tlopex opened a new pull request, #19742:
URL: https://github.com/apache/tvm/pull/19742

   This pr fixes `InjectSoftwarePipeline` to rewrite opaque buffer offsets for 
legacy PTX MMA intrinsics, including `ptx_ldmatrix_legacy`, `ptx_mma_legacy`, 
`mma_store_legacy`, and `mma_fill_legacy`.
   
   ### Failure
   
   `test_async_nested_pipeline_mma_gemm_ideal_annotation` failed with a real 
numerical mismatch during the final GEMM result check:
   
   ```text
   Mismatched elements: 6744298 / 16777216 (40.2%)
   Max absolute difference: 6.706421
   Max relative difference: 0.00653978
   ```
   
   A diagnostic run with a fixed seed reproduced the same class of error:
   
   ```text
   nested_full bad 6826830 / 16777216
   max_abs 7.0405273
   mean_abs 0.9874781
   ```
   
   The adjacent simple pipeline test passed, and disabling async copy did not 
change the mismatch, so this was not a cp.async lowering issue, tolerance 
noise, or GPU flakiness.
   
   ### Root Cause
   
   `InjectSoftwarePipeline` may add a leading version dimension to pipeline 
buffers. Normal buffer loads/stores and newer opaque PTX intrinsics already had 
their offsets rewritten to include the pipeline version slot.
   
   However, the legacy MMA intrinsics were not covered. In this test, the warp 
buffers became multi-versioned, but the legacy `ldmatrix`/`mma` offsets still 
pointed to the original slot. As a result, the second `ldmatrix` stage 
overwrote the first warp fragment, and both MMA stages read the same fragment. 
That skipped one K fragment and duplicated another, producing the numerical 
mismatch above.


-- 
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]

Reply via email to