strided-fused

Runtime-DAG fused elementwise execution. The typed FusedPlan and dtype-erased ErasedFusedPlan implementations use strided-basic's shared execution policy and generic traversal without depending on strided-kernel.

use strided_basic::StridedArray;
use strided_fused::{fused_elementwise_into, FusedInst, FusedOp, FusedPlan};
let input = StridedArray::<f64>::from_parts(vec![2.0, 3.0], &[2], &[1], 0).unwrap();
let mut output = StridedArray::<f64>::col_major(&[2]);
let plan = FusedPlan {
    input_count: 1,
    outputs: vec![1],
    ops: vec![FusedInst { op: FusedOp::Negate, inputs: vec![0] }],
};
fused_elementwise_into(&mut [output.view_mut()], &[input.view()], &plan).unwrap();
assert_eq!(output.get(&[1]), -3.0);

simd is enabled by default; parallel enables bounded parallel replay and forwards to the shared foundation. This crate does not create an independent thread pool. An assembling backend can retain automatic fusion by calling this crate with its existing context and output storage; users need not select plans manually when their runtime already performs that planning.

See kernel boundaries and the packaged license notices. This new package has not yet been published.