Parallel GPU Optimization for Fluid Dynamic Solver

An optimized fluid solver originally implemented in C/C++ without parallelization.
The code was restructured to leverage GPU acceleration and other performance improvements, reducing runtime from several weeks to just a few seconds.

---

📄 Supporting Report

At the root of this repository, there is a PDF file containing a detailed report in Portuguese.
The report explains and justifies the entire GPU optimization process, including the modifications made to the original C/C++ code, performance improvements, and the testing methodology.
It also includes instructions on how to implement an OpenMP optimization on the original code.
The original code is available at 3DFluid GitHub repository.

📂 Where is the code?

The main implementation is located in the .cu files:

• main.cu – entry point of the program
• fluid_solver.cu – GPU-optimized fluid solver
• EventManager.cpp – event handling utilities

---

⚙️ Requirements

To build and run the solver you will need:

• gcc (GNU Compiler Collection)
• nvcc (NVIDIA CUDA Compiler)
• An NVIDIA GPU compatible with your chosen CUDA architecture

---

🚀 How to Build and Run

Inside the src/ directory there is a Makefile.
Then you can use make to compile, this will return executable under the name of fluid_sim. By default, it compiles the solver with the following command:

nvcc -O3 -arch=sm_35 -Wno-deprecated-gpu-targets -use_fast_math -maxrregcount=32 main.cu fluid_solver.cu EventManager.cpp -o fluid_sim

⚠️ Note: The flags can be modified depending on your GPU model. The provided configuration is tuned for NVIDIA Kepler GPUs.

Run locally

On Linux, you can run:

./fluid_sim

If you want to check execution time:

time ./fluid_sim

Run on SLURM

The Makefile also includes a target to submit jobs via SLURM. Use:

make run

This will call the run.sh script with sbatch.

🧹 Cleaning Up

To remove the compiled binary and clean the project, run:

make clean

This will delete the fluid_sim executable and any temporary build files.

📌 Notes

• The project was migrated from a sequential CPU version to a massively parallel GPU version.
• Performance gain: weeks → seconds depending on the simulation size and GPU hardware.
• CUDA compiler flags (-O3 -arch=sm_35 -use_fast_math -maxrregcount=32) can be adjusted for different GPU architectures to achieve optimal performance.
• The project includes a Makefile that simplifies compilation and execution, including support for SLURM job submission.