public void OptimizeTest()
{
float max_err = 0;
foreach (int batch in new int[] { 1, 2 })
{
foreach (int inchannels in new int[] { 1, 2, 3, 4, 5, 10, 15, 20 })
{
foreach (int outchannels in new int[] { 7, 13 })
{
float[] yval = (new float[outchannels * batch]).Select((_, idx) => idx * 1e-3f).ToArray();
float[] wval = (new float[inchannels * outchannels]).Select((_, idx) => idx * 1e-3f).Reverse().ToArray();
Map0D y = new Map0D(outchannels, batch, yval);
Filter0D w = new Filter0D(inchannels, outchannels, 1, wval);
Map0D x = Reference(y, w);
Map0D x_optimized = OptimizedReference(y, w);
float[] x_expect = x.ToArray();
float[] x_actual = x_optimized.ToArray();
AssertError.Tolerance(x_expect, x_actual, 1e-7f, 1e-5f, ref max_err, $"mismatch value {inchannels},{outchannels},{batch}");
Console.WriteLine($"pass: {inchannels},{outchannels},{batch}");
}
}
}
Console.WriteLine($"maxerr:{max_err}");
}
public void ReferenceTest()
{
int inchannels = 7, outchannels = 11, batch = 2;
float[] xval = (new float[inchannels * batch]).Select((_, idx) => idx * 1e-3f).ToArray();
float[] wval = (new float[outchannels * inchannels]).Select((_, idx) => idx * 1e-3f).Reverse().ToArray();
Map0D x = new Map0D(inchannels, batch, xval);
Filter0D w = new Filter0D(inchannels, outchannels, 1, wval);
Map0D y = Reference(x, w);
float[] y_expect =
{
1.5050e-03f, 1.3580e-03f, 1.2110e-03f, 1.0640e-03f, 9.1700e-04f,
7.7000e-04f, 6.2300e-04f, 4.7600e-04f, 3.2900e-04f, 1.8200e-04f,
3.5000e-05f, 5.0820e-03f, 4.5920e-03f, 4.1020e-03f, 3.6120e-03f,
3.1220e-03f, 2.6320e-03f, 2.1420e-03f, 1.6520e-03f, 1.1620e-03f,
6.7200e-04f, 1.8200e-04f
};
float[] y_actual = y.ToArray();
AssertError.Tolerance(y_expect, y_actual, 1e-7f, 1e-5f, $"mismatch value {inchannels},{outchannels},{batch}");
}
public static Map0D OptimizedReference(Map0D y, Filter0D w)
{
int outchannels = y.Channels, inchannels = w.InChannels, batch = y.Batch;
Map0D x = new Map0D(inchannels, batch);
for (int th = 0; th < batch; th++)
{
for (int inch = 0; inch < inchannels; inch++)
{
double sum = 0;
int inmap_idx = outchannels * th;
int outmap_idx = inch + inchannels * th;
int kernel_idx = inch;
for (int outch = 0; outch < outchannels; outch++)
{
sum += y[inmap_idx] * w[kernel_idx];
inmap_idx++;
kernel_idx += inchannels;
}
x[outmap_idx] = sum;
}
}
return(x);
}
public static Map0D OptimizedReference(Map0D x, Filter0D w)
{
int inchannels = x.Channels, outchannels = w.OutChannels, batch = x.Batch;
Map0D y = new Map0D(outchannels, batch);
for (int th = 0; th < batch; th++)
{
int inmap_org = th * inchannels;
int outmap_idx = th * outchannels;
int kernel_idx = 0;
for (int outch = 0; outch < outchannels; outch++)
{
double sum = 0;
int inmap_idx = inmap_org;
for (int inch = 0; inch < inchannels; inch++)
{
sum += x[inmap_idx] * w[kernel_idx];
inmap_idx++;
kernel_idx++;
}
y[outmap_idx] = sum;
outmap_idx++;
}
}
return(y);
}
public void ExecuteTest()
{
float max_err = 0;
foreach (int batch in new int[] { 1, 2 })
{
foreach (int inchannels in new int[] { 1, 2, 3, 4, 5, 10, 15, 20 })
{
foreach (int outchannels in new int[] { 7, 13 })
{
float[] yval = (new float[outchannels * batch]).Select((_, idx) => idx * 1e-3f).ToArray();
float[] wval = (new float[inchannels * outchannels]).Select((_, idx) => idx * 1e-3f).Reverse().ToArray();
Map0D y = new Map0D(outchannels, batch, yval);
Filter0D w = new Filter0D(inchannels, outchannels, 1, wval);
Map0D x = Reference(y, w);
OverflowCheckedTensor y_tensor = new OverflowCheckedTensor(Shape.Map0D(outchannels, batch), yval);
OverflowCheckedTensor w_tensor = new OverflowCheckedTensor(Shape.Kernel0D(inchannels, outchannels), wval);
OverflowCheckedTensor x_tensor = new OverflowCheckedTensor(Shape.Map0D(inchannels, batch));
TransposeDense ope = new TransposeDense(outchannels, inchannels, batch);
ope.Execute(y_tensor, w_tensor, x_tensor);
float[] x_expect = x.ToArray();
float[] x_actual = x_tensor.State;
CollectionAssert.AreEqual(yval, y_tensor.State);
CollectionAssert.AreEqual(wval, w_tensor.State);
AssertError.Tolerance(x_expect, x_actual, 1e-7f, 1e-5f, ref max_err, $"mismatch value {inchannels},{outchannels},{batch}");
Console.WriteLine($"pass: {inchannels},{outchannels},{batch}");
}
}
}
Console.WriteLine($"maxerr:{max_err}");
}
public void ExecuteTest()
{
float max_err = 0;
foreach (int batch in new int[] { 1, 2 })
{
foreach (int inchannels in new int[] { 1, 2, 3, 4, 5, 10, 15, 20 })
{
foreach (int outchannels in new int[] { 7, 13 })
{
float[] xval = (new float[inchannels * batch]).Select((_, idx) => idx * 1e-3f).ToArray();
float[] yval = (new float[outchannels * batch]).Select((_, idx) => idx * 1e-3f).Reverse().ToArray();
Map0D x = new Map0D(inchannels, batch, xval);
Map0D y = new Map0D(outchannels, batch, yval);
Filter0D gw = Reference(x, y);
OverflowCheckedTensor x_tensor = new OverflowCheckedTensor(Shape.Map0D(inchannels, batch), xval);
OverflowCheckedTensor y_tensor = new OverflowCheckedTensor(Shape.Map0D(outchannels, batch), yval);
OverflowCheckedTensor gw_tensor = new OverflowCheckedTensor(Shape.Kernel0D(inchannels, outchannels));
KernelProduct ope = new KernelProduct(inchannels, outchannels, batch);
ope.Execute(x_tensor, y_tensor, gw_tensor);
float[] gw_expect = gw.ToArray();
float[] gw_actual = gw_tensor.State;
CollectionAssert.AreEqual(xval, x_tensor.State);
CollectionAssert.AreEqual(yval, y_tensor.State);
AssertError.Tolerance(gw_expect, gw_actual, 1e-7f, 1e-5f, ref max_err, $"mismatch value {inchannels},{outchannels},{batch}");
Console.WriteLine($"pass: {inchannels},{outchannels},{batch}");
}
}
}
Console.WriteLine($"maxerr:{max_err}");
}
public static Map0D Reference(Map0D y, Filter0D w)
{
int outchannels = y.Channels, inchannels = w.InChannels, batch = y.Batch;
Map0D x = new Map0D(inchannels, batch);
for (int th = 0; th < batch; th++)
{
for (int inch = 0; inch < inchannels; inch++)
{
double sum = 0;
for (int outch = 0; outch < outchannels; outch++)
{
sum += y[outch, th] * w[inch, outch, 0];
}
x[inch, th] = sum;
}
}
return(x);
}
public void ReferenceTest()
{
int inchannels = 7, outchannels = 11, batch = 2;
float[] yval = (new float[outchannels * batch]).Select((_, idx) => idx * 1e-3f).ToArray();
float[] wval = (new float[outchannels * inchannels]).Select((_, idx) => idx * 1e-3f).Reverse().ToArray();
Map0D y = new Map0D(outchannels, batch, yval);
Filter0D w = new Filter0D(inchannels, outchannels, 1, wval);
Map0D x = Reference(y, w);
float[] x_expect =
{
1.4850e-03f, 1.4300e-03f, 1.3750e-03f, 1.3200e-03f, 1.2650e-03f,
1.2100e-03f, 1.1550e-03f, 6.4460e-03f, 6.2700e-03f, 6.0940e-03f,
5.9180e-03f, 5.7420e-03f, 5.5660e-03f, 5.3900e-03f
};
float[] x_actual = x.ToArray();
AssertError.Tolerance(x_expect, x_actual, 1e-7f, 1e-5f, $"mismatch value {inchannels},{outchannels},{batch}");
}
public static Filter0D Reference(Map0D x, Map0D y)
{
int inchannels = x.Channels, outchannels = y.Channels, batch = x.Batch;
Filter0D w = new Filter0D(inchannels, outchannels, 1);
for (int inch, outch = 0; outch < outchannels; outch++)
{
for (inch = 0; inch < inchannels; inch++)
{
double sum = 0;
for (int th = 0; th < batch; th++)
{
sum += x[inch, th] * y[outch, th];
}
w[inch, outch, 0] = sum;
}
}
return(w);
}
public static Map0D Reference(Map0D x, Filter0D w)
{
int inchannels = x.Channels, outchannels = w.OutChannels, batch = x.Batch;
Map0D y = new Map0D(outchannels, batch);
for (int th = 0; th < batch; th++)
{
for (int outch = 0; outch < outchannels; outch++)
{
double sum = 0;
for (int inch = 0; inch < inchannels; inch++)
{
sum += x[inch, th] * w[inch, outch, 0];
}
y[outch, th] = sum;
}
}
return(y);
}
public void ReferenceTest()
{
int inchannels = 7, outchannels = 11, batch = 2;
float[] xval = (new float[batch * inchannels]).Select((_, idx) => idx * 1e-3f).ToArray();
float[] yval = (new float[batch * outchannels]).Select((_, idx) => idx * 1e-3f).Reverse().ToArray();
Map0D x = new Map0D(inchannels, batch, xval);
Map0D y = new Map0D(outchannels, batch, yval);
Filter0D gw = Reference(x, y);
float[] gw_expect =
{
7.0000e-05f, 1.0100e-04f, 1.3200e-04f, 1.6300e-04f, 1.9400e-04f,
2.2500e-04f, 2.5600e-04f, 6.3000e-05f, 9.2000e-05f, 1.2100e-04f,
1.5000e-04f, 1.7900e-04f, 2.0800e-04f, 2.3700e-04f, 5.6000e-05f,
8.3000e-05f, 1.1000e-04f, 1.3700e-04f, 1.6400e-04f, 1.9100e-04f,
2.1800e-04f, 4.9000e-05f, 7.4000e-05f, 9.9000e-05f, 1.2400e-04f,
1.4900e-04f, 1.7400e-04f, 1.9900e-04f, 4.2000e-05f, 6.5000e-05f,
8.8000e-05f, 1.1100e-04f, 1.3400e-04f, 1.5700e-04f, 1.8000e-04f,
3.5000e-05f, 5.6000e-05f, 7.7000e-05f, 9.8000e-05f, 1.1900e-04f,
1.4000e-04f, 1.6100e-04f, 2.8000e-05f, 4.7000e-05f, 6.6000e-05f,
8.5000e-05f, 1.0400e-04f, 1.2300e-04f, 1.4200e-04f, 2.1000e-05f,
3.8000e-05f, 5.5000e-05f, 7.2000e-05f, 8.9000e-05f, 1.0600e-04f,
1.2300e-04f, 1.4000e-05f, 2.9000e-05f, 4.4000e-05f, 5.9000e-05f,
7.4000e-05f, 8.9000e-05f, 1.0400e-04f, 7.0000e-06f, 2.0000e-05f,
3.3000e-05f, 4.6000e-05f, 5.9000e-05f, 7.2000e-05f, 8.5000e-05f,
0.0000e+00f, 1.1000e-05f, 2.2000e-05f, 3.3000e-05f, 4.4000e-05f,
5.5000e-05f, 6.6000e-05f
};
float[] gw_actual = gw.ToArray();
AssertError.Tolerance(gw_expect, gw_actual, 1e-7f, 1e-5f, $"mismatch value {inchannels},{outchannels},{batch}");
}
