feat[dace]: Updated DaCe Transformations #1639
Draft
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
No real improvement though but it is now a bit clearer and does not use the DaCe properties for storing the intermediate sets it computes.
This is that we can continue until the transformations are fully merged.
edopao
reviewed
Sep 12, 2024
..._tests/program_processor_tests/runners_tests/dace/transformation_tests/test_loop_blocking.py
Outdated
Show resolved
Hide resolved
..._tests/program_processor_tests/runners_tests/dace/transformation_tests/test_loop_blocking.py
Outdated
Show resolved
Hide resolved
..._tests/program_processor_tests/runners_tests/dace/transformation_tests/test_loop_blocking.py
Outdated
Show resolved
Hide resolved
| gtx_transformations.LoopBlocking(blocking_size=10, blocking_parameter="j"), | ||
| validate=True, | ||
| validate_all=True, | ||
| ) | ||
| assert count > 0 |
There was a problem hiding this comment.
Is there a reason why we would not like to assert on the exact number we expect? I mean count == 1 or 2 (applies also below)
There was a problem hiding this comment.
There is actually no particular reason for it.
However you are right using the exact number is a bit better, because it ensures that the transformation is not applied twice in the same place.
| # this preconditions on SDFG should ensure that, but there are a few | ||
| # super hard edge cases. | ||
| # TODO(phimuell): Add an intermediate here in this case | ||
| assert isinstance(independent_node, dace_nodes.AccessNode) |
There was a problem hiding this comment.
I am not sure whether we should have an assert or an exception. Probably a runtime exception?
There was a problem hiding this comment.
You are right a NotImplementedError is better.
Before the transformation only did the promotion, but it now also does the fusion in one go.
It is now possible to specify values for maps fo different dimensions. The `gt_set_gpu_blocksize()` function still has the old interface, if the values are used then they are used as default and can be overwritten.
I slightly refactored the function.
…`wcr_nonatomic` properties of edges. The code that does this was now relocated to the auto optimizer directly.
This also applies to some other transformations.
It was not checked that the whole array was translated.
The case of sequential trivial map was not handled. This lead to on infinite loop since a trivial map was promoted to a trivial map. The fix is that such promition are rejected and handled later when the promottion was propagated.
There was a forgotten `not`.
However, It seems that it is not needed.
In some cases we might generate dead dataflow. This is an effect of how fusion currently and probably also in the future works.
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.
The initial version of the optimization pipeline only contained a rough draft.
Currently this PR contains a copy of the map fusion transformations from DaCe that are currently under review. As soon as that PR is merged and DaCe was updated in GT4Py these files will be deleted.
This PR collects some general improvements:
LoopBlockingtransformation (with tests).MapFusionParallelTodo:
can_be_applied_to()as soon as DaCe is updated (TrivialGPUMapElimination,SerialMapPromoter).BaseMapPromoter; as probably nobody needs that