Optimization: Struct of Arrays (SoA)

Author: Linus
Predecessor: variant_inlining_7.cpp
File(s): variants/variant_inlining_7_SoA.cpp
Status: TBD

What this changes

Replaced the Pedestrian struct with parallel vector<double> arrays, one per field (pos_x, pos_y, e_x, e_y, sb_e_x, sb_e_y, sb_sq, ...). Board obstacles are left as AoS since there are only a handful of them.

Mostly a setup step for AVX2: with SoA the inner ped-ped loop reads contiguous chunks of pos_x[j..], which is what we'll want for aligned vector loads.

Code (relevant snippet)

vector<double> pos_x, pos_y;
vector<double> sb_sq, sb_e_x, sb_e_y;
vector<double> e_x, e_y;
vector<double> curr_vel_x, curr_vel_y;
// ...

See variants/variant_inlining_7_SoA.cpp.

Benchmark

Conditions: 2 runs, N_ITER=1000, -O3 -march=native -flto. Diff is vs variant_inlining_7. On my local machine i had an improvement of about 10-15% for low N and about 5% speed up for higher N

x86

Machine: TBD

Frequency: TBD

N	W(n) [Flops]	T(n) [Cycles]	Performance [Flops/Cycle]	Diff vs Predecessor
50
100
200
500
1000

ARM (Apple)

Machine: TBD

Frequency: TBD

N	W(n) [Flops]	T(n) [Cycles]	Performance [Flops/Cycle]	Diff vs Predecessor
50
100
200
500
1000

Analysis

No in depth Analysis done, just thought i wanna share it so you can also use it for further optimizations

Takeaway

Layout change, not a performance win by itself. Sets up the AVX2 variant.

2.5 KiB Raw Blame History