Add ready event to Tensor and TensorList. #5673
Conversation
- move SharedEventLease to core - add more complete shared_ptr interface to SharedEventLease - add tests for SharedEventLease - add (set_)ready_event to Tensor and TensorList - minor refactoring in TensorList - remove OperatorIO::event in favor of TensorList's ready_event in exec2 Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
|
CI MESSAGE: [19219483]: BUILD STARTED |
|
CI MESSAGE: [19219483]: BUILD FAILED |
Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
|
CI MESSAGE: [19220893]: BUILD STARTED |
|
CI MESSAGE: [19220893]: BUILD PASSED |
Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
|
CI MESSAGE: [19248455]: BUILD STARTED |
|
CI MESSAGE: [19248455]: BUILD PASSED |
Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
|
CI MESSAGE: [19327083]: BUILD STARTED |
| if (!ptr->shares_data()) { | ||
| if (AccessOrder consumer_order = OutputConsumerStream(o)) | ||
| ptr->set_order(consumer_order, false); | ||
| if (ptr->is_pinned()) { |
There was a problem hiding this comment.
This fixes a bug - previously we would set the consumer order for non-pinned host buffers - this PR adds an assert that verifies that such buffers are always in host order.
| // Hack: use shared_ptr<void> to store a CUDA event - shared_ptr doesn't care whether the pointer | ||
| // it manages is a real pointer or something else as long as: | ||
| // - null value is equivalent to nullptr | ||
| // - the provided deleter can free the object. |
There was a problem hiding this comment.
Just to make sure: is this guaranteed by the standard, or is it an implementation detail of shared_ptr?
There was a problem hiding this comment.
Let's put it differently: we've already been using it this way for 6 years - we use shared_ptr to manage device memory, which is also non-dereferenceble on host, so it's effectively an opaque handle.
|
CI MESSAGE: [19327083]: BUILD PASSED |
Category:
New feature (non-breaking change which adds functionality)
Refactoring (Redesign of existing code that doesn't affect functionality)
Description:
Preliminary work for (cleaner) DLPack support.
DLPack needs to synchronize an stream so that the tensor is ready for use in that stream.
We can't use stream-to-stream synchronization because of prefetching. This PR adds ready_event to Tensor and TensorList to address that.
Additional information:
Affected modules and functionalities:
Executor2
Tensor
TensorList
Key points relevant for the review:
Tests:
Checklist
Documentation
DALI team only
Requirements
REQ IDs: N/A
JIRA TASK: N/A