no argument against it, except who's going to do e2e local math optimizer and how long it will take. maybe, both is possible -- faster local gains we know current algorithms need, and longer term overhaul that included e2e in-core plans.
On Mon, Jun 9, 2014 at 4:16 PM, Ted Dunning <[email protected]> wrote: > On Mon, Jun 9, 2014 at 4:12 PM, Dmitriy Lyubimov <[email protected]> > wrote: > > > One open question i had was,well, since the outlined approach avoids > > logical plans still for in-core matrix, it would be difficult to optimize > > the expressions in this context (i.e. figure out best-working _output_ > > structure). > > > > This is the single biggest flaw that I see. To do higher level > optimizations, the lower levels have to expose costs for possible plans. > > Taking a major whack at the local cases for element-wise and matrix > multiply is still a good thing to do. >
