[stdlib] fast path for UMBP.initialize<C>(from: C) when C is a Slice #38677

glessard · 2021-07-28T21:09:52Z

UnsafeMutableBufferPointer.initialize(from:) initializes a buffer from a Sequence. Many standard library collections have a fast path that implements the multi-element copy in a single call, but that fast path is currently stymied when the collection is wrapped in a Slice. This PR implements Slice._copyContents, which attempts a fast copy via _withContiguousStorageIfAvailable before reverting to an element-by-element copy.

Resolves SR-14491 (rdar://76728166)

`Sequence._copyContents(initializing:)` is the function relied on by `UnsafeMutableBufferPointer` for performant initialization from Collections. Until now, `Slice<Base: Collection>` has not had its own implementation, and therefore fell back to the default version implemented on `Sequence`. This implementation adds an attempted fast path, using `withContiguousStorageIfAvailable`. If that fails, the `Sequence` algorithm is used, as before. This resolves https://bugs.swift.org/browse/SR-14491

glessard · 2021-07-28T21:24:00Z

@swift-ci please benchmark

swift-ci · 2021-07-28T22:08:09Z

Performance (x86_64): -O

Regression	OLD	NEW	DELTA	RATIO
DictionaryOfAnyHashableStrings_insert	2800	5096	+82.0%	0.55x
Set.isDisjoint.Box25	323	454	+40.6%	0.71x (?)
Breadcrumbs.MutatedUTF16ToIdx.ASCII	3	4	+33.3%	0.75x (?)
ObjectiveCBridgeStubFromNSDateRef	3750	4540	+21.1%	0.83x (?)
DictionaryKeysContainsNative	19	23	+21.1%	0.83x (?)
String.data.Medium	88	104	+18.2%	0.85x (?)
ObjectiveCBridgeStubFromNSDate	5550	6330	+14.1%	0.88x (?)
String.data.LargeUnicode	94	107	+13.8%	0.88x (?)

Improvement	OLD	NEW	DELTA	RATIO
BufferFillFromSlice	77	41	-46.8%	1.88x
FlattenListLoop	2200	1453	-34.0%	1.51x (?)
Data.init.Sequence.64kB.Count.I	52	35	-32.7%	1.49x
Data.append.Sequence.64kB.Count.I	52	35	-32.7%	1.49x
Data.append.Sequence.64kB.Count	52	35	-32.7%	1.49x
Data.init.Sequence.64kB.Count	52	35	-32.7%	1.49x
Data.init.Sequence.2047B.Count.I	89	63	-29.2%	1.41x
Data.init.Sequence.2049B.Count.I	89	63	-29.2%	1.41x
Data.init.Sequence.809B.Count	81	61	-24.7%	1.33x
Data.init.Sequence.809B.Count.I	81	61	-24.7%	1.33x
FlattenListFlatMap	5673	4411	-22.2%	1.29x (?)
Data.init.Sequence.511B.Count.I	93	74	-20.4%	1.26x (?)
Data.append.Sequence.809B.Count.I	98	78	-20.4%	1.26x (?)
Data.append.Sequence.809B.Count	97	78	-19.6%	1.24x (?)
Data.init.Sequence.513B.Count.I	93	78	-16.1%	1.19x (?)
RemoveWhereMoveInts	14	12	-14.3%	1.17x
LessSubstringSubstring	38	35	-7.9%	1.09x
EqualSubstringSubstring	38	35	-7.9%	1.09x (?)
EqualStringSubstring	38	35	-7.9%	1.09x
EqualSubstringSubstringGenericEquatable	38	35	-7.9%	1.09x (?)
EqualSubstringString	38	35	-7.9%	1.09x
LessSubstringSubstringGenericComparable	38	35	-7.9%	1.09x
Set.subtracting.Empty.Box	14	13	-7.1%	1.08x (?)
SortStringsUnicode	2805	2610	-7.0%	1.07x
Array2D	6480	6048	-6.7%	1.07x (?)

Code size: -O

Regression	OLD	NEW	DELTA	RATIO
RemoveWhere.o	16155	16571	+2.6%	0.97x
Queue.o	12119	12263	+1.2%	0.99x

Improvement	OLD	NEW	DELTA	RATIO
BufferFill.o	7940	7764	-2.2%	1.02x

Performance (x86_64): -Osize

Regression	OLD	NEW	DELTA	RATIO
Breadcrumbs.MutatedUTF16ToIdx.ASCII	3	4	+33.3%	0.75x (?)
ObjectiveCBridgeStubFromNSDate	5660	6400	+13.1%	0.88x (?)
ObjectiveCBridgeFromNSArrayAnyObjectForced	4160	4480	+7.7%	0.93x (?)
MapReduceShortString	13	14	+7.7%	0.93x (?)

Improvement	OLD	NEW	DELTA	RATIO
BufferFillFromSlice	77	41	-46.8%	1.88x
FlattenListFlatMap	5250	3677	-30.0%	1.43x (?)
Hanoi	3910	3550	-9.2%	1.10x
LessSubstringSubstring	38	35	-7.9%	1.09x
EqualSubstringSubstring	38	35	-7.9%	1.09x
EqualStringSubstring	38	35	-7.9%	1.09x (?)
EqualSubstringString	38	35	-7.9%	1.09x
LessSubstringSubstringGenericComparable	38	35	-7.9%	1.09x
ParseFloat.Double.Exp	13	12	-7.7%	1.08x (?)
Array2D	6736	6224	-7.6%	1.08x (?)
SortStringsUnicode	2810	2615	-6.9%	1.07x (?)

Code size: -Osize

Regression	OLD	NEW	DELTA	RATIO
BufferFill.o	7135	8118	+13.8%	0.88x
UTF8Decode.o	21531	21770	+1.1%	0.99x

Performance (x86_64): -Onone

Regression	OLD	NEW	DELTA	RATIO
ObjectiveCBridgeFromNSSetAnyObjectToStringForced	80000	87000	+8.7%	0.92x (?)
StringToDataMedium	5800	6250	+7.8%	0.93x (?)

Improvement	OLD	NEW	DELTA	RATIO
BufferFillFromSlice	51061	42	-99.9%	1215.71x
DictionaryBridgeToObjC_Access	1066	915	-14.2%	1.17x (?)

Code size: -swiftlibs

How to read the data

The tables contain differences in performance which are larger than 8% and differences in code size which are larger than 1%.

If you see any unexpected regressions, you should consider fixing the
regressions before you merge the PR.

Noise: Sometimes the performance results (not code size!) contain false
alarms. Unexpected regressions which are marked with '(?)' are probably noise.
If you see regressions which you cannot explain you can try to run the
benchmarks again. If regressions still show up, please consult with the
performance team (@eeckstein).

Hardware Overview

  Model Name: Mac Pro
  Model Identifier: MacPro6,1
  Processor Name: 8-Core Intel Xeon E5
  Processor Speed: 3 GHz
  Number of Processors: 1
  Total Number of Cores: 8
  L2 Cache (per Core): 256 KB
  L3 Cache: 25 MB
  Memory: 64 GB

glessard · 2021-07-28T23:12:58Z

@swift-ci Apple Silicon benchmark

glessard · 2021-07-28T23:13:13Z

@swift-ci please test

swift-ci · 2021-07-29T00:59:14Z

Performance (arm64): -O

Regression	OLD	NEW	DELTA	RATIO
DictionaryOfAnyHashableStrings_insert	1330	2254	+69.5%	0.59x (?)
Set.isDisjoint.Box25	114	193	+69.3%	0.59x (?)
Set.isDisjoint.Int25	86	130	+51.2%	0.66x (?)
Set.isDisjoint.Int50	86	127	+47.7%	0.68x (?)
StringRemoveDupes	114	142	+24.6%	0.80x (?)
CharIteration_punctuatedJapanese_unicodeScalars	240	280	+16.7%	0.86x (?)
SetIsSubsetBox25	84	95	+13.1%	0.88x (?)
ArrayAppend	390	440	+12.8%	0.89x (?)
ConvertFloatingPoint.MockFloat64ToDouble	9	10	+11.1%	0.90x (?)
DictionaryRemove	1790	1930	+7.8%	0.93x (?)

Improvement	OLD	NEW	DELTA	RATIO
BufferFillFromSlice	46	11	-76.1%	4.18x (?)
LessSubstringSubstring	21	18	-14.3%	1.17x (?)
EqualSubstringSubstring	21	18	-14.3%	1.17x (?)
EqualStringSubstring	21	18	-14.3%	1.17x (?)
EqualSubstringSubstringGenericEquatable	21	18	-14.3%	1.17x (?)
EqualSubstringString	21	18	-14.3%	1.17x (?)
LessSubstringSubstringGenericComparable	21	18	-14.3%	1.17x (?)
Set.filter.Int100.24k	855	784	-8.3%	1.09x (?)
Dict.CopyKeyValue.24k	863	795	-7.9%	1.09x (?)
Set.filter.Int100.20k	731	678	-7.3%	1.08x (?)

Code size: -O

Regression	OLD	NEW	DELTA	RATIO
RemoveWhere.o	13988	14348	+2.6%	0.97x
Queue.o	11757	11877	+1.0%	0.99x

Improvement	OLD	NEW	DELTA	RATIO
BufferFill.o	7314	7142	-2.4%	1.02x

Performance (arm64): -Osize

Improvement	OLD	NEW	DELTA	RATIO
BufferFillFromSlice	46	11	-76.1%	4.18x
ArrayAppendOptionals	930	410	-55.9%	2.27x (?)
EqualStringSubstring	21	18	-14.3%	1.17x (?)
EqualSubstringString	21	18	-14.3%	1.17x (?)
LessSubstringSubstring	22	19	-13.6%	1.16x (?)
EqualSubstringSubstring	22	19	-13.6%	1.16x (?)
EqualSubstringSubstringGenericEquatable	22	19	-13.6%	1.16x (?)
LessSubstringSubstringGenericComparable	22	19	-13.6%	1.16x (?)

Code size: -Osize

Regression	OLD	NEW	DELTA	RATIO
BufferFill.o	7646	8818	+15.3%	0.87x

Performance (arm64): -Onone

Regression	OLD	NEW	DELTA	RATIO
ParseFloat.Float.Exp	7	8	+14.3%	0.88x (?)
ArrayAppendOptionals	380	420	+10.5%	0.90x (?)
RecursiveOwnedParameter	3097	3421	+10.5%	0.91x (?)
DistinctClassFieldAccesses	580	624	+7.6%	0.93x (?)

Improvement	OLD	NEW	DELTA	RATIO
BufferFillFromSlice	20494	11	-99.9%	1862.92x
IterateData	1613	1417	-12.2%	1.14x
DataAppendDataLargeToLarge	13800	12600	-8.7%	1.10x (?)
LessSubstringSubstring	39	36	-7.7%	1.08x (?)
EqualSubstringSubstring	39	36	-7.7%	1.08x (?)
EqualStringSubstring	39	36	-7.7%	1.08x (?)
EqualSubstringSubstringGenericEquatable	39	36	-7.7%	1.08x (?)
EqualSubstringString	39	36	-7.7%	1.08x (?)
LessSubstringSubstringGenericComparable	39	36	-7.7%	1.08x (?)

Code size: -swiftlibs

How to read the data

The tables contain differences in performance which are larger than 8% and differences in code size which are larger than 1%.

If you see any unexpected regressions, you should consider fixing the
regressions before you merge the PR.

Noise: Sometimes the performance results (not code size!) contain false
alarms. Unexpected regressions which are marked with '(?)' are probably noise.
If you see regressions which you cannot explain you can try to run the
benchmarks again. If regressions still show up, please consult with the
performance team (@eeckstein).

Hardware Overview

  Model Name: Mac mini
  Model Identifier: Macmini9,1
  Total Number of Cores: 8 (4 performance and 4 efficiency)
  Memory: 16 GB

lorentey · 2021-07-29T01:02:42Z

Nice! Cc @karwa -- he has been interested in getting these fixed.

stdlib/public/core/Sequence.swift

stdlib/public/core/Slice.swift

swift-ci · 2021-07-29T01:57:20Z

Build failed
Swift Test OS X Platform
Git Sha - 2bc053f

glessard · 2021-08-02T18:21:28Z

@swift-ci please smoke test

glessard · 2021-08-02T22:13:26Z

@swift-ci please test macOS platform

karwa · 2021-08-09T15:08:58Z

Thanks a lot for taking this on, @glessard!

[test] test UnsafeMutableBufferPointer.initialize(from: Slice<_>)

23bed31

glessard requested a review from lorentey July 28, 2021 21:09

glessard added 2 commits July 28, 2021 15:23

[stdlib] make Sequence's _copyContents call into a non-protocol function

a9e0af4

glessard force-pushed the sr14491-v1 branch from 463b650 to 2bc053f Compare July 28, 2021 21:23

lorentey reviewed Jul 29, 2021

View reviewed changes

stdlib/public/core/Sequence.swift Outdated Show resolved Hide resolved

stdlib/public/core/Slice.swift Outdated Show resolved Hide resolved

stdlib/public/core/Slice.swift Outdated Show resolved Hide resolved

lorentey approved these changes Jul 29, 2021

View reviewed changes

glessard added 2 commits August 2, 2021 12:20

[stdlib] simplify indexing

f6f36fa

[stdlib] avoid ABI impact

3302750

glessard merged commit 92335b1 into apple:main Aug 3, 2021

glessard deleted the sr14491-v1 branch August 3, 2021 17:28

glessard mentioned this pull request Jul 28, 2021

[SR-14491] UnsafeMutableBufferPointer.initialize(from:) is slow with slices #56846

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[stdlib] fast path for UMBP.initialize<C>(from: C) when C is a Slice #38677

[stdlib] fast path for UMBP.initialize<C>(from: C) when C is a Slice #38677

glessard commented Jul 28, 2021 •

edited

glessard commented Jul 28, 2021

swift-ci commented Jul 28, 2021

glessard commented Jul 28, 2021

glessard commented Jul 28, 2021

swift-ci commented Jul 29, 2021

lorentey commented Jul 29, 2021

swift-ci commented Jul 29, 2021

glessard commented Aug 2, 2021

glessard commented Aug 2, 2021

karwa commented Aug 9, 2021

[stdlib] fast path for UMBP.initialize<C>(from: C) when C is a Slice #38677

[stdlib] fast path for UMBP.initialize<C>(from: C) when C is a Slice #38677

Conversation

glessard commented Jul 28, 2021 • edited

glessard commented Jul 28, 2021

swift-ci commented Jul 28, 2021

Performance (x86_64): -O

Code size: -O

Performance (x86_64): -Osize

Code size: -Osize

Performance (x86_64): -Onone

Code size: -swiftlibs

glessard commented Jul 28, 2021

glessard commented Jul 28, 2021

swift-ci commented Jul 29, 2021

Performance (arm64): -O

Code size: -O

Performance (arm64): -Osize

Code size: -Osize

Performance (arm64): -Onone

Code size: -swiftlibs

lorentey commented Jul 29, 2021

swift-ci commented Jul 29, 2021

glessard commented Aug 2, 2021

glessard commented Aug 2, 2021

karwa commented Aug 9, 2021

glessard commented Jul 28, 2021 •

edited