Skip to content

JIT: Faster vector == Vector128.Zero on arm64 #65632

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
EgorBo merged 11 commits into from
Feb 23, 2022
Merged

JIT: Faster vector == Vector128.Zero on arm64 #65632

Show file tree
Hide file tree
Changes from 8 commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
466afe6
Optimize vec == Vector.Zero for arm64
EgorBo Feb 20, 2022
3139218
Clean up
EgorBo Feb 20, 2022
a4143c4
Update lowerarmarch.cpp
EgorBo Feb 20, 2022
f9d45e0
Update lowerarmarch.cpp
EgorBo Feb 20, 2022
51f727f
Update lowerarmarch.cpp
EgorBo Feb 20, 2022
70e18b7
Apply Zoltan's patch
EgorBo Feb 21, 2022
1550865
Merge branch 'main' of https://github.com/dotnet/runtime into arm-fas…
EgorBo Feb 21, 2022
c1c7831
Address feedback
EgorBo Feb 21, 2022
5c1a77e
Merge branch 'main' of https://github.com/dotnet/runtime into arm-fas…
EgorBo Feb 22, 2022
59609c4
Address feedback
EgorBo Feb 22, 2022
bc9220b
use UINT for V128
EgorBo Feb 22, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
47 changes: 47 additions & 0 deletions src/coreclr/jit/lowerarmarch.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -868,6 +868,53 @@ void Lowering::LowerHWIntrinsicCmpOp(GenTreeHWIntrinsic* node, genTreeOps cmpOp)
GenTree* op1 = node->Op(1);
GenTree* op2 = node->Op(2);

// Optimize comparison against Vector64/128<>.Zero via UMAX:
//
// bool eq = v == Vector128<integer>.Zero
//
// to:
//
// bool eq = AdvSimd.Arm64.MaxAcross(v.AsByte()).ToScalar() == 0;
//
GenTree* op = nullptr;
GenTree* opZero = nullptr;
if (op1->IsVectorZero())
{
op = op2;
opZero = op1;
}
else if (op2->IsVectorZero())
{
op = op1;
opZero = op2;
}

if (!varTypeIsFloating(simdBaseType) && (op != nullptr))
{
GenTree* cmp =
comp->gtNewSimdHWIntrinsicNode(simdType, op, NI_AdvSimd_Arm64_MaxAcross, CORINFO_TYPE_UBYTE, simdSize);
Copy link
Contributor

@echesakov echesakov Feb 22, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

According to Arm® Cortex®-A76 Software Optimization Guide:
UMAXV, 16B has Exec latency 6 and Execution throughput 1/2
while UMAXV, 4H/4S has Exec latency 3 and Execution throughput 1

Do we want CORINFO_TYPE_USHORT/CORINFO_TYPE_UINT as a base type instead?

Copy link
Member Author

@EgorBo EgorBo Feb 22, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sure, let's change, Although, I had this in mind when I was benchmarking it and saw zero difference, but I now I see why:
image
(Apple M1, Firestorm core)

BlockRange().InsertBefore(node, cmp);
LowerNode(cmp);
BlockRange().Remove(opZero);

GenTree* val = comp->gtNewSimdHWIntrinsicNode(TYP_INT, cmp, NI_Vector128_ToScalar, CORINFO_TYPE_UINT, simdSize);
BlockRange().InsertAfter(cmp, val);
LowerNode(val);

GenTree* cmpZroCns = comp->gtNewIconNode(0, TYP_INT);
BlockRange().InsertAfter(val, cmpZroCns);

node->ChangeOper(cmpOp);
node->gtType = TYP_INT;
node->AsOp()->gtOp1 = val;
node->AsOp()->gtOp2 = cmpZroCns;
LowerNodeCC(node, (cmpOp == GT_EQ) ? GenCondition::EQ : GenCondition::NE);
node->gtType = TYP_VOID;
node->ClearUnusedValue();
LowerNode(node);
return;
}

NamedIntrinsic cmpIntrinsic;

switch (simdBaseType)
Expand Down
2 changes: 2 additions & 0 deletions src/mono/mono/mini/simd-intrinsics.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -773,6 +773,8 @@ emit_sri_vector (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsi
}
case SN_Create: {
MonoType *etype = get_vector_t_elem_type (fsig->ret);
if (!MONO_TYPE_IS_INTRINSICS_VECTOR_PRIMITIVE (etype))
return NULL;
if (fsig->param_count == 1 && mono_metadata_type_equal (fsig->params [0], etype))
return emit_simd_ins (cfg, klass, type_to_expand_op (etype), args [0]->dreg, -1);
else if (is_create_from_half_vectors_overload (fsig))
Expand Down
88 changes: 88 additions & 0 deletions src/tests/JIT/HardwareIntrinsics/General/HwiOp/CompareVectorWithZero.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,88 @@
// Licensed to the .NET Foundation under one or more agreements.
// The .NET Foundation licenses this file to you under the MIT license.

using System;
using System.Runtime.CompilerServices;
using System.Runtime.Intrinsics;

public class CompareVectorWithZero
{
public static int Main()
{
Test(Vector128.Create(0));
Test(Vector128.Create(0.0f));
Test(Vector128.Create(-0.0f));
Test(Vector128.Create(0.0));
Test(Vector128.Create(-0.0));

TestReversed(Vector128.Create(0));
TestReversed(Vector128.Create(0.0f));
TestReversed(Vector128.Create(-0.0f));
TestReversed(Vector128.Create(0.0));
TestReversed(Vector128.Create(-0.0));

Test(Vector128.Create(-10));
Test(Vector128.Create(10));
Test(Vector128.Create((sbyte)-10));
Test(Vector128.Create((ushort)10));
Test(Vector64.Create(0));
Test(Vector64.Create(0.0f));
Test(Vector64.Create(-0.0f));
Test(Vector64.Create(0.0));
Test(Vector64.Create(-0.0));
Test(Vector64.Create(-10));
Test(Vector64.Create(10));
Test(Vector64.Create((sbyte)-10));
Test(Vector64.Create((ushort)10));
Test(Vector128.Create(0, 0, 0, 0, 0, 0, 0, 1));
Test(Vector128.Create(0, 0, 0, 0, 0, 0, 0, -1));
Test(Vector64.Create(0, 0, 0, 0, 0, 0, 0, 1));
Test(Vector64.Create(0, 0, 0, 0, 0, 0, 0, -1));

Test(Vector128.Create(0, 0, 0, 0, 0, 0, 1, 0));
Test(Vector128.Create(0, 0, 0, 0, 0, 0, 1, 0));
Test(Vector64.Create(0, 0, 0, 0, 0, 0, -1, 0));
Test(Vector64.Create(0, 0, 0, 0, 0, 0, -1, 0));

Test(Vector128.Create(0, 0, 0, 1, 0, 0, 0, 1));
Test(Vector128.Create(0, 0, 0, -1, 0, 0, 0, -1));
Test(Vector64.Create(0, 0, 0, 1, 0, 0, 0, 1));
Test(Vector64.Create(0, 0, 0, -1, 0, 0, 0, -1));

TestReversed(Vector128.Create(0, 0, 0, 1, 0, 0, 0, 1));
TestReversed(Vector128.Create(0, 0, 0, -1, 0, 0, 0, -1));
TestReversed(Vector64.Create(0, 0, 0, 1, 0, 0, 0, 1));
TestReversed(Vector64.Create(0, 0, 0, -1, 0, 0, 0, -1));
return 100;
}

[MethodImpl(MethodImplOptions.NoInlining)]
public static T ToVar<T>(T t) => t;

[MethodImpl(MethodImplOptions.NoInlining)]
public static void AssertTrue(bool expr)
{
if (!expr)
throw new InvalidOperationException();
}

[MethodImpl(MethodImplOptions.NoInlining)]
public static void Test<T>(Vector128<T> v) where T : unmanaged =>
AssertTrue((v == Vector128<T>.Zero) ==
(v == Vector128.Create(ToVar(default(T)))));

[MethodImpl(MethodImplOptions.NoInlining)]
public static void Test<T>(Vector64<T> v) where T : unmanaged =>
AssertTrue((v == Vector64<T>.Zero) ==
(v == Vector64.Create(ToVar(default(T)))));

[MethodImpl(MethodImplOptions.NoInlining)]
public static void TestReversed<T>(Vector128<T> v) where T : unmanaged =>
AssertTrue((Vector128<T>.Zero == v) ==
(v == Vector128.Create(ToVar(default(T)))));

[MethodImpl(MethodImplOptions.NoInlining)]
public static void TestReversed<T>(Vector64<T> v) where T : unmanaged =>
AssertTrue((Vector64<T>.Zero == v) ==
(v == Vector64.Create(ToVar(default(T)))));
}
9 changes: 9 additions & 0 deletions src/tests/JIT/HardwareIntrinsics/General/HwiOp/CompareVectorWithZero.csproj
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
<Project Sdk="Microsoft.NET.Sdk">
<PropertyGroup>
<OutputType>Exe</OutputType>
<Optimize>True</Optimize>
</PropertyGroup>
<ItemGroup>
<Compile Include="$(MSBuildProjectName).cs" />
</ItemGroup>
</Project>