polybench-rs

Rust port of PolyBench/C - a collection of numerical computations from various domains.

This reimplementation was primarily motivated by benchmarking LLVM Polly as applied to Rust.

Usage

Individual benchmarks can be run with:

cargo run --bin <benchmark> --release

Configuration

The benchmarks are implemented as generic functions that accept the problem size as const generics. The data type to the benchmarks can be modified in src/config.rs.

LLVM flags can be set through the RUSTFLAGS environment variable:

set RUSTFLAGS=-Cllvm-args=--polly -Cllvm-args=--polly-vectorizer=stripmine

Results

The Polly optimization shows improvement on the correlation, covariance, gemm, gemver, trmm, 2mm, 3mm, gramschmidt and mvt.

Performance on gesummv, atax, bicg, durbin, deriche, fdtd_2d and jacobi_2d decreased. gesummv, atax, bicg and jacobi_2d are particularly bad, being about twice as slow with or without vectorization. ludcmp and jacobi_1d also have worse performance but only on vectorized builds. floyd_warshall is initially slowed but recovers on increased input sizes.

Benchmarks that show no significant (less than 5%) change includes symm, syr2k, syrk, doitgen, cholesky, lu, nussinov, adi, heat_3d and seidel_2d.

trisolv does not show any clear trend.

Data size dependency

Repeated here are the benchmarks that showed sensitivity to the input size. The speedups on correlation, covariance, gemm, trmm, 2mm, 3mm, gramschmidt and mvt increases with increasing input size indicative of better cache utilization compared to the baseline. floyd_warshall behaves more like a cutoff where below a certain size, Polly causes a regression and after that is parity with the default optimization.