For example copy the second two columns to the first: [1 2 3; 1 2 3; 1 2 3] becomes [2 3 3; 2 3 3; 2 3 3]

Both the view and copyto! solutions suggested above allocate. Try this to avoid allocations: copyto!(m, 1, m, 1 + size(m,1), (size(m,2) - 1) * size(m,1)) Or use a for loop.

I think you misunderstand: since I don’t see this documented, this is something I don’t think I can rely on. Whether this works or not is not relevant, since if this is not part of the interface then it could silently stop working or introduce incorrect results.

I totally get that, and I fully understand the concern: how can you copy overlapping memory portions without making an extra copy first? I’m actually not sure how copyto! does it. However I’d point out that the documentation of copyto! does not specify that source and destination have to be differen…

[image] pablosanjose: the documentation of copyto! does not specify that source and destination have to be different, right? I think it would be worthwhile to open an issue about clarifying this. I get the impression that it is supposed to work, eg copy! is …

[image] pablosanjose: nice one! Pity it only works for full columns, unlike the CartesianIndices approach. Yes, for other ranges, I’d just consider a for loop. Either loop over each individual element, or a loop of copyto!. It’s a bit frustrating how it’s sometimes so hard to avoid allocation…

Thanks everyone. I should have specified strictly non-allocating methods! Copyto seems like the way to go, I was imagining something like memmove existed but couldn’t find it.

[image] bennedich: Yes, for other ranges, I’d just consider a for loop. Either loop over each individual element, or a loop of copyto! . It’s a bit frustrating how it’s sometimes so hard to avoid allocations for simple operations like this in Julia. Is it clear why copyto! with CartesianIndic…

Hey again, I’ve checked that the bit that allocates when using copyto! with CartesianIndices is precisely the unalias part, which in this case needs to make a copy of m, I guess. The interesting question is then how can memmove (which gets called directly when using @bennedich ’s copyto! solution) a…

[image] tim.holy: See also Understanding performance using @btime and @code_warntype, @code_llvm, etc , which appears to apply here. You mean this? julia> @btime copyto!($m, CartesianIndices((1:3, 1:2)), $m, CartesianIndices((1:3,2:3))); 42.432 ns (1 allocation: 160 bytes) julia> g(m) = co…

No, I mean the size check inside copyto! itself uses string interpolation. Factor that out into a separate function and it might get faster.

What is the quickest way to move a subsection of a matrix?

General Usage

tim.holy November 26, 2018, 11:15am 14

Topic		Replies	Views
Avoiding allocations from views: copyto! between arrays of different shape General Usage	14	4181	October 31, 2018
Why making `a[range] = b[range]` seems to make unnecessary allocations? Performance	4	1024	September 30, 2019
Memory allocation with `view` and array assignment General Usage array , memory-allocation	12	1371	December 17, 2021
Preallocating the result of getindex(...) General Usage question	8	1464	July 13, 2017
Proving that copyto! is allocation-free on a view, using AllocCheck General Usage views , allocations , alloccheck	7	237	October 15, 2024

What is the quickest way to move a subsection of a matrix?

Related topics