AMReX STL functionality

[sshivank]

Hello
I was trying to read an STL file using the AMReX EB STL utilities. I prescribe the eb2 geometry type as "stl" in my input script, and I add the STL file name. Then I call the function EB2::Build(geom, required_level, max_level, 5, true). The code runs on a single core. Volume fractions look correct, and AMR levels work as well. But when I change the number of processors, the STL is not correctly recognized. Only a certain portion of the geometry is recognized. Does the STL functionality work with just one core? If not, could you please suggest what I am doing incorrectly? After testing it for my geometry, I tried the Airfoil stl provided in the AMReX Tutorials. That shows the same behavior as well.

[hsitaram]

Hi @sshivank ,

The STL functionality is not perfect yet, but thank you for spotting this error. The problem is that the triangle data is not getting broadcasted across processors, only getting allocated here -

amrex/Src/EB/AMReX_EB_STL_utils.cpp

Line 243 in 4fab19a

m_tri_pts_h.resize(m_num_tri);

If the stl read was done by all processors, this should work fine, but that can get memory intense for large stl data. Things work fine If you turn off this line and similar in the read functions -

amrex/Src/EB/AMReX_EB_STL_utils.cpp

Line 98 in 4fab19a

if (ParallelDescriptor::IOProcessor()) {

Let me see if there is a cleaner way to fix this. @WeiqunZhang - do you know of an easy amrexy way to Bcast GPU-pinnedvector of structs. Bcast(vec.dataPtr, vec.size) didn't work out.

[WeiqunZhang]

@hsitaram If that works, could you submit a PR to fix this?

[hsitaram]

Yes, that worked great! will submit a PR soon

[sshivank]

Hi @hsitaram, for now, turning off the if (ParallelDescriptor::IOProcessor()) { lines worked. Thanks for the help.

[hsitaram]

@sshivank , we have fixed this issue. Please try again using the latest amrex.

[sshivank]

It worked. Thank you.

[rmrsk]

@sshivank If you are interested, we have functionality for turning STLs (and other surface meshes) into signed distance fields, which lets you use the STL geometry like any other implicit function. It'll also be faster than native AMReX for larger triangle meshes since it uses a more clever lookup algorithm (AMReX goes linearly through these, I think).