Ppl-accel mailing list

["Dokania, Harshit" <hdokani2 AT illinois.edu>]

Hello Dave,

I am reading your phd thesis. I have few questions regarding accel entry 
method for GPGPUS.

CkIndex_xxxx class  which is used as a hook for the runtime system to call 
into the application code for entry methods.

In case of an AEM this function is actually broken up into two stages, a 
general stage and a device-specific stage.

I looked into the def files but couldn't find cuda device specific functions. 
There were general functions for accel entry methods.

I have a charm build of cuda, which currently performs offload operations and 
kernel executions through GPU Manager.

It would be nice if you can guide where to look.
 
Also where can I find code for accel manager decision making for AEM 
executions on host or device which is done on runtime.

I have more questions about when AEM's are batched then how are kernel thread 
index mapped for different kernel launches or AEM.  I read that it is one to 
one mapping for AEMs in batch but in that case if a cuda kernel launch has 
large number of threads <<<grid dim,block dim>>> how is that. Are all the 
AEM's in batched set AEM are executed concurrently on device. 

Note: Currently the charm build of cuda only takes GPU Manager into account 
and calls these methods initHybridAPI(), gpuProgressFn(), and ExitHybridAPI() 
from src/conv-core/convcore.c.

Regards,
Harshit