failing panzer tests on cuda
Created by: bathmatt
@rppawlo @kyungjoo-kim IntegrationValues2 is failing.. I'm dogging down the failing tests on cuda. Here is one I could use some help with. Not sure what is going on.
In integration_values2.cpp we have a test
TEST_FLOATING_EQUALITY(int_values_vol.ip_coordinates(0,0,0), realspace_x_coord_1, 1.0e-8);
The first value is NaN. It is computed here
int_values_vol.evaluateValues(node_coordinates);
This calls down to a call in Intrepid2_CubatureControlVolumeDef.hpp which looks like this
` // numNodesPerCell is maximum 8; this repeated run is necessary because of cell tools input consideration for (ordinal_type node=0;node<numNodesPerCell;++node) { auto subcvJacobianNode = Kokkos::subdynrankview(subcvJacobian, Kokkos::ALL(), node, Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL()); auto subcvCoordsNode = Kokkos::subdynrankview(subcvCoords, Kokkos::ALL(), node, Kokkos::ALL(), Kokkos::ALL()); auto subcvJacobianDetNode = Kokkos::subdynrankview(subcvJacobianDet, Kokkos::ALL(), node, Kokkos::ALL());
CellTools<SpT>::setJacobian(subcvJacobianNode, // C, P, D, D
subcvCubaturePoints_, // P, D
subcvCoordsNode, // C, N, D
subcvCellTopo_);
CellTools<SpT>::setJacobianDet(subcvJacobianDetNode, // C, P
subcvJacobianNode); // C, P, D, D
}
`
The setJacobian call returns an invalid jacobian
(cuda-gdb) p subcvJacobianNode.m_map .m_handle [0] $95 = -nan(0x8000000000000) (cuda-gdb)
The points look to be correct though, i know they are set correctly.
This code should be tested somewhere, is there a good test for getCubature somewhere? Have either of you seen it?
I have bounds checking on btw.