Vectorize output store in ublkcp DeviceTransform kernel#9481

Open

nanan-nvidia wants to merge 13 commits into

NVIDIA:mainfrom

nanan-nvidia:device_transform_align_16

Commits on Jun 23, 2026

cub: add tunable store_vec to ublkcp vectorized transform store
nanan-nvidia
committed
cub: test ublkcp vectorized transform store on generic counts (narrowing/widening)
nanan-nvidia
committed
cub: store_vec=1 disables ublkcp vectorized store; cap to 16B
nanan-nvidia
committed
cub: guard host vectorize check on contiguous output (fix c.parallel type-erased iterator)
nanan-nvidia
committed
cub: avoid default-constructing the ublkcp store buffer (support non-default-constructible outputs)
nanan-nvidia
committed
cub: gate host vectorize check on power-of-two element sizes (fix uchar3 is_aligned assert)
nanan-nvidia
committed
cub: update copyright year on vectorized transform test
nanan-nvidia
committed
cub: gate ublkcp store_vec on power-of-two width; add non-pow2 store_vec regression test
nanan-nvidia
committed
cub: clang-format vectorized transform store changes
nanan-nvidia
committed