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
tileops
Defines TileOps: tile-level operations built from layout.atoms.
A TileOp moves or computes a whole tile by issuing atoms. It derives the
memory tile's placement, a BitLayout from logical coordinate bits to
element offsets, and matches it against the atom's operand maps, converted
from the operand's address encoding to the engine's element offsets: every
index bit the atom fixes on the memory side must be played by one logical
bit that reaches the same offset.
The logical bits played by the atom's lane and warp bits say which thread
holds which elements. The remaining logical bits, in order, are this
thread's register tile, whose placement must put the atom's value bits on
the registers its register map names. Logical bits no index bit plays are
issued over: the TileOp issues the atom once per value of them. Each issue
passes the memory at those bits and this thread's warp; the instruction
adds the lane and value parts, as tcgen05.ld does. An atom without lane
or warp bits splits nothing, and every thread of the caller moves its own
tile.
A TileOp never waits: only a building block does. load and store issue
their atoms and return; the building block that called them waits with the
atom's completion, such as TMemLoadWait.wait(), when it needs the
registers or the memory, so that other work can run while the issues are in
flight.
You can import these APIs from the layout.tileops package:
from layout.tileops import can_load, can_store, load, storeFunctions
-
can_load: Returns whetherloadcan move a memory tile of layoutLMin engineEMinto register tiles of layoutLRwith atomA. -
can_store: Returns whetherstorecan move register tiles of layoutLRinto a memory tile of layoutLMin engineEMwith atomA. -
load: Loads a memory tile into the issuing threads' register tiles withatom. -
store: Stores the issuing threads' register tiles into a memory tile withatom.