FrozenGene commented on issue #4828: [QNN][TFLite] TFLite rounding mode support URL: https://github.com/apache/incubator-tvm/pull/4828#issuecomment-610853659 > @LiangHao151941 Was wondering if you ever got a chance to look at the performance of TFLite rounding? I am looking at the performance of UPWARD rounding and surprisingly Requantize was quite expensive in e2e models like mobilenet. I suspect TFLite rounding will be even slower. > > So, wondering if you got a chance to check performance. I worry that so much efforts in this PR might give us perfect functionality but not so good performance. And therefore, in future, we might decide to tradeoff performance with accuracy, taking us back to square 1. IMO, I think we should keep functionality first, then we should consider performance. Currently, our way is to lower many ops, I suspect this leads worse performance, because we can not control all computation in registers. I think TFLite rounding will not the bottleneck of performance. Let us put the effort to keep functionality firstly, next, we should reconsider our code and find the performance bottleneck leverage the performance tool (like DS5 Streamline).
---------------------------------------------------------------- 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: [email protected] With regards, Apache Git Services
