Update toolchains on tioga, lassen, ruby and poodle

[adrienbernede]

• Remove Intel 19
• Update to Intel 2023
• Update corona to ROCm 6.0.2 -> same error as with tioga, need ROCm 6.1.x , reverted to ROCm 5.7.1
• Update tioga to ROCm 6.1.2
• Remove CUDA 10 job (blueos system default is now 11.2.0)

Errors

• gcc 11.2.1 + cuda 11.8.0

[ 69%] Linking CXX executable ../../bin/managed_array_unit_tests
tmpxft_0001363d_00000000-6_managed_array_tests.cudafe1.stub.c:35:1: error: redefinition of 'const char _ZTSZN4chai12ManagedArrayIdE8allocateEmNS_14ExecutionSpaceERKSt8functionIFvPKNS_13PointerRecordENS_6ActionES2_EEEd_UlS6_S7_S2_E_ []'
tmpxft_0001363d_00000000-6_managed_array_tests.cudafe1.stub.c:35:1: note: 'const char _ZTSZN4chai12ManagedArrayIdE8allocateEmNS_14ExecutionSpaceERKSt8functionIFvPKNS_13PointerRecordENS_6ActionES2_EEEd_UlS6_S7_S2_E_ [125]' previously defined here
tmpxft_0001363d_00000000-6_managed_array_tests.cudafe1.stub.c:35:1: error: redefinition of 'const char _ZTSZN4chai12ManagedArrayIfE8allocateEmNS_14ExecutionSpaceERKSt8functionIFvPKNS_13PointerRecordENS_6ActionES2_EEEd_UlS6_S7_S2_E_ []'
tmpxft_0001363d_00000000-6_managed_array_tests.cudafe1.stub.c:35:1: note: 'const char _ZTSZN4chai12ManagedArrayIfE8allocateEmNS_14ExecutionSpaceERKSt8functionIFvPKNS_13PointerRecordENS_6ActionES2_EEEd_UlS6_S7_S2_E_ [125]' previously defined here
gmake[2]: *** [tests/integration/CMakeFiles/managed_array_tests.dir/build.make:76: tests/integration/CMakeFiles/managed_array_tests.dir/managed_array_tests.cpp.o] Error 1
gmake[1]: *** [CMakeFiles/Makefile2:1724: tests/integration/CMakeFiles/managed_array_tests.dir/all] Error 2
gmake[1]: *** Waiting for unfinished jobs....

[adayton1]

LGTM

[adrienbernede]

@adayton1 regarding the failure in the gcc 11.2.1 + cuda 11.8 build, do you think I should ignore the failure or wait for a fix?

[adayton1]

@adayton1 regarding the failure in the gcc 11.2.1 + cuda 11.8 build, do you think I should ignore the failure or wait for a fix?

I'm looking into this again.

[adayton1]

I think nvcc preprocesses std::function incorrectly. This isn't the first time I've seen issues with nvcc and the std library header.

[adayton1]

I think nvcc preprocesses std::function incorrectly. This isn't the first time I've seen issues with nvcc and the std library header.

I think we might just have to allow failure for now on this job.

[adrienbernede]

@adayton1 I was trying to understand the issue better, and here is my conclusion:

In src/chai/ManagedArray.inl at line 444,

template<typename T>
template<typename Idx>
CHAI_INLINE
CHAI_HOST_DEVICE T& ManagedArray<T>::operator[](const Idx i) const {
  return m_active_pointer[i];
}

This operator appears to be defined for both __host__ and __device__, while there exists an implementation of the class ManagedArray that is only defined for __host__,
In src/chai/ManagedArray.inl at line 82,

template<typename T>
CHAI_INLINE
CHAI_HOST ManagedArray<T>::ManagedArray(PointerRecord* record, ExecutionSpace space):
[...]

I think this explains the warning.

Then, the error itself appears to be related in the following way:

tmpxft_00021924_00000000-6_managed_array_tests.cudafe1.stub.c:35:1: error: redefinition of 'const char _ZTSZN4chai12ManagedArrayIdE8allocateEmNS_14ExecutionSpaceERKSt8functionIFvPKNS_13PointerRecordENS_6ActionES2_EEEd_UlS6_S7_S2_E_ []'
tmpxft_00021924_00000000-6_managed_array_tests.cudafe1.stub.c:35:1: note: 'const char _ZTSZN4chai12ManagedArrayIdE8allocateEmNS_14ExecutionSpaceERKSt8functionIFvPKNS_13PointerRecordENS_6ActionES2_EEEd_UlS6_S7_S2_E_ [125]' previously defined here

It looks like there is a mismatch between class implementations.
In tests/unit/managed_array_unit_tests.cpp at line ,

TEST(ManagedArray, SpaceConstructorCPU)
{
  chai::ManagedArray<float> array(10, chai::CPU);
  ASSERT_EQ(array.size(), 10u);
  array.free();
}

I think the intent is for the test to instantiate

CHAI_HOST_DEVICE ManagedArray<T>::ManagedArray(
    size_t elems,
    ExecutionSpace space) :
  ManagedArray()

defined at line 53 of src/chai/ManagedArray.inl, but the error clearly refers to the implementation involved in the above warning, the __host__ only one (line 84):

template<typename T>
CHAI_INLINE
CHAI_HOST ManagedArray<T>::ManagedArray(PointerRecord* record, ExecutionSpace space):
[...]

Could it be that PointerRecord* and size_t are not distinguishable ? I’m already beyond my limits... ;)

[adrienbernede]

@adayton1 You may now merge the PR.