strided-rs v0.4.7

strided-rs v0.4.7

Performance

• Permutation engine (#292).
  The parallel 2D transpose splits the whole row x column tile grid instead of
  only the row tiles, so a near-vector-shaped transpose (one extent below the
  micro-kernel block) uses every worker. The grouped high-rank tile target is
  widened to two micro-kernel blocks so the destination-contiguous inner run
  spans several cache lines, and the group length is bounded so one large axis
  cannot absorb the permutation and drop its outer parallel loop. A rootless
  ConstStride1 plan (one strided stream) is split across workers.
• Uninitialized full-overwrite copies. copy_into_uninit routes
  non-contiguous f32/f64 copies through the permutation engine, serial or
  parallel, under the active execution policy; equal contiguous layouts and
  every other dtype keep the element map. copy_into_par proves the destination
  offsets are separated before splitting tiles and otherwise keeps the serial
  engine, and a bounded policy below the pool size uses the policy-aware element
  map.
• Axis reductions. The output-block axis-reduction kernel folds each
  reduced step into the adjacent-output accumulator block with a SIMD vector
  operation for the sum, product, and sum-of-squares families. Results stay
  bitwise identical to the scalar fold.

Measured on the tenferro-rs CPU benchmark cases (four threads, AMD EPYC, pinned
to four CPUs): permutation materialization improves by up to about 1.9x
(reverse_23d_2 28.0 -> 16.6 ms, transpose_3d_256_201 37.5 -> 24.5 ms) and
axis reduce_sum / reduce_prod by about 1.25x and 1.28x at four threads.

Compatibility

No public API changes; the release is source compatible with 0.4.6.

Full comparison: v0.4.6...v0.4.7