Skip to content

remblais

remblais logo

remblais computes geometric losses between measures on NVIDIA (CUDA), AMD (ROCm/HIP) and Intel (SYCL) GPUs: Sinkhorn divergences (entropic optimal transport), kernel norms (MMD) and Hausdorff divergences between point clouds, images and volumes. It is a drop-in replacement for GeomLoss 0.3.1, built on kernox instead of PyKeOps: write import remblais as geomloss and your code runs unchanged.

Why remblais

  • Same API as GeomLoss. SamplesLoss, sinkhorn_divergence, ImagesBarycenter and the ot module have the GeomLoss 0.3.1 signatures, calling conventions and output shapes. With backend="tensorized", values and gradients match GeomLoss bit for bit, apart from the documented fixes.
  • Faster. Every softmin is one fused kernox reduction, compiled once and reused for all values of ε; the debiased Sinkhorn loop fuses its four softmins into two reductions per iteration.
  • NVIDIA, AMD and Intel GPUs. The online and multiscale backends, which GeomLoss runs through PyKeOps on NVIDIA GPUs only, run here on CUDA, ROCm/HIP and SYCL through kernox. remblais's own test suite runs on NVIDIA and Intel GPUs; on AMD GPUs it relies on kernox's tested HIP backend.
  • Fixes, not surprises. A few GeomLoss bugs are fixed (the gradient of unbalanced transport, "hausdorff" without a kernel, ...). Every difference is listed in Coming from GeomLoss.

Installation

remblais needs Python 3.10 or later, PyTorch and kernox. Install PyTorch for your platform, then kernox with the extra that matches your GPU, then remblais, both from PyPI. For an NVIDIA GPU:

pip install torch --index-url https://download.pytorch.org/whl/cu126
pip install "kernox[cuda]"
pip install remblais

For AMD and Intel GPUs, install PyTorch from the rocm or xpu index and kernox with the rocm or xpu extra. Intel GPUs also need a SYCL runtime matching the host icpx compiler, see Intel GPUs. The details are in the kernox installation guide and in Getting started.

Performance

The goal of remblais is to be at least as fast as GeomLoss wherever GeomLoss runs. In the cases measured so far, it is already faster.

Sinkhorn divergence, p = 2, FP32, forward and backward passes, on an RTX 2080 Ti; remblais with fast_math, its default setting, against GeomLoss 0.3.1 in its best configuration (each PyKeOps kernel compiled alone in a fresh process). Time ratios remblais / GeomLoss, lower than 1 when remblais is faster:

Case N = 10⁴ N = 10⁵ N = 10⁶
multiscale, D = 2 ×0.45 ×0.40 ×0.79
multiscale, D = 3 ×0.26 ×0.92 ×0.96
online, D = 2 ×0.33 ×0.84 ×0.89
online, D = 3 ×0.31 ×0.85 ×0.90
online, D = 64 ×0.45 ×0.54 not measured

For small clouds (N = 10³ to 3·10⁴), the online backend runs at ×0.31 to ×0.61 of the GeomLoss time in D = 2 and 3, and at ×0.17 to ×0.45 in D = 64.

These figures cover only the cases listed; they are not a general claim of superiority over GeomLoss on every workload or card. The full tables and the protocol are on the Performance page.

In 1781, Gaspard Monge asked how to move a pile of earth, the déblais, into a hole of the same volume, the remblais, at the lowest cost: the founding problem of optimal transport. remblais is where the mass ends up. The logo shows the ox of symox and kernox, harnessed this time to a scraper, carrying the déblais to the remblais: Monge's optimal transport.

License

At your option, MIT or Apache-2.0. Code derived from GeomLoss stays under the MIT license of GeomLoss; its provenance and notice are in THIRD_PARTY.md.