Fluid Solver Optimization (CPU, OpenMP, CUDA)

Results

• CPU$30.85\,\mathrm{s} \to 3.69\,\mathrm{s}$ ($8.36\times$) via loop reordering + cache tiling + compiler/ILP optimizations.
• OpenMPpeak performance around $16\text{–}24$ threads ($\approx 1.38\,\mathrm{s}$, $\approx 10.77\times$); analyzed plateau from contention/sync effects.
• CUDAlin_solve kernel dominated ($\approx 89.6\%$); achieved $\approx 90.16\,\mathrm{GiB/s}$ and minimized transfer overhead with persistent device allocations.

Overview

My work

• Identified hotspots with gprof/perf (lin_solve as primary target) and set up repeatable benchmarks across grid sizes.
• Implemented CPU-side locality/ILP improvements (loop order, tiling, hoisted reciprocals) and validated speedups.
• Parallelized solver loops with OpenMP (parallel for + collapse; reductions for convergence metrics) and evaluated strong scaling.
• Implemented CUDA kernels for solver steps (including Red/Black Gauss-Seidel updates) and tuned block geometry for coalescing (e.g., $32 \times 4 \times 1$).
• Used Nsight Systems/Compute to attribute time, verify memory behavior, and document trade-offs (kernel launch + sync constraints).

Technologies

• C/C++
• OpenMP
• CUDA
• Linux
• perf
• gprof
• Nsight Systems
• Nsight Compute

Source material