Trait OpInput

Source

pub trait OpInput<'a, T: Dist> { }

Expand description

This trait is used to represent the input to a batched LamellarArray element-wise operation.

§Contents

Overview
Batch-compatible input types and type conversion methods

§Overview

Valid inputs to batched operations are essentially “safe” list-like data structures, such as Vec<T> and slice &[T]. We have also provided functions on the LamellarArray types that allow you to access a PE’s local data. These functions appropriately protect the local data so the the safety guarantees of the given array type are maintained. See a list of batch-compatible input types and type conversion methods below.

Currently it is not recommended to try to implement this for your types. Rather you should try to convert to a slice if possible.

§Batch-compatible input types and type conversion methods

Methods that return a batch-compatible input type (and the array type that provides it):

local_data- (AtomicArray, ReadOnlyArray)
mut_local_data - (AtomicArray, ReadOnlyArray)
read_local_data - (LocalLockArray,GlobalLockArray)
write_local_data -(LocalLockArray,GlobalLockArray)

Batch-compatible input types:

T and &T
&[T]
- the local data of a ReadOnlyArray is directly the underlying slice
- read_only_array.local_data();
Vec[T] and &Vec[T]
AtomicLocalData
- atomic_array.local_data();
LocalLockLocalData
- local_lock_array.read_local_data();
- local_lock_array.write_local_data();
GlobalLockLocalData
- global_lock_array.read_local_data();
- global_lock_array.write_local_data();

It is possible to use a LamellarMemoryRegion or UnsafeArray as the parent source, but these will require unsafe calls to retrieve the underlying slices. The retrieved slices can then be used for the batched operation

 unsafe { onesided_mem_region.as_slice().expect("not on allocating PE") };
 unsafe { shared_mem_region.as_slice() };
 unsafe { unsafe_array.local_data() };