Lines Matching refs:P_1D
19 inline void ContractX1d(const CeedInt P_1D, const CeedInt Q_1D, private const CeedScalar *restrict … in ContractX1d() argument
28 for (CeedInt i = 0; i < P_1D; i++) { in ContractX1d()
29 *V += B[i + item_id_x * P_1D] * scratch[i]; // Contract x direction in ContractX1d()
38 inline void ContractTransposeX1d(const CeedInt P_1D, const CeedInt Q_1D, private const CeedScalar *… in ContractTransposeX1d() argument
46 if (item_id_x < P_1D) { in ContractTransposeX1d()
48 *V += B[item_id_x + i * P_1D] * scratch[i]; // Contract x direction in ContractTransposeX1d()
57 inline void Interp1d(const CeedInt NUM_COMP, const CeedInt P_1D, const CeedInt Q_1D, private const … in Interp1d() argument
60 ContractX1d(P_1D, Q_1D, r_U + comp, s_B, r_V + comp, scratch); in Interp1d()
67 inline void InterpTranspose1d(const CeedInt NUM_COMP, const CeedInt P_1D, const CeedInt Q_1D, priva… in InterpTranspose1d() argument
70 ContractTransposeX1d(P_1D, Q_1D, r_U + comp, s_B, r_V + comp, scratch); in InterpTranspose1d()
77 inline void Grad1d(const CeedInt NUM_COMP, const CeedInt P_1D, const CeedInt Q_1D, private const Ce… in Grad1d() argument
80 ContractX1d(P_1D, Q_1D, r_U + comp, s_G, r_V + comp, scratch); in Grad1d()
87 inline void GradTranspose1d(const CeedInt NUM_COMP, const CeedInt P_1D, const CeedInt Q_1D, private… in GradTranspose1d() argument
90 ContractTransposeX1d(P_1D, Q_1D, r_U + comp, s_G, r_V + comp, scratch); in GradTranspose1d()
109 inline void ContractX2d(const CeedInt P_1D, const CeedInt Q_1D, private const CeedScalar *restrict … in ContractX2d() argument
118 if (item_id_x < Q_1D && item_id_y < P_1D) { in ContractX2d()
119 for (CeedInt i = 0; i < P_1D; i++) { in ContractX2d()
120 *V += B[i + item_id_x * P_1D] * scratch[i + item_id_y * T_1D]; // Contract x direction in ContractX2d()
129 inline void ContractY2d(const CeedInt P_1D, const CeedInt Q_1D, private const CeedScalar *restrict … in ContractY2d() argument
139 for (CeedInt i = 0; i < P_1D; i++) { in ContractY2d()
140 *V += B[i + item_id_y * P_1D] * scratch[item_id_x + i * T_1D]; // Contract y direction in ContractY2d()
149 inline void ContractTransposeY2d(const CeedInt P_1D, const CeedInt Q_1D, private const CeedScalar *… in ContractTransposeY2d() argument
158 if (item_id_x < Q_1D && item_id_y < P_1D) { in ContractTransposeY2d()
160 *V += B[item_id_y + i * P_1D] * scratch[item_id_x + i * T_1D]; // Contract y direction in ContractTransposeY2d()
169 inline void ContractTransposeX2d(const CeedInt P_1D, const CeedInt Q_1D, private const CeedScalar *… in ContractTransposeX2d() argument
178 if (item_id_x < P_1D && item_id_y < P_1D) { in ContractTransposeX2d()
180 *V += B[item_id_x + i * P_1D] * scratch[i + item_id_y * T_1D]; // Contract x direction in ContractTransposeX2d()
189 inline void ContractTransposeAddX2d(const CeedInt P_1D, const CeedInt Q_1D, private const CeedScala… in ContractTransposeAddX2d() argument
197 if (item_id_x < P_1D && item_id_y < P_1D) { in ContractTransposeAddX2d()
199 *V += B[item_id_x + i * P_1D] * scratch[i + item_id_y * T_1D]; // Contract x direction in ContractTransposeAddX2d()
208 inline void InterpTensor2d(const CeedInt NUM_COMP, const CeedInt P_1D, const CeedInt Q_1D, private … in InterpTensor2d() argument
213 ContractX2d(P_1D, Q_1D, r_U + comp, s_B, r_t, scratch); in InterpTensor2d()
214 ContractY2d(P_1D, Q_1D, r_t, s_B, r_V + comp, scratch); in InterpTensor2d()
221 inline void InterpTransposeTensor2d(const CeedInt NUM_COMP, const CeedInt P_1D, const CeedInt Q_1D,… in InterpTransposeTensor2d() argument
226 ContractTransposeY2d(P_1D, Q_1D, r_U + comp, s_B, r_t, scratch); in InterpTransposeTensor2d()
227 ContractTransposeX2d(P_1D, Q_1D, r_t, s_B, r_V + comp, scratch); in InterpTransposeTensor2d()
234 inline void GradTensor2d(const CeedInt NUM_COMP, const CeedInt P_1D, const CeedInt Q_1D, private co… in GradTensor2d() argument
240 ContractX2d(P_1D, Q_1D, r_U + comp, s_G, r_t, scratch); in GradTensor2d()
241 ContractY2d(P_1D, Q_1D, r_t, s_B, r_V + comp + 0 * NUM_COMP, scratch); in GradTensor2d()
242 ContractX2d(P_1D, Q_1D, r_U + comp, s_B, r_t, scratch); in GradTensor2d()
243 ContractY2d(P_1D, Q_1D, r_t, s_G, r_V + comp + 1 * NUM_COMP, scratch); in GradTensor2d()
250 inline void GradTransposeTensor2d(const CeedInt NUM_COMP, const CeedInt P_1D, const CeedInt Q_1D, p… in GradTransposeTensor2d() argument
256 ContractTransposeY2d(P_1D, Q_1D, r_U + comp + 0 * NUM_COMP, s_B, r_t, scratch); in GradTransposeTensor2d()
257 ContractTransposeX2d(P_1D, Q_1D, r_t, s_G, r_V + comp, scratch); in GradTransposeTensor2d()
258 ContractTransposeY2d(P_1D, Q_1D, r_U + comp + 1 * NUM_COMP, s_G, r_t, scratch); in GradTransposeTensor2d()
259 ContractTransposeAddX2d(P_1D, Q_1D, r_t, s_B, r_V + comp, scratch); in GradTransposeTensor2d()
280 inline void ContractX3d(const CeedInt P_1D, const CeedInt Q_1D, private const CeedScalar *restrict … in ContractX3d() argument
286 for (CeedInt i = 0; i < P_1D; i++) { in ContractX3d()
287 r_B[i] = B[i + item_id_x * P_1D]; in ContractX3d()
290 for (CeedInt k = 0; k < P_1D; k++) { in ContractX3d()
295 if (item_id_x < Q_1D && item_id_y < P_1D) { in ContractX3d()
296 for (CeedInt i = 0; i < P_1D; i++) { in ContractX3d()
307 inline void ContractY3d(const CeedInt P_1D, const CeedInt Q_1D, private const CeedScalar *restrict … in ContractY3d() argument
313 for (CeedInt i = 0; i < P_1D; i++) { in ContractY3d()
314 r_B[i] = B[i + item_id_y * P_1D]; in ContractY3d()
317 for (CeedInt k = 0; k < P_1D; k++) { in ContractY3d()
323 for (CeedInt i = 0; i < P_1D; i++) { in ContractY3d()
334 inline void ContractZ3d(const CeedInt P_1D, const CeedInt Q_1D, private const CeedScalar *restrict … in ContractZ3d() argument
342 for (CeedInt i = 0; i < P_1D; i++) { in ContractZ3d()
343 V[k] += B[i + k * P_1D] * U[i]; // Contract z direction in ContractZ3d()
352 inline void ContractTransposeZ3d(const CeedInt P_1D, const CeedInt Q_1D, private const CeedScalar *… in ContractTransposeZ3d() argument
357 for (CeedInt k = 0; k < P_1D; k++) { in ContractTransposeZ3d()
361 V[k] += B[k + i * P_1D] * U[i]; // Contract z direction in ContractTransposeZ3d()
370 inline void ContractTransposeY3d(const CeedInt P_1D, const CeedInt Q_1D, private const CeedScalar *… in ContractTransposeY3d() argument
377 r_B[i] = B[item_id_y + i * P_1D]; in ContractTransposeY3d()
380 for (CeedInt k = 0; k < P_1D; k++) { in ContractTransposeY3d()
385 if (item_id_x < Q_1D && item_id_y < P_1D) { in ContractTransposeY3d()
397 inline void ContractTransposeAddY3d(const CeedInt P_1D, const CeedInt Q_1D, private const CeedScala… in ContractTransposeAddY3d() argument
404 r_B[i] = B[item_id_y + i * P_1D]; in ContractTransposeAddY3d()
407 for (CeedInt k = 0; k < P_1D; k++) { in ContractTransposeAddY3d()
410 if (item_id_x < Q_1D && item_id_y < P_1D) { in ContractTransposeAddY3d()
422 inline void ContractTransposeX3d(const CeedInt P_1D, const CeedInt Q_1D, private const CeedScalar *… in ContractTransposeX3d() argument
429 r_B[i] = B[item_id_x + i * P_1D]; in ContractTransposeX3d()
432 for (CeedInt k = 0; k < P_1D; k++) { in ContractTransposeX3d()
436 if (item_id_x < P_1D && item_id_y < P_1D) { in ContractTransposeX3d()
448 inline void ContractTransposeAddX3d(const CeedInt P_1D, const CeedInt Q_1D, private const CeedScala… in ContractTransposeAddX3d() argument
455 r_B[i] = B[item_id_x + i * P_1D]; in ContractTransposeAddX3d()
458 for (CeedInt k = 0; k < P_1D; k++) { in ContractTransposeAddX3d()
462 if (item_id_x < P_1D && item_id_y < P_1D) { in ContractTransposeAddX3d()
474 inline void InterpTensor3d(const CeedInt NUM_COMP, const CeedInt P_1D, const CeedInt Q_1D, private … in InterpTensor3d() argument
480 ContractX3d(P_1D, Q_1D, r_U + comp * P_1D, s_B, r_t1, scratch); in InterpTensor3d()
481 ContractY3d(P_1D, Q_1D, r_t1, s_B, r_t2, scratch); in InterpTensor3d()
482 ContractZ3d(P_1D, Q_1D, r_t2, s_B, r_V + comp * Q_1D, scratch); in InterpTensor3d()
489 inline void InterpTransposeTensor3d(const CeedInt NUM_COMP, const CeedInt P_1D, const CeedInt Q_1D,… in InterpTransposeTensor3d() argument
495 ContractTransposeZ3d(P_1D, Q_1D, r_U + comp * Q_1D, s_B, r_t1, scratch); in InterpTransposeTensor3d()
496 ContractTransposeY3d(P_1D, Q_1D, r_t1, s_B, r_t2, scratch); in InterpTransposeTensor3d()
497 ContractTransposeX3d(P_1D, Q_1D, r_t2, s_B, r_V + comp * P_1D, scratch); in InterpTransposeTensor3d()
504 inline void GradTensor3d(const CeedInt NUM_COMP, const CeedInt P_1D, const CeedInt Q_1D, private co… in GradTensor3d() argument
511 ContractX3d(P_1D, Q_1D, r_U + comp * P_1D, s_G, r_t1, scratch); in GradTensor3d()
512 ContractY3d(P_1D, Q_1D, r_t1, s_B, r_t2, scratch); in GradTensor3d()
513 ContractZ3d(P_1D, Q_1D, r_t2, s_B, r_V + comp * Q_1D + 0 * NUM_COMP * Q_1D, scratch); in GradTensor3d()
514 ContractX3d(P_1D, Q_1D, r_U + comp * P_1D, s_B, r_t1, scratch); in GradTensor3d()
515 ContractY3d(P_1D, Q_1D, r_t1, s_G, r_t2, scratch); in GradTensor3d()
516 ContractZ3d(P_1D, Q_1D, r_t2, s_B, r_V + comp * Q_1D + 1 * NUM_COMP * Q_1D, scratch); in GradTensor3d()
517 ContractX3d(P_1D, Q_1D, r_U + comp * P_1D, s_B, r_t1, scratch); in GradTensor3d()
518 ContractY3d(P_1D, Q_1D, r_t1, s_B, r_t2, scratch); in GradTensor3d()
519 ContractZ3d(P_1D, Q_1D, r_t2, s_G, r_V + comp * Q_1D + 2 * NUM_COMP * Q_1D, scratch); in GradTensor3d()
526 inline void GradTransposeTensor3d(const CeedInt NUM_COMP, const CeedInt P_1D, const CeedInt Q_1D, p… in GradTransposeTensor3d() argument
533 ContractTransposeZ3d(P_1D, Q_1D, r_U + comp * Q_1D + 0 * NUM_COMP * Q_1D, s_B, r_t1, scratch); in GradTransposeTensor3d()
534 ContractTransposeY3d(P_1D, Q_1D, r_t1, s_B, r_t2, scratch); in GradTransposeTensor3d()
535 ContractTransposeX3d(P_1D, Q_1D, r_t2, s_G, r_V + comp * P_1D, scratch); in GradTransposeTensor3d()
536 ContractTransposeZ3d(P_1D, Q_1D, r_U + comp * Q_1D + 1 * NUM_COMP * Q_1D, s_B, r_t1, scratch); in GradTransposeTensor3d()
537 ContractTransposeY3d(P_1D, Q_1D, r_t1, s_G, r_t2, scratch); in GradTransposeTensor3d()
538 ContractTransposeAddX3d(P_1D, Q_1D, r_t2, s_B, r_V + comp * P_1D, scratch); in GradTransposeTensor3d()
539 ContractTransposeZ3d(P_1D, Q_1D, r_U + comp * Q_1D + 2 * NUM_COMP * Q_1D, s_G, r_t1, scratch); in GradTransposeTensor3d()
540 ContractTransposeY3d(P_1D, Q_1D, r_t1, s_B, r_t2, scratch); in GradTransposeTensor3d()
541 ContractTransposeAddX3d(P_1D, Q_1D, r_t2, s_B, r_V + comp * P_1D, scratch); in GradTransposeTensor3d()
548 inline void GradTensorCollocated3d(const CeedInt NUM_COMP, const CeedInt P_1D, const CeedInt Q_1D, … in GradTensorCollocated3d() argument
555 ContractX3d(P_1D, Q_1D, r_U + comp * P_1D, s_B, r_t1, scratch); in GradTensorCollocated3d()
556 ContractY3d(P_1D, Q_1D, r_t1, s_B, r_t2, scratch); in GradTensorCollocated3d()
557 ContractZ3d(P_1D, Q_1D, r_t2, s_B, r_t1, scratch); in GradTensorCollocated3d()
567 inline void GradTransposeTensorCollocated3d(const CeedInt NUM_COMP, const CeedInt P_1D, const CeedI… in GradTransposeTensorCollocated3d() argument
577 ContractTransposeZ3d(P_1D, Q_1D, r_t2, s_B, r_t1, scratch); in GradTransposeTensorCollocated3d()
578 ContractTransposeY3d(P_1D, Q_1D, r_t1, s_B, r_t2, scratch); in GradTransposeTensorCollocated3d()
579 ContractTransposeX3d(P_1D, Q_1D, r_t2, s_B, r_V + comp * P_1D, scratch); in GradTransposeTensorCollocated3d()