I've been interested in Mono's direct support for SIMD operations. As far as I can tell, it does make a difference in performance. Even things like NumPy (for Python) effectively use SIMD under the covers, if I understand correctly. But, apparently, we don't need support for SIMD in Java, because HotSpot will (at some magical point) make all the pain go away.
If you can tell, I'm disappointed in the Sun response to the problem. And how hard would it be to recognize when low-level 'for' loops are parallelizable with SIMD operations? My guess is that it would be very difficult to recognize automatically compared with the ease of writing code in parallel form to begin with.
I'm rather convinced that basic, and maybe some additional advanced, matrix operations ought to be in the core JRE. After some time in MATLAB, I'm convinced it's a better level of abstraction than 'for' loop everywhere, anyway. And I like pure Java. If it requires native code outside the JRE, it's uncool. I'm also tired of having no clear path for matrix support in Java.
If scientific work (or other number crunching such as for games) takes people to MATLAB or Python or Mono or C, that leaves a huge hole for Java.
Maybe someday they'll care about this hole. Really, a matrix absraction layer allows for easier, high level coding and easier optimization. Pretty please, Sun or IBM or whatever?
Bonus points for including GPGPU implementations and abstractions.
This comment has been removed by the author.
ReplyDeleteIt's indeed a shame not to have high performance math availlabe with Java.
ReplyDeleteThey could provide 16 aligned bytes types and SIMD functions, that would be compiled by the JVM either to SIMD instructions if available, and to standard ones otherwise...