public void SpeedTest()
{
int inwidth = 32, inchannels = 33, outchannels = 33, ksize = 3, stride = 2;
int outwidth = (inwidth - ksize) / stride + 1;
OverflowCheckedTensor x_tensor = new OverflowCheckedTensor(Shape.Map1D(inchannels, inwidth));
OverflowCheckedTensor y_tensor = new OverflowCheckedTensor(Shape.Map1D(outchannels, outwidth));
OverflowCheckedTensor w_tensor = new OverflowCheckedTensor(Shape.Kernel1D(inchannels / 3 * 4, outchannels / 3, ksize));
OverflowCheckedTensor gw_tensor = new OverflowCheckedTensor(Shape.Kernel1D(inchannels / 3 * 4, outchannels / 3, ksize));
TrivectorKernelProduct1D ope = new TrivectorKernelProduct1D(inwidth, inchannels, outchannels, ksize, stride);
Stopwatch sw = new Stopwatch();
sw.Start();
ope.Execute(x_tensor, y_tensor, w_tensor, gw_tensor);
ope.Execute(x_tensor, y_tensor, w_tensor, gw_tensor);
ope.Execute(x_tensor, y_tensor, w_tensor, gw_tensor);
ope.Execute(x_tensor, y_tensor, w_tensor, gw_tensor);
sw.Stop();
Console.WriteLine($"{sw.ElapsedMilliseconds / 4} msec");
}
public void OverflowTest()
{
foreach (bool transpose in new bool[] { false, true })
{
foreach (int batch in new int[] { 1, 2, 3 })
{
foreach (int inchannels in new int[] { 3, 6, 9, 12 })
{
foreach (int outchannels in new int[] { 3, 6, 9, 12 })
{
foreach (int kwidth in new int[] { 1, 3, 5 })
{
foreach (int stride in new int[] { 1, 2, 3 })
{
foreach (int inwidth in new int[] { 8, 9, 13, 17 })
{
int outwidth = (inwidth - kwidth) / stride + 1;
float[] xval = (new float[inwidth * inchannels * batch]).Select((_, idx) => idx * 1e-3f).ToArray();
float[] yval = (new float[outwidth * outchannels * batch]).Select((_, idx) => idx * 1e-3f).Reverse().ToArray();
float[] wval = (new float[kwidth * inchannels * outchannels / 9 * 4]).Select((_, idx) => idx * 1e-3f).Reverse().ToArray();
OverflowCheckedTensor x_tensor = new OverflowCheckedTensor(Shape.Map1D(inchannels, inwidth, batch), xval);
OverflowCheckedTensor y_tensor = new OverflowCheckedTensor(Shape.Map1D(outchannels, outwidth, batch), yval);
OverflowCheckedTensor w_tensor = new OverflowCheckedTensor(Shape.Kernel1D(inchannels / 3 * 4, outchannels / 3, kwidth), wval);
OverflowCheckedTensor gw_tensor = new OverflowCheckedTensor(Shape.Kernel1D(inchannels / 3 * 4, outchannels / 3, kwidth));
TrivectorKernelProduct1D ope = new TrivectorKernelProduct1D(inwidth, inchannels, outchannels, kwidth, stride, transpose, batch);
ope.Execute(x_tensor, y_tensor, w_tensor, gw_tensor);
CollectionAssert.AreEqual(xval, x_tensor.State);
CollectionAssert.AreEqual(yval, y_tensor.State);
CollectionAssert.AreEqual(wval, w_tensor.State);
gw_tensor.CheckOverflow();
Console.WriteLine($"pass: {inchannels},{outchannels},{kwidth},{stride},{inwidth},{batch},{transpose}");
}
}
}
}
}
}
}
}
public void ExecuteTest()
{
float max_err = 0;
foreach (int batch in new int[] { 1, 2, 3 })
{
foreach (int inchannels in new int[] { 3, 6, 9, 12 })
{
foreach (int outchannels in new int[] { 3, 6, 9, 12 })
{
foreach (int kwidth in new int[] { 1, 3, 5 })
{
foreach (int stride in new int[] { 1, 2, 3 })
{
foreach (int inwidth in new int[] { 8, 9, 13, 17 })
{
int outwidth = (inwidth - kwidth) / stride + 1;
float[] xval = (new float[inwidth * inchannels * batch]).Select((_, idx) => idx * 1e-3f).ToArray();
float[] yval = (new float[outwidth * outchannels * batch]).Select((_, idx) => idx * 1e-3f).Reverse().ToArray();
float[] wval = (new float[kwidth * inchannels * outchannels / 9 * 4]).Select((_, idx) => idx * 1e-3f).Reverse().ToArray();
Trivector[] xcval = (new Trivector[xval.Length / 3])
.Select((_, idx) => new Trivector(xval[idx * 3], xval[idx * 3 + 1], xval[idx * 3 + 2])).ToArray();
Trivector[] ycval = (new Trivector[yval.Length / 3])
.Select((_, idx) => new Trivector(yval[idx * 3], yval[idx * 3 + 1], yval[idx * 3 + 2])).ToArray();
Quaternion.Quaternion[] wcval = (new Quaternion.Quaternion[wval.Length / 4])
.Select((_, idx) => new Quaternion.Quaternion(wval[idx * 4], wval[idx * 4 + 1], wval[idx * 4 + 2], wval[idx * 4 + 3])).ToArray();
TrivectorMap1D x = new TrivectorMap1D(inchannels / 3, inwidth, batch, xcval);
TrivectorMap1D y = new TrivectorMap1D(outchannels / 3, outwidth, batch, ycval);
Quaternion.QuaternionFilter1D w = new Quaternion.QuaternionFilter1D(inchannels / 3, outchannels / 3, kwidth, wcval);
Quaternion.QuaternionFilter1D gw = Reference(x, y, w, kwidth, stride);
OverflowCheckedTensor x_tensor = new OverflowCheckedTensor(Shape.Map1D(inchannels, inwidth, batch), xval);
OverflowCheckedTensor y_tensor = new OverflowCheckedTensor(Shape.Map1D(outchannels, outwidth, batch), yval);
OverflowCheckedTensor w_tensor = new OverflowCheckedTensor(Shape.Kernel1D(inchannels / 3 * 4, outchannels / 3, kwidth), wval);
OverflowCheckedTensor gw_tensor = new OverflowCheckedTensor(Shape.Kernel1D(inchannels / 3 * 4, outchannels / 3, kwidth));
TrivectorKernelProduct1D ope = new TrivectorKernelProduct1D(inwidth, inchannels, outchannels, kwidth, stride, transpose: false, batch);
ope.Execute(x_tensor, y_tensor, w_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);
CollectionAssert.AreEqual(wval, w_tensor.State);
AssertError.Tolerance(gw_expect, gw_actual, 1e-7f, 1e-5f, ref max_err, $"mismatch value {inchannels},{outchannels},{kwidth},{stride},{inwidth},{batch}");
Console.WriteLine($"pass: {inchannels},{outchannels},{kwidth},{stride},{inwidth},{batch}");
}
}
}
}
}
}
Console.WriteLine($"maxerr:{max_err}");
}