bug[next]: itir.embedded: fix shift inside scan pass #1280
Conversation
| @@ -1196,7 +1195,7 @@ def can_deref(self) -> bool: | |||
| return self.wrapped_iter.can_deref() | |||
|
|
|||
| def shift(self, *offsets: OffsetPart) -> ScanArgIterator: | |||
| return ScanArgIterator(self.wrapped_iter, self.k_pos, offsets=[*offsets, *self.offsets]) | |||
| return ScanArgIterator(self.wrapped_iter.shift(*offsets), self.k_pos) | |||
There was a problem hiding this comment.
This is the important part.
src/gt4py/next/iterator/embedded.py
Outdated
| if isinstance(res, tuple): | ||
| out.field_setitem( | ||
| ordered_indices, tuple(res[i][k] for i in range(len(res))) | ||
| ) # TODO(tehrengruber): only works for scalars | ||
| else: | ||
| out.field_setitem(ordered_indices, res[k]) |
There was a problem hiding this comment.
is this fixed in https://github.com/GridTools/gt4py/pull/1141/files? or is it another instance of that problem?
There was a problem hiding this comment.
Not sure actually, might be two different things. This is not clean however as it only works for tuples not nested tuples.
| @@ -46,7 +46,7 @@ def shifted(inp): | |||
|
|
|||
| @fundef | |||
| def wrapped(inp): | |||
| return scan(sum, True, 0.0)(inp, lift(shifted)(inp)) | |||
| return scan(sum, True, 0.0)(inp, shift(Koff, 1)(inp)) | |||
There was a problem hiding this comment.
This change breaks something else, if we undo we could merge.
There was a problem hiding this comment.
I addressed this issue by adding another inline_into_scan pass after temporaries are created
There was a problem hiding this comment.
We concluded that this is fine for now even though we are not sure whether the ordering of the passes makes sense anymore. This will be addressed as a whole in another PR.
tests/next_tests/integration_tests/feature_tests/ffront_tests/test_execution.py
Outdated
Show resolved
Hide resolved
|
The test failures seem to be a result of removing https://github.com/GridTools/gt4py/pull/1280/files/d4160bb32c7e091e94afe983fbe747c32c9f7514#r1250968455 again. |
havogt
changed the title
Shifts were ignored in the
ScanArgIterator, i.e. all shifts in the scan_pass were ignored. This partially fixes thetest_icon_like_scan.Additional:
InlineIntoScanafter temporary creation to inline remaining operations that block bubbling scans to the top.