search

tenstorrent / tt-metal

:metal: TT-NN operator library, and TT-Metalium low level kernel programming model.
Apache License 2.0
421 stars 54 forks source link

Shallow conv issues #4914

Closed tt-nshanker closed 1 month ago

tt-nshanker commented 8 months ago
  1. Padding of channel (< 8) to 8 for 16 B alignment.
  2. If C * window_w % 32 != 0, add support to offset by padding of 32 when writing to activation block. We already have support to pad the weights inner dim with 0s for this scenario. Activations can have garbage in inner dim.
  3. Re-work split reader implementation to read activations in parallel into the same CB. This will prevent this hard constraint - (act_block_h / out_subblock_h) % 2 == 0. Currently, activations are split at the subblock granularity and activations are split into 2 CBs. The compute waits and pops from 2 cbs which is not necessary.
tt-nshanker commented 7 months ago

image