We should expose an env to force loop-back communication on in a given dimension, as we have are able to using an API at present. This would allow us to switch on communication when QUDA is driven by an external application, e.g., Chroma or MILC. This would have the great benefit of enabling a reduction in node-hours spent on autotuning by doing most of the tuning at small scale.
maddyscientist
commented
3 weeks ago
We should expose an env to force loop-back communication on in a given dimension, as we have are able to using an API at present. This would allow us to switch on communication when QUDA is driven by an external application, e.g., Chroma or MILC. This would have the great benefit of enabling a reduction in node-hours spent on autotuning by doing most of the tuning at small scale.