Several errors in triton-to-linalg pass

[yuanfz98]

Hello,

We are trying to convert from ttirs to linalg in nanoGPT model and have encountered some edge cases that triton-to-linalg pass failed to lower.

https://drive.google.com/drive/folders/1oBIJFhdL8TQO2PIF_PsUaMJRPVMPJ6G7?usp=sharing

Thank you for your help.

[manbearian]

Thank you for the report. it looks like there are a number of unfinished/incomplete items you're running into with this collection of inputs.

here's a log of all the failures:
bugs.log

I'm seeing the following unique failures:

1. bad cast: decltype(auto) llvm::cast(const From&) [with To = mlir::MemRefType; From = mlir::Type]
2. "Only asserts on scalars are currently supported"
3. "Only support lowering reduction with body containing 1 maxf or addf."
4. "Scalar load is currently not supported"
5. failed to legalize operation 'tt.load'
6. "other value used in masked load produced by unsupported instruction"
7. "UNREACHABLE executed at src/triton/third_party/triton_shared/lib/Analysis/PtrAnalysis.cpp:377!

These all indicated not yet supported cases.

• Errors 2, 3, 4, and 7 are known issues i recognize.
• Error 6 looks like a missing case in mask-analysis.
• I'm not sure what exactly the errors 1 and 5 and pointing at and further investigation is required to understand.

Of the known issues (2,3,4,7): 3, 4, and 7 are likely to be addressed soon (next updated perhaps, not sure yet of our schedule on this) while item 2 is not on our planned set of features right now.

[makslevental]

@manbearian if you create a tracking issue (or whatever) for these (and others), I'm sure some of us adoring fans would be glad to pitch in (including appropriate RFC and design discussion and etc. of course). At least I know I would.

[manbearian]

@manbearian if you create a tracking issue (or whatever) for these (and others), I'm sure some of us adoring fans would be glad to pitch in (including appropriate RFC and design discussion and etc. of course). At least I know I would.

done!

From the items i created supporting additional reductions seem the most straightforward if you're looking for a nice on ramp :)

[nhat-nguyen]

Hello,

We are trying to convert from ttirs to linalg in nanoGPT model and have encountered some edge cases that triton-to-linalg pass failed to lower.

https://drive.google.com/drive/folders/1oBIJFhdL8TQO2PIF_PsUaMJRPVMPJ6G7?usp=sharing

Thank you for your help.

@yuanfz98 Thanks for sharing these! Would you mind also sharing us the triton code that produces these IR? It would help us a lot in understanding the usage of certain triton language features.