Ambient Occlusion Benchmark

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.

Instructions

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}

To build and run the Intel SPMD Program Compiler (ISPC) version:

> cd volta && make
> ao 800 600

You can download ISPC for free for all major architectures from the ISPC website.

Results

On a dual core AVX1 i5 @1.8 GHz:

800 x 600	time [ms] Rust	speedup [-] Rust-vs-Rust	time [ms] ISPC 1.9.2	speedup [-] ISPC-vs-Rust
scalar	6266	1.0x	4976	clang7: 1.3x
vector	2369	2.6x	1157	2.0x
scalar_par	2443	2.5x	-	-
vector_par	983	6.4x	454	2.2x

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

800 x 600	time [ms] Rust	speedup [-] Rust-vs-Rust	time [ms] ISPC 1.9.2	speedup [-] ISPC-vs-Rust
scalar	3234	1.0x	3536	clang7: 0.9x
vector	1096	3.0x	525	2.1x
scalar_par	132	24.5x	-	-
vector_par	76	42.6x	20	3.8x

4096 x 4096	time [ms] Rust	speedup [-] Rust-vs-Rust	time [ms] ISPC 1.9.2	speedup [-] ISPC-vs-Rust
scalar	116121	1.0x	127161	clang7: 0.9x
vector	40076	2.9x	19707	2.0x
scalar_par	3273	35.5x	-	-
vector_par	1398	83.1x	644	2.2x

Overview

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.