A benchmark comparing OpenGL Compute Shaders against OpenCL on the same hardware.

• Task: $1024 \times 1024$ Matrix Multiplication ($C = A \times B$).
• Hardware: Sony Vaio SVF15N2C5E (i7-4500U, Intel HD 4400, NVIDIA GT 735M).

1. OpenGL ≈ OpenCL on NVIDIA: When running both APIs on the GT 735M, performance is nearly identical (~5% difference), confirming that OpenGL Compute Shaders incur no significant penalty compared to OpenCL.

2. Mesa outperforms Beignet on Intel: OpenGL via Mesa (~227 ms) is 35% faster than OpenCL via Beignet (~307 ms) on the HD 4400. Beignet is a legacy unmaintained driver with known performance limitations.

3. GPU vs CPU: Both GPUs provide 17-36× speedup over the CPU for this compute-bound workload.

• The benchmark dynamically adjusts OpenCL work group sizes based on device limits (Beignet's HD 4400 is limited to 512 work items vs 1024 on NVIDIA).
• Rusticl (Mesa's modern OpenCL) requires the iris driver (Gen8+/Broadwell), so it's unavailable for Haswell.
• The legacy Intel OpenCL runtime (NEO) doesn't support Haswell on modern Linux.

• OpenCL Runtime:
  • nvidia-driver-470 (for NVIDIA)
  • beignet-opencl-icd (for Intel Haswell GPU)
  • pocl-opencl-icd (for CPU fallback)
  • ocl-icd-opencl-dev (Headers/Loader)

# Install dependencies sudo apt install ocl-icd-opencl-dev opencl-headers clinfo pocl-opencl-icd # Build cd examples/002_OpenCL cmake -B build -S . -DCMAKE_TOOLCHAIN_FILE=/path/to/vcpkg.cmake cmake --build build --config Release

Run on integrated Intel GPU (default):

./build/gpgpu_vs_cl

Run on NVIDIA (via PRIME offload):

__NV_PRIME_RENDER_OFFLOAD=1 __GLX_VENDOR_LIBRARY_NAME=nvidia ./build/gpgpu_vs_cl