Fix BFloat16.Significand to be byte-sized and correct TryWriteSignificand methods #123288

Copilot · 2026-01-16T20:57:02Z

Description

BFloat16.TryWriteSignificandBigEndian and TryWriteSignificandLittleEndian initially reported bytesWritten = 4 (sizeof(uint)) instead of the correct value. Upon further investigation, it was discovered that BFloat16's significand should be byte-sized (8 bits: 7 trailing bits + 1 implicit bit), not ushort-sized (16 bits).

Changes

BFloat16.cs:
- Changed Significand property from ushort to byte to correctly represent the 8-bit significand
- Updated GetSignificandByteCount() to return sizeof(byte) (1) instead of sizeof(ushort) (2)
- Updated TryWriteSignificandBigEndian to write 1 byte directly (endianness is irrelevant for single byte)
- Updated TryWriteSignificandLittleEndian to write 1 byte directly (endianness is irrelevant for single byte)
BFloat16Tests.cs: Added test coverage for TryWriteSignificandBigEndian and TryWriteSignificandLittleEndian using [Theory] with [MemberData] to provide BFloat16 test values directly. Tests verify correct byte count (1 byte) and output values for various BFloat16 values (NegativeInfinity, MinValue, NaN, Epsilon, Zero, MaxValue, PositiveInfinity). Empty destination edge cases are tested in separate [Fact] methods. Test data methods use modern C# collection literal syntax [ ... ] for cleaner, more maintainable code.

Technical Details: BFloat16 format has 1 sign bit + 8 exponent bits + 7 trailing significand bits. The Significand property includes the implicit leading 1 bit for normal numbers, resulting in 8 bits total (1 byte). For a single byte value, endianness doesn't apply, so both BigEndian and LittleEndian methods write the byte directly to destination[0].

Note: Tests use [MemberData] instead of [InlineData] to pass BFloat16 objects directly, avoiding float-to-BFloat16 conversion issues that could produce different values than expected.

Original prompt

This section details on the original issue you should resolve

<issue_title>BFloat16.TryWriteSignificand{Big,Little}Endian report 4 for bytesWritten</issue_title>
<issue_description>The BFloat16 implementation of TryWriteSignificandBigEndian and TryWriteSignificandLittleEndian write 16-bit integers, but report 4 as the number of bytes written (32-bit), even though only two bytes were written:

runtime/src/libraries/System.Private.CoreLib/src/System/Numerics/BFloat16.cs

Lines 1047 to 1051 in 0bb2889

bool IFloatingPoint<BFloat16>.TryWriteSignificandBigEndian(Span<byte> destination, out int bytesWritten)

{

if (BinaryPrimitives.TryWriteUInt16BigEndian(destination, Significand))

{

bytesWritten = sizeof(uint);

I'm not exactly a numerics person but should this be a ushort? (That is what Half does).</issue_description>

<agent_instructions>Please fix bytesWritten for BFloat16, add tests</agent_instructions>

Comments on the Issue (you are @copilot in this section)

Fixes BFloat16.TryWriteSignificand{Big,Little}Endian report 4 for bytesWritten #123286

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.

- Changed bytesWritten from sizeof(uint) to sizeof(ushort) in TryWriteSignificandBigEndian - Changed bytesWritten from sizeof(uint) to sizeof(ushort) in TryWriteSignificandLittleEndian - Added comprehensive tests for both methods Co-authored-by: EgorBo <523221+EgorBo@users.noreply.github.com>

Copilot

Pull request overview

This PR fixes a bug in BFloat16.TryWriteSignificandBigEndian and TryWriteSignificandLittleEndian where these methods incorrectly reported writing 4 bytes (sizeof(uint)) instead of 2 bytes (sizeof(ushort)) for the significand.

Changes:

Fixed bytesWritten to report sizeof(ushort) instead of sizeof(uint) in both methods
Added comprehensive test coverage for both methods to verify correct byte counts and output values

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
BFloat16.cs	Corrected bytesWritten from sizeof(uint) to sizeof(ushort) in TryWriteSignificandBigEndian and TryWriteSignificandLittleEndian
BFloat16Tests.cs	Added comprehensive test methods TryWriteSignificandBigEndianTest and TryWriteSignificandLittleEndianTest with 11 test cases each

src/libraries/System.Runtime/tests/System.Runtime.Tests/System/Numerics/BFloat16Tests.cs

- Converted TryWriteSignificandBigEndianTest to [Theory] with [InlineData] attributes - Converted TryWriteSignificandLittleEndianTest to [Theory] with [InlineData] attributes - Separated empty destination tests into dedicated [Fact] methods - Removed duplicate test cases (same significand for ±0.0 and ±Epsilon) Co-authored-by: stephentoub <2642209+stephentoub@users.noreply.github.com>

EgorBo · 2026-01-17T00:39:04Z

@copilot the tests you added fail on CI. e.g.

Assert.Equal() Failure: Collections differ
↓ (pos 0)
Expected: [255, 0]
Actual:   [128, 0]
↑ (pos 0)

- Changed from [InlineData] with float values to [MemberData] with BFloat16 values - Use BFloat16.MinValue/MaxValue instead of float.MinValue/MaxValue conversions - This ensures test data matches the actual BFloat16 representation - Fixes CI test failures where float-to-BFloat16 conversions produced different values Co-authored-by: EgorBo <523221+EgorBo@users.noreply.github.com>

Copilot · 2026-01-17T00:45:47Z

@copilot the tests you added fail on CI. e.g.
Assert.Equal() Failure: Collections differ

...

Fixed in commit b44d1de. The issue was using float literals that converted to different BFloat16 values than expected. Changed to use BFloat16 constants directly (e.g., BFloat16.MinValue instead of (BFloat16)(-3.40282347E+38f)) and switched from [InlineData] to [MemberData] to support BFloat16 parameters.

EgorBo · 2026-01-17T01:04:36Z

@copilot use the collection literal syntax [ ... ] instead of yield return new object[] { ... }

- Changed from yield return new object[] { ... } to collection literal [ ... ] syntax - Applied to both TryWriteSignificandBigEndianTest_TestData and TryWriteSignificandLittleEndianTest_TestData - Uses expression-bodied member syntax with => for cleaner code Co-authored-by: EgorBo <523221+EgorBo@users.noreply.github.com>

Copilot · 2026-01-17T01:08:57Z

@copilot use the collection literal syntax [ ... ] instead of yield return new object[] { ... }

Done in commit 062a536. Updated both test data methods to use collection literal syntax [ ... ] instead of yield return new object[] { ... }.

skyoxZ · 2026-01-17T15:30:23Z

I believe BFloat16 has 8-bits significand? Then many codes in BFloat16 need to be corrected.

EgorBo · 2026-01-17T15:35:47Z

I believe BFloat16 has 8-bits significand? Then many codes in BFloat16 need to be corrected.

thanks! What do you mean by many codes?

skyoxZ · 2026-01-17T15:40:48Z

What do you mean by many code?

BFloat16.Significand should be a byte, and there're several (not too many :) ) related codes.

EgorBo · 2026-01-17T15:59:06Z

What do you mean by many code?

BFloat16.Significand should be a byte, and there're several (not too many :) ) related codes.

I'm a bit confused, @tannergooding @huoyaoyuan can you clarify the behavior here. So TryWriteSignificand used to report 4 bytes were written (which is clearly a bug), but the Significant itself is defined as ushort, while it looks like it's actually should be byte? (7 bits to be precise, hence 1 byte written)?

skyoxZ · 2026-01-17T16:19:10Z

Significand is an internal property so it doesn't affect behavior directly. However I noticed something confusing:

float.IFloatingPoint<float>.TryWriteExponentBigEndian: bytesWritten = 1, which I think it's correct.
float.IFloatingPoint<float>.TryWriteSignificandBigEndian: bytesWritten = 4. Maybe it should be 3?

huoyaoyuan · 2026-01-17T16:42:48Z

I copied the logic from Half blindly without taking much attention here. Since the TryWrite API is for supporting various lengths, I think we should require least space that's sufficient for significand (1 byte).

EgorBo · 2026-01-17T23:51:37Z

@copilot it turns out Significand has to be byte-sized, can you changes all the relevant bits in BFloat16 ?

- Changed Significand property from ushort to byte (BFloat16 has 8-bit significand) - Updated GetSignificandByteCount to return sizeof(byte) instead of sizeof(ushort) - Updated TryWriteSignificandBigEndian to write 1 byte directly (no endianness needed) - Updated TryWriteSignificandLittleEndian to write 1 byte directly (no endianness needed) - Updated test data to expect 1 byte written instead of 2 bytes - Updated test destination buffers from [0, 0] to [0] Co-authored-by: EgorBo <523221+EgorBo@users.noreply.github.com>

Copilot · 2026-01-17T23:58:10Z

@copilot it turns out Significand has to be byte-sized, can you changes all the relevant bits in BFloat16 ?

Done in commit ca5fb35. Changed Significand property from ushort to byte (8-bit), updated GetSignificandByteCount() to return 1, and modified both TryWriteSignificand methods to write a single byte. Since it's only 1 byte, endianness doesn't matter, so both methods now simply write destination[0] = Significand. Tests updated to expect 1 byte written.

tannergooding · 2026-01-19T16:55:44Z

Significand is an internal property so it doesn't affect behavior directly. However I noticed something confusing:

float.IFloatingPoint.TryWriteExponentBigEndian: bytesWritten = 1, which I think it's correct.

float.IFloatingPoint.TryWriteSignificandBigEndian: bytesWritten = 4. Maybe it should be 3?

We generally want the "size written" to be a primitive size and otherwise a multiple of 4 or 8. That is, 3 is essentially a "worst possible size" to write out because it means no primitive can trivially work with the data.

So we have:

Type	Sign	Exponent	Significand
half	1-bit	5-bits	10-bits
bfloat16	1-bit	8-bits	7-bits
single	1-bit	8-bits	23-bits
double	1-bit	11-bits	52-bits
quad	1-bit	15-bits	112-bits

Which means we want:

Type	Sign	Exponent	Significand
half	bool	1 byte	2 bytes
bfloat16	bool	1 byte	1 byte
single	bool	1 byte	4 bytes
double	bool	2 bytes	8 bytes
quad	bool	2 bytes	16 bytes

Edit: sorry, hit enter too soon and it posted.

It's then worth noting that despite there being that many "stored" bits, that is not how many bits actually exist. There is an implicit significand bit that exists, so the actual amount is:

Type	Actual Significand
half	11-bits
bfloat16	8-bits
single	1-bit
double	1-bit
quad	113-bits

This could theoretically bump the limit over to the next size in some cases. It doesn't for any of these, however, so the fixes in this PR should be correct.

Initial plan

1124773

Copilot AI assigned Copilot and EgorBo Jan 16, 2026

Copilot started work on behalf of EgorBo January 16, 2026 20:57 View session

Copilot AI changed the title ~~[WIP] Fix bytesWritten reporting in BFloat16 TryWriteSignificand methods~~ Fix BFloat16.TryWriteSignificand bytesWritten reporting 4 instead of 2 Jan 16, 2026

Copilot AI requested a review from EgorBo January 16, 2026 21:29

Copilot finished work on behalf of EgorBo January 16, 2026 21:29

EgorBo marked this pull request as ready for review January 16, 2026 21:59

Copilot AI review requested due to automatic review settings January 16, 2026 21:59

Copilot started reviewing on behalf of EgorBo January 16, 2026 22:00 View session

Copilot AI reviewed Jan 16, 2026

View reviewed changes

stephentoub reviewed Jan 16, 2026

View reviewed changes

src/libraries/System.Runtime/tests/System.Runtime.Tests/System/Numerics/BFloat16Tests.cs Outdated Show resolved Hide resolved

Copilot started work on behalf of stephentoub January 16, 2026 22:20 View session

Copilot AI requested a review from stephentoub January 16, 2026 22:39

Copilot finished work on behalf of stephentoub January 16, 2026 22:39

Copilot started work on behalf of EgorBo January 17, 2026 00:39 View session

Copilot finished work on behalf of EgorBo January 17, 2026 00:47

Copilot started work on behalf of EgorBo January 17, 2026 01:05 View session

Copilot finished work on behalf of EgorBo January 17, 2026 01:10

Update BFloat16Tests.cs

84166d8

This was referenced Jan 17, 2026

Unable to pull image from mcr.microsoft.com #117164

Open

browser-wasm linux LibraryTests build failure - NETSDK1022: Duplicate 'Content' items #123237

Closed

Copilot started work on behalf of EgorBo January 17, 2026 15:36 View session

This comment was marked as outdated.

Sign in to view

Copilot finished work on behalf of EgorBo January 17, 2026 15:41

Copilot started work on behalf of EgorBo January 17, 2026 23:52 View session

Copilot AI changed the title ~~Fix BFloat16.TryWriteSignificand bytesWritten reporting 4 instead of 2~~ Fix BFloat16.Significand to be byte-sized and correct TryWriteSignificand methods Jan 17, 2026

Copilot finished work on behalf of EgorBo January 17, 2026 23:59

tannergooding approved these changes Jan 19, 2026

View reviewed changes

tannergooding merged commit de88fe7 into main Jan 19, 2026
142 of 146 checks passed

tannergooding deleted the copilot/fix-bfloat16-byteswritten branch January 19, 2026 17:28

dotnet-maestro bot mentioned this pull request Jan 20, 2026

[main] Source code updates from dotnet/runtime dotnet/dotnet#4277

Merged

	bool IFloatingPoint<BFloat16>.TryWriteSignificandBigEndian(Span<byte> destination, out int bytesWritten)
	{
	if (BinaryPrimitives.TryWriteUInt16BigEndian(destination, Significand))
	{
	bytesWritten = sizeof(uint);

Fix BFloat16.Significand to be byte-sized and correct TryWriteSignificand methods #123288

Fix BFloat16.Significand to be byte-sized and correct TryWriteSignificand methods #123288

Uh oh!

Conversation

Copilot AI commented Jan 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Changes

Comments on the Issue (you are @copilot in this section)

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

EgorBo commented Jan 17, 2026

Uh oh!

Copilot AI commented Jan 17, 2026

Uh oh!

EgorBo commented Jan 17, 2026

Uh oh!

Copilot AI commented Jan 17, 2026

Uh oh!

skyoxZ commented Jan 17, 2026

Uh oh!

EgorBo commented Jan 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

skyoxZ commented Jan 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

This comment was marked as outdated.

EgorBo commented Jan 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

skyoxZ commented Jan 17, 2026

Uh oh!

huoyaoyuan commented Jan 17, 2026

Uh oh!

EgorBo commented Jan 17, 2026

Uh oh!

Copilot AI commented Jan 17, 2026

Uh oh!

tannergooding commented Jan 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

Copilot AI commented Jan 16, 2026 •

edited

Loading

EgorBo commented Jan 17, 2026 •

edited

Loading

skyoxZ commented Jan 17, 2026 •

edited

Loading

EgorBo commented Jan 17, 2026 •

edited

Loading

tannergooding commented Jan 19, 2026 •

edited

Loading