+1, I was just looking at that this week too (and yes there are parts hardcoded). I’d even be curious about flushing out hard coding of both of these and picking some other negative value initially (but the -1 is much easier one that slips in)
jpienaar
4
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| Tensor Ops with dynamic sizes - which behaviour is "more correct"? | 8 | 324 | October 21, 2024 | |
| Question on whether box value could be dynamic(unknown in compile-time) of nvgpu.tma.create.descriptor | 0 | 75 | March 9, 2024 | |
| [RFC] Update valueBoundsOpInterface to compute bounds of dynamic stride/offset sizes | 11 | 268 | August 5, 2024 | |
| [RFC] Support scalable dimensions in ShapedTypes | 21 | 1048 | December 4, 2023 | |
| Dynamic MemRef Cast | 2 | 820 | June 8, 2020 |