---- several quads a thread in flight (side warps) ----
`wide` quads a thread an iteration: T, T + S, .., T + (wide - 1) S (S the
threads), every bank's loads issued before any update, so a thread keeps
`wide` times the bytes in flight -- what a side warp, a few to an SM,
needs to stream at the memory's rate. Each element's update is ahUpdate's
instructions (ahWideUpdate), its quad's stores ahStores'. Registers: quad
q's banks at 32 + 16 q + 4 b + e, the temporaries after the last quad's
(shared: the updates run one after another), each quad's half pairs after
those; quad q's guard P0 (q 0) or P(q + 1) (P1 counts the loop). At most
four quads (P2..P4).
227def ahwValue = (lambda unrestricted q : Nat . (lambda unrestricted e : Nat . (lambda unrestricted b : Nat .
228 (naturalAdd 32 (naturalAdd (naturalMultiply 16 q) (naturalAdd (naturalMultiply 4 b) e))))))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.