torch-fps 0.1.2

Native PyTorch farthest point sampling for point cloud workloads

torch-fps

Optimized standard farthest point sampling (FPS) for PyTorch written in C++.

pip install torch-fps

Note: Ensure gcc > 9 and < 14.

Usage

import torch
from torch_fps import farthest_point_sampling, farthest_point_sampling_with_knn

# Create example inputs
points = torch.randn(4, 1000, 3)     # [B, N, D] - batch of point clouds
mask = torch.ones(4, 1000, dtype=torch.bool)  # [B, N] - valid point mask
K = 512  # Number of samples per batch (must be <= number of valid points)

# Perform farthest point sampling
idx = farthest_point_sampling(points, mask, K)  # [B, K] - selected point indices

# Use indices to gather sampled points
sampled_points = points.gather(1, idx.unsqueeze(-1).expand(-1, -1, 3))  # [B, K, D]

# Fused FPS + kNN: get centroids and their k nearest neighbors in one pass
centroid_idx, neighbor_idx = farthest_point_sampling_with_knn(
    points, mask, K=512, k_neighbors=32
)  # centroid_idx: [B, K], neighbor_idx: [B, K, k_neighbors]

Performance

Benchmarked on AMD Threadripper 7970X and NVIDIA RTX 5090. Values show CPU / CUDA measurements.

FPS:

B	N	K	Baseline (ms)	Optimized (ms)	Speedup
4	100	20	0.45 / 1.40	0.05 / 0.24	8.5x / 5.9x
8	512	64	2.85 / 4.04	0.66 / 0.31	4.3x / 13.0x
16	1024	128	33.31 / 7.78	4.52 / 0.59	7.4x / 13.1x
32	2048	256	158.18 / 15.56	33.18 / 1.66	4.8x / 9.4x

FPS+kNN:

B	N	K	k	Baseline (ms)	Optimized (ms)	Speedup
4	100	16	8	0.43 / 1.21	0.08 / 0.24	5.4x / 5.0x
8	512	64	16	4.98 / 4.07	2.16 / 0.33	2.3x / 12.3x
16	1024	128	16	39.00 / 7.96	12.00 / 0.61	3.3x / 13.0x
32	2048	256	16	180.14 / 16.81	76.60 / 1.36	2.4x / 12.4x

Implementation

Farthest Point Sampling

Standard greedy algorithm maintaining minimum distances to selected centroids. Each iteration selects the point farthest from all previously selected points.

• CPU: Sequential selection with parallel batch processing. O(K·N·D) time, O(N) space.
• CUDA: Cooperative parallel reduction within thread blocks. O(K·N·D) time, O(N) space per batch.

Fused FPS + k-Nearest Neighbors

Combines FPS and kNN by reusing distance computations from the FPS phase.

• CPU: Incremental heap tracking during FPS. Maintains top-k neighbors per centroid using max-heaps. O(K·N·log(k)) time, O(K·k) space.
• CUDA: Stores all centroid distances during FPS, then applies PyTorch's optimized topk. O(K·N·D + K·N·log(k)) time, O(K·N) space per batch.

Both implementations eliminate redundant distance calculations compared to separate FPS and kNN operations.