public static void ReferenceThenCopyThenAssignGpu()
{
const int n = 1000;
var arrayA = GenerateRandomDoubleData(n, 1, 100);
var arrayB = new double[n];
var cpuA = arrayA.AsTensor();
var cpuB = arrayB.AsTensor();
var gpuA = gpu.Device.Allocate <double>(Shape.Create(n));
var gpuB = gpu.Device.Allocate <double>(Shape.Create(n));
gpu.Copy(gpuA, cpuA);
gpu.Assign(gpuB, gpuA);
// this copy need to sync, since cpuB is just a reference
gpu.Copy(cpuB, gpuB).Wait();
gpuB.Print();
Assert.IsTrue(arrayB.SequenceEqual(arrayA));
}
public static Tensor <T> Allocate <T>(this Context context, T[,,] initValues)
{
var shape = Shape.GetArrayShape(initValues);
var tensor = context.Device.Allocate <T>(shape);
var initTensor = initValues.AsTensor();
context.Copy(tensor, initTensor);
return(tensor);
}
public static T[,] ToArray2D <T>(this Context context, Tensor <T> tensor)
{
Util.EnsureTrue(tensor.Layout.Rank >= 2);
if (!tensor.Layout.IsInnerChangeMostFullyPacked)
{
var tempTensor = context.Device.Allocate <T>(tensor.Layout, tensor.Buffer.Memory.Length);
context.Copy(tempTensor, tensor).Wait();
tensor = tempTensor;
}
var rows = tensor.Layout.Shape[0];
var cols = tensor.Layout.Shape.Skip(1).Aggregate(ScalarOps.Mul);
var array = new T[rows, cols];
var cpuTensor = array.AsTensor();
context.Copy(cpuTensor, tensor).Wait();
return(array);
}
public static T[] ToArray <T>(this Context context, Tensor <T> tensor)
{
// if the tensor's layout is C Style, then we just make one copy and return it
if (tensor.Layout.IsInnerChangeMostFullyPacked)
{
var layout = tensor.Layout;
var shape = layout.Shape;
var array = new T[shape.Length];
var cpuTensor = array.AsTensor(shape);
context.Copy(cpuTensor, tensor).Wait();
return(array);
}
// not c style layout, we need allocate a temp tensor on that device, assign it, then copy it back
throw new NotImplementedException();
}
public static void Print <T>(this Context context, Tensor <T> tensor, bool all = false)
{
Task task;
// print must work on cpu side, if the tensor is on cpu side already, we
// can directly print it (using RDE to sync the resource)
if (tensor.Device == Device.CpuDevice)
{
task = new Task(() => tensor.Layout.Print(tensor.Buffer.RawReader, all));
}
else
{
// if the tensor is not on cpu, then we need make a copy of it and print it.
var cpuTensor = Device.CpuDevice.Allocate <T>(tensor.Layout, tensor.Memory.Length);
context.Copy(cpuTensor, tensor).Wait();
task = new Task(() => tensor.Layout.Print(cpuTensor.Buffer.RawReader, all));
}
// TODO:@RDE
task.Start();
task.Wait();
}
public static Task Assign <T>(this Context context, Tensor <T> tensor, T[] newValues)
{
var newValuesTensor = newValues.AsTensor();
return(context.Copy(tensor, newValuesTensor));
}