faster bin index for uniform edge #613

Moelf · 2020-10-29T19:03:54Z

before:

julia> @benchmark fit(Histogram, $(rand(10^6)), 0:0.05:1; closed=:right)
BenchmarkTools.Trial:
  memory estimate:  320 bytes
  allocs estimate:  2
  --------------
  minimum time:     25.080 ms (0.00% GC)
  median time:      25.262 ms (0.00% GC)
  mean time:        25.308 ms (0.00% GC)
  maximum time:     27.498 ms (0.00% GC)
  --------------
  samples:          198
  evals/sample:     1

julia> @benchmark fit(Histogram, $(rand(10^6)), 0:0.05:1; closed=:left)
BenchmarkTools.Trial:
  memory estimate:  320 bytes
  allocs estimate:  2
  --------------
  minimum time:     25.962 ms (0.00% GC)
  median time:      26.176 ms (0.00% GC)
  mean time:        26.764 ms (0.00% GC)
  maximum time:     28.579 ms (0.00% GC)
  --------------
  samples:          187
  evals/sample:     1

after:

julia> @benchmark fit(Histogram, $(rand(10^6)), 0:0.05:1; closed=:right)
BenchmarkTools.Trial:
  memory estimate:  320 bytes
  allocs estimate:  2
  --------------
  minimum time:     9.181 ms (0.00% GC)
  median time:      9.534 ms (0.00% GC)
  mean time:        9.539 ms (0.00% GC)
  maximum time:     10.606 ms (0.00% GC)
  --------------
  samples:          525
  evals/sample:     1

julia> @benchmark fit(Histogram, $(rand(10^6)), 0:0.05:1; closed=:left)
BenchmarkTools.Trial:
  memory estimate:  320 bytes
  allocs estimate:  2
  --------------
  minimum time:     15.343 ms (0.00% GC)
  median time:      18.779 ms (0.00% GC)
  mean time:        18.766 ms (0.00% GC)
  maximum time:     18.947 ms (0.00% GC)
  --------------
  samples:          267
  evals/sample:     1

Moelf · 2020-10-29T19:18:31Z

notice this also makes multi-dimensional faster automatically. I hope @ianshmean is as happy as I am rn ;)

julia> d = rand(3,10^6);

julia> @btime fit(Histogram, (d[1,:], d[2,:], d[3,:]), (0:0.05:1, 0:0.05:1, 0:0.05:1));
  88.310 ms (15 allocations: 22.95 MiB)

julia> @btime fit(Histogram, (d[1,:], d[2,:], d[3,:]), (0:0.05:1, 0:0.05:1, 0:0.05:1));
  48.421 ms (15 allocations: 22.95 MiB)

Octogonapus · 2020-10-30T18:32:43Z

I gave this a try locally and it is a performance regression in our use case. Our bins are a StepRange{UInt8,UInt8} 0x00:0x10:0xf0 so maybe something about the non-unit step causes this? I also had to clamp n between the bounds of the edge range to avoid an OOB exception (caused by the non-unit step in the range): n = clamp(round(Int, (x - first(edge)) / step(edge) + 1), 1, length(edge)).

Moelf · 2020-10-30T19:03:04Z

@Octogonapus interesting, any chance to paste a MWE here?

Octogonapus · 2020-10-30T20:03:28Z

using Images, TestImages, FileIO, Rotations, CoordinateTransformations, StatsBase, BenchmarkTools

img = Gray.(testimage("lighthouse"));
tfm = Translation(125,250) ∘ LinearMap(RotMatrix(pi/50)) ∘ Translation(-125,-250);
img_rotated = warp(img, tfm);

fixed = collect(reinterpret(UInt8, img[50:300, 50:300]));
moving = collect(reinterpret(UInt8, img_rotated[50:300, 50:300]));

step = UInt8(16);
h = fit(Histogram, (vec(fixed), vec(moving)), (0x00:step:0xff, 0x00:step:0xff));
@benchmark append!(h, (vec(fixed), vec(moving)))

Baseline:

BenchmarkTools.Trial: 
  memory estimate:  192 bytes
  allocs estimate:  5
  --------------
  minimum time:     859.135 μs (0.00% GC)
  median time:      924.734 μs (0.00% GC)
  mean time:        959.305 μs (0.00% GC)
  maximum time:     1.570 ms (0.00% GC)
  --------------
  samples:          5208
  evals/sample:     1

This patch without my fix:

ERROR: InexactError: trunc(UInt8, 256)
Stacktrace:
 [1] throw_inexacterror(::Symbol, ::Type{UInt8}, ::Int64) at ./boot.jl:558
 [2] checked_trunc_uint at ./boot.jl:588 [inlined]
 [3] toUInt8 at ./boot.jl:650 [inlined]
 [4] UInt8 at ./boot.jl:710 [inlined]
 [5] convert at ./number.jl:7 [inlined]
 [6] getindex at ./range.jl:658 [inlined]
 [7] _edge_binindex at /home/salmon/Documents/StatsBase.jl/src/hist.jl:255 [inlined]
 [8] (::StatsBase.var"#131#132"{Histogram{Int64,2,Tuple{StepRange{UInt8,UInt8},StepRange{UInt8,UInt8}}}})(::StepRange{UInt8,UInt8}, ::UInt8) at /home/salmon/Documents/StatsBase.jl/src/hist.jl:227
 [9] map at ./tuple.jl:177 [inlined]
 [10] binindex at /home/salmon/Documents/StatsBase.jl/src/hist.jl:226 [inlined]
 [11] push! at /home/salmon/Documents/StatsBase.jl/src/hist.jl:311 [inlined]
 [12] append!(::Histogram{Int64,2,Tuple{StepRange{UInt8,UInt8},StepRange{UInt8,UInt8}}}, ::Tuple{Array{UInt8,1},Array{UInt8,1}}) at /home/salmon/Documents/StatsBase.jl/src/hist.jl:332
 [13] ##core#853() at /home/salmon/.julia/packages/BenchmarkTools/eCEpo/src/execution.jl:371
 [14] ##sample#854(::BenchmarkTools.Parameters) at /home/salmon/.julia/packages/BenchmarkTools/eCEpo/src/execution.jl:377
 [15] _run(::BenchmarkTools.Benchmark{Symbol("##benchmark#852")}, ::BenchmarkTools.Parameters; verbose::Bool, pad::String, kwargs::Base.Iterators.Pairs{Symbol,Integer,NTuple{4,Symbol},NamedTuple{(:samples, :evals, :gctrial, :gcsample),Tuple{Int64,Int64,Bool,Bool}}}) at /home/salmon/.julia/packages/BenchmarkTools/eCEpo/src/execution.jl:405
 [16] (::Base.var"#inner#2"{Base.Iterators.Pairs{Symbol,Integer,NTuple{5,Symbol},NamedTuple{(:verbose, :samples, :evals, :gctrial, :gcsample),Tuple{Bool,Int64,Int64,Bool,Bool}}},typeof(BenchmarkTools._run),Tuple{BenchmarkTools.Benchmark{Symbol("##benchmark#852")},BenchmarkTools.Parameters}})() at ./essentials.jl:713
 [17] #invokelatest#1 at ./essentials.jl:714 [inlined]
 [18] #run_result#37 at /home/salmon/.julia/packages/BenchmarkTools/eCEpo/src/execution.jl:32 [inlined]
 [19] run(::BenchmarkTools.Benchmark{Symbol("##benchmark#852")}, ::BenchmarkTools.Parameters; progressid::Nothing, nleaves::Float64, ndone::Float64, kwargs::Base.Iterators.Pairs{Symbol,Integer,NTuple{5,Symbol},NamedTuple{(:verbose, :samples, :evals, :gctrial, :gcsample),Tuple{Bool,Int64,Int64,Bool,Bool}}}) at /home/salmon/.julia/packages/BenchmarkTools/eCEpo/src/execution.jl:94
 [20] #warmup#45 at /home/salmon/.julia/packages/BenchmarkTools/eCEpo/src/execution.jl:141 [inlined]
 [21] warmup(::BenchmarkTools.Benchmark{Symbol("##benchmark#852")}) at /home/salmon/.julia/packages/BenchmarkTools/eCEpo/src/execution.jl:141
 [22] top-level scope at /home/salmon/.julia/packages/BenchmarkTools/eCEpo/src/execution.jl:287

This patch with my fix:

BenchmarkTools.Trial: 
  memory estimate:  192 bytes
  allocs estimate:  5
  --------------
  minimum time:     1.565 ms (0.00% GC)
  median time:      1.671 ms (0.00% GC)
  mean time:        1.679 ms (0.00% GC)
  maximum time:     2.385 ms (0.00% GC)
  --------------
  samples:          2977
  evals/sample:     1

Moelf · 2021-08-28T12:15:12Z

implemented in https://github.com/Moelf/FHist.jl

julia> a = rand(10^5);

julia> using BenchmarkTools

julia> using FHist, StatsBase

julia> @benchmark Hist1D($a, -3:0.01:3) 
BenchmarkTools.Trial: 9057 samples with 1 evaluation.
 Range (min … max):  117.353 μs …  1.412 ms  ┊ GC (min … max): 0.00% … 82.64%
 Time  (median):     547.890 μs              ┊ GC (median):    0.00%
 Time  (mean ± σ):   550.589 μs ± 32.415 μs  ┊ GC (mean ± σ):  0.07% ±  1.59%

                                                          █     
  ▂▂▂▂▁▂▂▁▁▁▂▁▁▁▂▁▁▂▂▁▁▁▂▁▁▁▁▁▁▂▁▂▁▁▁▂▁▁▂▁▁▁▂▁▂▁▁▁▂▂▂▁▁▂▁▃█▆█▅ ▂
  117 μs          Histogram: frequency by time          572 μs <

 Memory estimate: 14.58 KiB, allocs estimate: 6.

julia> @benchmark fit(Histogram, $a, -3:0.01:3) 
BenchmarkTools.Trial: 1995 samples with 1 evaluation.
 Range (min … max):  2.457 ms …  2.635 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     2.482 ms              ┊ GC (median):    0.00%
 Time  (mean ± σ):   2.505 ms ± 40.487 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

        ▄█▅▁                                                  
  ▂▂▃▃▃▅████▅▅▅▅▄▄▃▃▃▃▃▃▂▂▂▂▂▂▂▁▂▁▂▂▂▂▂▃▃▄▅▆▆▄▃▃▃▃▃▂▃▂▂▂▂▂▂▂ ▃
  2.46 ms        Histogram: frequency by time        2.61 ms <

 Memory estimate: 4.89 KiB, allocs estimate: 2.

faster bin index for uniform edge

7a1eff0

Moelf force-pushed the faster_edge branch from 71abf9c to 7a1eff0 Compare October 29, 2020 19:09

Moelf closed this Aug 28, 2021

Moelf mentioned this pull request Aug 31, 2021

'Shortcut' in fitting Histogram if binning is uniform / even #563

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

faster bin index for uniform edge #613

faster bin index for uniform edge #613

Moelf commented Oct 29, 2020

Moelf commented Oct 29, 2020

Octogonapus commented Oct 30, 2020

Moelf commented Oct 30, 2020 •

edited

Octogonapus commented Oct 30, 2020

Moelf commented Aug 28, 2021 •

edited

faster bin index for uniform edge #613

faster bin index for uniform edge #613

Conversation

Moelf commented Oct 29, 2020

Moelf commented Oct 29, 2020

Octogonapus commented Oct 30, 2020

Moelf commented Oct 30, 2020 • edited

Octogonapus commented Oct 30, 2020

Moelf commented Aug 28, 2021 • edited

Moelf commented Oct 30, 2020 •

edited

Moelf commented Aug 28, 2021 •

edited