[eigen] Feature request: Add compile-time access to Eigen::Map dimensionality

[Kasper Marstal]

Hello,

I would like to propose compile-time access to Eigen::Map dimensionality to
align the Map API with the TensorMap API. Possible implementation:

const static int NumDimensions = IsVectorAtCompileTime ? 1 : 2;

(If there are more Eigen-idiomatic methods to achieve this functionality
suggestions are most welcome)

My use case is this: I have a class Inputs<Ts...> that is templated over a
list of tensor maps. The class is constructed with a corresponding number
of Tensors, and these Tensors sliced via Eigen::TensorMap and processed in
parallel by a thread pool. NumDimensions is used to determine what are
spatial dimensions and what are batch dimensions. The leading dimensions
(dimension zero to NumIndices) are treated as batch dimensions, while the
rest are treated as spatial dimensions.

With this proposal, Eigen::Map and Eigen::TensorMap share the same API for
type and dimension (Scalar and NumDimensions), so that Eigen::Map can used
together with TensorMaps without meta programming.

In addition, compile time access to image dimensionality is important for
medical image processing libraries (my background).

I am happy to contribute a patch with whatever the maintainers deem the
best name (NumDimensions or NumIndices or something else), the most
appropriate implementation, and the best place to put it (e.g. in matrix
class or only in map class).

Kind Regards,
Kasper

[Kasper Marstal]

Hello,

I would like to propose compile-time access to Eigen::Map dimensionality to
align the Map API with the TensorMap API. Possible implementation:

const static int NumDimensions = IsVectorAtCompileTime ? 1 : 2;

(If there are more Eigen-idiomatic methods to achieve this functionality
suggestions are most welcome)

My use case is this: I have a class Inputs<Ts...> that is templated over a
list of tensor maps. The class is constructed with a corresponding number
of Tensors, and these Tensors sliced via Eigen::TensorMap and processed in
parallel by a thread pool. NumDimensions is used to determine what are
spatial dimensions and what are batch dimensions. The leading dimensions
(dimension zero to NumIndices) are treated as batch dimensions, while the
rest are treated as spatial dimensions.

With this proposal, Eigen::Map and Eigen::TensorMap share the same API for
type and dimension (Scalar and NumDimensions), so that Eigen::Map can used
together with TensorMaps without meta programming.

In addition, compile time access to image dimensionality is important for
medical image processing libraries (my background).

I am happy to contribute a patch with whatever the maintainers deem the
best name (NumDimensions or NumIndices or something else), the most
appropriate implementation, and the best place to put it (e.g. in matrix
class or only in map class).

Kind Regards,
Kasper