Struct BlockPlan 

pub struct BlockPlan {
    pub per_site_m: usize,
    pub per_site_n: usize,
    pub grid: (usize, usize),
    pub inner: InnerLoop,
    pub block: BlockShape,
    pub acc: AccPolicy,
    pub pack_accumulate: bool,
    pub engines: EnginePacking,
    pub granularity: Granularity,
}

The per-site block plan the tile pass fixes.

Fields

per_site_m: usize

per_site_n: usize

grid: (usize, usize)

(n_y, n_x) blocks.

inner: InnerLoop

block: BlockShape

acc: AccPolicy

pack_accumulate: bool

The packer accumulates partial sums in place in the intermediate buffer, so the register file is reloaded once (at the last K block) instead of once per block.

engines: EnginePacking

How the site’s movement work is packed onto its engines.

granularity: Granularity

Whether the compute engine is handed a whole sub-block per call or one tile at a time.

Implementations

impl BlockPlan

pub fn site_block( &self, mt: usize, nt: usize, i: usize, j: usize, ) -> (usize, usize, usize, usize)

★ THE BLOCK THE SITE AT GRID POSITION (i, j) ACTUALLY COVERS — one definition, shared.

A grid extent need not divide the tile count: the last row and column carry what is LEFT. place emits each site’s patterns from this and work prices them from the same function, because two copies of a boundary rule disagree the first time either moves — and a planner that ranks by a closed form the emitter does not honour is the defect this campaign already removed once.

The sub-block must DIVIDE the extent it walks, or the store writes past a row and wraps into the next. Raggedness must not compound, so the largest divisor of the extent that is no bigger than the nominal sub-block is taken; 1 always divides, so a legal choice always exists.

Returns (rows, cols, sub_block_h, sub_block_w). On a grid whose extents DO divide, every site returns the nominal block and this is the identity.

pub fn site_blocks( &self, mt: usize, nt: usize, ) -> Vec<((usize, usize), (usize, usize, usize, usize))>

Every (i, j) in the grid that carries work, with its block — the boundary included.

Trait Implementations

impl Clone for BlockPlan

fn clone(&self) -> BlockPlan

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for BlockPlan

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for BlockPlan

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Hash for BlockPlan

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 BlockPlan

fn eq(&self, other: &BlockPlan) -> 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 Serialize for BlockPlan

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.

impl Copy for BlockPlan

impl Eq for BlockPlan

impl StructuralPartialEq for BlockPlan