[Opengcd-discuss] missing features in gcd/xdispatch

[David M. <Dav...@or...>]

Hello Everyone,

I just finished a comparative study of xdispatch , boost::asio, Intel
TBB and c++11 thread constructs. What I did was write multiple
implementations of an OpenCL based data-flow scheduler.  As a result of
this project I discovered a few things missing in gcd.  

1.       The ability to get the number of items currently waiting on a
queue.  This is important for load-balancing.

2.       A way to create thread-specific data. GCD has queue specific
data, but since a queue isn't bound to a thread it's not the same thing.
I found this to be important for efficiently doing task parallelism on
GPU's, where the GPU driver itself can handle multiple streams, but
those streams need to be bound to threads.

3.       A way to limit the number of threads that a work item can be
dispatched on.  This is also relevant to GPU operations, where (at least
on Nvidia) you want a thread per "real" GPU core and a stream (or OpenCL
queue) attached to each core.

 

My question is whether it is worth my while to make these changes and
try to get them accepted and what is the procedure, forums etc. Where
these issues can be discussed. I don't want to do this work unless there
is some sort of consensus on acceptance.  Considering that Apple is
driving GCD, I don't expect much of a hearing, but I imagine the
xdispatch community would be more receptive.  Some of these issues (like
queue size) can only be resolved at the GCD level, but others could
conceivably be implemented as xdispatch features.

 

Please advise,

David Minor

Orbotech