silu: serve f16 from a table instead of widening every chunk to f32

[czoli1976]

The aarch64 f16 SiLU (arm64simd_silu_f16_4n) converts each chunk into an f32 scratch, runs the NEON f32 kernel over it and converts back, on every call — it carried a // TODO: Change this SiLU kernel once we have a native-FP16 one. Since SiLU is a unary f16 -> f16 map, the whole thing collapses into a 2^16-entry table: one load per element.

Why it is safe

The table is filled by calling the same f32 kernel the roundtrip path runs, over an aligned whole number of nr-blocks — which is exactly the slice the element-wise frame hands that kernel directly. f16 -> f32 is exact and the store rounds to nearest-even, so the table reproduces the roundtrip bit for bit rather than re-deriving SiLU from a formula. lut_matches_the_f32_roundtrip_on_every_f16 checks that over the full domain.

Calling the kernel raw rather than through the frame is also load-bearing, not a shortcut: map_slice_with_alignment holds a borrow_mut on the frame's thread-local scratch for the duration of a kernel call, so building the table through the boxed ElementWise panics with RefCell already borrowed the first time a model hits the op. Worth knowing before anyone adds a second table this way.

It is built lazily behind a OnceLock, so a model with no f16 SiLU never allocates the 128 KiB.

Numbers

Kernel, all three variants in one process so there is no cross-run drift:

elements	generic	f32 roundtrip	lut
1024	541 Melem/s	1.560 Gelem/s	3.670 Gelem/s
65536	537 Melem/s	1.574 Gelem/s	3.626 Gelem/s
1048576	536 Melem/s	1.528 Gelem/s	3.116 Gelem/s

2.0-2.35x over the path it replaces. End to end, f16 through a real plan:

block	roundtrip	lut
SwiGLU — down(silu(gate(x)) * up(x))	2.767 ms	2.604 ms	-5.9%
plain FFN — matmul -> silu -> matmul	1.683 ms	1.592 ms	-5.4%

The SwiGLU shape is the one that matters, since that is what Llama/Qwen/Mistral actually run — SiLU is the activation inside the gate, not the block itself, so the extra projection dilutes the kernel win. Arms were interleaved within one session; on this box a criterion baseline compared across processes drifts far enough to invert the sign of a result this size.

x86 is untouched — it has its own x86_64_avx512_silu_f16_16n.

Tests

tract-linalg 4415, tract-core 270, test-f16 2378, test-unit-core 816 — green. cargo fmt --all clean; cargo clippy clean (the one warning in benches/silu.rs is pre-existing, in the scalar reference helper).

🍍