Repository navigation
Component shuffling for RGB(A)-like formats with SIMD Intrinsics #1354
Description
Activity
- added this to the milestone
on Sep 16, 2020 - changed the title
[-]Component shuffling for RGB(A) like formats with SIMD Intrinsics[/-][+]Component shuffling for RGB(A)-like formats with SIMD Intrinsics[/+]on Sep 16, 2020 I'm gonna have a look at this.
Reacted by Anton FirszovBtw, the
Span<Vector4> <=> Span<TPixel>conversions are delegating the (pad/slice) shuffle work toPixelOperations<TPixel>conversion methods:Which means that the last two steps will be done automatically, if I'm not missing anything. Can't wait to see the
Vector4 <=> Rgb24before/after comparison. (Which is the main goal of this issue because of the conversion steps in Jpeg decoder and ResizeProcessor.)Reacted by James Jackson-South@antonfirsov need to add a specific benchmark for that but I know for certain that even my
Rgba32 <==> Rgb24fallback is a lot faster than the original.@antonfirsov Here you go!
ToVector4_Rgb24
BenchmarkDotNet=v0.12.1, OS=Windows 10.0.19041.572 (2004/?/20H1) Intel Core i7-8650U CPU 1.90GHz (Kaby Lake R), 1 CPU, 8 logical and 4 physical cores .NET Core SDK=3.1.403 [Host] : .NET Core 3.1.9 (CoreCLR 4.700.20.47201, CoreFX 4.700.20.47203), X64 RyuJIT Job-OIBEDX : .NET Framework 4.8 (4.8.4250.0), X64 RyuJIT Job-OPAORC : .NET Core 2.1.23 (CoreCLR 4.6.29321.03, CoreFX 4.6.29321.01), X64 RyuJIT Job-VPSIRL : .NET Core 3.1.9 (CoreCLR 4.700.20.47201, CoreFX 4.700.20.47203), X64 RyuJIT IterationCount=3 LaunchCount=1 WarmupCount=3Master
Method Job Runtime Count Mean Error StdDev Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated PixelOperations_Base Job-BPVKME .NET 4.7.2 64 278.4 ns 6.89 ns 0.38 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-BPVKME .NET 4.7.2 64 310.2 ns 87.58 ns 4.80 ns 1.11 0.02 - - - - PixelOperations_Base Job-FBIBGB .NET Core 2.1 64 231.9 ns 254.31 ns 13.94 ns 1.00 0.00 0.0052 - - 24 B PixelOperations_Specialized Job-FBIBGB .NET Core 2.1 64 230.2 ns 27.06 ns 1.48 ns 0.99 0.05 - - - - PixelOperations_Base Job-CETXPV .NET Core 3.1 64 215.5 ns 52.24 ns 2.86 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-CETXPV .NET Core 3.1 64 236.7 ns 553.01 ns 30.31 ns 1.10 0.13 - - - - PixelOperations_Base Job-BPVKME .NET 4.7.2 256 1,022.3 ns 3,570.70 ns 195.72 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-BPVKME .NET 4.7.2 256 622.5 ns 26.76 ns 1.47 ns 0.62 0.11 - - - - PixelOperations_Base Job-FBIBGB .NET Core 2.1 256 762.3 ns 103.78 ns 5.69 ns 1.00 0.00 0.0048 - - 24 B PixelOperations_Specialized Job-FBIBGB .NET Core 2.1 256 498.1 ns 70.87 ns 3.88 ns 0.65 0.00 - - - - PixelOperations_Base Job-CETXPV .NET Core 3.1 256 754.0 ns 37.92 ns 2.08 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-CETXPV .NET Core 3.1 256 436.8 ns 21.88 ns 1.20 ns 0.58 0.00 - - - - PixelOperations_Base Job-BPVKME .NET 4.7.2 2048 5,679.3 ns 1,454.37 ns 79.72 ns 1.00 0.00 - - - 24 B PixelOperations_Specialized Job-BPVKME .NET 4.7.2 2048 3,460.6 ns 273.43 ns 14.99 ns 0.61 0.01 - - - - PixelOperations_Base Job-FBIBGB .NET Core 2.1 2048 6,033.8 ns 8,785.67 ns 481.57 ns 1.00 0.00 - - - 24 B PixelOperations_Specialized Job-FBIBGB .NET Core 2.1 2048 3,421.3 ns 376.64 ns 20.64 ns 0.57 0.04 - - - - PixelOperations_Base Job-CETXPV .NET Core 3.1 2048 5,542.3 ns 790.31 ns 43.32 ns 1.00 0.00 - - - 24 B PixelOperations_Specialized Job-CETXPV .NET Core 3.1 2048 2,972.2 ns 70.72 ns 3.88 ns 0.54 0.00 - - - - Branch
Method Job Runtime Count Mean Error StdDev Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated PixelOperations_Base Job-OIBEDX .NET 4.7.2 64 298.4 ns 33.63 ns 1.84 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-OIBEDX .NET 4.7.2 64 355.5 ns 908.51 ns 49.80 ns 1.19 0.17 - - - - PixelOperations_Base Job-OPAORC .NET Core 2.1 64 220.1 ns 13.77 ns 0.75 ns 1.00 0.00 0.0055 - - 24 B PixelOperations_Specialized Job-OPAORC .NET Core 2.1 64 228.5 ns 41.41 ns 2.27 ns 1.04 0.01 - - - - PixelOperations_Base Job-VPSIRL .NET Core 3.1 64 213.6 ns 12.47 ns 0.68 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-VPSIRL .NET Core 3.1 64 217.0 ns 9.95 ns 0.55 ns 1.02 0.01 - - - - PixelOperations_Base Job-OIBEDX .NET 4.7.2 256 829.0 ns 242.93 ns 13.32 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-OIBEDX .NET 4.7.2 256 448.9 ns 4.04 ns 0.22 ns 0.54 0.01 - - - - PixelOperations_Base Job-OPAORC .NET Core 2.1 256 863.0 ns 1,253.26 ns 68.70 ns 1.00 0.00 0.0048 - - 24 B PixelOperations_Specialized Job-OPAORC .NET Core 2.1 256 309.2 ns 66.16 ns 3.63 ns 0.36 0.03 - - - - PixelOperations_Base Job-VPSIRL .NET Core 3.1 256 737.0 ns 253.90 ns 13.92 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-VPSIRL .NET Core 3.1 256 212.3 ns 1.07 ns 0.06 ns 0.29 0.01 - - - - PixelOperations_Base Job-OIBEDX .NET 4.7.2 2048 5,625.6 ns 404.35 ns 22.16 ns 1.00 0.00 - - - 24 B PixelOperations_Specialized Job-OIBEDX .NET 4.7.2 2048 1,974.1 ns 229.84 ns 12.60 ns 0.35 0.00 - - - - PixelOperations_Base Job-OPAORC .NET Core 2.1 2048 5,467.2 ns 537.29 ns 29.45 ns 1.00 0.00 - - - 24 B PixelOperations_Specialized Job-OPAORC .NET Core 2.1 2048 1,985.5 ns 4,714.23 ns 258.40 ns 0.36 0.05 - - - - PixelOperations_Base Job-VPSIRL .NET Core 3.1 2048 5,888.2 ns 1,622.23 ns 88.92 ns 1.00 0.00 - - - 24 B PixelOperations_Specialized Job-VPSIRL .NET Core 3.1 2048 1,165.0 ns 191.71 ns 10.51 ns 0.20 0.00 - - - - FromVector4_Rgb24
BenchmarkDotNet=v0.12.1, OS=Windows 10.0.19041.572 (2004/?/20H1) Intel Core i7-8650U CPU 1.90GHz (Kaby Lake R), 1 CPU, 8 logical and 4 physical cores .NET Core SDK=3.1.403 [Host] : .NET Core 3.1.9 (CoreCLR 4.700.20.47201, CoreFX 4.700.20.47203), X64 RyuJIT Job-XYEQXL : .NET Framework 4.8 (4.8.4250.0), X64 RyuJIT Job-HSXNJV : .NET Core 2.1.23 (CoreCLR 4.6.29321.03, CoreFX 4.6.29321.01), X64 RyuJIT Job-YUREJO : .NET Core 3.1.9 (CoreCLR 4.700.20.47201, CoreFX 4.700.20.47203), X64 RyuJIT IterationCount=3 LaunchCount=1 WarmupCount=3Master
Method Job Runtime Count Mean Error StdDev Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated PixelOperations_Base Job-BPNZYS .NET 4.7.2 64 317.7 ns 125.40 ns 6.87 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-BPNZYS .NET 4.7.2 64 316.4 ns 70.42 ns 3.86 ns 1.00 0.03 - - - - PixelOperations_Base Job-NYROHY .NET Core 2.1 64 232.7 ns 82.61 ns 4.53 ns 1.00 0.00 0.0055 - - 24 B PixelOperations_Specialized Job-NYROHY .NET Core 2.1 64 238.9 ns 106.11 ns 5.82 ns 1.03 0.01 - - - - PixelOperations_Base Job-LSNAID .NET Core 3.1 64 228.4 ns 15.16 ns 0.83 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-LSNAID .NET Core 3.1 64 250.3 ns 22.79 ns 1.25 ns 1.10 0.01 - - - - PixelOperations_Base Job-BPNZYS .NET 4.7.2 256 975.5 ns 1,646.67 ns 90.26 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-BPNZYS .NET 4.7.2 256 1,051.3 ns 170.43 ns 9.34 ns 1.08 0.11 0.0172 - - 72 B PixelOperations_Base Job-NYROHY .NET Core 2.1 256 793.0 ns 69.24 ns 3.80 ns 1.00 0.00 0.0048 - - 24 B PixelOperations_Specialized Job-NYROHY .NET Core 2.1 256 846.8 ns 117.07 ns 6.42 ns 1.07 0.01 0.0172 - - 72 B PixelOperations_Base Job-LSNAID .NET Core 3.1 256 797.2 ns 342.02 ns 18.75 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-LSNAID .NET Core 3.1 256 640.2 ns 19.74 ns 1.08 ns 0.80 0.02 0.0172 - - 72 B PixelOperations_Base Job-BPNZYS .NET 4.7.2 2048 6,178.4 ns 1,537.81 ns 84.29 ns 1.00 0.00 - - - 24 B PixelOperations_Specialized Job-BPNZYS .NET 4.7.2 2048 4,551.2 ns 257.65 ns 14.12 ns 0.74 0.01 0.0153 - - 72 B PixelOperations_Base Job-NYROHY .NET Core 2.1 2048 6,621.8 ns 13,533.62 ns 741.82 ns 1.00 0.00 - - - 24 B PixelOperations_Specialized Job-NYROHY .NET Core 2.1 2048 4,390.4 ns 184.24 ns 10.10 ns 0.67 0.07 0.0153 - - 72 B PixelOperations_Base Job-LSNAID .NET Core 3.1 2048 6,357.9 ns 964.20 ns 52.85 ns 1.00 0.00 - - - 24 B PixelOperations_Specialized Job-LSNAID .NET Core 3.1 2048 2,979.4 ns 132.81 ns 7.28 ns 0.47 0.00 0.0153 - - 72 B Branch
Method Job Runtime Count Mean Error StdDev Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated PixelOperations_Base Job-XYEQXL .NET 4.7.2 64 343.2 ns 305.91 ns 16.77 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-XYEQXL .NET 4.7.2 64 320.8 ns 19.93 ns 1.09 ns 0.94 0.05 - - - - PixelOperations_Base Job-HSXNJV .NET Core 2.1 64 234.3 ns 17.98 ns 0.99 ns 1.00 0.00 0.0052 - - 24 B PixelOperations_Specialized Job-HSXNJV .NET Core 2.1 64 246.0 ns 82.34 ns 4.51 ns 1.05 0.02 - - - - PixelOperations_Base Job-YUREJO .NET Core 3.1 64 222.3 ns 39.46 ns 2.16 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-YUREJO .NET Core 3.1 64 243.4 ns 33.58 ns 1.84 ns 1.09 0.01 - - - - PixelOperations_Base Job-XYEQXL .NET 4.7.2 256 824.9 ns 32.77 ns 1.80 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-XYEQXL .NET 4.7.2 256 967.0 ns 39.09 ns 2.14 ns 1.17 0.01 0.0172 - - 72 B PixelOperations_Base Job-HSXNJV .NET Core 2.1 256 756.9 ns 94.43 ns 5.18 ns 1.00 0.00 0.0048 - - 24 B PixelOperations_Specialized Job-HSXNJV .NET Core 2.1 256 1,003.3 ns 3,192.09 ns 174.97 ns 1.32 0.22 0.0172 - - 72 B PixelOperations_Base Job-YUREJO .NET Core 3.1 256 748.6 ns 248.03 ns 13.60 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-YUREJO .NET Core 3.1 256 437.0 ns 36.48 ns 2.00 ns 0.58 0.01 0.0172 - - 72 B PixelOperations_Base Job-XYEQXL .NET 4.7.2 2048 5,751.6 ns 704.24 ns 38.60 ns 1.00 0.00 - - - 24 B PixelOperations_Specialized Job-XYEQXL .NET 4.7.2 2048 4,391.6 ns 718.17 ns 39.37 ns 0.76 0.00 0.0153 - - 72 B PixelOperations_Base Job-HSXNJV .NET Core 2.1 2048 6,202.0 ns 1,815.18 ns 99.50 ns 1.00 0.00 - - - 24 B PixelOperations_Specialized Job-HSXNJV .NET Core 2.1 2048 4,225.6 ns 1,004.03 ns 55.03 ns 0.68 0.01 0.0153 - - 72 B PixelOperations_Base Job-YUREJO .NET Core 3.1 2048 6,157.1 ns 2,516.98 ns 137.96 ns 1.00 0.00 - - - 24 B PixelOperations_Specialized Job-YUREJO .NET Core 3.1 2048 1,822.7 ns 1,764.43 ns 96.71 ns 0.30 0.02 0.0172 - - 72 B Reacted by Anton Firszov@antonfirsov Pushed an update that affects
FromVector4_Rgb24based on a suggestion by @Sergio0694 and It's squeezed a few more percentage performance.Method Job Runtime Count Mean Error StdDev Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated PixelOperations_Base Job-RPFDIH .NET 4.7.2 64 327.7 ns 179.42 ns 9.83 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-RPFDIH .NET 4.7.2 64 320.8 ns 21.37 ns 1.17 ns 0.98 0.03 - - - - PixelOperations_Base Job-QSMZGQ .NET Core 2.1 64 254.8 ns 337.29 ns 18.49 ns 1.00 0.00 0.0052 - - 24 B PixelOperations_Specialized Job-QSMZGQ .NET Core 2.1 64 245.3 ns 37.22 ns 2.04 ns 0.97 0.07 - - - - PixelOperations_Base Job-YJZMFY .NET Core 3.1 64 232.2 ns 189.28 ns 10.37 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-YJZMFY .NET Core 3.1 64 255.4 ns 52.39 ns 2.87 ns 1.10 0.04 - - - - PixelOperations_Base Job-RPFDIH .NET 4.7.2 256 910.1 ns 293.07 ns 16.06 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-RPFDIH .NET 4.7.2 256 974.4 ns 490.48 ns 26.89 ns 1.07 0.05 0.0172 - - 72 B PixelOperations_Base Job-QSMZGQ .NET Core 2.1 256 849.7 ns 1,654.73 ns 90.70 ns 1.00 0.00 0.0048 - - 24 B PixelOperations_Specialized Job-QSMZGQ .NET Core 2.1 256 759.9 ns 77.06 ns 4.22 ns 0.90 0.09 0.0172 - - 72 B PixelOperations_Base Job-YJZMFY .NET Core 3.1 256 816.1 ns 56.07 ns 3.07 ns 1.00 0.00 0.0057 - - 24 B PixelOperations_Specialized Job-YJZMFY .NET Core 3.1 256 493.0 ns 216.79 ns 11.88 ns 0.60 0.01 0.0172 - - 72 B PixelOperations_Base Job-RPFDIH .NET 4.7.2 2048 6,394.6 ns 2,077.05 ns 113.85 ns 1.00 0.00 - - - 24 B PixelOperations_Specialized Job-RPFDIH .NET 4.7.2 2048 4,139.6 ns 276.41 ns 15.15 ns 0.65 0.01 0.0153 - - 72 B PixelOperations_Base Job-QSMZGQ .NET Core 2.1 2048 6,249.9 ns 799.28 ns 43.81 ns 1.00 0.00 - - - 24 B PixelOperations_Specialized Job-QSMZGQ .NET Core 2.1 2048 4,020.5 ns 3,211.19 ns 176.02 ns 0.64 0.03 0.0153 - - 72 B PixelOperations_Base Job-YJZMFY .NET Core 3.1 2048 6,403.5 ns 4,626.86 ns 253.61 ns 1.00 0.00 - - - 24 B PixelOperations_Specialized Job-YJZMFY .NET Core 3.1 2048 1,588.3 ns 912.20 ns 50.00 ns 0.25 0.01 0.0172 - - 72 B Looks great!
@JimBobSquarePants there is one other important metric we need to check to set our expectations for #1410. Can be done by defining 2 new simple benchmark classes (baseline VS SIMD with Count = 2048):
- Comparing
Vector4 -> Rgba32(baseline) toVector4 -> Rgb24(Jpeg decoder pipeline last step) - Comparing
Rgba32 -> Vector4(baseline) toRgb24 -> Vector4(Resize pipeline first step)
The smaller the difference the bigger the happiness.
- Comparing
Couldn't we pack directly into
Rgb24by not scaling down in the color converter and packing the planar values as bytes?Vector4=>Rgba32is still 2.5x faster on .NET Core 3.1 thanVector4=>Rgb24.We'd need to have a method that goes direct to be able to cut into that since the pipeline is
Vector4=>Rgba32=>Rgb24. Not difficult with hardware intrinsics based on my existing code but likely not fun with the old stuff.What I dream of is a combination of shuffle + convert.
Vector4=>Rgba32is still 2.5x faster on .NET Core 3.1 thanVector4=>Rgb24.That's bad news :(
I don't think going
Vector4 => Rgb24is the best approach here:- It's a huge refactor strongly interfering with Vectorize (AVX2) JPEG Color Converter #1411.
- A single pipeline step would do way too much (convert colorspace + convert to
byte+ pack), making it very hard to maintain the code - Ensuring that the non-HwIntrinsics path does not regress is also very expensive (The
Vector<float>->Vector<byte>conversion shall be directly integrated into the colorspace conversion code.
I'd rather suggest to do the following:
- Color converters should convert and pack to a "
Vector3" buffer (Span<float>of RGB components only, no padding for alpha) - Apply
SimdUtils.NormalizedFloatToByteSaturateto get theRgb24buffer - Convert it further with
PixelOperations<TPixel>ifTPixel != Rgb24
=> Pro: likely still very fast, much more predictible amount of work, no regressions on old platforms
(But still a lot of work!)Yeah... All my shuffle code has touched conversion between pixel formats only. The
Vector4pipeline remains untouched (other than a speedup for converting to/fromRgba32)What do you mean by Color converters? The jpeg ones?
What do you mean by Color converters? The jpeg ones?
Yes I meant that.
Couldn't we pack directly into Rgb24
I think I misunderstood you on this one. I thought you want Jpeg color converters to convert directly into
Rgb24, my #1354 (comment) is listing arguments against doing that.A one-step
Vector4 => Rgb24method can help a bit, but wouldn't expect big miracle from it, choose wisely if you want to invest your time into it or not.Reacted by James Jackson-SouthJust realized this was still open.
We need cheap bulk-conversion between the following formats:
Argb32,Bgra32,Rgba32,Bgr24,Rgb24We are especially interested in
Rgb24, Bgr24 <=> Rgba32. I'm having sweat dreams about a community PR dealing with this problem. Should be easy for anyone having basic knowledge ofSystem.Runtime.Intrinsics.Implementation should be added to new span-based methods in
PixelConverter.cs, those could be invoked than from T4 generatedPixelOperations<TPixel>implementors like here:https://github.com/SixLabors/ImageSharp/blob/78a584e8482b052d7a9885682299e2f37518d83d/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Rgb24.PixelOperations.Generated.cs
If a PR would only add the
PixelConverterhelpers + tests, I'm happy to provide guidance or even finish the code for the rest of the work.@john-h-k @Sergio0694 any chance you are interested?
Tasks:
SimdUtilsSimdUtilsSimdUtilsRgba32compatible pixel operations to utilize new shuffle methods.Rgb24compatible pixel operations to utilize new shuffle methods.