support multiple RHS

[lukeolson]

Finish support for multiple RHS in the solve.

[nschloe]

Need that, too. :)

[lukeolson]

@jbschroder @nschloe I did a prototype here:
https://github.com/pyamg/pyamg/tree/multiple-rhs

For most variants of AMG, only the solve phase in impacted (adaptive AMG is one case that needs modification in the setup). For the solve, I added a jacobi_m that executes relaxation on blocked input:
https://github.com/pyamg/pyamg/blob/multiple-rhs/pyamg/amg_core/relaxation.h#L246

Finishing this will take a bit (but not a ton) on two fronts:

1. updating the kernels to blocked format (like jacobi_m) or at least restricting relaxation in this case to say jacobi and Gs.
2. adding better checks /more consistent shapes. right now we mostly use (M,) but even this isn't consistent. Since the cost is low, we can go to (M,1) and modify from there.

(3. more testing...)

[jbschroder]

Thanks for looking into this Luke! I had been thinking that the change could be more trivial, but I guess not.

…

On Tue, Jul 10, 2018 at 9:41 AM, Luke Olson ***@***.***> wrote: @jbschroder <https://github.com/jbschroder> @nschloe <https://github.com/nschloe> I did a prototype here: https://github.com/pyamg/pyamg/tree/multiple-rhs For most variants of AMG, only the solve phase in impacted (adaptive AMG is one case that needs modification in the setup). For the solve, I added a jacobi_m that executes relaxation on blocked input: https://github.com/pyamg/pyamg/blob/multiple-rhs/pyamg/ amg_core/relaxation.h#L246 Finishing this will take a bit (but not a ton) on two fronts: 1. updating the kernels to blocked format (like jacobi_m) or at least restricting relaxation in this case to say jacobi and Gs. 2. adding better checks /more consistent shapes. right now we *mostly* use (M,) but even this isn't consistent. Since the cost is low, we can go to (M,1) and modify from there. (3. more testing...) — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#215 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AEjD41L2vPxe4sG8YWw9aIQdrUgLWdkrks5uFNkmgaJpZM4RmHS6> .

[lukeolson]

It's not too bad ... the tricky part is that we've been sloppy with (M,) and (M,1) sizes.

[CebFu]

I tried it, but seemed not accurate. Are there some demos that can be provided?