Cryolitia commented on issue #12681:
URL: https://github.com/apache/arrow/issues/12681#issuecomment-6974614049

   I can still reproduce this with unmodified Arrow 26.0.1 using 
`extra-riscv64-build` in a RISC-V chroot under QEMU, with GCC 16.2.1. The build 
retained `_GLIBCXX_ASSERTIONS`, `-march=rv64gc -mabi=lp64d`, and LTO.
   
   Running `arrow-compute-aggregate-test` gives:
   
   ```text
   [==========] 345 tests from 222 test suites ran. (2072 ms total)
   [  PASSED  ] 341 tests.
   [  FAILED  ] 4 tests, listed below:
   [  FAILED  ] TestRandomInt64QuantileKernel.Overlapped
   [  FAILED  ] TestRandomInt64QuantileKernel.Sliced
   [  FAILED  ] TestRandomFloatQuantileKernel.Exact
   [  FAILED  ] TestRandomFloatQuantileKernel.Sliced
   ```
   
   All seven reported numerical mismatches are exactly 1 ULP apart. For 
example, `TestRandomFloatQuantileKernel.Exact` reports:
   
   ```text
     Actual Value: 30.264954399603461
   Expected Value: 30.264954399603457
   ```
   
   The failure occurs at the [exact-equality 
assertion](https://github.com/apache/arrow/blob/apache-arrow-26.0.1/cpp/src/arrow/compute/kernels/aggregate_test.cc#L3852-L3855).
   
   Separately, on x86_64 with `-mfma`, the same four tests fail with 
`_GLIBCXX_ASSERTIONS` enabled. All 345 tests pass either after removing that 
macro, or after retaining it and replacing the equality check with 
`AssertWithinUlp(..., 1)`. This is consistent with the FMA rounding differences 
discussed earlier.


-- 
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]

Reply via email to