public unsafe void CouldCallHiLevelSgemm()
{
var mnk = 2;
var x = new float[mnk * mnk];
for (int i = 0; i < mnk * mnk; i++)
{
x[i] = (i + 1);
}
var h = x.AsMemory().Pin();
var c = new float[mnk * mnk];
var hc = c.AsMemory().Pin();
CBLAS.Sgemm(LAYOUT.RowMajor, TRANSPOSE.NoTrans, TRANSPOSE.NoTrans,
mnk, mnk, mnk, alpha: 1f, (float *)h.Pointer, mnk, (float *)h.Pointer, mnk, beta: 0, (float *)hc.Pointer, mnk);
}
public static unsafe void Sample()
{
BlasNumThreads = 1;
var mnk = 2;
var x = new float[mnk * mnk];
var c = new float[mnk * mnk];
for (int i = 0; i < mnk * mnk; i++)
{
x[i] = (i + 1);
}
fixed(float *xP = &x[0])
fixed(float *cP = &c[0])
{
CBLAS.Sgemm(LAYOUT.RowMajor, TRANSPOSE.NoTrans, TRANSPOSE.NoTrans,
mnk, mnk, mnk, alpha: 1f, xP, mnk, xP, mnk, beta: 0, cP, mnk);
}
}
public static MemoryAlignedMatrix Power(this MemoryAlignedMatrix a, int n)
{
MemoryAlignedMatrix z = null;
MemoryAlignedMatrix result = null;
MemoryAlignedMatrix output = new MemoryAlignedMatrix(a.Size);
while (n > 0)
{
if (z == null)
{
z = a.Copy();
}
else
{
CBLAS.Square(z, output);
Swap(ref z, ref output);
}
int bit = n % 2;
n /= 2;
if (bit == 1)
{
if (result == null)
{
result = z.Copy();
}
else
{
CBLAS.Multiply(result, z, output);
Swap(ref result, ref output);
}
}
}
return(result);
}
public void Test_cblasSgemm()
{
/* Data initialization */
int Order = CBLAS.ORDER.ColMajor;
int TransA = CBLAS.TRANSPOSE.NoTrans;
int TransB = CBLAS.TRANSPOSE.NoTrans;
Random rand = new Random();
int rows_a = rand.Next(2, 10);
int cols_a = rand.Next(2, 10);
int rows_b = cols_a;
int cols_b = rand.Next(2, 10);
int rows_c = rows_a;
int cols_c = cols_b;
float[] A = new float[rows_a * cols_a];
float[] B = new float[rows_b * cols_b];
float[] C = new float[rows_c * cols_c];
for (int i = 0; i < A.Length; i++)
{
A[i] = Convert.ToSingle(rand.Next(0, 10));
}
for (int i = 0; i < B.Length; i++)
{
B[i] = Convert.ToSingle(rand.Next(0, 10));
}
for (int i = 0; i < C.Length; i++)
{
C[i] = Convert.ToSingle(rand.Next(0, 10));
}
var alpha = Convert.ToSingle(rand.Next(0, 10));
var beta = Convert.ToSingle(rand.Next(0, 10));
var mA = Matrix <float> .Build.Dense(rows_a, cols_a, A);
var mB = Matrix <float> .Build.Dense(rows_b, cols_b, B);
var mC = Matrix <float> .Build.Dense(rows_c, cols_c, C).Clone();
Console.WriteLine("MKL cblas_sgemm example");
Console.WriteLine();
/* Print initial data */
Console.WriteLine("alpha=" + alpha);
Console.WriteLine("beta=" + beta);
Console.WriteLine("Matrix A");
Console.WriteLine(mA.ToString());
Console.WriteLine("Matrix B");
Console.WriteLine(mB.ToString());
Console.WriteLine("Initial C");
Console.WriteLine(mC.ToString());
/* Computation */
CBLAS.sgemm(Order, TransA, TransB, rows_a, cols_b, cols_a, alpha, A, rows_a, B, rows_b, beta, ref C, rows_c);
mC = alpha * mA * mB + beta * mC; //sgemm
var expected = mC.ToColumnWiseArray(); //reference result by Math.Net
for (int i = 0; i < C.Length; i++)
{
Assert.AreEqual(expected[i], C[i]);
}
/* Print the result */
var mCResult = Matrix <float> .Build.Dense(rows_c, cols_c, C);
Console.WriteLine("expected");
Console.WriteLine(mC.ToString());
Console.WriteLine("Result");
Console.WriteLine(mCResult.ToString());
Console.WriteLine("TEST PASSED");
Console.WriteLine();
}
public void CouldUseSMFormulaCorrectly()
{
BLAS.BlasNumThreads = 1;
// Calculate (X'X)^-1 for M and M-1
var dataSpan = new Span <double>(Data, Columns * Rows);
var rounds = 3;
var count = 20;
var rows = new[] { 2, 4, 6, 8, 10, 15, 20, 30, 40, 50, 60, 70, 80, 90, 100, 200, 300, 400, 500 };
var columns = new[] { 2, 4, 6, 8, 10, 15, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 200, 300, 400, 500 };
var syrkVsGemm = new bool[rows.Length, columns.Length];
Console.Write($" ");
foreach (var co in columns)
{
Console.Write($"{co} ");
}
Console.WriteLine();
foreach (var ro in rows)
{
Console.Write($"{ro}: ");
foreach (var co in columns)
{
Rows = ro;
Columns = co;
Marshal.FreeHGlobal((IntPtr)Data);
SetUp();
var xxN = (double *)Marshal.AllocHGlobal(Columns * Columns * 8);
var xxN1 = (double *)Marshal.AllocHGlobal(Columns * Columns * 8);
var xxNSpan = new Span <double>(xxN, Columns * Columns);
var xxN1Span = new Span <double>(xxN1, Columns * Columns);
var flopsMult = 1L * Columns * Columns * Rows * 2;
for (int r = 0; r < rounds; r++)
{
var statGemm = Benchmark.Run($"Dgemm_{ro}x{co}", count * flopsMult, silent: true);
{
for (int i = 0; i < count; i++)
{
var data = Data;
CBLAS.Dgemm(LAYOUT.RowMajor, TRANSPOSE.Trans, TRANSPOSE.NoTrans, Columns, Columns, Rows, alpha: 1.0,
data, Columns,
data, Columns,
beta: 0.0,
xxN, Columns);
}
}
statGemm.Dispose();
var starSyrk = Benchmark.Run($"Dsyrk_{ro}x{co}", count * flopsMult, silent: true);
{
for (int i = 0; i < count; i++)
{
var data = Data;
CBLAS.Dsyrk(LAYOUT.RowMajor, UPLO.Upper, TRANSPOSE.Trans, Columns, Rows, 1.0,
data, Columns, 0.0, xxN, Columns);
}
}
starSyrk.Dispose();
if (r == rounds - 1)
{
// Console.WriteLine($"starSyrk.MOPS {starSyrk.MOPS} - starGemm.MOPS {statGemm.MOPS}");
if (starSyrk.MOPS > statGemm.MOPS)
{
// syrkVsGemm[ro, co] = true;
Console.Write(1 + " ");
}
else
{
Console.Write(0 + " ");
}
}
}
}
Console.WriteLine();
}
// Benchmark.Dump();
// CBLAS.Dsyrk(LAYOUT.RowMajor, UPLO.Lower, TRANSPOSE.Trans, Columns, Rows - 1, 1.0, Data, Rows, 0.0, xxN1, Columns);
}
private void Form1_Load(object sender, EventArgs e)
{
CBLAS.detectCBLAS();
prependLine("OpenBLAS: " + (CBLAS.openbAvailable ? "" : "not ") + "available");
prependLine("Intel MKL: " + (CBLAS.imklAvailable ? "" : "not ") + "available");
}
