use zero(FT) instead of zero(field) in broadcasts

[juliasloan25]

Purpose

We commonly use the pattern @. field = ifelse(area_fraction ≈ 0, zero(field), field). I was under the impression that the zero(field) would be equivalent to zero(eltype(field)) when broadcasted, but this isn't the case. See the allocation and timing information below

Content

• replace zero(field) with 0 in all broadcasted ifelse calls
• some other misc. cleanup of ifelse calls

Timing comparison

Despite having the most allocations, zero(FT) is the fastest to run since it the type of the zero doesn't need to be promoted (as it does for e.g. 0).

julia> @btime @. field1 = ifelse(area_fraction == 0, zero(field1), field1)
  199.083 μs (3 allocations: 128 bytes)

julia> @btime @. field1 = ifelse(area_fraction == 0, zero(eltype(field1)), field1)
  198.125 μs (4 allocations: 160 bytes)

julia> @btime @. field1 = ifelse(area_fraction == 0, 0, field1);
  222.667 μs (2 allocations: 96 bytes)

julia> @btime @. field1 = ifelse(area_fraction == 0, zero(FT), field1)
  174.708 μs (8 allocations: 288 bytes)

Allocation comparison

julia> @allocated @. field1 = ifelse(area_fraction == 0, zero(field1), field1)
128

julia> @allocated @. field1 = ifelse(area_fraction == 0, zero(eltype(field1)), field1)
160

julia> @allocated @. field1 = ifelse(area_fraction == 0, 0, field1)
96

The allocations are even worse when the whole expression isn't broadcasted, which we were doing in a few places:

julia> @allocated field1 .= ifelse.(area_fraction .== 0, zero(field1), field1)
6368

julia> @allocated field1 .= ifelse.(area_fraction .== 0, zero(eltype(field1)), field1)
96

julia> @allocated field1 .= ifelse.(area_fraction .== 0, 0, field1)
96

---

• I have read and checked the items on the review checklist.

[juliasloan25]

cc @petebachant @szy21 @kmdeck
We use this pattern a lot in the integrated land code, so I expect this will speed up that model. It was used in source too so the performance of everything should get better. We'll see by how much :)

[ph-kev]

I am not sure if this matter, but do we need to do something like zero(FT) or something similar to that?

[juliasloan25]

zero(FT) can be slightly more type-stable and doesn't require type promotion of the Int to FT. I'll switch to that

[juliasloan25]

I also profiled the different options and saw that zero(FT) is faster even though it has fewer allocations, probably because when we use 0 the type has to be promoted.

julia> @btime @. field1 = ifelse(area_fraction == 0, zero(field1), field1)
  199.083 μs (3 allocations: 128 bytes)

julia> @btime @. field1 = ifelse(area_fraction == 0, zero(eltype(field1)), field1)
  198.125 μs (4 allocations: 160 bytes)

julia> @btime @. field1 = ifelse(area_fraction == 0, 0, field1);
  222.667 μs (2 allocations: 96 bytes)

julia> @btime @. field1 = ifelse(area_fraction == 0, zero(FT), field1)
  174.708 μs (8 allocations: 288 bytes)