33   __shared__ CeedScalar s_B[BASIS_P * BASIS_Q];  in __launch_bounds__()
34   LoadMatrix<BASIS_P, BASIS_Q>(data, c_B, s_B);  in __launch_bounds__()
39 …ReadElementStrided1d<BASIS_NUM_COMP, BASIS_P>(data, elem, 1, BASIS_P * num_elem, BASIS_P, d_U, r_U…  in __launch_bounds__()
40     InterpNonTensor<BASIS_NUM_COMP, BASIS_P, BASIS_Q, BASIS_T_1D>(data, r_U, s_B, r_V);  in __launch_bounds__()
60   __shared__ CeedScalar s_B[BASIS_P * BASIS_Q];  in __launch_bounds__()
61   LoadMatrix<BASIS_P, BASIS_Q>(data, c_B, s_B);  in __launch_bounds__()
67     InterpTransposeNonTensor<BASIS_NUM_COMP, BASIS_P, BASIS_Q, BASIS_T_1D>(data, r_U, s_B, r_V);  in __launch_bounds__()
68 …WriteElementStrided1d<BASIS_NUM_COMP, BASIS_P>(data, elem, 1, BASIS_P * num_elem, BASIS_P, r_V, d_…  in __launch_bounds__()
87   __shared__ CeedScalar s_B[BASIS_P * BASIS_Q];  in __launch_bounds__()
88   LoadMatrix<BASIS_P, BASIS_Q>(data, c_B, s_B);  in __launch_bounds__()
94     InterpTransposeNonTensor<BASIS_NUM_COMP, BASIS_P, BASIS_Q, BASIS_T_1D>(data, r_U, s_B, r_V);  in __launch_bounds__()
95 …SumElementStrided1d<BASIS_NUM_COMP, BASIS_P>(data, elem, 1, BASIS_P * num_elem, BASIS_P, r_V, d_V);  in __launch_bounds__()
117   __shared__ CeedScalar s_G[BASIS_P * BASIS_Q * BASIS_DIM];  in __launch_bounds__()
118   LoadMatrix<BASIS_P, BASIS_Q * BASIS_DIM>(data, c_G, s_G);  in __launch_bounds__()
123 …ReadElementStrided1d<BASIS_NUM_COMP, BASIS_P>(data, elem, 1, BASIS_P * num_elem, BASIS_P, d_U, r_U…  in __launch_bounds__()
124     GradNonTensor<BASIS_NUM_COMP, BASIS_DIM, BASIS_P, BASIS_Q, BASIS_T_1D>(data, r_U, s_G, r_V);  in __launch_bounds__()
144   __shared__ CeedScalar s_G[BASIS_P * BASIS_Q * BASIS_DIM];  in __launch_bounds__()
145   LoadMatrix<BASIS_P, BASIS_Q * BASIS_DIM>(data, c_G, s_G);  in __launch_bounds__()
151 …GradTransposeNonTensor<BASIS_NUM_COMP, BASIS_DIM, BASIS_P, BASIS_Q, BASIS_T_1D>(data, r_U, s_G, r_…  in __launch_bounds__()
152 …WriteElementStrided1d<BASIS_NUM_COMP, BASIS_P>(data, elem, 1, BASIS_P * num_elem, BASIS_P, r_V, d_…  in __launch_bounds__()
171   __shared__ CeedScalar s_G[BASIS_P * BASIS_Q * BASIS_DIM];  in __launch_bounds__()
172   LoadMatrix<BASIS_P, BASIS_Q * BASIS_DIM>(data, c_G, s_G);  in __launch_bounds__()
178 …GradTransposeNonTensor<BASIS_NUM_COMP, BASIS_DIM, BASIS_P, BASIS_Q, BASIS_T_1D>(data, r_U, s_G, r_…  in __launch_bounds__()
179 …SumElementStrided1d<BASIS_NUM_COMP, BASIS_P>(data, elem, 1, BASIS_P * num_elem, BASIS_P, r_V, d_V);  in __launch_bounds__()
200     WeightNonTensor<BASIS_P, BASIS_Q>(data, q_weight, r_W);  in __launch_bounds__()