Fixes incorrect output in dpctl.tensor.where strided implementation #1171
Conversation
- Implemented FourOffsets_StridedIndexer and FourOffsets - Implemented simplify_iteration_space_4 and simplify_iteration_four_strides - Made std::min calls in strided_iters more readable - Added test case for #1170
ndgrigorian
force-pushed
the
where-fix-gh-1170
branch
from
287846e to
c4cb96b
Compare
- Matches other contract_iter functions
|
View rendered docs @ https://intelpython.github.io/dpctl/pulls/1171/index.html |
|
Array API standard conformance tests for dpctl=0.14.3dev0=py310h76be34b_108 ran successfully. |
ndgrigorian
force-pushed
the
where-fix-gh-1170
branch
from
c4cb96b to
98c6651
Compare
|
Array API standard conformance tests for dpctl=0.14.3dev0=py310h76be34b_108 ran successfully. |
|
I have added a test: def test_where_invariants():
test_sh = (6, 8,)
mask = dpt.asarray(np.random.choice([True, False], size=test_sh))
p = dpt.ones(test_sh, dtype=dpt.int16)
m = dpt.full(test_sh, -1, dtype=dpt.int16)
inds_list = [(np.s_[:3], np.s_[::2],), (np.s_[::2], np.s_[::2],), (np.s_[::-1], np.s_[:],),]
for ind in inds_list:
r1 = dpt.where(mask, p, m)[ind]
r2 = dpt.where(mask[ind], p[ind], m[ind])
assert (dpt.asnumpy(r1) == dpt.asnumpy(r2)).all() |
| py::ssize_t &, | ||
| py::ssize_t &); | ||
|
|
||
| void simplify_iteration_space_4(int &, |
There was a problem hiding this comment.
One day, it would be nice to figure out how to make this nicer, perhaps by packing data about src1, etc. into structs.
There was a problem hiding this comment.
And of course, this same comment applies to simplify_iteration_space_3 and others.
|
Array API standard conformance tests for dpctl=0.14.3dev0=py310h76be34b_109 ran successfully. |
|
Deleted rendered PR docs from intelpython.github.com/dpctl, latest should be updated shortly. 🤞 |
|
Array API standard conformance tests for dpctl=0.14.3dev0=py310h76be34b_109 ran successfully. |
3 participants
This PR implements simplify_iteration_space_4 and indexers for offsets in 4 arrays to resolve #1170.
Additionally, implements contract_iter4 and simplifies calls to std::min in strided_iters.
The issue was caused by an oversight in the mapping of elements between arrays in the strided where kernel. As the destination array was not being simplified, there was not a guarantee that the nth element of the destination corresponded to the correct element in x1 or x2.