English | 中文
6th-place (bronze medal) solution of the 2024 OpenAtom Open Source Contest — Solving Large-Scale Asymmetric Indefinite Banded Systems of Linear Equations.
An ILU(0)-preconditioned stabilized bi-conjugate gradient method (BiCGStab), where the incomplete LU factorization is computed via dependency analysis.
The entire framework is implemented from scratch in C/C++ with CUDA kernels at every algorithmic bottleneck — no vendor library hot path is reused:
- Dependency-analysis-driven ILU(0) factorization for the preconditioner
- Custom triangular solves for large sparse L/U systems
- 10–20× beyond common baseline implementations on key kernels
- The L-shaped sparse triangular solve reaches ~50× cuSPARSE (NVIDIA CUDA Toolkit) within this contest's scenario
GPU: NVIDIA V100 32 GB · CPU: 8-core Xeon · RAM: 64 GB
mkdir build && cd build
cmake ../
make
./megasolve <data_dir> <matrix_file> <rhs_file> <init_guess_file> \
<num_grids> <num_nonzeros> <output_file> [max_iterations]