-
Notifications
You must be signed in to change notification settings - Fork 32
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
#1011 - Convert to Vector in polyhedron's LP #1013
Conversation
src/HPolyhedron.jl
Outdated
@@ -152,7 +152,7 @@ function σ(d::AbstractVector{N}, P::HPoly{N}) where {N<:Real} | |||
end | |||
|
|||
function σ_helper(d::AbstractVector{N}, P::HPoly{N}) where {N<:Real} | |||
c = -d | |||
c = convert(Vector{N}, -d) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Hm, this looks like it creates two vectors (I am not 100% sure, though).
julia> f(d) = convert(Vector{N}, -d);
julia> @code_lowered f(d)
CodeInfo(
1 1 ─ %1 = (Core.apply_type)(Main.Vector, Main.N) │
│ %2 = -d │
│ %3 = (Main.convert)(%1, %2) │
└── return %3 │
)
Since this is a critical function, I would care for such things. What about this:
julia> g(d) = [-x for x in d];
julia> @code_lowered g(d)
CodeInfo(
1 1 ─ %1 = (Base.Generator)(Main.:-, d) │
│ %2 = (Base.collect)(%1) │
└── return %2 │
)
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Consider these 3 ways:
using SparseArrays, BenchmarkTools
function f1(d::AbstractVector{N}) where {N}
return convert(Vector{N}, -d)
end
function f2(d::AbstractVector{N}) where {N}
return [-x for x in d]
end
function f3(d::AbstractVector{N}) where {N}
return Vector(-d)
end
(in julia v1.0.2) they give similar but not the same timings, especially in higher dims:
for n in [10, 100, 1000, 10000]
println("\n n = $n ")
d = sprandn(n, 0.1)
@btime f1($d)
@btime f2($d)
@btime f3($d)
end
n = 10
122.701 ns (4 allocations: 352 bytes)
64.208 ns (2 allocations: 176 bytes)
118.216 ns (4 allocations: 352 bytes)
n = 100
489.149 ns (4 allocations: 1.19 KiB)
495.479 ns (2 allocations: 912 bytes)
162.628 ns (4 allocations: 1.19 KiB)
n = 1000
11.803 μs (4 allocations: 9.94 KiB)
8.526 μs (2 allocations: 7.95 KiB)
796.970 ns (4 allocations: 9.94 KiB)
n = 10000
161.891 μs (5 allocations: 94.11 KiB)
171.660 μs (3 allocations: 78.22 KiB)
7.230 μs (5 allocations: 94.11 KiB)
Seems like the 2nd option is preferable for "small" vectors (< 100) and the third one for > 100 vectors.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
A faster version for sparse vectors:
function f4(d::SparseVector{N}) where {N}
c = zeros(length(d))
for (ni, i) in enumerate(d.nzind)
@inbounds c[i] = -d.nzval[ni]
end
return c
end
# (see 4th row)
>
> n = 10
> 143.405 ns (4 allocations: 384 bytes)
> 73.122 ns (2 allocations: 176 bytes)
> 118.938 ns (4 allocations: 384 bytes)
> 42.025 ns (1 allocation: 160 bytes)
>
> n = 100
> 520.623 ns (4 allocations: 1.19 KiB)
> 542.469 ns (2 allocations: 912 bytes)
> 162.908 ns (4 allocations: 1.19 KiB)
> 82.381 ns (1 allocation: 896 bytes)
>
> n = 1000
> 8.543 μs (4 allocations: 9.41 KiB)
> 7.568 μs (2 allocations: 7.95 KiB)
> 735.254 ns (4 allocations: 9.41 KiB)
> 542.567 ns (1 allocation: 7.94 KiB)
>
> n = 10000
> 159.176 μs (5 allocations: 93.48 KiB)
> 168.338 μs (3 allocations: 78.22 KiB)
> 7.569 μs (5 allocations: 93.48 KiB)
> 6.690 μs (2 allocations: 78.20 KiB)
>
>
EDIT: combining these codes for dense/sparse vectors we arrive at this solution:
@inline function _to_minus_vector(d::SparseVector)
c = zeros(length(d))
for (ni, i) in enumerate(d.nzind)
@inbounds c[i] = -d.nzval[ni]
end
return c
end
@inline function _to_minus_vector(d::AbstractVector{N}) where {N}
return convert(Vector{N}, -d)
end
function f5(d::AbstractVector{N}) where {N}
return _to_minus_vector(d)
end
for n in [10, 100, 1000, 10000]
println("\n n = $n ")
println("\n dense\n")
d = randn(n)
@btime f1($d)
@btime f2($d)
@btime f3($d)
@btime f5($d)
println("\n sparse\n")
d = sprandn(n, 0.1)
@btime f1($d)
@btime f2($d)
@btime f3($d)
@btime f5($d)
end
n = 10
dense
41.935 ns (1 allocation: 160 bytes)
48.410 ns (2 allocations: 176 bytes)
84.796 ns (2 allocations: 320 bytes)
41.692 ns (1 allocation: 160 bytes)
sparse
140.612 ns (4 allocations: 384 bytes)
79.601 ns (2 allocations: 176 bytes)
121.206 ns (4 allocations: 384 bytes)
40.700 ns (1 allocation: 160 bytes)
n = 100
dense
80.613 ns (1 allocation: 896 bytes)
82.760 ns (2 allocations: 912 bytes)
161.918 ns (2 allocations: 1.75 KiB)
81.906 ns (1 allocation: 896 bytes)
sparse
576.656 ns (4 allocations: 1.25 KiB)
555.898 ns (2 allocations: 912 bytes)
168.000 ns (4 allocations: 1.25 KiB)
82.758 ns (1 allocation: 896 bytes)
n = 1000
dense
539.077 ns (1 allocation: 7.94 KiB)
523.103 ns (2 allocations: 7.95 KiB)
1.179 μs (2 allocations: 15.88 KiB)
514.635 ns (1 allocation: 7.94 KiB)
sparse
8.839 μs (4 allocations: 9.56 KiB)
7.727 μs (2 allocations: 7.95 KiB)
752.535 ns (4 allocations: 9.56 KiB)
591.906 ns (1 allocation: 7.94 KiB)
n = 10000
dense
5.393 μs (2 allocations: 78.20 KiB)
5.553 μs (3 allocations: 78.22 KiB)
9.688 μs (4 allocations: 156.41 KiB)
5.403 μs (2 allocations: 78.20 KiB)
sparse
162.125 μs (5 allocations: 93.86 KiB)
171.141 μs (3 allocations: 78.22 KiB)
7.579 μs (5 allocations: 93.86 KiB)
6.501 μs (2 allocations: 78.20 KiB)
src/HPolyhedron.jl
Outdated
@@ -151,8 +151,23 @@ function σ(d::AbstractVector{N}, P::HPoly{N}) where {N<:Real} | |||
end | |||
end | |||
|
|||
@inline function _to_minus_vector(d::SparseVector) | |||
c = zeros(length(d)) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
use zeros(N, ...)
Closes #1011.