Crate strided_perm 

Cache-efficient tensor permutation / transpose.

This crate provides optimized copy and permutation operations for strided multidimensional arrays. It is designed as a single-responsibility crate sitting between strided-view (data structures) and strided-kernel (general map/reduce/broadcast operations).

Dependency graph

strided-view -> strided-perm -> strided-kernel

Threading

Every copy takes a threadteam::Exec placement token as its last parameter and splits the permutation within that budget. The width is decided by the token: the serial time of the copy (copy_ns_per_element times its elements) against the dispatch cost of one section (Exec::width_for), so a copy too small to repay a fan-out runs on the calling thread whatever the budget. The crate never reaches a global pool; pass Exec::serial() for a copy on the calling thread.

Structs

BilateralFusionPlan

Validated result of fusing adjacent dimensions that are contiguous in both source and destination layouts.

Enums

FusionPlanError

Failure while validating or constructing a bilateral fusion plan.

HostClass

The class of the machine, keyed by the kernel dispatch it will take.

Constants

BLOCK_MEMORY_SIZE

CACHE_LINE_SIZE

Functions

copy_into

Copy elements from source to destination: dest[i] = src[i].

copy_into_col_major

Copy elements to a col-major destination.

copy_into_uninit

Copy elements into an uninitialized destination: dest[i] = src[i].

copy_ns_per_element

Serial throughput of a copy in nanoseconds per element on this host (copy_ns_per_element_of of HostClass::detect).

copy_ns_per_element_of

Serial throughput of a copy in nanoseconds per element, of a host class.

plan_bilateral_fusion

Plan bilateral dimension fusion for source and destination layouts.