Inherit thrust::transform_iterator traversal from base iterator traversal#5883
Conversation
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
|
/ok to test |
|
/ok to test cf00e5e |
| use_default, | ||
| typename ::cuda::std::iterator_traits<Iterator>::iterator_category, | ||
| use_default, // pick system from Iterator | ||
| use_default, // pick traversal from Iterator |
There was a problem hiding this comment.
This line is the fix to #5860 and I also think it's an overall cleaner behavior. I don't know though why the old behavior was chosen. Need to do some digging.
There was a problem hiding this comment.
Found a hint: https://github.com/NVIDIA/cccl/pull/1915/files#diff-866ad8d2535a5dd3e954db56809b4673fcce2b064fc630fe24c872c5e8195a81L61-L67
typedef thrust::iterator_adaptor<transform_iterator<UnaryFunc, Iterator, Reference, Value>,
Iterator,
value_type,
thrust::use_default // Leave the system alone
//, thrust::use_default // Leave the traversal alone
// use the Iterator's category to let any system iterators remain
// random access even though transform_iterator's reference type
// may not be a reference
// XXX figure out why only iterators whose reference types are
// true references are random access
,
typename thrust::iterator_traits<Iterator>::iterator_category,
reference>It seems category was used to work around any system iterators loosing their random accessness. Let's see whether this problem still exists.
There was a problem hiding this comment.
I added tests to check whether the properties of a random access any system iterator are retained through a transform iterator. And they are. So I think the fix is legit.
|
/ok to test |
This comment has been minimized.
This comment has been minimized.
bernhardmgruber
force-pushed
the
transform_traversal_fix
branch
2 times, most recently
from
8c71eea to
1fa341e
Compare
This comment has been minimized.
This comment has been minimized.
…rsal Instead of from the base iterator's iterator category. Fixes: NVIDIA#5860
bernhardmgruber
force-pushed
the
transform_traversal_fix
branch
from
fe93a03 to
1ed55d1
Compare
🥳 CI Workflow Results🟩 Finished in 3h 11m: Pass: 100%/159 | Total: 6d 01h | Max: 3h 02m | Hits: 65%/186335See results here. |
…versal (NVIDIA#5883) Instead of from the base iterator's iterator category. Fixes: NVIDIA#5860
2 participants
Instead of from the base iterator's iterator category.
iterator_category_to_systemfor device iterator tags #5880Fixes: #5860