229 …CK_SIZE) __global__ void Grad(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *c_G,  in __launch_bounds__()
247   LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G);  in __launch_bounds__()
273 …void GradCollocated(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *c_G, const Ce…  in __launch_bounds__()
289   LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G);  in __launch_bounds__()
314 …void GradTranspose(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *c_G, const Cee…  in __launch_bounds__()
332   LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G);  in __launch_bounds__()
358 …Transpose(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *c_G, const CeedScalar *…  in __launch_bounds__()
374   LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G);  in __launch_bounds__()
399 …void GradTransposeAdd(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *c_G, const …  in __launch_bounds__()
417   LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G);  in __launch_bounds__()
443 …nsposeAdd(const CeedInt num_elem, const CeedScalar *c_B, const CeedScalar *c_G, const CeedScalar *…  in __launch_bounds__()
459   LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G);  in __launch_bounds__()