search

lattice / quda

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

individual residua for multi-shift solver #8

Closed bjoo closed 11 years ago

bjoo commented 13 years ago

Hi, Mike has a nice algorithmic trick for RHMC where we can ask for different residua for different shifts in the multi-shift algorithm. Typically the ill conditioned small shift systems contribute less to the overall force and only the higher shifts need to be solved more accurately. It would be nice if we could specify a desired residuum for each shift, rather than having a master one for the lowest shift.

This may need some kind of modification to the invParams tho. Let us get our thinking caps on as to how to specify this in the interface.

jpfoley commented 12 years ago

I hadn't noticed this issue until recently but it seems I've been assigned to work on it, so I'll make a few points. First off, QudaInvertParam already has member variables tol and tol_offset, where tol is a double and tol_offset is an array of doubles, one for each offset. As far as I can see, tol_offset is completely disregarded at the moment, In Balint's example, the system with the smallest shift needs to be solved less accurately than larger-shift systems. So one could, for example, run multi-shift until the tolerance for the lowest shift had been reached and then polish the higher-shift results individually until they reached their respective target residuals. On the other hand, with the HISQ action, you usually want to solve the low-shift results to a higher accuracy. The high-shift results don't incorporate the quark-mass dependence of the naik term, and they need to be polished with the correct operator. One way of making the multi-shift solver more flexible would be to use tol to specify a base tolerance for the standard multi-shift inverter and tol_offset to specify a set of residuals for the polishing step. This would be a trivial change to the mixed-precision multi-shift solver. Of course, more flexibility can mean more confusion. Any comments?