turagalab / splinepsf Goto Github PK

1.0 1.0 3.0 160 KB

License: GNU General Public License v3.0

CMake 3.04% C 23.89% Cuda 50.60% C++ 17.15% Batchfile 0.16% Shell 0.12% Python 5.03%

splinepsf's Introduction

SplinePSF

This is a factored out implementation repository of the spline function implemented in C++/C and CUDA. Currently it may only be used in conjunction with the DeepSMLM package. However we are planning on small standalone wrapper for Python / MATLAB; this is also the reason why we opted for a separate repository.

Installation

conda install -c haydnspass -c conda-forge spline

Known Limitations

C++/CUDA implementation is currently not ready for standalone use
Updated: CUDA supports arbitrary ROI size now. ~~CUDA does not support ROIs bigger than 32 x 32 = 1024 pixels (this is due to the maximum number of threads per block). Please refer to the CPU version in this case~~

Build and Install Python package for Local Use

python setup.py install

Buildwheels

python setup.py bdist_wheel

Build and Deploy with conda

# recommended: create a new conda build environment
conda create --name build_clean conda-build
conda activate build_clean

# navigate to [repo]/conda
cd dist_tools/conda
conda-build -c conda-forge spline

splinepsf's People

Contributors

Stargazers

Watchers

Forkers

jimver jackzhousz axtr7550

splinepsf's Issues

Add CUDA Device index selection

In .cu spline constructor

cudaSetDevice(0);

Change device manager to same behaviour as pytorch in decode.

Pip / PyPi Deployment

Hey,

Great to hear that we have some help to figure out how to deploy the stuff via Pip / Pypi as a second option to conda! :) We do this mainly because pip support is more native on Google Colab. I don't know if you have looked at the code/deployment but it is roughly structured like this:

C++ / CUDA is built with the help of CMake. Building with CUDA might require to set environment variables to point to the right CUDA version / compiler etc.; not building the CUDA part on purpose can be done with appropriate flags.
On top of the CMake there is a wrapper as a python setup.py which basically just calls the CMake.
The actual conda deployment happens in the respective conda files / directories which does all the dependency handling and then calls the python setup.py.

I imagine the only change/addition would be in the setup.py which is commonly used to deploy to PyPi right? There one would specify the dependencies again?

The platforms we want to deploy to are macOS, Linux and Windows. Last time I did this to conda I excluded CUDA support for macOS (because Nvidia and Apple ...) and for Windows because there were some weird bugs. It would be good to have CUDA support for windows as well though.

Differentiate between CUDA compiled and CUDA available

Currently used the wrong flag only by compilation mode, but CUDA compiled code can also run on CPU only machine if the code does not call any device code.

MacOS Compiler / SDK on Github

@olbris I think the problem with the macOS runner is the version of the SDK in combination with the clang version I set in the build config. I had to change the version on my local machine as well.

Let's do it like this: https://github.com/adc-connect/adcc/blob/041cdcb16ec60d1655db8d562ef2f3b53038c120/.github/workflows/publish.yaml#L48

Then I will need to set the CONDA_BUILD_SYSROOT and I need to pass this to CMake.
This way it should work irrespective of any changes of the macOS runner.

CMAKE > 3.18 with multiple cuda architectures does not work

import spline
Traceback (most recent call last):
File "", line 1, in
ImportError: /home/lucas/xconda3/envs/decode_dev_cuda_dev_spline/lib/python3.8/site-packages/spline-0.10.0a0-py3.8-linux-x86_64.egg/spline.cpython-38-x86_64-linux-gnu.so: undefined symbol: fatbinData