Struct ErasedUpdatePlan 

pub struct ErasedUpdatePlan { /* private fields */ }

Dtype-erased wrapper around UpdatePlan.

The layout traversal is compiled once and stored without a scalar type; each execute names the concrete scalar type, resolves the runtime element op once per call, and then replays monomorphized loops.

Implementations

impl ErasedUpdatePlan

pub fn compile( dtype: KernelDType, dims: &[usize], dst_strides: &[isize], src_strides: &[isize], ) -> Result<Self>

Compile an update plan for one dtype and layout pair.

pub fn dtype(&self) -> KernelDType

pub fn plan(&self) -> &UpdatePlan

The compiled typed traversal.

pub fn execute<T>( &self, ctx: &ExecContext, dest: &mut ErasedRawStridedMut<'_>, src: &ErasedRawStridedRef<'_>, alpha: T, beta: T, op: UpdateOp, ) -> Result<()>

where T: Copy + MaybeSendSync + KernelStorageElement + Add<Output = T> + Mul<Output = T> + Zero + One + PartialEq + ElementOpApply,

dest = alpha * op(src) + beta * dest into an erased destination.

alpha, beta, and op are resolved once per call; the layout check and replay are the same allocation-free UpdatePlan path.

Examples

use strided_basic::{
    ErasedRawStridedMut, ErasedRawStridedRef, ErasedUpdatePlan, ExecContext, KernelDType,
    UpdateOp,
};

let plan = ErasedUpdatePlan::compile(KernelDType::F64, &[2], &[1], &[1]).unwrap();
let src_data = [1.0_f64, 2.0];
let mut dst_data = [10.0_f64, 20.0];
let src = ErasedRawStridedRef::from_slice(&src_data, &[2], &[1], 0).unwrap();
let mut dst = ErasedRawStridedMut::from_slice_mut(&mut dst_data, &[2], &[1], 0).unwrap();
// dst = 2 * src + 3 * dst
plan.execute(&ExecContext::serial(), &mut dst, &src, 2.0, 3.0, UpdateOp::Identity)
    .unwrap();
drop(dst);
assert_eq!(dst_data, [32.0, 64.0]);

pub fn execute_uninit<T>( &self, ctx: &ExecContext, dest: &mut ErasedRawStridedUninitMut<'_>, src: &ErasedRawStridedPtr<'_>, alpha: T, op: UpdateOp, ) -> Result<()>

where T: Copy + MaybeSendSync + KernelStorageElement + Mul<Output = T> + One + PartialEq + ElementOpApply,

dest = alpha * op(src) into an uninitialized erased destination.

This is execute with beta = 0: the destination is never read, so a NaN-filled or uninitialized output cannot contaminate the result. On success every reachable destination element is written; validation errors are returned before any write. A panic during execution may leave a partially initialized MaybeUninit destination, which is still safely droppable; no readable value is promised for unwritten reachable slots.

Examples

use core::mem::MaybeUninit;
use strided_basic::{
    ErasedRawStridedPtr, ErasedRawStridedRef, ErasedRawStridedUninitMut, ErasedUpdatePlan,
    ExecContext, KernelDType, UpdateOp,
};

let plan = ErasedUpdatePlan::compile(KernelDType::F64, &[2], &[1], &[1]).unwrap();
let src_data = [1.0_f64, 2.0];
let mut dst_data = [MaybeUninit::<f64>::uninit(); 2];
let src = ErasedRawStridedRef::from_slice(&src_data, &[2], &[1], 0).unwrap();
let src_ptr = ErasedRawStridedPtr::from_ref(&src);
let mut dst =
    ErasedRawStridedUninitMut::from_uninit_slice(&mut dst_data, &[2], &[1], 0).unwrap();
// dst = 2 * src
plan.execute_uninit(&ExecContext::serial(), &mut dst, &src_ptr, 2.0, UpdateOp::Identity)
    .unwrap();
let values: Vec<f64> = dst_data.iter().map(|v| unsafe { v.assume_init() }).collect();
assert_eq!(values, [2.0, 4.0]);

Errors

Returns DTypeMismatch or PlanLayoutMismatch when a descriptor does not match the plan, and OverlappingInputOutput when the source overlaps the destination allocation, all before any destination write.

Trait Implementations

impl Clone for ErasedUpdatePlan

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ErasedUpdatePlan

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

Formats the value using the given formatter.