Replace `tilize`/`untilize` optional args in BMM to inferring from the input/output tensor format as `TILE` or `ROW_MAJOR` instead

tenstorrent / tt-metal

:metal: TT-NN operator library, and TT-Metalium low level kernel programming model.

Apache License 2.0

428 stars 57 forks source link

Replace `tilize`/`untilize` optional args in BMM to inferring from the input/output tensor format as `TILE` or `ROW_MAJOR` instead #965

Closed mywoodstock closed 8 months ago

mywoodstock commented 1 year ago

The optional tilize/untilize arguments input/output activations can be removed, and can directly infer whether to tilize/untilize based on the input/output activations tensor formats. This will simplify the code and remove extra compatibility checks.

davorchap commented 8 months ago

Inference of args is done at ttnn level.