Replace device_vector with device_uvector in null_mask #7715
Conversation
harrism
added
3 - Ready for Review
| @@ -53,7 +53,7 @@ void set_null_mask(bitmask_type *bitmask, | |||
| * @param[in] stream CUDA stream used for device memory operations and kernel launches. | |||
| */ | |||
| std::vector<size_type> segmented_count_set_bits(bitmask_type const *bitmask, | |||
| std::vector<size_type> const &indices, | |||
| host_span<size_type const> indices, | |||
There was a problem hiding this comment.
I assume the host_span being passed by value here now is not a performance concern since it's much smaller than a vector (just a wrapper around a raw pointer)? Perhaps more for my edification than anything else: is there any particular reason to prefer not using the reference in this case?
There was a problem hiding this comment.
I think Vukasin covered this in his Better Code presentation on spans, but in general one of the points of views is you pass by value, though I don't remember why. It's just a pointer and a size. We don't pass column_view by value because they are not true views. There is an open issue for that one...
|
@gpucibot merge |
|
Rerun tests. Previous CI test error: |
Codecov Report
@@ Coverage Diff @@
## branch-0.19 #7715 +/- ##
===============================================
+ Coverage 81.86% 82.53% +0.66%
===============================================
Files 101 101
Lines 16884 17453 +569
===============================================
+ Hits 13822 14404 +582
+ Misses 3062 3049 -13
Continue to review full report at Codecov.
|
Replaces remaining
device_vectorinstances in null_mask.cu withdevice_uvector. Change the interface ofsegmented_count_[un]set_bitsto takehost_spaninstead ofstd::vector.