build instructions unclear #92
Comments
|
I build FahBench on Windows and it worked to build OpenMM on its own and FahBench on its own and then just copy the OpenMM dlls to FahBench folder. Maybe something similar works on Linux too? |
|
I am having similar issues trying to compile this on linux and the instructions are sorely lacking. can someone update it? @akostadinov how did you get boost in there? when i download the package nothing is in the boost folder. the instructions say that boost will be automatically downloaded, but it is not, the folder remains empty. i've tried manually downloading boost, but there is no cmake list file there and it still gives an error. |
|
Hiya @IanSteveC Please note that I have reached out to someone so let's see what happens 🤞 |
|
@PantherX Thanks, I'm looking forward to it, I'd love to be able to compile the FAH bench app with core22 and the latest CUDA support for some testing. Or if we can get someone to compile and release a new build of FAHbench with these features, then I wont even have to bother trying to build it myself. Whatever works. |
|
I can only give it to you unofficial build for Windows with Core22 mixed precision OpenCL / CUDA https://gofile.io/d/qdIqlR Only download file if you trust me. This is a Windows executable file and could in theory be a computer virus. |
|
@informatorius, what version of CUDA is this? I need the latest CUDA 11.1 for full RTX 3000 Ampere support. |
|
It is CUDA 9.2 compiler. I build it using CUDA 11.1 but the runtime needs the CUDA 9.2 dlls for kernel compile. I got it from a real FAH work unit running with CUDA. I can also provide build guidance for Windows if you want to try it yourself. I was wrong, it works with CUDA 11 compiler too. |
|
If you can provide full instructions for Windows I’ll give it a try. |
|
I build FAHbench and OpenMM with Release x64
I had issue with some environment variables not set and some includes and libs path not correct. After fixing that it builds everything Same instructions are valid for the FahBench source code. But it needs additional QT library for GUI.
If both FahBench and OpenMM are build then copy OpenMM.dll to FahBench root folder and OpenMMCPU.dll, OpenMMCUDA.dll, OpenMMCPU.dll to FahBench/OpenMM subfolder For CUDA you will also need cufft64_92.dll, nvrtc64_92.dll, nvrtc-builtins64_92.dll copy to FahBench/OpenMM subfolder e.g. from CUDA toolkit folder |
|
I was wrong FahBench runs with CUDA 11 only too. I will update the zip package. |
|
Thanks I’ll try that one. Did you use CUDA 11.0 or 11.1? |
|
@informatorius I'm getting an error trying to run your FAH bench version. I did install the CUDA 11.1 toolkit on my system. i get "Error loading CUDA module: CUDA_ERROR_FILE_NOT_FOUND (301)" |
|
if i remove the CUDA toolkit I get: "error compiling program: nvrtc: error: failed to open nvrtc-builtins64_111.dll. another thing that is interesting is that I think it should be trying compute_86, not compute 70. is this a parameter you changed when you built this? my RTX 3070 has CC 8.6, not 7.0 |
|
That sounds like it did not like my renaming of cuda files from 111 to 92 which works on my PC. Then I have to rebuild FahBench with 111 in code. I did not change compute_86 but I guess FahBench/OpenMM does not use newer CUDA features anyway |
|
thanks, I got it working by using your first link with cuda92 DLLs and then adding/renaming the cuda111 DLLs from a freshly installed CUDA toolkit (then uninstalled CUDA toolkit). |
|
but i hope someone can update the linux instructions and procedures. I'm more comfortable with linux compiling and use. |
|
I'm more familiar with Windows but it should also be easy on Linux, download source, download dependencies devel, cmake and build. I can try it in a Linux VM but don't know if this can run CUDA... |
|
this part: "3. From a clean build directory, use ccmake to configure the build. Make sure the right OpenCL and Qt libraries are being found." the first issue is that these instructions are incredibly vague. there really needs to be a more specific direction on how exactly to use ccmake to get the desired output. saying " just use it" is a lot different than saying "run this exact command with these parameters". when doing this, the configure fails because it cannot find boost, claiming there is no boost cmake lists file. and indeed there isnt since the boost directory comes empty. in another version of the instructions here: http://fahbench.github.io/build-and-install.html it says "Boost (automatically downloaded and built)". with no explanation on how boost is supposed to be automatically downloaded and installed if there is no cmake file or nothing in the code that seems to perform this action. i've tried downloading boost manually, but there seems to be no way to link that into this install scheme. there is no cmake lists file for boost when you download their source from the vendor. in general, their Linux instructions have a lot of missing steps or implied steps making the whole thing very unclear. that's what triggered this whole discussion. |
|
For the boost part I can say on Windows it gets downloaded when building the solution, because there is a boost project which purpose is to download and build boost on the fly during build. I will try it on Linux when I have time. |
|
boost and openmm are included as submodules in this repository. Make sure you force git to actually check them out. This will obviate the need to download them separately |
|
Can you elaborate? Maybe provide a command? How do you “force git to actually check them out” In my previous attempts I just downloaded the zip file. And the openmm and boost folders are empty in this case. |
|
git clone --recursive |
Hi, current build instructions are not very helpful. After I managed to put proper
boostI'm stuck on openmm setup. I can't find any patches to be applied. So I just cloned7.4.1tag intoopenmmdir. It builds fine on its own. But as part of fahbench build it failsI copied cmake files to
cmakedir ti have them found, then build failed:Apparently some paths are confused. So I guess I'm missing some patches. Existing documentation does not help me much though. Thanks.
The text was updated successfully, but these errors were encountered: