Skip to content

[gsplat4d] add a physics-aware 4D Gaussian splatting benchmark - #324

Open
zjin-lcf wants to merge 2 commits into
masterfrom
gsplat4d
Open

[gsplat4d] add a physics-aware 4D Gaussian splatting benchmark#324
zjin-lcf wants to merge 2 commits into
masterfrom
gsplat4d

Conversation

@zjin-lcf

Copy link
Copy Markdown
Collaborator

Summary

  • Add a new gsplat4d benchmark (CUDA, HIP, SYCL, OpenMP) implementing the forward simulation and rendering pipeline from Physics-Aware 4D Dynamic Endoscopic Scene Simulations via MLLM-Guided Gaussian Splatting (https://arxiv.org/abs/2605.16022).
  • Three independently timed and validated stages: MLS-MPM particle advection with neo-Hookean stress, 4D Gaussian conditioning with EWA projection and spherical harmonics shading, and tile-based alpha-blending rasterization.
  • Each stage is verified against a host reference implementation; the benchmark uses a procedurally generated scene (no external data dependencies).
  • Add gsplat4d to the Simulation category in README.md and to src/CMakeLists.txt.

Test plan

  • CUDA: compiled and verified on NVIDIA GPU (all three stages PASS)
  • HIP: compiled and verified on AMD MI300A with ROCm 7.2.4 (all three stages PASS)
  • OpenMP: compiled and verified on AMD MI300A with ROCm 7.2.4 (linker warnings from nested team-local memory but runs correctly, all three stages PASS)
  • SYCL: compiled and verified (all three stages PASS)
  • Production workload: 200,000 Gaussians, 1280×720 image, 100 repetitions

Made with Cursor

zjin-lcf and others added 2 commits August 21, 2026 16:41
The benchmark follows the pipeline of the physics-aware endoscopic
simulators built on 4D Gaussian splatting and the material point method,
and times its three stages separately:

  1. an MLS-MPM step (p2g, grid update, g2p) that advects the Gaussians
     as neo-Hookean material particles;
  2. the 4D preprocess, which conditions every 4D Gaussian on the render
     time, projects the conditional 3D covariance to a 2D conic with the
     EWA Jacobian and evaluates the degree 3 spherical harmonics;
  3. the tile rasterizer, which alpha blends each 16x16 tile front to
     back with early termination.

The scatter of the particle-to-grid stage accumulates in a block local
grid tile rather than in the global grid, which needs the particles
binned by cell block; the binning and the per tile Gaussian lists are
host side setup shared by every backend, so all four see identical
input. Every stage is verified against a host reference.

Verified on a Tesla M40 (cuda), an MI100 (hip, omp) and an Intel GPU
(sycl, omp), plus sycl on the cuda backend.

Co-authored-by: Cursor <cursoragent@cursor.com>
…ence

Update the gsplat4d benchmark implementations (CUDA, HIP, SYCL, OpenMP)
with production-size workload parameters, CMake test configuration, and
individual stage PASS/FAIL validation. Add gsplat4d to the Simulation
category in README.md with the full paper reference.

Co-authored-by: Cursor <cursoragent@cursor.com>

Copilot AI left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

Adds a physics-aware 4D Gaussian splatting simulation benchmark across CUDA, HIP, OpenMP, and SYCL.

Changes:

  • Implements MLS-MPM, Gaussian preprocessing, and tile rasterization.
  • Adds procedural inputs and host-reference validation.
  • Registers the benchmark in CMake and documentation.

Reviewed changes

Copilot reviewed 17 out of 17 changed files in this pull request and generated 5 comments.

Show a summary per file
File Description
README.md Documents the benchmark.
src/CMakeLists.txt Registers gsplat4d.
src/gsplat4d-cuda/CMakeLists.txt Configures CUDA build/test.
src/gsplat4d-cuda/Makefile Adds CUDA Make build.
src/gsplat4d-cuda/main.cu Implements CUDA benchmark.
src/gsplat4d-cuda/reference.h Provides shared scene generation and references.
src/gsplat4d-hip/CMakeLists.txt Configures HIP build.
src/gsplat4d-hip/Makefile Adds HIP Make build.
src/gsplat4d-hip/main.cu Implements HIP benchmark.
src/gsplat4d-omp/CMakeLists.txt Configures OpenMP build.
src/gsplat4d-omp/Makefile Adds Intel OpenMP build.
src/gsplat4d-omp/Makefile.aomp Adds AOMP build.
src/gsplat4d-omp/Makefile.nvc Adds NVIDIA HPC build.
src/gsplat4d-omp/main.cpp Implements OpenMP benchmark.
src/gsplat4d-sycl/CMakeLists.txt Configures SYCL build.
src/gsplat4d-sycl/Makefile Adds SYCL Make build.
src/gsplat4d-sycl/main.cpp Implements SYCL benchmark.

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

Comment thread src/gsplat4d-cuda/main.cu
Comment on lines +631 to +634
const int n = atoi(argv[1]);
const int width = atoi(argv[2]);
const int height = atoi(argv[3]);
const int repeat = atoi(argv[4]);
Comment thread src/gsplat4d-hip/main.cu
Comment on lines +631 to +634
const int n = atoi(argv[1]);
const int width = atoi(argv[2]);
const int height = atoi(argv[3]);
const int repeat = atoi(argv[4]);
Comment thread src/gsplat4d-omp/main.cpp
Comment on lines +623 to +626
const int n = atoi(argv[1]);
const int width = atoi(argv[2]);
const int height = atoi(argv[3]);
const int repeat = atoi(argv[4]);
Comment on lines +610 to +613
const int n = atoi(argv[1]);
const int width = atoi(argv[2]);
const int height = atoi(argv[3]);
const int repeat = atoi(argv[4]);
SOURCES main.cu
CATEGORIES simulation
TEST_ARGS 200000 1280 720 100
TEST_REGEX "Rasterizer: PASS"
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants