apache
[TOPI][Hexagon] Implement quantized avgpool
Thanks for contributing to TVM! Please refer to guideline https://tvm.apache.org/docs/contribute/ for useful information and tips. After the pull request is submitted, please request code reviews from Reviewers by @ them in the pull request thread.
cc @mehrdadh
Please let me know if there are any additional comments. If not, can someone approve and merge it for me please?
@jverma-quic : Could you please click the "Ready for re-review" link next to my name? I think that might be necessary to separate my old vs. new review comments.
This PR introduces several TOPI-related functions (qnn_avg_pool2d_compute and qnn_avg_pool2d_schedule). Does this PR make these functions available for compile-time consideration by TOPI?
I'm not very familiar with the mechanisms TVM uses for this, so apologies if I'm just missing how it happens.
This PR introduces several TOPI-related functions (
qnn_avg_pool2d_computeandqnn_avg_pool2d_schedule). Does this PR make these functions available for compile-time consideration by TOPI?I'm not very familiar with the mechanisms TVM uses for this, so apologies if I'm just missing how it happens.
That's correct. The PR does introduce several TOPI related functions. However, since they require inputs and outputs to be in 2d discontiguous buffer, they aren't yet available for use by OpStrategy. One of the issues here is that Relay is unable to handle complex layout needed for these discontiguous buffers and requires some additional work.
Thanks @jverma-quic , the PR is looking really good. I left one small suggestion, feel free to ignore it.
Once you're satisifed, I'm happy with getting this merged. You may want to ping @mehrdadh for that official review; I think he was waiting for my review to finish.
Thanks @cconvey! I really appreciate you taking time to review this PR and your detailed comments.
@cconvey, @mehrdadh, @kparzysz-quic, @TejashShah : I'm waiting to merge this PR. Unless there are additional comments, can someone approve and merge it for me please? Thanks!
@jverma-quic PR is merged! Thanks for your contribution! Moving forward please use a meaningful PR description.
@cconvey thanks for the review!