Matthew-Whisenant
commented
1 year ago As a follow-up, benchmarking did not see a significant performance increase.
New method
if N == 1
# If it is a one dimensional array, just return it as is
return v
else
# Otherwise, reshape into two-dimensional array
return reshape(v, :, N)
end Test
using ReadVTK
using BenchmarkTools
file = VTKFile("celldata_inline_binary_uncompressed.vtu")
@benchmark get_points($file)Current method
BenchmarkTools.Trial: 10000 samples with 1 evaluation. Range (min … max): 112.900 μs … 4.918 ms ┊ GC (min … max): 0.00% … 97.19% Time (median): 123.700 μs ┊ GC (median): 0.00% Time (mean ± σ): 147.404 μs ± 117.758 μs ┊ GC (mean ± σ): 6.24% ± 8.98%
▇█▆▅▅▄▃▂▁▁▁▃▃▃▂▁▁▁▁ ▂ ████████████████████▇▇▆▇▆▅▅▄▅▄▃▃▃▁▃▁▁▁▁▁▁▁▃▁▁▁▁▁▁▁▁▁▁▁▁▁▁▃▄▇▆ █ 113 μs Histogram: log(frequency) by time 577 μs <
Memory estimate: 581.57 KiB, allocs estimate: 2341.
New method
BenchmarkTools.Trial: 10000 samples with 1 evaluation. Range (min … max): 114.200 μs … 2.912 ms ┊ GC (min … max): 0.00% … 94.71% Time (median): 123.600 μs ┊ GC (median): 0.00% Time (mean ± σ): 137.880 μs ± 76.972 μs ┊ GC (mean ± σ): 5.86% ± 9.25%
█▇▅▅▄▂▁ ▂ ███████▇▇█▇▇▆▆▅▆▃▃▃▄▁▃▃▁▁▁▃▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▃▆▇ █ 114 μs Histogram: log(frequency) by time 685 μs <
Memory estimate: 581.57 KiB, allocs estimate: 2341.
sloede
In both FormatBinary and FormatAppended versions of get_data, we have:
https://github.com/JuliaVTK/ReadVTK.jl/blob/85db2535c3da7b0c3fffe2f5a2764011671d845c/src/ReadVTK.jl#L681-L687
Is there a reason for the type instability here? Such as Vector type for scalar fields and Matrix type for vector fields.
My workflow includes getting data from multiple field keys at the same time, such as fluid velocity and fluid pressure. But its harder to hcat/vcat or postprocess when one get_data returns a 3xN Matrix and one is a Nx1 Vector where the major dimension is different.
Seems to me it would make more sense for it to return Nx3 Matrix and Nx1 Vector or return 3xN Matrix and 1xN Matrix. The former would preserve current typing and make the output Matrix column-major just like Julia is.
Again I'm not too concerned about the typing, moreso that to cat the arrays I have to process vectors and scalar fields differently because of the major dimension mismatch.
I can do a PR of this if you all don't forsee this breaking anything in your Trixi2Vtk.jl package. Or if its more of a unnecessary or tedious change it can just be on my local ReadVTK. Just let me know.