Improve unpickling performance when reading from memory buffer

[adamsitnik]

When the Unpickler reads from a Stream it has non-trivial overhead compared to when reading from a memory buffer. This PR improves unpickling performance for that case.

The difference

• stream.ReadByte with an additional check (if the byte is not -1) is just array[position++].
• stream.Read(buffer, 0, size) is just array.Slice(0, size). This helps to avoid calling Array.Copy and copying the data more than needed.

Perf trick: the types that implement IInputReader are structs to enable for full interface method devirtualization and inlining. You can read about it here. I have added the necessary comment in the code that explains that.

The gains are quite big. From 20% to 2 times faster when unpickling from array. Unpickling from Stream has not been regressed. I added micro-benchmarks to measure and ensure this.

Method	Count	WAS	IS netcoreapp2.0	IS netcoreapp3.0
DoublesFromStream	100	7.151 us	7.174 us	5.594 us
IntegersFromStream	100	4.859 us	4.427 us	4.240 us
BooleansFromStream	100	2.183 us	1.917 us	1.824 us
StringsFromStream	100	14.648 us	12.766 us	10.439 us
DoublesFromArray	100	7.246 us	5.020 us	3.912 us
IntegersFromArray	100	4.802 us	3.897 us	3.822 us
BooleansFromArray	100	2.154 us	1.653 us	1.626 us
StringsFromArray	100	13.254 us	9.353 us	7.298 us
DoublesFromStream	1000	61.817 us	60.708 us	48.213 us
IntegersFromStream	1000	52.808 us	50.321 us	39.714 us
BooleansFromStream	1000	18.297 us	15.809 us	14.560 us
StringsFromStream	1000	151.480 us	142.836 us	112.647 us
DoublesFromArray	1000	60.844 us	41.348 us	31.345 us
IntegersFromArray	1000	52.648 us	40.204 us	31.525 us
BooleansFromArray	1000	18.346 us	13.245 us	13.369 us
StringsFromArray	1000	153.078 us	99.069 us	72.801 us

/cc @stephentoub, @imback82, @eerhardt, @rapoth

[irmen]

I have been on holiday so haven't been able to follow this PR until now. Will have a look at all the stuff that was talked about tomorrow, I think

[adamsitnik]

@irmen I have pushed all the code fixes, the code is ready for review.

[irmen]

I hope to be able to review all this in the following few days, but it's clear to me that you guys are a lot more knowledgeable about these low level optimizations than I am. So I don't think I'll be able to comment on the function, but rather only on the form (if any)

[irmen]

thanks a lot for these PRs and it's exiting that it makes Pyrolite a key library in .NET Spark

[adamsitnik]

@irmen thank you for the awesome library!

BTW please don't release a new NuGet package yet, tomorrow I plan to finish a PR that improves the Pickler performance.

Preview:

Method	Count	Mean WAS	Mean IS	WAS Alloc	IS Allocated
DoublesToByteArray	100	3.138 us	0.911 us	3184 B	304 B
StringsToByteArray	100	22.235 us	16.828 us	16424 B	10280 B
DoublesToByteArray	1000	31.298 us	7.269 us	42105 B	304 B
StringsToByteArray	1000	234.243 us	169.381 us	180100 B	102306 B

[irmen]

@adamsitnik Those are incredible results! I'm very curious what wizardry you're performing here

I'm just in the process right now of spicing up the readme.