Is it really better to introduce VML as a temporary hack than it is to fix 
LLVM, even if the latter takes a little longer?

In Matlab, until quite recently X.^2 was slower than X.*X. It was that way for 
something like 25 years before they fixed it.

--Tim

On Monday, September 08, 2014 10:50:53 PM Jeff Waller wrote:
> I feel that x^2 must among the things that are convenient and fast; it's
> one of those language elements that people simply depend on.  If it's
> inconvenient, then people are going to experience that over and over.  If
> it's not fast enough people are going experience slowness over and over.
> 
> Like Simon says it's already a thing
> <https://github.com/JuliaLang/julia/issues/2741>.  Maybe use of something
> like VML is an option or necessary, or maybe extending inference.jl or
> maybe even it eventually might not be necessary
> <http://llvm.org/docs/Vectorizers.html>.
> 
> Julia is a fundamental improvement, but give Matlab and Octave their due,
> that syntax is great.  When I say essentially C, to me it means
> 
> Option A use built in infix:
> 
> X.^y
> 
> versus Option B go write a function
> 
> pow(x,y)
>    for i = 1 ...
>       for j = ....
>    ...
>    etc
> 
> Option A is just too good, it has to be supported.
> 
> What to do?  Is this a fundamental flaw?  No I don't think so.  Is this a
> one time only thing?  It feels like no, this is one of many things that
> will occasionally occur.  Is it possible to make this a hassle?  Like I
> think Tony is saying, can/should there be a "branch of immediate
> optimizations?"  Stuff that would eventually be done in a better way but
> needs to be completed now.  It's a branch so things can be collected and
> retracted more coherently.

Reply via email to