CUDA-WalkOnSolver

Project Overview

1. CUDA implementation of Monte Carlo Geometry Processing

Included Features

1. GPU based Linear Bounding Volume Hierarchy, Closest point query
2. Laplace and Poisson PDE solver based on Walk on Sphere algorithm
3. Consideration of only Diriclet Boundary Condition
4. Parallel execution of 1's algorithm

Future works

• Other BVH method
• Other Walk on method (Walk on Star shape, Walk on Boundary)
• Shader excution reordering

Build dependencies

CMake, C++17, CUDA 11.8, Visual Studio 2019 (Windows 10, 11)

Third-party dependencies

tinyply, tiny-cuda-nn(reference), Polyscope (for visualiziation)

Demo & Usage

• Visualization of the tested solver by Polyscope is implemented in demo/wosolver.cu. Please refer this.

Windows

mkdir build
cd build && cmake ..
cmake -B build
.\build\test_wos.exe

• Define Boundary Condition and Source Term for PDE The functioncallity in this repo works by functor. Just define the interesting boundary condition and source term as below.

struct BoundaryCondition {
    HOST_DEVICE float operator()(vec3 x) {
        return cos(2.f*M_PI*x.x) * sin(2.f*M_PI*x.y);
    }
};

struct SourceTerm {
    HOST_DEVICE float operator()(vec3 x) {
        return (M_PI*M_PI) * cos(2.f*M_PI*x.x) * sin(2.f*M_PI*x.y);
    }
};

int main() {
    WoSolver<TriangleLBVH> solver(mesh.v(), mesh.f(), stream);

    Laplace<BoundaryCondition> laplace;
    Poisson<BoundaryCondition, SourceTerm> poisson;
    
    ... make cuda stream interesting points
    ...

    auto gpu_result = solver.estimate_features(stream, gpu_points.mut_span(), laplace, poisson);
}