Advanced Topic: C vs Fortran Order

The Precomputed standard specifies that "raw" encoded chunks be stored in Fortran (column-major) order. Initially, we translated downloaded images into C order in order to be compatible with most people's expectations. However, this translation both takes time and sometimes memory to perform and Numpy makes most operations in Python transparent to the underlying representation. Usually the difference becomes important only with certain C extensions, the users of which are hopefully knowledgable about these issues.

Therefore, rather than forcing all users to pay the cost of transforming Fortran order data into C order data, I decided to defer the C vs. Fortran decision to the end user, a person that has better information about what tradeoffs are acceptable for their application. Here's how to switch to C order if you need it:

1. Using Numpy's Function

This is the typical way to do it, however it will incur 2x memory overhead.

img = vol[...]
img = np.ascontiguousarray(img)

2. Using fastremap

We were getting three channel float32 C order data from a GPU and needed to transform it to Fortran order for upload. The 2x memory pressure incurred would kill the process in some configurations. Therefore, I added an in-place transposition function to the Cython library fastremap.

$ pip install fastremap

import fastremap

img = GPU(...)
img = fastremap.asfortranarray(img)
vol[...] = img

Notably, if your data are 2D or 3D and square (cubic), the in-place transposition is also faster than the numpy function.