Experience of SYCL-Based Development on Granular Material Large Deformation

Authors: Shuang Li, Xiangyu Hu

Introduction

Aim

This project focuses on heterogeneous SPH simulation of granular materials, particularly leveraging GPU acceleration.

Development Route

We rewrote the existing plasticity-related functions in SYCL to achieve efficient GPU computation.

What You Can Gain

  • Access to an open-source heterogeneous SPH simulation framework for granular materials.
  • Insights into SYCL-based development, including errors I met and solutions.

New Test Cases

1. test_2d_column_collapse_sycl and test_2d_column_collapse_ck

2. test_3d_repose_angle_sycl and test_3d_repose_angle_ck

Note: These cases are SYCL-based implementations of their original versions.

GPU Acceleration Efficiency

For a simulation with 1.6 million particles (dp = 0.0015 m), the computation time for 0.1s is:

Device Time (with ParticleSort) Time (without ParticleSort)
CPU (Intel Xeon Platinum 8377C) ~40 min N/A
GTX 2080Ti ~16 min N/A
GTX 1030 ~210 min ~280 min

Conclusion: GPU acceleration significantly outperforms CPU computation, making it crucial for large-scale field simulations.

Development Experience

(1) Variable Initialization Order vs. Definition Order

Issue: The initialization order of variables differed from their definition order, leading to potential bugs.

Solution: Add the following flag in CMAKEPRESETS.json to enforce correct ordering:

CMAKE_CXX_FLAGS: “-Werror=reorder”

(2) Build Error on Linux in GitHub CI

Issue: The test passes on Windows but fails on Linux.

Solution: The GCC compiler on Linux enforces stricter rules and treats warnings as errors. To ensure compatibility, first build the project locally using g++ and resolve any warnings before pushing to GitHub.

(3) Math Function Linking Errors in SYCL

Issue: Functions like sin() caused linking errors in SYCL computations.

Solution: Use SYCL’s built-in math functions explicitly.

Modification in base_data_type.h

if SPHINXSYS_USE_FLOAT

using Real = float;
using UnsignedInt = u_int32_t;
namespace math = sycl;

else

using Real = double;
using UnsignedInt = size_t;
namespace math = std;

endif

Usage in Code

Real sin_x = math::sin(x);

(4) Regression Testing on GitHub

Issue: GitHub CI/CD only supports CPU-based testing, but the project requires SYCL-based regression tests.

Solution:

  1. Generate the regression test database using CPU.

  2. Modify CMAKEPRESETS.json by removing the following line:

    “SPHINXSYS_SYCL_TARGETS”: “nvptx64-nvidia-cuda”

  3. With SPHINXSYS_USE_SYCL="ON", the SYCL functions will still be executed, but on the CPU instead of GPU.

Conclusion

This work demonstrates the feasibility and advantages of SYCL-based GPU acceleration in granular material simulations.