Parallel Computing — CPU and GPU Optimization

Profiling and parallelizing a physics/graphics engine across CPU cores with OpenMP and the GPU with OpenCL, ending up with as much as a 34x speedup over the original.

Overview

The starting point was a single-threaded engine. From there I profiled to find where the time was actually going, then applied optimizations one layer at a time — compiler vectorization first, then multicore CPU parallelism with OpenMP, and finally a GPU port with OpenCL. Measuring after each step made it clear which changes actually paid off.

Optimization stages

1. Baseline: the single-threaded reference implementation
2. Compiler and SIMD: -O3 -ftree-vectorize -ffast-math, with AVX2
3. CPU multicore: OpenMP (-fopenmp) on the parallelizable loops
4. GPU: an OpenCL kernel (parallel.cl) offloading the hot path

Results

Version	Frametime (80x80 workload)	Speedup
Optimized CPU	537 ms	baseline
OpenCL GPU	69 ms	about 7.8x over the optimized CPU version
End to end	—	up to 34x over the original

Build

mkdir -p build && cd build
cmake ..
make
./parallel

Dependencies: OpenGL, GLUT, an OpenCL SDK, and an OpenMP-capable compiler.

Repository layout

├── parallel.c        host program (OpenMP plus OpenCL host code)
├── parallel.cl       OpenCL GPU kernel
├── CMakeLists.txt
└── build/

Skills demonstrated

OpenMP, OpenCL, GPU computing, SIMD and AVX2, performance profiling, CMake, C, and parallel algorithm design.