BigInteger parsing optimizations by jfd16 · Pull Request #47842 · dotnet/runtime

jfd16 · 2021-02-04T09:22:48Z

Improves performance of BigInteger parsing from decimal and hex strings (both speed and memory allocations), according to the benchmark that I have included. The difference is particularly significant for large decimal strings, given that the current implementation allocates two BigInteger instances for each input digit.

Benchmark code

usingSystem.Collections.Generic;usingSystem.Globalization;usingSystem.Linq;usingBenchmarkDotNet.Attributes;publicclassBigIntegerParseBenchmarks{publicIEnumerable<object>NumberStrings=>newstring[]{"123",int.MinValue.ToString(),string.Concat(Enumerable.Repeat("1234567890",20)),string.Concat(Enumerable.Repeat("654719003",57)),};publicIEnumerable<object>NumberStringsHex=>newstring[]{"12A",int.MinValue.ToString("X"),string.Concat(Enumerable.Repeat("1234567890abcdefffff",10)),string.Concat(Enumerable.Repeat("18a473f5d",57)),};[Benchmark][ArgumentsSource(nameof(NumberStrings))]publicBigIntegerParse(stringnumberString)=>BigInteger.Parse(numberString);[Benchmark][ArgumentsSource(nameof(NumberStringsHex))]publicBigIntegerParseHex(stringnumberString)=>BigInteger.Parse(numberString,NumberStyles.HexNumber);}

Benchmark results (on master branch)

BenchmarkDotNet=v0.12.1.1466-nightly, OS=Windows 10.0.18363.1316 (1909/November2019Update/19H2)
Intel Core i7-8550U CPU 1.80GHz (Kaby Lake R), 1 CPU, 8 logical and 4 physical cores
.NET SDK=5.0.102
[Host] : .NET 6.0.0 (6.0.21.10203), X64 RyuJIT
Job-NSGODO : .NET 6.0.0 (42.42.42.42424), X64 RyuJIT
PowerPlanMode=00000000-0000-0000-0000-000000000000 Arguments=/p:DebugType=portable Toolchain=CoreRun IterationTime=250.0000 ms MaxIterationCount=20 MinIterationCount=15 WarmupCount=1

Method	numberString	Mean	Error	StdDev	Median	Min	Max	Gen 0	Gen 1	Gen 2	Allocated
Parse	123	186.5 ns	3.43 ns	2.87 ns	187.0 ns	181.9 ns	192.6 ns	0.0246	-	-	104 B
Parse	-2147483648	418.0 ns	16.69 ns	18.55 ns	419.0 ns	359.6 ns	442.6 ns	0.0312	-	-	136 B
Parse	123456789012(...)901234567890 [200]	17,866.9 ns	193.43 ns	180.93 ns	17,901.3 ns	17,527.7 ns	18,163.3 ns	13.1653	-	-	55,176 B
Parse	654719003654(...)003654719003 [513]	69,284.7 ns	643.04 ns	536.97 ns	69,112.3 ns	68,447.0 ns	70,545.0 ns	65.5556	-	-	274,272 B
ParseHex	12A	121.8 ns	0.64 ns	0.50 ns	121.7 ns	121.3 ns	123.2 ns	0.0323	-	-	136 B
ParseHex	`8000000`	183.5 ns	1.28 ns	1.07 ns	183.6 ns	182.4 ns	185.5 ns	0.0319	-	-	136 B
ParseHex	1234567890ab(...)90abcdefffff [200]	2,520.9 ns	17.84 ns	14.90 ns	2,511.0 ns	2,505.7 ns	2,544.7 ns	0.2609	-	-	1,128 B
ParseHex	18a473f5d18a(...)f5d18a473f5d [513]	6,069.5 ns	83.14 ns	77.77 ns	6,045.6 ns	5,938.0 ns	6,202.7 ns	0.7287	-	-	3,128 B

Results after changes

BenchmarkDotNet=v0.12.1.1466-nightly, OS=Windows 10.0.18363.1316 (1909/November2019Update/19H2)
Intel Core i7-8550U CPU 1.80GHz (Kaby Lake R), 1 CPU, 8 logical and 4 physical cores
.NET SDK=5.0.102
[Host] : .NET 6.0.0 (6.0.21.10203), X64 RyuJIT
Job-WMUKDU : .NET 6.0.0 (42.42.42.42424), X64 RyuJIT
PowerPlanMode=00000000-0000-0000-0000-000000000000 Arguments=/p:DebugType=portable Toolchain=CoreRun IterationTime=250.0000 ms MaxIterationCount=20 MinIterationCount=15 WarmupCount=1

Method	numberString	Mean	Error	StdDev	Median	Min	Max	Gen 0	Gen 1	Gen 2	Allocated
Parse	123	126.0 ns	1.81 ns	1.51 ns	125.8 ns	123.7 ns	128.9 ns	0.0248	-	-	104 B
Parse	-2147483648	176.5 ns	3.55 ns	3.65 ns	174.7 ns	171.9 ns	182.6 ns	0.0322	-	-	136 B
Parse	123456789012(...)901234567890 [200]	1,843.5 ns	21.83 ns	20.42 ns	1,843.6 ns	1,819.9 ns	1,875.5 ns	0.2339	-	-	984 B
Parse	654719003654(...)003654719003 [513]	5,197.3 ns	23.14 ns	20.51 ns	5,206.4 ns	5,134.6 ns	5,210.8 ns	0.6656	-	-	2,792 B
ParseHex	12A	116.3 ns	1.23 ns	1.09 ns	116.2 ns	114.9 ns	118.8 ns	0.0245	-	-	104 B
ParseHex	`8000000`	166.1 ns	1.76 ns	1.56 ns	165.9 ns	164.2 ns	169.3 ns	0.0320	-	-	136 B
ParseHex	1234567890ab(...)90abcdefffff [200]	1,705.1 ns	10.51 ns	8.77 ns	1,703.0 ns	1,692.7 ns	1,715.7 ns	0.2362	-	-	1,000 B
ParseHex	18a473f5d18a(...)f5d18a473f5d [513]	3,905.5 ns	34.60 ns	30.67 ns	3,907.6 ns	3,866.8 ns	3,975.6 ns	0.6660	-	-	2,840 B

ghost · 2021-02-04T09:22:56Z

Tagging subscribers to this area: @tannergooding, @pgovind
See info in area-owners.md if you want to be subscribed.

Issue Details

Improves performance of BigInteger parsing from decimal and hex strings (both speed and memory allocations), according to the benchmark that I have included. The difference is particularly significant for large decimal strings, given that the current implementation allocates two BigInteger instances for each input digit.

Benchmark code

usingSystem.Collections.Generic;usingSystem.Globalization;usingSystem.Linq;usingBenchmarkDotNet.Attributes;publicclassBigIntegerParseBenchmarks{publicIEnumerable<object>NumberStrings=>newstring[]{"123",int.MinValue.ToString(),string.Concat(Enumerable.Repeat("1234567890",20)),string.Concat(Enumerable.Repeat("654719003",57)),};publicIEnumerable<object>NumberStringsHex=>newstring[]{"12A",int.MinValue.ToString("X"),string.Concat(Enumerable.Repeat("1234567890abcdefffff",10)),string.Concat(Enumerable.Repeat("18a473f5d",57)),};[Benchmark][ArgumentsSource(nameof(NumberStrings))]publicBigIntegerParse(stringnumberString)=>BigInteger.Parse(numberString);[Benchmark][ArgumentsSource(nameof(NumberStringsHex))]publicBigIntegerParseHex(stringnumberString)=>BigInteger.Parse(numberString,NumberStyles.HexNumber);}

Benchmark results (on master branch)

BenchmarkDotNet=v0.12.1.1466-nightly, OS=Windows 10.0.18363.1316 (1909/November2019Update/19H2)
Intel Core i7-8550U CPU 1.80GHz (Kaby Lake R), 1 CPU, 8 logical and 4 physical cores
.NET SDK=5.0.102
[Host] : .NET 6.0.0 (6.0.21.10203), X64 RyuJIT
Job-NSGODO : .NET 6.0.0 (42.42.42.42424), X64 RyuJIT
PowerPlanMode=00000000-0000-0000-0000-000000000000 Arguments=/p:DebugType=portable Toolchain=CoreRun IterationTime=250.0000 ms MaxIterationCount=20 MinIterationCount=15 WarmupCount=1

Method	numberString	Mean	Error	StdDev	Median	Min	Max	Gen 0	Gen 1	Gen 2	Allocated
Parse	123	186.5 ns	3.43 ns	2.87 ns	187.0 ns	181.9 ns	192.6 ns	0.0246	-	-	104 B
Parse	-2147483648	418.0 ns	16.69 ns	18.55 ns	419.0 ns	359.6 ns	442.6 ns	0.0312	-	-	136 B
Parse	123456789012(...)901234567890 [200]	17,866.9 ns	193.43 ns	180.93 ns	17,901.3 ns	17,527.7 ns	18,163.3 ns	13.1653	-	-	55,176 B
Parse	654719003654(...)003654719003 [513]	69,284.7 ns	643.04 ns	536.97 ns	69,112.3 ns	68,447.0 ns	70,545.0 ns	65.5556	-	-	274,272 B
ParseHex	12A	121.8 ns	0.64 ns	0.50 ns	121.7 ns	121.3 ns	123.2 ns	0.0323	-	-	136 B
ParseHex	`8000000`	183.5 ns	1.28 ns	1.07 ns	183.6 ns	182.4 ns	185.5 ns	0.0319	-	-	136 B
ParseHex	1234567890ab(...)90abcdefffff [200]	2,520.9 ns	17.84 ns	14.90 ns	2,511.0 ns	2,505.7 ns	2,544.7 ns	0.2609	-	-	1,128 B
ParseHex	18a473f5d18a(...)f5d18a473f5d [513]	6,069.5 ns	83.14 ns	77.77 ns	6,045.6 ns	5,938.0 ns	6,202.7 ns	0.7287	-	-	3,128 B

Results after changes

BenchmarkDotNet=v0.12.1.1466-nightly, OS=Windows 10.0.18363.1316 (1909/November2019Update/19H2)
Intel Core i7-8550U CPU 1.80GHz (Kaby Lake R), 1 CPU, 8 logical and 4 physical cores
.NET SDK=5.0.102
[Host] : .NET 6.0.0 (6.0.21.10203), X64 RyuJIT
Job-WMUKDU : .NET 6.0.0 (42.42.42.42424), X64 RyuJIT
PowerPlanMode=00000000-0000-0000-0000-000000000000 Arguments=/p:DebugType=portable Toolchain=CoreRun IterationTime=250.0000 ms MaxIterationCount=20 MinIterationCount=15 WarmupCount=1

Method	numberString	Mean	Error	StdDev	Median	Min	Max	Gen 0	Gen 1	Gen 2	Allocated
Parse	123	126.0 ns	1.81 ns	1.51 ns	125.8 ns	123.7 ns	128.9 ns	0.0248	-	-	104 B
Parse	-2147483648	176.5 ns	3.55 ns	3.65 ns	174.7 ns	171.9 ns	182.6 ns	0.0322	-	-	136 B
Parse	123456789012(...)901234567890 [200]	1,843.5 ns	21.83 ns	20.42 ns	1,843.6 ns	1,819.9 ns	1,875.5 ns	0.2339	-	-	984 B
Parse	654719003654(...)003654719003 [513]	5,197.3 ns	23.14 ns	20.51 ns	5,206.4 ns	5,134.6 ns	5,210.8 ns	0.6656	-	-	2,792 B
ParseHex	12A	116.3 ns	1.23 ns	1.09 ns	116.2 ns	114.9 ns	118.8 ns	0.0245	-	-	104 B
ParseHex	`8000000`	166.1 ns	1.76 ns	1.56 ns	165.9 ns	164.2 ns	169.3 ns	0.0320	-	-	136 B
ParseHex	1234567890ab(...)90abcdefffff [200]	1,705.1 ns	10.51 ns	8.77 ns	1,703.0 ns	1,692.7 ns	1,715.7 ns	0.2362	-	-	1,000 B
ParseHex	18a473f5d18a(...)f5d18a473f5d [513]	3,905.5 ns	34.60 ns	30.67 ns	3,907.6 ns	3,866.8 ns	3,975.6 ns	0.6660	-	-	2,840 B

Author:	jfd16
Assignees:	-
Labels:	`area-System.Numerics`
Milestone:	-

dnfadmin · 2021-02-04T09:23:05Z

All CLA requirements met.

danmoseley · 2021-02-05T20:39:29Z

The use of ArrayPool is much more common than ArrayPool

Grepped under libraries/*/src/**cs for interest and I see

object - 2
IntPtr, GCHandle - 4 each
char - 118
byte - 301

there are no uses I see for int or uint.

tannergooding · 2021-02-05T20:43:57Z

Grepped under libraries/*/src/**cs for interest and I see

I wonder if there is something we can do here to promote better sharing. Particularly when using Span the underlying data type doesn't really matter too much (provided it has the correct alignment)

jfd16 · 2021-02-05T20:52:43Z

The use of ArrayPool is much more common than ArrayPool
Grepped under libraries/*/src/**cs for interest and I see
object - 2
IntPtr, GCHandle - 4 each
char - 118
byte - 301
there are no uses I see for int or uint.

ArrayPool<int> is being used indirectly in some places through ValueListBuilder: https://github.com/dotnet/runtime/search?q=ValueListBuilder+path%3Asrc

tannergooding · 2021-02-09T16:01:08Z

+
+ try
 {
- char c = number.digits[i];


@stephentoub, is this an opportunity for an analyzer? Indexing into StringBuilder can be deceptively slow due to its internal chunking.

tannergooding · 2021-02-09T16:04:42Z

- isNegative = true;
+ foreach (ReadOnlyMemory<char> digitsChunkMem in number.digits.GetChunks())
+ {
+ ReadOnlySpan<char> chunkDigits = digitsChunkMem.Span;


Likewise, I wonder if we should have a (internal for now) CurrentSpan property that is slightly more efficient.
We don't really need to return a ReadOnlyMemory when we only want to extract the underlying Span<T>, particularly when the ROM<T> is just being constructed over a backing array.
It's, in practice, no more unsafe than say CollectionsMarshal.AsSpan

tannergooding

LGTM. I think it can be merged provided @stephentoub has no additional feedback

danmoseley · 2021-02-11T16:21:33Z

I wonder if there is something we can do here to promote better sharing. Particularly when using Span the underlying data type doesn't really matter too much (provided it has the correct alignment

Worth an issue do you think @tannergooding

jeffhandley · 2021-05-09T04:33:42Z

/azp run runtime

azure-pipelines · 2021-05-09T04:34:07Z

Azure Pipelines successfully started running 1 pipeline(s).

Optimize BigInteger parsing
7ea7474

ghost added the area-System.Numerics label Feb 4, 2021

Use ArrayPool<int> instead of ArrayPool<uint>.
5657cfc

tannergooding reviewed Feb 5, 2021
View reviewed changes

Comment threadsrc/libraries/System.Runtime.Numerics/src/System/Numerics/BigNumber.cs

tannergooding reviewed Feb 5, 2021
View reviewed changes

Comment threadsrc/libraries/System.Runtime.Numerics/src/System/Numerics/BigNumber.cs Outdated

stephentoub reviewed Feb 8, 2021
View reviewed changes

jfd16 force-pushed the bigint-parse-perf branch 3 times, most recently from 3041198 to 2383bd3Compare February 9, 2021 01:42

Additional changes
7b5eb62

jfd16 force-pushed the bigint-parse-perf branch from 2383bd3 to 7b5eb62Compare February 9, 2021 01:44