Hi all, I am trying my first steps with CUDA.jl and I have a question about memory layout. I have a simple unmutable structure, for example: struct Angles x::SVector{3, Float64} end I have operations like function add!(r::Angles, x::Angles, y:.Angles) for k in 1:3 r.x[i] = x.x[i] + y.x[i…

Hi, you can have a look at : [image] GitHub - JuliaArrays/StructArrays.jl: Efficient implementation of struct... Efficient implementation of struct arrays in Julia - GitHub - JuliaArrays/StructArrays.jl: Efficient implementation of struct arrays in Julia

StructArrays should work for this. The main caveat is that SVector has a nested layout (it has a Tuple inside that actually holds the elements), so you should tell StructArrays to “unwrap” that: using StructArrays, StaticArrays x = StructArray([SA[1, 2, 3], SA[4, 5, 6]], unwrap = t -> t <: Tuple) O…

Many thanks! I would like to use Tuples, but the problem is that being inmutable I would not be able to update its values (like for example, in the add! or myexp! functions above). Is this right?

You may perform some performance experiments but I guess that returning “new” immutable angles will not cause any trouble because it will only imply stack (or even register) business.

SVector is also immutable, so that wouldn’t be affected. Other than the comment above though (the julia compiler is quite good with immutable struct) it’s important to note that, given the particular “struct of arrays” layout, you can cheat and modify tuples. For example: julia> using StructArrays …

Hi Laurent, piever Many thanks. I only have a couple of questions regarding your comments. About the solution proposed by Laurent: Can I assume that immutables types will not cause extra allocations when running on the GPU. (I have tested and on the CPU this works fine). About the solution pr…

I think that @maleadt can answer more precisely but I think that it wrote this here: [image] Thread local storage is not implemented GPU StaticArrays are implemented as structs containing tuples, so it’ll be using registers and not PTX’s local memory. You can inspect gene…

[image] shiroghost: About the solution proposed by piever: Would this trick/cheat perform well on the GPU? I haven’t used StructArrays directly in CUDA kernels much, so I don’t really have a strong intuition here, but in principle if CUDA can figure out that it should turn the components of t…

Hi again, Many thanks for the suggestions. Certainly StructArrays sounds ver interesting, but I cannot manage to make it work in CUDA kernels. A simple example: using StructArrays, CUDA, BenchmarkTools struct pt x::Float64 y::Float64 end function initpt!(a) index = (blockIdx().x - 1)…

You need to assign the result of replace_storage. Furthermore, also always use @btime CUDA.@sync ....

Array of structures vs. Structures of arrays

Specific Domains GPU

LaurentPlagne March 11, 2021, 2:16pm 8

I think that @maleadt can answer more precisely but I think that it wrote this here:

1 Like

Topic		Replies	Views
New to Julia and Formatting StructArrays for GPU use with mutable scalars GPU staticarrays , structarrays	2	673	February 13, 2023
How to properly pass structs into GPU? (MWE included) GPU	6	1431	January 29, 2023
Best way to store data for GPU use GPU	7	1738	February 16, 2019
CUDA.@cuStaticSharedMem returning a Structure of Arrays GPU	0	346	July 26, 2021
Passing mutable struct to kernel GPU	7	2240	January 29, 2023

Array of structures vs. Structures of arrays

Related topics