Cascade is a specialized data transfer between horizontally adjacent tiles in a specific direction depending on the row. For first generation AI Engines, cascades flow in the following way:
- Even rows - East to West
- Odd rows - West to East
Note: The first row starts at index 0 and is the shim row. Therefore, the first regular tile row is actually an odd row at row index 1 and flows West to East.
Note: The last tile in each row cascades into the tile directly North of it which makes the full cascade chain snake from bottom to top.
As far as mlir-aie syntax support for cascade connections, we support cascade functionality by simply placing tiles directly horizontally adjacent to one another. No additional operators are needed in mlir-aie. There are core ops that push and pull data from the cascade ports
AIE.get_cascade()
AIE.put_cascade($cascadeValue : type($cascadeValue))
This allows us to push and pull data off the cascade ports.
NOTE: This feature is currently not working so we rely instead on external compiled kernel functions to push and pull cascade data in our example design
-
Read through the aie.mlir design. Based on the tile locations, which tile is pushing cascade data and which tile is pulling it?
-
Verify correct functionality by compiling the design and then simulating the design via make.
make; make -C aie.mlir.prj/sim -
Move the design to an even row and make the necessary code changes. Verify your design works in simulation.
- Move the design one more time to the end of the row where the two tiles are in different rows and verify correct behavior in simulation.