Jevin Jiang 9e5edb7015 [Mosaic TPU] Support packed type matmul with arbitrary shapes. ...

This cl removes all the shape constrains in matmul for all types.

We only need to mask out subelement on contracting dim. Instead of unpacking data and applying masks, we create a VREG-sized i32 "mask" which contains subelement mask info to logical and with target vreg. Through this way, in order to mask sub-elements, each target vreg only needs to apply 1 op (logical_and) instead of 3 ops (unpacking + select + packing).

PiperOrigin-RevId: 702480077

2024-12-03 14:58:42 -08:00

..

dialect

[Mosaic TPU] Support packed type matmul with arbitrary shapes.

2024-12-03 14:58:42 -08:00

gpu

[mosaic_gpu] The profiler now uses FFI calls for creating events and computing elapsed time

2024-11-12 11:01:59 -08:00

python

[MOSAIC:GPU] Add async_load, async_store, and supporting attributes to the MLIR Mosaic GPU Dialect.

2024-11-08 14:34:23 -08:00

BUILD

[Mosaic] Add target core type parameter to tpu.sem_signal

2024-11-19 13:40:13 -08:00