BLAS test failures with non-ETI-only testing

[brian-kelley]

With non-ETI testing enabled, some BLAS unit tests are failing on CUDA:

1: [  FAILED  ] 28 tests, listed below:
1: [  FAILED  ] Cuda.axpby_unification_float
1: [  FAILED  ] Cuda.axpby_mv_unification_float
1: [  FAILED  ] Cuda.axpby_unification_double
1: [  FAILED  ] Cuda.axpby_mv_unification_double
1: [  FAILED  ] Cuda.axpby_unification_complex_double
1: [  FAILED  ] Cuda.axpby_mv_unification_complex_double
1: [  FAILED  ] Cuda.axpby_unification_int
1: [  FAILED  ] Cuda.axpby_mv_unification_int
1: [  FAILED  ] Cuda.axpby_unification_double_int
1: [  FAILED  ] Cuda.axpby_double_mv_unification_int
1: [  FAILED  ] Cuda.ger_float
1: [  FAILED  ] Cuda.ger_complex_float
1: [  FAILED  ] Cuda.ger_double
1: [  FAILED  ] Cuda.ger_complex_double
1: [  FAILED  ] Cuda.ger_int
1: [  FAILED  ] Cuda.ger_double_int_float
1: [  FAILED  ] Cuda.syr_float
1: [  FAILED  ] Cuda.syr_complex_float
1: [  FAILED  ] Cuda.syr_double
1: [  FAILED  ] Cuda.syr_complex_double
1: [  FAILED  ] Cuda.syr_int
1: [  FAILED  ] Cuda.syr_int_float
1: [  FAILED  ] Cuda.syr2_float
1: [  FAILED  ] Cuda.syr2_complex_float
1: [  FAILED  ] Cuda.syr2_double
1: [  FAILED  ] Cuda.syr2_complex_double
1: [  FAILED  ] Cuda.syr2_int
1: [  FAILED  ] Cuda.syr2_int_float_double

The error message for the ger failures:

1: C++ exception with description "Error: Kokkos::deep_copy with no available copy mechanism: Y to Y_mirror" thrown in the test body.

For the others:

1: C++ exception with description "Error: Kokkos::deep_copy with no available copy mechanism: X to X_mirror" thrown in the test body.

This was with the configuration

../kokkos-kernels/cm_generate_makefile.bash \
  --with-devices=cuda \
  --arch=SKX,Pascal61 --release \
  --disable-test-eti-only \

or for raw CMake, it's the flags -DKokkosKernels_ENABLE_TESTS=ON -DKokkosKernels_TEST_ETI_ONLY=OFF.

@eeprude I think this is happening because the test is trying to copy non-contiguous views. Lines like

Kokkos::deep_copy(x.h_view, x.d_view);

should be

Kokkos::deep_copy(x.h_base, x.d_base);

h_view and d_view are subviews of h_base and d_base respectively, and h_base and d_base are always contiguous. The view_stride_adapter manages that so that we can test kernels more easily with LayoutStride views.

[eeprude]

@brian-kelley Thanks for the hint. I will work on it later today or this Thursday.