Boost logo

Boost :

Subject: [boost] [compute] Some remarks
From: Thomas Heller (thom.heller_at_[hidden])
Date: 2015-01-01 13:55:51

Hey list,

These are just a few remarks I wanted to share that I made during the compute
review phase. Unfortunately, I don't have the time to write a full review. As
I mostly skimmed over everything so take it with a grain of salt.

First of all, I really like the general idea of Boost.Compute. That is, I like
the idea of having a proper API for OpenCL capable devices and I like the idea
of having parallel algorithms. With that being said, i think it is a waste of
opportunities to have the two tied together so closely especially considering
N4104 with the combination of different execution policies which would allow
for a wonderful API for using different parallelization backends (where
boost.compute could be the reference implementation or some such).
Additionally, I think that the author and other reviewers are running in
circles when it comes to synchronization. IMHO, the event alone is enough and
perfectly fine. An OpenCL event has interesting similarities to
std::future<void>, and the clEnqueue functions to std::async. I'd consider the
command queue however similar to std::thread (and threads should be considered
harmful...). With that being said, a event should be perfectly fine to express
synchronization points. Together with executors, it would even fit into the
more higher level stdlib parallel algorithms. It would be very nice to get rid
of a publicly exposed event class though and have one future to rule them all
(one can dream...). But having a future and event in Boost.compute is, IMHO,
not a good idea.


Boost list run by bdawes at, gregod at, cpdaniel at, john at