static unsafe int Main(string[] args)
{
int testResult = Pass;
if (Sse3.IsSupported)
{
using (TestTable <float> floatTable = new TestTable <float>(new float[4] {
1, -5, 100, 0
}, new float[4]))
{
var vf1 = Sse.LoadVector128((float *)(floatTable.inArrayPtr));
var vf2 = Sse3.MoveHighAndDuplicate(vf1);
Unsafe.Write(floatTable.outArrayPtr, vf2);
if (BitConverter.SingleToInt32Bits(floatTable.inArray[1]) != BitConverter.SingleToInt32Bits(floatTable.outArray[0]) ||
BitConverter.SingleToInt32Bits(floatTable.inArray[1]) != BitConverter.SingleToInt32Bits(floatTable.outArray[1]) ||
BitConverter.SingleToInt32Bits(floatTable.inArray[3]) != BitConverter.SingleToInt32Bits(floatTable.outArray[2]) ||
BitConverter.SingleToInt32Bits(floatTable.inArray[3]) != BitConverter.SingleToInt32Bits(floatTable.outArray[3]))
{
Console.WriteLine("Sse3 MoveHighAndDuplicate failed on float:");
foreach (var item in floatTable.outArray)
{
Console.Write(item + ", ");
}
Console.WriteLine();
testResult = Fail;
}
}
}
return(testResult);
}
static int Main()
{
s_success = true;
// We expect the AOT compiler generated HW intrinsics with the following characteristics:
//
// * TRUE = IsSupported assumed to be true, no runtime check
// * NULL = IsSupported is a runtime check, code should be behind the check or bad things happen
// * FALSE = IsSupported assumed to be false, no runtime check, PlatformNotSupportedException if used
//
// The test is compiled with multiple defines to test this.
#if BASELINE_INTRINSICS
bool vectorsAccelerated = true;
int byteVectorLength = 16;
bool?Sse2AndBelow = true;
bool?Sse3Group = null;
bool?AesLzPcl = null;
bool?Sse4142 = null;
bool?PopCnt = null;
bool?Avx12 = false;
bool?FmaBmi12 = false;
bool?Avxvnni = false;
#elif NON_VEX_INTRINSICS
bool vectorsAccelerated = true;
int byteVectorLength = 16;
bool?Sse2AndBelow = true;
bool?Sse3Group = true;
bool?AesLzPcl = null;
bool?Sse4142 = true;
bool?PopCnt = null;
bool?Avx12 = false;
bool?FmaBmi12 = false;
bool?Avxvnni = false;
#elif VEX_INTRINSICS
bool vectorsAccelerated = true;
int byteVectorLength = 32;
bool?Sse2AndBelow = true;
bool?Sse3Group = true;
bool?AesLzPcl = null;
bool?Sse4142 = true;
bool?PopCnt = null;
bool?Avx12 = true;
bool?FmaBmi12 = null;
bool?Avxvnni = null;
#else
#error Who dis?
#endif
if (vectorsAccelerated != Vector.IsHardwareAccelerated)
{
throw new Exception($"Vectors HW acceleration state unexpected - expected {vectorsAccelerated}, got {Vector.IsHardwareAccelerated}");
}
if (byteVectorLength != Vector <byte> .Count)
{
throw new Exception($"Unexpected vector length - expected {byteVectorLength}, got {Vector<byte>.Count}");
}
Check("Sse", Sse2AndBelow, &SseIsSupported, Sse.IsSupported, () => Sse.Subtract(Vector128 <float> .Zero, Vector128 <float> .Zero).Equals(Vector128 <float> .Zero));
Check("Sse.X64", Sse2AndBelow, &SseX64IsSupported, Sse.X64.IsSupported, () => Sse.X64.ConvertToInt64WithTruncation(Vector128 <float> .Zero) == 0);
Check("Sse2", Sse2AndBelow, &Sse2IsSupported, Sse2.IsSupported, () => Sse2.Extract(Vector128 <ushort> .Zero, 0) == 0);
Check("Sse2.X64", Sse2AndBelow, &Sse2X64IsSupported, Sse2.X64.IsSupported, () => Sse2.X64.ConvertToInt64(Vector128 <double> .Zero) == 0);
Check("Sse3", Sse3Group, &Sse3IsSupported, Sse3.IsSupported, () => Sse3.MoveHighAndDuplicate(Vector128 <float> .Zero).Equals(Vector128 <float> .Zero));
Check("Sse3.X64", Sse3Group, &Sse3X64IsSupported, Sse3.X64.IsSupported, null);
Check("Ssse3", Sse3Group, &Ssse3IsSupported, Ssse3.IsSupported, () => Ssse3.Abs(Vector128 <short> .Zero).Equals(Vector128 <ushort> .Zero));
Check("Ssse3.X64", Sse3Group, &Ssse3X64IsSupported, Ssse3.X64.IsSupported, null);
Check("Sse41", Sse4142, &Sse41IsSupported, Sse41.IsSupported, () => Sse41.Max(Vector128 <int> .Zero, Vector128 <int> .Zero).Equals(Vector128 <int> .Zero));
Check("Sse41.X64", Sse4142, &Sse41X64IsSupported, Sse41.X64.IsSupported, () => Sse41.X64.Extract(Vector128 <long> .Zero, 0) == 0);
Check("Sse42", Sse4142, &Sse42IsSupported, Sse42.IsSupported, () => Sse42.Crc32(0, 0) == 0);
Check("Sse42.X64", Sse4142, &Sse42X64IsSupported, Sse42.X64.IsSupported, () => Sse42.X64.Crc32(0, 0) == 0);
Check("Aes", AesLzPcl, &AesIsSupported, Aes.IsSupported, () => Aes.KeygenAssist(Vector128 <byte> .Zero, 0).Equals(Vector128.Create((byte)99)));
Check("Aes.X64", AesLzPcl, &AesX64IsSupported, Aes.X64.IsSupported, null);
Check("Avx", Avx12, &AvxIsSupported, Avx.IsSupported, () => Avx.Add(Vector256 <double> .Zero, Vector256 <double> .Zero).Equals(Vector256 <double> .Zero));
Check("Avx.X64", Avx12, &AvxX64IsSupported, Avx.X64.IsSupported, null);
Check("Avx2", Avx12, &Avx2IsSupported, Avx2.IsSupported, () => Avx2.Abs(Vector256 <int> .Zero).Equals(Vector256 <uint> .Zero));
Check("Avx2.X64", Avx12, &Avx2X64IsSupported, Avx2.X64.IsSupported, null);
Check("Bmi1", FmaBmi12, &Bmi1IsSupported, Bmi1.IsSupported, () => Bmi1.AndNot(0, 0) == 0);
Check("Bmi1.X64", FmaBmi12, &Bmi1X64IsSupported, Bmi1.X64.IsSupported, () => Bmi1.X64.AndNot(0, 0) == 0);
Check("Bmi2", FmaBmi12, &Bmi2IsSupported, Bmi2.IsSupported, () => Bmi2.MultiplyNoFlags(0, 0) == 0);
Check("Bmi2.X64", FmaBmi12, &Bmi2X64IsSupported, Bmi2.X64.IsSupported, () => Bmi2.X64.MultiplyNoFlags(0, 0) == 0);
Check("Fma", FmaBmi12, &FmaIsSupported, Fma.IsSupported, () => Fma.MultiplyAdd(Vector128 <float> .Zero, Vector128 <float> .Zero, Vector128 <float> .Zero).Equals(Vector128 <float> .Zero));
Check("Fma.X64", FmaBmi12, &FmaX64IsSupported, Fma.X64.IsSupported, null);
Check("Lzcnt", AesLzPcl, &LzcntIsSupported, Lzcnt.IsSupported, () => Lzcnt.LeadingZeroCount(0) == 32);
Check("Lzcnt.X64", AesLzPcl, &LzcntX64IsSupported, Lzcnt.X64.IsSupported, () => Lzcnt.X64.LeadingZeroCount(0) == 64);
Check("Pclmulqdq", AesLzPcl, &PclmulqdqIsSupported, Pclmulqdq.IsSupported, () => Pclmulqdq.CarrylessMultiply(Vector128 <long> .Zero, Vector128 <long> .Zero, 0).Equals(Vector128 <long> .Zero));
Check("Pclmulqdq.X64", AesLzPcl, &PclmulqdqX64IsSupported, Pclmulqdq.X64.IsSupported, null);
Check("Popcnt", PopCnt, &PopcntIsSupported, Popcnt.IsSupported, () => Popcnt.PopCount(0) == 0);
Check("Popcnt.X64", PopCnt, &PopcntX64IsSupported, Popcnt.X64.IsSupported, () => Popcnt.X64.PopCount(0) == 0);
Check("AvxVnni", Avxvnni, &AvxVnniIsSupported, AvxVnni.IsSupported, () => AvxVnni.MultiplyWideningAndAdd(Vector128 <int> .Zero, Vector128 <byte> .Zero, Vector128 <sbyte> .Zero).Equals(Vector128 <int> .Zero));
Check("AvxVnni.X64", Avxvnni, &AvxVnniX64IsSupported, AvxVnni.X64.IsSupported, null);
return(s_success ? 100 : 1);
}
public static Vector128 <float> _mm_movehdup_ps(Vector128 <float> source)
{
return(Sse3.MoveHighAndDuplicate(source));
}
public static __m128 _mm_movehdup_ps(__m128 a) => Sse3.MoveHighAndDuplicate(a);
unsafe void IConvolver.ConvolveSourceLine(byte *istart, byte *tstart, int cb, byte *mapxstart, int smapx, int smapy)
{
float *tp = (float *)tstart, tpe = (float *)(tstart + cb);
float *pmapx = (float *)mapxstart;
int kstride = smapx * channels;
int tstride = smapy * 4;
int vcnt = smapx / Vector128 <float> .Count;
while (tp < tpe)
{
int ix = *(int *)pmapx++;
int lcnt = vcnt;
float *ip = (float *)istart + ix * channels;
float *mp = pmapx;
pmapx += kstride;
Vector128 <float> av0, av1, av2;
if (Avx.IsSupported && lcnt >= 2)
{
Vector256 <float> ax0 = Vector256 <float> .Zero, ax1 = ax0, ax2 = ax0;
for (; lcnt >= 2; lcnt -= 2)
{
var iv0 = Avx.LoadVector256(ip);
var iv1 = Avx.LoadVector256(ip + Vector256 <float> .Count);
var iv2 = Avx.LoadVector256(ip + Vector256 <float> .Count * 2);
ip += Vector256 <int> .Count * channels;
if (Fma.IsSupported)
{
ax0 = Fma.MultiplyAdd(Avx.LoadVector256(mp), iv0, ax0);
ax1 = Fma.MultiplyAdd(Avx.LoadVector256(mp + Vector256 <float> .Count), iv1, ax1);
ax2 = Fma.MultiplyAdd(Avx.LoadVector256(mp + Vector256 <float> .Count * 2), iv2, ax2);
}
else
{
ax0 = Avx.Add(ax0, Avx.Multiply(iv0, Avx.LoadVector256(mp)));
ax1 = Avx.Add(ax1, Avx.Multiply(iv1, Avx.LoadVector256(mp + Vector256 <float> .Count)));
ax2 = Avx.Add(ax2, Avx.Multiply(iv2, Avx.LoadVector256(mp + Vector256 <float> .Count * 2)));
}
mp += Vector256 <float> .Count * channels;
}
av0 = Sse.Add(ax0.GetLower(), ax1.GetUpper());
av1 = Sse.Add(ax0.GetUpper(), ax2.GetLower());
av2 = Sse.Add(ax1.GetLower(), ax2.GetUpper());
}
else
{
av0 = av1 = av2 = Vector128 <float> .Zero;
}
for (; lcnt != 0; lcnt--)
{
var iv0 = Sse.LoadVector128(ip);
var iv1 = Sse.LoadVector128(ip + Vector128 <float> .Count);
var iv2 = Sse.LoadVector128(ip + Vector128 <float> .Count * 2);
ip += Vector128 <float> .Count * channels;
if (Fma.IsSupported)
{
av0 = Fma.MultiplyAdd(Sse.LoadVector128(mp), iv0, av0);
av1 = Fma.MultiplyAdd(Sse.LoadVector128(mp + Vector128 <float> .Count), iv1, av1);
av2 = Fma.MultiplyAdd(Sse.LoadVector128(mp + Vector128 <float> .Count * 2), iv2, av2);
}
else
{
av0 = Sse.Add(av0, Sse.Multiply(iv0, Sse.LoadVector128(mp)));
av1 = Sse.Add(av1, Sse.Multiply(iv1, Sse.LoadVector128(mp + Vector128 <float> .Count)));
av2 = Sse.Add(av2, Sse.Multiply(iv2, Sse.LoadVector128(mp + Vector128 <float> .Count * 2)));
}
mp += Vector128 <float> .Count * channels;
}
var avs0 = Sse.Add(Sse.Add(
Sse.Shuffle(av0, av0, 0b_00_10_01_11),
Sse.Shuffle(av1, av1, 0b_00_01_11_10)),
Sse.Shuffle(av2, av2, 0b_00_11_10_01)
);
var avs1 = Sse3.IsSupported ?
Sse3.MoveHighAndDuplicate(avs0) :
Sse.Shuffle(avs0, avs0, 0b_11_11_01_01);
var avs2 = Sse.UnpackHigh(avs0, avs0);
tp[0] = Sse.AddScalar(av0, avs0).ToScalar();
tp[1] = Sse.AddScalar(av1, avs1).ToScalar();
tp[2] = Sse.AddScalar(av2, avs2).ToScalar();
tp += tstride;
}
}
