How to properly pass structs into GPU? (MWE included)

[Ahmed_Salih]

@mcabbott

The mutability of the structs is irrelevant here. The faster path is mutating CuArrays within (and would work fine with an immutable struct); the slower path with @set is allocating new arrays.

Could you explain to me how one would do that? I am not aware of how to change the “CuArrays from within a struct”. I thought immutable meant I had to use a special package like Setfield to do it

EDIT: I think I got it! Basically I was trying to change an immutable struct, but instead I can change the mutable container it holds, in this case the array, using .=.

Link with more info: Julia: How much can we change the objects in immutable struct type? - Stack Overflow

---

@mcabbott

You should probably care about how many GPU operations you perform, this code is 3:

    V.xᵢⱼ²    .= dot.(V.xᵢⱼ,V.xᵢⱼ)
    V.dᵢⱼ     .= sqrt.(V.xᵢⱼ²)
    V.qᵢⱼ     .= V.dᵢⱼ ./ H

which could be 1, like V.dᵢⱼ .= sqrt.(dot.(V.xᵢⱼ,V.xᵢⱼ)) ./ H. I think that’s why your kernel is faster. For operations taking 30μs, you may find that you’re just timing the number of operations.

Okay, that is a good take away! In my case I wish to save the intermediate results as well (since I will be using them in other places), but for cases where I don’t need to I will try merging everything into one operation.

Thank you for your answer