akaashrp opened a new pull request, #20409: URL: https://github.com/apache/tvm/pull/20409
#20378 moved block iterator, access region, and match buffer rewriting into `s_tir::IndexDataTypeNormalizer`, so `tirx.transform.ForceNarrowIndexToInt32` only handles functions whose blocks have been lowered. Callers that narrow scheduled TIR, such as MLC's compilation pipeline, get tensorized blocks with int64 iterators and int32 domains, which later fail in `CompactBufferAllocation` with a `mismatched types. int64 vs. int32` error. Every q4f16_1 Metal build in MLC currently hits this. Add `s_tir.transform.ForceNarrowIndexToInt32` for functions that still contain blocks, sharing the narrowing rules with the TIRX pass and restoring the check on int64 block allocations. The TIRX pass now rejects S-TIR blocks instead of producing inconsistent IR, and the scan dispatch tests that narrow dispatched TIR use the new pass. -- 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] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
