AlphaGitHub
Source/Packages

Realization.Nvidia.SM86.StreamingAttentionSM86

packages/realizations/cooperative/nvidia-sm86/src/Realization/Nvidia/SM86/StreamingAttentionSM86.alpha

1,203 lines221 declarations70.9 KiBSHA-256 23d4e5e2aa2a
Download exact fileView at revision0dc6acfc09dcaa39d8

def · lines 551–551

sqOutputGradientLoadPointer

Full file
Q loads may still use R14/R15 when the output-gradient address is formed. Give dO its own pair; otherwise some row terms read the wrong operand and vary across identical physical runs. The dO loads finish before the row- term stores, which can reuse this pair. Both finish before the dP loop assigns these words, keeping register demand below 256.
551def sqOutputGradientLoadPointer : Nat = 240

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.