Improve RGB Min Max evaluation performance by using 2 or 3 comparison… #50622

L2 · 2021-04-01T23:31:18Z

…s instead of 4

Replaces 2 calls each to Math.Min and Math.Max (a total of 4 required
comparisons) with a single call to MinMaxRgb (at least 2 required
comparisons and at worst 3).

…s instead of 4 * Replaces 2 calls each to Math.Min and Math.Max (a total of 4 required comparisons) with a single call to MinMaxRgb (at least 2 required comparisons and at worst 3).

ghost · 2021-04-01T23:31:24Z

Tagging subscribers to this area: @safern, @tarekgh
See info in area-owners.md if you want to be subscribed.

Issue Details

…s instead of 4

Replaces 2 calls each to Math.Min and Math.Max (a total of 4 required
comparisons) with a single call to MinMaxRgb (at least 2 required
comparisons and at worst 3).

Author:	L2
Assignees:	-
Labels:	`area-System.Drawing`
Milestone:	-

L2 · 2021-04-01T23:41:43Z

dotnet/performance microbenchmarks filtered with System.Drawing.Tests.Perf_Color*:

Base = .NET6 Before this changeset
Diff = .NET6 After this changeset

No Slower results for the provided threshold = 2% and noise filter = 25ns.

Faster	base/diff	Base Median (ns)	Diff Median (ns)
System.Drawing.Tests.Perf_Color.GetHue	1.23	1756.53	1432.47
System.Drawing.Tests.Perf_Color.GetBrightness	1.09	1544.74	1415.17
System.Drawing.Tests.Perf_Color.GetSaturation	1.09	1602.51	1476.70

iamcarbon · 2021-04-01T23:44:24Z

src/libraries/System.Drawing.Primitives/src/System/Drawing/Color.cs

+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private void MinMaxRgb(out int min, out int max, int r, int g, int b)
+        {
+            if (b < g) {


Move trailing braces to their own lines for consistency with existing .NET code.

Thanks, updated.

danmoseley · 2021-04-02T04:29:15Z

Do we have sufficient functional test coverage?

L2 · 2021-04-02T05:46:29Z

Do we have sufficient functional test coverage?

Thanks @danmoseley . There were a couple of test cases missing which I've just added:

All Possible Cases:

label for existing test	case
(1)	r == g == b
(2)	r == g < b
(3)	r == g > b
(4)	r < g == b
(missing)	r < g < b
	r < g > b
(5)	where [r == b]
(missing)	where [r > b]
(missing)	where [r < b]
(7)	r > g == b
	r > g < b
(6)	where [r == b]
(missing)	where [r > b]
(missing)	where [r < b]
(missing)	r > g > b

Existing Test Coverage:

test label	case
(1)	[InlineData(0, 0, 0, 0f)]
	[InlineData(255, 255, 255, 1f)]
(7)	[InlineData(255, 0, 0, 0.5f)]
(5)	[InlineData(0, 255, 0, 0.5f)]
(2)	[InlineData(0, 0, 255, 0.5f)]
(3)	[InlineData(255, 255, 0, 0.5f)]
(6)	[InlineData(255, 0, 255, 0.5f)]
(4)	[InlineData(0, 255, 255, 0.5f)]
	[InlineData(51, 255, 255, 0.6f)]
	[InlineData(255, 51, 255, 0.6f)]
	[InlineData(255, 255, 51, 0.6f)]
	[InlineData(255, 51, 51, 0.6f)]
	[InlineData(51, 255, 51, 0.6f)]
	[InlineData(51, 51, 255, 0.6f)]
	[InlineData(51, 51, 51, 0.2f)]

For:

public void GetBrightness(int r, int g, int b, float expected)
{
    Assert.Equal(expected, Color.FromArgb(r, g, b).GetBrightness());
}

L2 · 2021-04-02T05:49:28Z

Also I wasn't sure if I should put this new method in Math.cs and make it a general use case to find the min and max of three values so I've left it inside Color.cs for now.

danmoseley · 2021-04-02T14:21:52Z

wasn't sure if I should put this new method in Math.cs

Drawing wouldn't be able to use anything from Math unless it was public: the libraries can only use each other's public surface. And to make new public API involves a review process, which would include consideration of how broadly useful the API would be. So I think it is in the right place.

safern · 2021-04-02T21:04:11Z

src/libraries/System.Drawing.Primitives/src/System/Drawing/Color.cs

+                    if (r < g)
+                        max = g;
+                    else
+                        max = r;


We could simplify this to use ? assignment?

Suggested change

if (r < g)

max = g;

else

max = r;

max = r < g ? g : r;

Thanks, updated.

safern · 2021-04-02T21:05:10Z

src/libraries/System.Drawing.Primitives/src/System/Drawing/Color.cs

+                    if (g < r)
+                        min = g;
+                    else
+                        min = r;


Ditto:

Suggested change

if (g < r)

min = g;

else

min = r;

min = g < r ? g : r;

Thanks, updated.

safern · 2021-04-02T21:05:33Z

src/libraries/System.Drawing.Primitives/src/System/Drawing/Color.cs

+            int min, max;
+            MinMaxRgb(out min, out max, r, g, b);


Suggested change

int min, max;

MinMaxRgb(out min, out max, r, g, b);

MinMaxRgb(out int min, out int max, r, g, b);

Thanks, updated.

safern · 2021-04-02T21:05:56Z

src/libraries/System.Drawing.Primitives/src/System/Drawing/Color.cs

+            int min, max;
+            MinMaxRgb(out min, out max, r, g, b);


Suggested change

int min, max;

MinMaxRgb(out min, out max, r, g, b);

MinMaxRgb(out int min, out int max, r, g, b);

Thanks, updated.

safern · 2021-04-02T21:06:13Z

src/libraries/System.Drawing.Primitives/src/System/Drawing/Color.cs

+            int min, max;
+            MinMaxRgb(out min, out max, r, g, b);


Suggested change

int min, max;

MinMaxRgb(out min, out max, r, g, b);

MinMaxRgb(out int min, out int max, r, g, b);

Thanks, updated.

safern

Other than small style improvements, looks good. Thanks!

L2 · 2021-04-02T22:02:36Z

Other than small style improvements, looks good. Thanks!

Thanks @safern , I've updated it with the requested changes.

ghost · 2021-04-02T22:03:36Z

Hello @safern!

Because this pull request has the auto-merge label, I will be glad to assist with helping to merge this pull request once all check-in policies pass.

p.s. you can customize the way I help with merging this pull request, such as holding this pull request until a specific person approves. Simply @mention me (`@msftbot`) and give me an instruction to get started! Learn more here.

stephentoub · 2021-04-03T23:46:58Z

src/libraries/System.Drawing.Primitives/src/System/Drawing/Color.cs

+        private void MinMaxRgb(out int min, out int max, int r, int g, int b)
+        {
+            if (b < g)
+            {
+                if (b < r)
+                {
+                    min = b;
+                    max = r < g ? g : r;
+                }
+                else
+                {
+                    min = r;
+                    max = g;
+                }
+            }
+            else
+            {
+                if (r < b)
+                {
+                    max = b;
+                    min = g < r ? g : r;
+                }
+                else
+                {
+                    max = r;
+                    min = g;
+                }
+            }
+        }


This can be simplified to instead be:

Suggested change

private void MinMaxRgb(out int min, out int max, int r, int g, int b)

{

if (b < g)

{

if (b < r)

{

min = b;

max = r < g ? g : r;

}

else

{

min = r;

max = g;

}

}

else

{

if (r < b)

{

max = b;

min = g < r ? g : r;

}

else

{

max = r;

min = g;

}

}

}

[MethodImpl(MethodImplOptions.AggressiveInlining)]

private static void MinMaxRgb(out int min, out int max, int r, int g, int b)

{

if (r > g)

{

max = r;

min = g;

}

else

{

max = g;

min = r;

}

if (b > max)

{

max = b;

}

else if (b < min)

{

min = b;

}

}

Thanks @stephentoub , I ran the perf_color microbenchmarks against 1) before this PR changeset, 2) the current PR's changeset, 3) proposed simplification

**.NET6 Before VS .NET6 with current PR changeset: Run1**
summary: better: 3, geomean: 1.151 total diff: 3
No Slower results for the provided threshold = 2% and noise filter = 25ns.

Faster base/diff Base Median (ns) Diff Median (ns) Modality

System.Drawing.Tests.Perf_Color.GetBrightness 1.23 1549.96 1259.50

System.Drawing.Tests.Perf_Color.GetHue 1.18 1753.24 1484.57

System.Drawing.Tests.Perf_Color.GetSaturation 1.05 1451.59 1383.79

**.NET6 Before VS .NET6 with current PR changeset: Run2**
summary: better: 3, geomean: 1.208 total diff: 3
No Slower results for the provided threshold = 2% and noise filter = 25ns.

Faster base/diff Base Median (ns) Diff Median (ns) Modality

System.Drawing.Tests.Perf_Color.GetBrightness 1.23 1543.19 1259.71

System.Drawing.Tests.Perf_Color.GetSaturation 1.22 1682.03 1384.17

System.Drawing.Tests.Perf_Color.GetHue 1.18 1758.68 1485.53

**.NET6 Before VS .NET6 with current PR changeset: Run3**
summary: better: 3, geomean: 1.189 total diff: 3
No Slower results for the provided threshold = 2% and noise filter = 25ns.

Faster base/diff Base Median (ns) Diff Median (ns) Modality

System.Drawing.Tests.Perf_Color.GetSaturation 1.22 1682.42 1384.03

System.Drawing.Tests.Perf_Color.GetBrightness 1.19 1497.18 1259.28

System.Drawing.Tests.Perf_Color.GetHue 1.16 1725.25 1483.41

**.NET6 Before VS .NET6 with proposed simplification: Run1**
summary: better: 2, geomean: 1.123 total diff: 2
No Slower results for the provided threshold = 2% and noise filter = 25ns.

Faster base/diff Base Median (ns) Diff Median (ns) Modality

System.Drawing.Tests.Perf_Color.GetBrightness 1.13 1549.96 1375.52

System.Drawing.Tests.Perf_Color.GetHue 1.12 1753.24 1565.75

**.NET6 Before VS .NET6 with proposed simplification: Run2**
summary: better: 3, geomean: 1.135 total diff: 3
No Slower results for the provided threshold = 2% and noise filter = 25ns.

Faster base/diff Base Median (ns) Diff Median (ns) Modality

System.Drawing.Tests.Perf_Color.GetSaturation 1.16 1682.03 1448.45

System.Drawing.Tests.Perf_Color.GetHue 1.12 1758.68 1566.17

System.Drawing.Tests.Perf_Color.GetBrightness 1.12 1543.19 1375.17

**.NET6 Before VS .NET6 with proposed simplification: Run3**
summary: better: 2, geomean: 1.132 worse: 1, geomean: 1.050 total diff: 3

Slower diff/base Base Median (ns) Diff Median (ns) Modality

System.Drawing.Tests.Perf_Color.GetBrightness 1.05 1497.18 1572.36

Faster base/diff Base Median (ns) Diff Median (ns) Modality

System.Drawing.Tests.Perf_Color.GetSaturation 1.16 1682.42 1446.59

System.Drawing.Tests.Perf_Color.GetHue 1.10 1725.25 1565.42

---------------------- ----------------------
Some of the benchmarks seem to get a bit slower on average after the simplification. I think what is happening may be two causes:

There are paths with an extra store (we store max & min in the first conditional and then if the second conditional is taken one of these is overwritten).

Just my guess, but processor pipelining with branch prediction may find it easier to go through each of the conditionals in the current implementation as there are no dependencies. The simplification introduces a dependency on the second conditional as it needs the values from the first? Just a guess here.

I will follow yours and @safern 's recommendation on which to keep. Thanks again.

I didn't look at the disassembly but there might be an aliasing complication also, if the JIT compiler has to account for the possibility that min and max refer to the same location. Although perhaps it can disprove that after inlining.

I actually see them get faster on my machine. Regardless, though, I'd rather keep the code smaller, especially with it triplicated due to the aggressive inlining.

One more possible improvement - using local valiables instead of out parameters:

private static void MinMaxRgb(out int min, out int max, int r, int g, int b) { int minLoc, maxLoc; if (r > g) { maxLoc = r; minLoc = g; } else { maxLoc = g; minLoc = r; } if (b > maxLoc) { maxLoc = b; } else if (b < minLoc) { minLoc = b; } max = maxLoc; min = minLoc; }

I actually see them get faster on my machine. Regardless, though, I'd rather keep the code smaller, especially with it triplicated due to the aggressive inlining.

Sounds good @stephentoub , I've updated it with the requested changes. (I'm not sure how crediting works in this project as this is my second pull request, so I've added your username to the commit details as the originator of the simplification).

One more possible improvement - using local valiables instead of out parameters:

Thanks @hypeartist, I'm wondering if these local variables will be eliminated due to inlining, so performance will be about the same?

I'm not sure how crediting works in this project as this is my second pull request, so I've added your username to the commit details as the originator of the simplification

Thanks, but not necessary... I comment on a lot of PRs :)

I'm wondering if these local variables will be eliminated due to inlining, so performance will be about the same?

Yes, e.g.
https://sharplab.io/#v2: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=

* Credit for this simplification: stephentoub

danmoseley · 2021-04-06T04:55:12Z

Thanks for the contribution @L2 - perhaps you'd like to pick up something else?

…shim_mono # By Aaron Robinson (10) and others # Via GitHub * upstream/main: (108 commits) [mbr] Add Apple sample (dotnet#50740) make EstablishProxyTunnelAsync throw on failure status code from proxy (dotnet#50763) Improve RGB Min Max evaluation performance by using 2 or 3 comparison… (dotnet#50622) [mono] More domain cleanups (dotnet#50479) Fix Crossgen2 of PlatformDefaultMemberFunction methods and calls. (dotnet#50754) Disable EventSource generator in design-time builds (dotnet#50741) Fix X509 test failures on Android (dotnet#50301) Do not confuse fgDispBasicBlocks in fgMorphBlocks (dotnet#50703) Enforce 64KB event payload size limit on EventPipe (dotnet#50600) Reorganize CoreCLR native build to reduce CMake reconfigures when the build system is untouched (dotnet#49906) [mbr] Turn on hot reload for iOS, tvOS and MacCatalyst (dotnet#50458) improve connection scavenge logic by doing zero-byte read (dotnet#50545) Resolve call mdtokens when making tier 1 inline observations (dotnet#50675) Annotate APIs in System.Private.Xml (dotnet#49682) Support compiling against OpenSSL 3 headers Change Configuration.Json to use a regular Dictionary. (dotnet#50611) Remove unused BigNumFromBinary P/Invoke (dotnet#50670) Make Ninja the default CMake generator on Windows for the repo (dotnet#49715) [AppleAppBuilder] Entitlements to run tests on catalyst using the JIT (dotnet#50637) [mono] Fix delegate invokes to dynamic methods in mixed mode. (dotnet#50547) ... # Conflicts: # src/mono/dlls/mscordbi/CMakeLists.txt

Improve RGB Min Max evaluation performance by using 2 or 3 comparison…

7a0fabe

…s instead of 4 * Replaces 2 calls each to Math.Min and Math.Max (a total of 4 required comparisons) with a single call to MinMaxRgb (at least 2 required comparisons and at worst 3).

dotnet-issue-labeler bot added the area-System.Drawing label Apr 1, 2021

iamcarbon reviewed Apr 1, 2021

View reviewed changes

Conform trailing braces to .NET style guide

59d5387

Add missing functional test cases for MinMaxRgb

8341fe2

safern reviewed Apr 2, 2021

View reviewed changes

safern approved these changes Apr 2, 2021

View reviewed changes

Use ternary conditional operator and out var initialization

03644b6

safern added the auto-merge label Apr 2, 2021

stephentoub reviewed Apr 3, 2021

View reviewed changes

Simplify by removing nested conditionals

5c109b7

* Credit for this simplification: stephentoub

ghost removed the auto-merge label Apr 5, 2021

safern approved these changes Apr 5, 2021

View reviewed changes

safern added the auto-merge label Apr 5, 2021

This was referenced Apr 5, 2021

Builds legs getting abandoned due to agent notification issues #35223

Closed

Runtime jobs being abandoned due to infra #50746

Closed

safern merged commit 89498d8 into dotnet:main Apr 6, 2021

L2 deleted the rgbMinMax branch April 6, 2021 15:05

ghost locked as resolved and limited conversation to collaborators May 9, 2021

karelz added this to the 6.0.0 milestone May 20, 2021

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Improve RGB Min Max evaluation performance by using 2 or 3 comparison… #50622

Improve RGB Min Max evaluation performance by using 2 or 3 comparison… #50622

L2 commented Apr 1, 2021 •

edited by safern

Loading

ghost commented Apr 1, 2021

L2 commented Apr 1, 2021

iamcarbon Apr 1, 2021

L2 Apr 1, 2021

danmoseley commented Apr 2, 2021

L2 commented Apr 2, 2021 •

edited

Loading

L2 commented Apr 2, 2021

danmoseley commented Apr 2, 2021

safern Apr 2, 2021

L2 Apr 2, 2021

safern Apr 2, 2021

L2 Apr 2, 2021

safern Apr 2, 2021

L2 Apr 2, 2021

safern Apr 2, 2021

L2 Apr 2, 2021

safern Apr 2, 2021

L2 Apr 2, 2021

safern left a comment

L2 commented Apr 2, 2021

ghost commented Apr 2, 2021

stephentoub Apr 3, 2021

L2 Apr 5, 2021

KalleOlaviNiemitalo Apr 5, 2021

stephentoub Apr 5, 2021

hypeartist Apr 5, 2021

L2 Apr 5, 2021 •

edited

Loading

stephentoub Apr 6, 2021

stephentoub Apr 6, 2021

danmoseley commented Apr 6, 2021

	int min, max;
	MinMaxRgb(out min, out max, r, g, b);
	MinMaxRgb(out int min, out int max, r, g, b);

Faster	base/diff	Base Median (ns)	Diff Median (ns)
System.Drawing.Tests.Perf_Color.GetBrightness	1.23	1549.96	1259.50
System.Drawing.Tests.Perf_Color.GetHue	1.18	1753.24	1484.57
System.Drawing.Tests.Perf_Color.GetSaturation	1.05	1451.59	1383.79

Faster	base/diff	Base Median (ns)	Diff Median (ns)
System.Drawing.Tests.Perf_Color.GetBrightness	1.23	1543.19	1259.71
System.Drawing.Tests.Perf_Color.GetSaturation	1.22	1682.03	1384.17
System.Drawing.Tests.Perf_Color.GetHue	1.18	1758.68	1485.53

Faster	base/diff	Base Median (ns)	Diff Median (ns)
System.Drawing.Tests.Perf_Color.GetSaturation	1.22	1682.42	1384.03
System.Drawing.Tests.Perf_Color.GetBrightness	1.19	1497.18	1259.28
System.Drawing.Tests.Perf_Color.GetHue	1.16	1725.25	1483.41

Faster	base/diff	Base Median (ns)	Diff Median (ns)	Modality
System.Drawing.Tests.Perf_Color.GetBrightness	1.13	1549.96	1375.52
System.Drawing.Tests.Perf_Color.GetHue	1.12	1753.24	1565.75

Improve RGB Min Max evaluation performance by using 2 or 3 comparison… #50622

Improve RGB Min Max evaluation performance by using 2 or 3 comparison… #50622

Conversation

L2 commented Apr 1, 2021 • edited by safern Loading

ghost commented Apr 1, 2021

L2 commented Apr 1, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

danmoseley commented Apr 2, 2021

L2 commented Apr 2, 2021 • edited Loading

L2 commented Apr 2, 2021

danmoseley commented Apr 2, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

safern left a comment

Choose a reason for hiding this comment

L2 commented Apr 2, 2021

ghost commented Apr 2, 2021

p.s. you can customize the way I help with merging this pull request, such as holding this pull request until a specific person approves. Simply @mention me (@msftbot) and give me an instruction to get started! Learn more here.

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

L2 Apr 5, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

danmoseley commented Apr 6, 2021

L2 commented Apr 1, 2021 •

edited by safern

Loading

L2 commented Apr 2, 2021 •

edited

Loading

p.s. you can customize the way I help with merging this pull request, such as holding this pull request until a specific person approves. Simply @mention me (`@msftbot`) and give me an instruction to get started! Learn more here.

L2 Apr 5, 2021 •

edited

Loading