RiemannEngine

A native C++20 library for sparse 3D deep learning on LibTorch and CUDA. It covers sparse convolutional networks, adaptive (mixed-resolution) voxelization and differentiable 3D Gaussian splatting, and needs no Python.

Get started Download a release API reference

---

What’s inside

Area	Highlights	Guide
Sparse tensors	SparseTensor / TensorField with batched integer coordinates, coordinate manager, quantization modes, dense/sparse conversion	Sparse tensors & layers
Sparse networks	Convolution, transposed and generative convolution, pooling, batch/instance norm, about 30 activations, residual and squeeze-excite blocks, broadcast and union ops	Sparse tensors & layers
Adaptive voxelization	Mixed-resolution octree-style leaves (AdaptiveSparseTensor), refinement by region/density/variance/predicate, adaptive convolution and pooling, GPU sphere voxelizer	Adaptive voxelization
3D Gaussian splats	GaussianTensor / GaussianModel, PLY I/O, SH degree 0-3, fused CUDA tile rasterizer, a 3DGS-style trainer with densification and pruning	Gaussian splats
4D Gaussian splats	Gaussian4DModel with linear motion and temporal opacity, Gaussian4DTrainer, D-NeRF loading, full-data reconstruct4d runner, .4dgs export, Python orbit viewer and looping GIFs	4D Gaussian splats
Training	Optimizer factory, sparse losses, datasets and collation, gradient accumulation, LR scheduling, model and tensor persistence	Training

Where to start

1. Installation: install a prebuilt MSI or .deb, or use the Docker image.
2. Quick start: build and run your first program in a few minutes.
3. Docker: develop and deploy in the GPU-ready container, with nothing installed on the host.
4. Core concepts: coordinates, tensor strides and the coordinate manager. Read this before writing networks.
5. Examples: complete classification, segmentation, adaptive and Gaussian programs.

Supported platforms

OS	CUDA	LibTorch	Packages
Windows 10/11 x64	13.2, 13.4	2.15.0	.msi, .zip
Ubuntu 22.04 x64	12.4	2.5.1 (CXX11 ABI)	.deb, .tar.gz, Docker cu124
Ubuntu 24.04 x64	13.2, 13.4	2.15.0	.deb, .tar.gz, Docker cu132 / cu134

See Installation for the details of each variant.

A taste

#include <riemann.hpp>
#include <iostream>

int main() {
    auto coords = riemann::batched_coordinates(
        {torch::randint(0, 32, {1000, 3}, torch::kInt32)}, torch::kInt32, torch::kCUDA);
    riemann::SparseTensor x(torch::rand({coords.size(0), 3}, torch::kCUDA), coords);

    riemann::RiemannConvolution conv(3, 32, /*kernel_size=*/3, /*stride=*/2);
    conv->to(torch::kCUDA);
    auto y = conv->forward(x);           // stride-2 sparse convolution on the GPU
    std::cout << y.F().sizes() << "\n";  // [M, 32]
}