Originally written by Syoyo Fujita: https://github.com/syoyo/aobench

aoench is a small ambient occlusion renderer for benchmarking realworld floating point performance in various languages.

To run it with the default target options (replace ${NAME} with an algorithm name):

> cargo run --release -- 800 600 --algo ${NAME}

Use RUSTFLAGS to set the target CPU, for example:

> RUSTFLAGS="-C target-cpu=native" cargo run --release -- 800 600 --algo ${NAME}

./benchmark.sh

On a dual core AVX1 i5 @1.8 GHz:

tiled_par is 1.02x slower than ispc_tasks.

On a 28 core Xeon CPU E5-2690 v4 @ 2.60GHz:

tiled_par is 1.53x faster than ispc_tasks.

On a 40 core Xeon Gold 6148 CPU @ 2.40GHz:

tiled_par is 1.5x faster than ispc_tasks.

There are 4 main pieces in the aobench benchmark:

• ray-plane intersection algorithm: source
• ray-sphere intersection algorithm: source
• ambient occlusion algorithm: source
• ray-casting the pixels:
  • scalar serial: source
  • scalar parallel: source
  • vector serial: source
  • vector parallel: source

The scalar and vectorized implementations of the intersection and ao algorithms are in the same file so that they can be easily compared.

As a comparison, the ISPC sources of the same benchmark are here.