# GPU-friendly one-hot encoding

**URL:** <https://discourse.julialang.org/t/gpu-friendly-one-hot-encoding/49312>\
**Category:** General Usage\
**Tags:** gpu, tullio\
**Created:** [October 30, 2020, 3:53pm UTC](https://discourse.julialang.org/t/gpu-friendly-one-hot-encoding/49312 "2020-10-30T15:53:51Z")\
**Posts on this page:** 1\
**Showing post:** 2

<div class="post-metadata">

**Author:** ![Mason](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/mason/32/2423_2.png) [@Mason](https://discourse.julialang.org/u/Mason)\
**Post date:** [October 30, 2020, 6:28pm UTC](https://discourse.julialang.org/t/gpu-friendly-one-hot-encoding/49312/2 "2020-10-30T18:28:30Z")

</div>

The simple answer is don’t worry about indexing yourself, let Tullio handle it:

```julia
julia> using CUDA, Tullio, KernelAbstractions

julia> function onehot(s::AbstractVector, n_dims=maximum(s))
           x = similar(s, n_dims, length(s))
           @tullio x[i, j] = (i == s[j]) (i ∈ 1:n_dims)
       end
onehot (generic function with 2 methods)

julia> onehot([1,2,3,4])
4×4 Array{Int64,2}:
 1 0 0 0
 0 1 0 0
 0 0 1 0
 0 0 0 1

julia> onehot(cu([1,2,3,4]))
4×4 CuArray{Int64,2}:
 1 0 0 0
 0 1 0 0
 0 0 1 0
 0 0 0 1

```

When you do `using KernelAbstractions` before defining `onehot`, Tullio is able to do efficient GPU codegen. If you do `using LoopVectorization` as well, you’ll get faster CPU codegen on hardware native types like `Int`.

Of course, making an actual dense array for a one-hot vector is often a bad idea but I’ll let others who know what they’re talking about chime in there.

---

_[View the full topic](https://discourse.julialang.org/t/gpu-friendly-one-hot-encoding/49312)._