the thread's eight L (or D) words: column 8 nt + 2 (lane % 4) + b, through
`pointer`. Each load names SB5 as its read barrier and waits for the one
before it, so the pointer can be re-formed for the next sub-tile.
(Shared memory would serve these too, but SM86 LDS/STS are encoded with
the .X4 address scale: see the 2026-10-02 shared-attention fixture.)
2828def sksLoadRowTerms = (lambda unrestricted pointer : Nat . (lambda unrestricted tail : (family SM86Program) .
2829 (saFor 8 (lambda unrestricted i : Nat .
2830 (saOp (constructor SM86InstructionBody SM86LoadGlobal (saR (skTmp i)) (saR pointer)
2831 (saU (naturalAdd (naturalMultiply (naturalDivideUnchecked i 2) 32) (naturalMultiply (naturalModuloUnchecked i 2) 4)))
2832 (constructor SM86Control SM86ControlValue (byte 15) (constructor SM86YieldMode SM86Continue)
2833 saSB1 saSB5 (nat-to-byte saWait5) (byte 0)))))
2834 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.