Re: GPU and "memory is not contiguous" - you need to take a look at 
e.g.http://www.anandtech.com/show/7677/amd-kaveri-review-a8-7600-a10-7850k/6

"GPU can now access the entire CPU address space without any copies"

Heterogeneous System Architecture (HSA) is already here. The h/w and 
libraries are available for e.g. C++. Java should be able to leverage this 
directly in Java9 next year via Sumatra/Graal/Okra. I would like to make 
sure that Clojure hits the ground running with this technology and has a 
working integration from the moment that Java9 is delivered. I think that 
it is a game-changer. We'll just have to wait and see....

Jules


On Monday, 17 March 2014 12:15:37 UTC, Jean Niklas L'orange wrote:
>
> Hello,
>
> On Saturday, March 15, 2014 11:37:11 PM UTC+1, Jules wrote:
>>
>> 2. I've had a look at rrb-vector - very fast for catvec - agreed, but 
>> there does appear to be a performance penalty in terms of conj[!]-ing - I 
>> can post my results if you  are interested. I read the Bagwell paper on 
>> which your work was based and came away with the impression that he did not 
>> think he was introducing significant overhead with his improvements for 
>> constant tiime concatenation - so maybe this is due to rrb being impled in 
>> clojure and builting vec in java ? 
>>
>
> I've not looked at the implementation done by MichaƂ in detail (lack of 
> documentation makes it a tad hard to grok), but a properly implemented 
> variant should have about the same performance as a normal vector when no 
> concatenation has been done on the vector, and would only require one extra 
> check. Transients which aren't concatenated should have same performance 
> guarantees as transient Clojure vectors, but only if the node size is fixed 
> to 32 elements.
>
> However, I've found it rather hard to perform the general conj-case 
> efficiently, so you're right that conjing will be slower on average in an 
> RRB-vector after concatenations. I'm trying to find ways to cut off large 
> parts of the constant factor off, and hope to get some results dished out 
> in some months time.
>
> Another factor here is that the Clojure impl *may* suffer from suboptimal 
> JVM-emitted code, although that is purely speculation from my side.
>
> As for 3, I've not looked into GPU programming of these things as the 
> memory is not contiguous and would probably be messy to transfer over.
>
> -- JN
>

-- 
You received this message because you are subscribed to the Google
Groups "Clojure" group.
To post to this group, send email to clojure@googlegroups.com
Note that posts from new members are moderated - please be patient with your 
first post.
To unsubscribe from this group, send email to
clojure+unsubscr...@googlegroups.com
For more options, visit this group at
http://groups.google.com/group/clojure?hl=en
--- 
You received this message because you are subscribed to the Google Groups 
"Clojure" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to clojure+unsubscr...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Reply via email to