Opt examples run significantly slower than CUDA solver

[gerwang]

I am using Opt with VS2015 and LLVM 6.0.1 on Windows. When I ran those examples, the Opt method is significantly slower than the corresponding CUDA solver, but the paper claims than Opt should be faster. Is it an issue with my platform or that is what it should be?

[Mx7f]

What version of CUDA are you using? The values in the paper were generated with VS2013, CUDA 7.5, and LLVM 3.8.

[gerwang]

I am using CUDA 10.0. I noticed that CUDA 10.0 should generate PTX 6.3, my LLVM actually supports PTX 6.0 (I applied the commit on https://marc.info/?l=llvm-commits&m=153783073315460&w=2 manually to get my LLVM to work), is that a problem?

[Mx7f]

Sorry for the delay on this, I will be trying to upgrade my personal development machines to CUDA 10 this month, so hopefully can give you a real answer soon.

[mfratarcangeli]

I also have the same problem. I have compiled Terra from source using LLVM 6.01. I use Windows 10, CUDA 10.1. The cuda solver runs ~20X faster than Opt.

[ProfFan]

Hi, @marfr960

Can you try using terra master with LLVM 7?

Fan