For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
Mojo module
pointer
Defines atoms for plain pointer loads and stores.
PointerLoad and PointerStore wrap one vector access through a pointer:
width contiguous elements, such as ld.global.v4.f32 or
st.shared.v2.b32 on NVIDIA GPUs, global_load_dwordx4 or ds_write_b128
on AMD GPUs, and a vector load or store on a CPU. The access claims align
bytes of alignment, by default the whole vector's; a caller whose addresses
are only element aligned claims less, and the backend then picks accesses
that need no more. A load is never invariant: an immutable view only says
the access reads, and other agents may write the memory. Each thread issues its own access at the base of the memory
view it passes, so the memory map fixes the value bits and leaves lanes
free. The registers are ready, and the memory written, when the call
returns.
You can import these APIs from the layout.atoms.pointer module:
from layout.atoms.pointer import PointerLoad, PointerStoreStructsโ
- โ
PointerLoad: A vector load ofwidthcontiguous elements into registers. - โ
PointerStore: A vector store ofwidthregisters into contiguous elements.