Source/Packages

Realization.Nvidia.SM86.GreedyArgmaxSM86

packages/realizations/cooperative/nvidia-sm86/src/Realization/Nvidia/SM86/GreedyArgmaxSM86.alpha

1,985 lines222 declarations73.3 KiBSHA-256 a4ae889c3756

def · lines 228–229

greedyArgmaxSM86SharedBytes

Full file
The block reduction's shared records: warp w's (value, token, invalid count) at word 12 w. LDS and STS as the encoder emits them count the address register in words (SM86.MachineModel.machineSharedAddress: byte 4 R + offset), so record w starts at byte 48 w and the eight span 384 bytes, not the 96 the records' own twelve bytes each would suggest -- the figure this was until 2026-10-12, which only an allocation granularity above it hid on the card.
228def greedyArgmaxSM86SharedBytes =
229  (naturalMultiply 4 (naturalMultiply 12 8))

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.