iree: ScheduleAllocation crashes when compiling included sparse program

Segmentation fault inside of ScheduleAllocation (iree-stream-schedule-allocation) when compiling the following sparse program:

https://gist.github.com/rsuderman/28ad59022e46137903746557e6d05dd4

About this issue

Original URL
State: closed
Created a year ago
Comments: 22 (8 by maintainers)

Commits related to this issue

Allocation workaround for unused write-only dispatch results. The stream allocation pass is currently just looking at the top level of each execution region and values nested under concurrency regions... — committed to iree-org/iree by benvanik a year ago
Allocation workaround for unused write-only dispatch results. (#13748) The stream allocation pass is currently just looking at the top level of each execution region and values nested under concurre... — committed to iree-org/iree by benvanik a year ago
[StableHLO] Port gather to linalg lowering pattern This pattern fell through the cracks during the initial porting of hlo-to-linalg lowering in https://github.com/openxla/iree/pull/12957. With this ... — committed to kuhar/iree by kuhar a year ago
[StableHLO] Port gather to linalg lowering pattern (#13779) This pattern fell through the cracks during the initial porting of hlo-to-linalg lowering in https://github.com/openxla/iree/pull/12957. ... — committed to iree-org/iree by kuhar a year ago
Allocation workaround for unused write-only dispatch results. (#13748) The stream allocation pass is currently just looking at the top level of each execution region and values nested under concurre... — committed to NatashaKnk/iree by benvanik a year ago
[StableHLO] Port gather to linalg lowering pattern (#13779) This pattern fell through the cracks during the initial porting of hlo-to-linalg lowering in https://github.com/openxla/iree/pull/12957. ... — committed to NatashaKnk/iree by kuhar a year ago
Allocation workaround for unused write-only dispatch results. (#13748) The stream allocation pass is currently just looking at the top level of each execution region and values nested under concurre... — committed to plaidml/iree by benvanik a year ago
[StableHLO] Port gather to linalg lowering pattern (#13779) This pattern fell through the cracks during the initial porting of hlo-to-linalg lowering in https://github.com/openxla/iree/pull/12957. ... — committed to plaidml/iree by kuhar a year ago

Most upvoted comments

For argmax-like things we should be able to do much better than a globally-sized heap allocation (eventually) - as we’re regularly out-of-memory’ing on LLM’s I’d err on the side of assuming we’ll have to some day eliminate all such cases of globally-sized heap allocations that can be done any other way 😃

Argmax are handled today without globally-sizes heap allocation… they are handled through bounded stack allocations.

MaheshRavishankar on May 23, 2023

Yeah, the repro from @silvasean in #13637 (ir19 from #13543) also has this issue.

benvanik on May 23, 2023