Using AsFermi for KernelGen on Kepler

[Dmitry N. Mikushin]

Dear colleagues,

First of all, thank you again for working on AsFermi!

Today on behalf of group I presented yet another report on our
compiler present state, which includes some tests on Kepler (slides 15
and 16). Overall, perf on Kepler is currently worse than on Fermi. One
unusual thing: it turns we mix non-scheduled code of dynamic loader
with scheduled ptx-generated code of inserted kernel code. Does it
make sense? Do we need to start every code insertion from
particularly-aligned address to make scheduler happy?

- D.

[Hou Yunqing]

Hi D.,

Do you mean you haven't tried dynamically loading ptxas-scheduled
instructions? I guess it should work if it were properly aligned (to
0x40 boundries most probably).

Yunqing