Skip to content

3 <==> 4 Channel Shuffling with Hardware Intrinsics #1409

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Merged
merged 23 commits into from
Nov 6, 2020
Merged

3 <==> 4 Channel Shuffling with Hardware Intrinsics #1409

merged 23 commits into from
Nov 6, 2020

Conversation

@JimBobSquarePants
Copy link
Member

@JimBobSquarePants JimBobSquarePants commented Oct 30, 2020 •
edited
Loading

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 matches 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

Adds three methods to SimdUtils:

  1. Pad3Shuffle4 Pads a buffer representing 3 channel pixels to 4 channels and shuffles the result.
  2. Shuffle4Slice3 Shuffles a buffer representing 4 channel pixels and slices it to 3 channels.
  3. Shuffle3 Shuffles a buffer representing 3 channel pixels.

All fallbacks are better than or equal two current implementations.

~I've completed a basic implementation for now with results in 2.5-3x speedup. I looked into more loop unrolling based on this example but kept messing up my offsets so I left it. @saucecontrol if you fancy helping me here please do. Fixed, thanks @saucecontrol!

Once the initial implementation is reviewed I will make the methods generic and add optimized fallback versions of XYZW shuffling as per @antonfirsov previous suggestions in 4 channel shuffling which I can use custom structs similar to Rgb24 and Rgba32 to provide no loss in current performance on older platforms. Done!

Current Benchmarks

Pad3Shuffle4

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
1. No HwIntrinsics : .NET Core 3.1.9 (CoreCLR 4.700.20.47201, CoreFX 4.700.20.47203), X64 RyuJIT
2. AVX             : .NET Core 3.1.9 (CoreCLR 4.700.20.47201, CoreFX 4.700.20.47203), X64 RyuJIT
3. SSE             : .NET Core 3.1.9 (CoreCLR 4.700.20.47201, CoreFX 4.700.20.47203), X64 RyuJIT

Runtime=.NET Core 3.1
Method Job EnvironmentVariables Count Mean Error StdDev Median Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated
Pad3Shuffle4 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 96 120.64 ns 7.190 ns 21.200 ns 114.26 ns 1.00 0.00 - - - -
Pad3Shuffle4 2. AVX Empty 96 23.63 ns 0.175 ns 0.155 ns 23.65 ns 0.15 0.01 - - - -
Pad3Shuffle4 3. SSE COMPlus_EnableAVX=0 96 25.25 ns 0.356 ns 0.298 ns 25.27 ns 0.17 0.01 - - - -
Pad3Shuffle4FastFallback 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 96 14.80 ns 0.358 ns 1.032 ns 14.64 ns 1.00 0.00 - - - -
Pad3Shuffle4FastFallback 2. AVX Empty 96 24.84 ns 0.376 ns 0.333 ns 24.74 ns 1.57 0.06 - - - -
Pad3Shuffle4FastFallback 3. SSE COMPlus_EnableAVX=0 96 24.58 ns 0.471 ns 0.704 ns 24.38 ns 1.60 0.09 - - - -
Pad3Shuffle4 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 384 258.92 ns 4.873 ns 4.069 ns 257.95 ns 1.00 0.00 - - - -
Pad3Shuffle4 2. AVX Empty 384 41.41 ns 0.859 ns 1.204 ns 41.33 ns 0.16 0.00 - - - -
Pad3Shuffle4 3. SSE COMPlus_EnableAVX=0 384 40.74 ns 0.848 ns 0.793 ns 40.48 ns 0.16 0.00 - - - -
Pad3Shuffle4FastFallback 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 384 74.50 ns 0.490 ns 0.383 ns 74.49 ns 1.00 0.00 - - - -
Pad3Shuffle4FastFallback 2. AVX Empty 384 40.74 ns 0.624 ns 0.584 ns 40.72 ns 0.55 0.01 - - - -
Pad3Shuffle4FastFallback 3. SSE COMPlus_EnableAVX=0 384 38.28 ns 0.534 ns 0.417 ns 38.22 ns 0.51 0.01 - - - -
Pad3Shuffle4 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 768 503.91 ns 6.466 ns 6.048 ns 501.58 ns 1.00 0.00 - - - -
Pad3Shuffle4 2. AVX Empty 768 62.86 ns 0.332 ns 0.277 ns 62.80 ns 0.12 0.00 - - - -
Pad3Shuffle4 3. SSE COMPlus_EnableAVX=0 768 64.59 ns 0.469 ns 0.415 ns 64.62 ns 0.13 0.00 - - - -
Pad3Shuffle4FastFallback 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 768 110.51 ns 0.592 ns 0.554 ns 110.33 ns 1.00 0.00 - - - -
Pad3Shuffle4FastFallback 2. AVX Empty 768 64.72 ns 1.306 ns 1.090 ns 64.51 ns 0.59 0.01 - - - -
Pad3Shuffle4FastFallback 3. SSE COMPlus_EnableAVX=0 768 62.11 ns 0.816 ns 0.682 ns 61.98 ns 0.56 0.01 - - - -
Pad3Shuffle4 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 1536 1,005.84 ns 13.176 ns 12.325 ns 1,004.70 ns 1.00 0.00 - - - -
Pad3Shuffle4 2. AVX Empty 1536 110.05 ns 0.256 ns 0.214 ns 110.04 ns 0.11 0.00 - - - -
Pad3Shuffle4 3. SSE COMPlus_EnableAVX=0 1536 110.23 ns 0.545 ns 0.483 ns 110.09 ns 0.11 0.00 - - - -
Pad3Shuffle4FastFallback 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 1536 220.37 ns 1.601 ns 1.419 ns 220.13 ns 1.00 0.00 - - - -
Pad3Shuffle4FastFallback 2. AVX Empty 1536 111.54 ns 2.173 ns 2.901 ns 111.27 ns 0.51 0.01 - - - -
Pad3Shuffle4FastFallback 3. SSE COMPlus_EnableAVX=0 1536 110.23 ns 0.456 ns 0.427 ns 110.25 ns 0.50 0.00 - - - -

Shuffle4Slice3

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
1. No HwIntrinsics : .NET Core 3.1.9 (CoreCLR 4.700.20.47201, CoreFX 4.700.20.47203), X64 RyuJIT
2. AVX             : .NET Core 3.1.9 (CoreCLR 4.700.20.47201, CoreFX 4.700.20.47203), X64 RyuJIT
3. SSE             : .NET Core 3.1.9 (CoreCLR 4.700.20.47201, CoreFX 4.700.20.47203), X64 RyuJIT

Runtime=.NET Core 3.1
Method Job EnvironmentVariables Count Mean Error StdDev Median Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated
Shuffle4Slice3 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 128 56.44 ns 2.843 ns 8.382 ns 56.70 ns 1.00 0.00 - - - -
Shuffle4Slice3 2. AVX Empty 128 27.15 ns 0.556 ns 0.762 ns 27.34 ns 0.41 0.03 - - - -
Shuffle4Slice3 3. SSE COMPlus_EnableAVX=0 128 26.36 ns 0.321 ns 0.268 ns 26.26 ns 0.38 0.02 - - - -
Shuffle4Slice3FastFallback 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 128 25.85 ns 0.494 ns 0.462 ns 25.84 ns 1.00 0.00 - - - -
Shuffle4Slice3FastFallback 2. AVX Empty 128 26.15 ns 0.113 ns 0.106 ns 26.16 ns 1.01 0.02 - - - -
Shuffle4Slice3FastFallback 3. SSE COMPlus_EnableAVX=0 128 25.57 ns 0.078 ns 0.061 ns 25.56 ns 0.99 0.02 - - - -
Shuffle4Slice3 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 256 97.47 ns 0.327 ns 0.289 ns 97.35 ns 1.00 0.00 - - - -
Shuffle4Slice3 2. AVX Empty 256 32.61 ns 0.107 ns 0.095 ns 32.62 ns 0.33 0.00 - - - -
Shuffle4Slice3 3. SSE COMPlus_EnableAVX=0 256 33.21 ns 0.169 ns 0.150 ns 33.15 ns 0.34 0.00 - - - -
Shuffle4Slice3FastFallback 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 256 52.34 ns 0.779 ns 0.729 ns 51.94 ns 1.00 0.00 - - - -
Shuffle4Slice3FastFallback 2. AVX Empty 256 32.16 ns 0.111 ns 0.104 ns 32.16 ns 0.61 0.01 - - - -
Shuffle4Slice3FastFallback 3. SSE COMPlus_EnableAVX=0 256 33.61 ns 0.342 ns 0.319 ns 33.62 ns 0.64 0.01 - - - -
Shuffle4Slice3 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 512 210.74 ns 3.825 ns 5.956 ns 207.70 ns 1.00 0.00 - - - -
Shuffle4Slice3 2. AVX Empty 512 51.03 ns 0.535 ns 0.501 ns 51.18 ns 0.24 0.01 - - - -
Shuffle4Slice3 3. SSE COMPlus_EnableAVX=0 512 66.60 ns 1.313 ns 1.613 ns 65.93 ns 0.31 0.01 - - - -
Shuffle4Slice3FastFallback 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 512 119.12 ns 1.905 ns 1.689 ns 118.52 ns 1.00 0.00 - - - -
Shuffle4Slice3FastFallback 2. AVX Empty 512 50.33 ns 0.382 ns 0.339 ns 50.41 ns 0.42 0.01 - - - -
Shuffle4Slice3FastFallback 3. SSE COMPlus_EnableAVX=0 512 49.25 ns 0.555 ns 0.492 ns 49.26 ns 0.41 0.01 - - - -
Shuffle4Slice3 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 1024 423.55 ns 4.891 ns 4.336 ns 423.27 ns 1.00 0.00 - - - -
Shuffle4Slice3 2. AVX Empty 1024 77.13 ns 1.355 ns 2.264 ns 76.19 ns 0.19 0.01 - - - -
Shuffle4Slice3 3. SSE COMPlus_EnableAVX=0 1024 79.39 ns 0.103 ns 0.086 ns 79.37 ns 0.19 0.00 - - - -
Shuffle4Slice3FastFallback 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 1024 226.57 ns 2.930 ns 2.598 ns 226.10 ns 1.00 0.00 - - - -
Shuffle4Slice3FastFallback 2. AVX Empty 1024 80.25 ns 1.647 ns 2.082 ns 80.98 ns 0.35 0.01 - - - -
Shuffle4Slice3FastFallback 3. SSE COMPlus_EnableAVX=0 1024 84.99 ns 1.234 ns 1.155 ns 85.60 ns 0.38 0.01 - - - -
Shuffle4Slice3 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 2048 794.96 ns 1.735 ns 1.538 ns 795.15 ns 1.00 0.00 - - - -
Shuffle4Slice3 2. AVX Empty 2048 128.41 ns 0.417 ns 0.390 ns 128.24 ns 0.16 0.00 - - - -
Shuffle4Slice3 3. SSE COMPlus_EnableAVX=0 2048 127.24 ns 0.294 ns 0.229 ns 127.23 ns 0.16 0.00 - - - -
Shuffle4Slice3FastFallback 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 2048 382.97 ns 1.064 ns 0.831 ns 382.87 ns 1.00 0.00 - - - -
Shuffle4Slice3FastFallback 2. AVX Empty 2048 126.93 ns 0.382 ns 0.339 ns 126.94 ns 0.33 0.00 - - - -
Shuffle4Slice3FastFallback 3. SSE COMPlus_EnableAVX=0 2048 149.36 ns 1.875 ns 1.754 ns 149.33 ns 0.39 0.00 - - - -

Shuffle3

Note: No fast fallback implementation here. I experimented with casting to uint and using shuffling but it was a shade slower than the default.

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
1. No HwIntrinsics : .NET Core 3.1.9 (CoreCLR 4.700.20.47201, CoreFX 4.700.20.47203), X64 RyuJIT
2. AVX             : .NET Core 3.1.9 (CoreCLR 4.700.20.47201, CoreFX 4.700.20.47203), X64 RyuJIT
3. SSE             : .NET Core 3.1.9 (CoreCLR 4.700.20.47201, CoreFX 4.700.20.47203), X64 RyuJIT

Runtime=.NET Core 3.1
Method Job EnvironmentVariables Count Mean Error StdDev Median Ratio RatioSD Gen 0 Gen 1 Gen 2 Allocated
Shuffle3 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 96 48.46 ns 1.034 ns 2.438 ns 47.46 ns 1.00 0.00 - - - -
Shuffle3 2. AVX Empty 96 32.42 ns 0.537 ns 0.476 ns 32.34 ns 0.66 0.04 - - - -
Shuffle3 3. SSE COMPlus_EnableAVX=0 96 32.51 ns 0.373 ns 0.349 ns 32.56 ns 0.66 0.03 - - - -
Shuffle3 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 384 199.04 ns 1.512 ns 1.180 ns 199.17 ns 1.00 0.00 - - - -
Shuffle3 2. AVX Empty 384 71.20 ns 2.654 ns 7.784 ns 69.60 ns 0.41 0.02 - - - -
Shuffle3 3. SSE COMPlus_EnableAVX=0 384 63.23 ns 0.569 ns 0.505 ns 63.21 ns 0.32 0.00 - - - -
Shuffle3 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 768 391.28 ns 5.087 ns 3.972 ns 391.22 ns 1.00 0.00 - - - -
Shuffle3 2. AVX Empty 768 109.12 ns 2.149 ns 2.010 ns 108.66 ns 0.28 0.01 - - - -
Shuffle3 3. SSE COMPlus_EnableAVX=0 768 106.51 ns 0.734 ns 0.613 ns 106.56 ns 0.27 0.00 - - - -
Shuffle3 1. No HwIntrinsics COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 1536 773.70 ns 5.516 ns 4.890 ns 772.96 ns 1.00 0.00 - - - -
Shuffle3 2. AVX Empty 1536 190.41 ns 1.090 ns 0.851 ns 190.38 ns 0.25 0.00 - - - -
Shuffle3 3. SSE COMPlus_EnableAVX=0 1536 190.94 ns 0.985 ns 0.769 ns 190.85 ns 0.25 0.00 - - - -

@JimBobSquarePants JimBobSquarePants added this to the 1.1.0 milestone Oct 30, 2020
@JimBobSquarePants JimBobSquarePants requested a review from a team October 30, 2020 23:20
@JimBobSquarePants JimBobSquarePants changed the title [Draft] 3 <==> 4 Channel Shuffling with Harware Intrinsics [Draft] 3 <==> 4 Channel Shuffling with Hardware Intrinsics Oct 30, 2020
@codecov
Copy link

codecov bot commented Oct 30, 2020 •
edited
Loading

Codecov Report

Merging #1409 (3cda066) into master (cf9cc6b) will increase coverage by 0.14%.
The diff coverage is 99.33%.

Impacted file tree graph

@@            Coverage Diff             @@
##           master    #1409      +/-   ##
==========================================
+ Coverage   82.99%   83.14%   +0.14%     
==========================================
  Files         692      695       +3     
  Lines       31189    31484     +295     
  Branches     3578     3586       +8     
==========================================
+ Hits        25884    26176     +292     
- Misses       4582     4585       +3     
  Partials      723      723
Flag Coverage Δ
unittests 83.14% <99.33%> (+0.14%) ⬆️

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

Impacted Files Coverage Δ
src/ImageSharp/Common/Helpers/SimdUtils.Shuffle.cs 92.30% <83.33%> (-4.99%) ⬇️
...eSharp/Common/Helpers/Shuffle/IComponentShuffle.cs 100.00% <100.00%> (ø)
...ImageSharp/Common/Helpers/Shuffle/IPad3Shuffle4.cs 100.00% <100.00%> (ø)
src/ImageSharp/Common/Helpers/Shuffle/IShuffle3.cs 100.00% <100.00%> (ø)
...ageSharp/Common/Helpers/Shuffle/IShuffle4Slice3.cs 100.00% <100.00%> (ø)
...mageSharp/Common/Helpers/SimdUtils.HwIntrinsics.cs 97.80% <100.00%> (+1.28%) ⬆️
...ions/Generated/Argb32.PixelOperations.Generated.cs 100.00% <100.00%> (ø)
...tions/Generated/Bgr24.PixelOperations.Generated.cs 100.00% <100.00%> (ø)
...ions/Generated/Bgra32.PixelOperations.Generated.cs 100.00% <100.00%> (ø)
...tions/Generated/Rgb24.PixelOperations.Generated.cs 100.00% <100.00%> (ø)
... and 5 more

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 cf9cc6b...3cda066. Read the comment docs.

saucecontrol
saucecontrol reviewed Oct 31, 2020
View reviewed changes
saucecontrol
saucecontrol reviewed Oct 31, 2020
View reviewed changes
@antonfirsov antonfirsov mentioned this pull request Nov 2, 2020
Closed
@JimBobSquarePants JimBobSquarePants changed the title [Draft] 3 <==> 4 Channel Shuffling with Hardware Intrinsics 3 <==> 4 Channel Shuffling with Hardware Intrinsics Nov 2, 2020
@JimBobSquarePants JimBobSquarePants marked this pull request as ready for review November 2, 2020 16:50
@antonfirsov
Copy link
Member

antonfirsov commented Nov 2, 2020

@JimBobSquarePants I want to do a proper review here, but not sure about the timing. Lucky case: by Wednesday evening, worst case: by Sunday. Is this acceptable for you?

@JimBobSquarePants
Copy link
Member Author

JimBobSquarePants commented Nov 2, 2020

@antonfirsov No worries. Hopefully someone else from the team can also have a look in the interim.

I added the Rgb24/Vector4 benchmarks to #1354 (comment)

We're looking at a very healthy speedup on all target frameworks.

@antonfirsov
Copy link
Member

antonfirsov commented Nov 2, 2020 •
edited
Loading

@JimBobSquarePants I tried to destile some derived numbers to make conclusions from #1354 (comment):

ToVector4_Rgb24

Count    Runtime Master Branch Speedup (Master/Branch)
64 .NET 4.7.2 310.2 355.5 0.87257384
64 .NET Core 2.1 230.2 228.5 1.007439825
64 .NET Core 3.1 236.7 217 1.09078341
256 .NET 4.7.2 622.5 448.9 1.386723101
256 .NET Core 2.1 498.1 309.2 1.610931436
256 .NET Core 3.1 436.8 212.3 2.05746585
2048 .NET 4.7.2 3,460.60 1,974.10 1.753001368
2048 .NET Core 2.1 3,421.30 1,985.50 1.723142785
2048 .NET Core 3.1 2,972.20 1,165.00 2.551244635

FromVector4_Rgb24

Count  Runtime Master Branch Speedup (Master/Branch)
64 .NET 4.7.2 316.4 320.8 0.986284289
64 .NET Core 2.1 238.9 246 0.971138211
64 .NET Core 3.1 250.3 243.4 1.028348398
256 .NET 4.7.2 1,051.30 967 1.087176836
256 .NET Core 2.1 846.8 1,003.30 0.844014751
256 .NET Core 3.1 640.2 437 1.464988558
2048 .NET 4.7.2 4,551.20 4,391.60 1.036342108
2048 .NET Core 2.1 4,390.40 4,225.60 1.039000379
2048 .NET Core 3.1 2,979.40 1,822.70 1.634607999

Things I don't understand:

  • Why do the benchmarks show speedup for 2.1 and .NET Framework for Count >= 256? There is no SIMD, and it doesn't look like we changed something there.
  • Is there a regression in FromVector4_Rgb24 for 2.1 with Count == 256? (Or is it noise?)

@JimBobSquarePants
Copy link
Member Author

JimBobSquarePants commented Nov 2, 2020 •
edited
Loading

@antonfirsov the speedup is due to this change. Our original per pixel implementation turned out to be quite slow.

https://github.com/SixLabors/ImageSharp/pull/1409/files#diff-92b35c7d3a5d6901602e57d9dfd618b0d6ceb8cf3442fcfa044111c0e1f509dfR73

The regression is just noise. There’s massive error on that run for some reason

@peter-dolkens
Copy link

peter-dolkens commented Nov 3, 2020 •
edited
Loading

The "Fastfallback" is a regression in performance - presumably because it is no longer falling back, but doing it properly, quickly.

Is this only ever triggered algorithmically?

Could people be running this in fallback mode deliberately as it's faster and close enough for their needs?

Should you be retaining the "Fastfallback" path, or happy to discontinue that support.

For simplicity, I'd certainly lean towards making it obsolete, but you know your library better than anyone.

Not sure if it's the kind of thing that might deliberately get used in low-quality bulk tasks like thumbnail generation.

@JimBobSquarePants
Copy link
Member Author

JimBobSquarePants commented Nov 4, 2020 •
edited
Loading

@peter-dolkens if you check the linked benchmarks the error for that regression run is super high. Likely due to my laptop throttling. I’ll post updated ones in the morning

Those fast fallbacks are tuned algorithms for particular shuffle operations. They only kick in on target frameworks without intrinsics and when there are remaining items we cannot process in bulk as they don’t fit within a vector.

@JimBobSquarePants
Copy link
Member Author

JimBobSquarePants commented Nov 6, 2020

Oh wow! I really broke this!

@JimBobSquarePants JimBobSquarePants merged commit 9f51a92 into master Nov 6, 2020
@JimBobSquarePants JimBobSquarePants deleted the js/Shuffle3Channel branch November 6, 2020 21:32
JimBobSquarePants added a commit that referenced this pull request Mar 13, 2021
@JimBobSquarePants
Merge pull request #1409 from SixLabors/js/Shuffle3Channel
522a91e 
3 <==> 4 Channel Shuffling with Hardware Intrinsics
This was referenced Jul 30, 2025
Merged
Open
Open
Open
Open
Open
Open
Open
Open
Open
Open
Open
Open
Open
Open
@dependabot dependabot bot mentioned this pull request Sep 23, 2025
Open
This was referenced Oct 3, 2025
Open
Merged
Closed
Open
@dependabot dependabot bot mentioned this pull request Oct 19, 2025
Open
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@antonfirsov antonfirsov antonfirsov approved these changes

+1 more reviewer

@saucecontrol saucecontrol saucecontrol left review comments

Reviewers whose approvals may not affect merge requirements

Assignees

No one assigned

Labels

area:performance

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

5 participants

@JimBobSquarePants @antonfirsov @peter-dolkens @saucecontrol