Arm64: Use VectorTableLookup/VectorTableLookupExtension APIs in libraries

[kunalspathak]

#80297 added support for Arm64's VectorTableLookup and VectorTableLookupExtension APIs. We should start using them in libraries methods.

[kunalspathak]

@a74nh @SwapnilGaikwad

Tagging subscribers to this area: @dotnet/area-system-runtime
See info in area-owners.md if you want to be subscribed.

Issue Details

---

#80297 added support for Arm64's VectorTableLookup and VectorTableLookupExtension APIs. We should start using them in libraries methods.

Author:	kunalspathak
Assignees:	-
Labels:	area-System.Runtime, untriaged, needs-area-label
Milestone:	-

[a74nh]

Looking through the libraries the current uses of shuffle/vectortablelookup can be improved:

• System.Private.CoreLib/src/System/IndexOfAnyValues/ProbabilisticMap.cs

  • Two consecutive shuffles can be combined into a vectorlookup2.
  • Use double VectorTableLookup on ARM in ProbabilisticMap #85189

• System.Private.CoreLib/src/System/IndexOfAnyValues/IndexOfAnyAsciiSearcher.cs

  • Can combine shuffles, but only the instances where there are two bitmaps in a single function.
  • WIP in https://github.com/dotnet/runtime/pull/88183/files

• System.Private.CoreLib/src/System/Buffers/Text/Utf8Formatter/Utf8Formatter.Guid.cs

  • Should be able to turn vector z into a single vectorlookup2.
  • Done in Combine two table lookups into one while formatting Guid on Arm64 #87126

• System.Private.CoreLib/src/System/Buffers/Binary/BinaryPrimitives.ReverseEndianness.cs
  System.Private.CoreLib/src/System/SpanHelpers.Byte.cs
  System.Private.CoreLib/src/System/SpanHelpers.cs
  Common/src/System/HexConverter.cs
  System.IO.Hashing/src/System/IO/Hashing/XxHashShared.cs

  • Technically out of scope of this issue, but these are all using hardcoded index values, so can switch from shuffle to shuffleunsafe.

• EncodeToUtf8.cs
  DecodeFromUtf8.cs

  • Once LD4 and ST3 available, then they can be rewritten to use to vectortablelookup2.

There may be additional things, but these were the obvious ones.

[MihaZupan]

System.Private.CoreLib/src/System/IndexOfAnyValues/IndexOfAnyAsciiSearcher.cs

• Can combine shuffles, but only the instances where there are two bitmaps in a single function.

Do you mean

runtime/src/libraries/System.Private.CoreLib/src/System/IndexOfAnyValues/IndexOfAnyAsciiSearcher.cs

Lines 912 to 920 in 4ebecb0

	Vector128<byte> row0 = Vector128.ShuffleUnsafe(bitmapLookup0, lowNibbles);
	Vector128<byte> row1 = Vector128.ShuffleUnsafe(bitmapLookup1, lowNibbles);
	Vector128<byte> bitmask = Vector128.ShuffleUnsafe(Vector128.Create(0x8040201008040201).AsByte(), highNibbles);
	Vector128<byte> mask = Vector128.GreaterThan(highNibbles.AsSByte(), Vector128.Create((sbyte)0x7)).AsByte();
	Vector128<byte> bitsets = Vector128.ConditionalSelect(mask, row1, row0);
	Vector128<byte> result = Vector128.Equals(bitsets & bitmask, bitmask);

?
Which operations would you combine here?

[kunalspathak]

@JulieLeeMSFT - FYI.

[a74nh]

Do you mean

runtime/src/libraries/System.Private.CoreLib/src/System/IndexOfAnyValues/IndexOfAnyAsciiSearcher.cs

Lines 912 to 920 in 4ebecb0

	Vector128<byte> row0 = Vector128.ShuffleUnsafe(bitmapLookup0, lowNibbles);
	Vector128<byte> row1 = Vector128.ShuffleUnsafe(bitmapLookup1, lowNibbles);
	Vector128<byte> bitmask = Vector128.ShuffleUnsafe(Vector128.Create(0x8040201008040201).AsByte(), highNibbles);
	Vector128<byte> mask = Vector128.GreaterThan(highNibbles.AsSByte(), Vector128.Create((sbyte)0x7)).AsByte();
	Vector128<byte> bitsets = Vector128.ConditionalSelect(mask, row1, row0);
	Vector128<byte> result = Vector128.Equals(bitsets & bitmask, bitmask);

?
Which operations would you combine here?

Unless there's something I'm missing, bitsets can be created just via:
Vector128 bitsets = AdvSimd.VectorTableLookup2(bitmapLookup0, bitmapLookup1, lowNibbles);

[SwapnilGaikwad]

System.Private.CoreLib/src/System/Buffers/Binary/BinaryPrimitives.ReverseEndianness.cs
System.Private.CoreLib/src/System/SpanHelpers.Byte.cs
System.Private.CoreLib/src/System/SpanHelpers.cs
Common/src/System/HexConverter.cs
System.IO.Hashing/src/System/IO/Hashing/XxHashShared.cs

* Technically out of scope of this issue, but these are all using hardcoded index values, so can switch from shuffle to shuffleunsafe.

For these cases, changing from Shuffle to ShuffleUnsafe would not give any performance improvements.
When the indices are hardcoded, they get optimised during the importation phase and avoid the bounds check logic. Therefore, we get the same JIT-ed assembly sequence.

Even when the indices are out of the bounds of the vector, if the positions vector is hard-coded then the overhead of bounds-check is avoided with the shuffle.
In the following example, the index for the first position is 25 so with Vector128.Shuffle we expect value 0 in dest[0] but with Ssse3.Shuffle we expect value src[9] (25%16=9).

Vector128.Shuffle(sourceVec, Vector128.Create((byte)25, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0));

Instead of getting a bounds check, after the importation phase, we see an updated index FF for the first position that would set value 0 in dest[0].
With Vector128.ShuffleUnsafe we get index 19 instead.

; Assembly listing for method System.Text.Tests.ShuffleUnsafeTest:ReverseWithShuffle(byref,ulong)
; Emitting BLENDED_CODE for X64 CPU with AVX512 - Unix
; optimized code
; rsp based frame
; partially interruptible
; No PGO data
; Final local variable assignments
;
;  V00 arg0         [V00,T00] (  4,  4   )   byref  ->  rdi         single-def
;* V01 arg1         [V01    ] (  0,  0   )    long  ->  zero-ref    single-def
;# V02 OutArgs      [V02    ] (  1,  1   )  struct ( 0) [rsp+00H]   do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
;  V03 tmp1         [V03,T01] (  2,  4   )  simd16  ->  mm0         "Spilling op1 side effects for HWIntrinsic"
;
; Lcl frame size = 0

G_M46485_IG01:  ;; offset=0000H
       vzeroupper
                        ;; size=3 bbWeight=1 PerfScore 1.00
G_M46485_IG02:  ;; offset=0003H
       vmovups  xmm0, xmmword ptr [rdi]
       vpshufb  xmm0, xmm0, xmmword ptr [reloc @RWD00]
       vmovups  xmmword ptr [rdi], xmm0
                        ;; size=17 bbWeight=1 PerfScore 8.00
G_M46485_IG03:  ;; offset=0014H
       ret
                        ;; size=1 bbWeight=1 PerfScore 1.00
RWD00   dq  08090A0B0C0D0EFFh, 0001020304050607h

Assembly for `Vector128.ShuffleUnsafe(sourceVec, Vector128.Create((byte)25, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0));`.

; Assembly listing for method System.Text.Tests.ShuffleUnsafeTest:ReverseWithShuffleUnsafe(byref,ulong)
; Emitting BLENDED_CODE for X64 CPU with AVX512 - Unix
; optimized code
; rsp based frame
; partially interruptible
; No PGO data
; 0 inlinees with PGO data; 0 single block inlinees; 1 inlinees without PGO data
; Final local variable assignments
;
;  V00 arg0         [V00,T00] (  4,  4   )   byref  ->  rdi         single-def
;* V01 arg1         [V01    ] (  0,  0   )    long  ->  zero-ref    single-def
;# V02 OutArgs      [V02    ] (  1,  1   )  struct ( 0) [rsp+00H]   do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
;  V03 tmp1         [V03,T01] (  2,  4   )  simd16  ->  mm0         "Spilling op1 side effects for HWIntrinsic"
;* V04 tmp2         [V04    ] (  0,  0   )  simd16  ->  zero-ref    "Inlining Arg"
;
; Lcl frame size = 0

G_M34751_IG01:  ;; offset=0000H
       vzeroupper
                        ;; size=3 bbWeight=1 PerfScore 1.00
G_M34751_IG02:  ;; offset=0003H
       vmovups  xmm0, xmmword ptr [rdi]
       vpshufb  xmm0, xmm0, xmmword ptr [reloc @RWD00]
       vmovups  xmmword ptr [rdi], xmm0
                        ;; size=17 bbWeight=1 PerfScore 8.00
G_M34751_IG03:  ;; offset=0014H
       ret
                        ;; size=1 bbWeight=1 PerfScore 1.00
RWD00   dq  08090A0B0C0D0E19h, 0001020304050607h


; Total bytes of code 21, prolog size 3, PerfScore 12.10, instruction count 5, allocated bytes for code 21 (MethodHash=64be7840) for method System.Text.Tests.ShuffleUnsafeTest:ReverseWithShuffleUnsafe(byref,ulong)
; ============================================================