mooreniemi / dot_product

Intel MKL vs Rust

Test Dot Product in C++, MKL and Rust

Note from Alex:

I've been working on the Rust side in this fork, porting it to current library versions.

The original implementation tested one extremely large vector dot product, which is not really representative since the horizontal sum is only done once. In practice, vectors tend to be maybe 32-1024 dim. A more representative benchmark then is to run on many vectors of (say) dim=256.

For dim=256 and n=1_000_000, the output on my XPS13 was:

Rust: (naive/brute force)
printf: 254.178934319s seconds
Rust: (SIMD f64x4)
printf: 183.175553805s seconds
Rust: (SIMD f64x8)
printf: 161.44382163s seconds
Rust: (parallelism)
printf: 720.338301011s seconds
Rust: (parallelism + SIMD)
printf: 711.990750934s seconds

This worse performance actually makes more sense because we're not doing the overall dot products in parallel, just the operations of each dot product in parallel.

This is retained in the bin "original". Further changes happened in the bin "fork". Note that indicatif progress bar effects timings, so a verbose flag is at the top of "fork". The below had dim=256, n=100_000.

# indicatif on
Rust: (naive/brute force)
printf: 18.10648789s
Rust: (SIMD f64x8)
printf: 17.184435915s
Rust: (par_iter parallelism + SIMD f64x8)
printf: 21.560859663s

# no indicatif on
Rust: (naive/brute force)
printf: 1.002513099s
Rust: (SIMD f64x8)
printf: 733.738948ms
Rust: (par_iter parallelism + SIMD f64x8)
printf: 309.76286ms

Fork, with blas ndarray:

Rust: (naive/brute force)
took: 1.077836212s
Rust: (SIMD f64x8)
took: 686.731468ms
Rust: (par_iter parallelism + SIMD f64x8)
took: 421.287732ms
Rust: (ndarray)
took: 108.83656ms
Rust: (par_iter parallelism + ndarray)
took: 113.398612ms

ndarray WITHOUT blas:

Rust: (ndarray)
took: 1.120636321s
Rust: (par_iter parallelism + ndarray)
took: 453.773839ms

Requirements:

• Visual Studio 2017
• Intel C++ compiler
• Intel MKL
• Armadillo
• Eigen
• Rust

Performance Results (in seconds):

Programs	First Time	Second Time	Third Time	Avg. Time
C++ + MKL	0.0313977	0.0302247	0.0303358	0.030652733
C++ + Armadillo + MKL	0.030196	0.0317884	0.0320485	0.0313443
C++ + Eigen + MKL	0.868	0.849	0.878	0.865
Rust	0.088117961	0.095113075	0.094697091	0.092642709
Rust (SIMD with f64x4)	0.075088962	0.075810549	0.069293656	0.073397722
Rust ( SIMD with f64x8)	0.068858522	0.075586333	0.075168754	0.073204536
Rust: (parallism)	0.035086217	0.035334371	0.035818443	0.03541301
Rust: (parallism + SIMD)	0.025660633	0.025662228	0.025888573	0.025737145