[michalharakal]

Problem

DefaultCpuOpsBase.split built every output element through dataFactory.init { outIdx -> val inIdx = outIdx.copyOf(); …; tensor.data.get(*inIdx) }. That's two IntArray copies (the copyOf and the vararg spread) plus a boxed value per element. RoPE.applyRoPESplitHalfFull calls split every decode step, and once the Vector API fallback was gone (#1360) this was the top decode allocation site in a BitNet 2B4T JVM profile: about 7.8 GB in a 45 s window.

Change

A dense-FP32 fast path, using the same floatWindowOf / floatResult funnel as the other hot ops (plain and slab-backed data, scope-aware adoption):

• each chunk is outer contiguous runs of size × inner floats, copied with copyInto;
• other data keeps the generic path unchanged.

Tests

• New split_rank4_every_dim_matches_element_mapping: a 2×3×5×4 tensor split on every axis with a remainder chunk. Every element of every part is checked against the source mapping; it also checks negative dim.
• :skainet-backends:skainet-backend-cpu:jvmTest 491/491 pass; compileKotlinLinuxX64 passes.

Measured

Arduino Ventuno Q, 4× Cortex-A78C, JVM, BitNet 2B4T I2_S, 128 greedy tokens. The patch was backported onto 0.61.0 together with #1360 and the mmap bulk-copy PR:

• the Arrays.copyOf(int[]) ← split ← RoPE allocation site disappears from the JFR allocation-by-site list;
• greedy output is byte-identical (md5 52f57d1d…);
• decode goes 2.34 → 2.39 tok/s (≈ +2 %; on the JVM this mostly reduces allocation, not wall time).

Refs #1357

🤖 Generated with Claude Code