Recommend to use Conda/Mamba to compile TomoATT on NSCC

[JingChen-Thu]

Due to the routine update of NSCC environment, TomoATT compilation instruction usually become invalid. So, I recommend a easy way to compile TomoATT on NSCC using conda environment.

First, we can clean loaded modules and load conda environment using

module purge
module load miniforge3

After that, we will have "(base)" at the beginning, as follows:

There are two ways to install TomoATT:

1. The most convenient way to install TomoATT of main branch, See Installation instructions via Conda/Mamba for details.

Here lists a example using conda.

Create a new environment and apply a quick installation

conda create --name tomoatt
conda activate tomoatt
conda install -c conda-forge tomoatt pytomoatt

2. Compile source code

If users want to compile source code by themselves, we can use the following instruction to install dependence
Also given here:

conda create --name tomoatt
conda activate tomoatt
conda install cxx-compiler cmake yaml-cpp openmpi "hdf5=*=mpi_openmpi*" -c conda-forge

(PS: this process may take several or tens minutes sometimes)

Now, we can activate "tomoatt" environment using

conda activate tomoatt

or quit to base using

conda deactivate

After installing the dependence, we can simply download and compile TomoATT under the "tomoatt" environment, as follows

git clone https://github.com/TomoATT/TomoATT.git
mkdir build && cd build
cmake .. 
make -j

Then, we will generate the executive file in path/to/TomoATT/build/bin/TOMOATT

NOTICE:
Please compile TomoATT under the created "tomoatt" environment. We also need to go to the "tomoatt" environment when we run TOMOATT. For example, when we submit a script to NSCC, we need

module purge
module load miniforge3
conda activate tomoatt

before

mpirun -n 128 ${TOMOATT_path} -I input_param.yaml

[xumi1993]

• If conga-forge works on NSCC, why do not install TomoATT directly using

  mamba install tomoatt
  

  Is there any issues about this command?

• For miniforge environment, I recommend using mamba rather than conda to install dependencies, due to its much faster environmental solving.

• I remember the Cray compiler is the internal compiler. Have you guys tried to compile TomoATT using this complier, which may be faster than GCC from Conda environment.

[mnagaso]

Hi @xumi1993
Would the TomoATT binary with this command be a compiled one with SIMD for each architecture of installation targets?
For running a large scale calculation, it is better to make it available as ~x4 speed up is expected if the machine architecture supports AVX256.

[xumi1993]

@JingChen-Thu, the current solver does not support SIMD, but we can test the legacy solver with SIMD and compare the time consumption and accuracy with the current solver, especially on large scales.

[JingChen-Thu]

Based on my vague memory, SIMD is only applicable for third order non-upwind solver and inapplicable for first order upwind solver?

To be honest, I am not quite sure. But if so, first order upwind solver usually has higher efficiency due to less iteration (for eikonal solver) and also higher accuracy due to upwind scheme. I fully recommend to use upwind solver.

By the way, will SIMD introduce additional memory cost? I also remember one user encountered a problem that TomoATT require much additional memory capacity when using domain decomposition or hyperplane stepping parallelization. However, these two parallel methods do not require much additional memory, as illustrated in our 2025 paper published on C&G. (I am also not quit sure since it happens about 1 year ago).

[mnagaso]

Thanks for your comments!

SIMD is not about the memory, but the thing which accelerate the computing speed. So user may prefer SIMD for speed up if their machine has plenty memory or simulation size is not very large. While it requires additional memory (maybe double or so) to put the memory in 1d arrays for optimize the memory accessing speed onto L1/L2 cache.

So i will add SIMD functions for the recent solvers.

Thank you very much again for your continuous development and adding useful functions!

[JingChen-Thu]

@mnagaso Thank you, Masaru! It would be wonderful if SIMD can be used for the recent solver (first order upwind solver) for higher efficiency! We can also provide an optional for user to use SIMD or not when compiling tomoatt, to fit their environment.

If you have any problem about the solver when add SIMD functions, please feel free to tell me. We can discuss it together in detail.

[JingChen-Thu]

Sure, this command is more convenient for tomoatt installation. I have added your notation. Thanks for this suggestion.

[xumi1993]

Is mamba available for this installation workflow?

[JingChen-Thu]

Is mamba available for this installation workflow?

Actually, I have not not try mamba on NSCC yet. I will test mamba on NSCC soon and revise our instruction depending on test results.

[JingChen-Thu]

Is mamba available for this installation workflow?

@xumi1993

I failed to use mamba on HPC. Although mamba has been installed on NSCC, however, I failed to use it to create env or activate env right now :(