v0.9.0

• New CPU backend (OpenMP-accelerated) for both DISCO convolution and attention layers
• Pre-compiled manylinux wheels for multiple PyTorch and CUDA versions, available on PyPI and pypi.nvidia.com
• Revised truncation logic for the SHT: centralized in new truncation.py module, enforcing triangular truncation (lmax = min(lmax, mmax)) across all SHT classes. Note: truncation for equiangular/equidistant grids changed from nlat to (nlat+1)//2
• SHT performance improvements: contraction dimensions are now transposed to be stride-1 before einsum, and real/imaginary parts are split into separate contiguous tensors
• New fft.py wrapper module with proper Hermitian symmetry enforcement in irfft and explicit mode truncation in rfft
• Full PyTorch 2 custom operator compatibility for DISCO and attention layers using torch.library registration, enabling torch.compile and torch.export
• Restructured DISCO convolution and attention code into proper subpackages (torch_harmonics/disco/, torch_harmonics/attention/)
• Added double precision support for DISCO convolution
• Fixed Schmidt normalization for derivatives of associated Legendre polynomials
• Fixed up/downsampling in attention layers when input and output shapes differ
• Fixed GaussianRandomFieldS2 to use isht.lmax/isht.mmax for compatibility with revised truncation logic
• Distributed module: added shape verification for transpose and gather operations, controllable via TORCH_HARMONICS_DISTRIBUTED_DEBUG
• Distributed module: fixed finalize() bug where process group was not properly destroyed
• Query functions torch_harmonics.disco.optimized_kernels_is_available and torch_harmonics.attention.optimized_kernels_is_available for checking optimized layer availability
• Quadrature helper functions precompute_latitudes and precompute_longitudes are now public API
• added new tests:
  • Comprehensive SHT test suite now covering vector SHT, Schmidt normalization, batch dimensions, and multiple grid types
  • New test suites for SpectralConvS2, QuadratureS2, GaussianRandomFieldS2, and ResampleS2
    Enhanced DISCO convolution tests covering different input/output channel counts and double precision
  • Enhanced attention tests with up/downsampling and opcheck integration
  • New distributed tests for primitives, quadrature, and spectral convolution
  • Shared test utilities module (testutils.py)

v0.8.0

• Adding spherical attention and spherical neighborhood attention
• Custom CUDA kerneles for spherical neighborhood attention
• New datasets for segmentation and depth estimation on the sphere based on the 2D3DS dataset
• added new spherical architectures and corresponding baselines
  • S2 Transformer
  • S2 Segformer
  • S2 U-Net
• Reorganized examples folder, including new examples based on the 2d3ds dataset
• Added spherical loss functions to examples
• Added plotting module
• Updated docstrings

v0.7.4

• New filter basis normalization in DISCO convolutions
• More robust pre-computation of DISCO convolution tensor
• Reworked DISCO filter basis datastructure
• Support for new filter basis types
• Added Zernike polynomial basis on a disk
• Added Morlet wavelet basis functions on a spherical disk
• Cleaning up the SFNO example and adding new Local Spherical Neural Operator model
• Updated resampling module to extend input signal to the poles if needed
• Added slerp interpolation to the resampling module
• Added distributed resampling module

v0.7.3

• Changing default grid in all SHT routines to equiangular
• Hotfix to the numpy version requirements

v0.7.2

• Added resampling modules for convenience
• Changing behavior of distributed SHT to use dim=-3 as channel dimension
• Fixing SHT unittests to test SHT and ISHT individually, rather than the roundtrip
• Changing the way custom CUDA extensions are handled

v0.7.1

• Hotfix to AMP in SFNO example

v0.7.0

• CUDA-accelerated DISCO convolutions
• Updated DISCO convolutions to support even number of collocation points across the diameter
• Distributed DISCO convolutions
• Fused quadrature into multiplication with the Psi tensor to lower memory footprint
• Removed DISCO convolution in the plane to focus on the sphere
• Updated unit tests which now include tests for the distributed convolutions

v0.6.5

• Discrete-continuous (DISCO) convolutions on the sphere and in two dimensions
• DISCO supports isotropic and anisotropic kernel functions parameterized as hat functions
• Supports regular and transpose convolutions
• Experimental accelerated spherical DISCO convolutions on GPU via Triton Kernel
• Unittests for DISCO convolutions in tests/test_convolution.py

v0.6.4

• reworking distributed to allow for uneven split tensors, effectively removing the necessity of padding the transformed tensors
• distributed SHT tests are now using unittest. Test extended to vector SHT versions. Tests are defined in torch_harmonics/distributed/distributed_tests.py
• base pytorch container version bumped up to 23.11 in Dockerfile

v0.6.3

• Updated SFNO example
• Fixes NCCL issues with distributed SHT
• other bugixes