public static Vector128 <int>[,] Procedure(int[,] A, int[,] B, int a, Vector128 <int> x, Vector128 <int> y)
{
unsafe
{
int length = A.GetLength(0);
int length2 = B.GetLength(1);
int length3 = 0;
int length4 = 0;
try
{
length3 = length;
length4 = length2 / 4;
}
catch (Exception e)
{
Console.WriteLine("Error");
}
Vector128 <int>[,] MatrixA = new Vector128 <int> [length3, length4];
Vector128 <int>[,] MatrixB = new Vector128 <int> [length3, length4];
Vector128 <int>[,] Matrix = new Vector128 <int> [length3, length4];
int c = 0;
fixed(int *ptr = A)
{
fixed(int *ptr2 = B)
{
for (int i = 0; i < length3; i++)
{
for (int j = 0; j < length4; j++)
{
var v = Sse41.LoadVector128(ptr + c);
var v2 = Sse41.LoadVector128(ptr2 + c);
c += 4;
MatrixA[i, j] = Sse41.MultiplyLow(v, x);
MatrixA[i, j] = Sse41.MultiplyLow(MatrixA[i, j], Vector128.Create(a));
MatrixB[i, j] = Sse41.MultiplyLow(v2, y);
Matrix[i, j] = Sse41.Add(MatrixA[i, j], MatrixB[i, j]);
}
}
}
}
return(Matrix);
}
}
public static Vector128 <int>[,] MultiplyMatrixSSE(int[,] a, int[,] b)
{
unsafe
{
int length = a.GetLength(0);
int length2 = a.GetLength(1);
int length3 = 0;
int length4 = 0;
try
{
length3 = length;
length4 = length2 / 4;
}
catch (Exception e)
{
Console.WriteLine("Error");
}
Vector128 <int>[,] A = new Vector128 <int> [length3, length4];
int c = 0;
fixed(int *ptr = a)
{
fixed(int *ptr2 = b)
{
for (int i = 0; i < length3; i++)
{
for (int j = 0; j < length4; j++)
{
var v = Sse41.LoadVector128(ptr + c);
var v2 = Sse41.LoadVector128(ptr2 + c);
c += 4;
A[i, j] = Sse41.MultiplyLow(v, v2);
}
}
}
}
return(A);
}
}
private static unsafe void FillRow3(ushort *previousRow, int length)
{
ushort *bytes = stackalloc ushort[] { 1, 2, 3, 4, 5, 6, 7, 8 };
var counter = Sse41.LoadVector128(bytes);
var step = Vector128.Create((ushort)8);
int i = 0;
for (; i < (length & 7); i += 8)
{
Sse42.Store(previousRow + i, counter);
counter = Sse42.AddSaturate(counter, step);
}
//step = Sse42.ShiftRightArithmetic
step = Vector128.Create((ushort)1);
for (; i < length; ++i)
{
previousRow[i] = counter.GetElement(0);
counter = Sse42.AddSaturate(counter, step);
}
}
private static unsafe void FillRowSSeVector2(int *previousRow, int length)
{
const int LENGHT = 4;
int i = 0;
//int initialCount = Math.Min(count, previousRow.Length);
for (i = 0; i < 2 * LENGHT;)
{
previousRow[i] = ++i;
}
var counter1 = Sse41.LoadVector128(previousRow);
var counter2 = Sse41.LoadVector128(previousRow + LENGHT);
var step = Vector128.Create(i);
int *pDest = previousRow + i;
for (; i < (length - (2 * LENGHT - 1)); i += 2 * LENGHT)
{
counter1 = Sse42.Add(counter1, step);
counter2 = Sse42.Add(counter2, step);
Sse42.Store(pDest, counter1);
Sse42.Store(pDest + LENGHT, counter2);
pDest += 2 * LENGHT;
}
if (i < (length - (LENGHT - 1)))
{
counter1 = Sse42.Add(counter1, step);
Sse42.Store(previousRow + i, counter1);
i += LENGHT;
}
for (; i < length;)
{
previousRow[i] = ++i;
}
}
public unsafe int FastMax(Span <int> span)
{
int max = span[0];
int i = 0;
int vectorSize = 4;
if (span.Length >= (vectorSize * 2))
{
fixed(int *ptr = &span[0])
{
var maxVector = Sse41.LoadVector128(ptr + i);
for (i = vectorSize; i < span.Length - vectorSize; i += vectorSize)
{
var current = Sse41.LoadVector128(ptr + i);
maxVector = Sse41.Max(maxVector, current);
}
Span <int> buffer = stackalloc int[vectorSize];
maxVector.AsVector().CopyTo(buffer);
for (int k = 0; k < buffer.Length; k++)
{
if (buffer[k] > max)
{
max = buffer[k];
}
}
}
}
for (; i < span.Length; i++)
{
if (span[i] > max)
{
max = span[i];
}
}
return(max);
}
