antonmks / Alenka

GPU database engine
Other
1.17k stars 120 forks source link

Device Errors #100

Open dakami opened 7 years ago

dakami commented 7 years ago

I've been trying for a bit to get Alenka to run against these Titan X Pascals. I just get device errors akin to:

terminate called after throwing an instance of 'thrust::system::system_error' what(): function_attributes(): after cudaFuncGetAttributes: invalid device function

Things I've tried have included iterating across Ubuntu 14 and 16, running against CUDA 7.5 in a nvidia-docker instance, manipulating compute version, trying to compile with clang/gpucc, trying both master and development, commenting chunks of moderngpu out... it just doesn't want to work. Any ideas?

antonmks commented 7 years ago

Which compute version have you used ?

dakami commented 7 years ago

20, 35, 50, 62

antonmks commented 7 years ago

I don't know. I don't have 1080 so I can't test it myself, but there was a guy couple of days ago with 1080 - https://github.com/marklit - I think he managed to compile and run it. I myself tested alenka on Gtx Titan with CUDA 7.5 and 8.0 on Ubuntu and Windows.

marklit commented 7 years ago

I got a successful compile for my GTX 1080 using CUDA 8 and only enabling compute 6.1. I originally tried CUDA 7.5 but ran into a few issues there.

https://github.com/antonmks/Alenka/issues/99#issuecomment-252738069

I'm not sure if 6.2 could be the issue, does 6.1 in isolation with CUDA 8 compile properly?

dakami commented 7 years ago

OK, I've got Alenka running on Pascal/CUDA 8. I think crashes are now because of drift between your TCPH / SSB generator, and what I'm putting together. I'm going to try to assemble some of my own tests.

I can't get much to work in TCPH. SSB absorbs and bitmaps, but queries past SS11 segfault or cause device errors.