Adding the convenience Avx.Compare* overloads #34100
Conversation
|
Note regarding the This serves as a reminder for when your PR is modifying a ref *.cs file and adding/modifying public APIs, to please make sure the API implementation in the src *.cs file is documented with triple slash comments, so the PR reviewers can sign off that change. |
Dotnet-GitSync-Bot
added
area-CodeGen-coreclr
|
I found #34094 while implementing these APIs. |
src/libraries/System.Private.CoreLib/src/System/Runtime/Intrinsics/X86/Avx.cs
Show resolved
Hide resolved
| /// CMPPS ymm, ymm/m256, imm8(0) | ||
| /// The above native signature does not exist. We provide this additional overload for completeness. | ||
| /// </summary> | ||
| public static Vector256<float> CompareEqual(Vector256<float> left, Vector256<float> right) => Compare(left, right, FloatComparisonMode.OrderedEqualNonSignaling); |
There was a problem hiding this comment.
@tannergooding Can you explain why CompareEqual corresponds to OrderedEqualNonSignaling while CompareGreaterThan corresponds to OrderedGreaterThanSignaling?
There was a problem hiding this comment.
The simplest explanation is because that is what the immediate values available to Sse/Sse2 are (imm8 == 0x00 through 0x07, inclusive; 0x08 through 0x1F are Avx+):
The lengthier explanation is that Signaling when floating-point exceptions are supported (we don't support and explicitly disable them in .NET) throw for SNaN inputs but not for QNaN. The comparison operations like ==, !=, >, >=, <, and <= don't expect NaN as inputs and so they signal by default; while the Ordered/Unordered comparisons do expect NaN inputs and so they don't.
ghost
locked as resolved and limited conversation to collaborators
This resolves #31193 by implementing the approved
Avx.Compare*convenience overloads. There is no JIT side work for these as they just forward to the existing implementation that takes theFloatComparisonModeenumCC. @CarolEidt, @echesakovMSFT