The mode FAST_COMPILE should include the stability optimisation
Hi,
I think that all mode(including FAST_COMPILE that don't do it now) should return the same numerical result, but could take more or less time. Currently FAST_COMPILE don't include the stability optimization. We need to think about it, because some of them could depend on other optimization before they can be applied. This mean that we should also put all the dependent optimization in FAST_COMPILE.
I put it major now, but this can wait after the deep teaching as we don't regulary use FAST_COMPILE.
I think that all mode(including FAST_COMPILE that don't do it now) should return the same numerical result, but could take more or less time. Currently FAST_COMPILE don't include the stability optimization. We need to think about it, because some of them could depend on other optimization before they can be applied. This mean that we should also put all the dependent optimization in FAST_COMPILE.
I put it major now, but this can wait after the deep teaching as we don't regulary use FAST_COMPILE.
Leave a comment
We should inplace all inplace that work in linear time like dimshuffle(to have transpose in place) and maybe subtensor.
To have those opt in linear time, we will maybe need to modify them to don't always do a validate after each modification to don't call toposort that is slow and not linear.
To have those opt in linear time, we will maybe need to modify them to don't always do a validate after each modification to don't call toposort that is slow and not linear.