slow write performance for abstract array

[johnnychen94]

using TestImages, FileIO, ImageCore
using BenchmarkTools

img = Float64.(testimage("cameraman"))

@btime save("tmp.tiff", img) #  1.840 ms (140 allocations: 519.67 KiB)
@btime save("tmp.tiff", reinterpret(Gray{Float64}, img)) # 20.758 ms (235 allocations: 2.01 MiB)

julia> versioninfo()
Julia Version 1.7.2
Commit bf53498635 (2022-02-06 15:21 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: 12th Gen Intel(R) Core(TM) i9-12900K
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.1 (ORCJIT, goldmont)

[tlnagy]

That's a wild performance drop. My best guess is that there's an inference failure?

[johnnychen94]

This seems to be version related

Julia Version	type	time
1.6.5	Array	7.388 ms (138 allocations: 519.14 KiB)
1.6.5	ReinterpretArray	22.285 ms (238 allocations: 2.01 MiB)
1.6.5	MappedArray	10.035 ms (236 allocations: 2.01 MiB)
1.7.2	Array	1.498 ms (138 allocations: 519.14 KiB)
1.7.2	ReinterpretArray	19.275 ms (238 allocations: 2.01 MiB)
1.7.2	MappedArray	7.658 ms (235 allocations: 2.01 MiB)
1.8.0-beta1	Array	1.565 ms (181 allocations: 520.90 KiB)
1.8.0-beta1	ReinterpretArray	7.637 ms (234 allocations: 2.01 MiB)
1.8.0-beta1	MappedArray	6.346 ms (234 allocations: 2.01 MiB)

• MappedArray: of_eltype(Gray{Float64}, img)
• ReinterpretArray: reinterpret(Gray{Float64}, img)

The ReinterpretArray performance might be due to JuliaLang/julia#43287, which was fixed in Julia 1.8. This also causes ImageMagick's performance drop for JPEG JuliaIO/ImageMagick.jl#208

[tlnagy]

Thanks for thoroughly checking this. Glad to hear that this was fixed upstream. Can I go ahead and close?

[johnnychen94]

Given that eagerly collecting the data into contiguous memory takes less than 1ms in the same machine, I think there's more room for optimization to further close the gap.

julia> @btime collect(reinterpret(Gray{Float64}, $img));
  251.000 μs (2 allocations: 2.00 MiB)

[tlnagy]

Writing to disk will almost always have orders of magnitude of slowdown. I'm sure there's performance to be had though.

[johnnychen94]

I mean, if save(file, img) is slower than save(file, collect(img)), then it indicates that we're hitting a slow method.

[tlnagy]

That's fair.

[tlnagy]

I ran into this performance slowdown again today, it appears to be especially egregious when using memory-mapped OMETIFFs and leads to ~100x slowdown versus writing the image lazily to disk frame-by-frame.

[tlnagy]

After looking into this more, I think this is mostly a base Julia problem (correct me if I'm wrong here 😅), but even using Base.write leads to this delta.

Base case

julia> img = Float64.(testimage("cameraman"));

julia> @btime write("tmp.tiff", img)
  287.581 μs (13 allocations: 728 bytes)
2097152

Gray type

julia> b = reinterpret(Gray{Float64}, img);

julia> @btime write("tmp.tiff", b) 
ERROR: MethodError: no method matching write(::IOStream, ::Gray{Float64})
Closest candidates are:
  write(::IO, ::Any) at io.jl:672
  write(::IO, ::Any, ::Any...) at io.jl:673
  write(::TiffFile, ::Any) at ~/.julia/packages/TiffImages/FEcMg/src/files.jl:117
  ...
Stacktrace:
  [1] write(io::IOStream, x::Gray{Float64})
    @ Base ./io.jl:672


julia> @btime write("tmp.tiff", reinterpret(UInt8, b)) # need to reinterpret
  31.452 ms (14 allocations: 776 bytes)
2097152

Now, I'm forced to do the extra reinterpret because of issue highlighted here: JuliaLang/julia#24234 (comment) and it seems that reinterpret arrays are slow at writing: JuliaLang/julia#39781

Now interestingly if you reinterpret back to a Float64 instead of UInt8, it's faster

julia> @btime write("tmp.tiff", reinterpret(Float64, b))
  286.599 μs (13 allocations: 728 bytes)
2097152

But that doesn't generalize well:

julia> img = RGB{Float64}.(testimage("cameraman")); # RGB colors this time

julia> @btime write("tmp.tiff", img) # takes about three times longer as expected
  878.243 μs (13 allocations: 728 bytes)
6291456

julia> @btime write("tmp.tiff", reinterpret(Float64, img)) # now this is slow again
  21.970 ms (786446 allocations: 12.00 MiB)
6291456

[vtjnash]

Fixed by JuliaLang/julia#42593?