Split JsonWriter into UTF-8 and UTF-16 specific types.

[ahsonkhan]

[WIP] Changes are on top of #2291, and hence need to be re-based first, before review.

This improves performance by 20-50% for some cases (small json payloads).

Focus only on the following commit, while reviewing: b220180

• Split the JsonWriter into JsonWriterUtf8 and JsonWriterUtf16 types.
• The JsonWriter no longer supports ITextBufferWriter (now supports IBufferWriter instead). Removed the SymbolTable and the "Slow" (non-invariant) encoding paths.
• Fixed a bug in prettyPrint bytesNeeded calculation for WriteObject/ArrayStart (to account for white space), which was ported back to Optimized the JsonWriter APIs and expanded the unit test. #2291.
• Added prettyPrint = true/false tests for both UTF-8 and UTF-16

BenchmarkDotNet=v0.10.14.534-nightly, OS=Windows 10.0.16299.431 (1709/FallCreatorsUpdate/Redstone3)
Intel Xeon CPU E5-1620 v2 3.70GHz, 1 CPU, 8 logical and 4 physical cores
Frequency=3604598 Hz, Resolution=277.4234 ns, Timer=TSC
.NET Core SDK=2.1.300-rtm-008838
  [Host]     : .NET Core 2.1.0-rtm-26514-02 (CoreCLR 4.6.26514.02, CoreFX 4.6.26514.02), 64bit RyuJIT
  DefaultJob : .NET Core 2.1.0-rtm-26514-02 (CoreCLR 4.6.26514.02, CoreFX 4.6.26514.02), 64bit RyuJIT

Method	Target	Formatted	Mean	Error	StdDev	Gen 0	Allocated
WriterSystemTextJsonBasic	InvariantUtf16	False	14,072.96 ns	259.4878 ns	254.8517 ns	-	0 B
WriterNewtonsoftBasic	InvariantUtf16	False	17,213.43 ns	337.5822 ns	401.8676 ns	0.0610	416 B
WriterSystemTextJsonHelloWorld	InvariantUtf16	False	97.62 ns	1.8345 ns	1.7160 ns	-	0 B
WriterNewtonsoftHelloWorld	InvariantUtf16	False	290.87 ns	1.5428 ns	1.3677 ns	0.0486	256 B
WriterSystemTextJsonBasic	InvariantUtf16	True	17,008.83 ns	119.0800 ns	78.7641 ns	-	0 B
WriterNewtonsoftBasic	InvariantUtf16	True	31,596.69 ns	87.0094 ns	81.3886 ns	0.0610	584 B
WriterSystemTextJsonHelloWorld	InvariantUtf16	True	103.20 ns	0.5058 ns	0.4223 ns	-	0 B
WriterNewtonsoftHelloWorld	InvariantUtf16	True	389.65 ns	5.6618 ns	4.7278 ns	0.0806	424 B
WriterSystemTextJsonBasic	InvariantUtf8	False	14,463.19 ns	31.6241 ns	24.6900 ns	-	0 B
WriterNewtonsoftBasic	InvariantUtf8	False	17,268.86 ns	75.3866 ns	58.8569 ns	0.0610	416 B
WriterSystemTextJsonHelloWorld	InvariantUtf8	False	161.86 ns	2.8533 ns	2.6690 ns	-	0 B
WriterNewtonsoftHelloWorld	InvariantUtf8	False	287.88 ns	0.9145 ns	0.6612 ns	0.0486	256 B
WriterSystemTextJsonBasic	InvariantUtf8	True	15,764.73 ns	71.1455 ns	63.0686 ns	-	0 B
WriterNewtonsoftBasic	InvariantUtf8	True	32,678.38 ns	204.6518 ns	147.9767 ns	0.0610	584 B
WriterSystemTextJsonHelloWorld	InvariantUtf8	True	166.55 ns	1.0146 ns	0.7336 ns	-	0 B
WriterNewtonsoftHelloWorld	InvariantUtf8	True	389.79 ns	4.3153 ns	3.8254 ns	0.0806	424 B

[ahsonkhan]

@adamsitnik, please review the changes in JsonWriterPerf.cs.

Is there a way for me to avoid adding this if check? It doesn't have much impact on the end result, though.

[adamsitnik]

The extra branch should not impact this particular benchmark (it's not a nano-benchmark).

To avoid this check you would have to remove EncoderTarget param and write a benchmark method per case

[adamsitnik]

the benchmark changes look good to me! 👍

[adamsitnik]

The extra branch should not impact this particular benchmark (it's not a nano-benchmark).

To avoid this check you would have to remove EncoderTarget param and write a benchmark method per case

[Tornhoof]

May I suggest looking into moving the pretty print part out of the writing methods?

• If someone starts pretty printing, performance is probably not the issue anymore
• Both SpanJson and Utf8Json have a "post-processor" pretty printer, which reads the normal output an reformat it prettily.
• It might be possible to have the same pretty print code for all supported encodings.

[ahsonkhan]

May I suggest looking into moving the pretty print part out of the writing methods?

How would that work for writing to a stream, etc.? We can't have a post-processor, if we are sending the bytes as we write them. Given the fact that the pretty print version has relatively low cost, it might not be worth it (looking at the delta between Formatted = false vs true). Still, I will file an issue because that is an interesting optimization we can investigate. If you are interested in giving it a shot, go for it :)

We need to expand the Writer Perf tests to include more variety of json writing scenarios (similar to the JsonReader, for instance heavy nesting, where pretty print might end up costing a lot more).

#2295