Julia: Parsing as Float32 half as fast as parsing as Float64 and converting to Float32

[Luapulu]

So, I just discovered this:

julia> using BenchmarkTools
julia> @btime parse(Float64, "50.303")
  32.163 ns (0 allocations: 0 bytes)
50.303

julia> @btime parse(Float32, "50.303")
  71.524 ns (0 allocations: 0 bytes)
50.303f0

julia> @btime convert(Float32, parse(Float64, "50.303"))
  32.430 ns (0 allocations: 0 bytes)
50.303f0

If parsing as Float64 and converting to Float32 is twice as fast, why not do it that way by default?

I googled around a bit and couldn't find anything on this so hopefully I'm not missing something obvious here.

[Luapulu]

My version info:

julia> versioninfo()
Julia Version 1.5.0
Commit 96786e22cc (2020-08-01 23:44 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin18.7.0)
  CPU: Intel(R) Core(TM) i5-8257U CPU @ 1.40GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-9.0.1 (ORCJIT, skylake)
Environment:
  JULIA_NUM_THREADS = 4

I also just installed julia fresh on a windows machine and got:

julia> using BenchmarkTools
julia> @btime parse(Float64, "50.303")
  530.521ns (0 allocations: 0 bytes)
50.303

julia> @btime parse(Float32, "50.303")
  532.105 ns (0 allocations: 0 bytes)
50.303f0

julia> @btime convert(Float32, parse(Float64, "50.303"))
  530.521 ns (0 allocations: 0 bytes)
50.303f0

julia> versioninfo()
Julia Version 1.5.1
Commit 697e782ab8 (2020-08-25 20:08 UTC)
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: Intel(R) Core(TM) i7-8565U CPU @ 1.80GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-9.0.1 (ORCJIT, skylake)

So on that machine it seems to make no difference I guess.

[yuyichao]

AFAICT this calls libc function so that seems to be a property of your libc parse function...

[Luapulu]

Yes, it does call a libc function. But if it's the case that it's always either faster or just as fast to parse as Float64 and convert, then why not do that?

[quinnj]

Quick plug for the Parsers.jl package which has native julia parsing implementations:

julia> @btime Parsers.parse(Float64, "50.303")
  13.697 ns (0 allocations: 0 bytes)
50.303

julia> @btime Parsers.parse(Float32, "50.303")
  13.997 ns (0 allocations: 0 bytes)
50.303f0

julia> @btime convert(Float32, Parsers.parse(Float64, "50.303"))
  13.586 ns (0 allocations: 0 bytes)
50.303f0

[oscardssmith]

Any reason why these aren't in Base?

[quinnj]

It's just a lot of work to get things into base/stdlib, much more than just a PR.