Struct Pending 

pub struct Pending<T: PendingValue> { /* private fields */ }

Work that completes later, and the T it produces then.

Implementations

impl<T: PendingValue> Pending<T>

pub fn wait(self) -> T

Waits for the work to complete, and returns what it produced.

pub fn __expand_wait( scope: &Scope, this: <Self as CubeType>::ExpandType, ) -> <T as CubeType>::ExpandType

Waits for the work to complete, and returns what it produced.

impl Pending<()>

pub fn wait_complete(self)

Waits for the work to fully complete. A TMA store has then also performed its writes to global memory, which a fence or an atomic after the wait publishes to other cubes. Anything else completes as Pending::wait waits for it.

pub fn __expand_wait_complete( scope: &Scope, this: <Self as CubeType>::ExpandType, )

Waits for the work to fully complete. A TMA store has then also performed its writes to global memory, which a fence or an atomic after the wait publishes to other cubes. Anything else completes as Pending::wait waits for it.

impl<CD: Numeric> Pending<Accumulator<CD>>

pub fn downcast<O: Numeric>(&self) -> Pending<Accumulator<O>>

This accumulator named at O, panicking unless O is the element it holds.

Like a slice’s downcast, this only satisfies the Rust type system: the registers and the wait are the same, so it costs nothing.

pub unsafe fn downcast_unchecked<O: Numeric>(&self) -> Pending<Accumulator<O>>

This accumulator named at O, whatever element it holds.

Safety

Only for keeping the accumulator at a dummy type, to be named again with downcast where it is used, as a slice’s downcast_unchecked is.

impl<CD: Numeric> Pending<Accumulator<CD>>

pub fn execute<A: Scalar, B: Scalar>( &mut self, a: &MatrixDescriptor<A>, b: &MatrixDescriptor<B>, )

Issues D = A * B + D, with A and B in shared memory. Each MMA reads 64 rows of A and n of B, and WgmmaTileLayout::k_step elements of K, from where the descriptors point.

pub fn execute_registers<A: Scalar, B: Scalar>( &mut self, a: &Fragment<A>, b: &MatrixDescriptor<B>, )

Issues D = A * B + D, with A in registers and B in shared memory. The fragment must stay unchanged until the MMA’s group completes.

pub fn reset(&mut self)

Waits for every MMA issued into the accumulator, zeroes it, and hands it back to the MMAs to come, in the same registers: what Accumulator::new then Accumulator::start give, with no copy written after the fence the MMAs to come read past.

pub fn commit(&mut self) -> Pending<()>

Commits the MMAs issued since the last commit into a group, and returns its completion. Once it resolves, the MMAs are done reading their tiles and fragments.

pub fn __expand_execute<A: Scalar, B: Scalar>( scope: &Scope, this: &mut <Self as CubeType>::ExpandType, a: &<MatrixDescriptor<A> as CubeType>::ExpandType, b: &<MatrixDescriptor<B> as CubeType>::ExpandType, )

Issues D = A * B + D, with A and B in shared memory. Each MMA reads 64 rows of A and n of B, and WgmmaTileLayout::k_step elements of K, from where the descriptors point.

pub fn __expand_execute_registers<A: Scalar, B: Scalar>( scope: &Scope, this: &mut <Self as CubeType>::ExpandType, a: &<Fragment<A> as CubeType>::ExpandType, b: &<MatrixDescriptor<B> as CubeType>::ExpandType, )

Issues D = A * B + D, with A in registers and B in shared memory. The fragment must stay unchanged until the MMA’s group completes.

pub fn __expand_reset(scope: &Scope, this: &mut <Self as CubeType>::ExpandType)

Waits for every MMA issued into the accumulator, zeroes it, and hands it back to the MMAs to come, in the same registers: what Accumulator::new then Accumulator::start give, with no copy written after the fence the MMAs to come read past.

pub fn __expand_commit( scope: &Scope, this: &mut <Self as CubeType>::ExpandType, ) -> <Pending<()> as CubeType>::ExpandType

Commits the MMAs issued since the last commit into a group, and returns its completion. Once it resolves, the MMAs are done reading their tiles and fragments.

Trait Implementations

impl<T: PendingValue + Clone> Clone for Pending<T>

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 Pending<()>

impl<T: PendingValue> CubeType for Pending<T>

type ExpandType = PendingExpand<T>