xref: /libCEED/rust/libceed-sys/c-src/backends/blocked/ceed-blocked-operator.c (revision c4e3f59b2ea5a0c95cc0118aa5026c447cce3092)
13d8e8822SJeremy L Thompson // Copyright (c) 2017-2022, Lawrence Livermore National Security, LLC and other CEED contributors.
23d8e8822SJeremy L Thompson // All Rights Reserved. See the top-level LICENSE and NOTICE files for details.
34a2e7687Sjeremylt //
43d8e8822SJeremy L Thompson // SPDX-License-Identifier: BSD-2-Clause
54a2e7687Sjeremylt //
63d8e8822SJeremy L Thompson // This file is part of CEED:  http://github.com/ceed
74a2e7687Sjeremylt 
849aac155SJeremy L Thompson #include <ceed.h>
9ec3da8bcSJed Brown #include <ceed/backend.h>
103d576824SJeremy L Thompson #include <stdbool.h>
113d576824SJeremy L Thompson #include <stddef.h>
123d576824SJeremy L Thompson #include <stdint.h>
132b730f8bSJeremy L Thompson 
144a2e7687Sjeremylt #include "ceed-blocked.h"
154a2e7687Sjeremylt 
16f10650afSjeremylt //------------------------------------------------------------------------------
17f10650afSjeremylt // Setup Input/Output Fields
18f10650afSjeremylt //------------------------------------------------------------------------------
192b730f8bSJeremy L Thompson static int CeedOperatorSetupFields_Blocked(CeedQFunction qf, CeedOperator op, bool is_input, CeedElemRestriction *blk_restr, CeedVector *e_vecs_full,
202b730f8bSJeremy L Thompson                                            CeedVector *e_vecs, CeedVector *q_vecs, CeedInt start_e, CeedInt num_fields, CeedInt Q) {
212b730f8bSJeremy L Thompson   CeedInt  num_comp, size, P;
22d2643443SJeremy L Thompson   CeedSize e_size, q_size;
23aedaa0e5Sjeremylt   Ceed     ceed;
242b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetCeed(op, &ceed));
25d1bcdac9Sjeremylt   CeedBasis           basis;
26d1bcdac9Sjeremylt   CeedElemRestriction r;
27d1d35e2fSjeremylt   CeedOperatorField  *op_fields;
28d1d35e2fSjeremylt   CeedQFunctionField *qf_fields;
294fc1f125SJeremy L Thompson   if (is_input) {
302b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorGetFields(op, NULL, &op_fields, NULL, NULL));
312b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_fields, NULL, NULL));
324fc1f125SJeremy L Thompson   } else {
332b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorGetFields(op, NULL, NULL, NULL, &op_fields));
342b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionGetFields(qf, NULL, NULL, NULL, &qf_fields));
35fe2413ffSjeremylt   }
36d1d35e2fSjeremylt   const CeedInt blk_size = 8;
374a2e7687Sjeremylt 
384a2e7687Sjeremylt   // Loop over fields
394fc1f125SJeremy L Thompson   for (CeedInt i = 0; i < num_fields; i++) {
40d1d35e2fSjeremylt     CeedEvalMode eval_mode;
412b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_fields[i], &eval_mode));
424a2e7687Sjeremylt 
43d1d35e2fSjeremylt     if (eval_mode != CEED_EVAL_WEIGHT) {
442b730f8bSJeremy L Thompson       CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &r));
452b730f8bSJeremy L Thompson       CeedCallBackend(CeedElemRestrictionGetCeed(r, &ceed));
46e79b91d9SJeremy L Thompson       CeedSize l_size;
47e79b91d9SJeremy L Thompson       CeedInt  num_elem, elem_size, comp_stride;
482b730f8bSJeremy L Thompson       CeedCallBackend(CeedElemRestrictionGetNumElements(r, &num_elem));
492b730f8bSJeremy L Thompson       CeedCallBackend(CeedElemRestrictionGetElementSize(r, &elem_size));
502b730f8bSJeremy L Thompson       CeedCallBackend(CeedElemRestrictionGetLVectorSize(r, &l_size));
512b730f8bSJeremy L Thompson       CeedCallBackend(CeedElemRestrictionGetNumComponents(r, &num_comp));
52bd33150aSjeremylt 
533ac43b2cSJeremy L Thompson       bool strided;
542b730f8bSJeremy L Thompson       CeedCallBackend(CeedElemRestrictionIsStrided(r, &strided));
553ac43b2cSJeremy L Thompson       if (strided) {
563ac43b2cSJeremy L Thompson         CeedInt strides[3];
572b730f8bSJeremy L Thompson         CeedCallBackend(CeedElemRestrictionGetStrides(r, &strides));
582b730f8bSJeremy L Thompson         CeedCallBackend(
592b730f8bSJeremy L Thompson             CeedElemRestrictionCreateBlockedStrided(ceed, num_elem, elem_size, blk_size, num_comp, l_size, strides, &blk_restr[i + start_e]));
603ac43b2cSJeremy L Thompson       } else {
61bd33150aSjeremylt         const CeedInt *offsets = NULL;
622b730f8bSJeremy L Thompson         CeedCallBackend(CeedElemRestrictionGetOffsets(r, CEED_MEM_HOST, &offsets));
632b730f8bSJeremy L Thompson         CeedCallBackend(CeedElemRestrictionGetCompStride(r, &comp_stride));
642b730f8bSJeremy L Thompson         CeedCallBackend(CeedElemRestrictionCreateBlocked(ceed, num_elem, elem_size, blk_size, num_comp, comp_stride, l_size, CEED_MEM_HOST,
652b730f8bSJeremy L Thompson                                                          CEED_COPY_VALUES, offsets, &blk_restr[i + start_e]));
662b730f8bSJeremy L Thompson         CeedCallBackend(CeedElemRestrictionRestoreOffsets(r, &offsets));
673ac43b2cSJeremy L Thompson       }
682b730f8bSJeremy L Thompson       CeedCallBackend(CeedElemRestrictionCreateVector(blk_restr[i + start_e], NULL, &e_vecs_full[i + start_e]));
694a2e7687Sjeremylt     }
704a2e7687Sjeremylt 
71d1d35e2fSjeremylt     switch (eval_mode) {
724a2e7687Sjeremylt       case CEED_EVAL_NONE:
732b730f8bSJeremy L Thompson         CeedCallBackend(CeedQFunctionFieldGetSize(qf_fields[i], &size));
74d2643443SJeremy L Thompson         q_size = (CeedSize)Q * size * blk_size;
752b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i]));
76aedaa0e5Sjeremylt         break;
77aedaa0e5Sjeremylt       case CEED_EVAL_INTERP:
784a2e7687Sjeremylt       case CEED_EVAL_GRAD:
79a915a514Srezgarshakeri       case CEED_EVAL_DIV:
80*c4e3f59bSSebastian Grimberg       case CEED_EVAL_CURL:
812b730f8bSJeremy L Thompson         CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis));
822b730f8bSJeremy L Thompson         CeedCallBackend(CeedQFunctionFieldGetSize(qf_fields[i], &size));
832b730f8bSJeremy L Thompson         CeedCallBackend(CeedBasisGetNumNodes(basis, &P));
842b730f8bSJeremy L Thompson         CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp));
85d2643443SJeremy L Thompson         e_size = (CeedSize)P * num_comp * blk_size;
862b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorCreate(ceed, e_size, &e_vecs[i]));
87d2643443SJeremy L Thompson         q_size = (CeedSize)Q * size * blk_size;
882b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i]));
894a2e7687Sjeremylt         break;
904a2e7687Sjeremylt       case CEED_EVAL_WEIGHT:  // Only on input fields
912b730f8bSJeremy L Thompson         CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis));
92d2643443SJeremy L Thompson         q_size = (CeedSize)Q * blk_size;
932b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i]));
942b730f8bSJeremy L Thompson         CeedCallBackend(CeedBasisApply(basis, blk_size, CEED_NOTRANSPOSE, CEED_EVAL_WEIGHT, CEED_VECTOR_NONE, q_vecs[i]));
954a2e7687Sjeremylt         break;
964a2e7687Sjeremylt     }
974a2e7687Sjeremylt   }
98e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
994a2e7687Sjeremylt }
1004a2e7687Sjeremylt 
101f10650afSjeremylt //------------------------------------------------------------------------------
102f10650afSjeremylt // Setup Operator
103f10650afSjeremylt //------------------------------------------------------------------------------
1044a2e7687Sjeremylt static int CeedOperatorSetup_Blocked(CeedOperator op) {
1058c1105f8SJeremy L Thompson   bool is_setup_done;
1062b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorIsSetupDone(op, &is_setup_done));
1078c1105f8SJeremy L Thompson   if (is_setup_done) return CEED_ERROR_SUCCESS;
108aedaa0e5Sjeremylt   Ceed ceed;
1092b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetCeed(op, &ceed));
1104ce2993fSjeremylt   CeedOperator_Blocked *impl;
1112b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetData(op, &impl));
1124ce2993fSjeremylt   CeedQFunction qf;
1132b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetQFunction(op, &qf));
114d1d35e2fSjeremylt   CeedInt Q, num_input_fields, num_output_fields;
1152b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q));
1162b730f8bSJeremy L Thompson   CeedCallBackend(CeedQFunctionIsIdentity(qf, &impl->is_identity_qf));
117d1d35e2fSjeremylt   CeedOperatorField *op_input_fields, *op_output_fields;
1182b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields));
119d1d35e2fSjeremylt   CeedQFunctionField *qf_input_fields, *qf_output_fields;
1202b730f8bSJeremy L Thompson   CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields));
1214a2e7687Sjeremylt 
1224a2e7687Sjeremylt   // Allocate
1232b730f8bSJeremy L Thompson   CeedCallBackend(CeedCalloc(num_input_fields + num_output_fields, &impl->blk_restr));
1242b730f8bSJeremy L Thompson   CeedCallBackend(CeedCalloc(num_input_fields + num_output_fields, &impl->e_vecs_full));
1254a2e7687Sjeremylt 
1262b730f8bSJeremy L Thompson   CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->input_states));
1272b730f8bSJeremy L Thompson   CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->e_vecs_in));
1282b730f8bSJeremy L Thompson   CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->e_vecs_out));
1292b730f8bSJeremy L Thompson   CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->q_vecs_in));
1302b730f8bSJeremy L Thompson   CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->q_vecs_out));
1314a2e7687Sjeremylt 
1324fc1f125SJeremy L Thompson   impl->num_inputs  = num_input_fields;
1334fc1f125SJeremy L Thompson   impl->num_outputs = num_output_fields;
134aedaa0e5Sjeremylt 
1354a2e7687Sjeremylt   // Set up infield and outfield pointer arrays
1364a2e7687Sjeremylt   // Infields
1372b730f8bSJeremy L Thompson   CeedCallBackend(
1382b730f8bSJeremy L Thompson       CeedOperatorSetupFields_Blocked(qf, op, true, impl->blk_restr, impl->e_vecs_full, impl->e_vecs_in, impl->q_vecs_in, 0, num_input_fields, Q));
1394a2e7687Sjeremylt   // Outfields
1402b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorSetupFields_Blocked(qf, op, false, impl->blk_restr, impl->e_vecs_full, impl->e_vecs_out, impl->q_vecs_out,
1412b730f8bSJeremy L Thompson                                                   num_input_fields, num_output_fields, Q));
142aedaa0e5Sjeremylt 
14316911fdaSjeremylt   // Identity QFunctions
1440b454692Sjeremylt   if (impl->is_identity_qf) {
145d1d35e2fSjeremylt     CeedEvalMode        in_mode, out_mode;
146d1d35e2fSjeremylt     CeedQFunctionField *in_fields, *out_fields;
1472b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &in_fields, NULL, &out_fields));
1482b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionFieldGetEvalMode(in_fields[0], &in_mode));
1492b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionFieldGetEvalMode(out_fields[0], &out_mode));
150d1d35e2fSjeremylt 
1510b454692Sjeremylt     if (in_mode == CEED_EVAL_NONE && out_mode == CEED_EVAL_NONE) {
1520b454692Sjeremylt       impl->is_identity_restr_op = true;
1530b454692Sjeremylt     } else {
1542b730f8bSJeremy L Thompson       CeedCallBackend(CeedVectorDestroy(&impl->q_vecs_out[0]));
1550b454692Sjeremylt       impl->q_vecs_out[0] = impl->q_vecs_in[0];
1562b730f8bSJeremy L Thompson       CeedCallBackend(CeedVectorAddReference(impl->q_vecs_in[0]));
15716911fdaSjeremylt     }
15816911fdaSjeremylt   }
15916911fdaSjeremylt 
1602b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorSetSetupDone(op));
1614a2e7687Sjeremylt 
162e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
1634a2e7687Sjeremylt }
1644a2e7687Sjeremylt 
165f10650afSjeremylt //------------------------------------------------------------------------------
166f10650afSjeremylt // Setup Operator Inputs
167f10650afSjeremylt //------------------------------------------------------------------------------
1682b730f8bSJeremy L Thompson static inline int CeedOperatorSetupInputs_Blocked(CeedInt num_input_fields, CeedQFunctionField *qf_input_fields, CeedOperatorField *op_input_fields,
1694fc1f125SJeremy L Thompson                                                   CeedVector in_vec, bool skip_active, CeedScalar *e_data_full[2 * CEED_FIELD_MAX],
170a0162de9SJeremy L Thompson                                                   CeedOperator_Blocked *impl, CeedRequest *request) {
171d1d35e2fSjeremylt   CeedEvalMode eval_mode;
172d1bcdac9Sjeremylt   CeedVector   vec;
17316c359e6Sjeremylt   uint64_t     state;
1744a2e7687Sjeremylt 
175d1d35e2fSjeremylt   for (CeedInt i = 0; i < num_input_fields; i++) {
1761d102b48SJeremy L Thompson     // Get input vector
1772b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec));
1781d102b48SJeremy L Thompson     if (vec == CEED_VECTOR_ACTIVE) {
1792b730f8bSJeremy L Thompson       if (skip_active) continue;
1802b730f8bSJeremy L Thompson       else vec = in_vec;
1811d102b48SJeremy L Thompson     }
1821d102b48SJeremy L Thompson 
1832b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode));
184d1d35e2fSjeremylt     if (eval_mode == CEED_EVAL_WEIGHT) {  // Skip
1854a2e7687Sjeremylt     } else {
1864a2e7687Sjeremylt       // Restrict
1872b730f8bSJeremy L Thompson       CeedCallBackend(CeedVectorGetState(vec, &state));
1884fc1f125SJeremy L Thompson       if (state != impl->input_states[i] || vec == in_vec) {
1892b730f8bSJeremy L Thompson         CeedCallBackend(CeedElemRestrictionApply(impl->blk_restr[i], CEED_NOTRANSPOSE, vec, impl->e_vecs_full[i], request));
1904fc1f125SJeremy L Thompson         impl->input_states[i] = state;
19116c359e6Sjeremylt       }
1924a2e7687Sjeremylt       // Get evec
1932b730f8bSJeremy L Thompson       CeedCallBackend(CeedVectorGetArrayRead(impl->e_vecs_full[i], CEED_MEM_HOST, (const CeedScalar **)&e_data_full[i]));
1944a2e7687Sjeremylt     }
1954a2e7687Sjeremylt   }
196e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
1974a2e7687Sjeremylt }
1984a2e7687Sjeremylt 
199f10650afSjeremylt //------------------------------------------------------------------------------
200f10650afSjeremylt // Input Basis Action
201f10650afSjeremylt //------------------------------------------------------------------------------
2022b730f8bSJeremy L Thompson static inline int CeedOperatorInputBasis_Blocked(CeedInt e, CeedInt Q, CeedQFunctionField *qf_input_fields, CeedOperatorField *op_input_fields,
203d1d35e2fSjeremylt                                                  CeedInt num_input_fields, CeedInt blk_size, bool skip_active,
2044fc1f125SJeremy L Thompson                                                  CeedScalar *e_data_full[2 * CEED_FIELD_MAX], CeedOperator_Blocked *impl) {
205a915a514Srezgarshakeri   CeedInt             elem_size, size, num_comp;
206d1d35e2fSjeremylt   CeedElemRestriction elem_restr;
207d1d35e2fSjeremylt   CeedEvalMode        eval_mode;
2081d102b48SJeremy L Thompson   CeedBasis           basis;
2091d102b48SJeremy L Thompson 
210d1d35e2fSjeremylt   for (CeedInt i = 0; i < num_input_fields; i++) {
2111d102b48SJeremy L Thompson     // Skip active input
212d1d35e2fSjeremylt     if (skip_active) {
2131d102b48SJeremy L Thompson       CeedVector vec;
2142b730f8bSJeremy L Thompson       CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec));
2152b730f8bSJeremy L Thompson       if (vec == CEED_VECTOR_ACTIVE) continue;
2161d102b48SJeremy L Thompson     }
2171d102b48SJeremy L Thompson 
218d1d35e2fSjeremylt     // Get elem_size, eval_mode, size
2192b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &elem_restr));
2202b730f8bSJeremy L Thompson     CeedCallBackend(CeedElemRestrictionGetElementSize(elem_restr, &elem_size));
2212b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode));
2222b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionFieldGetSize(qf_input_fields[i], &size));
2234a2e7687Sjeremylt     // Basis action
224d1d35e2fSjeremylt     switch (eval_mode) {
2254a2e7687Sjeremylt       case CEED_EVAL_NONE:
2262b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorSetArray(impl->q_vecs_in[i], CEED_MEM_HOST, CEED_USE_POINTER, &e_data_full[i][e * Q * size]));
2274a2e7687Sjeremylt         break;
2284a2e7687Sjeremylt       case CEED_EVAL_INTERP:
2294a2e7687Sjeremylt       case CEED_EVAL_GRAD:
230a915a514Srezgarshakeri       case CEED_EVAL_DIV:
231*c4e3f59bSSebastian Grimberg       case CEED_EVAL_CURL:
232a915a514Srezgarshakeri         CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis));
233a915a514Srezgarshakeri         CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp));
234a915a514Srezgarshakeri         CeedCallBackend(CeedVectorSetArray(impl->e_vecs_in[i], CEED_MEM_HOST, CEED_USE_POINTER, &e_data_full[i][e * elem_size * num_comp]));
235*c4e3f59bSSebastian Grimberg         CeedCallBackend(CeedBasisApply(basis, blk_size, CEED_NOTRANSPOSE, eval_mode, impl->e_vecs_in[i], impl->q_vecs_in[i]));
236a915a514Srezgarshakeri         break;
2374a2e7687Sjeremylt       case CEED_EVAL_WEIGHT:
2384a2e7687Sjeremylt         break;  // No action
2394a2e7687Sjeremylt     }
24089c6efa4Sjeremylt   }
241e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
24289c6efa4Sjeremylt }
2434a2e7687Sjeremylt 
244f10650afSjeremylt //------------------------------------------------------------------------------
245f10650afSjeremylt // Output Basis Action
246f10650afSjeremylt //------------------------------------------------------------------------------
2472b730f8bSJeremy L Thompson static inline int CeedOperatorOutputBasis_Blocked(CeedInt e, CeedInt Q, CeedQFunctionField *qf_output_fields, CeedOperatorField *op_output_fields,
2482b730f8bSJeremy L Thompson                                                   CeedInt blk_size, CeedInt num_input_fields, CeedInt num_output_fields, CeedOperator op,
2492b730f8bSJeremy L Thompson                                                   CeedScalar *e_data_full[2 * CEED_FIELD_MAX], CeedOperator_Blocked *impl) {
250a915a514Srezgarshakeri   CeedInt             elem_size, num_comp;
251d1d35e2fSjeremylt   CeedElemRestriction elem_restr;
252d1d35e2fSjeremylt   CeedEvalMode        eval_mode;
2531d102b48SJeremy L Thompson   CeedBasis           basis;
2541d102b48SJeremy L Thompson 
255d1d35e2fSjeremylt   for (CeedInt i = 0; i < num_output_fields; i++) {
256d1d35e2fSjeremylt     // Get elem_size, eval_mode, size
2572b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &elem_restr));
2582b730f8bSJeremy L Thompson     CeedCallBackend(CeedElemRestrictionGetElementSize(elem_restr, &elem_size));
2592b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode));
2604a2e7687Sjeremylt     // Basis action
261d1d35e2fSjeremylt     switch (eval_mode) {
2624a2e7687Sjeremylt       case CEED_EVAL_NONE:
2634a2e7687Sjeremylt         break;  // No action
2644a2e7687Sjeremylt       case CEED_EVAL_INTERP:
2654a2e7687Sjeremylt       case CEED_EVAL_GRAD:
266a915a514Srezgarshakeri       case CEED_EVAL_DIV:
267*c4e3f59bSSebastian Grimberg       case CEED_EVAL_CURL:
268a915a514Srezgarshakeri         CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis));
269a915a514Srezgarshakeri         CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp));
270a915a514Srezgarshakeri         CeedCallBackend(
271a915a514Srezgarshakeri             CeedVectorSetArray(impl->e_vecs_out[i], CEED_MEM_HOST, CEED_USE_POINTER, &e_data_full[i + num_input_fields][e * elem_size * num_comp]));
272*c4e3f59bSSebastian Grimberg         CeedCallBackend(CeedBasisApply(basis, blk_size, CEED_TRANSPOSE, eval_mode, impl->q_vecs_out[i], impl->e_vecs_out[i]));
273a915a514Srezgarshakeri         break;
274c042f62fSJeremy L Thompson       // LCOV_EXCL_START
275bbfacfcdSjeremylt       case CEED_EVAL_WEIGHT: {
2764ce2993fSjeremylt         Ceed ceed;
2772b730f8bSJeremy L Thompson         CeedCallBackend(CeedOperatorGetCeed(op, &ceed));
2782b730f8bSJeremy L Thompson         return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT cannot be an output evaluation mode");
279bbfacfcdSjeremylt         // LCOV_EXCL_STOP
2804a2e7687Sjeremylt       }
28189c6efa4Sjeremylt     }
28289c6efa4Sjeremylt   }
283e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
2841d102b48SJeremy L Thompson }
2851d102b48SJeremy L Thompson 
286f10650afSjeremylt //------------------------------------------------------------------------------
287f10650afSjeremylt // Restore Input Vectors
288f10650afSjeremylt //------------------------------------------------------------------------------
2892b730f8bSJeremy L Thompson static inline int CeedOperatorRestoreInputs_Blocked(CeedInt num_input_fields, CeedQFunctionField *qf_input_fields, CeedOperatorField *op_input_fields,
2902b730f8bSJeremy L Thompson                                                     bool skip_active, CeedScalar *e_data_full[2 * CEED_FIELD_MAX], CeedOperator_Blocked *impl) {
291d1d35e2fSjeremylt   CeedEvalMode eval_mode;
2921d102b48SJeremy L Thompson 
293d1d35e2fSjeremylt   for (CeedInt i = 0; i < num_input_fields; i++) {
2941d102b48SJeremy L Thompson     // Skip active inputs
295d1d35e2fSjeremylt     if (skip_active) {
2961d102b48SJeremy L Thompson       CeedVector vec;
2972b730f8bSJeremy L Thompson       CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec));
2982b730f8bSJeremy L Thompson       if (vec == CEED_VECTOR_ACTIVE) continue;
2991d102b48SJeremy L Thompson     }
3002b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode));
301d1d35e2fSjeremylt     if (eval_mode == CEED_EVAL_WEIGHT) {  // Skip
3021d102b48SJeremy L Thompson     } else {
3032b730f8bSJeremy L Thompson       CeedCallBackend(CeedVectorRestoreArrayRead(impl->e_vecs_full[i], (const CeedScalar **)&e_data_full[i]));
3041d102b48SJeremy L Thompson     }
3051d102b48SJeremy L Thompson   }
306e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
3071d102b48SJeremy L Thompson }
3081d102b48SJeremy L Thompson 
309f10650afSjeremylt //------------------------------------------------------------------------------
310f10650afSjeremylt // Operator Apply
311f10650afSjeremylt //------------------------------------------------------------------------------
3122b730f8bSJeremy L Thompson static int CeedOperatorApplyAdd_Blocked(CeedOperator op, CeedVector in_vec, CeedVector out_vec, CeedRequest *request) {
3131d102b48SJeremy L Thompson   CeedOperator_Blocked *impl;
3142b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetData(op, &impl));
315d1d35e2fSjeremylt   const CeedInt blk_size = 8;
316d1d35e2fSjeremylt   CeedInt       Q, num_input_fields, num_output_fields, num_elem, size;
3172b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem));
3182b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q));
319d1d35e2fSjeremylt   CeedInt       num_blks = (num_elem / blk_size) + !!(num_elem % blk_size);
3201d102b48SJeremy L Thompson   CeedQFunction qf;
3212b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetQFunction(op, &qf));
322d1d35e2fSjeremylt   CeedOperatorField *op_input_fields, *op_output_fields;
3232b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields));
324d1d35e2fSjeremylt   CeedQFunctionField *qf_input_fields, *qf_output_fields;
3252b730f8bSJeremy L Thompson   CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields));
326d1d35e2fSjeremylt   CeedEvalMode eval_mode;
3271d102b48SJeremy L Thompson   CeedVector   vec;
3284fc1f125SJeremy L Thompson   CeedScalar  *e_data_full[2 * CEED_FIELD_MAX] = {0};
3291d102b48SJeremy L Thompson 
3301d102b48SJeremy L Thompson   // Setup
3312b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorSetup_Blocked(op));
3321d102b48SJeremy L Thompson 
3330b454692Sjeremylt   // Restriction only operator
3340b454692Sjeremylt   if (impl->is_identity_restr_op) {
3352b730f8bSJeremy L Thompson     CeedCallBackend(CeedElemRestrictionApply(impl->blk_restr[0], CEED_NOTRANSPOSE, in_vec, impl->e_vecs_full[0], request));
3362b730f8bSJeremy L Thompson     CeedCallBackend(CeedElemRestrictionApply(impl->blk_restr[1], CEED_TRANSPOSE, impl->e_vecs_full[0], out_vec, request));
3370b454692Sjeremylt     return CEED_ERROR_SUCCESS;
3380b454692Sjeremylt   }
3390b454692Sjeremylt 
3401d102b48SJeremy L Thompson   // Input Evecs and Restriction
3412b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorSetupInputs_Blocked(num_input_fields, qf_input_fields, op_input_fields, in_vec, false, e_data_full, impl, request));
3421d102b48SJeremy L Thompson 
3431d102b48SJeremy L Thompson   // Output Evecs
344d1d35e2fSjeremylt   for (CeedInt i = 0; i < num_output_fields; i++) {
3452b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorGetArrayWrite(impl->e_vecs_full[i + impl->num_inputs], CEED_MEM_HOST, &e_data_full[i + num_input_fields]));
3461d102b48SJeremy L Thompson   }
3471d102b48SJeremy L Thompson 
3481d102b48SJeremy L Thompson   // Loop through elements
349d1d35e2fSjeremylt   for (CeedInt e = 0; e < num_blks * blk_size; e += blk_size) {
3501d102b48SJeremy L Thompson     // Output pointers
351d1d35e2fSjeremylt     for (CeedInt i = 0; i < num_output_fields; i++) {
3522b730f8bSJeremy L Thompson       CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode));
353d1d35e2fSjeremylt       if (eval_mode == CEED_EVAL_NONE) {
3542b730f8bSJeremy L Thompson         CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[i], &size));
3552b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[i], CEED_MEM_HOST, CEED_USE_POINTER, &e_data_full[i + num_input_fields][e * Q * size]));
3561d102b48SJeremy L Thompson       }
3571d102b48SJeremy L Thompson     }
3581d102b48SJeremy L Thompson 
35916911fdaSjeremylt     // Input basis apply
3602b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorInputBasis_Blocked(e, Q, qf_input_fields, op_input_fields, num_input_fields, blk_size, false, e_data_full, impl));
36116911fdaSjeremylt 
3621d102b48SJeremy L Thompson     // Q function
3630b454692Sjeremylt     if (!impl->is_identity_qf) {
3642b730f8bSJeremy L Thompson       CeedCallBackend(CeedQFunctionApply(qf, Q * blk_size, impl->q_vecs_in, impl->q_vecs_out));
36516911fdaSjeremylt     }
3661d102b48SJeremy L Thompson 
3671d102b48SJeremy L Thompson     // Output basis apply
3682b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorOutputBasis_Blocked(e, Q, qf_output_fields, op_output_fields, blk_size, num_input_fields, num_output_fields, op,
3692b730f8bSJeremy L Thompson                                                     e_data_full, impl));
3701d102b48SJeremy L Thompson   }
37189c6efa4Sjeremylt 
37289c6efa4Sjeremylt   // Output restriction
373d1d35e2fSjeremylt   for (CeedInt i = 0; i < num_output_fields; i++) {
37489c6efa4Sjeremylt     // Restore evec
3752b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorRestoreArray(impl->e_vecs_full[i + impl->num_inputs], &e_data_full[i + num_input_fields]));
376d1bcdac9Sjeremylt     // Get output vector
3772b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec));
37889c6efa4Sjeremylt     // Active
3792b730f8bSJeremy L Thompson     if (vec == CEED_VECTOR_ACTIVE) vec = out_vec;
3804a2e7687Sjeremylt     // Restrict
3812b730f8bSJeremy L Thompson     CeedCallBackend(
3822b730f8bSJeremy L Thompson         CeedElemRestrictionApply(impl->blk_restr[i + impl->num_inputs], CEED_TRANSPOSE, impl->e_vecs_full[i + impl->num_inputs], vec, request));
3834a2e7687Sjeremylt   }
3844a2e7687Sjeremylt 
3854a2e7687Sjeremylt   // Restore input arrays
3862b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorRestoreInputs_Blocked(num_input_fields, qf_input_fields, op_input_fields, false, e_data_full, impl));
3871d102b48SJeremy L Thompson 
388e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
3891d102b48SJeremy L Thompson }
3901d102b48SJeremy L Thompson 
391f10650afSjeremylt //------------------------------------------------------------------------------
39270a7ffb3SJeremy L Thompson // Core code for assembling linear QFunction
393f10650afSjeremylt //------------------------------------------------------------------------------
3942b730f8bSJeremy L Thompson static inline int CeedOperatorLinearAssembleQFunctionCore_Blocked(CeedOperator op, bool build_objects, CeedVector *assembled,
395a0162de9SJeremy L Thompson                                                                   CeedElemRestriction *rstr, CeedRequest *request) {
3961d102b48SJeremy L Thompson   CeedOperator_Blocked *impl;
3972b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetData(op, &impl));
398d1d35e2fSjeremylt   const CeedInt blk_size = 8;
399d1d35e2fSjeremylt   CeedInt       Q, num_input_fields, num_output_fields, num_elem, size;
4002b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem));
4012b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q));
402d1d35e2fSjeremylt   CeedInt       num_blks = (num_elem / blk_size) + !!(num_elem % blk_size);
4031d102b48SJeremy L Thompson   CeedQFunction qf;
4042b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetQFunction(op, &qf));
405d1d35e2fSjeremylt   CeedOperatorField *op_input_fields, *op_output_fields;
4062b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields));
407d1d35e2fSjeremylt   CeedQFunctionField *qf_input_fields, *qf_output_fields;
4082b730f8bSJeremy L Thompson   CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields));
4094fc1f125SJeremy L Thompson   CeedVector  vec, l_vec = impl->qf_l_vec;
4102b730f8bSJeremy L Thompson   CeedInt     num_active_in = impl->num_active_in, num_active_out = impl->num_active_out;
411d2643443SJeremy L Thompson   CeedSize    q_size;
412bb219a0fSJeremy L Thompson   CeedVector *active_in = impl->qf_active_in;
4131d102b48SJeremy L Thompson   CeedScalar *a, *tmp;
4141d102b48SJeremy L Thompson   Ceed        ceed;
4152b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetCeed(op, &ceed));
4164fc1f125SJeremy L Thompson   CeedScalar *e_data_full[2 * CEED_FIELD_MAX] = {0};
4171d102b48SJeremy L Thompson 
4181d102b48SJeremy L Thompson   // Setup
4192b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorSetup_Blocked(op));
4201d102b48SJeremy L Thompson 
42116911fdaSjeremylt   // Check for identity
4222b730f8bSJeremy L Thompson   if (impl->is_identity_qf) {
42316911fdaSjeremylt     // LCOV_EXCL_START
4242b730f8bSJeremy L Thompson     return CeedError(ceed, CEED_ERROR_BACKEND, "Assembling identity QFunctions not supported");
42516911fdaSjeremylt     // LCOV_EXCL_STOP
4262b730f8bSJeremy L Thompson   }
42716911fdaSjeremylt 
4281d102b48SJeremy L Thompson   // Input Evecs and Restriction
4292b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorSetupInputs_Blocked(num_input_fields, qf_input_fields, op_input_fields, NULL, true, e_data_full, impl, request));
4301d102b48SJeremy L Thompson 
4311d102b48SJeremy L Thompson   // Count number of active input fields
432bb219a0fSJeremy L Thompson   if (!num_active_in) {
433d1d35e2fSjeremylt     for (CeedInt i = 0; i < num_input_fields; i++) {
4341d102b48SJeremy L Thompson       // Get input vector
4352b730f8bSJeremy L Thompson       CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec));
4361d102b48SJeremy L Thompson       // Check if active input
4371d102b48SJeremy L Thompson       if (vec == CEED_VECTOR_ACTIVE) {
4382b730f8bSJeremy L Thompson         CeedCallBackend(CeedQFunctionFieldGetSize(qf_input_fields[i], &size));
4392b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorSetValue(impl->q_vecs_in[i], 0.0));
4402b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorGetArray(impl->q_vecs_in[i], CEED_MEM_HOST, &tmp));
4412b730f8bSJeremy L Thompson         CeedCallBackend(CeedRealloc(num_active_in + size, &active_in));
4421d102b48SJeremy L Thompson         for (CeedInt field = 0; field < size; field++) {
443d2643443SJeremy L Thompson           q_size = (CeedSize)Q * blk_size;
4442b730f8bSJeremy L Thompson           CeedCallBackend(CeedVectorCreate(ceed, q_size, &active_in[num_active_in + field]));
4452b730f8bSJeremy L Thompson           CeedCallBackend(CeedVectorSetArray(active_in[num_active_in + field], CEED_MEM_HOST, CEED_USE_POINTER, &tmp[field * Q * blk_size]));
4461d102b48SJeremy L Thompson         }
447d1d35e2fSjeremylt         num_active_in += size;
4482b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorRestoreArray(impl->q_vecs_in[i], &tmp));
4491d102b48SJeremy L Thompson       }
4501d102b48SJeremy L Thompson     }
4514fc1f125SJeremy L Thompson     impl->num_active_in = num_active_in;
452bb219a0fSJeremy L Thompson     impl->qf_active_in  = active_in;
453bb219a0fSJeremy L Thompson   }
4541d102b48SJeremy L Thompson 
4551d102b48SJeremy L Thompson   // Count number of active output fields
456bb219a0fSJeremy L Thompson   if (!num_active_out) {
457d1d35e2fSjeremylt     for (CeedInt i = 0; i < num_output_fields; i++) {
4581d102b48SJeremy L Thompson       // Get output vector
4592b730f8bSJeremy L Thompson       CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec));
4601d102b48SJeremy L Thompson       // Check if active output
4611d102b48SJeremy L Thompson       if (vec == CEED_VECTOR_ACTIVE) {
4622b730f8bSJeremy L Thompson         CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[i], &size));
463d1d35e2fSjeremylt         num_active_out += size;
4641d102b48SJeremy L Thompson       }
4651d102b48SJeremy L Thompson     }
4664fc1f125SJeremy L Thompson     impl->num_active_out = num_active_out;
467bb219a0fSJeremy L Thompson   }
4681d102b48SJeremy L Thompson 
4691d102b48SJeremy L Thompson   // Check sizes
4702b730f8bSJeremy L Thompson   if (!num_active_in || !num_active_out) {
4711d102b48SJeremy L Thompson     // LCOV_EXCL_START
4722b730f8bSJeremy L Thompson     return CeedError(ceed, CEED_ERROR_BACKEND, "Cannot assemble QFunction without active inputs and outputs");
4731d102b48SJeremy L Thompson     // LCOV_EXCL_STOP
4742b730f8bSJeremy L Thompson   }
4751d102b48SJeremy L Thompson 
476d1d35e2fSjeremylt   // Setup Lvec
4774fc1f125SJeremy L Thompson   if (!l_vec) {
478d2643443SJeremy L Thompson     CeedSize l_size = (CeedSize)num_blks * blk_size * Q * num_active_in * num_active_out;
4792b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorCreate(ceed, l_size, &l_vec));
4804fc1f125SJeremy L Thompson     impl->qf_l_vec = l_vec;
481bb219a0fSJeremy L Thompson   }
4822b730f8bSJeremy L Thompson   CeedCallBackend(CeedVectorGetArrayWrite(l_vec, CEED_MEM_HOST, &a));
4831d102b48SJeremy L Thompson 
48470a7ffb3SJeremy L Thompson   // Build objects if needed
485d1d35e2fSjeremylt   CeedInt strides[3] = {1, Q, num_active_in * num_active_out * Q};
48670a7ffb3SJeremy L Thompson   if (build_objects) {
48770a7ffb3SJeremy L Thompson     // Create output restriction
4882b730f8bSJeremy L Thompson     CeedCallBackend(CeedElemRestrictionCreateStrided(ceed, num_elem, Q, num_active_in * num_active_out, num_active_in * num_active_out * num_elem * Q,
4892b730f8bSJeremy L Thompson                                                      strides, rstr));
4901d102b48SJeremy L Thompson     // Create assembled vector
491d2643443SJeremy L Thompson     CeedSize l_size = (CeedSize)num_elem * Q * num_active_in * num_active_out;
4922b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorCreate(ceed, l_size, assembled));
49370a7ffb3SJeremy L Thompson   }
4941d102b48SJeremy L Thompson 
4951d102b48SJeremy L Thompson   // Loop through elements
496d1d35e2fSjeremylt   for (CeedInt e = 0; e < num_blks * blk_size; e += blk_size) {
4971d102b48SJeremy L Thompson     // Input basis apply
4982b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorInputBasis_Blocked(e, Q, qf_input_fields, op_input_fields, num_input_fields, blk_size, true, e_data_full, impl));
4991d102b48SJeremy L Thompson 
5001d102b48SJeremy L Thompson     // Assemble QFunction
501d1d35e2fSjeremylt     for (CeedInt in = 0; in < num_active_in; in++) {
5021d102b48SJeremy L Thompson       // Set Inputs
5032b730f8bSJeremy L Thompson       CeedCallBackend(CeedVectorSetValue(active_in[in], 1.0));
504d1d35e2fSjeremylt       if (num_active_in > 1) {
5052b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorSetValue(active_in[(in + num_active_in - 1) % num_active_in], 0.0));
50642ea3801Sjeremylt       }
5071d102b48SJeremy L Thompson       // Set Outputs
508d1d35e2fSjeremylt       for (CeedInt out = 0; out < num_output_fields; out++) {
5091d102b48SJeremy L Thompson         // Get output vector
5102b730f8bSJeremy L Thompson         CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[out], &vec));
5111d102b48SJeremy L Thompson         // Check if active output
5121d102b48SJeremy L Thompson         if (vec == CEED_VECTOR_ACTIVE) {
5132b730f8bSJeremy L Thompson           CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[out], CEED_MEM_HOST, CEED_USE_POINTER, a));
5142b730f8bSJeremy L Thompson           CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[out], &size));
515d1d35e2fSjeremylt           a += size * Q * blk_size;  // Advance the pointer by the size of the output
5161d102b48SJeremy L Thompson         }
5171d102b48SJeremy L Thompson       }
5181d102b48SJeremy L Thompson       // Apply QFunction
5192b730f8bSJeremy L Thompson       CeedCallBackend(CeedQFunctionApply(qf, Q * blk_size, impl->q_vecs_in, impl->q_vecs_out));
5204a2e7687Sjeremylt     }
5214a2e7687Sjeremylt   }
5224a2e7687Sjeremylt 
5231d102b48SJeremy L Thompson   // Un-set output Qvecs to prevent accidental overwrite of Assembled
524d1d35e2fSjeremylt   for (CeedInt out = 0; out < num_output_fields; out++) {
5251d102b48SJeremy L Thompson     // Get output vector
5262b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[out], &vec));
5271d102b48SJeremy L Thompson     // Check if active output
5281d102b48SJeremy L Thompson     if (vec == CEED_VECTOR_ACTIVE) {
5292b730f8bSJeremy L Thompson       CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[out], CEED_MEM_HOST, CEED_COPY_VALUES, NULL));
5301d102b48SJeremy L Thompson     }
5311d102b48SJeremy L Thompson   }
5321d102b48SJeremy L Thompson 
5331d102b48SJeremy L Thompson   // Restore input arrays
5342b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorRestoreInputs_Blocked(num_input_fields, qf_input_fields, op_input_fields, true, e_data_full, impl));
5351d102b48SJeremy L Thompson 
5361d102b48SJeremy L Thompson   // Output blocked restriction
5372b730f8bSJeremy L Thompson   CeedCallBackend(CeedVectorRestoreArray(l_vec, &a));
5382b730f8bSJeremy L Thompson   CeedCallBackend(CeedVectorSetValue(*assembled, 0.0));
539bb219a0fSJeremy L Thompson   CeedElemRestriction blk_rstr = impl->qf_blk_rstr;
540bb219a0fSJeremy L Thompson   if (!impl->qf_blk_rstr) {
5412b730f8bSJeremy L Thompson     CeedCallBackend(CeedElemRestrictionCreateBlockedStrided(ceed, num_elem, Q, blk_size, num_active_in * num_active_out,
5422b730f8bSJeremy L Thompson                                                             num_active_in * num_active_out * num_elem * Q, strides, &blk_rstr));
543bb219a0fSJeremy L Thompson     impl->qf_blk_rstr = blk_rstr;
544bb219a0fSJeremy L Thompson   }
5452b730f8bSJeremy L Thompson   CeedCallBackend(CeedElemRestrictionApply(blk_rstr, CEED_TRANSPOSE, l_vec, *assembled, request));
5461d102b48SJeremy L Thompson 
547e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
5484a2e7687Sjeremylt }
5494a2e7687Sjeremylt 
550f10650afSjeremylt //------------------------------------------------------------------------------
55170a7ffb3SJeremy L Thompson // Assemble Linear QFunction
55270a7ffb3SJeremy L Thompson //------------------------------------------------------------------------------
5532b730f8bSJeremy L Thompson static int CeedOperatorLinearAssembleQFunction_Blocked(CeedOperator op, CeedVector *assembled, CeedElemRestriction *rstr, CeedRequest *request) {
5542b730f8bSJeremy L Thompson   return CeedOperatorLinearAssembleQFunctionCore_Blocked(op, true, assembled, rstr, request);
55570a7ffb3SJeremy L Thompson }
55670a7ffb3SJeremy L Thompson 
55770a7ffb3SJeremy L Thompson //------------------------------------------------------------------------------
55870a7ffb3SJeremy L Thompson // Update Assembled Linear QFunction
55970a7ffb3SJeremy L Thompson //------------------------------------------------------------------------------
5602b730f8bSJeremy L Thompson static int CeedOperatorLinearAssembleQFunctionUpdate_Blocked(CeedOperator op, CeedVector assembled, CeedElemRestriction rstr, CeedRequest *request) {
5612b730f8bSJeremy L Thompson   return CeedOperatorLinearAssembleQFunctionCore_Blocked(op, false, &assembled, &rstr, request);
56270a7ffb3SJeremy L Thompson }
56370a7ffb3SJeremy L Thompson 
56470a7ffb3SJeremy L Thompson //------------------------------------------------------------------------------
565f10650afSjeremylt // Operator Destroy
566f10650afSjeremylt //------------------------------------------------------------------------------
567f10650afSjeremylt static int CeedOperatorDestroy_Blocked(CeedOperator op) {
568f10650afSjeremylt   CeedOperator_Blocked *impl;
5692b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetData(op, &impl));
570f10650afSjeremylt 
5714fc1f125SJeremy L Thompson   for (CeedInt i = 0; i < impl->num_inputs + impl->num_outputs; i++) {
5722b730f8bSJeremy L Thompson     CeedCallBackend(CeedElemRestrictionDestroy(&impl->blk_restr[i]));
5732b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorDestroy(&impl->e_vecs_full[i]));
574f10650afSjeremylt   }
5752b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl->blk_restr));
5762b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl->e_vecs_full));
5772b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl->input_states));
578f10650afSjeremylt 
5794fc1f125SJeremy L Thompson   for (CeedInt i = 0; i < impl->num_inputs; i++) {
5802b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorDestroy(&impl->e_vecs_in[i]));
5812b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorDestroy(&impl->q_vecs_in[i]));
582f10650afSjeremylt   }
5832b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl->e_vecs_in));
5842b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl->q_vecs_in));
585f10650afSjeremylt 
5864fc1f125SJeremy L Thompson   for (CeedInt i = 0; i < impl->num_outputs; i++) {
5872b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorDestroy(&impl->e_vecs_out[i]));
5882b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorDestroy(&impl->q_vecs_out[i]));
589f10650afSjeremylt   }
5902b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl->e_vecs_out));
5912b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl->q_vecs_out));
592f10650afSjeremylt 
593bb219a0fSJeremy L Thompson   // QFunction assembly data
5944fc1f125SJeremy L Thompson   for (CeedInt i = 0; i < impl->num_active_in; i++) {
5952b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorDestroy(&impl->qf_active_in[i]));
596bb219a0fSJeremy L Thompson   }
5972b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl->qf_active_in));
5982b730f8bSJeremy L Thompson   CeedCallBackend(CeedVectorDestroy(&impl->qf_l_vec));
5992b730f8bSJeremy L Thompson   CeedCallBackend(CeedElemRestrictionDestroy(&impl->qf_blk_rstr));
600bb219a0fSJeremy L Thompson 
6012b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl));
602e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
603f10650afSjeremylt }
604f10650afSjeremylt 
605f10650afSjeremylt //------------------------------------------------------------------------------
606f10650afSjeremylt // Operator Create
607f10650afSjeremylt //------------------------------------------------------------------------------
6084a2e7687Sjeremylt int CeedOperatorCreate_Blocked(CeedOperator op) {
609fe2413ffSjeremylt   Ceed ceed;
6102b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetCeed(op, &ceed));
6114ce2993fSjeremylt   CeedOperator_Blocked *impl;
6124a2e7687Sjeremylt 
6132b730f8bSJeremy L Thompson   CeedCallBackend(CeedCalloc(1, &impl));
6142b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorSetData(op, impl));
615fe2413ffSjeremylt 
6162b730f8bSJeremy L Thompson   CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunction", CeedOperatorLinearAssembleQFunction_Blocked));
6172b730f8bSJeremy L Thompson   CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunctionUpdate", CeedOperatorLinearAssembleQFunctionUpdate_Blocked));
6182b730f8bSJeremy L Thompson   CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAdd_Blocked));
6192b730f8bSJeremy L Thompson   CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Blocked));
620e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
6214a2e7687Sjeremylt }
622f10650afSjeremylt //------------------------------------------------------------------------------
623