How to set NaN values in an RArray to a certain number?

[x66ccff]

Besides this method, are there any others?

@allowscalar xr = replace(xr, NaN => 1e9)

I found that using @allowscalar is particularly slow. I looked into the methods in Ops.jl, but I couldn't find a way to avoid @allowscalar.

[Pangoraw]

You can use something like this:

xr = Reactant.Ops.select(Reactant.Ops.is_finite(xr),
                         xr, fill!(similar(xr), 1e9))

We could add an overload to isnan for reactant arrays/numbers.

[x66ccff]

Thanks, this works for me!

using Reactant
x = rand(1, 100)
M = 1e9
x[2] = NaN
xr = Reactant.to_rarray(x)
f_is_finite = @compile Reactant.Ops.is_finite(xr)
f_fill = @compile fill!(similar(xr), M)
f_select = @compile Reactant.Ops.select(f_is_finite(xr), xr, f_fill(similar(xr), M))

f_select(f_is_finite(xr), xr, f_fill(similar(xr), 1e9))

1×100 ConcreteRArray{Float64, 2}:
 0.771444  1.0e9  0.510561  0.724406  0.0616827  0.00912104  0.286097  0.0725363  0.817941  0.419156  0.214779  0.222481  …  0.679535  0.118373  0.0639142  0.156426  0.692612  0.475274  0.565046  0.805544  0.175033  0.0221182  0.197093  0.0293862

[Pangoraw]

Note that it will also set -Inf and Inf to 1e9 (per stablehlo.is_finite).

[Pangoraw]

Following #525, you can now do:

julia> x = Reactant.to_rarray([1.0, NaN, Inf, -Inf, NaN]);

julia> replace_nans(x, v) = ifelse.(isnan.(x), fill!(similar(x), v), x)
replace_nans (generic function with 1 method)

julia> Reactant.@jit replace_nans(x, 1e9)
5-element ConcreteRArray{Float64, 1}:
   1.0
   1.0e9
  Inf
 -Inf
   1.0e9

which will also work outside Reactant.