wrongtest commented on pull request #9699:
URL: https://github.com/apache/tvm/pull/9699#issuecomment-999287677
Rebase, add more testcases and try to fix three detected issues:
1. My previous code is buggy on duplicate lowerbounds on same iteration, I
should not subtract `min` directly because it will break the assumption that
`mark->source` be a structural form `IterSumExpr`. Instead, set `-min` to base
as suggested.
2. Previously sum with single iter would not go into `TryFuseIters` thus if
we have the `i` in [0, 48) and predicate is `i < 10`, it seems ignore the
condition totally, result to an iteration of extent=48. A new testcase with
single var depict this.
```python
res = tvm.arith.detect_iter_map([i], var_dom([(i, 48)]), tvm.tir.all(i < 10))
assert_iter_sum_pattern(res[0], 10, 0)
```
3. If we fix (2), than the case below and several general unittests on tir
will fail because `TryFuseIters` can not fully match the predicate constraint
even all iters are visisted.
```python
res = tvm.arith.detect_iter_map(
[i, j, k],
var_dom([(i, 128), (j, 128), (k, 128)]),
tvm.tir.all(i * 16384 + j * 128 + k < 100),
)
```
The workaround currently is that if we fail to fully match the predicate
constraint, but a suffix of predicate is matched successfully, step into
fallback behaviors and skip this constraint. For example, if target iter expr
is `k` or `j*128+k` or `i*128+j`(fix me, this case will just not start a
constraint match), the predicate `i * 16384 + j * 128 + k < 100` will be
skipped.
Unfortunately, the originated `region_cover` issue now seems can not be
fully solved by this PR. I think the issue (2) make the corresponding case
incorrectly passed previously. There is only one iteration except free vars in
`hh_0 * 8 - 1 + ax0` so the produced region is inferred regardless of
predicate. A new case is added to intset test (previous case has two iters).
cc @spectrometerHBH
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
