Reshape improvements

[oleksandr-pavlyk]

Closes gh-1664.

Improves performance of dpt.reshape(X, new_shape, order="F") when copy is needed.

• Have you provided a meaningful PR description?
• Have you added a test, reproducer or referred to an issue with a reproducer?
• Have you tested your changes locally for CPU and GPU devices?
• Have you made sure that new changes do not introduce compiler warnings?
• Have you checked performance impact of proposed changes?
• If this PR is a work in progress, are you opening the PR as a draft?

[github-actions]

Deleted rendered PR docs from intelpython.github.com/dpctl, latest should be updated shortly. 🤞

[github-actions]

Array API standard conformance tests for dpctl=0.17.0dev0=py310h15de555_355 ran successfully.
Passed: 887
Failed: 18
Skipped: 91

[coveralls]

coverage: 87.951% (+0.003%) from 87.948%
when pulling 9d2633f on reshape-improvements
into 7bc3124 on master.

[antonwolfy]

53 tests affecting call of gemm function are failed in dpnp with that PR.
But I believe this might be due to some unproper internal logic in dpnp.

The first issue I've found is incorrect result returned below:

import dpnp
from dpnp.dpnp_utils.dpnp_utils_linearalgebra import _define_contig_flag

a = dpnp.arange(3).reshape((1, 3, 1))
_define_contig_flag(a)
# Out: False

but it's expected to return True, because a is both C-contiguous or F-contiguous along last two dimensions.
It will result in unexpected extra memory allocation to copy array a to temporary C-contiguous array.

The second one is somewhere in _gemm_batch implementation:

import dpctl, dpctl.tensor as dpt
import dpnp.backend.extensions.blas._blas_impl as bi

a = dpt.reshape(dpt.arange(60, dtype='f4'), (5, 4, 3))
b = dpt.reshape(dpt.arange(3, dtype='f4'), (1, 3, 1))
b = dpt.copy(b)
c = dpt.zeros((5, 4, 1))

a.strides, b.strides, c.strides
# Out: ((12, 3, 1), (3, 1, 1), (4, 1, 1))

ev,  _, _ = bi._gemm_batch(a.sycl_queue, a, b, c)
ev.wait()

c
# Out:
# usm_ndarray([[[ 5.],
#               [14.],
#               [23.],
#               [32.]],
# 
#              [[ 0.],
#               [ 0.],
#               [ 0.],
#               [ 0.]],
# 
#              [[ 0.],
#               [ 0.],
#               [ 0.],
#               [ 0.]],
# 
#              [[ 0.],
#               [ 0.],
#               [ 0.],
#               [ 0.]],
# 
#              [[ 0.],
#               [ 0.],
#               [ 0.],
#               [ 0.]]], dtype=float32)

@vtavana , could you please look on that?

[vtavana]

53 tests affecting call of gemm function are failed in dpnp with that PR. But I believe this might be due to some unproper internal logic in dpnp.

The necessary changes are implemented in dpnp-gh-1828. Relevant tests in dpnp are now passed with both master branch of dpctl and this branch.

[ndgrigorian]

The first issue I've found is incorrect result returned below:

import dpnp
from dpnp.dpnp_utils.dpnp_utils_linearalgebra import _define_contig_flag

a = dpnp.arange(3).reshape((1, 3, 1))
_define_contig_flag(a)
# Out: False

but it's expected to return True, because a is both C-contiguous or F-contiguous along last two dimensions. It will result in unexpected extra memory allocation to copy array a to temporary C-contiguous array.

First case is working correctly in dpctl

In [1]: import dpctl.tensor as dpt, dpctl, numpy as np

In [2]: x = dpt.reshape(dpt.arange(10), (1, 10, 1))

In [3]: x.flags.contiguous
Out[3]: True

In [4]: x.flags
Out[4]:
  C_CONTIGUOUS : True
  F_CONTIGUOUS : True
  WRITABLE : True

@vtavana does dpnp-gh-1828 address this case too?

[vtavana]

First case is working correctly in dpctl

In [1]: import dpctl.tensor as dpt, dpctl, numpy as np

In [2]: x = dpt.reshape(dpt.arange(10), (1, 10, 1))

In [3]: x.flags.contiguous
Out[3]: True

In [4]: x.flags
Out[4]:
  C_CONTIGUOUS : True
  F_CONTIGUOUS : True
  WRITABLE : True

@vtavana does dpnp-gh-1828 address this case too?

Yes, both examples provided by @antonwolfy also work fine in dpnp-gh-1828.

As a side note, _define_contig_flag from dpnp is used in batch calculation. And the goal is to check if each 2D array that forms the N-D array is f-contiguous or c-contiguous. So, we do not use built-in flag there.

[ndgrigorian]

Tested this out, this LGTM!