Recover default backend handling #42
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
Move `gt4py.next.iterator.ir_makers` and `gt4py.next.iterator.transforms.common_pattern_matcher` into a new module named `gt4py.next.iterator.ir_utils`. Just a small refactoring in preparation of GridTools#1350.
…ridTools#1370) Full-inlining of unrolled reduce in DaCe backend requires lambda arguments to be inlined, in order to generate the corresponding taskgraph. This PR adds an option to the existing ITIR transformation InlineLambdas to enable this additional transformation pass, disabled by default.
* Adapt state struct codegen for indexing of dace:* stencil backend * Add new "Default" schedule type for dace <> gt4py schedule mapping * Missing key for make template render * Fix typo * Typo of the typo (tm)
…expressions (GridTools#1377) This PR adds support to DaCe backend for sparse fields and reductions over lift expressions.
) During integration of icon4py stencils with the DaCe backend, it was found that reduce-unroll can generate an ITIR containing can_deref on a scalar value. Such expression should always evaluate to true, so it can be evaluated at compile-time. Note that in theory such case could be detected by the ITIR pass, once ITIR type inference is replaced by a new solution. At that time, the solution proposed here should be removed.
This commit adds positional arguments to the generated SDFG. It also improves the naming of some automatically generated symbols, such as the shape and stride.
…s#1379) Before it was only possible to translate ITIR to SDFG and execute it and it was not possible to extract the SDFG. This commits splits this task into two parts and thus allows to perform the ITIR to SDFG translation without executing it.
Most part of this PR is about cleaning up and refactoring the code in DaCe backend for translation of neighbor reduction. As part of the refactoring, the backend is now using the reduce library node from DaCe library. However, this PR also contains one functional change, which is a fix. Neighbor reduction should check for validity of neighbor index. This means for neighbor-tables to check that the neighbor index stored in the table is not -1. For neighbor strided offsets, we should check that the neighbor index does not access the origin field out of boundary.
…ridTools#1393) Some icon4py stencils require support for if expressions with tuple arguments. This PR adds support to the DaCe backend in the visitor of builtin_if function. Additionally, this PR contains one fix in the result of builtin_tuple_get, which should return a list.
Adds the scalar scan operator for embedded field view.
Fix for broken DaCe test in baseline: - use `flatten_list` to get `ValueExpr` arguments to numeric builtin function Additionally, enable test for DaCe backend (left-over from PR GridTools#1393).
- add style guide to the coding guidelines - fix existing error messages in next - deal with ensuing qa errorrs / test updates - unshadow one test and fix the code it wasn't testing Co-authored-by: Rico Häuselmann <ricoh@cscs.ch> Co-authored-by: Enrique González Paredes <enriqueg@cscs.ch>
Changes to the quickstart guide to use `field.asnumpy()` (introduced in GridTools#1366) instead of `np.asarray(field)`. The quickstart guide is still broken though since the embedded backend (used by default) does not support skip neighbors connectivities.
- Introduce a better Infinity - Make UnitRange Generic to express finite, infinite, left-finite, right-finite properly. - Remove `Set` from UnitRange
Added a function to get the arguments to call an SDFG. This commit adds a function that allows to generate the arguments needed to call an SDFG, before this was part of `run_dace_iterator()`. This made it very complex to run an SDFG outside this function. One should consider this as an amend to [PR GridTools#1379](GridTools#1379).
Implementation of array slicing in DaCe backend changed to a mapped tasklet. Tested on GPU. CUDA code generation did not support the previous implementation, based on memlet in nested-SDFG.
Modified how the SDFG arguments are computed. It was noticed that some transformations, especially the `SDFG.apply_gpu_transformation()`, to the SDFG, added new arguments to the SDFG. But, since a lot of functions build on the `SDFG.arg_names` member and this member was populated before the transformation, an error occurred. Thus it was changed such that `SDFG.arg_names` was only populated with the arguments also known to the Fencil.
Currently only supported in field view embedded.
Co-authored-by: Till Ehrengruber <t.ehrengruber@me.com>
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.
Description
Describe the content of the PR and links to related issues, bugs of features.
Requirements
Before submitting this PR, please make sure:
Additionally, if this PR contains code authored by new contributors:
signed by the employer if needed, provided to ETH Zurich
version of the AUTHORS.rst file included in the PR