NDTensor is a C++ header library for N-Dimensional tensors. The templated Tensor class implements a number of functions for interacting with tensors, most of which are written generically (i.e. they are generated at compile time given the known dimensionality and type of the tensor).
A Tensor can be used to reference memory managed elsewhere, as in the following example:
std::vector<float> tensorData(width * height);
NDT::Tensor<2,float> image({width,height}, tensorData.data());
The library also provides a ManagedTensor class which automatically allocates memory in the constructor given the tensor dimensions, and deallocates the memory in the destructor:
NDT::ManagedTensor<2,float> image({width,height});
The ManagedTensor class both derives from and provides implicit conversion to a Tensor of the same time and dimensionality, such that a ManagedTensor class can be safely passed to a function expecting either a pointer or reference to a Tensor or an actual Tensor object. Copy construction and assignment for ManagedTensor are disabled, but move construction and assignment are available and transfer ownership of the allocated memory.
There are aliases for low-dimensional tensors according to more commonly used names, i.e. Tensor<1,T> is also known as Vector<T>, Tensor<2,T> is also known as Image<T> and Tensor<3,T> is also known as Volume<T>. The corresponding aliases also exist for ManagedTensor.
The Tensor class overloads operator(), which is called with a number of parameters depending on the tensor dimensionality, or an Eigen vector with the corresponding length. For example:
NDT::ManagedVector<float> vector(10);
vector(5) = 0.f;
NDT::ManagedImage<Eigen::Vector3f> image({10,10});
image(5,5) = Eigen::Vector3f(0,0,0);
image(Eigen::Vector2i(3,3)) = Eigen::Vector3f(0,0,0);
NDT::ManagedVolume<uint> volume({10,10,10});
volume(5,5,5) = 0;
volume(Eigen::Vector3i(3,3,3)) = 0;
The Interpolate function can be used to do linear interpolation:
NDT::ManagedImage<float> image({10,10});
std::cout << image.Interpolate(4.3f, 2.1f) << std::endl;
If compiled with CUDA support, a third tensor allows a Tensor or a ManagedTensor to reference GPU memory. For ManagedTensor, memory is automatically allocated on the host or device depending on this template parameter, but for Tensor, it is up to the library user to ensure that the pointer passed into the constructor is in the appropriate parameter space. Copying between host and device is easy, as shown in the following example:
// allocate memory
NDT::ManagedVolume<int> hVolume({4,4,4});
NDT::ManagedDeviceVolume<int> dVolume({4,4,4});
// fill the host volume
...
// copy from host to device
dVolume.CopyFrom(hVolume);
Normal rules for element access on host / device apply, i.e. host memory can only be accessed from code executing on the host and device memory can only be accessed from a CUDA kernel executing on the device.
The Tensor class assumes packed memory, such that the stride of the first dimension is the size of one element and the total size of the memory referenced is the product of the dimensions times the size of one element. The TensorView class, on the other hand, decouples the stride and dimensions of the tensor. A TensorView is created using the SubTensor member function of the Tensor class.
float data[16] = { 0, 1, 2, 3,
4, 5, 6, 7,
8, 9, 10, 11,
12, 13, 14, 15 };
NDT::Image<float> image({4,4}, data);
NDT::TensorView<2,float> imageView(2, 1, 2, 2);
std::cout << imageView(0,0) << std::endl; // prints 6
std::cout << imageView(1,1) << std::endl; // prints 11
One particularly useful use case is in copying to/from a part of a tensor:
NDT::ManagedVector<float> vector(10);
NDT::ManagedVector<float> smallerVector(2);
smallerVector(0) = 3.f;
smallerVector(1) = 4.f;
vector.SubTensor( 3, 2 ).CopyFrom(smallerVector);
NDT::ManagedImage<int> image({100,100});
NDT::ManagedImage<int> patch({10,10});
patch.CopyFrom(image.SubTensor(17, 25, 10, 10));