Add timers for Hubbard symmetrization #1008
Conversation
|
I'm almost sure it's the same issue that we solved in symmetrize_mt_function(): generation of |
simonpintarelli
force-pushed
the
hubbard-profile
branch
from
3e31e77 to
379a5b2
Compare
|
There is some room for improvment here ^^. I've cherry picked the changes and compiled on todi, At least it seems to fix the bulk of the performance issue. |
| @@ -46,6 +48,19 @@ symmetrize_occupation_matrix(Occupation_matrix& om__) | |||
| } | |||
| } | |||
|
|
|||
| std::vector<std::vector<mdarray<double, 2>>> rotms(sym.size()); | |||
There was a problem hiding this comment.
but why not passing it as argument? it is already computed in sim.ctx.
There was a problem hiding this comment.
Aren't the ones in sim ctx different? https://github.com/electronic-structure/SIRIUS/blob/sirius-md/src/context/simulation_context.cpp#L834
It does the following at the end:
https://github.com/electronic-structure/SIRIUS/blob/sirius-md/src/core/sht/sht.cpp#L98
In symmetrize_occupation_matrix for each symmetry
https://github.com/electronic-structure/SIRIUS/blob/sirius-md/src/core/sht/sht.cpp#L115
was called, which returns a vector of rotations matrices.
I didn't look at the for loops below, perhaps one can rearrange things and use the ctx.rotm().
There was a problem hiding this comment.
I checked the code and I found a TODO that perfectly fits this refactoring. I'll add my changes to this PR
This reverts commit 84c492b.
…IUS into hubbard-profile
|
cscs-ci run |
|
@simonpintarelli @gsavva is it good to be merged? |
* profile symmetrize_occupation_matrix * use precomputed block rotation matrices for real spherical harmonics everywhere --------- Co-authored-by: Anton Kozhevnikov <toxa81@gmail.com>
Add
sirus::symmetrize_occupation_matrix.@gsavva encountered a SCF+U calculation which spent the bulk of the time in
symmetrize_occupation_matrixinDensity::generate.