Lunderberg commented on PR #13936: URL: https://github.com/apache/tvm/pull/13936#issuecomment-1492029560
Well, first mystery solved. The test failure was in `test_cuda_t2d`, but I had been comparing against the expected result in `test_cuda_c2d`. The `test_cuda_t2d` does start with the transpose, so it isn't an issue of a schedule primitive erroneously inserting a transposition. After that, I was able to track down and fix the test breakage. The main issue in understanding the test was that I assumed that it was comparing the module in `actual[0].mod` to the expected module. Instead, it was only using the trace from the sketch, re-applying it with the hard-coded decisions, and then comparing that result to the expected module. After that, the only difference was resolved by replacing `(var % 2 + 1) // 2` with `(var % 2)`, exactly the simplification enabled by this PR. -- 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]
