Struct Conv2dCost 

pub struct Conv2dCost {
    pub batch: usize,
    pub channels_in: usize,
    pub spatial_in: [usize; 2],
    pub channels_out: usize,
    pub kernel: [usize; 2],
    pub spatial_out: [usize; 2],
    pub bias_elems: usize,
    pub elems: MatmulGlobalElems,
}

Minimal representation of a 2D convolution’s cost: the shapes it moves and the element types it moves them in.

Fields

batch: usize

Batch, input channels, and the input’s spatial extent.

channels_in: usize

spatial_in: [usize; 2]

channels_out: usize

Output channels and the filter’s spatial extent.

kernel: [usize; 2]

spatial_out: [usize; 2]

The output’s spatial extent, which stride, padding and dilation decide.

bias_elems: usize

Elements of the bias, zero when the convolution has none.

elems: MatmulGlobalElems

Global element types of the operands.

Implementations

impl Conv2dCost

pub fn work(&self) -> Work

Calculates the compute operations and compulsory memory traffic for the convolution.

pub fn compute_ops(&self) -> usize

Operations the contraction performs, 2 * k - 1 per output element.

pub fn traffic(&self) -> (usize, usize)

Compulsory global traffic in bytes, split by direction, which work sums.

The maps and the filter are each moved once. An unfolded operand would re-read every overlapping window, which is the implementation’s choice rather than the convolution’s cost.

pub fn compute_key(&self, client: &Client) -> ThroughputKey

The probe the contraction’s arithmetic runs on, which is the matmul’s: an accelerated convolution issues the same MMA a gemm does.

Trait Implementations

impl Clone for Conv2dCost

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Conv2dCost

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

Formats the value using the given formatter.