k64_kernel

Function k64_kernel 

Source
pub fn k64_kernel(gpu: &dyn GpuBackend) -> Result<KernelHandle>
Expand description

Resolve the k64 deep-K tile GEMM, preferring the 3-deep weight-pipeline variant. ON by default; ATLAS_NO_K64_PIPELINE3 (presence — =0 is NOT “off”) falls back to the 2-stage parent.

The parent issues one cp.async group then wait_alls it before the dequant phase, so with a small grid there are ZERO outstanding loads across that phase. The out_proj/o_proj shapes (N=5120, K=6144) launch 40 CTAs on 48 SMs — exactly 1 CTA/SM — so nothing covers the drain, and they measure ~38% of achievable while lm_head (1938 CTAs) reaches 83% on the identical loop. _p3 keeps step i+2’s loads in flight across dequant(i+1). Bit-identical.