GitLab

* initial implementation of distributed DISCO layer

* working distributed convolution

* working refactored serial conv transpose with torch kernel

* working distributed conv and transposed conv when using the python kernel

* working distributed convolution with torch kernel

* fixed triton kernel tests

* adding print statement to debug CI

* adjusting tolerances in local convolution unittest

---------
Co-authored-by: Boris Bonev <bbonev@nvidia.com>

* Cleaned up DISCO convolutions

Changed the code to only implicitly use sparse tensors in the modules, in order to enable compatibility with DDP

* Moved convolutions and exposed them directly

* Added transposition to the unit test

* Minor bugfix in CPU version of DISCO transpose code

* Adding convolution tests to CI

* Added gradient check

* Checking the weight grad as well

* Added test for anisotropic kernels

* Fixing the precomputation of the Psi matrix

* moving tests to a parent folder

* removing tqdm

* deleting deprecated distributed tests

* Moving distributed test

* adapting workfolow

* Added some more comments to make the code more understandable

* more detailed explanation for the derivation of the rotation angles

* added another comment

---------
Co-authored-by: Thorsten Kurth <tkurth@nvidia.com>

* fixing banding issues with corrected computation of kernel size
* fixing Parameters
* changed normalization of isotropic kernels

---------
Co-authored-by: Thorsten Kurth <tkurth@nvidia.com>

* working distributed fwd SHT with flexible sharding and without padding

* fixing distributed sht with new logic

* fixed distributed SHT and ISHT with flexible padding

* working unittest for distributed fwd SHT

* distributed tests converted to unit tests

* bumping version number up to 0.6.4

* fixing small splitting bug in th distributed

* updated changeloc, removed tests folder

* Adding prototype implementation of disco convolutions

* adding s2convolutions.py

* Somewhat functional first implementation in Triton

* Somewhat functional first implementation in Triton

* Made batched execution work on GPU and refactored precomputation of kernel values and supports.

* Adding the Lobatto grid to the grid selection method in quadrature

* fixing bug in triton kernel when iterating over non-zeros

* Updating with working custom triton implementation of DISCO convolution

* Merged both forward and backward DISCO contraction into a single kernel

* Some cleanup and minor bugfix

* bugfixes to the reference implementation

* adding weights to s2conv

* removing unnecessary imports

* suggestion for torch harmonics, math should be checked thoroughly

* Intermediate working reference implementation for the transpose DISCO convolution. Fixed normalization of kernels

* adjusting cutoff frequency in disco convolution

* fixing transpose DISCO contraction

* moving triton to install_requires

* enabling triton by default

---------
Co-authored-by: Thorsten Kurth <tkurth@nvidia.com>

* Adding prototype implementation of disco convolutions

* adding s2convolutions.py

* Somewhat functional first implementation in Triton

* Somewhat functional first implementation in Triton

* Made batched execution work on GPU and refactored precomputation of kernel values and supports.

* Adding the Lobatto grid to the grid selection method in quadrature

* fixing bug in triton kernel when iterating over non-zeros

* Updating with working custom triton implementation of DISCO convolution

* Merged both forward and backward DISCO contraction into a single kernel

* Some cleanup and minor bugfix

* bugfixes to the reference implementation

* adding weights to s2conv

* removing unnecessary imports

* suggestion for torch harmonics, math should be checked thoroughly

* Intermediate working reference implementation for the transpose DISCO convolution. Fixed normalization of kernels

* adjusting cutoff frequency in disco convolution

* fixing transpose DISCO contraction

* moving triton to install_requires

---------
Co-authored-by: Thorsten Kurth <tkurth@nvidia.com>

Major Cleanup of SFNO. Retiring non-linear architecture and fixing initialization. Adding scripts for training and validation.

* Updated Changelog

* improved readability of legendre.py

* bugfix in computation of dlegpoly

* Moving conversion from numpy to torch from the legendre module to the sht module

* renaming to vdm in tests.py

* fixing contig in distributed sht

* make it normally contiguous before all to all

* commented the code to be a WAR

* reworked SFNO example

* updated changelog

* Added gradient check to test suite
* reduced size of the unit test
* switched to parametrized for unittests

Without this fix, importing torch_harmonics had run into a module import error
when trying to execute:
```import examples```

Don't import sfno by default (or try/except importing apex)