Enum ConvKind 

#[non_exhaustive]
#[repr(u16)]
pub enum ConvKind {
    Conv2d = 0,
    Conv2dBackwardData = 1,
    Conv2dBackwardFilter = 2,
    Conv1d = 3,
    Conv1dBackwardData = 4,
    Conv1dBackwardFilter = 5,
    Conv3d = 6,
    Conv3dBackwardData = 7,
    Conv3dBackwardFilter = 8,
    ConvTranspose2d = 9,
    ConvTranspose2dBackward = 10,
    DepthwiseConv2d = 11,
    Unfold = 12,
    Fold = 13,
    ConvTranspose1d = 14,
    ConvTranspose1dBackwardData = 15,
    ConvTranspose1dBackwardFilter = 16,
    ConvTranspose2dBackwardData = 17,
    ConvTranspose2dBackwardFilter = 18,
    ConvTranspose3d = 19,
    ConvTranspose3dBackwardData = 20,
    ConvTranspose3dBackwardFilter = 21,
    Im2Col2d = 22,
    Im2Col1d = 23,
    Col2Im1d = 24,
}

Convolution-family op discriminant — Category I from the comprehensive plan.

Stored as u16 in crate::KernelSku::op when category == OpCategory::Convolution. Each variant maps to a distinct cuDNN exec path (forward, data-gradient, filter-gradient) of the underlying convolution descriptor. The dimensional axis (1-D / 2-D / 3-D), padding / stride / dilation, and depthwise / transposed flavors live on the per-plan descriptor — they don’t fan out a separate enum slot here.

Today wired: Conv2d × {f32, f64, f16, bf16} (FW + BW data + BW filter) via cuDNN. Conv1d / Conv3d / ConvTranspose* / depthwise / unfold / fold are reserved discriminants for fanout milestones.

Variants (Non-exhaustive)

Non-exhaustive enums could have additional variants added in future. Therefore, when matching against variants of non-exhaustive enums, an extra wildcard arm must be added to account for any future variants.

Conv2d = 0

2-D convolution forward pass. PyTorch torch.nn.functional.conv2d. Trailblazer for Phase 7.

Conv2dBackwardData = 1

2-D convolution data-gradient pass (computes dx from dy and the filter w). PyTorch’s autograd-internal conv2d_backward_input.

Conv2dBackwardFilter = 2

2-D convolution filter-gradient pass (computes dw from x and dy). PyTorch’s autograd-internal conv2d_backward_weight.

Conv1d = 3

1-D convolution forward. Reserved.

Conv1dBackwardData = 4

1-D convolution data-gradient. Reserved.

Conv1dBackwardFilter = 5

1-D convolution filter-gradient. Reserved.

Conv3d = 6

3-D convolution forward. Reserved.

Conv3dBackwardData = 7

3-D convolution data-gradient. Reserved.

Conv3dBackwardFilter = 8

3-D convolution filter-gradient. Reserved.

ConvTranspose2d = 9

2-D transposed convolution (fractionally-strided / “deconv”). Forward pass.

ConvTranspose2dBackward = 10

2-D transposed convolution backward. Reserved — backward is dispatched through the same plan via run_bw_data / run_dw.

DepthwiseConv2d = 11

Depthwise 2-D convolution (groups == c_in). Today callers route through the generic Conv2dPlan with groups set on the descriptor — cuDNN’s cudnnSetConvolutionGroupCount detects the depthwise path automatically.

Unfold = 12

torch.nn.functional.unfold — extract sliding windows. Reserved.

Fold = 13

torch.nn.functional.fold — inverse of unfold. Reserved.

ConvTranspose1d = 14

1-D transposed convolution forward.

ConvTranspose1dBackwardData = 15

1-D transposed convolution data-gradient.

ConvTranspose1dBackwardFilter = 16

1-D transposed convolution filter-gradient.

ConvTranspose2dBackwardData = 17

2-D transposed convolution data-gradient.

ConvTranspose2dBackwardFilter = 18

2-D transposed convolution filter-gradient.

ConvTranspose3d = 19

3-D transposed convolution forward.

ConvTranspose3dBackwardData = 20

3-D transposed convolution data-gradient.

ConvTranspose3dBackwardFilter = 21

3-D transposed convolution filter-gradient.

Im2Col2d = 22

2-D im2col — torch.nn.functional.unfold (Phase 19.3). Extracts sliding windows from an NCHW input into an [N, C·kh·kw, h_out·w_out] column-shaped matrix. Distinct from the reserved Self::Unfold discriminant for forward-source- compat; the 19.3 wiring routes through this discriminant.

Im2Col1d = 23

1-D im2col (NCL → [N, C·kl, l_out]).

Col2Im1d = 24

1-D col2im — inverse of Self::Im2Col1d. Atomic-add scatter.

Trait Implementations

impl Clone for ConvKind

fn clone(&self) -> ConvKind

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for ConvKind

impl Debug for ConvKind

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Eq for ConvKind

impl Hash for ConvKind

fn hash<__H: Hasher>(&self, state: &mut __H)

Feeds this value into the given Hasher.

fn hash_slice<H>(data: &[Self], state: &mut H)

where H: Hasher, Self: Sized,

Feeds a slice of this type into the given Hasher.

impl PartialEq for ConvKind

fn eq(&self, other: &ConvKind) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl StructuralPartialEq for ConvKind