Re: Fwd: Re: Errors while installing ESPResSO

[Michael Kuron]

Hi Angelos and Jean-Noel,

judging from the backtrace, the problem is this: OpenMPI calls into
hwloc, hwloc calls into opencl and opencl is the AMD-provided version
of the library. For some reason, that segfaults. You will have this
problem with any MPI program, not just Espresso. I think you might be
able to work around it by setting the environment variable
HWLOC_PLUGINS_BLACKLIST=hwloc_opencl .

Other than that, installing both AMD and Nvidia graphics drivers is
calling for trouble. Both provide OpenCL and OpenGL libraries that may
conflict. I recommend uninstalling all rocm-* packages (or
alternatively, uninstall the nvidia-* packages and remove the Nvidia
card, it's no faster than your integrated AMD GPU).

Michael

On Wed, 2020-01-15 at 11:59 +0100, Jean-Noël Grad wrote:
> Hi,
> 
> @Angelos do you also get an ompi_mpi_init error message if you
> compile 
> and run a small MPI program, like the one attached?
> 
> @mkuron do you know if this is normal for espresso compiled with cuda
> to 
> call rocm/opencl shared objects on a machine with both an AMD
> chipset 
> and an NVIDIA GPU?
> 
> Best,
> JN
> 
> On 1/15/20 11:56 AM, Jean-Noël Grad wrote:
> > I'm forwarding this to the mailing list as the attached backtrace
> > can be 
> > useful to other people. You can click "Reply List" in your email 
> > software to reply to address@hidden instead of me,
> > this 
> > will help other maintainers following this discussion.
> > 
> > JN
> > 
> > -------- Forwarded Message --------
> > Subject: Re: Errors while installing ESPResSO
> > Date: Wed, 15 Jan 2020 11:26:44 +0200
> > From: Angelos Mourkas <address@hidden>
> > To: Jean-Noël Grad <address@hidden>
> > 
> > Hello again,
> > 
> > 
> > I did as you asked, and it detects NVIDIA graphics card and CUDA
> > drivers 
> > without a problem. But when i run "make check" i have multiple
> > errors 
> > and some tests that were unable to complete. I guess the MPI
> > service 
> > could not be started. Also, if i try to import again espressomd
> > now 
> > gives me errors. I have the errors attached.
> > 
> > 
> > Best Regards
> > 
> > Angelos Mourkas
> > 
> > 
> > On 14/1/20 1:08 μ.μ., Jean-Noël Grad wrote:
> > > Hi,
> > > 
> > > According to the CMake output, the HIP compiler was selected to 
> > > compile espresso for your AMD chipset. If this is indeed what
> > > you 
> > > intended, instructions to install the ROCm SDK can be found at
> > > the end 
> > > of section 2.1.1. Installing requirements on Ubuntu Linux:
> > > 
http://espressomd.org/html/doc/installation.html#installing-requirements-on-ubuntu-linux
> > >  
> > > 
> > > 
> > > If however you intended to compile espresso for the NVIDIA
> > > graphics 
> > > card, you'll need to modify the espresso CMakeLists.txt file.
> > > CMake 
> > > will not detect the main GPU, instead it will try all available
> > > GPU 
> > > compilers on your system and select the first one that works, in
> > > your 
> > > case the HIP compiler. To avoid that, please try this: in 
> > > /home/agelosmo/Desktop/Simulations/espresso/CMakeLists.txt at
> > > line 
> > > 181, replace "if(HIP_FOUND)" by "if(0)" to skip the HIP
> > > compiler. 
> > > CMake will then try to find NVIDIA CUDA 9.0. If CMake then says
> > > it 
> > > couldn't find a GPU compiler, it means that your installation of 
> > > NVIDIA CUDA is not properly detected by CMake.
> > > 
> > > Best regards,
> > > JN
> > > 
> > > On 1/14/20 8:44 AM, Angelos Mourkas wrote:
> > > > Greetings Jean-Noël,
> > > > 
> > > > 
> > > > Thank you for the quick respond to my problem. Attached you can
> > > > find 
> > > > two txt files with the cmake output and the make compilation
> > > > error. 
> > > > Also i have attached the features error.
> > > > 
> > > > At the moment, i am using my nvidia card as the main graphics
> > > > card. 
> > > > The log files are from 4.1 version of ESPResSO, but i had the
> > > > same 
> > > > problem with 4.1.2 version.
> > > > 
> > > > Thank you very much for your help.
> > > > 
> > > > 
> > > > Best Regards
> > > > 
> > > > Angelos Mourkas
> > > > 
> > > > On 13/1/20 7:16 μ.μ., Jean-Noël Grad wrote:
> > > > > Hi,
> > > > > 
> > > > > Compilation errors will interrupt the generation of the
> > > > > Python 
> > > > > bindings. You might still be able to import espressomd, but
> > > > > won't be 
> > > > > able to import its submodules. Please also note the rocrand
> > > > > library 
> > > > > is specific to AMD graphic cards; espresso compiled for AMD
> > > > > will not 
> > > > > execute on an NVIDIA card.
> > > > > 
> > > > > To help us diagnostic your issue, please attach a copy of
> > > > > the 
> > > > > compilation error to your email, as well as a copy of the
> > > > > CMake 
> > > > > output, which contains information about your compiler and 
> > > > > libraries, and the version of espresso you're compiling.
> > > > > 
> > > > > Best regards,
> > > > > JN
> > > > > 
> > > > > On 1/13/20 1:07 PM, Angelos Mourkas wrote:
> > > > > > Greetings,
> > > > > > 
> > > > > > 
> > > > > > My name is Angelos Mourkas, a new user of ESPResSO from
> > > > > > Greece. I 
> > > > > > am doing my PhD at the University of Ioannina, Department
> > > > > > of 
> > > > > > Material Science and Engineering.
> > > > > > 
> > > > > > I have a problem in getting the program to work. I use
> > > > > > Ubuntu 18.04 
> > > > > > LTS. I believe i have installed all the necessary packages
> > > > > > for the 
> > > > > > program to work. I have a Ryzen 3 2200g (integrated graphic
> > > > > > card 
> > > > > > vega 8) and a discrete graphic card (nvidia 1030gt). I
> > > > > > have 
> > > > > > downloaded and installed nvidia drivers (this is the card i
> > > > > > use) 
> > > > > > and also i have installed the cuda drivers through the
> > > > > > commands 
> > > > > > from the documentation.
> > > > > > 
> > > > > > Everything is fine until i run the "make" command, where it
> > > > > > reports 
> > > > > > an error concerning <rocrand/rocrand_kernel.h> and another
> > > > > > one at 
> > > > > > EspressoCore_generated_lb_inertialess_tracers_cuda.cu.o.cma
> > > > > > ke:120.
> > > > > > 
> > > > > > If I ignore the message and continue to the python
> > > > > > enviroment, when 
> > > > > > i enter the command "print(espressomd.features())" it
> > > > > > returns the 
> > > > > > message
> > > > > > 
> > > > > > Traceback (most recent call last):
> > > > > >    File "<stdin>", line 1, in <module>
> > > > > > AttributeError: module 'espressomd' has no attribute
> > > > > > 'features'
> > > > > > 
> > > > > > 
> > > > > > Can you please help me installing the program?
> > > > > > 
> > > > > > 
> > > > > > Best Regards
> > > > > > 
> > > > > > Angelos Mourkas
> > > > > >