search

omnia-md / omnia-linux-anvil

Experimental version of conda-forge/linux-anvil to build openmm
0 stars 3 forks source link

Create a CUDA 9.1 build #7

Closed Lnaden closed 6 years ago

Lnaden commented 6 years ago

Issue to track the build of a AMD3.0 + TexLive 2017 + CUDA 9.1 build

I am working on building this.

@peastman I am looking at the CUDA installers here and there are three patches for this version of CUDA, does the docker image need these patches installed for OpenMM to build correctly against, or is un-patched the base version fine?

Specific details of the patches:

  1. cuBLAS Patch Update: This update to CUDA 9.1 includes new GEMM kernels optimized for the Volta architecture and improved heuristics to select GEMM kernels for given input sizes.
  2. CUDA Compiler Patch Update: This update to CUDA 9.1 includes a bug fix to the PTX assembler (ptxas). The fix resolves an issue when compiling code that performs address calculations using large immediate operands.
  3. cuBLAS Patch: This CUDA 9.1 patch includes fixes to GEMM optimizations for convolutional sequence to sequence (seq2seq) models.
peastman commented 6 years ago

It sounds like those are just bug fixes, so probably the base version is fine. As long as they don't change the SONAME of any libraries they should be compatible.

Lnaden commented 6 years ago

Great! That does not sound like something a cuda patch would do, so I don't think they do since they are not introducing new libraries

Lnaden commented 6 years ago

This has been created in the texlive-amd30-cuda91 branch.