CUDA error with tree neighbor list

[Jonas Lehnen]

Dear all,

I am trying to use the tree neighbor list to improve the performance of my simulation, but I am running into a CUDA error when changing the nlist from cell to tree.

The example script works fine for

lattice_size<=16 but fails for larger values.

I am using a GPU compiled conda version of hoomd 2.9.7 on a Nvidia RTX 3080

Jonas Lehnen

[Michael Howard]

What is your build configuration? Your script runs fine for me on an RTX A4000, which is the same compute capability as RTX 3080, using gcc 9.4, CUDA 11.7, and a double-precision build of HOOMD.

Can you also reproduce the error using HOOMD 3.x? HOOMD 2.9.7 is no longer supported, so you are not going to get a bug fix even if you do identify one.

Regards,

Mike

[Jonas Lehnen]

Dear Mike

I tried it with the conda-forge version:

HOOMD-blue 2.9.7 CUDA (10.2) DOUBLE HPMC_MIXED TBB SSE SSE2 SSE3 Compiled: 11/08/2021

The same is true for HOOMD 2.9.6 compiled with GCC 10.2.0 and CUDA 11.1.1 and double-precision. The same error occurs for both versions.

The error does not appear for HOOMD 3.5. I will try to compile HOOMD 2.9.7 with the suggested settings and otherwise switch to HOOMD 3.5. 

Thanks for your quick respones.

Regards,

Jonas

[Michael Howard]

Hi Jonas,

I'm not sure what would cause that behavior, but maybe it is an issue with how the PTX code is being translated for your GPU architecture? When you built the code for yourself, did you include compute capability 8.6? I'm not sure if that will work with CUDA 11.1.1, but you could try it and see if it fixes the issue, if you haven't done so already.

It would also be helpful for me if you could report the result of:

ctest --output-on-failure -R neighbor

for your HOOMD 2.9.6 build that gave the error. We can't do any fixes on HOOMD 2.x, but this will help identify any potential bugs in the underlying library used by the tree neighbor list (which is also used in HOOMD 3.x).

Regards,

Mike