Published September 6, 2026 | Version v1.0.4

SCARF: A Scene-Adaptive Depth-Guided G-3DGS Encoder Accelerator with Semantic Reuse and Fused Dataflow

  • 1. Institute of Computing Technology, Chinese Academy of Sciences
  • 2. ROR icon University of Chinese Academy of Sciences
  • 3. School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences

Description

This artifact accompanies SCARF: A Scene-Adaptive Depth-Guided G-3DGS Encoder Accelerator with Semantic Reuse and Fused Dataflow, presented at the 59th IEEE/ACM International Symposium on Microarchitecture (MICRO 2026), Athens, Greece, October 31–November 4, 2026.

SCARF accelerates depth-guided generalizable 3D Gaussian Splatting encoders. Feature Similarity Depth Reuse (FSDR) exploits semantic similarity to narrow depth searches, while Scene-Adaptive Early Sparsification (SAES) uses local 3D continuity to select Gaussian generation paths. A fused dataflow integrates these mechanisms to reduce computation and data movement.

Artifact contents

  • Python functional and cycle-level simulators, Chisel RTL, and generated SystemVerilog.
  • Model adapters and evaluation workflows for TranSplat, MVSplat, and DepthSplat on Re10K, ACID, and DL3DV.
  • A Docker workflow, deterministic synthetic quick fixture, validators, frozen mechanism configuration, and provenance manifests.
  • Hardware evaluation scripts and a public physical-design proxy based on iFlow and ASAP7.

Getting started

Start with README.md and ARTIFACT_EVALUATION.md in the archive. The quick workflow uses x86-64 Linux, Docker with NVIDIA GPU support, and a CUDA 12.1-compatible GPU with at least 8 GB VRAM. Quality and mechanism evaluations benefit from at least 12 GB VRAM. The documented performance measurements use a Jetson Orin NX 16 GB in MAXN mode with locked clocks. RTL evaluation uses JDK 11 or newer, sbt 1.9 or newer, Chisel 6.6, and Verilator 5 or newer.

The archive includes the synthetic fixture. Third-party datasets and checkpoints are obtained separately through the documented upstream sources and download scripts. The original project code is distributed under Apache License 2.0, as specified in LICENSE. Third-party components retain their respective licenses, listed in THIRD_PARTY.md.

Paper and source repository

Zirui Ma, Zhihua Fan, Haibin Wu, Wenxing Li, Fulin Zhang, Dongrui Fan, Xiaochun Ye, and Wenming Li. SCARF: A Scene-Adaptive Depth-Guided G-3DGS Encoder Accelerator with Semantic Reuse and Fused Dataflow. MICRO 2026, pp. 2512–2528. DOI: 10.1109/MICRO67795.2026.00172.

Source repository: https://github.com/MAdrid1011/SCARF. Corresponding authors: Zhihua Fan and Wenming Li.

Acknowledgments

This work was supported by the National Key R&D Program of China (Grant No. 2023YFB4503500), the National Natural Science Foundation of China (Grant No. 62502498), the Beijing Natural Science Foundation (Grant No. L234078), and Innovation Funding of ICT, CAS (Grant No. E661120).

Files

Files (3.3 MB)

Name Size
md5:69474f967f5c83513c9c65783906778c
3.3 MB Download
md5:bf46f8a3263d9d0243ca01fd45a9d2b6
111 Bytes Download

Additional details

Related works

Is supplement to
Conference paper: 10.1109/MICRO67795.2026.00172 (DOI)
Is version of
Software: https://github.com/MAdrid1011/SCARF (URL)

Funding

National Key R&D Program of China
2023YFB4503500
National Natural Science Foundation of China
62502498
Beijing Natural Science Foundation
L234078
Institute of Computing Technology, Chinese Academy of Sciences
Innovation Funding of ICT, CAS E661120

Software

Repository URL
https://github.com/MAdrid1011/SCARF
Programming language
Python , Scala , SystemVerilog
Development Status
Active