Hoist some of the calculations from loops of 3DMoments() by kunalspathak · Pull Request #1818 · SixLabors/ImageSharp

kunalspathak · 2021-11-10T04:16:07Z

Prerequisites

I have written a descriptive pull-request title
I have verified that there are no overlapping pull-requests open
I have verified that I am following the existing coding patterns and practice as demonstrated in the repository. These follow strict Stylecop rules 👮.
I have provided test coverage for my change (where applicable)

Description

The 4-level nested for loop in 3DMoments is very expensive, the inner most loop executes almost 1.1M times. We call GetPaletteIndex() method that does bunch of calculation using index variables of all the 4-loops.

ImageSharp/src/ImageSharp/Processing/Processors/Quantization/WuQuantizer{TPixel}.cs

Lines 418 to 444 in 255226b

for(intr=1;r<IndexCount;r++)
{
volumeSpan.Clear();
for(intg=1;g<IndexCount;g++)
{
areaSpan.Clear();
for(intb=1;b<IndexCount;b++)
{
Momentline=default;
for(inta=1;a<IndexAlphaCount;a++)
{
intind1=GetPaletteIndex(r,g,b,a);
line+=momentSpan[ind1];
areaSpan[a]+=line;
intinv=(b*IndexAlphaCount)+a;
volumeSpan[inv]+=areaSpan[a];
intind2=ind1-baseIndex;
momentSpan[ind1]=momentSpan[ind2]+volumeSpan[inv];
}
}
}

There is a potential to hoist some of the invariants calculation outside the loop. I verified the benchmark and see around 2-3% improvement.

Here is the assembly difference: https://www.diffchecker.com/I4uqCtBO

codecov · 2021-11-10T04:39:35Z

Codecov Report

Merging #1818 (0db87a5) into master (21d95a7) will increase coverage by 0%.
The diff coverage is 100%.

@@ Coverage Diff @@## master #1818 +/- ##
=======================================
Coverage 87% 87% =======================================
Files 936 936 Lines 48164 48191 +27 Branches 6037 6037 =======================================
+ Hits 41970 42097 +127 + Misses 5190 5093 -97 + Partials 1004 1001 -3

Flag	Coverage Δ
unittests	`87% <100%> (+<1%)`	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
...ing/Processors/Quantization/WuQuantizer{TPixel}.cs	`91% <100%> (+<1%)`	⬆️
...ImageSharp/Formats/Webp/Lossless/Vp8LBitEntropy.cs	`98% <0%> (-2%)`	⬇️
.../ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs	`97% <0%> (+8%)`	⬆️
...ageSharp/Formats/Webp/Lossless/PredictorEncoder.cs	`98% <0%> (+8%)`	⬆️

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 21d95a7...0db87a5. Read the comment docs.

JimBobSquarePants · 2021-11-11T02:50:32Z

@kunalspathak Thanks for this it is very interesting. I've just read your linked runtime issue also.

If we can add some supplementary comments to explain the optimization and refactor some of the variable naming (prefer ind1R over ind1_r ) then that would be a great addition to the codebase! 👍

kunalspathak · 2021-11-11T07:53:06Z

I have also noticed other places where similar technique could be use although I didn't measure its performance.
For eg, if I just search for all the references of GetPaletteIndex:

ImageSharp/src/ImageSharp/Processing/Processors/Quantization/WuQuantizer{TPixel}.cs

Line 217 in 7d74c4c

privatestaticintGetPaletteIndex(intr,intg,intb,inta)

, once of the reference is :

ImageSharp/src/ImageSharp/Processing/Processors/Quantization/WuQuantizer{TPixel}.cs

Lines 233 to 249 in 7d74c4c

privatestaticMomentVolume(refBoxcube,ReadOnlySpan<Moment>moments)
=>moments[GetPaletteIndex(cube.RMax,cube.GMax,cube.BMax,cube.AMax)]
-moments[GetPaletteIndex(cube.RMax,cube.GMax,cube.BMax,cube.AMin)]
-moments[GetPaletteIndex(cube.RMax,cube.GMax,cube.BMin,cube.AMax)]
+moments[GetPaletteIndex(cube.RMax,cube.GMax,cube.BMin,cube.AMin)]
-moments[GetPaletteIndex(cube.RMax,cube.GMin,cube.BMax,cube.AMax)]
+moments[GetPaletteIndex(cube.RMax,cube.GMin,cube.BMax,cube.AMin)]
+moments[GetPaletteIndex(cube.RMax,cube.GMin,cube.BMin,cube.AMax)]
-moments[GetPaletteIndex(cube.RMax,cube.GMin,cube.BMin,cube.AMin)]
-moments[GetPaletteIndex(cube.RMin,cube.GMax,cube.BMax,cube.AMax)]
+moments[GetPaletteIndex(cube.RMin,cube.GMax,cube.BMax,cube.AMin)]
+moments[GetPaletteIndex(cube.RMin,cube.GMax,cube.BMin,cube.AMax)]
-moments[GetPaletteIndex(cube.RMin,cube.GMax,cube.BMin,cube.AMin)]
+moments[GetPaletteIndex(cube.RMin,cube.GMin,cube.BMax,cube.AMax)]
-moments[GetPaletteIndex(cube.RMin,cube.GMin,cube.BMax,cube.AMin)]
-moments[GetPaletteIndex(cube.RMin,cube.GMin,cube.BMin,cube.AMax)]
+moments[GetPaletteIndex(cube.RMin,cube.GMin,cube.BMin,cube.AMin)];

Here, r related calculation for cube.RMax and cube.RMin can be precalculated and reused.

Another interesting method is Mark(). What is the common range of RMin ~ RMax, GMin ~ GMax, etc. If it is >= 32, we are already calculating GetPaletteIndex() million times and some of the calculation can be hoisted out the way its been done in this PR. Let me know and I can do for Mark() method as well.

ImageSharp/src/ImageSharp/Processing/Processors/Quantization/WuQuantizer{TPixel}.cs

Lines 629 to 641 in 7d74c4c

for(intr=cube.RMin+1;r<=cube.RMax;r++)
{
for(intg=cube.GMin+1;g<=cube.GMax;g++)
{
for(intb=cube.BMin+1;b<=cube.BMax;b++)
{
for(inta=cube.AMin+1;a<=cube.AMax;a++)
{
tagSpan[GetPaletteIndex(r,g,b,a)]=label;
}
}
}
}

kunalspathak · 2021-11-11T07:53:34Z

 {
 using var memoryStream = new MemoryStream();
- var options = new PngEncoder { Quantizer = new WuQuantizer(new QuantizerOptions { Dither = null }) };
+ var options = new PngEncoder { Quantizer = new WuQuantizer(new QuantizerOptions { Dither = null }), ColorType = PngColorType.Palette };


I made this change so I can get perf coverage for it. Should I keep it or revert it?

Yep. That's an oversight from us.
Very happy to see what other optimizations you can bring as long as they are well commented. Thanks! 👍

Sure, happy to contribute. I think it will be more productive for us to contribute or investigate the performance issues if these benchmarks are in our system. Any progress on #1795?

Leaving that one for @antonfirsov Likely after we complete #1730 and release V2.

Answered:
#1795 (reply in thread)

kunalspathak · 2021-11-12T15:42:12Z

This should be ready for review now.

JimBobSquarePants

LGTM, thanks! 👍

kunalspathak · 2022-05-15T03:58:43Z

This workaround will not be needed anymore because codegen will take care of it as part of dotnet/runtime#68061. Here is the diff for the earlier version of C# code:

Full diff: https://www.diffchecker.com/DKGAZKq0

JimBobSquarePants · 2022-05-15T04:46:01Z

That’s .NET 7 though yeah? We’ll be sticking to LTS release targets from now on so just 6 for now.

I am planning on revising many parts of the codebase though to simplify things where possible now we have a much more modern single target.

kunalspathak · 2022-05-15T05:41:57Z

That’s .NET 7 though yeah?

Yes.

We’ll be sticking to LTS release targets from now on so just 6 for now.

Sure, this was just FYI :)

JimBobSquarePants · 2022-05-15T06:26:09Z

It’s really good to know!!

hoist some calculations out
6885d97

kunalspathak mentioned this pull request Nov 10, 2021
Hoisting the invariant out of multi-level nested loops dotnet/runtime#61420
Closed

kunalspathak commented Nov 11, 2021
View reviewed changes

kunalspathak added 2 commits November 10, 2021 23:55

Better variable names
c0d5dfb

Optimize Mark
57357b0

kunalspathak marked this pull request as ready for review November 12, 2021 15:41

brianpopow added the area:performance label Nov 12, 2021

Merge branch 'master' into 3DMoments
0db87a5

JimBobSquarePants approved these changes Nov 13, 2021
View reviewed changes

JimBobSquarePants merged commit a39fe4c into SixLabors:masterNov 13, 2021

This was referenced Sep 23, 2025
Bump the nuget group with 5 updates norschel-archive/enterJSWebSecurity2025-Demo1#8
Open
Bump SixLabors.ImageSharp from 1.0.3 to 2.1.11 sajadbz/Sbz-clean-architecture#10
Open

dependabotBot mentioned this pull request Oct 19, 2025
Bump the nuget group with 1 update ewdlop/Raytracing#1
Open

dependabotBot mentioned this pull request Feb 21, 2026
Bump SixLabors.ImageSharp from 1.0.1 to 2.1.11 sipsorcery-org/sipsorcery#1533
Merged

dependabotBot mentioned this pull request Mar 27, 2026
Bump the nuget group with 1 update gongfuture/afdian-action#1
Open

dependabotBot mentioned this pull request Apr 18, 2026
Bump the nuget group with 3 updates bakerboy448/Whisparr#5
Closed

dependabotBot mentioned this pull request Apr 26, 2026
Bump SixLabors.ImageSharp and System.Text.Encoding.CodePages johnwatson484/sugar-free-healthy-diet#4
Merged

This was referenced Jul 1, 2026
Bump the nuget group with 5 updates norschel-conferences/DWX2026-Prep-GHAS#12
Closed
Bump the nuget group with 14 updates norschel-conferences/DWX2026-Prep-GHAS#13
Open
Bump the nuget group with 14 updates norschel-conferences/DWX2026-GHAS#8
Open

This was referenced Jul 17, 2026
Bump SixLabors.ImageSharp from 1.0.4 to 2.1.11 MinistryOfWelfare/SibylClient#2
Open
Bump the nuget group with 4 updates felipetruman/OpenBullet2#1
Closed
Bump the nuget group with 2 updates felipetruman/OpenBullet2#2
Closed

dependabotBot mentioned this pull request Aug 2, 2026
Bump SixLabors.ImageSharp from 1.0.0-beta0006 to 2.1.11 mohaom/TailwindTraders-Website#2
Open

dependabotBot mentioned this pull request Aug 16, 2026
Bump SixLabors.ImageSharp from 1.0.0-beta0006 to 2.1.11 cc-prod-org-01/TailwindTraders-Website#1
Open

	for(intr=1;r<IndexCount;r++)
	{
	volumeSpan.Clear();

	for(intg=1;g<IndexCount;g++)
	{
	areaSpan.Clear();

	for(intb=1;b<IndexCount;b++)
	{
	Momentline=default;

	for(inta=1;a<IndexAlphaCount;a++)
	{
	intind1=GetPaletteIndex(r,g,b,a);
	line+=momentSpan[ind1];

	areaSpan[a]+=line;

	intinv=(b*IndexAlphaCount)+a;
	volumeSpan[inv]+=areaSpan[a];

	intind2=ind1-baseIndex;
	momentSpan[ind1]=momentSpan[ind2]+volumeSpan[inv];
	}
	}
	}

Uh oh!

Uh oh!

Conversation

kunalspathak commented Nov 10, 2021

Prerequisites

Description

Uh oh!

codecovBot commented Nov 10, 2021• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

JimBobSquarePants commented Nov 11, 2021

Uh oh!

kunalspathak commented Nov 11, 2021

Uh oh!

kunalspathakNov 11, 2021

Choose a reason for hiding this comment

Uh oh!

JimBobSquarePantsNov 12, 2021

Choose a reason for hiding this comment

Uh oh!

kunalspathakNov 12, 2021

Choose a reason for hiding this comment

Uh oh!

JimBobSquarePantsNov 13, 2021

Choose a reason for hiding this comment

Uh oh!

antonfirsovNov 13, 2021

Choose a reason for hiding this comment

Uh oh!

kunalspathak commented Nov 12, 2021

Uh oh!

JimBobSquarePants left a comment

Choose a reason for hiding this comment

Uh oh!

kunalspathak commented May 15, 2022

Uh oh!

JimBobSquarePants commented May 15, 2022

Uh oh!

kunalspathak commented May 15, 2022

Uh oh!

JimBobSquarePants commented May 15, 2022

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

codecovBot commented Nov 10, 2021•
edited
Loading