MicroGPUBench

A cross-platform FP32 compute micro-benchmark for OpenCL and Vulkan.

• make

• OpenCL headers and libOpenCL.so
• Vulkan headers and libvulkan.so
• glslangValidator (for Vulkan shader build)

• ./opencl_bench --gpu
• ./opencl_bench --cpu
• ./opencl_bench --work-items=1048576 --iters=1048576 --chunk-iters=4096 --accum=4 --runs=5 --warmup=3

• ./vulkan_bench
• ./vulkan_bench --device=0 --work-items=1048576 --iters=1048576 --chunk-iters=4096 --accum=4 --runs=5 --warmup=3
• ./vulkan_bench --shader=path/to/custom.spv

• Work items are rounded up to a multiple of the local size.
• The shader uses 4 accumulators by default (ACCUM=4). Each iteration is one fused multiply-add per accumulator, counted as 2 FLOPs each.

• OpenCL builds with fast-math flags enabled for peak throughput.
• Shader compilation uses basic SPIR-V generation flags for compatibility across glslangValidator versions.
• chunk-iters splits the total iterations into smaller dispatches to avoid GPU watchdog timeouts and device loss.
• Use --accum to select 1-1024 accumulators per work-item.