Interesting but the examples seem much simpler than real algoritms

Id like to see a compiler work out the  UTF8 SIMD encoding the link to
which  i pasted earlier . I bet it will beat a  new std lib code by a
factor of 10 .

I think the key point from the paper is not it can be auto optomized (i
have my doubts except for basic add / multiply and copy /cmp ) but that
 when we are talking about 5-15% difference to C with GC etc its entireyly
pointless - you want fast you use SIMD.. If the language cant do that it
will not be useful especially with 512 bit SIMD around the corner. Make it
SIMD friendly and eaiser for programmers and you will somethign that will
really attract  c and C++ devs even with a 20% GC cost.  Maybe the language
should force /assist greater vectorization ,  structures of arrays etc
which then allow a better compiler.

The array of structures reminds me of some  matrix  op cach hit
optomizations , interesting it should be applied further.

Ben


On Thu, Aug 1, 2013 at 2:41 PM, David Jeske <[email protected]> wrote:

> This is a pretty interesting look at automatic compiler
> parallelization.... Through mostly-automatic technique they closed the
> performance gap between serial and hand-coded parallel implementations from
> 50x to 1.2x.
>
>
> http://www.intel.com/content/dam/www/public/us/en/documents/technology-briefs/intel-labs-closing-ninja-gap-paper.pdf
>
> A technique worth noting in here is a strategy where they block
> data-structures to fit in the cache by converting an array-of-structures to
> a structure-of-arrays. This reminds me of column-oriented databases, which
> also reminds me research showing merely using column-oriented local
> block-layouts provides most of the benefits of column-oriented layout
> without many of the drawbacks..
>
> http://www.vldb.org/conf/2002/S12P03.pdf
>
> _______________________________________________
> bitc-dev mailing list
> [email protected]
> http://www.coyotos.org/mailman/listinfo/bitc-dev
>
>
_______________________________________________
bitc-dev mailing list
[email protected]
http://www.coyotos.org/mailman/listinfo/bitc-dev

Reply via email to