An odd key tile's 32 keys against the 32 queries below its diagonal lie
inside the dQ product's diagonal 64 x 64 tile but are never walked here:
every one is masked (P = +0), so the query pass's dS there is a signed
zero, and a zero product leaves the dQ accumulators' words unchanged (they
start at +0, which no sum of zeros turns into -0). Write +0 there, through
the offset the loop leaves one sub-tile below the diagonal.
2768def sksZeroBelowDiagonal = (lambda unrestricted seq : Nat . (lambda unrestricted tail : (family SM86Program) .
2769 (saAnd sksT2 saTileIndex saOne
2770 (saGreater saP0 sksT2 0
2771 (sfWide sksScoreGradientPointer sksScoreGradientOffset (saArgument streamingAttentionKeyScoreGradientArgument) saWait5
2772 (saFor 8 (lambda unrestricted i : Nat .
2773 (saWhen saP0 (sksScoreGradientWord seq (lambda unrestricted k : Nat . sm86RegisterZero) i)))
2774 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.