# Help with Custom Struct for High Dimensional COO Arrays

**URL:** https://discourse.julialang.org/t/help-with-custom-struct-for-high-dimensional-coo-arrays/98934
**Category:** GPU
**Created:** [May 16, 2023, 2:49pm UTC](https://discourse.julialang.org/t/help-with-custom-struct-for-high-dimensional-coo-arrays/98934 "2023-05-16T14:49:09Z")
**Posts on this page:** 2
**Page:** 1

<div class="post-metadata">

### Author: ![ejmeitz](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/ejmeitz/32/45283_2.png) [@ejmeitz](https://discourse.julialang.org/u/ejmeitz)
#### Post date: [May 16, 2023, 2:49pm UTC](https://discourse.julialang.org/t/help-with-custom-struct-for-high-dimensional-coo-arrays/98934/1 "2023-05-16T14:49:09Z")

</div>

Hello,

I have an application where I need \>2 dimensional sparse arrays and I would like to test running them on the GPU as the application is embarrassingly parallel. My plan was to make a CuArray of the type TensorVal type using the function below. This function does run but only if I use @allowscalar. From my understanding its ok that this runs on the CPU because I’m just initializing the array. However, everything breaks when I try to use this array in the `cuda_sparse` function and I get index out of bounds error and the REPL crashes. I’m not sure if the issue is in the initialization or in the execution. Any help would be appreciated! If there are better ways to do this I’d love to hear them.

```julia
struct TensorVal
    i::Int32
    j::Int32
    k::Int32
    val::Float32
end

function get_non_zero_gpu(F3::SparseArray)
    num_nonzero = length(nonzero_values(F3))
    F3_non_zero = CuArray{TensorVal}(undef,(num_nonzero,))
    for (idx, val) in nonzero_pairs(F3)
        F3_non_zero[idx] = TensorVal(idx[1], idx[2], idx[3], val)
    end
    return F3_non_zero
end

```

For reference this is how I am trying to use the array (basically a high dimensional dot product):

```julia
function cuda_sparse(cuF3_sparse, cuPhi1, cuPhi2, cuPhi3)
    f = (f3_data) -> f3_data.val * cuPhi1[f3_data.i] * cuPhi2[f3_data.j] * cuPhi3[f3_data.k]
   return mapreduce(f, +, cuF3_sparse)
end

```

---

<div class="post-metadata">

### Author: ![ejmeitz](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/ejmeitz/32/45283_2.png) [@ejmeitz](https://discourse.julialang.org/u/ejmeitz)
#### Post date: [May 16, 2023, 2:57pm UTC](https://discourse.julialang.org/t/help-with-custom-struct-for-high-dimensional-coo-arrays/98934/2 "2023-05-16T14:57:14Z")

</div>

As it turns out I was being stupid, the `get_non_zero_gpu()` function should have been the code below. I would still love to hear if people have recommendations as to how I should go about this or if this approach seems optimal.

```julia
function get_non_zero_gpu(F3::SparseArray)
    num_nonzero = length(nonzero_values(F3))
    F3_non_zero = CuArray{TensorVal}(undef,(num_nonzero,))
    count = 1
    for (idx, val) in nonzero_pairs(F3)
        F3_non_zero[count] = TensorVal(idx[1], idx[2], idx[3], val)
        count += 1
    end
    return F3_non_zero
end

```