Consider this function: pure function so2nn_inlined(Hso,Nspin,Norb,L) result(Hnn) !! Local version of so2nn_reshape, doing all iω at once !! > with high optimization levels it should be inlined integer,intent(in) :: Nspin,Norb,L complex(8),intent(in) :: Hso(Nspin*Norb,…

https://fortran-lang.org/en/learn/best_practices/arrays/ Explicit-shape Array and Assumed-size Array don’t check the size and the rank. It is the reason why modern fortran suggests Assumed-shape Array.

[image] rgba: especially since I never imagined I would need to check for size mismatch in a function call. I thought the compiler would do that for me. Checking the sizes would require some tests at runtime (and an explicit interface), which is something that we want to avoid in general: if …

kargl: A minimum working example. Fair enough, here you have: program MWE_for_Kargl implicit none integer, parameter :: Norb=3, Nspin=2, Lmats=1000, Nk=5 complex(8) :: dDelta(Nspin,Nspin,Norb,Norb,Lmats,Nk) complex(8) :: Htmp(Nspin*Norb,Nspin*Norb) in…

[image] rgba: better than what I see on the bigger example Concerning the evidence for what I claim to happen in my actual use case, I cannot really dump here the whole repository, but consider this (the whole body of grad_delta_replica, you just have to trust me that so2nn_inlined is placed …

[image] rgba: program MWE_for_Kargl I think this is the fix you’re looking for: link to godbolt EDIT: uhm I guess I hadn’t read the lengthy thread well enough. With assumed-size arrays, the subroutine only receives an address to the beginning of the variable, and then decides it will have th…

Oh… thanks for spending time on this, that’s indeed the catch (well actually the output Hnn must be rank 2 too, and the whole call being done in a loop over that index). Maybe I should have been clear, I know (now) what was wrong and how to fix it, I was more asking about why the compiler didn’t hel…

Yeah I this is another case of Knuth’s gold rule, “premature optimization is the root of all evil”. You were just trying to code the reshape function yourself!

Oh to all, sorry, I am actually silly after all. Crucial info I forgot to add (cause I totally forgot about, in first place): Nspin, Norb, Ldelta are only known at runtime so yeah, no matter what, it was silly to expect the compiler to check anything for me. And, being speed the main concern here, I…

I see your struggle, I think if you want to improve performance maybe you could try refactor some of your variables to reduce the number of jumps between non-contiguous chunks of memory: Here for example, you are copying the same value across L=size(Hnn,5) floats which are far apart in memory: Hn…

[image] rgba: This is a little perplexing to me, the so2nn_inlined function is a module procedure in my original code (or at least I thought so… it is placed in the contains section of the same module where grad_delta_replica is). I thought an explicit interface was in place, am I wrong on this? …

[image] FedericoPerini: to reduce the number of jumps between non-contiguous chunks of memory Yes, sure. Well, probably my actual sin has been to over-optimize. Profiling with valgrind (callgrind) I saw that switching from calling inv (which is just an high level wrapper to the suitable BLAS …

Do functions check size/rank of input arrays? If not why?

Help

FortranFan October 27, 2022, 3:59pm 15

@rgba, you want to go through this thread if you haven’t already:

Topic		Replies	Views
What's the purpose of array size inside subroutine arguments? Help	33	1377	February 1, 2024
Matrix index-pointer confusion Help	21	471	August 11, 2024
Fortran Best Practice Minibook Announcements	84	4750	September 20, 2021
Array element actual argument & assumed size dummy Help	17	488	May 1, 2025
Assumed-Size Dummy Argument Not Working in Module Procedure Help	37	1347	May 3, 2023

Do functions check size/rank of input arrays? If not why?

Related topics