Source/Packages

Accelerator.SM86.Operands

packages/hardware/architectures/nvidia-sm86/src/Accelerator/SM86/Operands.alpha

645 lines51 declarations53.6 KiBSHA-256 e735be07682e

def · lines 414–414

sm86FixedLatencyCycles

Full file
ASSUMED, and stated as an assumption: the cycles after issue before a fixed-latency form's destination (register or predicate) may be read -- 6 for every such form, above the 4 published for Ampere's FP32 and integer pipes. Variable-latency forms (loads, S2R, SHFL, MUFU, HMMA) deliver through a scoreboard barrier instead, and forms with no destination deliver nothing (0). A program accepted under this table that a card computes wrongly refutes it: silicon runs of stall-compacted programs are its evidence.
414def sm86FixedLatencyCycles : Nat = 6

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.