[ydl-gen] Java and SPUs

[Moore, Randy]

Ian,
 
Great Article! They seem to address  both the PPE and SPE more than I had been hoping. Being able to specify which threads run on the SPE and which run on the PPE would seem to be a must have. You don't want the SPEs running a bunch of code with conditionals in it.  As I get smarter on Cell programming, it seems that if one really wants to get every ounce of power out of the Cell, custom coding in C/C++ or assembly is the way to go.  Interpreting byte codes in the JVM would probably have to be replaced with native compilation, and Cell specific optimizations would be needed to take advantage of instruction and data pipelining, shuffling, etc. 
 
A second look at JNI may be in order so that I can maximize the use of the SPE's to do what they do best. To maintain portability of the main java application, JNI code could need to be sensitive to the architecture that it was running on so that it could invoke mathematically intensive code on SPEs on the Cell, or normal C/C++ optimized code on lets say a PPC or X86 machine.
 
Thanks Again,
 
Randy
 
>>Hi Randy,

>>I think this may interest you:
http://www.ics.uci.edu/~franz/Site/pubs-pdf/ICS-TR-06-17.pdf
>>This system is built from the interpreter only JamVM but similar systems
>>should be possible with cluster/distributed JVMs such as dJVM and JESSICA.

>>Regards,
>>Ian


-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/ms-tnef
Size: 4694 bytes
Desc: not available
Url : http://lists.terrasoftsolutions.com/pipermail/yellowdog-general/attachments/20070510/e803950e/attachment-0001.bin