Specialize copy_from_numpy_into_usm_ndarray for contig case #1829
Conversation
|
Deleted rendered PR docs from intelpython.github.com/dpctl, latest should be updated shortly. 🤞 |
for contiguous case. Tensor implementation module temporarily exports both _copy_numpy_into_usm_ndarray_legacy, and _copy_numpy_into_usm_ndarray functions to performance comparison.
oleksandr-pavlyk
force-pushed
the
specialize-1d-case-for-numpy-to-usm-ndarray
branch
from
e043932 to
1869f06
Compare
|
Array API standard conformance tests for dpctl=0.19.0dev0=py310hdf72452_9 ran successfully. |
|
Array API standard conformance tests for dpctl=0.19.0dev0=py310hdf72452_9 ran successfully. |
…ernel Used unitrace to verify that the contiguous kernel is exercised in both direct call to asarray (for C-contiguous numpy array) with type casting, and when assigning to usm-ndarray from numpy of different data type when both arrays are dense, but strides may be negative.
|
Array API standard conformance tests for dpctl=0.19.0dev0=py310hdf72452_24 ran successfully. |
|
Array API standard conformance tests for dpctl=0.19.0dev0=py310hdf72452_25 ran successfully. |
dpctl/tensor/libtensor/source/copy_numpy_ndarray_into_usm_ndarray.cpp
Outdated
Show resolved
Hide resolved
|
@oleksandr-pavlyk |
|
Array API standard conformance tests for dpctl=0.19.0dev0=py310hdf72452_27 ran successfully. |
There was a problem hiding this comment.
LGTM, this is a nice improvement, thank you @oleksandr-pavlyk
oleksandr-pavlyk
deleted the
specialize-1d-case-for-numpy-to-usm-ndarray
branch
Specialize
_copy_from_numpy_into_usm_ndarrayfor contiguous case.Tensor implementation module temporarily exports both
_copy_numpy_into_usm_ndarray_legacy, and_copy_numpy_into_usm_ndarrayfunctions to performance comparison.The specialization reuses the same kernel, instantiated with
NoOpIndexer, reducing the number of kernel arguments.It also eliminates small allocation of USM temporaries needed to pass shape/strides information to the general kernel, and the cost of copying from host to these temporary allocations.