Enum OpSize 

pub enum OpSize {
    MatMul {
        m: usize,
        n: usize,
        k: usize,
    },
    Elementwise(usize),
    Reduce {
        rows: usize,
        cols: usize,
    },
    ReduceTo {
        rows: usize,
        cols: usize,
    },
    Attention {
        b: usize,
        hq: usize,
        hkv: usize,
        d: usize,
        k_len: usize,
    },
}

Per-op size descriptor. Exposed (not private) so tests can supply a shrunk size_plan_override; not in the persisted profile report shape — the report uses the bucketed SizeClass instead.

Variants

MatMul

Dense matmul [m, k] @ [k, n] -> [m, n].

Fields

m: usize

n: usize

k: usize

Elementwise(usize)

Flat 1-D shape used by unary, binary, scalar/affine, clamp, powi kernels. Probe builds inputs of shape [n].

Reduce

Reduction along the last dim of a 2-D shape [rows, cols]. Total element count = rows × cols. Used for SumReduce / MaxReduce / MinReduce / MeanReduce.

Fields

rows: usize

cols: usize

ReduceTo

Reduce-to-broadcast-target from [rows, cols] → [1, cols] (sum-reduce-along-rows). Used for ReduceSumTo / ReduceMaxTo.

Fields

rows: usize

cols: usize

Attention

Decode-shaped GQA attention for OpKind::FlashAttn. seq_q is fixed at 1 (autoregressive decode — the regime slice-4’s flash-vs-decomposed ranking cares about). Builds q = [b, hq, 1, d], k = v = [b, hkv, k_len, d] (hq % hkv == 0, GQA), attending the full k_len prefix. Keyed via SizeClass::attention(hq, k_len, d).

Fields

b: usize

hq: usize

hkv: usize

d: usize

k_len: usize

Trait Implementations

impl Clone for OpSize

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for OpSize

impl Debug for OpSize

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

Formats the value using the given formatter.

impl Eq for OpSize

impl PartialEq for OpSize

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

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for OpSize