59 template <int NUM_COMP, int P_1D>
61   const CeedInt target_comp = n / P_1D;  in SetEVecStandard1d_Single()
62   const CeedInt target_node = n % P_1D;  in SetEVecStandard1d_Single()
72 template <int NUM_COMP, int COMP_STRIDE, int P_1D>
75   if (data.t_id_x < P_1D) {  in ReadLVecStandard1d()
77     const CeedInt ind  = indices[node + elem * P_1D];  in ReadLVecStandard1d()
86 template <int NUM_COMP, int P_1D, int STRIDES_NODE, int STRIDES_COMP, int STRIDES_ELEM>
88   if (data.t_id_x < P_1D) {  in ReadLVecStrided1d()
99 template <int NUM_COMP, int COMP_STRIDE, int P_1D>
102   if (data.t_id_x < P_1D) {  in WriteLVecStandard1d()
104     const CeedInt ind  = indices[node + elem * P_1D];  in WriteLVecStandard1d()
110 template <int NUM_COMP, int COMP_STRIDE, int P_1D>
114   const CeedInt target_comp = n / P_1D;  in WriteLVecStandard1d_Single()
115   const CeedInt target_node = n % P_1D;  in WriteLVecStandard1d_Single()
118     const CeedInt ind = indices[target_node + elem * P_1D];  in WriteLVecStandard1d_Single()
127 template <int NUM_COMP, int COMP_STRIDE, int P_1D>
130   const CeedInt in_comp    = in / P_1D;  in WriteLVecStandard1d_Assembly()
131   const CeedInt in_node    = in % P_1D;  in WriteLVecStandard1d_Assembly()
132   const CeedInt e_vec_size = P_1D * NUM_COMP;  in WriteLVecStandard1d_Assembly()
134   if (data.t_id_x < P_1D) {  in WriteLVecStandard1d_Assembly()
138 …lem * e_vec_size * e_vec_size + (in_comp * NUM_COMP + comp) * P_1D * P_1D + out_node * P_1D + in_n…  in WriteLVecStandard1d_Assembly()
161 template <int NUM_COMP, int P_1D, int STRIDES_NODE, int STRIDES_COMP, int STRIDES_ELEM>
164   if (data.t_id_x < P_1D) {  in WriteLVecStrided1d()
179 template <int NUM_COMP, int P_1D>
181   const CeedInt target_comp   = n / (P_1D * P_1D);  in SetEVecStandard2d_Single()
182   const CeedInt target_node_x = n % P_1D;  in SetEVecStandard2d_Single()
183   const CeedInt target_node_y = (n % (P_1D * P_1D)) / P_1D;  in SetEVecStandard2d_Single()
193 template <int NUM_COMP, int COMP_STRIDE, int P_1D>
196   if (data.t_id_x < P_1D && data.t_id_y < P_1D) {  in ReadLVecStandard2d()
197     const CeedInt node = data.t_id_x + data.t_id_y * P_1D;  in ReadLVecStandard2d()
198     const CeedInt ind  = indices[node + elem * P_1D * P_1D];  in ReadLVecStandard2d()
207 template <int NUM_COMP, int P_1D, int STRIDES_NODE, int STRIDES_COMP, int STRIDES_ELEM>
209   if (data.t_id_x < P_1D && data.t_id_y < P_1D) {  in ReadLVecStrided2d()
210     const CeedInt node = data.t_id_x + data.t_id_y * P_1D;  in ReadLVecStrided2d()
220 template <int NUM_COMP, int COMP_STRIDE, int P_1D>
223   if (data.t_id_x < P_1D && data.t_id_y < P_1D) {  in WriteLVecStandard2d()
224     const CeedInt node = data.t_id_x + data.t_id_y * P_1D;  in WriteLVecStandard2d()
225     const CeedInt ind  = indices[node + elem * P_1D * P_1D];  in WriteLVecStandard2d()
231 template <int NUM_COMP, int COMP_STRIDE, int P_1D>
235   const CeedInt target_comp   = n / (P_1D * P_1D);  in WriteLVecStandard2d_Single()
236   const CeedInt target_node_x = n % P_1D;  in WriteLVecStandard2d_Single()
237   const CeedInt target_node_y = (n % (P_1D * P_1D)) / P_1D;  in WriteLVecStandard2d_Single()
240     const CeedInt node = data.t_id_x + data.t_id_y * P_1D;  in WriteLVecStandard2d_Single()
241     const CeedInt ind  = indices[node + elem * P_1D * P_1D];  in WriteLVecStandard2d_Single()
250 template <int NUM_COMP, int COMP_STRIDE, int P_1D>
253   const CeedInt elem_size  = P_1D * P_1D;  in WriteLVecStandard2d_Assembly()
255   const CeedInt in_node_x  = in % P_1D;  in WriteLVecStandard2d_Assembly()
256   const CeedInt in_node_y  = (in % elem_size) / P_1D;  in WriteLVecStandard2d_Assembly()
259   if (data.t_id_x < P_1D && data.t_id_y < P_1D) {  in WriteLVecStandard2d_Assembly()
260     const CeedInt in_node  = in_node_x + in_node_y * P_1D;  in WriteLVecStandard2d_Assembly()
261     const CeedInt out_node = data.t_id_x + data.t_id_y * P_1D;  in WriteLVecStandard2d_Assembly()
289 template <int NUM_COMP, int P_1D, int STRIDES_NODE, int STRIDES_COMP, int STRIDES_ELEM>
292   if (data.t_id_x < P_1D && data.t_id_y < P_1D) {  in WriteLVecStrided2d()
293     const CeedInt node = data.t_id_x + data.t_id_y * P_1D;  in WriteLVecStrided2d()
307 template <int NUM_COMP, int P_1D>
309   const CeedInt target_comp   = n / (P_1D * P_1D * P_1D);  in SetEVecStandard3d_Single()
310   const CeedInt target_node_x = n % P_1D;  in SetEVecStandard3d_Single()
311   const CeedInt target_node_y = ((n % (P_1D * P_1D * P_1D)) / P_1D) % P_1D;  in SetEVecStandard3d_Single()
312   const CeedInt target_node_z = (n % (P_1D * P_1D * P_1D)) / (P_1D * P_1D);  in SetEVecStandard3d_Single()
315     r_v[target_node_z + target_comp * P_1D] = value;  in SetEVecStandard3d_Single()
322 template <int NUM_COMP, int COMP_STRIDE, int P_1D>
325   if (data.t_id_x < P_1D && data.t_id_y < P_1D) {  in ReadLVecStandard3d()
326     for (CeedInt z = 0; z < P_1D; z++) {  in ReadLVecStandard3d()
327       const CeedInt node = data.t_id_x + data.t_id_y * P_1D + z * P_1D * P_1D;  in ReadLVecStandard3d()
328       const CeedInt ind  = indices[node + elem * P_1D * P_1D * P_1D];  in ReadLVecStandard3d()
330 …for (CeedInt comp = 0; comp < NUM_COMP; comp++) r_u[z + comp * P_1D] = d_u[ind + COMP_STRIDE * com…  in ReadLVecStandard3d()
338 template <int NUM_COMP, int P_1D, int STRIDES_NODE, int STRIDES_COMP, int STRIDES_ELEM>
340   if (data.t_id_x < P_1D && data.t_id_y < P_1D) {  in ReadLVecStrided3d()
341     for (CeedInt z = 0; z < P_1D; z++) {  in ReadLVecStrided3d()
342       const CeedInt node = data.t_id_x + data.t_id_y * P_1D + z * P_1D * P_1D;  in ReadLVecStrided3d()
345 …for (CeedInt comp = 0; comp < NUM_COMP; comp++) r_u[z + comp * P_1D] = d_u[ind + comp * STRIDES_CO…  in ReadLVecStrided3d()
382 template <int NUM_COMP, int COMP_STRIDE, int P_1D>
385   if (data.t_id_x < P_1D && data.t_id_y < P_1D) {  in WriteLVecStandard3d()
386     for (CeedInt z = 0; z < P_1D; z++) {  in WriteLVecStandard3d()
387       const CeedInt node = data.t_id_x + data.t_id_y * P_1D + z * P_1D * P_1D;  in WriteLVecStandard3d()
388       const CeedInt ind  = indices[node + elem * P_1D * P_1D * P_1D];  in WriteLVecStandard3d()
390 …comp = 0; comp < NUM_COMP; comp++) atomicAdd(&d_v[ind + COMP_STRIDE * comp], r_v[z + comp * P_1D]);  in WriteLVecStandard3d()
395 template <int NUM_COMP, int COMP_STRIDE, int P_1D>
399   const CeedInt target_comp   = n / (P_1D * P_1D * P_1D);  in WriteLVecStandard3d_Single()
400   const CeedInt target_node_x = n % P_1D;  in WriteLVecStandard3d_Single()
401   const CeedInt target_node_y = ((n % (P_1D * P_1D * P_1D)) / P_1D) % P_1D;  in WriteLVecStandard3d_Single()
402   const CeedInt target_node_z = (n % (P_1D * P_1D * P_1D)) / (P_1D * P_1D);  in WriteLVecStandard3d_Single()
405     const CeedInt node = data.t_id_x + data.t_id_y * P_1D + target_node_z * P_1D * P_1D;  in WriteLVecStandard3d_Single()
406     const CeedInt ind  = indices[node + elem * P_1D * P_1D * P_1D];  in WriteLVecStandard3d_Single()
408     atomicAdd(&d_v[ind + COMP_STRIDE * target_comp], r_v[target_node_z + target_comp * P_1D]);  in WriteLVecStandard3d_Single()
415 template <int NUM_COMP, int COMP_STRIDE, int P_1D>
418   const CeedInt elem_size  = P_1D * P_1D * P_1D;  in WriteLVecStandard3d_Assembly()
420   const CeedInt in_node_x  = in % P_1D;  in WriteLVecStandard3d_Assembly()
421   const CeedInt in_node_y  = (in % (P_1D * P_1D)) / P_1D;  in WriteLVecStandard3d_Assembly()
422   const CeedInt in_node_z  = (in % elem_size) / (P_1D * P_1D);  in WriteLVecStandard3d_Assembly()
425   if (data.t_id_x < P_1D && data.t_id_y < P_1D) {  in WriteLVecStandard3d_Assembly()
426     const CeedInt in_node = in_node_x + in_node_y * P_1D + in_node_z * P_1D * P_1D;  in WriteLVecStandard3d_Assembly()
427     for (CeedInt z = 0; z < P_1D; z++) {  in WriteLVecStandard3d_Assembly()
428       const CeedInt out_node = data.t_id_x + data.t_id_y * P_1D + z * P_1D * P_1D;  in WriteLVecStandard3d_Assembly()
433         d_v[elem * e_vec_size * e_vec_size + index] += r_v[z + comp * P_1D];  in WriteLVecStandard3d_Assembly()
459 template <int NUM_COMP, int P_1D, int STRIDES_NODE, int STRIDES_COMP, int STRIDES_ELEM>
462   if (data.t_id_x < P_1D && data.t_id_y < P_1D) {  in WriteLVecStrided3d()
463     for (CeedInt z = 0; z < P_1D; z++) {  in WriteLVecStrided3d()
464       const CeedInt node = data.t_id_x + data.t_id_y * P_1D + z * P_1D * P_1D;  in WriteLVecStrided3d()
467 …(CeedInt comp = 0; comp < NUM_COMP; comp++) d_v[ind + comp * STRIDES_COMP] += r_v[z + comp * P_1D];  in WriteLVecStrided3d()