matrix to array-of-vectors and vice-versa

[bob-carpenter]

Summary:

Implement functions that will look like this in Stan:

// to_matrix(rv)[i, j] = rv[i, j]
matrix to_matrix(row_vector[] rv);  

// to_matrix(v)[i, j] = v[j, i]
matrix to_matrix(vector[] x);

// array_of_rows(y)[i, j] = y[i, j]
row_vector[] array_of_rows(matrix y);

// array_of_columns(y)[i, j] = y[j, i]
vector[] array_of_columns(matrix y);

Additional Info

Also need to add mc-stan/stan.

Current Version:

v2.17.0

[wds15]

would these functions work row / column wise?

so array_of_columns would take the matrix, then slice it by column and make them an array?

... and yes, these functions would be very useful...

[bob-carpenter]

This is the problem with these definition---not clear what they mean by just looking at them. I'll edit to make the intention clear.

[Inferrator]

In the context of using map_rect this is actually quite relevant. In plain c++, I would be able to create pointers, i.e. create an Eigen or Armadillo vector/matrix based on an array of reals or whatever.

[bob-carpenter]

@Inferrator: Thanks for commenting. map_rect is one of the main applications, as are simple things like multivariate distributions where you often want to access the data by column or by row.

Internally, we can wrap the data pointer of a std::vector to create an Eigen::Map<VectorXd>, for example, but Stan works by copy, not by reference sharing, so we'll have to copy anyway. It's not so easy to do the other way because we haven't customized std::vector allocators yet.

The bigger question is whether we can do things like program analysis in the compiler to figure out when we can get by without copying. But that's for the stanc3 repo...

[Inferrator]

Bob: thanks for clearing this up, that is very helpful.
Copies do become an issue when you have to pass on covariance matrices of dim 20 or above - it's a lot of copying.

[bob-carpenter]

Copying in Eigen or std::vector hits the C++ heap using malloc. It doesn't hit autodiff memory. That's good in that it doesn't increase memory headroom requirements, but bad in that it uses malloc, which is expensive. The copying is often dominated by something like a matrix solve or matrix multiply (cubic), or even a matrix-vector multiply (quadratic, but with autodiff). We're trying to reduce that on the inside by passing expression templates further along and inside the GPU by merging kernels.

…

On Feb 25, 2020, at 10:05 PM, Inferrator ***@***.***> wrote: Bob: thanks for clearing this up, that is very helpful. Copies do become an issue when you have to pass on covariance matrices of dim 20 or above - it's a lot of copying. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.