Build and test with CUDA 12.9.0

[bdice]

This PR uses CUDA 12.9.0 to build and test.

xref: rapidsai/build-planning#173

[jakirkham]

Thanks Bradley! 🙏

Do we want to update the Spark image to CUDA 12.9 as well ( cc @sameerz )?

cudf/.github/workflows/spark-rapids-jni.yaml

Line 10 in a02b4fa

image: rapidsai/ci-spark-rapids-jni:rockylinux8-cuda12.8.0

Also it looks like pre-commit got stuck on git checkout. Is there an easy way to restart it?

Edit: We noted offline that there was a GitHub incident today. Likely that caused the pre-commit issue. It has since cleared up

[jakirkham]

Seeing the following build error in CI

 │ │ $BUILD_PREFIX/bin/../lib/gcc/x86_64-conda-linux-gnu/13.3.0/../../../../x86_64-conda-linux-gnu/bin/ld: cannot find -lnvToolsExt: No such file or directory
 │ │ collect2: error: ld returned 1 exit status
 │ │ [3/8] Building CUDA object CMakeFiles/custom_p

The nvToolsExt library was dropped in CUDA 12.9

That said, think we are just using NVTX headers (like nvtx3/nvtoolsext.h). So we don't need the library to be linked. And can disable it

Think lines like this...

cudf/cpp/examples/parquet_io/CMakeLists.txt

Line 27 in 874ecb4

target_link_libraries(parquet_io PRIVATE cudf::cudf nvToolsExt $<TARGET_OBJECTS:parquet_io_utils>)

...can either be replaced with...

nvtx3::nvtx3-cpp

...as is done here

cudf/cpp/CMakeLists.txt

Line 998 in 874ecb4

PRIVATE $<BUILD_LOCAL_INTERFACE:nvtx3::nvtx3-cpp> cuco::cuco ZLIB::ZLIB nvcomp::nvcomp

[jakirkham]

After discussing with the Spark team offline, they have filed a CUDA 12.9 update issue: NVIDIA/spark-rapids#12679

[jakirkham]

Seeing the following test failure in conda C++ tests:

[  FAILED  ] 3 tests, listed below:
[  FAILED  ] CollectTestFixedWidth/22.CollectSet, where TypeParam = numeric::fixed_point<__int128,(numeric::Radix)10>
[  FAILED  ] ReductionHistogramTest/8.Histogram, where TypeParam = numeric::fixed_point<__int128,(numeric::Radix)10>
[  FAILED  ] ReductionHistogramTest/8.MergeHistogram, where TypeParam = numeric::fixed_point<__int128,(numeric::Radix)10>

 3 FAILED TESTS
CMake Error at run_gpu_test.cmake:35 (execute_process):
  execute_process failed command indexes:

    1: "Child return code: 1"

[jakirkham]

Also seeing the following test failure in conda and wheel Python tests on CI:

FAILED tests/test_dataframe.py::test_decimal_quantile[Decimal128Dtype-higher-1] - AssertionError: DataFrame.iloc[:, 1] (column name="val") are different

DataFrame.iloc[:, 1] (column name="val") values are different (100.0 %)
[index]: [1.0]
[left]:  [98.14]
[right]: [453.23]
At positional index 0, first diff: 98.14 != 453.23
FAILED tests/test_dataframe.py::test_decimal_quantile[Decimal128Dtype-higher-q3] - AssertionError: DataFrame.iloc[:, 1] (column name="val") are different

[jakirkham]

@pxLi have bumped the Spark RAPIDS CI image to CUDA 12.9.0 above

Is there anything else we need to do?

[wence-]

These failures are all with 128bit decimals which smells like miscompilation somewhere.