Topic

SIMD kernel optimization

Hand-tuning compute kernels to use a processor's vector instructions, registers and pipelines, usually scored against the chip's theoretical peak floating-point rate.

Current clusters

build1 publisher

One fmla in eleven instructions keeps a NEON matmul at 7% of the M2's 112 GFlops

A dev.to post takes a single-threaded f32 matmul from 2.09 to 7.86 GFlops in hand-written NEON on a MacBook Air M2. The inner loop runs eleven instructions for every one multiply.

Publishers:dev.to

Reality

Evidence60
Adoption: Insufficient
Hype gap−20
Incentives18
Confidence65