Source/Packages

Realization.Nvidia.SM86.ScaledGradientUpdateSM86

packages/realizations/cooperative/nvidia-sm86/src/Realization/Nvidia/SM86/ScaledGradientUpdateSM86.alpha

116 lines7 declarations5.1 KiBSHA-256 9e5f93b54aae

def · lines 17–26

scaledUpdate64Control

Full file
Realizations of Learning.ScaledGradientUpdate's two independent operations. Every arithmetic instruction rounds separately to binary64, nearest-even. No contraction or conversion to binary32 is permitted. Values are supplied as register pairs by the enclosing kernel's layout; no model facts live here. This fragment reserves SB0 (write) and SB1 (read). Entry waits retire prior work on both; each issue replaces both reservations. The continuation must wait on mask 3 before consuming a result or overwriting a source. A caller using other barriers must first retire them. The scoreboard checks the complete enclosing program, including the continuation.
17def scaledUpdate64Control =
18  (constructor
19    SM86Control
20    SM86ControlValue
21    (byte 15)
22    (constructor SM86YieldMode SM86Continue)
23    (constructor SM86Barrier SM86Barrier0)
24    (constructor SM86Barrier SM86Barrier1)
25    (byte 3)
26    (byte 0))

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.