Speed up group_by_color
#116
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #116 +/- ##
=========================================
Coverage 100.00% 100.00%
=========================================
Files 12 12
Lines 878 884 +6
=========================================
+ Hits 878 884 +6 ☔ View full report in Codecov by Sentry. |
Benchmark Results
|
|
On second thought it is a kind of bucket sort, maybe we could just use |
|
An example of the benefits, for the following code: using SparseMatrixColorings
using SparseArrays
problem = ColoringProblem(; structure=:nonsymmetric, partition=:column)
algo = GreedyColoringAlgorithm(; decompression=:direct)
A = sprand(Bool, 1000, 1000, 0.02)
coloring(A, problem, algo)
@profview_allocs for _ in 1:10000; coloring(A, problem, algo); endBefore Allocation count
Allocation size
After Allocation count
Allocation size
|
|
@gdalle Your argument is that we have a better frame graph? Sincerely, I don't think we should do this. I think we should focus on more important issues for now and potentially revisit this PR later if you really want to keep it. |
No, it's a combination of things (profiling was just one of those):
First of all, let's remain civil please.
Benchmarks are noisy and they don't tell the whole story, otherwise your manual transposition would have been clearly superior in #107. On certain use cases this
I haven't seen you make a strong case for this either: what are your arguments?
Here's the thing though: the right way to decide about important issues on performance is to profile the code to find the bottlenecks. If the profile is shitty to read because there's one source of allocation taking up all the space, this makes our life harder for no reason. |
|
TLDR: my approach is slightly worse when there are very very few colors (3), and faster otherwise (>=10). Here's a benchmark taking only the grouping into account: using BenchmarkTools
function compute_group_sizes(colors::Vector{Int})
cmax = maximum(colors)
group_sizes = zeros(Int, cmax)
for c in colors
group_sizes[c] += 1
end
return group_sizes
end
function split_vecvec(colors::Vector{Int})
group_sizes = compute_group_sizes(colors)
groups = [Vector{Int}(undef, group_sizes[c]) for c in eachindex(group_sizes)]
fill!(group_sizes, 0)
for (k, c) in enumerate(colors)
group_sizes[c] += 1
pos = group_sizes[c]
groups[c][pos] = k
end
return groups
end
function split_vecview(colors::Vector{Int})
group_sizes = compute_group_sizes(colors)
group_offsets = cumsum(group_sizes)
groups_flat = similar(colors)
for (k, c) in enumerate(colors)
i = group_offsets[c] - group_sizes[c] + 1
groups_flat[i] = k
group_sizes[c] -= 1
end
TV = typeof(view(groups_flat, 1:1))
groups = Vector{TV}(undef, length(group_sizes)) # allocation 4, size cmax
for c in eachindex(group_sizes)
i = 1 + (c == 1 ? 0 : group_offsets[c - 1])
j = group_offsets[c]
groups[c] = view(groups_flat, i:j)
end
return groups
endAnd the benchmarking results (> 1 means the approach with views is better): julia> for n in 10 .^ (2, 3, 4, 5), cmax in (3, 10, 30, 100)
yield()
bench_vecvec = @benchmark split_vecvec(_colors) setup = (_colors = rand(1:($cmax), $n))
bench_vecview = @benchmark split_vecview(_colors) setup = (
_colors = rand(1:($cmax), $n)
)
ratios = (
time=minimum(bench_vecvec).time / minimum(bench_vecview).time,
memory=minimum(bench_vecvec).memory / minimum(bench_vecview).memory,
allocs=minimum(bench_vecvec).allocs / minimum(bench_vecview).allocs,
)
@info "Vecvec / vecview ratios - n=$n, cmax=$cmax" ratios.time ratios.memory ratios.allocs
end
┌ Info: Vecvec / vecview ratios - n=100, cmax=3
│ ratios.time = 0.9921126179496853
│ ratios.memory = 0.922077922077922
└ ratios.allocs = 1.25
┌ Info: Vecvec / vecview ratios - n=100, cmax=10
│ ratios.time = 1.4329145256099618
│ ratios.memory = 0.9714285714285714
└ ratios.allocs = 3.0
┌ Info: Vecvec / vecview ratios - n=100, cmax=30
│ ratios.time = 1.9765313592357387
│ ratios.memory = 1.1
└ ratios.allocs = 7.5
┌ Info: Vecvec / vecview ratios - n=100, cmax=100
│ ratios.time = 2.646371976647206
│ ratios.memory = 1.243718592964824
└ ratios.allocs = 23.0
┌ Info: Vecvec / vecview ratios - n=1000, cmax=3
│ ratios.time = 1.1245857661353953
│ ratios.memory = 1.00945179584121
└ ratios.allocs = 1.25
┌ Info: Vecvec / vecview ratios - n=1000, cmax=10
│ ratios.time = 1.099305752912996
│ ratios.memory = 1.007181328545781
└ ratios.allocs = 3.0
┌ Info: Vecvec / vecview ratios - n=1000, cmax=30
│ ratios.time = 1.1895478015959944
│ ratios.memory = 1.0316957210776545
└ ratios.allocs = 8.0
┌ Info: Vecvec / vecview ratios - n=1000, cmax=100
│ ratios.time = 1.5270439790191834
│ ratios.memory = 1.1164383561643836
└ ratios.allocs = 25.5
┌ Info: Vecvec / vecview ratios - n=10000, cmax=3
│ ratios.time = 0.9527100549951475
│ ratios.memory = 0.9990047770700637
└ ratios.allocs = 1.6
┌ Info: Vecvec / vecview ratios - n=10000, cmax=10
│ ratios.time = 1.050451626197754
│ ratios.memory = 1.0096991290577988
└ ratios.allocs = 2.4
┌ Info: Vecvec / vecview ratios - n=10000, cmax=30
│ ratios.time = 1.2483034681177096
│ ratios.memory = 1.0349200156067109
└ ratios.allocs = 6.4
┌ Info: Vecvec / vecview ratios - n=10000, cmax=100
│ ratios.time = 1.2643422354104847
│ ratios.memory = 1.0528372093023255
└ ratios.allocs = 20.4
┌ Info: Vecvec / vecview ratios - n=100000, cmax=3
│ ratios.time = 0.9579135843188993
│ ratios.memory = 0.9999000479769711
└ ratios.allocs = 1.6
┌ Info: Vecvec / vecview ratios - n=100000, cmax=10
│ ratios.time = 1.0529008105578626
│ ratios.memory = 0.9999800207783904
└ ratios.allocs = 4.4
┌ Info: Vecvec / vecview ratios - n=100000, cmax=30
│ ratios.time = 1.0253449503339322
│ ratios.memory = 1.0005785420739737
└ ratios.allocs = 12.4
┌ Info: Vecvec / vecview ratios - n=100000, cmax=100
│ ratios.time = 1.0237008826453011
│ ratios.memory = 1.0133598014888336
└ ratios.allocs = 20.4 |
|
@amontoison thoughts on this one? When we benchmark the grouping function on its own, as you see above, the benefits are clear as soon as we go beyond |
|
Closing temporarily because bicoloring will require rethinking this grouping function. We can reoptimize it afterwards |
|
Actually figured out a way to keep the same grouping behavior in the bicoloring branch, so we can merge this one |
gdalle
changed the title
group_by_colorgroup_by_color
|
@amontoison this is a very quick review and while the global benchmarks don't show much of a difference, the specific benchmarks in this comment strongly support this change. What do you think? |
amontoison
approved these changes
Oct 7, 2024
|
Merged 😉 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
group_by_color, instead of allocating one vector per color group, allocate a single common vector and return one view for each group. That way we only need