CuPy to support accepting `order` as a parameter to functions other than `cupy.array`

Would it be possible that functions can accept an order parameter? For example, I would love to use https://docs-cupy.chainer.org/en/stable/reference/generated/cupy.random.standard_normal.html#cupy.random.standard_normal with order='F'. As it stands, the only way to convert a C matrix to F is by doing X = cp.array(X, order='F'). This temporarily blows up memory usage to 2x as it is most likely creating a copy of the transpose, I believe. For memory intensive applications, this is not the best as OOMs are common.

[divyegala]

Apparently if we wanna draw 1D samples, we should not specify size in multivariate_normal:

>>> import cupy as cp
>>>
>>> size = 1000
>>> mean = cp.random.random(size)
>>> cov_matrix = cp.diag(cp.ones(size))
>>> a = cp.random.multivariate_normal(mean, cov_matrix, dtype='float32')
>>> a.shape
(1000,)
>>> a = a.reshape(4, 250, order='F')
>>> a.shape
(4, 250)
>>> a.flags
  C_CONTIGUOUS : False
  F_CONTIGUOUS : True
  OWNDATA : False

@leofang upon more testing, that works but this is also what works for me: X = X.T.reshape(X.shape, order='F')

The problem with the approach you suggested is that I am using the covariance matrix to verify whether my PCA algorithm worked correctly. Reshaping point samples of size n to form 1D samples would not reproduce the same covariance matrix for me from PCA, I believe, as then the columns would become features of their own. If you could verify for me if my approach is correct, then it would be really helpful! 😃

[emcastillo]

Hi! Thanks for the feedback. With random functions, all the elements are generated independently so just transposing the matrix should give you a Fortran order matrix without an additional copy.

For the random package, NumPy does not support the order argument for the routines. So in order to keep as close as possible to NumPy we won’t be adding such option. But we are working on adding Generalized Universal Functions, so that order will be an argument of several NumPy core routines.