105 …CeedScalar rU[1][BASIS_NUM_COMP][BASIS_P] = {0.0};  // for a non-fused operator BASIS_DIM is alway…  in __launch_bounds__()
115   CeedScalar *sTmp = sT + BASIS_P * BASIS_Q;  in __launch_bounds__()
116   sTmp += ty * (max(BASIS_P * BASIS_P * BASIS_MAX_P_Q, BASIS_P * BASIS_Q * BASIS_Q));  in __launch_bounds__()
120     read_T_notrans_gm2sm<BASIS_P, BASIS_Q>(tx, dT, sT);  in __launch_bounds__()
124   read_U_3d<CeedScalar, BASIS_P, 1, BASIS_NUM_COMP, BASIS_P, 0>(dU, cstrdU, rU, sTmp, tx);  in __launch_bounds__()
127 …magma_interp_3d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_P, BASIS_Q, BASIS_P, BASIS_Q>(sT, r…  in __launch_bounds__()
147 …CeedScalar rV[1][BASIS_NUM_COMP][BASIS_P] = {0.0};  // for a non-fused operator BASIS_DIM is alway…  in __launch_bounds__()
148   CeedScalar rTmp[BASIS_P]                  = {0.0};  in __launch_bounds__()
156   CeedScalar *sTmp = sT + BASIS_Q * BASIS_P;  in __launch_bounds__()
157   sTmp += ty * (max(BASIS_Q * BASIS_Q * BASIS_MAX_P_Q, BASIS_Q * BASIS_P * BASIS_P));  in __launch_bounds__()
161     read_T_trans_gm2sm<BASIS_Q, BASIS_P>(tx, dT, sT);  in __launch_bounds__()
168 …magma_interp_3d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_Q, BASIS_P, BASIS_Q, BASIS_P>(sT, r…  in __launch_bounds__()
172   write_V_3d<CeedScalar, BASIS_P, 1, BASIS_NUM_COMP, BASIS_P, 0>(dV, cstrdV, rV, tx);  in __launch_bounds__()
188 …CeedScalar rV[1][BASIS_NUM_COMP][BASIS_P] = {0.0};  // for a non-fused operator BASIS_DIM is alway…  in __launch_bounds__()
189   CeedScalar rTmp[BASIS_P]                  = {0.0};  in __launch_bounds__()
197   CeedScalar *sTmp = sT + BASIS_Q * BASIS_P;  in __launch_bounds__()
198   sTmp += ty * (max(BASIS_Q * BASIS_Q * BASIS_MAX_P_Q, BASIS_Q * BASIS_P * BASIS_P));  in __launch_bounds__()
202     read_T_trans_gm2sm<BASIS_Q, BASIS_P>(tx, dT, sT);  in __launch_bounds__()
209 …magma_interp_3d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_Q, BASIS_P, BASIS_Q, BASIS_P>(sT, r…  in __launch_bounds__()
213   sum_V_3d<CeedScalar, BASIS_P, 1, BASIS_NUM_COMP, BASIS_P, 0>(dV, cstrdV, rV, tx);  in __launch_bounds__()