You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The evaluation of splines is nested - one loop over blocks of splines, and an inner loop over the splines in that block. I think this was done to experiment with distributing the spline evaluation across (shared memory) processors.
This feels like an implementation detail that should not be in the base miniapp. For the base miniapp, there should be a single loop over splines.
A better way to do this is to create a spline interface, and then put the breaking of splines into blocks as one implementation behind that interface.
I think this is general. In fact, thinking about splitting the walkers, this feature provide another layer of parallelism. The QMCPACK CUDA code has some similar feature although only the computing not the memory is chunked.
Regarding the loop over blocks, one nice thing about it is that it helps with cache pressure as described in the IPDPS paper (or maybe SC). It turns out to be a good feature even if you are not letting different processing elements handle different blocks.
Arguing for the simple loop is that it will be easier for others to understand. The benefit of the loop over blocks is that it is well studied and generally turns out to be the best implementation for many platforms. I might suggest that the _ref version could have the simple loop that Mark suggests and the standard implementation retain the loop over blocks. This does imply some coupling of the data types as well, but this could also be handled easily.
Activity
ye-luo commentedon Jul 13, 2018
This is the advantage of miniQMC over QMCPACK. There are many use of this feature for exploration.
mdewing commentedon Jul 13, 2018
So the nested evaluation of splines is general enough to keep in the base miniapp?
ye-luo commentedon Jul 13, 2018
I think this is general. In fact, thinking about splitting the walkers, this feature provide another layer of parallelism. The QMCPACK CUDA code has some similar feature although only the computing not the memory is chunked.
lshulen commentedon Jul 13, 2018
Regarding the loop over blocks, one nice thing about it is that it helps with cache pressure as described in the IPDPS paper (or maybe SC). It turns out to be a good feature even if you are not letting different processing elements handle different blocks.
Arguing for the simple loop is that it will be easier for others to understand. The benefit of the loop over blocks is that it is well studied and generally turns out to be the best implementation for many platforms. I might suggest that the _ref version could have the simple loop that Mark suggests and the standard implementation retain the loop over blocks. This does imply some coupling of the data types as well, but this could also be handled easily.