this is the test I promised to write. It does the following things:
initialise two random distillation/A2A style vectors with VDIM components each: phi is dense and rho is nonzero only on TSRC (like in exact distillation)
compute 3 meson fields: phi-phi, phi-rho, rho-rho
save them to disk as a HDF5 file.
Using h5diff you should then be able to verify that the output hasn't changed when changing the code.
For distillation, a straightforward optimisation would be to accelerate the phi-rho and rho-rho meson fields by only computing them for t=TSRC (or even better, a list passed to the routine). All other components are predictably zero.
The other optimisation is obviously the loop order in the kernel. I'm happy to get involved with this as well if there is any way I can help you.
paboyle
commented
3 years ago
Hi Peter,
this is the test I promised to write. It does the following things:
Using h5diff you should then be able to verify that the output hasn't changed when changing the code.
For distillation, a straightforward optimisation would be to accelerate the phi-rho and rho-rho meson fields by only computing them for t=TSRC (or even better, a list passed to the routine). All other components are predictably zero.
The other optimisation is obviously the loop order in the kernel. I'm happy to get involved with this as well if there is any way I can help you.