GEMV unit test failures on Ponte Vecchio with oneMKL

[cwpearson]

The tests are failing on Blake with the Ponte Vecchio GPUs with oneMKL.
All failures are when beta=0: a possible cause is beta=0 is not clearing the input vector of NaN.

clone Kokkos to kokkos and Kokkos Kernels to kokkos-kernels

export KOKKOS_SRC="kokkos"
export KOKKOS_BUILD="kokkos-build"
export KOKKOS_INSTALL="kokkos-install"

export KERNELS_SRC="kokkos-kernels"
export KERNELS_BUILD="kokkos-kernels-build"

source /projects/x86-64-icelake-rocky8/spack-config/blake-setup-user-module-env.sh
module load cmake
module load intel-oneapi-compilers/2023.1.0 intel-oneapi-dpl/2022.1.0 intel-oneapi-mkl/2023.1.0
export ZES_ENABLE_SYSMAN=1

## Configure Kokkos
cmake -S "$KOKKOS_SRC" -B "$KOKKOS_BUILD" \
-DCMAKE_INSTALL_PREFIX="$KOKKOS_INSTALL" \
-DCMAKE_CXX_STANDARD=17 \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_CXX_COMPILER=icpx \
-DKokkos_ENABLE_SYCL=ON \
-DKokkos_ARCH_SPR=ON \
-DKokkos_ARCH_INTEL_PVC=ON \
-DKokkos_ENABLE_ONEDPL=OFF \
-DCMAKE_CXX_FLAGS="-fp-model=precise -fno-finite-math-only -mavx512f" \
-DBUILD_SHARED_LIBS=ON

## Build & Install Kokkos
cmake --build "$KOKKOS_BUILD" -j "$(nproc)" -t install

## Configure Kernels
cmake -S "$KERNELS_SRC" -B "$KERNELS_BUILD" \
-DKokkos_DIR="$KOKKOS_INSTALL/lib64/cmake/Kokkos" \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_CXX_COMPILER=icpx \
-DKokkosKernels_ENABLE_TESTS=ON \
-DKokkosKernels_ENABLE_PERFTESTS=ON \
-DKokkosKernels_ENABLE_BENCHMARK=ON \
-DKokkosKernels_ENABLE_TPL_MKL=ON \
-DCMAKE_CXX_FLAGS="-fp-model=precise -mavx512f" \
-DBUILD_SHARED_LIBS=ON

## Build Kernels
VERBOSE=1 make -C "$KERNELS_BUILD/blas/unit_test" -j "$(nproc)" 

srun -N 1 -p PV --exclude=blake15 -n1 -t 60 ctest --test-dir "$KERNELS_BUILD" -V -R blas |& tee "$LOG_DIR/ctest.log"

9: [ RUN      ] sycl_test.gemv_double
9: /projects/cwpears/kk-mkl/kokkos-kernels/blas/unit_test/Test_Blas2_gemv.hpp:141: expected(0)=-1.50128, h_y(0)=nan, eps=2.22045e-16, 1024*2*eps=4.54747e-13
9: /projects/cwpears/kk-mkl/kokkos-kernels/blas/unit_test/Test_Blas2_gemv.hpp:141: expected(1)=1.1445, h_y(1)=nan, eps=2.22045e-16, 1024*2*eps=4.54747e-13
9: /projects/cwpears/kk-mkl/kokkos-kernels/blas/unit_test/Test_Blas2_gemv.hpp:141: expected(2)=0.0141285, h_y(2)=nan, eps=2.22045e-16, 1024*2*eps=4.54747e-13
9: /projects/cwpears/kk-mkl/kokkos-kernels/blas/unit_test/Test_Blas2_gemv.hpp:141: expected(3)=-0.466376, h_y(3)=nan, eps=2.22045e-16, 1024*2*eps=4.54747e-13
9: /projects/cwpears/kk-mkl/kokkos-kernels/blas/unit_test/Test_Blas2_gemv.hpp:141: expected(4)=3.11782, h_y(4)=nan, eps=2.22045e-16, 1024*2*eps=4.54747e-13
9: /projects/cwpears/kk-mkl/kokkos-kernels/blas/unit_test/Test_Blas2_gemv.hpp:141: expected(5)=-2.10281, h_y(5)=nan, eps=2.22045e-16, 1024*2*eps=4.54747e-13
9: /projects/cwpears/kk-mkl/kokkos-kernels/blas/unit_test/Test_Blas2_gemv.hpp:141: expected(6)=-3.70048, h_y(6)=nan, eps=2.22045e-16, 1024*2*eps=4.54747e-13
9: /projects/cwpears/kk-mkl/kokkos-kernels/blas/unit_test/Test_Blas2_gemv.hpp:141: expected(7)=-3.07327, h_y(7)=nan, eps=2.22045e-16, 1024*2*eps=4.54747e-13
9: /projects/cwpears/kk-mkl/kokkos-kernels/blas/unit_test/Test_Blas2_gemv.hpp:141: expected(8)=-1.42057, h_y(8)=nan, eps=2.22045e-16, 1024*2*eps=4.54747e-13
9: /projects/cwpears/kk-mkl/kokkos-kernels/blas/unit_test/Test_Blas2_gemv.hpp:141: expected(9)=1.28458, h_y(9)=nan, eps=2.22045e-16, 1024*2*eps=4.54747e-13
9: /projects/cwpears/kk-mkl/kokkos-kernels/blas/unit_test/Test_Blas2_gemv.hpp:141: expected(10)=2.60594, h_y(10)=nan, eps=2.22045e-16, 1024*2*eps=4.54747e-13
9: /projects/cwpears/kk-mkl/kokkos-kernels/blas/unit_test/Test_Blas2_gemv.hpp:141: expected(11)=-5.60001, h_y(11)=nan, eps=2.22045e-16, 1024*2*eps=4.54747e-13
9: /projects/cwpears/kk-mkl/kokkos-kernels/blas/unit_test/Test_Blas2_gemv.hpp:141: expected(12)=-0.193979, h_y(12)=nan, eps=2.22045e-16, 1024*2*eps=4.54747e-13
9: /projects/cwpears/kk-mkl/kokkos-kernels/blas/unit_test/Test_Blas2_gemv.hpp:147: Failure
9: Value of: 0
9: Expected: numErrors
9: Which is: 13
9: beta = 0, input contains NaN, A is 13x13, mode N: gemv incorrect

[cwpearson]

@lucbv says this is reported to Intel, in the mean time we can apply beta=0 to the input vector ourselves as a workaround