This PR adds support for testing heavy-quark CG solves in the gtest solver testing. In doing so, I have seemingly discovered a bug in the HQ solver, where it fails to converge in mixed-precision.
The following command
tests/invert_test --dslash-type wilson --ngcrkrylov 8 --dim 2 4 6 8 --prec double --prec-sloppy single --tol 1e-7 --tolhq 1e-7
This is seemingly a mixed-precision bug in the HQ CG solver, as convergence is fine for the regular CG solver, and also proceeds as expected in the HQ solver in uni-precision.
@weinbe2 handing this to you to fix this bug, so we can get this test merged in.
This PR adds support for testing heavy-quark CG solves in the gtest solver testing. In doing so, I have seemingly discovered a bug in the HQ solver, where it fails to converge in mixed-precision.
The following command
fails to converge.
This is seemingly a mixed-precision bug in the HQ CG solver, as convergence is fine for the regular CG solver, and also proceeds as expected in the HQ solver in uni-precision.
@weinbe2 handing this to you to fix this bug, so we can get this test merged in.