Improve Binary Xml (XmlDictionaryWriter) performance

[Daniel-Svensson]

Improves performance the binary XmlDictionaryWriter mainly by using Span and Unsafe

• Do larger writes using Unsafe instead of writing bytes at a time and avoid bounds checks for each byte written
• Use Span instead of pointers in a few places
  • For array writes the usage of Spans means that it is no longer necessary to do extra copies separate byte arrays
• Remove allocations when writing Guids
• Behaviour change: 32bit floats (and doubles) will no longer be encoded as 64bit (or 32bit) int. Instead it will only ecode it as an int if it saves space.

Other:

• I did not find the tests for the Binary XmlDictionaryWriter so instead I did some manual tests to ensure double and long round tripped as well as ran OpenRiaServices tests (which includes tests roundtriping classes with most property types with the changes on the server
• update 2022-07-25: added tests cases for binary xml
• I considered adding a tests to verify that (float)(1L << 63) is serialized as float instead of long, if you think it should be added then please advice where such a test should be written.
• I also did som experriments just using Span and Cast instead of Unsafe but the improvements was smaller compared to using Unsafe (ratio 0.78 for span and 0.54 for Unsafe for calling WriteInt64Text directly with memorystream backing).

Benchmarks

"Before" is the official net7 preview.5 bits (with r2r) while After is local release build of System.Xml and "System.Private.DataContract..* dll.

*Benchmarks Code:*

usingBenchmarkDotNet.Attributes;usingSystem;usingSystem.Diagnostics;usingSystem.IO;usingSystem.Threading;usingSystem.Threading.Tasks;usingSystem.Xml;namespaceConsoleApp1;[MemoryDiagnoser]publicclassBinaryXmlBenchmarks{privateXmlDictionaryWriter_writer;privateGuid_guid=Guid.NewGuid();constintArraySize=100;privateGuid[]_guids=Enumerable.Repeat(Guid.NewGuid(),ArraySize).ToArray();privateDouble[]_doubles=Enumerable.Repeat(4.0/3.0,ArraySize).ToArray();[GlobalSetup]publicvoidSetup(){_writer=XmlDictionaryWriter.CreateBinaryWriter(newNullWriteStream());_writer.WriteStartDocument();_writer.WriteStartElement("root");varasm=_writer.GetType().Assembly;varversion=FileVersionInfo.GetVersionInfo(asm.Location);Console.WriteLine($"Assembly: {asm.Location}");Console.WriteLine($"Version: {version.FileVersion}");}[Benchmark]publicvoidWriteInt32(){_writer.WriteValue(unchecked((int)0xdeadbeef));}[Benchmark]publicvoidWriteInt64(){_writer.WriteValue(long.MaxValue);}[Benchmark]publicvoidWriteDouble(){_writer.WriteValue(1.0/3.0);}[Benchmark]publicvoidWriteDecimal(){_writer.WriteValue(1.2m);}[Benchmark]publicvoidWriteGuid(){_writer.WriteValue(_guid);}[Benchmark]publicvoidWriteGuidArray(){_writer.WriteArray(null,"a",null,_guids,0,_guids.Length);}[Benchmark]publicvoidWriteDoubleArray(){_writer.WriteArray(null,"a",null,_doubles,0,_doubles.Length);}}internalclassNullWriteStream:Stream{publicoverrideboolCanRead=>false;publicoverrideboolCanSeek=>false;publicoverrideboolCanWrite=>true;publicoverridelongLength=>thrownewNotSupportedException();publicoverridelongPosition{get=>thrownewNotSupportedException();set=>thrownewNotSupportedException();}publicoverridevoidFlush(){}publicoverrideintRead(byte[]buffer,intoffset,intcount){thrownewNotSupportedException();}publicoverridelongSeek(longoffset,SeekOriginorigin){thrownewNotSupportedException();}publicoverridevoidSetLength(longvalue){thrownewNotSupportedException();}publicoverridevoidWrite(byte[]buffer,intoffset,intcount){}publicoverridevoidWrite(ReadOnlySpan<byte>buffer){}publicoverrideTaskWriteAsync(byte[]buffer,intoffset,intcount,CancellationTokencancellationToken){returnTask.CompletedTask;}publicoverridevoidWriteByte(bytevalue){}publicoverrideValueTaskWriteAsync(ReadOnlyMemory<byte>buffer,CancellationTokencancellationToken=default){returnValueTask.CompletedTask;}}

BenchmarkDotNet=v0.13.1, OS=Windows 10.0.22000
AMD Ryzen 5 5600U with Radeon Graphics, 1 CPU, 12 logical and 6 physical cores
.NET SDK=7.0.100-preview.5.22307.18
[Host] : .NET 7.0.0 (7.0.22.30112), X64 RyuJIT
After : .NET 7.0.0 (7.0.22.30112), X64 RyuJIT
7.0-preview5 : .NET 7.0.0 (7.0.22.30112), X64 RyuJIT
MaxRelativeError=0.01 IterationTime=250.0000 ms 

Method	Job	Mean	Error	StdDev	Median	Ratio	RatioSD	TotalIssues/Op	TotalCycles/Op	InstructionRetired/Op	Timer/Op	Gen 0	Allocated
WriteInt32	7.0-preview5	16.26 ns	0.186 ns	0.207 ns	16.21 ns	1.00	0.00	88	29	95	0	-	-
WriteInt32	Unsafe	13.55 ns	0.149 ns	0.204 ns	13.52 ns	0.84	0.02	69	24	74	0	-	-
WriteInt32	Span	14.06 ns	0.163 ns	0.326 ns	13.99 ns	0.85	0.03	79	25	74	0	-
WriteInt64	7.0-preview5	21.66 ns	0.237 ns	0.446 ns	21.57 ns	1.00	0.00	117	41	126	0	-	-
WriteInt64	Unsafe	12.45 ns	0.150 ns	0.125 ns	12.47 ns	0.58	0.01	62	25	66	0	-	-
WriteInt64	Span	13.05 ns	0.154 ns	0.183 ns	13.02 ns	0.60	0.01	74	24	69	0	-	-
WriteDouble	7.0-preview5	24.43 ns	0.261 ns	0.290 ns	24.35 ns	1.00	0.00	120	47	128	0	-	-
WriteDouble	Unsafe	14.15 ns	0.162 ns	0.233 ns	14.19 ns	0.58	0.01	68	26	72	0	-	-
WriteDouble	Span	17.00 ns	0.682 ns	1.958 ns	16.34 ns	0.64	0.04	79	30	74	0	-	-
WriteDecimal	7.0-preview5	43.56 ns	0.688 ns	1.918 ns	42.75 ns	1.00	0.00	249	77	270	0	-	-
WriteDecimal	Unsafe	20.44 ns	0.219 ns	0.234 ns	20.40 ns	0.44	0.01	91	40	96	0	-	-
WriteDecimal	Span	21.81 ns	0.173 ns	0.200 ns	21.82 ns	0.50	0.01	103	45	97	0	-	-
WriteGuid	7.0-preview5	45.66 ns	0.451 ns	0.646 ns	45.69 ns	1.00	0.00	209	89	224	0	0.0047	40 B
WriteGuid*	Unsafe*	19.75 ns	0.180 ns	0.159 ns	19.80 ns	0.43	0.01	82	39	86	0	-	-
WriteGuid*	Span*	16.90 ns	0.189 ns	0.266 ns	16.84 ns	0.38	0.01	88	31	83	0	-	-
WriteTimespan	7.0-preview5	22.05 ns	0.219 ns	0.195 ns	22.01 ns	1.00	0.00	119	40	128	0	-	-
WriteTimespan	Unsafe	12.59 ns	0.150 ns	0.307 ns	12.59 ns	0.56	0.01	61	23	65	0	-	-
WriteTimespan	Span	14.49 ns	0.167 ns	0.245 ns	14.49 ns	0.65	0.01	72	28	68	0	-	-
WriteGuidArray	7.0-preview5	3,417.86 ns	33.230 ns	34.125 ns	3,426.72 ns	1.00	0.00	16,152	6,654	17,066	34	0.4704	4,000 B
WriteGuidArray	Unsafe	595.89 ns	5.973 ns	7.767 ns	594.17 ns	0.18	0.00	3,153	1,116	3,414	6	-	-
WriteGuidArray	Span	614.04 ns	5.926 ns	6.824 ns	612.00 ns	0.18	0.00	3,417	1,167	3,167	6	-	-
WriteDoubleArray	7.0-preview5	1,040.62 ns	10.192 ns	10.467 ns	1,039.25 ns	1.00	0.00	5,954	1,896	6,504	10	-	-
WriteDoubleArray	Unsafe	130.87 ns	1.337 ns	1.830 ns	131.20 ns	0.13	0.00	666	249	715	1	-	-
WriteDoubleArray	Span	130.76 ns	1.242 ns	1.526 ns	130.69 ns	0.13	0.00	715	249	666	1	-	-

Remarks: * The span and unsafe version of WriteGuid and Write...Array are identical so I do not know why there is a timing differecnce.

[Daniel-Svensson]

I did try using ReadOnlySpan but then it looked like to many methods called this overload instead of the string version so I used span instead to avoid extra string allocations.

However it might be possibly to merge this implementation with the string based method if using ReadOnlyMemory instead since it allows trying to retrieve the original string.

[jkotas]

many methods called this overload instead of the string version

That does not sound right. If you have a string and there is a string overload, the compiler should pick the string overload.

[Daniel-Svensson]

I suspect might have been a VisualStudio bug, but I did not want to take any chances in case it had anything to do with one method being an override or similar.

[Daniel-Svensson]

changed the code, will have to decompile the code and have a second look. Visual studio (17.2) still reports many references to the ReadOnlySpan version

[Daniel-Svensson]

The unexpected calls are still there in the compiled assembly, should I keep it?
You might know who to involve to find out why it is called instead of the string overload.

[stephentoub]

Hmm, this surprises me:
SharpLab
For some reason the WriteText(string) overload being an override is causing it to lose out to the span-based overload.

[Daniel-Svensson]

I solved this by creating a separate "implementation metod" which is not an override (I called it WriteTextImpl to follow what seems to be the current naming convention for implementation methods)

[jkotas]

This code was using safe C# with proper boundschecks. You are switching this to use unsafe code without boundschecks. It is increasing security risk profile of this code.

Can this stay with safe C# and Spans? You should be still able to get nice performance gains by doing that.

[Daniel-Svensson]

Yes there was quite much unsafe code and points present in the file so I assumed it was ok to be unsafe for the best performance sine GetBuffer does the bounds check.

I've refactored the code so that all new "unsafe" code is now only present in 2 methods to make it much easier to reason about and updated the benchmark with timings for both Unsafe and Span.

The span version does .AsSpan(offset, size) on the buffer and then uses MemoryMarshal.Write<T> which I just foud out about.. I hope that method is considered safe enough and works well with "unaligned writes".

Please have a look and let me know based on the new code and the benchmark results which version you want.
I can push the updated span version instead if that version is the prefered one.

[Daniel-Svensson]

Note: A side effect of moving to the span version is that it will probably be necessary to keep separate implementationa for writing 8bit (and maybe 16bit) nodes in order to not regress performance for them.

Some benchmarks would be required to determine the performance numbers for smaller elements first

[stephentoub]

Is there sufficient test coverage to fully exercise all of these changes and root out any corner-case issues?

[Daniel-Svensson]

I think I fixed all the review comments for now so I think next up it to determine what to do with tests.

Is there sufficient test coverage to fully exercise all of these changes and root out any corner-case issues?

@stephentoub I am a bit unsure about that.

I did not find the tests for the Binary XmlDictionaryWriter so instead I did some manual tests to ensure double and long round tripped as well as ran OpenRiaServices tests (which includes tests roundtriping classes with most property types with the changes on the server

I can look into adding a new tests I found some tests for the text version of dictionary writer I could write some tests for the binary writer (write primitive types and some arrays ?).
I am thinking about serializing some data and then checking the binary representation against expected binary, or do you prefer round trip kind tests (where the same data is read by binary reader)?

update: I've added a couple of testcases so I think relevant scenarios are covered.

[Daniel-Svensson]

I've added a couple of tests that I think cover relevant cases.

I was a bit unsure how to handle any potential big endian platform since it is only handled for single integers so the array test as it is now will fail on big endian platforms. it can easily be made to pass but it will not make the binary xml the same for different endiannes (this pr does not make any changes to the representation apart from the fix for floats described in the description).

[Daniel-Svensson]

The test will give problem on big-endian architectures just as @uweigand found out #73332 (comment)

I can take a look and add a commit to this PR (maybe even this week) to try to fix the implementation for big endian architectures if you think it is a good idea? @jkotas@mconnew

If you prefer this PR as is without any behavioral changes is there a recommended way of skipping the "problematic" test on big endian architectures?

[HongGit]

@StephenMolloy and @mconnew can you please take a look?

[StephenMolloy]

Test failure appears to be unrelated. #64227

[StephenMolloy]

lgtm

[StephenMolloy]

lgtm