mdspan cache policy accessors
#2487
Closed
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
miscco
reviewed
Oct 1, 2024
|
|
||
| _LIBCUDACXX_BEGIN_NAMESPACE_CUDA | ||
|
|
||
| enum class EvictionPolicy |
There was a problem hiding this comment.
Is that a publicly facing enumeration?
If so we should document it
|
|
||
| accessor_reference(accessor_reference&&) = delete; | ||
|
|
||
| _CCCL_HIDE_FROM_ABI _CCCL_DEVICE _CCCL_FORCEINLINE accessor_reference(const accessor_reference&) = default; |
There was a problem hiding this comment.
I am not a big fan of putting _CCCL_FORCEINLINE everywhere
I was investigating making it part of _CCCL_HIDE_FROM_ABI but that lead to a ton of compiler issues
There was a problem hiding this comment.
I think it makes sense for small functions, especially in CUDA
Comment on lines
113
to
116
| static_assert(!::cuda::std::is_array<_ElementType>::value, | ||
| "cache_policy_accessor: template argument may not be an array type"); | ||
| static_assert(!::cuda::std::is_abstract<_ElementType>::value, | ||
| "cache_policy_accessor: template argument may not be an abstract class"); |
There was a problem hiding this comment.
I am wondering why those constraints are not in the other case
There was a problem hiding this comment.
they are on both version (const/non-const) of cache_policy_accessor
Co-authored-by: Michael Schellenberger Costa <[email protected]>
Co-authored-by: Michael Schellenberger Costa <[email protected]>
…VIDIA#2483) * Fix `common_type` specialization for extended floating point types The machinery we had in place was not really suited to specialize `common_type` because it would take precendence over the actual implementation of `common_type` In that case, we only specialized `common_type<__half, __half>` but not `common_type<__half, __half&>` and so on. This shows how brittle the whole thing is and that it is not extensible. Rather than putting another bandaid over it, add a proper 5th step in the common_type detection that properly treats combinations of an extended floating point type with an arithmetic type. Allowing arithmetic types it necessary to keep machinery like `pow(__half, 2)` working. Fixes [BUG]: `is_common_type` trait is broken when mixing rvalue references NVIDIA#2419 * Work around MSVC declval bug
There is an incredible compiler bug reported in nvbug4867473 where the use of system header changes the way some types are instantiated. The culprit seems to be that within a system header the compiler accepts narrowing conversions that it should not accept Work around it by moving __is_non_narrowing_convertible to its own header that is included before we define the system header machinery
Signed-off-by: fbusato <[email protected]>
Signed-off-by: fbusato <[email protected]>
Signed-off-by: fbusato <[email protected]>
Signed-off-by: fbusato <[email protected]>
…erty (NVIDIA#2489) Currently we implicitly assumed that any resource that had no execution space property was host accessible. However, that is not a good design, as it provides a source of surprise and numerous challenges with proper type matching down the road. So rather than implicitly assuming that something is host accessible, we require the user to always provide at least one execution space property.
* Move builtin detection to its own file * Try to reenable more builtins * Address review comments
This is used in the `cudax::vector` PR and the only dependency change of libcu++ which blows up the CI
Signed-off-by: fbusato <[email protected]>
Signed-off-by: fbusato <[email protected]>
fbusato
force-pushed
the
mdspan-cache-operator-accessors
branch
from
92b9963 to
b72b013
Compare
1 task
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
closes #2472
Add custom CUDA
mdspanaccessors to enable cache operators.The PR covers the following features:
cache_policy_accessorfor load and store operationcache_policy_accessorfor load-only operationaccessor_referencefor dispatching load and store operation in different wayscub::ThreadLoadandcub::ThreadStore(related issue [FEA]: Improve and cleanupThreadLoad#2486 for improving the two methods)(names to finalize later)