Rewrite GetIndexOfFirstNonAsciiByte #104503

neon-sunset · 2024-07-06T06:58:58Z

Description

The purpose of this change is to simplify GetIndexOfFirstNonAsciiByte by servicing all targets with a single path, which is now possible, and achieving optimal throughput on longer lengths.
Per "Validating gigabytes of Unicode strings per second… in C#?", it appears that we leave a lot of performance on the table for ARM64 target.

Also closes #89924

Analysis

After addressing feedback, this PR now improves maximum non-ASCII byte search throughput by 3x for AVX2 (Zen 3) and AdvSimd (Neoverse N1), and 2x for AVX512 (Ice Lake). Short lengths improve by up to 25%, and the worst case regression is on AMD only for <5% for lengths ~65-200 (under 1ns impact).

Benchmark

[SimpleJob]
public class AsciiBench
{
    [ParamsSource(nameof(Lengths))]
    public int Length;
    public IEnumerable<int> Lengths => [..Enumerable.Range(1, 8), 12, 16, 25, 50, 70, 90, 200, 512, 16384];

    byte[] ascii = [];

    [GlobalSetup]
    public void Setup()
    {
        ascii = new byte[Length];
        ascii.AsSpan().Fill((byte)'f');
    }

    [Benchmark]
    public bool ValidateAscii()
    {
        return Utf8.IsValid(ascii);
    }
}

Results

Ice Lake, Zen 3 and N1: #104503 (comment)

M1 Pro

BenchmarkDotNet v0.13.12, macOS 15.0 (24A5289g) [Darwin 24.0.0]
Apple M1 Pro, 1 CPU, 8 logical and 8 physical cores
.NET SDK 9.0.100-preview.7.24354.8
  [Host]     : .NET 9.0.0 (9.0.24.35215), Arm64 RyuJIT AdvSIMD
  DefaultJob : .NET 9.0.0 (9.0.24.35215), Arm64 RyuJIT AdvSIMD
  Job-PJSXIE : .NET 9.0.0 (42.42.42.42424), Arm64 RyuJIT AdvSIMD

Method	Job	Toolchain	Length	Mean	Error	StdDev
ValidateAscii	DefaultJob	Default	1	2.336 ns	0.0087 ns	0.0073 ns
ValidateAscii	Job-PJSXIE	CoreRun	1	2.345 ns	0.0062 ns	0.0052 ns
ValidateAscii	DefaultJob	Default	2	2.478 ns	0.0026 ns	0.0025 ns
ValidateAscii	Job-PJSXIE	CoreRun	2	2.355 ns	0.0033 ns	0.0030 ns
ValidateAscii	DefaultJob	Default	3	2.634 ns	0.0059 ns	0.0050 ns
ValidateAscii	Job-PJSXIE	CoreRun	3	2.507 ns	0.0025 ns	0.0022 ns
ValidateAscii	DefaultJob	Default	4	2.397 ns	0.0030 ns	0.0026 ns
ValidateAscii	Job-PJSXIE	CoreRun	4	2.347 ns	0.0020 ns	0.0019 ns
ValidateAscii	DefaultJob	Default	5	2.537 ns	0.0029 ns	0.0024 ns
ValidateAscii	Job-PJSXIE	CoreRun	5	2.462 ns	0.0047 ns	0.0042 ns
ValidateAscii	DefaultJob	Default	6	2.638 ns	0.0054 ns	0.0050 ns
ValidateAscii	Job-PJSXIE	CoreRun	6	2.517 ns	0.0016 ns	0.0014 ns
ValidateAscii	DefaultJob	Default	7	2.645 ns	0.0044 ns	0.0042 ns
ValidateAscii	Job-PJSXIE	CoreRun	7	2.666 ns	0.0048 ns	0.0045 ns
ValidateAscii	DefaultJob	Default	8	2.476 ns	0.0024 ns	0.0022 ns
ValidateAscii	Job-PJSXIE	CoreRun	8	2.435 ns	0.0015 ns	0.0014 ns
ValidateAscii	DefaultJob	Default	12	2.641 ns	0.0126 ns	0.0105 ns
ValidateAscii	Job-PJSXIE	CoreRun	12	2.432 ns	0.0033 ns	0.0030 ns
ValidateAscii	DefaultJob	Default	16	2.720 ns	0.0048 ns	0.0045 ns
ValidateAscii	Job-PJSXIE	CoreRun	16	2.435 ns	0.0033 ns	0.0029 ns
ValidateAscii	DefaultJob	Default	25	3.307 ns	0.0056 ns	0.0047 ns
ValidateAscii	Job-PJSXIE	CoreRun	25	3.358 ns	0.0050 ns	0.0047 ns
ValidateAscii	DefaultJob	Default	50	4.541 ns	0.0055 ns	0.0046 ns
ValidateAscii	Job-PJSXIE	CoreRun	50	3.618 ns	0.0055 ns	0.0052 ns
ValidateAscii	DefaultJob	Default	70	5.022 ns	0.0054 ns	0.0050 ns
ValidateAscii	Job-PJSXIE	CoreRun	70	3.993 ns	0.0062 ns	0.0055 ns
ValidateAscii	DefaultJob	Default	90	5.486 ns	0.0083 ns	0.0077 ns
ValidateAscii	Job-PJSXIE	CoreRun	90	4.674 ns	0.0050 ns	0.0047 ns
ValidateAscii	DefaultJob	Default	200	8.652 ns	0.0082 ns	0.0073 ns
ValidateAscii	Job-PJSXIE	CoreRun	200	5.941 ns	0.0046 ns	0.0041 ns
ValidateAscii	DefaultJob	Default	512	17.469 ns	0.0360 ns	0.0319 ns
ValidateAscii	Job-PJSXIE	CoreRun	512	8.589 ns	0.0323 ns	0.0270 ns
ValidateAscii	DefaultJob	Default	16384	473.304 ns	2.2263 ns	1.9736 ns
ValidateAscii	Job-PJSXIE	CoreRun	16384	169.737 ns	0.2044 ns	0.1912 ns

- Deduplicate SIMD paths - there is no more need for that - Remove at least one branch from short length path - Use up to 128x4/256x2 SIMD unrolling

EgorBot · 2024-07-06T09:45:23Z

Benchmark results on Arm64

BenchmarkDotNet v0.13.12, Ubuntu 22.04.4 LTS (Jammy Jellyfish)
Unknown processor
  Job-IDDVGL : .NET 9.0.0 (42.42.42.42424), Arm64 RyuJIT AdvSIMD
  Job-BQBANQ : .NET 9.0.0 (42.42.42.42424), Arm64 RyuJIT AdvSIMD

Method	Toolchain	Length	Mean	Error	Ratio
ValidateAscii	Main	1	6.245 ns	0.0085 ns	1.00
ValidateAscii	PR	1	5.958 ns	0.0028 ns	0.95

ValidateAscii	Main	2	6.396 ns	0.0069 ns	1.00
ValidateAscii	PR	2	6.244 ns	0.0142 ns	0.98

ValidateAscii	Main	3	6.533 ns	0.0019 ns	1.00
ValidateAscii	PR	3	6.281 ns	0.0179 ns	0.96

ValidateAscii	Main	4	6.358 ns	0.0051 ns	1.00
ValidateAscii	PR	4	6.027 ns	0.0039 ns	0.95

ValidateAscii	Main	5	6.569 ns	0.0021 ns	1.00
ValidateAscii	PR	5	6.220 ns	0.0041 ns	0.95

ValidateAscii	Main	6	6.813 ns	0.0075 ns	1.00
ValidateAscii	PR	6	6.431 ns	0.0046 ns	0.94

ValidateAscii	Main	7	6.922 ns	0.0011 ns	1.00
ValidateAscii	PR	7	6.472 ns	0.0081 ns	0.93

ValidateAscii	Main	8	6.287 ns	0.0014 ns	1.00
ValidateAscii	PR	8	6.245 ns	0.0067 ns	0.99

ValidateAscii	Main	9	6.496 ns	0.0028 ns	1.00
ValidateAscii	PR	9	6.664 ns	0.0067 ns	1.03

ValidateAscii	Main	10	6.689 ns	0.0037 ns	1.00
ValidateAscii	PR	10	6.682 ns	0.0082 ns	1.00

ValidateAscii	Main	11	6.938 ns	0.0018 ns	1.00
ValidateAscii	PR	11	6.653 ns	0.0064 ns	0.96

ValidateAscii	Main	12	6.657 ns	0.0016 ns	1.00
ValidateAscii	PR	12	6.570 ns	0.0070 ns	0.99

ValidateAscii	Main	13	6.865 ns	0.0008 ns	1.00
ValidateAscii	PR	13	6.649 ns	0.0071 ns	0.97

ValidateAscii	Main	14	7.190 ns	0.0054 ns	1.00
ValidateAscii	PR	14	6.649 ns	0.0063 ns	0.92

ValidateAscii	Main	15	7.478 ns	0.0007 ns	1.00
ValidateAscii	PR	15	6.647 ns	0.0078 ns	0.89

ValidateAscii	Main	16	7.172 ns	0.0009 ns	1.00
ValidateAscii	PR	16	6.567 ns	0.0043 ns	0.92

ValidateAscii	Main	24	9.155 ns	0.0014 ns	1.00
ValidateAscii	PR	24	9.153 ns	0.0025 ns	1.00

ValidateAscii	Main	48	9.968 ns	0.0024 ns	1.00
ValidateAscii	PR	48	10.684 ns	0.0120 ns	1.07

ValidateAscii	Main	64	10.775 ns	0.0012 ns	1.00
ValidateAscii	PR	64	10.723 ns	0.0019 ns	1.00

ValidateAscii	Main	72	12.227 ns	0.0035 ns	1.00
ValidateAscii	PR	72	13.220 ns	0.0089 ns	1.08

ValidateAscii	Main	96	12.649 ns	0.0026 ns	1.00
ValidateAscii	PR	96	13.110 ns	0.0117 ns	1.04

ValidateAscii	Main	256	23.221 ns	0.0076 ns	1.00
ValidateAscii	PR	256	16.474 ns	0.0115 ns	0.71

ValidateAscii	Main	512	39.824 ns	0.0184 ns	1.00
ValidateAscii	PR	512	21.312 ns	0.0190 ns	0.54

ValidateAscii	Main	2048	139.264 ns	0.0350 ns	1.00
ValidateAscii	PR	2048	56.632 ns	0.0088 ns	0.41

ValidateAscii	Main	16384	1,072.289 ns	0.0625 ns	1.00
ValidateAscii	PR	16384	373.391 ns	0.1192 ns	0.35

BDN_Artifacts.zip

EgorBot · 2024-07-06T09:47:34Z

Benchmark results on Amd

BenchmarkDotNet v0.13.12, Ubuntu 22.04.4 LTS (Jammy Jellyfish)
AMD EPYC 7763, 1 CPU, 8 logical and 4 physical cores
  Job-HPNOVE : .NET 9.0.0 (42.42.42.42424), X64 RyuJIT AVX2
  Job-PZAFXR : .NET 9.0.0 (42.42.42.42424), X64 RyuJIT AVX2

Method	Toolchain	Length	Mean	Error	Ratio
ValidateAscii	Main	1	5.271 ns	0.0009 ns	1.00
ValidateAscii	PR	1	4.979 ns	0.0012 ns	0.94

ValidateAscii	Main	2	5.595 ns	0.0016 ns	1.00
ValidateAscii	PR	2	4.973 ns	0.0006 ns	0.89

ValidateAscii	Main	3	5.271 ns	0.0018 ns	1.00
ValidateAscii	PR	3	4.967 ns	0.0018 ns	0.94

ValidateAscii	Main	4	5.281 ns	0.0012 ns	1.00
ValidateAscii	PR	4	4.661 ns	0.0014 ns	0.88

ValidateAscii	Main	5	5.287 ns	0.0018 ns	1.00
ValidateAscii	PR	5	4.978 ns	0.0018 ns	0.94

ValidateAscii	Main	6	5.595 ns	0.0018 ns	1.00
ValidateAscii	PR	6	4.979 ns	0.0030 ns	0.89

ValidateAscii	Main	7	5.578 ns	0.0023 ns	1.00
ValidateAscii	PR	7	5.278 ns	0.0017 ns	0.95

ValidateAscii	Main	8	5.610 ns	0.0012 ns	1.00
ValidateAscii	PR	8	5.010 ns	0.0040 ns	0.89

ValidateAscii	Main	9	5.594 ns	0.0016 ns	1.00
ValidateAscii	PR	9	4.975 ns	0.0021 ns	0.89

ValidateAscii	Main	10	5.897 ns	0.0022 ns	1.00
ValidateAscii	PR	10	4.965 ns	0.0009 ns	0.84

ValidateAscii	Main	11	6.203 ns	0.0010 ns	1.00
ValidateAscii	PR	11	4.985 ns	0.0013 ns	0.80

ValidateAscii	Main	12	5.905 ns	0.0015 ns	1.00
ValidateAscii	PR	12	4.984 ns	0.0015 ns	0.84

ValidateAscii	Main	13	5.897 ns	0.0021 ns	1.00
ValidateAscii	PR	13	4.984 ns	0.0021 ns	0.85

ValidateAscii	Main	14	6.216 ns	0.0016 ns	1.00
ValidateAscii	PR	14	4.976 ns	0.0007 ns	0.80

ValidateAscii	Main	15	5.902 ns	0.0013 ns	1.00
ValidateAscii	PR	15	4.979 ns	0.0018 ns	0.84

ValidateAscii	Main	16	5.909 ns	0.0019 ns	1.00
ValidateAscii	PR	16	4.976 ns	0.0015 ns	0.84

ValidateAscii	Main	24	6.519 ns	0.0011 ns	1.00
ValidateAscii	PR	24	6.526 ns	0.0017 ns	1.00

ValidateAscii	Main	48	6.846 ns	0.0018 ns	1.00
ValidateAscii	PR	48	6.536 ns	0.0038 ns	0.95

ValidateAscii	Main	64	7.154 ns	0.0015 ns	1.00
ValidateAscii	PR	64	6.540 ns	0.0032 ns	0.91

ValidateAscii	Main	72	7.772 ns	0.0024 ns	1.00
ValidateAscii	PR	72	6.852 ns	0.0023 ns	0.88

ValidateAscii	Main	96	7.779 ns	0.0033 ns	1.00
ValidateAscii	PR	96	6.844 ns	0.0025 ns	0.88

ValidateAscii	Main	256	9.078 ns	0.0095 ns	1.00
ValidateAscii	PR	256	7.770 ns	0.0030 ns	0.86

ValidateAscii	Main	512	13.534 ns	0.1369 ns	1.00
ValidateAscii	PR	512	9.774 ns	0.0177 ns	0.72

ValidateAscii	Main	2048	41.238 ns	0.0177 ns	1.00
ValidateAscii	PR	2048	24.271 ns	0.0104 ns	0.59

ValidateAscii	Main	16384	333.180 ns	2.4574 ns	1.00
ValidateAscii	PR	16384	171.384 ns	0.0595 ns	0.51

BDN_Artifacts.zip

…512b loop, mitigate Dynamic PGO impact

EgorBot · 2024-07-09T14:37:54Z

Benchmark results on Arm64

BenchmarkDotNet v0.13.12, Ubuntu 22.04.4 LTS (Jammy Jellyfish)
Unknown processor
  Job-REYHIP : .NET 9.0.0 (42.42.42.42424), Arm64 RyuJIT AdvSIMD
  Job-CQPRHS : .NET 9.0.0 (42.42.42.42424), Arm64 RyuJIT AdvSIMD

Method	Toolchain	Length	Mean	Error	Ratio
ValidateAscii	Main	1	5.376 ns	0.0056 ns	1.00
ValidateAscii	PR	1	5.172 ns	0.0052 ns	0.96

ValidateAscii	Main	2	5.448 ns	0.0033 ns	1.00
ValidateAscii	PR	2	5.262 ns	0.0304 ns	0.97

ValidateAscii	Main	3	5.717 ns	0.0056 ns	1.00
ValidateAscii	PR	3	5.488 ns	0.0038 ns	0.96

ValidateAscii	Main	4	5.454 ns	0.0040 ns	1.00
ValidateAscii	PR	4	5.189 ns	0.0016 ns	0.95

ValidateAscii	Main	5	5.743 ns	0.0037 ns	1.00
ValidateAscii	PR	5	5.465 ns	0.0035 ns	0.95

ValidateAscii	Main	6	5.739 ns	0.0014 ns	1.00
ValidateAscii	PR	6	5.526 ns	0.0039 ns	0.96

ValidateAscii	Main	7	6.041 ns	0.0056 ns	1.00
ValidateAscii	PR	7	5.724 ns	0.0031 ns	0.95

ValidateAscii	Main	8	5.454 ns	0.0020 ns	1.00
ValidateAscii	PR	8	5.454 ns	0.0047 ns	1.00

ValidateAscii	Main	12	5.663 ns	0.0037 ns	1.00
ValidateAscii	PR	12	5.479 ns	0.0061 ns	0.97

ValidateAscii	Main	16	6.494 ns	0.0075 ns	1.00
ValidateAscii	PR	16	5.451 ns	0.0065 ns	0.84

ValidateAscii	Main	25	9.342 ns	0.0029 ns	1.00
ValidateAscii	PR	25	8.386 ns	0.0007 ns	0.90

ValidateAscii	Main	50	11.428 ns	0.0063 ns	1.00
ValidateAscii	PR	50	8.749 ns	0.0014 ns	0.77

ValidateAscii	Main	70	12.374 ns	0.0039 ns	1.00
ValidateAscii	PR	70	10.612 ns	0.0019 ns	0.86

ValidateAscii	Main	200	21.311 ns	0.0017 ns	1.00
ValidateAscii	PR	200	14.255 ns	0.0055 ns	0.67

ValidateAscii	Main	512	42.772 ns	0.0038 ns	1.00
ValidateAscii	PR	512	20.237 ns	0.0030 ns	0.47

ValidateAscii	Main	16384	1,065.931 ns	0.1022 ns	1.00
ValidateAscii	PR	16384	339.312 ns	0.0216 ns	0.32

BDN_Artifacts.zip

EgorBot · 2024-07-09T14:38:09Z

Benchmark results on Amd

BenchmarkDotNet v0.13.12, Ubuntu 22.04.4 LTS (Jammy Jellyfish)
AMD EPYC 7763, 1 CPU, 8 logical and 4 physical cores
  Job-UWXNYT : .NET 9.0.0 (42.42.42.42424), X64 RyuJIT AVX2
  Job-XOGXAK : .NET 9.0.0 (42.42.42.42424), X64 RyuJIT AVX2

Method	Toolchain	Length	Mean	Error	Ratio
ValidateAscii	Main	1	4.951 ns	0.0003 ns	1.00
ValidateAscii	PR	1	4.365 ns	0.0004 ns	0.88

ValidateAscii	Main	2	4.965 ns	0.0008 ns	1.00
ValidateAscii	PR	2	3.946 ns	0.0010 ns	0.79

ValidateAscii	Main	3	4.982 ns	0.0011 ns	1.00
ValidateAscii	PR	3	4.666 ns	0.0008 ns	0.94

ValidateAscii	Main	4	4.994 ns	0.0005 ns	1.00
ValidateAscii	PR	4	4.373 ns	0.0009 ns	0.88

ValidateAscii	Main	5	4.981 ns	0.0012 ns	1.00
ValidateAscii	PR	5	4.676 ns	0.0106 ns	0.94

ValidateAscii	Main	6	4.977 ns	0.0010 ns	1.00
ValidateAscii	PR	6	4.652 ns	0.0031 ns	0.93

ValidateAscii	Main	7	5.251 ns	0.0012 ns	1.00
ValidateAscii	PR	7	4.674 ns	0.0005 ns	0.89

ValidateAscii	Main	8	5.283 ns	0.0011 ns	1.00
ValidateAscii	PR	8	4.366 ns	0.0006 ns	0.83

ValidateAscii	Main	12	5.564 ns	0.0005 ns	1.00
ValidateAscii	PR	12	4.365 ns	0.0005 ns	0.78

ValidateAscii	Main	16	5.981 ns	0.0008 ns	1.00
ValidateAscii	PR	16	4.369 ns	0.0014 ns	0.73

ValidateAscii	Main	25	6.526 ns	0.0014 ns	1.00
ValidateAscii	PR	25	5.582 ns	0.0007 ns	0.86

ValidateAscii	Main	50	6.209 ns	0.0015 ns	1.00
ValidateAscii	PR	50	5.583 ns	0.0008 ns	0.90

ValidateAscii	Main	70	5.607 ns	0.0014 ns	1.00
ValidateAscii	PR	70	6.203 ns	0.0007 ns	1.11

ValidateAscii	Main	200	7.132 ns	0.0010 ns	1.00
ValidateAscii	PR	200	6.818 ns	0.0006 ns	0.96

ValidateAscii	Main	512	11.592 ns	0.0144 ns	1.00
ValidateAscii	PR	512	8.678 ns	0.0021 ns	0.75

ValidateAscii	Main	16384	324.580 ns	0.3885 ns	1.00
ValidateAscii	PR	16384	171.401 ns	0.0528 ns	0.53

BDN_Artifacts.zip

EgorBot · 2024-07-09T15:19:36Z

Benchmark results on Intel

BenchmarkDotNet v0.13.12, Ubuntu 22.04.4 LTS (Jammy Jellyfish)
Intel Xeon Platinum 8370C CPU 2.80GHz, 1 CPU, 8 logical and 4 physical cores
  Job-IRPCYL : .NET 9.0.0 (42.42.42.42424), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI
  Job-HIDRWE : .NET 9.0.0 (42.42.42.42424), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI

Method	Toolchain	Length	Mean	Error	Ratio
ValidateAscii	Main	1	4.712 ns	0.0203 ns	1.00
ValidateAscii	PR	1	3.828 ns	0.0180 ns	0.81

ValidateAscii	Main	2	4.492 ns	0.0068 ns	1.00
ValidateAscii	PR	2	3.904 ns	0.0065 ns	0.87

ValidateAscii	Main	3	4.923 ns	0.0087 ns	1.00
ValidateAscii	PR	3	4.172 ns	0.0233 ns	0.85

ValidateAscii	Main	4	4.313 ns	0.0294 ns	1.00
ValidateAscii	PR	4	3.945 ns	0.0329 ns	0.91

ValidateAscii	Main	5	4.739 ns	0.0089 ns	1.00
ValidateAscii	PR	5	4.049 ns	0.0112 ns	0.85

ValidateAscii	Main	6	4.494 ns	0.0056 ns	1.00
ValidateAscii	PR	6	4.263 ns	0.0196 ns	0.95

ValidateAscii	Main	7	5.355 ns	0.0211 ns	1.00
ValidateAscii	PR	7	4.102 ns	0.0139 ns	0.77

ValidateAscii	Main	8	4.714 ns	0.0135 ns	1.00
ValidateAscii	PR	8	3.894 ns	0.0118 ns	0.83

ValidateAscii	Main	12	4.848 ns	0.0116 ns	1.00
ValidateAscii	PR	12	3.875 ns	0.0107 ns	0.80

ValidateAscii	Main	16	5.186 ns	0.0170 ns	1.00
ValidateAscii	PR	16	3.889 ns	0.0133 ns	0.75

ValidateAscii	Main	25	5.818 ns	0.0089 ns	1.00
ValidateAscii	PR	25	5.105 ns	0.0159 ns	0.88

ValidateAscii	Main	50	6.527 ns	0.0339 ns	1.00
ValidateAscii	PR	50	5.260 ns	0.0022 ns	0.81

ValidateAscii	Main	70	6.859 ns	0.0139 ns	1.00
ValidateAscii	PR	70	6.169 ns	0.0032 ns	0.90

ValidateAscii	Main	200	9.229 ns	0.0114 ns	1.00
ValidateAscii	PR	200	6.453 ns	0.0147 ns	0.70

ValidateAscii	Main	512	9.452 ns	0.0124 ns	1.00
ValidateAscii	PR	512	10.454 ns	0.0289 ns	1.11

ValidateAscii	Main	16384	162.361 ns	0.0240 ns	1.00
ValidateAscii	PR	16384	163.184 ns	0.0489 ns	1.01

BDN_Artifacts.zip

src/libraries/System.Private.CoreLib/src/System/Text/Ascii.Utility.cs

neon-sunset · 2024-07-15T13:45:37Z

@EgorBot -amd -arm64

using System.Text.Unicode;
using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Running;

// Let's try again
BenchmarkRunner.Run<AsciiBench>(args: args);

public class AsciiBench
{
    [ParamsSource(nameof(Lengths))]
    public int Length;
    public IEnumerable<int> Lengths => [..Enumerable.Range(1, 8), 12, 16, 25, 50, 70, 200, 512, 16384];

    byte[] ascii = [];

    [GlobalSetup]
    public void Setup()
    {
        ascii = new byte[Length];
        ascii.AsSpan().Fill((byte)'f');
    }

    [Benchmark]
    public bool ValidateAscii()
    {
        return Utf8.IsValid(ascii);
    }
}

EgorBot · 2024-07-15T14:14:34Z

Benchmark results on Intel

BenchmarkDotNet v0.13.12, Ubuntu 22.04.4 LTS (Jammy Jellyfish)
Intel Xeon Platinum 8370C CPU 2.80GHz, 1 CPU, 8 logical and 4 physical cores
  Job-DGHARR : .NET 9.0.0 (42.42.42.42424), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI
  Job-CCKSVA : .NET 9.0.0 (42.42.42.42424), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI

Method	Toolchain	Length	Mean	Error	Ratio
ValidateAscii	Main	1	4.731 ns	0.0199 ns	1.00
ValidateAscii	PR	1	3.672 ns	0.0061 ns	0.78

ValidateAscii	Main	2	4.471 ns	0.0039 ns	1.00
ValidateAscii	PR	2	3.723 ns	0.0122 ns	0.83

ValidateAscii	Main	3	4.915 ns	0.0093 ns	1.00
ValidateAscii	PR	3	4.162 ns	0.0148 ns	0.85

ValidateAscii	Main	4	4.490 ns	0.0198 ns	1.00
ValidateAscii	PR	4	3.799 ns	0.0158 ns	0.85

ValidateAscii	Main	5	4.794 ns	0.0179 ns	1.00
ValidateAscii	PR	5	4.164 ns	0.0142 ns	0.87

ValidateAscii	Main	6	4.451 ns	0.0022 ns	1.00
ValidateAscii	PR	6	4.129 ns	0.0325 ns	0.93

ValidateAscii	Main	7	5.376 ns	0.0222 ns	1.00
ValidateAscii	PR	7	4.035 ns	0.0199 ns	0.75

ValidateAscii	Main	8	4.867 ns	0.0126 ns	1.00
ValidateAscii	PR	8	3.690 ns	0.0047 ns	0.76

ValidateAscii	Main	12	4.986 ns	0.0232 ns	1.00
ValidateAscii	PR	12	3.690 ns	0.0055 ns	0.74

ValidateAscii	Main	16	5.321 ns	0.0194 ns	1.00
ValidateAscii	PR	16	3.689 ns	0.0044 ns	0.69

ValidateAscii	Main	25	5.852 ns	0.0103 ns	1.00
ValidateAscii	PR	25	4.616 ns	0.0193 ns	0.79

ValidateAscii	Main	50	6.172 ns	0.0012 ns	1.00
ValidateAscii	PR	50	4.674 ns	0.0303 ns	0.76

ValidateAscii	Main	70	6.872 ns	0.0179 ns	1.00
ValidateAscii	PR	70	5.886 ns	0.0735 ns	0.86

ValidateAscii	Main	200	7.563 ns	0.0244 ns	1.00
ValidateAscii	PR	200	6.432 ns	0.0128 ns	0.85

ValidateAscii	Main	512	8.873 ns	0.1425 ns	1.00
ValidateAscii	PR	512	7.488 ns	0.0179 ns	0.84

ValidateAscii	Main	16384	162.573 ns	0.0717 ns	1.00
ValidateAscii	PR	16384	80.495 ns	0.0317 ns	0.50

BDN_Artifacts.zip

…D asserts (for now)

GrabYourPitchforks · 2024-08-07T06:12:47Z

FYI the original implementation of this method waaaaay back when did indeed use byrefs instead of raw pointers, but IIRC @jkotas had concerns with that approach and much preferred the simplicity of using true pointers. I'd recommend asking him for his thoughts on this PR.

jkotas · 2024-08-07T06:16:35Z

Yes, I still think that unmanaged pointers are less bug prone than byref arithmetics. We have been using byref arithmetics all over the place just because of we can - I have stopped commenting on it.

neon-sunset · 2024-08-07T21:55:20Z

My idea was to move the ASCII path over to byrefs first to allow moving the UTF-8 path to them later (once Egor's PR for adopting Lemire's UTF-8 validation is done, possibly adding ARM64 path if it doesn't happen in scope of it). Once done, this would have allowed to improve call codegen around Utf8.IsValid and GetIndexOfFirstInvalidUtf8Sequence that it forwards to for external callers - right now it is quite messy for something that is just Handle(span) -> nint and result < 0. Additionally, I hope it makes the path more GC friendly under heavy allocation traffic with reduced pinning.

However, if you think this PR should be changed back to pointer arithmetics - please let me know. It was an optional goal of the change, and I'm happy as long as the performance of this hot path improves. Thanks!

neon-sunset · 2024-09-09T14:16:06Z

CC @jeffhandley

As an area owner, PTAL when you have time. Thanks!

tarekgh · 2024-09-23T20:35:23Z

@tannergooding will you continue reviewing this change?

neon-sunset added 2 commits July 6, 2024 09:25

Rewrite GetIndexOfFirstNonAsciiByte

69a7473

- Deduplicate SIMD paths - there is no more need for that - Remove at least one branch from short length path - Use up to 128x4/256x2 SIMD unrolling

Improve lengths 32 and 64

f04b90c

dotnet-issue-labeler bot added the area-System.Text.Encoding label Jul 6, 2024

dotnet-policy-service bot added the community-contribution Indicates that the PR has been added by a community member label Jul 6, 2024

This comment was marked as outdated.

Sign in to view

build-analysis bot mentioned this pull request Jul 6, 2024

Nuget: Central Directory corrupt dotnet/dnceng#3099

Open

3 tasks

Reduce scalar path code size, use cheaper instructions on x86, align …

2033692

…512b loop, mitigate Dynamic PGO impact

This comment was marked as outdated.

Sign in to view

build-analysis bot mentioned this pull request Jul 7, 2024

[WASI][Mono-AOT] Invariant.Tests.dll test failure #102128

Closed

This comment was marked as outdated.

Sign in to view

dotnet deleted a comment from EgorBot Jul 9, 2024