CUDA 11.0: Unresolved extern function (fixed in CUDA 11.1)

[havogt]

In some conditions, GridTools code won't compile with CUDA 11 with the following error

ptxas fatal   : Unresolved extern function '_ZN9gridtools5impl_17tuple_leaf_getter3getILm0ENS_17positional_strideEEERKT0_RKNS0_10tuple_leafIXT_ES4_Lb1EEE'

This issues is partly worked-around in #1519 by disabling a static_assert.

NVIDIA is investigating the problem. We are waiting for feedback.

[havogt]

Bug should be fixed in CUDA 11.1u2.

[havogt]

Current default CUDA version on daint is 11.0 which is affected by this bug. The bug is triggered by regression_tests/boundary_conditions.cc.

[anstaf]

For the record: demangled symbol is:
gridtools::positional_stride const& gridtools::impl_::tuple_leaf_getter::get<0ul, gridtools::positional_stride>(gridtools::impl_::tuple_leaf<0ul, gridtools::positional_stride, true> const&)

[havogt]

In the issue with the boundary_conditions, the symbol is a different one (same function, but template instantiated with a different type). If I understand the bug correctly, somewhere in a static_assert the function is used, but not in run-time. Looking at ptx (where the symbol appears) didn't help me to figure out the location of the static_assert (if that's really the problem).

[havogt]

In the boundary_conditions problem it is _ZN9gridtools5impl_17tuple_leaf_getter3getILm0EiEERKT0_RKNS0_10tuple_leafIXT_ES3_Lb0EEE

int const& gridtools::impl_::tuple_leaf_getter::get<0ul, int>(gridtools::impl_::tuple_leaf<0ul, int, false> const&)