Skip to content

Add floating-point routines to SVE microbenchmark #5043

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

Open
ylpoonlg wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Add floating-point routines to SVE microbenchmark #5043

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
206 changes: 206 additions & 0 deletions src/benchmarks/micro/sve/Exponent.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,206 @@
using System;
using System.Diagnostics;
using System.Numerics;
using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.Arm;
using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Extensions;
using BenchmarkDotNet.Configs;
using BenchmarkDotNet.Filters;
using MicroBenchmarks;

namespace SveBenchmarks
{
[BenchmarkCategory(Categories.Runtime)]
[OperatingSystemsArchitectureFilter(allowed: true, System.Runtime.InteropServices.Architecture.Arm64)]
[Config(typeof(Config))]
public class Exponent
{
private class Config : ManualConfig
{
public Config()
{
AddFilter(new SimpleFilter(_ => Sve.IsSupported));
}
}

[Params(15, 127, 527, 10015)]
public int Size;

private float[] _input;
private float[] _data_sve;
private float[] _data_neon;
private float[] _output;

[GlobalSetup]
public virtual void Setup()
{
_input = ValuesGenerator.Array<float>(Size);

_data_sve = new float[] {
// c1, c3, inv_ln2
BitConverter.UInt32BitsToSingle(0x3f000000),
BitConverter.UInt32BitsToSingle(0x3d2aaab5),
BitConverter.UInt32BitsToSingle(0x3fb8aa3b),
// ln2_lo, c0, c2, c4
BitConverter.UInt32BitsToSingle(0x35bfbe8e),
BitConverter.UInt32BitsToSingle(0x3f800000),
BitConverter.UInt32BitsToSingle(0x3e2aaaab),
BitConverter.UInt32BitsToSingle(0x3c057330),
// ln2_hi, shift
BitConverter.UInt32BitsToSingle(0x3f317200),
BitConverter.UInt32BitsToSingle(0x48401fc0),
};

_data_neon = new float[] {
// inv_ln2, ln2_lo, c0, c2
BitConverter.UInt32BitsToSingle(0x3fb8aa3b),
BitConverter.UInt32BitsToSingle(0x35bfbe8e),
BitConverter.UInt32BitsToSingle(0x3c07cfce),
BitConverter.UInt32BitsToSingle(0x3e2aad40),
// ln2_hi, shift, c1, c3, c4
BitConverter.UInt32BitsToSingle(0x3f317200),
BitConverter.UInt32BitsToSingle(0x4b40007f),
BitConverter.UInt32BitsToSingle(0x3d2b9d0d),
BitConverter.UInt32BitsToSingle(0x3efffee3),
BitConverter.UInt32BitsToSingle(0x3f7ffffb),
};

_output = new float[Size];
}

[GlobalCleanup]
public virtual void Verify()
{
float[] current = (float[])_output.Clone();
Setup();
Scalar();
float[] scalar = (float[])_output.Clone();

// Check that the result is the same as scalar (within 3ULP).
for (int i = 0; i < Size; i++)
{
int e = (int)(BitConverter.SingleToUInt32Bits(scalar[i]) >> 23 & 0xff);
if (e == 0) e++;
float ulpScale = (float)Math.ScaleB(1.0, e - 127 - 23);
float ulpError = (float)Math.Abs(current[i] - scalar[i]) / ulpScale;
Debug.Assert(ulpError <= 3);
}
}

[Benchmark]
public unsafe void Scalar()
{
fixed (float* input = _input, output = _output)
{
for (int i = 0; i < Size; i++)
{
output[i] = (float)Math.Exp(input[i]);
}
}
}

[Benchmark]
public unsafe void Vector128Exponent()
{
// Algorithm based on Arm Optimized-Routines.
// https://github.com/ARM-software/optimized-routines/blob/v25.07/math/aarch64/advsimd/expf.c
fixed (float* input = _input, output = _output, d = _data_neon)
{
int i = 0;

Vector128<float> constVec = AdvSimd.LoadVector128(d);
Vector128<float> ln2hiVec = Vector128.Create(d[4]);
Vector128<float> shiftVec = Vector128.Create(d[5]);
Vector128<float> c1Vec = Vector128.Create(d[6]);
Vector128<float> c3Vec = Vector128.Create(d[7]);
Vector128<float> c4Vec = Vector128.Create(d[8]);

for (; i < Size - 4; i += 4)
{
Vector128<float> x = AdvSimd.LoadVector128(input + i);

// z = shift + x * 1/ln2
Vector128<float> z = AdvSimd.Arm64.FusedMultiplyAddBySelectedScalar(shiftVec, x, constVec, 0);
// -n = shift - z
Vector128<float> neg_n = AdvSimd.Subtract(shiftVec, z);
// scale = z << 23
Vector128<float> scale = AdvSimd.ShiftLeftLogical(z.AsUInt32(), 23).AsSingle();

// r = x - n * ln2_hi
Vector128<float> r = AdvSimd.FusedMultiplyAdd(x, neg_n, ln2hiVec);
// r = r - n * ln2_lo
r = AdvSimd.Arm64.FusedMultiplyAddBySelectedScalar(r, neg_n, constVec, 1);
Vector128<float> r2 = AdvSimd.Multiply(r, r);

// poly(r) = exp(r) - 1.
Vector128<float> p10 = AdvSimd.Arm64.FusedMultiplyAddBySelectedScalar(c1Vec, r, constVec, 2);
Vector128<float> p32 = AdvSimd.Arm64.FusedMultiplyAddBySelectedScalar(c3Vec, r, constVec, 3);
Vector128<float> p30 = AdvSimd.FusedMultiplyAdd(p32, r2, p10);
Vector128<float> p4 = AdvSimd.Multiply(r, c4Vec);
Vector128<float> poly = AdvSimd.FusedMultiplyAdd(p4, r2, p30);

// result = scale * (1 + poly).
Vector128<float> result = AdvSimd.FusedMultiplyAdd(scale, poly, scale);
AdvSimd.Store(output + i, result);
}
// Handle remaining elements.
for (; i < Size; i++)
{
output[i] = (float)Math.Exp(input[i]);
}
}
}

[Benchmark]
public unsafe void SveExponent()
{
fixed (float* input = _input, output = _output, d = _data_sve)
{
int i = 0;
int cntw = (int)Sve.Count32BitElements();

Vector<float> c1Vec = new Vector<float>(d[0]);
Vector<float> c3Vec = new Vector<float>(d[1]);
Vector<float> invln2Vec = new Vector<float>(d[2]);
Vector<float> shiftVec = new Vector<float>(d[8]);
Vector<float> ln2hiVec = new Vector<float>(d[7]);
Vector<float> constVec = Sve.LoadVector(Sve.CreateTrueMaskSingle(), &d[3]);

Vector<uint> pTrue = Sve.CreateTrueMaskUInt32();
Vector<uint> pLoop = Sve.CreateWhileLessThanMask32Bit(0, Size);
while (Sve.TestFirstTrue(pTrue, pLoop))
{
Vector<float> x = (Vector<float>)Sve.LoadVector(pLoop, (uint*)(input + i));

// n = round(x/(ln2/N)).
Vector<float> z = Sve.FusedMultiplyAdd(shiftVec, invln2Vec, x);
Vector<float> n = Sve.Subtract(z, shiftVec);

// r = x - n*ln2/N.
Vector<float> r = Sve.FusedMultiplySubtract(x, ln2hiVec, n);
r = Sve.FusedMultiplySubtractBySelectedScalar(r, n, constVec, 0);
// scale = 2^(n/N).
Vector<float> scale = Sve.FloatingPointExponentialAccelerator((Vector<uint>)z);

// poly(r) = exp(r) - 1.
Vector<float> p12 = Sve.FusedMultiplyAddBySelectedScalar(c1Vec, r, constVec, 2);
Vector<float> p34 = Sve.FusedMultiplyAddBySelectedScalar(c3Vec, r, constVec, 3);
Vector<float> r2 = Sve.Multiply(r, r);
Vector<float> p14 = Sve.FusedMultiplyAdd(p12, p34, r2);
Vector<float> p0 = Sve.MultiplyBySelectedScalar(r, constVec, 1);
Vector<float> poly = Sve.FusedMultiplyAdd(p0, r2, p14);

// result = scale * (1 + poly).
Vector<float> result = Sve.FusedMultiplyAdd(scale, poly, scale);
Sve.StoreAndZip(pLoop, (uint*)output + i, (Vector<uint>)result);

// Handle loop.
i += cntw;
pLoop = Sve.CreateWhileLessThanMask32Bit(i, Size);
}
}
}

}
}
Loading
Loading