Conversation
lkdvos
changed the title
VictorVanthilt
left a comment
VictorVanthilt
left a comment
There was a problem hiding this comment.
- The
dominant_eigsolvename is self-explanatory and I can't come up with a better name so I'm okay with it. - The usage of
Base.iteratein an iterative solver is indeed a readability improvement, it's way more clear what steps the different algorithms entail.
Could you point out where Boxed variables were showing up in closures? I want to be able to avoid ruining performance when using closures in the future.
|
I can, but this really does not "ruin" performance at all, you want to avoid this in cases where your iterations take ~1ms or less, sure, but for VUMPS that's really just not the case. The boxed variables appear whenever you are capturing a variable in a closure, but also modifying it outside of the closure. The exact mechanism of when it does and doesn't happen is also not fully clear to me, but it tends to be a combination of scoping, closures, and reassignment that does the trick. Here are some examples to give an idea: let
A = zeros(10)
for _ in 1:10
res = tmap(eachindex(A)) do i
A[i] = i
A[i]
end
A = zeros(10)
end
end # A boxed
let
A = zeros(10)
for _ in 1:10
A = zeros(10)
res = tmap(eachindex(A)) do i
A[i] = i
A[i]
end
end
end # A boxed
let
for _ in 1:10
A = zeros(10)
res = tmap(eachindex(A)) do i
A[i] = i
A[i]
end
end
end # fineNote that the easiest way to fix this, which you might have seen a couple times in different places, is to introduce a new let
A = zeros(10)
for _ in 1:10
res = let A = A
tmap(eachindex(A)) do i
A[i] = i
A[i]
end
end
A = zeros(10)
end
end # A boxed |
|
Thanks for the explanation! This seems like a bit of julia black magic you just have to know about. Tomo Soejima has a nice piece about closures on his website that helped me understand them a bit more. |
2 participants
I wanted to include OhMyThreads v0.8, which now throws errors for boxed variables in closures.
Since I had wanted to improve the code reuse a bit anyways, I went ahead and fully refactored VUMPS and VOMPS to make better use of the iterative solvers approach.
dominant_eigsolve, the internal function I now used to map bothleading_boundaryandfind_groundstateto? It definitely avoids some code duplication in the different VUMPS implementations. Naming suggestions also welcome of course.