=== Transformer layer pipeline: unimem vs standard ===
dim=512 heads=8 seq=128 ffn=1376
>>> STANDARD (Vec + separate allocs) <<<
alloc: 578.333µs
CPU rmsnorm: 19.708µs
AMX qkv: 260.083µs
CPU rope: 15.5µs
AMX attn: 12.875µs
CPU softmax: 6.875µs
AMX attn@V: 8.458µs
AMX ffn up: 309.292µs
CPU silu: 78.375µs
AMX ffn down: 129.084µs
ANE compile: 36.710666ms
ANE run: 241.584µs
TOTAL: 41.40875ms
pass: true
>>> CYB-MEM (tape, zero-copy) <<<
alloc (tape): 935.833µs (14.7 MB used / 15.7 MB cap)
CPU rmsnorm: 11.625µs
AMX qkv: 250.083µs
CPU rope: 13.417µs
AMX attn: 4.917µs
CPU softmax: 8.208µs
AMX attn@V: 5.375µs
AMX ffn up: 269.292µs
CPU silu: 94.375µs
AMX ffn down: 132.209µs
ANE compile: 18.081ms
ANE run: 212.083µs
TOTAL: 21.01775ms
pass: true
=== Comparison ===
Stage Standard unimem Speedup
----- -------- ------- -------
alloc + fill 578.3µs 935.8µs 0.6x
CPU: rmsnorm 19.7µs 11.6µs 1.7x
AMX: qkv projection 260.1µs 250.1µs 1.0x
CPU: rope 15.5µs 13.4µs 1.2x
AMX: attention matmul 12.9µs 4.9µs 2.6x
CPU: softmax 6.9µs 8.2µs 0.8x
AMX: attn @ V 8.5µs 5.4µs 1.6x
AMX: ffn up+gate 309.3µs 269.3µs 1.1x
CPU: silu 78.4µs 94.4µs 0.8x
AMX: ffn down 129.1µs 132.2µs 1.0x
ANE: compile+load 36.7ms 18.1ms 2.0x
ANE: matmul run 241.6µs 212.1µs 1.1x
TOTAL 41.4ms 21.0ms 2.0x
reference verified: true
unimem verified: true