S = Q K^T and O += P V as the register kernel's, the first product waiting
for the fragments (SB1, which also ordered Q's loads)
The first slice reads RZ for its accumulator, so no tile zeroes the
scores first: +0 + the products is the same word either way.
671def sfScores = (lambda unrestricted tail : (family SM86Program) .
672 (saFor 4 (lambda unrestricted k : Nat .
673 (saFor 8 (lambda unrestricted nt : Nat .
674 (saHmmaFrom (saS nt) (saQ k) (saB k nt)
675 (naturalSelect (naturalIsZero k) sm86RegisterZero (saS nt)) saSB2
676 (naturalSelect (naturalIsZero nt) (naturalSelect (naturalIsZero k) saWait1 saWait2) saWaitNone)))))
677 tail))The compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.