We provide wrappers for linear algebra routines that allow to pre-allocate memory for repeated executions of the same operations. For convenience, we also provide functors, that contain the necessary memory and can be called with no further allocations.
Running the code
using KWLinalg
using BenchmarkTools
m, n = 5, 3
dtype = Float64
A = rand(dtype, m, n)
AC = deepcopy(A)
f = svd_functor_divconquer(m, n, Float64)
function copy_and_svd_inplace!(A, AC, f)
AC .= A
f(AC)
return nothing
end
@benchmark $copy_and_svd_inplace!($A, $AC, $f)leads to following result:
BenchmarkTools.Trial: 10000 samples with 9 evaluations.
Range (min … max): 2.612 μs … 8.526 μs ┊ GC (min … max): 0.00% … 0.00%
Time (median): 2.659 μs ┊ GC (median): 0.00%
Time (mean ± σ): 2.672 μs ± 193.544 ns ┊ GC (mean ± σ): 0.00% ± 0.00%
▄▇█▆▁
▂▂▂▂▂▂▂▂▃▄▅██████▆▃▂▂▂▂▂▁▂▂▁▁▁▁▁▂▁▁▁▂▂▁▁▂▂▂▂▂▁▂▂▂▁▁▁▁▂▂▁▂▂▂ ▃
2.61 μs Histogram: frequency by time 2.81 μs <
Memory estimate: 0 bytes, allocs estimate: 0.In contrast, using the LinearAlgebra function svd! as in
using LinearAlgebra
using BenchmarkTools
m, n = 5, 3
dtype = Float64
A = rand(dtype, m, n)
AC = deepcopy(A)
function copy_and_svd!(A, AC)
AC .= A
svd!(AC)
return nothing
end
@benchmark $copy_and_svd!($A, $AC)leads to following result:
BenchmarkTools.Trial: 10000 samples with 8 evaluations.
Range (min … max): 3.101 μs … 96.520 μs ┊ GC (min … max): 0.00% … 94.23%
Time (median): 3.154 μs ┊ GC (median): 0.00%
Time (mean ± σ): 3.297 μs ± 1.473 μs ┊ GC (mean ± σ): 0.75% ± 1.64%
▃▇█▇▅▃▁ ▁▁▁▁▁▂▃▂▂▂▂▃▃▂▁▂▂▂▁▁ ▂
███████▄▄▄▁▅▁▅▇▇▅▇████████████████████████▇▆▅▆▅▅▆▅▄▅▆▆▆▆▆▆ █
3.1 μs Histogram: log(frequency) by time 4.14 μs <
Memory estimate: 2.45 KiB, allocs estimate: 6.The time-difference is mostly due to the overhead of garbage collection. KWLinalg provides a benefit, when linear algebra operations are performed on dense matrices of the same size for potentially millions of times (e.g. within an optimization loop).
KWLinalg can be installed via Pkg:
using Pkg
Pkg.add(url="https://github.com/Algopaul/KWLinalg.jl.git")For a detailed description of the package and its functionality, we refer to the documentation.