Module perm 

Cache-efficient tensor permutation and transpose routines.

Modules

block

Block size computation ported from Strided.jl

copy

Copy/permutation operations on strided views.

fuse

Dimension fusion logic ported from Strided.jl/src/mapreduce.jl

hptt

HPTT-faithful cache-efficient tensor permutation.

kernel

Block-based iteration engine for strided permutation operations.

order

Loop ordering algorithm ported from Strided.jl

Structs

KernelPlan

Constants

BLOCK_MEMORY_SIZE

CACHE_LINE_SIZE

SMALL_TENSOR_THRESHOLD

Maximum total elements for the small tensor fast path.

Functions

build_plan_fused

Build an execution plan with dimension fusion.

build_plan_fused_small

Simplified plan for small tensors that fit in L1 cache.

compress_dims

Remove size-1 dimensions from fused dims and all corresponding strides.

compute_costs

Compute the minimum stride cost for each dimension.

compute_importance

Compute the “importance” of each dimension for loop ordering.

compute_order

Compute the optimal iteration order for dimensions.

copy_into

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

copy_into_col_major

Copy elements to a col-major destination.

for_each_inner_block_preordered

Iterate over blocks with pre-ordered dimensions and initial offsets.

fuse_dims

Fuse contiguous dimensions across multiple arrays.

sort_by_importance

Get the permutation that sorts by importance (descending).

total_len

Utility: total number of elements.

try_fuse_group

Try to fuse a contiguous dimension group into a single (total_size, innermost_stride).