search

JuliaLang / LinearAlgebra.jl

Julia Linear Algebra standard library
Other
17 stars 4 forks source link

Performance regressions in linear algebra benchmarks with Hermitian and Triangular matrices #1077

Open Zentrik opened 5 months ago

Zentrik commented 5 months ago

I haven't been able to replicate locally so can't bisect. But the regression is persistent on NanoSoldier so likely real. The bad commit should be one of these ones, https://github.com/JuliaLang/julia/compare/16d70b53f3e2493310b1bca61b3dcc861284a248...b8e714dfa3d81edd56ccda853322b6e42d94cf60.

You can find the benchmark data at https://tealquaternion.camdvr.org/compare.html?name=%28Hermitian%7CTriangular%29&regex=true&start=16d70b53f3e2493310b1bca61b3dcc861284a248&stat=min-wall-time&end=b8e714dfa3d81edd56ccda853322b6e42d94cf60 and below. These are the min wall time results,

Summary

Range Mean Count
Regressions 3.24%, 78.76% 34.89% 11

Benchmarks

Benchmark % Change Significance Factor
linalg.arithmetic.(exp, typename(LinearAlgebra.Hermitian), 1024) 78.76% 72.72x
linalg.arithmetic.(log, typename(LinearAlgebra.Hermitian), 1024) 78.24% 46.84x
linalg.arithmetic.(sqrt, typename(LinearAlgebra.Hermitian), 1024) 77.74% 48.91x
linalg.arithmetic.(*, typename(LinearAlgebra.UpperTriangular), Vector, 1024) 27.45% 1.21x
linalg.arithmetic.(sqrt, typename(LinearAlgebra.Hermitian), 256) 27.23% 74.58x
linalg.arithmetic.(exp, typename(LinearAlgebra.Hermitian), 256) 27.19% 66.60x
linalg.arithmetic.(log, typename(LinearAlgebra.Hermitian), 256) 27.02% 95.91x
linalg.arithmetic.(\, typename(LinearAlgebra.UpperTriangular), Vector, 1024) 25.86% 1.39x
linalg.arithmetic.(\, typename(LinearAlgebra.LowerTriangular), Vector, 256) 5.82% 1.71x
linalg.arithmetic.(\, typename(LinearAlgebra.UpperTriangular), Vector, 256) 5.22% 1.05x
linalg.arithmetic.(*, typename(LinearAlgebra.LowerTriangular), Vector, 256) 3.24% 1.52x