This repository is a scratchspace of mine to play around with some Rust functionality (and measure its impact on runtime). Note that none of this code is intended to be used in production somewhere.
It currently tests different ways to sum up the entries of a 1024x1024 matrix of type Vec<Vec<f32>> which represents a collection of row-vectors. It has three different approaches implemented in lib.rs: column-wise summation (which has very bad cache locality), row-wise summation (much better cache locality) and SIMD-accelerated row-wise summation. Interestingly there is a factor of 10x difference on the runtime (on my machine):
$ cargo bench
running 3 tests
test bench_colwise_summation ... bench: 8,297,295 ns/iter (+/- 886,357)
test bench_rowwise_simd_summation_sse41 ... bench: 851,027 ns/iter (+/- 28,224)
test bench_rowwise_summation ... bench: 3,377,313 ns/iter (+/- 180,663)
If we install cargo-profiler, we can get some statistics about the cache misses incurred by our different implementations. The numbers confirm what we expected: the column-wise access pattern results in 8 times more L1 cache misses:
$ cargo profiler cachegrind --bin ./target/release/colwise
Profiling colwise with cachegrind...
Total Memory Accesses...46,824,375
Total L1 I-Cache Misses...332 (0%)
Total LL I-Cache Misses...260 (0%)
Total L1 D-Cache Misses...1,128,715 (2%)
Total LL D-Cache Misses...68,207 (0%)
$ cargo profiler cachegrind --bin ./target/release/rowwise
Profiling rowwise with cachegrind...
Total Memory Accesses...40,419,121
Total L1 I-Cache Misses...372 (0%)
Total LL I-Cache Misses...298 (0%)
Total L1 D-Cache Misses...140,739 (0%)
Total LL D-Cache Misses...69,659 (0%)
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent