Han-Chung Wang
Han-Chung Wang
Okay... I just found that it is enabled under some flags. I'm seeing that the benchmark is using additional flags (e.g., `--iree-flow-enable-aggressive-fusion`) which falls to experimental benchmark suites. I don't...
> > https://gist.github.com/IanWood1/827aee042ff6968070d2d0c6beccae4a > > To repro: run `iree-compile --output-format=vm-bytecode --compile-from=executable-sources ~/repro.mlir -o /tmp/a.vmfb`. @bangtianliu Mahesh mentioned that you could help on this. Can you take a look?
https://github.com/iree-org/iree/pull/17725 fixes the issue. The GPU specific issue is tracking in https://github.com/iree-org/iree/issues/17948 closing the issue.
> This is going to get tricky. Do we really need this on the encoding attribute? Could get out of hand pretty quickly. Yes, I think this is needed for...
If there are concerns, I can prototype all the things in a branch. And we can revisit the needs of this.
Converting it to draft because we need more information. We will iterate it on a branch, and land it to main later.
I can review after you fix all build failures. It looks not ready to me when CI is so red.
Please click `Re-request` item or ping me on discord when it is ready for the review. Thanks! :)
> Can we wait a few more days? The IREE folks are still integrating the last bufferization changes. They are one of the few projects that use the bufferization extensively...
There is a tensor.extract_slice created after distribution, which blocks the further TileAndFuse. So it generates large vectors. Someone needs to check why the extract_slice is created and fix it. If...