Skip to main content

AllocDispatch

atomr_accel_cuda::device

Trait AllocDispatch 

Source 
pub trait AllocDispatch: Send + 'static {
    // Required methods
    fn dtype(&self) -> DType;
    fn len(&self) -> usize;
    fn run(
        self: Box<Self>,
        stream: Option<&Arc<CudaStream>>,
        state: &Arc<DeviceState>,
        mock_mode: bool,
    );
}
Expand description

Trait object the DeviceActor stashes inside a single crate::device::DeviceMsg::Alloc variant. The actual T is erased at the box boundary; the receiving ContextActor calls AllocDispatch::run which downcasts back into the concrete AllocReq<T> and performs a typed stream.alloc_zeros::<T>(len).

Required Methods§

Source

fn dtype(&self) -> DType

Concrete dtype carried by this dispatcher.

Source

fn len(&self) -> usize

Element count being allocated.

Source

fn run( self: Box<Self>, stream: Option<&Arc<CudaStream>>, state: &Arc<DeviceState>, mock_mode: bool, )

Execute the allocation against the given context’s primary stream and reply on the embedded oneshot channel.

mock_mode == true makes this match the legacy mock-mode behaviour of super::context_actor::ContextActor::alloc — it always replies with GpuError::Unrecoverable("alloc not supported in mock mode") so existing tests are unaffected.

Implementors§