the candidates, as scripts/alpha/tiled-choices.py numbers them (block M x
block N x stages, warps along M x N, each warp's rows x columns):
0 128 x 128 x 4 2 x 4 64 x 32
1 128 x 128 x 2 2 x 4 64 x 32
2 64 x 64 x 4 2 x 2 32 x 32
3 64 x 64 x 2 2 x 2 32 x 32
4 128 x 64 x 4 4 x 2 32 x 32 (48 KiB: two blocks an SM)
5 64 x 128 x 4 2 x 4 32 x 32
6 256 x 128 x 2 4 x 2 64 x 64 (the most registers the SM86 file allows)
7 128 x 256 x 2 2 x 4 64 x 64
Candidates over a card's shared memory or registers exist but never serve
(coppeliusTiledTileServes). The cuBLAS tiles measured on the GB10
(research gemm/cublas-shapes.md: 128 x 96 .. 256 x 128, three stages) are
where 4 .. 7 came from.
529def coppeliusTiledCandidateCount : Nat = 8The compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.