Add Float16 to stdlib #30130

stephentyrone · 2020-02-28T22:19:52Z

Implement IEEE 754 binary16 as Float16 in the stdlib conforming to BinaryFloatingPoint and SIMDScalar.

Evolution proposal: https://github.com/apple/swift-evolution/blob/master/proposals/0277-float16.md
SE thread: https://forums.swift.org/t/se-0277-float16/33546

stephentyrone · 2020-02-28T22:20:02Z

@swift-ci please test

stephentyrone · 2020-02-28T23:00:58Z

@swift-ci please test

swift-ci · 2020-02-28T23:01:21Z

Build failed
Swift Test Linux Platform
Git Sha - a5ffc9f

swift-ci · 2020-02-28T23:05:04Z

Build failed
Swift Test OS X Platform
Git Sha - a5ffc9f

stephentyrone · 2020-02-28T23:38:07Z

@swift-ci please test

swift-ci · 2020-02-28T23:40:37Z

Build failed
Swift Test OS X Platform
Git Sha - 1d82e6c

swift-ci · 2020-02-28T23:40:39Z

Build failed
Swift Test Linux Platform
Git Sha - 1d82e6c

lorentey

Looks good! I caught one typo 🎣

lorentey · 2020-02-28T22:55:14Z

stdlib/public/core/FloatingPointTypes.swift.gyb

@@ -58,6 +67,7 @@ else:
 % end

 ${SelfDocComment}
+${Availability(bits)}


The extra line here won't interfere with tools extracting doc comments, right?

I don't think so.

stdlib/public/core/FloatingPointTypes.swift.gyb

stephentyrone · 2020-03-07T01:17:28Z

Update: ubuntu seems to be missing the float <-> float16 conversion functions in its libgcc, so we'll probably have to build them into the Swift runtime lib on Linux (and likely some other platforms, as well). I'll add a patch to do that sometime in the next few days.

stephentyrone · 2020-03-09T19:02:31Z

@swift-ci please smoke test Linux

stephentyrone · 2020-03-09T19:02:55Z

(Local build for Linux isn't working quite right, so rebuilding on CI to get a list of exactly the symbols required.)

stephentyrone · 2020-03-10T19:14:07Z

Looks like we need to provide:

__gnu_h2f_ieee (half -> float)
__gnu_f2h_ieee (float -> half)
__truncdfhf2 (double -> half)

stephentyrone · 2020-03-11T20:20:10Z

@swift-ci smoke test

stephentyrone · 2020-03-11T21:28:47Z

@tbkka can you review the new runtime/Float16Support.cpp?

tbkka

Minor suggestions.

stdlib/public/core/Codable.swift

tbkka · 2020-03-11T22:21:40Z

stdlib/public/runtime/Float16Support.cpp

@@ -0,0 +1,164 @@
+//===------------- Float16Support.cpp - Swift Float16 Support -------------===//


This seems like the right place for this.

stdlib/public/core/FloatingPointTypes.swift.gyb

stephentyrone · 2020-04-01T16:54:44Z

Getting back to this, let's run a test on Linux and see where we stand ...

stephentyrone · 2020-04-01T16:54:50Z

@swift-ci test

swift-ci · 2020-04-01T18:08:07Z

Build failed
Swift Test Linux Platform
Git Sha - 2d11922faede69623256062ea4f542d30c5eb140

This isn't quite right, because it introduces double-rounding, but it will do in the short-term.

xwu · 2020-04-02T02:19:02Z

@swift-ci test

swift-ci · 2020-04-02T02:21:35Z

Build failed
Swift Test OS X Platform
Git Sha - 2d11922faede69623256062ea4f542d30c5eb140

swift-ci · 2020-04-02T02:21:43Z

Build failed
Swift Test Linux Platform
Git Sha - 2d11922faede69623256062ea4f542d30c5eb140

Add an actual float16ToString operation, that we can give a real implemention for down the road.

stephentyrone · 2020-04-02T14:11:34Z

@swift-ci test

swift-ci · 2020-04-02T14:14:06Z

Build failed
Swift Test OS X Platform
Git Sha - 1aefc71

swift-ci · 2020-04-02T14:14:20Z

Build failed
Swift Test Linux Platform
Git Sha - 1aefc71

lib/SILOptimizer/Utils/ConstantFolding.cpp

stephentyrone · 2020-04-03T19:49:05Z

@swift-ci test

swift-ci · 2020-04-03T19:51:35Z

Build failed
Swift Test OS X Platform
Git Sha - 6705057

swift-ci · 2020-04-03T19:51:47Z

Build failed
Swift Test Linux Platform
Git Sha - 6705057

stephentyrone · 2020-04-06T19:21:29Z

@swift-ci please test

swift-ci · 2020-04-06T19:24:16Z

Build failed
Swift Test Linux Platform
Git Sha - 1188d20

swift-ci · 2020-04-06T19:24:42Z

Build failed
Swift Test OS X Platform
Git Sha - 1188d20

stephentyrone · 2020-04-06T19:31:34Z

@swift-ci please test

swift-ci · 2020-04-06T19:34:20Z

Build failed
Swift Test Linux Platform
Git Sha - 0030880bd91ae41c4f42f7baeba743319008d58a

swift-ci · 2020-04-06T19:34:52Z

Build failed
Swift Test OS X Platform
Git Sha - 0030880bd91ae41c4f42f7baeba743319008d58a

Implementation was merged to Swift in swiftlang/swift#30130.

compnerd · 2020-04-08T16:48:53Z

This caused a regression on the android builds :-(

Logs:
https://dev.azure.com/compnerd/3133d6ab-80a8-4996-ac4f-03df25cd3224/_apis/build/builds/29728/logs/289

The new Float16 support requires runtime support from the compiler runtime. In particular, the AEABI provides the runtime functions __aeabi_d2h to truncate a double to a half precision floating point which is missing in the compiler runtime. I think that the most expedient thing will be to provide an implementation in the standard library or in the shims :-(. Thoughts @stephentyrone?

stephentyrone · 2020-04-08T16:52:21Z

@compnerd Yes, you'll need to update stdlib/public/runtime/Float16Support.cpp to expose __truncdfhf2 under that name when building for Android. Host platforms' compiler-rt or libgcc are supposed to provide this operation, but they clearly don't on a bunch of non-Apple targets.

The implementation should work essentially unchanged, though you may need to replace the call to __gnu_f2h_ieee with either the appropriate float-to-half conversion or a cast if the compiler knows how to handle that conversion when targeting Android.

compnerd · 2020-04-08T18:05:46Z

Yeah, why would the system provide the (system) required symbols? sigh On the bright side, at least it is supposed to be there in future versions. I'll see if I can get the shims in place for ARMv7.

stephentyrone added the swift evolution approved Flag → feature: A feature that was approved through the Swift evolution process label Feb 28, 2020

stephentyrone mentioned this pull request Feb 28, 2020

[DNM] Implement Float16 #21738

Closed

stephentyrone requested a review from lorentey February 28, 2020 22:29

lorentey approved these changes Feb 29, 2020

View reviewed changes

stephentyrone requested a review from tbkka March 11, 2020 21:28

tbkka approved these changes Mar 11, 2020

View reviewed changes

stephentyrone force-pushed the SE-0277-Float16 branch from 7f8f62b to 2d11922 Compare March 13, 2020 15:16

stephentyrone added 7 commits April 1, 2020 16:47

Implement Float16

36b42c9

Add support for simd vectors of Float16, and Mirrors.

61b03df

Add conversions from Float16 to integer types.

fda59db

Quick implementation of FloatingPointParsing for Float16

af5446b

This isn't quite right, because it introduces double-rounding, but it will do in the short-term.

Add codable conformance for Float16, remove AnyHashableBox.

443cc9c

Add importer support for mapping __fp16 and _Float16 to Float16.

d3eba53

add _swift_stdlib_strtof16 shim

05f59ca

Incorporate suggestions from review.

1aefc71

stephentyrone force-pushed the SE-0277-Float16 branch from 2d11922 to 1aefc71 Compare April 1, 2020 20:48

Further review simplifications.

6705057

Add an actual float16ToString operation, that we can give a real implemention for down the road.

CodaFi reviewed Apr 2, 2020

View reviewed changes

lib/SILOptimizer/Utils/ConstantFolding.cpp Show resolved Hide resolved

Fix suggested by review.

1188d20

compiler-rt funcs need to take unsigned short for float16, not short.

232a78b

stephentyrone force-pushed the SE-0277-Float16 branch from 0030880 to 232a78b Compare April 6, 2020 19:31

stephentyrone merged commit 248c554 into swiftlang:master Apr 6, 2020

stephentyrone deleted the SE-0277-Float16 branch April 6, 2020 21:57

stephentyrone added a commit to swiftlang/swift-evolution that referenced this pull request Apr 7, 2020

Updated status of 0277. (#1139)

4e8e8e1

Implementation was merged to Swift in swiftlang/swift#30130.

owenv mentioned this pull request Apr 9, 2020

[157] Issue #157 - April 9, 2020 SwiftWeekly/swiftweekly.github.io#507

Closed

		@@ -0,0 +1,164 @@
		//===------------- Float16Support.cpp - Swift Float16 Support -------------===//

Add Float16 to stdlib #30130

Add Float16 to stdlib #30130

Uh oh!

Conversation

stephentyrone commented Feb 28, 2020

Uh oh!

stephentyrone commented Feb 28, 2020

Uh oh!

stephentyrone commented Feb 28, 2020

Uh oh!

swift-ci commented Feb 28, 2020

Uh oh!

swift-ci commented Feb 28, 2020

Uh oh!

stephentyrone commented Feb 28, 2020

Uh oh!

swift-ci commented Feb 28, 2020

Uh oh!

swift-ci commented Feb 28, 2020

Uh oh!

lorentey left a comment

Choose a reason for hiding this comment

Uh oh!

lorentey Feb 28, 2020

Choose a reason for hiding this comment

Uh oh!

stephentyrone Feb 29, 2020

Choose a reason for hiding this comment

Uh oh!

Uh oh!

stephentyrone commented Mar 7, 2020

Uh oh!

stephentyrone commented Mar 9, 2020

Uh oh!

stephentyrone commented Mar 9, 2020

Uh oh!

stephentyrone commented Mar 10, 2020

Uh oh!

stephentyrone commented Mar 11, 2020

Uh oh!

stephentyrone commented Mar 11, 2020

Uh oh!

tbkka left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

tbkka Mar 11, 2020

Choose a reason for hiding this comment

Uh oh!

Uh oh!

stephentyrone commented Apr 1, 2020

Uh oh!

stephentyrone commented Apr 1, 2020

Uh oh!

swift-ci commented Apr 1, 2020

Uh oh!

xwu commented Apr 2, 2020

Uh oh!

swift-ci commented Apr 2, 2020

Uh oh!

swift-ci commented Apr 2, 2020

Uh oh!

stephentyrone commented Apr 2, 2020

Uh oh!

swift-ci commented Apr 2, 2020

Uh oh!

swift-ci commented Apr 2, 2020

Uh oh!

Uh oh!

stephentyrone commented Apr 3, 2020

Uh oh!

swift-ci commented Apr 3, 2020

Uh oh!

swift-ci commented Apr 3, 2020

Uh oh!

stephentyrone commented Apr 6, 2020

Uh oh!

swift-ci commented Apr 6, 2020

Uh oh!

swift-ci commented Apr 6, 2020

stephentyrone commented Apr 8, 2020 •

edited

Loading

compnerd commented Apr 8, 2020 •

edited

Loading