GEMM Serial test failures

[ndellingwood]

The GEMM unit test is failing in serial builds on White with the xl compiler following merge of PR #524

The change likely causing the serial failures was the modification in
unit_test/blas/Test_Blas3_gemm.hpp

from this

    Kokkos::fill_random(A,rand_pool,ScalarA(10));
    Kokkos::fill_random(B,rand_pool,ScalarB(10));
    Kokkos::fill_random(C,rand_pool,ScalarC(10));

to this

    Kokkos::fill_random(A,rand_pool, Kokkos::rand<typename Kokkos::Random_XorShift64_Pool<execution_space>::generator_type,ScalarA>::max());
    Kokkos::fill_random(B,rand_pool, Kokkos::rand<typename Kokkos::Random_XorShift64_Pool<execution_space>::generator_type,ScalarB>::max());
    Kokkos::fill_random(C,rand_pool, Kokkos::rand<typename Kokkos::Random_XorShift64_Pool<execution_space>::generator_type,ScalarC>::max());

@seheracer do you think if we add if-guards etc to allow the new fill_random calls when Cuda is enabled and allow the previous behavior otherwise will achieve the same goal, or is another approach needed?

Edit: Adding link to failing Jenkins build

[ndellingwood]

Also failing with these builds:

KokkosKernels_White_XL_16_1_1_Serial_cpp14 
KokkosKernels_White_OpenMP_gcc_720_cpp14 
KokkosKernels_KokkosDev2_GCC720 
KokkosKernels_White_OpenMP_gcc_640
KokkosKernels_Apollo_Clang6_Threads
...

Edit: Nearly all host builds are seeing this test fail.

[ndellingwood]

I think this will get tests back to passing:

#ifdef KOKKOS_ENABLE_CUDA

if (std::is_same<execution_space,Kokkos::Cuda>::value) {
    Kokkos::fill_random(A,rand_pool, Kokkos::rand<typename Kokkos::Random_XorShift64_Pool<execution_space>::generator_type,ScalarA>::max());
    Kokkos::fill_random(B,rand_pool, Kokkos::rand<typename Kokkos::Random_XorShift64_Pool<execution_space>::generator_type,ScalarB>::max());
    Kokkos::fill_random(C,rand_pool, Kokkos::rand<typename Kokkos::Random_XorShift64_Pool<execution_space>::generator_type,ScalarC>::max());
}
else
#endif 
{
    Kokkos::fill_random(A,rand_pool,ScalarA(10));
    Kokkos::fill_random(B,rand_pool,ScalarB(10));
    Kokkos::fill_random(C,rand_pool,ScalarC(10));
}

I'll put in a PR shortly

[seheracer]

@ndellingwood Thanks! I shouldn't have skipped the extensive testing, lessons learned..

[ndellingwood]

@ndellingwood Thanks! I shouldn't have skipped the extensive testing, lessons learned..

Yeah, my mistake for requesting to skip it since the Trilinos tests passed which included host builds, lessons learned indeed...

[ndellingwood]

The change above worked, single host run on kokkos-dev-2 confirmed:

[ndellin@kokkos-dev-2 TestAllSandia]$ ../../scripts/test_all_sandia --spot-check gcc/7.3 --skip-hwloc
Running on machine: kokkos-dev-2
Going to test compilers:  gcc/7.3.0
Testing compiler gcc/7.3.0
  Starting job gcc-7.3.0-Pthread-release
  Starting job gcc-7.3.0-OpenMP-release
  PASSED gcc-7.3.0-OpenMP-release
  PASSED gcc-7.3.0-Pthread-release
#######################################################
PASSED TESTS
#######################################################
gcc-7.3.0-OpenMP-release build_time=218 run_time=272
gcc-7.3.0-Pthread-release build_time=204 run_time=455
#######################################################

I'll have a PR ready pretty quick here

[ndellingwood]

PR #529 issued.