try to port ASCIIUtility.NarrowUtf16ToAscii to x-plat intrinsics #73064
Conversation
|
Tagging subscribers to this area: @dotnet/area-system-text-encoding Issue DetailsFor both x64 and ARM64 I am observing 10-15% regression. x64DetailsBenchmarkDotNet=v0.13.1.1828-nightly, OS=Windows 11 (10.0.22000.795/21H2)
AMD Ryzen Threadripper PRO 3945WX 12-Cores, 1 CPU, 24 logical and 12 physical cores
.NET SDK=7.0.100-preview.6.22352.1
[Host] : .NET 7.0.0 (7.0.22.32404), X64 RyuJIT AVX2
Method=GetBytes
Here I was able to use VTune (I also tried uProf but it leaves a lot to desire): 
From what I can see the regression comes from additional 3
It seems that the first one comes from the new ARM64DetailsBenchmarkDotNet=v0.13.1.1828-nightly, OS=ubuntu 20.04
Unknown processor
.NET SDK=7.0.100-rc.1.22379.1
[Host] : .NET 7.0.0 (7.0.22.37802), Arm64 RyuJIT AdvSIMD
Method=GetBytes
|
|
|
||
| ref byte asciiBuffer = ref *pAsciiBuffer; | ||
| Vector128<byte> asciiVector = ExtractAsciiVector(utf16VectorFirst, utf16VectorFirst); | ||
| Vector128<byte> asciiVector = Vector128.Narrow(utf16VectorFirst, utf16VectorFirst); |
There was a problem hiding this comment.
@adamsitnik Narrow doesn't know that utf16VectorFirst is already ASCII at this point (from my understanding) so it applies a mask via AND to cut anything above 0xFF (not needed in this case). Consider this:
So for this case for better perf you probably want to keep ExtractAsciiVector
There was a problem hiding this comment.
Can we log an issue here for Narrow. There is possibly some codegen improvements we can make in .NET 8
| const ushort asciiMask = ushort.MaxValue - 127; // 0x7F80 | ||
| Vector128<ushort> zeroIsAscii = utf16Vector & Vector128.Create(asciiMask); | ||
| // If a non-ASCII bit is set in any WORD of the vector, we have seen non-ASCII data. | ||
| return !Vector128.EqualsAll(zeroIsAscii, Vector128<ushort>.Zero); |
There was a problem hiding this comment.
You said that it's expected to have a redundant AND here - why? 🙂
|
@tannergooding Since I was not able to get the same perf with |
|
Improvements - dotnet/perf-autofiling-issues#7322 |
ghost
locked as resolved and limited conversation to collaborators
4 participants
For both x64 and ARM64 I am observing 10-15% regression.
x64
Details
Here I was able to use VTune (I also tried uProf but it leaves a lot to desire):
From what I can see the regression comes from additional 3
vpandinstructions:It seems that the first one comes from the new
VectorContainsNonAsciiCharimplementation (this is expected). But the other two fromVector128.Narrow? @tannergooding is that expected?ARM64
Details