Created a temporary copy in case of overlap for unary function #1281
Conversation
|
Changes unknown when pulling 03a46e1 on unary_out_overlap into ** on master**. |
|
Array API standard conformance tests for dpctl=0.14.5dev1=py310h7bf5fec_7 ran successfully. |
|
View rendered docs @ https://intelpython.github.io/dpctl/pulls/1281/index.html |
|
Array API standard conformance tests for dpctl=0.14.5dev1=py310h7bf5fec_12 ran successfully. |
test_square.py::test_sqrt_out_overlap -> test_square.py::test_square_out_overlap
oleksandr-pavlyk
force-pushed
the
unary_out_overlap
branch
from
07015f1 to
fde917a
Compare
The call operator of this struct verifies whether two USM ND-arrays logically address the same memory elements. In the case when data-parallel read from and write to arrays that locally address the same memory elements there is no race condition and no additional copying is needed.
The predicate determines is argument arrays are the same (same dimension, shape, data type, pointer, strides). Used to determine if copying must be performed in case of overlap to avoid race condition.
Of out array is logically the same as input array, there is no race condition, so avoid performing the temporary copy.
oleksandr-pavlyk
force-pushed
the
unary_out_overlap
branch
from
fde917a to
09cd171
Compare
|
Array API standard conformance tests for dpctl=0.14.5dev1=py310h7bf5fec_15 ran successfully. |
|
Array API standard conformance tests for dpctl=0.14.5dev1=py310h7bf5fec_15 ran successfully. |
|
Array API standard conformance tests for dpctl=0.14.5dev1=py310h7bf5fec_18 ran successfully. |
|
Deleted rendered PR docs from intelpython.github.com/dpctl, latest should be updated shortly. 🤞 |
|
Array API standard conformance tests for dpctl=0.14.5dev1=py310h7bf5fec_18 ran successfully. |
The PR proposes to allocate a temporary buffer rather than to raise an exception in case when the memory overlapping is detected between
inandoutarrays in a call of unary function.The changed is intended to have the below code example as a valid:
The approach with temporary buffer was chosen as a start point while integrating support of the use case, since it is easiest way to implement this.
The next step might be to add separate kernels to handle in-place unary operations (when both
inandoutarrays point to the same memory) if the performance gain would be sensible.