Allow multiple inner products with the same vector in one kernel

[karlrupp]

Required for some flavors of GMRES.

[ptillet]

This shouldn't be a problem. However, we need to fully specify an upper
limit on the number of inner products that can be packed together,
considering each new inner product induces register pressure and local
memory usage :P

2013/5/31 Karl Rupp notifications@github.com

Required for some flavors of GMRES.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/32
.

[karlrupp]

Some time ago I implemented this for up to 8 different vectors (outside ViennaCL). Higher counts just repeatedly call these kernels then. Works like a charm :-)

[ptillet]

I prepared the field to handle dual-kernels with multiple operations. However, I am realizing packing multiple inner products together can only be done if there is no dependency. How/Where to check the dependency between operations? This should actually be a distinct tool, almost independent from the generator (since it may also be used in the scheduler or multi-devices)...

[karlrupp]

I'd prefer to check this in the scheduler logic. Thus, I'd like to have a 'no aliasing can be assumed' policy for the generator, because this makes this way the generator can really focus on performance without having to do too much bookkeeping. The alias checking is much better done by a scheduler logic on top of the generator.

[ptillet]

Awesome ! This makes my life much much easier if independent operations are in independent custom operations!

[karlrupp]

Implementation completed here:
0a1e5c9