Enum OpNode 

pub enum OpNode {
    Elementwise {
        f: EwKind,
    },
    MatMul {
        m: u32,
        n: u32,
        k: u32,
        tc: bool,
        trans_b: bool,
    },
    Gemv {
        m: u32,
        n: u32,
        tc: bool,
    },
    Fft {
        len: u32,
        inverse: bool,
    },
    Reduce {
        op: RedKind,
        axis: Axis,
    },
    GatherDequant {
        scheme: DType,
        block: u32,
    },
    Broadcast {
        shape: Shape,
    },
    Softmax {
        axis: Axis,
    },
    Stencil {
        kind: StencilKind,
        halo: u32,
        axis: Axis,
    },
    ScatterAccum {
        op: AccumKind,
    },
    Neighbor {
        kind: NbKind,
        k_or_r: f32,
        dims: u8,
        enc: NeighborEnc,
    },
    Slice {
        offset: u32,
        len: u32,
    },
    Rope {
        head_dim: u32,
        pos: u32,
        mode: u32,
        base_bits: u32,
    },
    Pool2d {
        c: u32,
        h: u32,
        w: u32,
        kh: u32,
        kw: u32,
        stride_h: u32,
        stride_w: u32,
    },
    Resize2d {
        c: u32,
        h_in: u32,
        w_in: u32,
        h_out: u32,
        w_out: u32,
    },
    Conv2d {
        c_in: u32,
        c_out: u32,
        h: u32,
        w: u32,
        kh: u32,
        kw: u32,
        stride_h: u32,
        stride_w: u32,
        pad_h: u32,
        pad_w: u32,
    },
}

An op-CLASS and its compile-time payload — the node label of the compute DAG (plan §2; Slice/Rope added for the real decode layer). Seed arms (Elementwise/MatMul/Gemv/ Fft) have a real lowering today; the rest are declared so the vocabulary is fixed, and lower to an explicit Err until their phase builds them.

Variants

Elementwise

Fields

f: EwKind

MatMul

Fields

m: u32

n: u32

k: u32

tc: bool

trans_b: bool

Gemv

Fields

m: u32

n: u32

tc: bool

Tensor-core (WMMA) eligible — the CUDA-C lowerer emits the wmma dequant-GEMV kernel when true, the plain f32 GEMV when false. Mirrors MatMul::tc.

Fft

Fields

len: u32

inverse: bool

Reduce

Fields

op: RedKind

axis: Axis

GatherDequant

Fields

scheme: DType

block: u32

Broadcast

Fields

shape: Shape

Softmax

Fields

axis: Axis

Stencil

Fields

kind: StencilKind

halo: u32

axis: Axis

ScatterAccum

Fields

op: AccumKind

Neighbor

Fields

kind: NbKind

k_or_r: f32

dims: u8

enc: NeighborEnc

Slice

Extract a contiguous sub-range input[offset .. offset+len] (1 input → [len]). The composable primitive for per-head slicing of projected q/K/V in multi-head attention.

Fields

offset: u32

len: u32

Rope

Rotary position embedding over a flat [.., head_dim] vector (1 input). pos is carried so the executor writes it into the kernel’s params buffer — the kernel source is independent of pos, so the pipeline cache stays warm across tokens. mode is 0=interleaved / 1=NeoX; base_bits is the f32 θ-base bit pattern. See graph_ops::stencil::{rope_wgsl, rope_cpu}.

Fields

head_dim: u32

pos: u32

mode: u32

base_bits: u32

Pool2d

NCHW max-pool 2D (vision). Inputs: feature map only; params packed in op.

Fields

c: u32

h: u32

w: u32

kh: u32

kw: u32

stride_h: u32

stride_w: u32

Resize2d

NCHW nearest resize 2D (vision).

Fields

c: u32

h_in: u32

w_in: u32

h_out: u32

w_out: u32

Conv2d

NCHW Conv2D f32 (vision). Inputs: activation, weight, bias (bias may be zeros).

Fields

c_in: u32

c_out: u32

h: u32

w: u32

kh: u32

kw: u32

stride_h: u32

stride_w: u32

pad_h: u32

pad_w: u32

Trait Implementations

impl Clone for OpNode

fn clone(&self) -> OpNode

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for OpNode

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

Formats the value using the given formatter.

impl PartialEq for OpNode

fn eq(&self, other: &OpNode) -> 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 Copy for OpNode

impl StructuralPartialEq for OpNode