You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[BUG] Improper tiling implementation in PULPOpen for Slice and ReduceMean operators #134
The tiling system currently does not support list values as replacement variables, which affects the Slice and ReduceMean operators. These operators use arbitrary numbers of axes for input and output, which results in their data_in and data_out dimensions producing lists as replacement values.
To Reproduce
Run:
cd DeeployTest
python testRunner_tiled_siracusa.py -t Tests/testFloatDemoTinyViT/ --cores=8 --l1 4000
When checking the generated Network.c file, no tiling for Slice or ReduceMean operations
Expected behavior
Successful tiling for Slice and ReduceMean operations, that will be obtained by adjustings the templates and tiling constraints.
Additional context
Will need to also implement extra testing for Slice and ReduceMean operations.
Exactly, once PR #140 is finished, we can mark this as solved. I did not have time to work on this recently, but I will try to get back to it this week or next week at the latest.
Activity
Apply commit review and temporary fix of Slice and ReduceMean PULP co…
Demo TinyViT compatibility with tiled Siracusa (#124)
FP32 ReduceMean operator improvement (#137)
Xeratec commentedon Feb 3, 2026
@diaconuccalin If I remember correctly, this issue cannot yet be closed as you only implemented the tiling for the ReduceMean operator, correct?
diaconuccalin commentedon Feb 3, 2026
Exactly, once PR #140 is finished, we can mark this as solved. I did not have time to work on this recently, but I will try to get back to it this week or next week at the latest.
diaconuccalin commentedon Nov 4, 2026
Will be solved by PR #190