FrozenGene edited a comment on issue #4828: [QNN][TFLite] TFLite rounding mode support URL: https://github.com/apache/incubator-tvm/pull/4828#issuecomment-583218692 > One request that I have is - While setting the rounding mode to TFLite in TFLite parser, it might be better to set it by adding an extra arg in `from_tflite` API. The main reason is that the lowering here might involve many more number of Relay ops, which might degrade performance. So, to enable TVM user to choose different accuracy-performance tradeoff points, we can add a new argument in the TFLite parser API, and then pass it along to the QNN ops. The default value of this arg can be tflite rounding. @FrozenGene Are you ok with this? Yes, this is my thought too. I am ok with it. > LGTM. I like the scope of this PR. IMO, we can get this in. And send other PRs for TFLite parser. IMO, I would prefer adding TFLite parser in this PR too. Because this will be organized into one complete solution for TFLite rounding. When others review the history, they don't need to check isolated PRs together to get complete support for TFLite rounding support. Maybe these PRs during time is long. It is right that we could make small prs and merge it in quickly so that we could get functionality supporting quickly and could locate the reason quickly. However, from this pr's perspective, TFLite's parser code will not be complex and could be contained in. It is acceptable in one single PR from my opinion. So I would suggest @LiangHao151941 adding TFLite parser in this PR too. If we think it is import, we even could port it back to release 0.6, one PR will make us do it easily too.
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
