Actually, the updated optimizer is now even capabile of (not in all
situations, but reasonably some) detecting and fusing unary elementwise
operations in a continuous expression like

5 + exp(drmX * drmX)

and convert it into plan that does x_i,j = 5+ exp(x_i,j*x_i,j) element-wise
in one physical step.

On Thu, Dec 11, 2014 at 3:08 PM, Dmitriy Lyubimov <[email protected]> wrote:
>
> Is there any expectation about release time frame?
>
> I am working (and i am hoping to reasonable resolution) to push a new
> installment of small optimizer enhancements for distributed algebra.
>
> it would help me to know if there's a sense of release time frame, it'd
> give me an idea what to shoot for.
>
>
> I want to offer few small things that we mostly already discussed.
>
> (1) optimizer techniques to detect unary elementwise operations such as
>
> drmX * drmX
> dexp(drmX), mexp(mxX), vexp(vec)  (in R simply exp(x))
> mxX ::= abs _ ( i.e sparse/dense elementwise evals)
>
> (2) better binds like 1 cbind drmX, drmX cbind 1 (enhancement relevant for
> "playing with shell" things)
>
> (3) smaller things like non-int keyed matrices to int-keyed matrix
> converstion with old 2 new key extraction
>
> (4) various bug fixes
>
> (5) whatever i forgot
>
> Nothing earth shuttering but even small things are difficult (but not
> unprecedented) to contribute here.
>
> -d
>
>
>
>

Reply via email to