d_with_indices
• max_pool3d_with_indices_backward
• max_unpool2d
• max_unpool2d_backward
• max_unpool3d
• max_unpool3d_backward
• reflection_pad1d - reflection_pad1: port to structured kernel #55531
• reflection_pad1d_backward - reflection_pad1d_backward: Port to structured #59103
• reflection_pad2d
• reflection_pad2d_backward
• replication_pad1d - replication_padding1d: port to structured #55481
• replication_pad1d_backward - ezyang
• replication_pad2d - asuhan
• replication_pad2d_backward
• replication_pad3d - ezyang
• replication_pad3d_backward
• upsample_linear1d
• upsample_linear1d_backward
• upsample_bilinear2d
• upsample_bilinear2d_backward
• upsample_bicubic2d
• upsample_bicubic2d_backward
• upsample_trilinear3d
• upsample_trilinear3d_backward
• upsample_nearest1d
• upsample_nearest1d_backward
• upsample_nearest2d
• upsample_nearest2d_backward
• upsample_nearest3d
• upsample_nearest3d_backward
• sigmoid_backward - [sigmoid_backward: Port to structured #60815]
• logit_backward - [logit_backward: Port to structured #60817]
• tanh_backward - [tanh_backward: Port to structured #60816]
• slow_conv_transpose2d - [Port slow_conv_transpose2d to structured #55503, but blocked on Optional[Tensor] support]
• slow_conv_transpose3d
• thnn_conv2d
• thnn_conv2d_forward
• thnn_conv_depthwise2d
• thnn_conv_depthwise2d_forward
• slow_conv3d
• slow_conv3d_forward
• col2im
• col2im_backward
• column_stack
• im2col
• im2col_backward
• isposinf (Port isposinf & isneginf kernel to structured kernels #60633)
• isneginf (Port isposinf & isneginf kernel to structured kernels #60633)
• special_entr Port all unary float functions to structured #56082
• special_gammaln
• special_erf
• special_erfc
• special_erfinv
• fft_fft
• fft_ifft
• fft_rfft
• fft_irfft
• fft_hfft
• fft_ihfft
• fft_fft2
• fft_ifft2
• fft_rfft2
• fft_irfft2
• fft_fftn
• fft_ifftn
• fft_rfftn
• fft_irfftn
• linalg_cholesky
• linalg_slogdet
• linalg_eigh
• linalg_householder_product
• linalg_inv
• inner
• outer
• ger
• linalg_norm
• linalg_norm.ord_str
• linalg_vector_norm
• linalg_svd
• linalg_cond
• linalg_cond.p_str
• linalg_pinv
• linalg_pinv.rcond_tensor
• linalg_solve
• linalg_tensorinv
• linalg_tensorsolve
• linalg_qr
• linalg_matrix_power
• linalg_matrix_rank

Things that require adding an out kernel and then can be made structured:

• _embedding_bag (high priority)

Issues marked (XLA) are supported by XLA and thus are higher priority to port to structured.

Functions with 1d/2d/3d in their name tend to be easier and more beginner friendly

Some of these functions may not be immediately portable. Common grounds for disqualification:

• It's a reduction (blocked on Don't allocate result Tensors in out overloads: Reduction Ops #53218)
• It's multi output (need to implement this)
• It's still a TH kernel (need to do the TH to ATen port first)
• It's an alias/composite kernel (requires design from Dispatch-less structured wrapper / composite / alias kernels #50953)
• It behaves differently in CPU and CUDA; you'll have to fix this bug first! (Most common situation is the memory format behavior is divergent)
• It has optional tensors