Remove tc thread private array

AMReX-Combustion / PelePhysics

A collection of physics databases and implementation code for use with the Pele suite of of codes

https://amrex-combustion.github.io/PelePhysics/

Other

60 stars 52 forks source link

Remove tc thread private array #435

Closed marchdf closed 1 year ago

marchdf commented 1 year ago

@malihass can you check to see if this gives correct results? I think it does but your tests are better ;)

jrood-nrel commented 1 year ago

This appears to have no performance implications.

malihass commented 1 year ago

I see a very small diff of temperature specifically when running the 0D ignition with the analytical Jacobian and the dodecane_lu_qss. I don't see any diff for dodecane_lu and ch4-2-steps. I also don't see diff for the qss mech with numerical Jacobian. Maybe something is off with the symbolic Jacobian?

Otherwise for performance, I suspect it will shine most on GPU. Did you see no diff when running on GPU @jrood-nrel ?

marchdf commented 1 year ago

I don't expect a performance change (it's basically what I want). I am not really doing this for performance... but i am hoping this gets me to the nasa9 polynomials easier

marchdf commented 1 year ago

I think there are a couple of things causing this minor diff in the analytical jacobian:

pow(x, 0.5) became sqrt(x)
logT, T2, etc are expressions instead of using tc[i]. All this might lead to a slightly different expression optimization path.

Thoughts?