291 …ned(SharedData_Cuda &data, const int t_id_x, const int t_id_y, const int t_id_z, const CeedScalar …  in ContractX3dFlattened()  argument
294   data.slice[t_id_x + t_id_y * T_1D + t_id_z * T_1D * T_1D] = *U;  in ContractX3dFlattened()
297   if (t_id_x < Q_1D && t_id_y < P_1D && t_id_z < P_1D) {  in ContractX3dFlattened()
299 …*V += B[i + t_id_x * P_1D] * data.slice[i + t_id_y * T_1D + t_id_z * T_1D * T_1D];  // Contract x …  in ContractX3dFlattened()
308 …ned(SharedData_Cuda &data, const int t_id_x, const int t_id_y, const int t_id_z, const CeedScalar …  in ContractY3dFlattened()  argument
311   data.slice[t_id_x + t_id_y * T_1D + t_id_z * T_1D * T_1D] = *U;  in ContractY3dFlattened()
314   if (t_id_x < Q_1D && t_id_y < Q_1D && t_id_z < P_1D) {  in ContractY3dFlattened()
316 …*V += B[i + t_id_y * P_1D] * data.slice[t_id_x + i * T_1D + t_id_z * T_1D * T_1D];  // Contract y …  in ContractY3dFlattened()
325 …ned(SharedData_Cuda &data, const int t_id_x, const int t_id_y, const int t_id_z, const CeedScalar …  in ContractZ3dFlattened()  argument
328   data.slice[t_id_x + t_id_y * T_1D + t_id_z * T_1D * T_1D] = *U;  in ContractZ3dFlattened()
331   if (t_id_x < Q_1D && t_id_y < Q_1D && t_id_z < Q_1D) {  in ContractZ3dFlattened()
333 …*V += B[i + t_id_z * P_1D] * data.slice[t_id_x + t_id_y * T_1D + i * T_1D * T_1D];  // Contract z …  in ContractZ3dFlattened()
342 …ned(SharedData_Cuda &data, const int t_id_x, const int t_id_y, const int t_id_z, const CeedScalar …  in ContractTransposeZ3dFlattened()  argument
345   data.slice[t_id_x + t_id_y * T_1D + t_id_z * T_1D * T_1D] = *U;  in ContractTransposeZ3dFlattened()
348   if (t_id_x < Q_1D && t_id_y < Q_1D && t_id_z < P_1D) {  in ContractTransposeZ3dFlattened()
350 …*V += B[t_id_z + i * P_1D] * data.slice[t_id_x + t_id_y * T_1D + i * T_1D * T_1D];  // Contract z …  in ContractTransposeZ3dFlattened()
359 …ansposeAddZ3dFlattened(SharedData_Cuda &data, const int t_id_x, const int t_id_y, const int t_id_z,  in ContractTransposeAddZ3dFlattened()  argument
362   data.slice[t_id_x + t_id_y * T_1D + t_id_z * T_1D * T_1D] = *U;  in ContractTransposeAddZ3dFlattened()
364   if (t_id_x < Q_1D && t_id_y < Q_1D && t_id_z < P_1D) {  in ContractTransposeAddZ3dFlattened()
366 …*V += B[t_id_z + i * P_1D] * data.slice[t_id_x + t_id_y * T_1D + i * T_1D * T_1D];  // Contract z …  in ContractTransposeAddZ3dFlattened()
375 …ned(SharedData_Cuda &data, const int t_id_x, const int t_id_y, const int t_id_z, const CeedScalar …  in ContractTransposeY3dFlattened()  argument
378   data.slice[t_id_x + t_id_y * T_1D + t_id_z * T_1D * T_1D] = *U;  in ContractTransposeY3dFlattened()
381   if (t_id_x < Q_1D && t_id_y < P_1D && t_id_z < P_1D) {  in ContractTransposeY3dFlattened()
383 …*V += B[t_id_y + i * P_1D] * data.slice[t_id_x + i * T_1D + t_id_z * T_1D * T_1D];  // Contract y …  in ContractTransposeY3dFlattened()
392 …ansposeAddY3dFlattened(SharedData_Cuda &data, const int t_id_x, const int t_id_y, const int t_id_z,  in ContractTransposeAddY3dFlattened()  argument
395   data.slice[t_id_x + t_id_y * T_1D + t_id_z * T_1D * T_1D] = *U;  in ContractTransposeAddY3dFlattened()
397   if (t_id_x < Q_1D && t_id_y < P_1D && t_id_z < P_1D) {  in ContractTransposeAddY3dFlattened()
399 …*V += B[t_id_y + i * P_1D] * data.slice[t_id_x + i * T_1D + t_id_z * T_1D * T_1D];  // Contract y …  in ContractTransposeAddY3dFlattened()
408 …ned(SharedData_Cuda &data, const int t_id_x, const int t_id_y, const int t_id_z, const CeedScalar …  in ContractTransposeX3dFlattened()  argument
411   data.slice[t_id_x + t_id_y * T_1D + t_id_z * T_1D * T_1D] = *U;  in ContractTransposeX3dFlattened()
414   if (t_id_x < P_1D && t_id_y < P_1D && t_id_z < P_1D) {  in ContractTransposeX3dFlattened()
416 …*V += B[t_id_x + i * P_1D] * data.slice[i + t_id_y * T_1D + t_id_z * T_1D * T_1D];  // Contract x …  in ContractTransposeX3dFlattened()
425 …ansposeAddX3dFlattened(SharedData_Cuda &data, const int t_id_x, const int t_id_y, const int t_id_z,  in ContractTransposeAddX3dFlattened()  argument
428   data.slice[t_id_x + t_id_y * T_1D + t_id_z * T_1D * T_1D] = *U;  in ContractTransposeAddX3dFlattened()
430   if (t_id_x < P_1D && t_id_y < P_1D && t_id_z < P_1D) {  in ContractTransposeAddX3dFlattened()
432 …*V += B[t_id_x + i * P_1D] * data.slice[i + t_id_y * T_1D + t_id_z * T_1D * T_1D];  // Contract x …  in ContractTransposeAddX3dFlattened()
441 …ack3d(SharedData_Cuda &data, const int t_id_x, const int t_id_y, const int t_id_z, CeedScalar *U) {  in QPack3d()  argument
446 …if (t_id_x < Q_1D && t_id_y < Q_1D) data.slice[t_id_x + t_id_y * T_1D + t_id_z * T_1D * T_1D] = U[…  in QPack3d()
453 …ack3d(SharedData_Cuda &data, const int t_id_x, const int t_id_y, const int t_id_z, CeedScalar *U) {  in QUnpack3d()  argument
460 …U[comp] = (t_id_x < Q_1D && t_id_y < Q_1D) ? data.slice[t_id_x + t_id_y * T_1D + t_id_z * T_1D * T…  in QUnpack3d()
470 …const CeedInt t_id_x = data.t_id_x % T_1D, t_id_y = (data.t_id_x / T_1D) % T_1D, t_id_z = data.t_i…  in InterpTensor3dFlattened()  local
473   if (P_1D != T_1D) QUnpack3d<NUM_COMP, P_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U);  in InterpTensor3dFlattened()
475 …ContractX3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, &r_U[comp], c_B, r_…  in InterpTensor3dFlattened()
476     ContractY3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_t1, c_B, r_t2);  in InterpTensor3dFlattened()
477 …ContractZ3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_t2, c_B, &r_V[com…  in InterpTensor3dFlattened()
480   if (P_1D != T_1D) QPack3d<NUM_COMP, P_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U);  in InterpTensor3dFlattened()
481   if (Q_1D != T_1D) QPack3d<NUM_COMP, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_V);  in InterpTensor3dFlattened()
490 …const CeedInt t_id_x = data.t_id_x % T_1D, t_id_y = (data.t_id_x / T_1D) % T_1D, t_id_z = data.t_i…  in InterpTransposeTensor3dFlattened()  local
493   if (Q_1D != T_1D) QUnpack3d<NUM_COMP, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U);  in InterpTransposeTensor3dFlattened()
495 …ContractTransposeZ3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, &r_U[comp]…  in InterpTransposeTensor3dFlattened()
496 …ContractTransposeY3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_t1, c_B,…  in InterpTransposeTensor3dFlattened()
497 …ContractTransposeX3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_t2, c_B,…  in InterpTransposeTensor3dFlattened()
500   if (P_1D != T_1D) QPack3d<NUM_COMP, P_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_V);  in InterpTransposeTensor3dFlattened()
509 …const CeedInt t_id_x = data.t_id_x % T_1D, t_id_y = (data.t_id_x / T_1D) % T_1D, t_id_z = data.t_i…  in InterpTensorCollocatedNodes3dFlattened()  local
511   if (P_1D != T_1D) QUnpack3d<NUM_COMP, P_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U);  in InterpTensorCollocatedNodes3dFlattened()
516   if (P_1D != T_1D) QPack3d<NUM_COMP, P_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U);  in InterpTensorCollocatedNodes3dFlattened()
517   if (Q_1D != T_1D) QPack3d<NUM_COMP, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_V);  in InterpTensorCollocatedNodes3dFlattened()
526 …const CeedInt t_id_x = data.t_id_x % T_1D, t_id_y = (data.t_id_x / T_1D) % T_1D, t_id_z = data.t_i…  in InterpTransposeTensorCollocatedNodes3dFlattened()  local
528   if (Q_1D != T_1D) QUnpack3d<NUM_COMP, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U);  in InterpTransposeTensorCollocatedNodes3dFlattened()
533   if (P_1D != T_1D) QPack3d<NUM_COMP, P_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_V);  in InterpTransposeTensorCollocatedNodes3dFlattened()
542 …const CeedInt t_id_x = data.t_id_x % T_1D, t_id_y = (data.t_id_x / T_1D) % T_1D, t_id_z = data.t_i…  in GradTensor3dFlattened()  local
545   if (P_1D != T_1D) QUnpack3d<NUM_COMP, P_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U);  in GradTensor3dFlattened()
547 …ContractX3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, &r_U[comp], c_G, r_…  in GradTensor3dFlattened()
548     ContractY3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_t1, c_B, r_t2);  in GradTensor3dFlattened()
549 …ContractZ3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_t2, c_B, &r_V[com…  in GradTensor3dFlattened()
550 …ContractX3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, &r_U[comp], c_B, r_…  in GradTensor3dFlattened()
551     ContractY3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_t1, c_G, r_t2);  in GradTensor3dFlattened()
552 …ContractZ3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_t2, c_B, &r_V[com…  in GradTensor3dFlattened()
553 …ContractX3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, &r_U[comp], c_B, r_…  in GradTensor3dFlattened()
554     ContractY3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_t1, c_B, r_t2);  in GradTensor3dFlattened()
555 …ContractZ3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_t2, c_G, &r_V[com…  in GradTensor3dFlattened()
558   if (P_1D != T_1D) QPack3d<NUM_COMP, P_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U);  in GradTensor3dFlattened()
559   if (Q_1D != T_1D) QPack3d<NUM_COMP * 3, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_V);  in GradTensor3dFlattened()
568 …const CeedInt t_id_x = data.t_id_x % T_1D, t_id_y = (data.t_id_x / T_1D) % T_1D, t_id_z = data.t_i…  in GradTransposeTensor3dFlattened()  local
571   if (Q_1D != T_1D) QUnpack3d<NUM_COMP * 3, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U);  in GradTransposeTensor3dFlattened()
573 …ContractTransposeZ3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, &r_U[comp …  in GradTransposeTensor3dFlattened()
574 …ContractTransposeY3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, r_t1, c_B,…  in GradTransposeTensor3dFlattened()
575 …ContractTransposeX3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, r_t2, c_G,…  in GradTransposeTensor3dFlattened()
576 …ContractTransposeZ3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, &r_U[comp …  in GradTransposeTensor3dFlattened()
577 …ContractTransposeY3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, r_t1, c_G,…  in GradTransposeTensor3dFlattened()
578 …ContractTransposeAddX3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, r_t2, c…  in GradTransposeTensor3dFlattened()
579 …ContractTransposeZ3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, &r_U[comp …  in GradTransposeTensor3dFlattened()
580 …ContractTransposeY3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, r_t1, c_B,…  in GradTransposeTensor3dFlattened()
581 …ContractTransposeAddX3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, r_t2, c…  in GradTransposeTensor3dFlattened()
584   if (P_1D != T_1D) QPack3d<NUM_COMP, P_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_V);  in GradTransposeTensor3dFlattened()
593 …const CeedInt t_id_x = data.t_id_x % T_1D, t_id_y = (data.t_id_x / T_1D) % T_1D, t_id_z = data.t_i…  in GradTensorCollocated3dFlattened()  local
596   if (P_1D != T_1D) QUnpack3d<NUM_COMP, P_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U);  in GradTensorCollocated3dFlattened()
598 …ContractX3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, &r_U[comp], c_B, r_…  in GradTensorCollocated3dFlattened()
599     ContractY3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_t1, c_B, r_t2);  in GradTensorCollocated3dFlattened()
600     ContractZ3dFlattened<NUM_COMP, P_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_t2, c_B, r_t1);  in GradTensorCollocated3dFlattened()
601 …ContractX3dFlattened<NUM_COMP, Q_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_t1, c_G, &r_V[com…  in GradTensorCollocated3dFlattened()
602 …ContractY3dFlattened<NUM_COMP, Q_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_t1, c_G, &r_V[com…  in GradTensorCollocated3dFlattened()
603 …ContractZ3dFlattened<NUM_COMP, Q_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_t1, c_G, &r_V[com…  in GradTensorCollocated3dFlattened()
606   if (P_1D != T_1D) QPack3d<NUM_COMP, P_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U);  in GradTensorCollocated3dFlattened()
607   if (Q_1D != T_1D) QPack3d<NUM_COMP * 3, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_V);  in GradTensorCollocated3dFlattened()
616 …const CeedInt t_id_x = data.t_id_x % T_1D, t_id_y = (data.t_id_x / T_1D) % T_1D, t_id_z = data.t_i…  in GradTransposeTensor3dFlattened()  local
619   if (Q_1D != T_1D) QUnpack3d<NUM_COMP * 3, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U);  in GradTransposeTensor3dFlattened()
621 …ContractTransposeZ3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, &r_U[comp …  in GradTransposeTensor3dFlattened()
622 …ContractTransposeY3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, r_t1, c_B,…  in GradTransposeTensor3dFlattened()
623 …ContractTransposeX3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, r_t2, c_G,…  in GradTransposeTensor3dFlattened()
624 …ContractTransposeZ3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, &r_U[comp …  in GradTransposeTensor3dFlattened()
625 …ContractTransposeY3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, r_t1, c_G,…  in GradTransposeTensor3dFlattened()
626 …ContractTransposeAddX3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, r_t2, c…  in GradTransposeTensor3dFlattened()
627 …ContractTransposeZ3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, &r_U[comp …  in GradTransposeTensor3dFlattened()
628 …ContractTransposeY3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, r_t1, c_B,…  in GradTransposeTensor3dFlattened()
629 …ContractTransposeAddX3dFlattened<NUM_COMP, t_id_x, t_id_y, t_id_z, P_1D, Q_1D, T_1D>(data, r_t2, c…  in GradTransposeTensor3dFlattened()
632   if (P_1D != T_1D) QPack3d<NUM_COMP, P_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_V);  in GradTransposeTensor3dFlattened()
641 …const CeedInt t_id_x = data.t_id_x % T_1D, t_id_y = (data.t_id_x / T_1D) % T_1D, t_id_z = data.t_i…  in GradTensorCollocatedNodes3dFlattened()  local
643   if (P_1D != T_1D) QUnpack3d<NUM_COMP, P_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U);  in GradTensorCollocatedNodes3dFlattened()
645 …ContractX3dFlattened<NUM_COMP, Q_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U[comp], c_G, &r_…  in GradTensorCollocatedNodes3dFlattened()
646 …ContractY3dFlattened<NUM_COMP, Q_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U[comp], c_G, &r_…  in GradTensorCollocatedNodes3dFlattened()
647 …ContractZ3dFlattened<NUM_COMP, Q_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U[comp], c_G, &r_…  in GradTensorCollocatedNodes3dFlattened()
650   if (P_1D != T_1D) QPack3d<NUM_COMP, P_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U);  in GradTensorCollocatedNodes3dFlattened()
651   if (Q_1D != T_1D) QPack3d<NUM_COMP * 3, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_V);  in GradTensorCollocatedNodes3dFlattened()
660 …const CeedInt t_id_x = data.t_id_x % T_1D, t_id_y = (data.t_id_x / T_1D) % T_1D, t_id_z = data.t_i…  in GradTransposeTensorCollocatedNodes3dFlattened()  local
662   if (Q_1D != T_1D) QUnpack3d<NUM_COMP * 3, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_U);  in GradTransposeTensorCollocatedNodes3dFlattened()
664 …ContractTransposeZ3dFlattened<NUM_COMP, Q_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, &r_U[comp …  in GradTransposeTensorCollocatedNodes3dFlattened()
665 …ContractTransposeAddY3dFlattened<NUM_COMP, Q_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, &r_U[co…  in GradTransposeTensorCollocatedNodes3dFlattened()
666 …ContractTransposeAddX3dFlattened<NUM_COMP, Q_1D, Q_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, &r_U[co…  in GradTransposeTensorCollocatedNodes3dFlattened()
669   if (P_1D != T_1D) QPack3d<NUM_COMP, P_1D, T_1D>(data, t_id_x, t_id_y, t_id_z, r_V);  in GradTransposeTensorCollocatedNodes3dFlattened()
677 …const CeedInt t_id_x = data.t_id_x % Q_1D, t_id_y = (data.t_id_x / Q_1D) % Q_1D, t_id_z = data.t_i…  in WeightTensor3dFlattened()  local
679 …= (t_id_x < Q_1D && t_id_y < Q_1D && t_id_z < Q_1D) ? q_weight_1d[t_id_x] * q_weight_1d[t_id_y] * …  in WeightTensor3dFlattened()