Lunderberg commented on issue #13508: URL: https://github.com/apache/tvm/issues/13508#issuecomment-1332722603
Testing on the GPU, with both cuda and vulkan backends (nvidia-driver-470 on ubuntu 21.04), it shows a pretty similar effect. It isn't quite as dramatic, only 50x slower instead of 1000x, but it's still quite a large effect. Both GPU tests were done with the same compute definition, but with `topi.cuda.injective.schedule_injective` instead of `topi.x86.injective.schedule_injective`  The specific fix here wasn't on the transformation side, but a change to the topi operator. The nice thing is that it can be a lot more general, and can convert floating point numbers to integer ratios (e.g. the `-0.75` in the example into `Fraction(-3, 4)`) before they get too folded to be recognized. The downside is that it isn't as general of a solution. -- 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. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
