static void DoMatrix()
{
int size = true ? 512 : 16;
Console.WriteLine("Matrix size: " + size);
Console.WriteLine("Vector size: " + Vector <float> .Count);
int bumper = 0;
var a = DataGenerator.Matrix(size, size + bumper);
var am = new FloatMatrix2d(a);
var b = DataGenerator.Matrix(size + bumper, size);
var bm = new FloatMatrix2d(b);
var mm = new MatrixMultiply();
if (size <= 20)
{
//EnsureCorrectness(a, b, mm.Naive, mm.NaiveSumUnsafe);
EnsureCorrectness(a, b, mm.TransposeSum, am, bm, mm.TransposeSumStridingVectorMat2d);
}
TimeIt("MM - " + nameof(mm.Naive), () => mm.Naive(a, b));
TimeIt("MM - " + nameof(mm.NaiveMat2d), () => mm.NaiveMat2d(am, bm));
TimeIt("MM - " + nameof(mm.NaiveSum), () => mm.NaiveSum(a, b));
TimeIt("MM - " + nameof(mm.NaiveSumUnsafe), () => mm.NaiveSumUnsafe(a, b));
TimeIt("MM - " + nameof(mm.TransposeSum), () => mm.TransposeSum(a, b));
TimeIt("MM - " + nameof(mm.TransposeSumMat2d), () => mm.TransposeSumMat2d(am, bm));
TimeIt("MM - " + nameof(mm.TransposeSumVectorMat2d), () => mm.TransposeSumVectorMat2d(am, bm));
TimeIt("MM - " + nameof(mm.TransposeSumStridingVectorMat2d), () => mm.TransposeSumStridingVectorMat2d(am, bm));
TimeIt("MM - " + nameof(mm.TransposeSumUnsafe), () => mm.TransposeSumUnsafe(a, b));
TimeIt("MM - " + nameof(mm.StridingSum), () => mm.StridingSum(a, b));
TimeIt("MM - " + nameof(mm.StridingSumUnsafe), () => mm.StridingSumUnsafe(a, b));
TimeIt("MM - " + nameof(mm.StridingUnrolledUnsafe), () => mm.StridingUnrolledUnsafe(a, b));
}
public FloatMatrix2d TransposeSumVectorMat2d(FloatMatrix2d a, FloatMatrix2d b)
{
var bT = MatrixTransposeMat2d(b);
var x = a.rows;
var z = a.cols;
var y = b.cols;
FloatMatrix2d result = new FloatMatrix2d(x, y);
for (int i = 0; i < x; i++)
{
for (int j = 0; j < y; j++)
{
float sum = 0f;
for (int k = 0; k < z; k += Vector <float> .Count)
{
var am = a.Vector(i, k);
var bm = bT.Vector(j, k);
//sum += a[i, k] * bT[j, k];
sum += Vector.Dot <float>(am, bm);
}
result[i, j] = sum;
}
}
return(result);
}
static void EnsureCorrectness(
float[,] a, float[,] b, Func <float[, ], float[, ], float[, ]> baseline,
FloatMatrix2d am, FloatMatrix2d bm, Func <FloatMatrix2d, FloatMatrix2d, FloatMatrix2d> hypothesis)
{
var bl = baseline(a, b);
var h = hypothesis(am, bm);
for (int i = 0; i < a.GetLength(0); i++)
{
for (int j = 0; j < a.GetLength(1); j++)
{
Console.WriteLine(i + ", " + j + " --- " + bl[i, j] + " --- " + h[i, j]);
}
}
}
public FloatMatrix2d MatrixTransposeMat2d(FloatMatrix2d m)
{
var x = m.rows;
var y = m.cols;
var result = new FloatMatrix2d(x, y);
for (int i = 0; i < x; i++)
{
for (int j = 0; j < y; j++)
{
result[j, i] = m[i, j];
}
}
return(result);
}
public FloatMatrix2d TransposeSumStridingVectorMat2d(FloatMatrix2d a, FloatMatrix2d b)
{
const int stride = 64 / sizeof(float);
var bT = MatrixTransposeMat2d(b);
var x = a.rows;
var z = a.cols;
var y = b.cols;
FloatMatrix2d result = new FloatMatrix2d(x, y);
for (int i = 0; i < x; i += stride)
{
for (int j = 0; j < y; j += stride)
{
for (int k = 0; k < z; k += stride)
{
for (int i2 = i; i2 < i + stride; i2++)
{
for (int j2 = j; j2 < j + stride; j2++)
{
float sum = 0f;
//for (int k2 = k; k2 < k+stride; k2++)
for (int k2 = k; k2 < k + stride; k2 += Vector <float> .Count)
{
try{
var am = a.Vector(i2, k2);
var bm = bT.Vector(j2, k2);
sum += Vector.Dot <float>(am, bm);
}catch (Exception)
{
Console.WriteLine("error: " + i2 + "," + k2 + "," + j2);
throw;
}
//sum += a[i2, k2] * bT[j2, k2];
}
result[i2, j2] = sum;
}
}
}
}
}
return(result);
}
public FloatMatrix2d NaiveMat2d(FloatMatrix2d a, FloatMatrix2d b)
{
var x = a.rows;
var z = a.cols;
var y = b.cols;
var result = new FloatMatrix2d(x, y);
for (int i = 0; i < x; i++)
{
for (int j = 0; j < y; j++)
{
for (int k = 0; k < z; k++)
{
result[i, j] += a[i, k] * b[k, j];
}
}
}
return(result);
}
public FloatMatrix2d TransposeSumMat2d(FloatMatrix2d a, FloatMatrix2d b)
{
var bT = MatrixTransposeMat2d(b);
var x = a.rows;
var z = a.cols;
var y = b.cols;
FloatMatrix2d result = new FloatMatrix2d(x, y);
for (int i = 0; i < x; i++)
{
for (int j = 0; j < y; j++)
{
float sum = 0f;
for (int k = 0; k < z; k++)
{
sum += a[i, k] * bT[j, k];
}
result[i, j] = sum;
}
}
return(result);
}
