search

lattice / quda

QUDA is a library for performing calculations in lattice QCD on GPUs.
https://lattice.github.io/quda
Other
287 stars 94 forks source link

Block dilution #1393

Closed maddyscientist closed 10 months ago

maddyscientist commented 1 year ago

This simple PR add supports for local disjoint hyper-cubic block dilution. I developed this as part of some experiments with smoothers. This effort is on ice, but since this is a nice self-contained feature addition, I figured it's best to get this upstreamed. This work makes some small changes to the colorspinor::FloatNOrder accessor, which will be helpful for multi-RHS.

Note this PR is dependent on #1384 and so should not be considered for merging until after that is merged.

weinbe2 commented 10 months ago

This is a nice PR @maddyscientist , thanks!

One concern I have is that, when running dilution_test by hand, there's an OOM error (on an 80GB card). This is reasonable on the merits because the default values are a volume of 24x24x24x24 and the default dilution size is 1x1x1x1. I know that "under the hood" the maximum number of dilution blocks is 128, but the OOM error hits during the allocations before that.

This is not an issue during ctest since the local volume, and the dilution block size, is sufficiently small.

In the interest of the test executables "just working", I recommend changing the default value for --dilution-block-size to 8 8 8 8, since it "just works", though I'm not married to any particular value so long as it doesn't OOM on a reasonable GPU and the number of dilution vectors is <= 128.