Struct Costs 

pub struct Costs {
    pub latency: f64,
    pub up: f64,
    pub down: f64,
    pub gpu_gemm: f64,
    pub gpu_direct: f64,
    pub gpu_col: f64,
    pub cpu_gemm: f64,
    pub cpu_conv: f64,
    pub cpu_act: f64,
    pub gpu_act: f64,
}

实测的开销，单位都是秒。

Fields

latency: f64

一次空的往返（交给工人、启动、等完成）。

up: f64

每字节上传、下载。

down: f64

gpu_gemm: f64

GPU 上每次乘加：SGEMM、深度卷积的直接内核；im2col 每个元素。

gpu_direct: f64

gpu_col: f64

cpu_gemm: f64

CPU 上每次乘加：矩阵乘、卷积。

cpu_conv: f64

cpu_act: f64

激活函数每个元素：CPU、GPU（以 GELU 为代表，它最贵）。

gpu_act: f64

Implementations

impl Costs

pub fn gemm_on_gpu( &self, (m, n, k): (usize, usize, usize), cached_b: bool, ) -> bool

一次矩阵乘走 GPU 是否更快。cached_b 表示 b 已在显存。

pub fn conv_on_gpu(&self, s: &ConvShape, cached_w: bool) -> bool

一次卷积走 GPU 是否更快。cached_w 表示权重和偏置已在显存。

pub fn act_on_gpu(&self, n: usize) -> bool

对 n 个元素做激活走 GPU 是否更快。

Trait Implementations

impl Clone for Costs

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 Costs

impl Debug for Costs

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

Formats the value using the given formatter.

impl PartialEq for Costs

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

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for Costs