mlcommons
[API] Enable access to data buffer
Hi,
For microcontrollers with limited memory, we can optimize memory usage by directly passing gp_buff to inference machine instead of copying the data to a separate memory.
For some microcontrollers this is a huge amount of memory and it's a matter of whether the model/runtime fits on device or not. I'll be happy to change the implementation if there is any suggestion.
Thanks.
MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅
![github-actions[bot] avatar](https://avatars.githubusercontent.com/in/15368?v=4 "Published by a pub.dev verified publisher"){kind=small}
This looks like a great addition!
My suggestion is that we add this change to the next version of the benchmark and keep v0.5 frozen for posterity.
We should be starting on a new version soon so I'll leave this PR up for now.
@colbybanbury that sounds great. Thanks! Also I have signed the CLA google form and I think my github account will be added soon.
Hi @colbybanbury, We could potentially add a tag to the current commit and label it as v0.5 and move forward with the development on main branch. wdyt? We can also discuss it in the next meeting.
Thanks.
Friendly reminder about this PR. We're using this change in our future submission and I appreciate if we can decide if we could merge this to the API. thanks!