xref: /libCEED/rust/libceed-sys/c-src/backends/blocked/ceed-blocked-operator.c (revision 77d1c127eaba12da4c1761ef74a16ca3fc16e493)
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;
62*77d1c127SSebastian Grimberg         const bool    *orients      = NULL;
63*77d1c127SSebastian Grimberg         const CeedInt *curl_orients = NULL;
642b730f8bSJeremy L Thompson         CeedCallBackend(CeedElemRestrictionGetOffsets(r, CEED_MEM_HOST, &offsets));
65*77d1c127SSebastian Grimberg         CeedCallBackend(CeedElemRestrictionGetOrientations(r, CEED_MEM_HOST, &orients));
66*77d1c127SSebastian Grimberg         CeedCallBackend(CeedElemRestrictionGetCurlOrientations(r, CEED_MEM_HOST, &curl_orients));
672b730f8bSJeremy L Thompson         CeedCallBackend(CeedElemRestrictionGetCompStride(r, &comp_stride));
68*77d1c127SSebastian Grimberg         if (!orients && !curl_orients) {
692b730f8bSJeremy L Thompson           CeedCallBackend(CeedElemRestrictionCreateBlocked(ceed, num_elem, elem_size, blk_size, num_comp, comp_stride, l_size, CEED_MEM_HOST,
702b730f8bSJeremy L Thompson                                                            CEED_COPY_VALUES, offsets, &blk_restr[i + start_e]));
71*77d1c127SSebastian Grimberg         } else if (!curl_orients) {
72*77d1c127SSebastian Grimberg           CeedCallBackend(CeedElemRestrictionCreateBlockedOriented(ceed, num_elem, elem_size, blk_size, num_comp, comp_stride, l_size, CEED_MEM_HOST,
73*77d1c127SSebastian Grimberg                                                                    CEED_COPY_VALUES, offsets, orients, &blk_restr[i + start_e]));
74*77d1c127SSebastian Grimberg         } else {
75*77d1c127SSebastian Grimberg           CeedCallBackend(CeedElemRestrictionCreateBlockedCurlOriented(ceed, num_elem, elem_size, blk_size, num_comp, comp_stride, l_size,
76*77d1c127SSebastian Grimberg                                                                        CEED_MEM_HOST, CEED_COPY_VALUES, offsets, curl_orients,
77*77d1c127SSebastian Grimberg                                                                        &blk_restr[i + start_e]));
78*77d1c127SSebastian Grimberg         }
792b730f8bSJeremy L Thompson         CeedCallBackend(CeedElemRestrictionRestoreOffsets(r, &offsets));
80*77d1c127SSebastian Grimberg         CeedCallBackend(CeedElemRestrictionRestoreOrientations(r, &orients));
81*77d1c127SSebastian Grimberg         CeedCallBackend(CeedElemRestrictionRestoreCurlOrientations(r, &curl_orients));
823ac43b2cSJeremy L Thompson       }
832b730f8bSJeremy L Thompson       CeedCallBackend(CeedElemRestrictionCreateVector(blk_restr[i + start_e], NULL, &e_vecs_full[i + start_e]));
844a2e7687Sjeremylt     }
854a2e7687Sjeremylt 
86d1d35e2fSjeremylt     switch (eval_mode) {
874a2e7687Sjeremylt       case CEED_EVAL_NONE:
882b730f8bSJeremy L Thompson         CeedCallBackend(CeedQFunctionFieldGetSize(qf_fields[i], &size));
89d2643443SJeremy L Thompson         q_size = (CeedSize)Q * size * blk_size;
902b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i]));
91aedaa0e5Sjeremylt         break;
92aedaa0e5Sjeremylt       case CEED_EVAL_INTERP:
934a2e7687Sjeremylt       case CEED_EVAL_GRAD:
94a915a514Srezgarshakeri       case CEED_EVAL_DIV:
95c4e3f59bSSebastian Grimberg       case CEED_EVAL_CURL:
962b730f8bSJeremy L Thompson         CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis));
972b730f8bSJeremy L Thompson         CeedCallBackend(CeedQFunctionFieldGetSize(qf_fields[i], &size));
982b730f8bSJeremy L Thompson         CeedCallBackend(CeedBasisGetNumNodes(basis, &P));
992b730f8bSJeremy L Thompson         CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp));
100d2643443SJeremy L Thompson         e_size = (CeedSize)P * num_comp * blk_size;
1012b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorCreate(ceed, e_size, &e_vecs[i]));
102d2643443SJeremy L Thompson         q_size = (CeedSize)Q * size * blk_size;
1032b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i]));
1044a2e7687Sjeremylt         break;
1054a2e7687Sjeremylt       case CEED_EVAL_WEIGHT:  // Only on input fields
1062b730f8bSJeremy L Thompson         CeedCallBackend(CeedOperatorFieldGetBasis(op_fields[i], &basis));
107d2643443SJeremy L Thompson         q_size = (CeedSize)Q * blk_size;
1082b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorCreate(ceed, q_size, &q_vecs[i]));
1092b730f8bSJeremy L Thompson         CeedCallBackend(CeedBasisApply(basis, blk_size, CEED_NOTRANSPOSE, CEED_EVAL_WEIGHT, CEED_VECTOR_NONE, q_vecs[i]));
1104a2e7687Sjeremylt         break;
1114a2e7687Sjeremylt     }
1124a2e7687Sjeremylt   }
113e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
1144a2e7687Sjeremylt }
1154a2e7687Sjeremylt 
116f10650afSjeremylt //------------------------------------------------------------------------------
117f10650afSjeremylt // Setup Operator
118f10650afSjeremylt //------------------------------------------------------------------------------
1194a2e7687Sjeremylt static int CeedOperatorSetup_Blocked(CeedOperator op) {
1208c1105f8SJeremy L Thompson   bool is_setup_done;
1212b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorIsSetupDone(op, &is_setup_done));
1228c1105f8SJeremy L Thompson   if (is_setup_done) return CEED_ERROR_SUCCESS;
123aedaa0e5Sjeremylt   Ceed ceed;
1242b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetCeed(op, &ceed));
1254ce2993fSjeremylt   CeedOperator_Blocked *impl;
1262b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetData(op, &impl));
1274ce2993fSjeremylt   CeedQFunction qf;
1282b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetQFunction(op, &qf));
129d1d35e2fSjeremylt   CeedInt Q, num_input_fields, num_output_fields;
1302b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q));
1312b730f8bSJeremy L Thompson   CeedCallBackend(CeedQFunctionIsIdentity(qf, &impl->is_identity_qf));
132d1d35e2fSjeremylt   CeedOperatorField *op_input_fields, *op_output_fields;
1332b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields));
134d1d35e2fSjeremylt   CeedQFunctionField *qf_input_fields, *qf_output_fields;
1352b730f8bSJeremy L Thompson   CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields));
1364a2e7687Sjeremylt 
1374a2e7687Sjeremylt   // Allocate
1382b730f8bSJeremy L Thompson   CeedCallBackend(CeedCalloc(num_input_fields + num_output_fields, &impl->blk_restr));
1392b730f8bSJeremy L Thompson   CeedCallBackend(CeedCalloc(num_input_fields + num_output_fields, &impl->e_vecs_full));
1404a2e7687Sjeremylt 
1412b730f8bSJeremy L Thompson   CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->input_states));
1422b730f8bSJeremy L Thompson   CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->e_vecs_in));
1432b730f8bSJeremy L Thompson   CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->e_vecs_out));
1442b730f8bSJeremy L Thompson   CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->q_vecs_in));
1452b730f8bSJeremy L Thompson   CeedCallBackend(CeedCalloc(CEED_FIELD_MAX, &impl->q_vecs_out));
1464a2e7687Sjeremylt 
1474fc1f125SJeremy L Thompson   impl->num_inputs  = num_input_fields;
1484fc1f125SJeremy L Thompson   impl->num_outputs = num_output_fields;
149aedaa0e5Sjeremylt 
1504a2e7687Sjeremylt   // Set up infield and outfield pointer arrays
1514a2e7687Sjeremylt   // Infields
1522b730f8bSJeremy L Thompson   CeedCallBackend(
1532b730f8bSJeremy 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));
1544a2e7687Sjeremylt   // Outfields
1552b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorSetupFields_Blocked(qf, op, false, impl->blk_restr, impl->e_vecs_full, impl->e_vecs_out, impl->q_vecs_out,
1562b730f8bSJeremy L Thompson                                                   num_input_fields, num_output_fields, Q));
157aedaa0e5Sjeremylt 
15816911fdaSjeremylt   // Identity QFunctions
1590b454692Sjeremylt   if (impl->is_identity_qf) {
160d1d35e2fSjeremylt     CeedEvalMode        in_mode, out_mode;
161d1d35e2fSjeremylt     CeedQFunctionField *in_fields, *out_fields;
1622b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &in_fields, NULL, &out_fields));
1632b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionFieldGetEvalMode(in_fields[0], &in_mode));
1642b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionFieldGetEvalMode(out_fields[0], &out_mode));
165d1d35e2fSjeremylt 
1660b454692Sjeremylt     if (in_mode == CEED_EVAL_NONE && out_mode == CEED_EVAL_NONE) {
1670b454692Sjeremylt       impl->is_identity_restr_op = true;
1680b454692Sjeremylt     } else {
169db002c03SJeremy L Thompson       CeedCallBackend(CeedVectorReferenceCopy(impl->q_vecs_in[0], &impl->q_vecs_out[0]));
17016911fdaSjeremylt     }
17116911fdaSjeremylt   }
17216911fdaSjeremylt 
1732b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorSetSetupDone(op));
1744a2e7687Sjeremylt 
175e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
1764a2e7687Sjeremylt }
1774a2e7687Sjeremylt 
178f10650afSjeremylt //------------------------------------------------------------------------------
179f10650afSjeremylt // Setup Operator Inputs
180f10650afSjeremylt //------------------------------------------------------------------------------
1812b730f8bSJeremy L Thompson static inline int CeedOperatorSetupInputs_Blocked(CeedInt num_input_fields, CeedQFunctionField *qf_input_fields, CeedOperatorField *op_input_fields,
1824fc1f125SJeremy L Thompson                                                   CeedVector in_vec, bool skip_active, CeedScalar *e_data_full[2 * CEED_FIELD_MAX],
183a0162de9SJeremy L Thompson                                                   CeedOperator_Blocked *impl, CeedRequest *request) {
184d1d35e2fSjeremylt   CeedEvalMode eval_mode;
185d1bcdac9Sjeremylt   CeedVector   vec;
18616c359e6Sjeremylt   uint64_t     state;
1874a2e7687Sjeremylt 
188d1d35e2fSjeremylt   for (CeedInt i = 0; i < num_input_fields; i++) {
1891d102b48SJeremy L Thompson     // Get input vector
1902b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec));
1911d102b48SJeremy L Thompson     if (vec == CEED_VECTOR_ACTIVE) {
1922b730f8bSJeremy L Thompson       if (skip_active) continue;
1932b730f8bSJeremy L Thompson       else vec = in_vec;
1941d102b48SJeremy L Thompson     }
1951d102b48SJeremy L Thompson 
1962b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode));
197d1d35e2fSjeremylt     if (eval_mode == CEED_EVAL_WEIGHT) {  // Skip
1984a2e7687Sjeremylt     } else {
1994a2e7687Sjeremylt       // Restrict
2002b730f8bSJeremy L Thompson       CeedCallBackend(CeedVectorGetState(vec, &state));
2014fc1f125SJeremy L Thompson       if (state != impl->input_states[i] || vec == in_vec) {
2022b730f8bSJeremy L Thompson         CeedCallBackend(CeedElemRestrictionApply(impl->blk_restr[i], CEED_NOTRANSPOSE, vec, impl->e_vecs_full[i], request));
2034fc1f125SJeremy L Thompson         impl->input_states[i] = state;
20416c359e6Sjeremylt       }
2054a2e7687Sjeremylt       // Get evec
2062b730f8bSJeremy L Thompson       CeedCallBackend(CeedVectorGetArrayRead(impl->e_vecs_full[i], CEED_MEM_HOST, (const CeedScalar **)&e_data_full[i]));
2074a2e7687Sjeremylt     }
2084a2e7687Sjeremylt   }
209e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
2104a2e7687Sjeremylt }
2114a2e7687Sjeremylt 
212f10650afSjeremylt //------------------------------------------------------------------------------
213f10650afSjeremylt // Input Basis Action
214f10650afSjeremylt //------------------------------------------------------------------------------
2152b730f8bSJeremy L Thompson static inline int CeedOperatorInputBasis_Blocked(CeedInt e, CeedInt Q, CeedQFunctionField *qf_input_fields, CeedOperatorField *op_input_fields,
216d1d35e2fSjeremylt                                                  CeedInt num_input_fields, CeedInt blk_size, bool skip_active,
2174fc1f125SJeremy L Thompson                                                  CeedScalar *e_data_full[2 * CEED_FIELD_MAX], CeedOperator_Blocked *impl) {
218a915a514Srezgarshakeri   CeedInt             elem_size, size, num_comp;
219d1d35e2fSjeremylt   CeedElemRestriction elem_restr;
220d1d35e2fSjeremylt   CeedEvalMode        eval_mode;
2211d102b48SJeremy L Thompson   CeedBasis           basis;
2221d102b48SJeremy L Thompson 
223d1d35e2fSjeremylt   for (CeedInt i = 0; i < num_input_fields; i++) {
2241d102b48SJeremy L Thompson     // Skip active input
225d1d35e2fSjeremylt     if (skip_active) {
2261d102b48SJeremy L Thompson       CeedVector vec;
2272b730f8bSJeremy L Thompson       CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec));
2282b730f8bSJeremy L Thompson       if (vec == CEED_VECTOR_ACTIVE) continue;
2291d102b48SJeremy L Thompson     }
2301d102b48SJeremy L Thompson 
231d1d35e2fSjeremylt     // Get elem_size, eval_mode, size
2322b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &elem_restr));
2332b730f8bSJeremy L Thompson     CeedCallBackend(CeedElemRestrictionGetElementSize(elem_restr, &elem_size));
2342b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode));
2352b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionFieldGetSize(qf_input_fields[i], &size));
2364a2e7687Sjeremylt     // Basis action
237d1d35e2fSjeremylt     switch (eval_mode) {
2384a2e7687Sjeremylt       case CEED_EVAL_NONE:
2392b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorSetArray(impl->q_vecs_in[i], CEED_MEM_HOST, CEED_USE_POINTER, &e_data_full[i][e * Q * size]));
2404a2e7687Sjeremylt         break;
2414a2e7687Sjeremylt       case CEED_EVAL_INTERP:
2424a2e7687Sjeremylt       case CEED_EVAL_GRAD:
243a915a514Srezgarshakeri       case CEED_EVAL_DIV:
244c4e3f59bSSebastian Grimberg       case CEED_EVAL_CURL:
245a915a514Srezgarshakeri         CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis));
246a915a514Srezgarshakeri         CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp));
247a915a514Srezgarshakeri         CeedCallBackend(CeedVectorSetArray(impl->e_vecs_in[i], CEED_MEM_HOST, CEED_USE_POINTER, &e_data_full[i][e * elem_size * num_comp]));
248c4e3f59bSSebastian Grimberg         CeedCallBackend(CeedBasisApply(basis, blk_size, CEED_NOTRANSPOSE, eval_mode, impl->e_vecs_in[i], impl->q_vecs_in[i]));
249a915a514Srezgarshakeri         break;
2504a2e7687Sjeremylt       case CEED_EVAL_WEIGHT:
2514a2e7687Sjeremylt         break;  // No action
2524a2e7687Sjeremylt     }
25389c6efa4Sjeremylt   }
254e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
25589c6efa4Sjeremylt }
2564a2e7687Sjeremylt 
257f10650afSjeremylt //------------------------------------------------------------------------------
258f10650afSjeremylt // Output Basis Action
259f10650afSjeremylt //------------------------------------------------------------------------------
2602b730f8bSJeremy L Thompson static inline int CeedOperatorOutputBasis_Blocked(CeedInt e, CeedInt Q, CeedQFunctionField *qf_output_fields, CeedOperatorField *op_output_fields,
2612b730f8bSJeremy L Thompson                                                   CeedInt blk_size, CeedInt num_input_fields, CeedInt num_output_fields, CeedOperator op,
2622b730f8bSJeremy L Thompson                                                   CeedScalar *e_data_full[2 * CEED_FIELD_MAX], CeedOperator_Blocked *impl) {
263a915a514Srezgarshakeri   CeedInt             elem_size, num_comp;
264d1d35e2fSjeremylt   CeedElemRestriction elem_restr;
265d1d35e2fSjeremylt   CeedEvalMode        eval_mode;
2661d102b48SJeremy L Thompson   CeedBasis           basis;
2671d102b48SJeremy L Thompson 
268d1d35e2fSjeremylt   for (CeedInt i = 0; i < num_output_fields; i++) {
269d1d35e2fSjeremylt     // Get elem_size, eval_mode, size
2702b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &elem_restr));
2712b730f8bSJeremy L Thompson     CeedCallBackend(CeedElemRestrictionGetElementSize(elem_restr, &elem_size));
2722b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode));
2734a2e7687Sjeremylt     // Basis action
274d1d35e2fSjeremylt     switch (eval_mode) {
2754a2e7687Sjeremylt       case CEED_EVAL_NONE:
2764a2e7687Sjeremylt         break;  // No action
2774a2e7687Sjeremylt       case CEED_EVAL_INTERP:
2784a2e7687Sjeremylt       case CEED_EVAL_GRAD:
279a915a514Srezgarshakeri       case CEED_EVAL_DIV:
280c4e3f59bSSebastian Grimberg       case CEED_EVAL_CURL:
281a915a514Srezgarshakeri         CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis));
282a915a514Srezgarshakeri         CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp));
283a915a514Srezgarshakeri         CeedCallBackend(
284a915a514Srezgarshakeri             CeedVectorSetArray(impl->e_vecs_out[i], CEED_MEM_HOST, CEED_USE_POINTER, &e_data_full[i + num_input_fields][e * elem_size * num_comp]));
285c4e3f59bSSebastian Grimberg         CeedCallBackend(CeedBasisApply(basis, blk_size, CEED_TRANSPOSE, eval_mode, impl->q_vecs_out[i], impl->e_vecs_out[i]));
286a915a514Srezgarshakeri         break;
287c042f62fSJeremy L Thompson       // LCOV_EXCL_START
288bbfacfcdSjeremylt       case CEED_EVAL_WEIGHT: {
2894ce2993fSjeremylt         Ceed ceed;
2902b730f8bSJeremy L Thompson         CeedCallBackend(CeedOperatorGetCeed(op, &ceed));
2912b730f8bSJeremy L Thompson         return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT cannot be an output evaluation mode");
292bbfacfcdSjeremylt         // LCOV_EXCL_STOP
2934a2e7687Sjeremylt       }
29489c6efa4Sjeremylt     }
29589c6efa4Sjeremylt   }
296e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
2971d102b48SJeremy L Thompson }
2981d102b48SJeremy L Thompson 
299f10650afSjeremylt //------------------------------------------------------------------------------
300f10650afSjeremylt // Restore Input Vectors
301f10650afSjeremylt //------------------------------------------------------------------------------
3022b730f8bSJeremy L Thompson static inline int CeedOperatorRestoreInputs_Blocked(CeedInt num_input_fields, CeedQFunctionField *qf_input_fields, CeedOperatorField *op_input_fields,
3032b730f8bSJeremy L Thompson                                                     bool skip_active, CeedScalar *e_data_full[2 * CEED_FIELD_MAX], CeedOperator_Blocked *impl) {
304d1d35e2fSjeremylt   CeedEvalMode eval_mode;
3051d102b48SJeremy L Thompson 
306d1d35e2fSjeremylt   for (CeedInt i = 0; i < num_input_fields; i++) {
3071d102b48SJeremy L Thompson     // Skip active inputs
308d1d35e2fSjeremylt     if (skip_active) {
3091d102b48SJeremy L Thompson       CeedVector vec;
3102b730f8bSJeremy L Thompson       CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec));
3112b730f8bSJeremy L Thompson       if (vec == CEED_VECTOR_ACTIVE) continue;
3121d102b48SJeremy L Thompson     }
3132b730f8bSJeremy L Thompson     CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode));
314d1d35e2fSjeremylt     if (eval_mode == CEED_EVAL_WEIGHT) {  // Skip
3151d102b48SJeremy L Thompson     } else {
3162b730f8bSJeremy L Thompson       CeedCallBackend(CeedVectorRestoreArrayRead(impl->e_vecs_full[i], (const CeedScalar **)&e_data_full[i]));
3171d102b48SJeremy L Thompson     }
3181d102b48SJeremy L Thompson   }
319e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
3201d102b48SJeremy L Thompson }
3211d102b48SJeremy L Thompson 
322f10650afSjeremylt //------------------------------------------------------------------------------
323f10650afSjeremylt // Operator Apply
324f10650afSjeremylt //------------------------------------------------------------------------------
3252b730f8bSJeremy L Thompson static int CeedOperatorApplyAdd_Blocked(CeedOperator op, CeedVector in_vec, CeedVector out_vec, CeedRequest *request) {
3261d102b48SJeremy L Thompson   CeedOperator_Blocked *impl;
3272b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetData(op, &impl));
328d1d35e2fSjeremylt   const CeedInt blk_size = 8;
329d1d35e2fSjeremylt   CeedInt       Q, num_input_fields, num_output_fields, num_elem, size;
3302b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem));
3312b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q));
332d1d35e2fSjeremylt   CeedInt       num_blks = (num_elem / blk_size) + !!(num_elem % blk_size);
3331d102b48SJeremy L Thompson   CeedQFunction qf;
3342b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetQFunction(op, &qf));
335d1d35e2fSjeremylt   CeedOperatorField *op_input_fields, *op_output_fields;
3362b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields));
337d1d35e2fSjeremylt   CeedQFunctionField *qf_input_fields, *qf_output_fields;
3382b730f8bSJeremy L Thompson   CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields));
339d1d35e2fSjeremylt   CeedEvalMode eval_mode;
3401d102b48SJeremy L Thompson   CeedVector   vec;
3414fc1f125SJeremy L Thompson   CeedScalar  *e_data_full[2 * CEED_FIELD_MAX] = {0};
3421d102b48SJeremy L Thompson 
3431d102b48SJeremy L Thompson   // Setup
3442b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorSetup_Blocked(op));
3451d102b48SJeremy L Thompson 
3460b454692Sjeremylt   // Restriction only operator
3470b454692Sjeremylt   if (impl->is_identity_restr_op) {
3482b730f8bSJeremy L Thompson     CeedCallBackend(CeedElemRestrictionApply(impl->blk_restr[0], CEED_NOTRANSPOSE, in_vec, impl->e_vecs_full[0], request));
3492b730f8bSJeremy L Thompson     CeedCallBackend(CeedElemRestrictionApply(impl->blk_restr[1], CEED_TRANSPOSE, impl->e_vecs_full[0], out_vec, request));
3500b454692Sjeremylt     return CEED_ERROR_SUCCESS;
3510b454692Sjeremylt   }
3520b454692Sjeremylt 
3531d102b48SJeremy L Thompson   // Input Evecs and Restriction
3542b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorSetupInputs_Blocked(num_input_fields, qf_input_fields, op_input_fields, in_vec, false, e_data_full, impl, request));
3551d102b48SJeremy L Thompson 
3561d102b48SJeremy L Thompson   // Output Evecs
357d1d35e2fSjeremylt   for (CeedInt i = 0; i < num_output_fields; i++) {
3582b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorGetArrayWrite(impl->e_vecs_full[i + impl->num_inputs], CEED_MEM_HOST, &e_data_full[i + num_input_fields]));
3591d102b48SJeremy L Thompson   }
3601d102b48SJeremy L Thompson 
3611d102b48SJeremy L Thompson   // Loop through elements
362d1d35e2fSjeremylt   for (CeedInt e = 0; e < num_blks * blk_size; e += blk_size) {
3631d102b48SJeremy L Thompson     // Output pointers
364d1d35e2fSjeremylt     for (CeedInt i = 0; i < num_output_fields; i++) {
3652b730f8bSJeremy L Thompson       CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode));
366d1d35e2fSjeremylt       if (eval_mode == CEED_EVAL_NONE) {
3672b730f8bSJeremy L Thompson         CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[i], &size));
3682b730f8bSJeremy 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]));
3691d102b48SJeremy L Thompson       }
3701d102b48SJeremy L Thompson     }
3711d102b48SJeremy L Thompson 
37216911fdaSjeremylt     // Input basis apply
3732b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorInputBasis_Blocked(e, Q, qf_input_fields, op_input_fields, num_input_fields, blk_size, false, e_data_full, impl));
37416911fdaSjeremylt 
3751d102b48SJeremy L Thompson     // Q function
3760b454692Sjeremylt     if (!impl->is_identity_qf) {
3772b730f8bSJeremy L Thompson       CeedCallBackend(CeedQFunctionApply(qf, Q * blk_size, impl->q_vecs_in, impl->q_vecs_out));
37816911fdaSjeremylt     }
3791d102b48SJeremy L Thompson 
3801d102b48SJeremy L Thompson     // Output basis apply
3812b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorOutputBasis_Blocked(e, Q, qf_output_fields, op_output_fields, blk_size, num_input_fields, num_output_fields, op,
3822b730f8bSJeremy L Thompson                                                     e_data_full, impl));
3831d102b48SJeremy L Thompson   }
38489c6efa4Sjeremylt 
38589c6efa4Sjeremylt   // Output restriction
386d1d35e2fSjeremylt   for (CeedInt i = 0; i < num_output_fields; i++) {
38789c6efa4Sjeremylt     // Restore evec
3882b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorRestoreArray(impl->e_vecs_full[i + impl->num_inputs], &e_data_full[i + num_input_fields]));
389d1bcdac9Sjeremylt     // Get output vector
3902b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec));
39189c6efa4Sjeremylt     // Active
3922b730f8bSJeremy L Thompson     if (vec == CEED_VECTOR_ACTIVE) vec = out_vec;
3934a2e7687Sjeremylt     // Restrict
3942b730f8bSJeremy L Thompson     CeedCallBackend(
3952b730f8bSJeremy L Thompson         CeedElemRestrictionApply(impl->blk_restr[i + impl->num_inputs], CEED_TRANSPOSE, impl->e_vecs_full[i + impl->num_inputs], vec, request));
3964a2e7687Sjeremylt   }
3974a2e7687Sjeremylt 
3984a2e7687Sjeremylt   // Restore input arrays
3992b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorRestoreInputs_Blocked(num_input_fields, qf_input_fields, op_input_fields, false, e_data_full, impl));
4001d102b48SJeremy L Thompson 
401e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
4021d102b48SJeremy L Thompson }
4031d102b48SJeremy L Thompson 
404f10650afSjeremylt //------------------------------------------------------------------------------
40570a7ffb3SJeremy L Thompson // Core code for assembling linear QFunction
406f10650afSjeremylt //------------------------------------------------------------------------------
4072b730f8bSJeremy L Thompson static inline int CeedOperatorLinearAssembleQFunctionCore_Blocked(CeedOperator op, bool build_objects, CeedVector *assembled,
408a0162de9SJeremy L Thompson                                                                   CeedElemRestriction *rstr, CeedRequest *request) {
4091d102b48SJeremy L Thompson   CeedOperator_Blocked *impl;
4102b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetData(op, &impl));
411d1d35e2fSjeremylt   const CeedInt blk_size = 8;
412d1d35e2fSjeremylt   CeedInt       Q, num_input_fields, num_output_fields, num_elem, size;
4132b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem));
4142b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q));
415d1d35e2fSjeremylt   CeedInt       num_blks = (num_elem / blk_size) + !!(num_elem % blk_size);
4161d102b48SJeremy L Thompson   CeedQFunction qf;
4172b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetQFunction(op, &qf));
418d1d35e2fSjeremylt   CeedOperatorField *op_input_fields, *op_output_fields;
4192b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields));
420d1d35e2fSjeremylt   CeedQFunctionField *qf_input_fields, *qf_output_fields;
4212b730f8bSJeremy L Thompson   CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields));
4224fc1f125SJeremy L Thompson   CeedVector  vec, l_vec = impl->qf_l_vec;
4232b730f8bSJeremy L Thompson   CeedInt     num_active_in = impl->num_active_in, num_active_out = impl->num_active_out;
424d2643443SJeremy L Thompson   CeedSize    q_size;
425bb219a0fSJeremy L Thompson   CeedVector *active_in = impl->qf_active_in;
4261d102b48SJeremy L Thompson   CeedScalar *a, *tmp;
4271d102b48SJeremy L Thompson   Ceed        ceed;
4282b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetCeed(op, &ceed));
4294fc1f125SJeremy L Thompson   CeedScalar *e_data_full[2 * CEED_FIELD_MAX] = {0};
4301d102b48SJeremy L Thompson 
4311d102b48SJeremy L Thompson   // Setup
4322b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorSetup_Blocked(op));
4331d102b48SJeremy L Thompson 
43416911fdaSjeremylt   // Check for identity
4356574a04fSJeremy L Thompson   CeedCheck(!impl->is_identity_qf, ceed, CEED_ERROR_BACKEND, "Assembling identity QFunctions not supported");
43616911fdaSjeremylt 
4371d102b48SJeremy L Thompson   // Input Evecs and Restriction
4382b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorSetupInputs_Blocked(num_input_fields, qf_input_fields, op_input_fields, NULL, true, e_data_full, impl, request));
4391d102b48SJeremy L Thompson 
4401d102b48SJeremy L Thompson   // Count number of active input fields
441bb219a0fSJeremy L Thompson   if (!num_active_in) {
442d1d35e2fSjeremylt     for (CeedInt i = 0; i < num_input_fields; i++) {
4431d102b48SJeremy L Thompson       // Get input vector
4442b730f8bSJeremy L Thompson       CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec));
4451d102b48SJeremy L Thompson       // Check if active input
4461d102b48SJeremy L Thompson       if (vec == CEED_VECTOR_ACTIVE) {
4472b730f8bSJeremy L Thompson         CeedCallBackend(CeedQFunctionFieldGetSize(qf_input_fields[i], &size));
4482b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorSetValue(impl->q_vecs_in[i], 0.0));
4492b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorGetArray(impl->q_vecs_in[i], CEED_MEM_HOST, &tmp));
4502b730f8bSJeremy L Thompson         CeedCallBackend(CeedRealloc(num_active_in + size, &active_in));
4511d102b48SJeremy L Thompson         for (CeedInt field = 0; field < size; field++) {
452d2643443SJeremy L Thompson           q_size = (CeedSize)Q * blk_size;
4532b730f8bSJeremy L Thompson           CeedCallBackend(CeedVectorCreate(ceed, q_size, &active_in[num_active_in + field]));
4542b730f8bSJeremy L Thompson           CeedCallBackend(CeedVectorSetArray(active_in[num_active_in + field], CEED_MEM_HOST, CEED_USE_POINTER, &tmp[field * Q * blk_size]));
4551d102b48SJeremy L Thompson         }
456d1d35e2fSjeremylt         num_active_in += size;
4572b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorRestoreArray(impl->q_vecs_in[i], &tmp));
4581d102b48SJeremy L Thompson       }
4591d102b48SJeremy L Thompson     }
4604fc1f125SJeremy L Thompson     impl->num_active_in = num_active_in;
461bb219a0fSJeremy L Thompson     impl->qf_active_in  = active_in;
462bb219a0fSJeremy L Thompson   }
4631d102b48SJeremy L Thompson 
4641d102b48SJeremy L Thompson   // Count number of active output fields
465bb219a0fSJeremy L Thompson   if (!num_active_out) {
466d1d35e2fSjeremylt     for (CeedInt i = 0; i < num_output_fields; i++) {
4671d102b48SJeremy L Thompson       // Get output vector
4682b730f8bSJeremy L Thompson       CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec));
4691d102b48SJeremy L Thompson       // Check if active output
4701d102b48SJeremy L Thompson       if (vec == CEED_VECTOR_ACTIVE) {
4712b730f8bSJeremy L Thompson         CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[i], &size));
472d1d35e2fSjeremylt         num_active_out += size;
4731d102b48SJeremy L Thompson       }
4741d102b48SJeremy L Thompson     }
4754fc1f125SJeremy L Thompson     impl->num_active_out = num_active_out;
476bb219a0fSJeremy L Thompson   }
4771d102b48SJeremy L Thompson 
4781d102b48SJeremy L Thompson   // Check sizes
4796574a04fSJeremy L Thompson   CeedCheck(num_active_in > 0 && num_active_out > 0, ceed, CEED_ERROR_BACKEND, "Cannot assemble QFunction without active inputs and outputs");
4801d102b48SJeremy L Thompson 
481d1d35e2fSjeremylt   // Setup Lvec
4824fc1f125SJeremy L Thompson   if (!l_vec) {
483d2643443SJeremy L Thompson     CeedSize l_size = (CeedSize)num_blks * blk_size * Q * num_active_in * num_active_out;
4842b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorCreate(ceed, l_size, &l_vec));
4854fc1f125SJeremy L Thompson     impl->qf_l_vec = l_vec;
486bb219a0fSJeremy L Thompson   }
4872b730f8bSJeremy L Thompson   CeedCallBackend(CeedVectorGetArrayWrite(l_vec, CEED_MEM_HOST, &a));
4881d102b48SJeremy L Thompson 
48970a7ffb3SJeremy L Thompson   // Build objects if needed
490d1d35e2fSjeremylt   CeedInt strides[3] = {1, Q, num_active_in * num_active_out * Q};
49170a7ffb3SJeremy L Thompson   if (build_objects) {
49270a7ffb3SJeremy L Thompson     // Create output restriction
4932b730f8bSJeremy L Thompson     CeedCallBackend(CeedElemRestrictionCreateStrided(ceed, num_elem, Q, num_active_in * num_active_out, num_active_in * num_active_out * num_elem * Q,
4942b730f8bSJeremy L Thompson                                                      strides, rstr));
4951d102b48SJeremy L Thompson     // Create assembled vector
496d2643443SJeremy L Thompson     CeedSize l_size = (CeedSize)num_elem * Q * num_active_in * num_active_out;
4972b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorCreate(ceed, l_size, assembled));
49870a7ffb3SJeremy L Thompson   }
4991d102b48SJeremy L Thompson 
5001d102b48SJeremy L Thompson   // Loop through elements
501d1d35e2fSjeremylt   for (CeedInt e = 0; e < num_blks * blk_size; e += blk_size) {
5021d102b48SJeremy L Thompson     // Input basis apply
5032b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorInputBasis_Blocked(e, Q, qf_input_fields, op_input_fields, num_input_fields, blk_size, true, e_data_full, impl));
5041d102b48SJeremy L Thompson 
5051d102b48SJeremy L Thompson     // Assemble QFunction
506d1d35e2fSjeremylt     for (CeedInt in = 0; in < num_active_in; in++) {
5071d102b48SJeremy L Thompson       // Set Inputs
5082b730f8bSJeremy L Thompson       CeedCallBackend(CeedVectorSetValue(active_in[in], 1.0));
509d1d35e2fSjeremylt       if (num_active_in > 1) {
5102b730f8bSJeremy L Thompson         CeedCallBackend(CeedVectorSetValue(active_in[(in + num_active_in - 1) % num_active_in], 0.0));
51142ea3801Sjeremylt       }
5121d102b48SJeremy L Thompson       // Set Outputs
513d1d35e2fSjeremylt       for (CeedInt out = 0; out < num_output_fields; out++) {
5141d102b48SJeremy L Thompson         // Get output vector
5152b730f8bSJeremy L Thompson         CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[out], &vec));
5161d102b48SJeremy L Thompson         // Check if active output
5171d102b48SJeremy L Thompson         if (vec == CEED_VECTOR_ACTIVE) {
5182b730f8bSJeremy L Thompson           CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[out], CEED_MEM_HOST, CEED_USE_POINTER, a));
5192b730f8bSJeremy L Thompson           CeedCallBackend(CeedQFunctionFieldGetSize(qf_output_fields[out], &size));
520d1d35e2fSjeremylt           a += size * Q * blk_size;  // Advance the pointer by the size of the output
5211d102b48SJeremy L Thompson         }
5221d102b48SJeremy L Thompson       }
5231d102b48SJeremy L Thompson       // Apply QFunction
5242b730f8bSJeremy L Thompson       CeedCallBackend(CeedQFunctionApply(qf, Q * blk_size, impl->q_vecs_in, impl->q_vecs_out));
5254a2e7687Sjeremylt     }
5264a2e7687Sjeremylt   }
5274a2e7687Sjeremylt 
5281d102b48SJeremy L Thompson   // Un-set output Qvecs to prevent accidental overwrite of Assembled
529d1d35e2fSjeremylt   for (CeedInt out = 0; out < num_output_fields; out++) {
5301d102b48SJeremy L Thompson     // Get output vector
5312b730f8bSJeremy L Thompson     CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[out], &vec));
5321d102b48SJeremy L Thompson     // Check if active output
5331d102b48SJeremy L Thompson     if (vec == CEED_VECTOR_ACTIVE) {
5342b730f8bSJeremy L Thompson       CeedCallBackend(CeedVectorSetArray(impl->q_vecs_out[out], CEED_MEM_HOST, CEED_COPY_VALUES, NULL));
5351d102b48SJeremy L Thompson     }
5361d102b48SJeremy L Thompson   }
5371d102b48SJeremy L Thompson 
5381d102b48SJeremy L Thompson   // Restore input arrays
5392b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorRestoreInputs_Blocked(num_input_fields, qf_input_fields, op_input_fields, true, e_data_full, impl));
5401d102b48SJeremy L Thompson 
5411d102b48SJeremy L Thompson   // Output blocked restriction
5422b730f8bSJeremy L Thompson   CeedCallBackend(CeedVectorRestoreArray(l_vec, &a));
5432b730f8bSJeremy L Thompson   CeedCallBackend(CeedVectorSetValue(*assembled, 0.0));
544bb219a0fSJeremy L Thompson   CeedElemRestriction blk_rstr = impl->qf_blk_rstr;
545bb219a0fSJeremy L Thompson   if (!impl->qf_blk_rstr) {
5462b730f8bSJeremy L Thompson     CeedCallBackend(CeedElemRestrictionCreateBlockedStrided(ceed, num_elem, Q, blk_size, num_active_in * num_active_out,
5472b730f8bSJeremy L Thompson                                                             num_active_in * num_active_out * num_elem * Q, strides, &blk_rstr));
548bb219a0fSJeremy L Thompson     impl->qf_blk_rstr = blk_rstr;
549bb219a0fSJeremy L Thompson   }
5502b730f8bSJeremy L Thompson   CeedCallBackend(CeedElemRestrictionApply(blk_rstr, CEED_TRANSPOSE, l_vec, *assembled, request));
5511d102b48SJeremy L Thompson 
552e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
5534a2e7687Sjeremylt }
5544a2e7687Sjeremylt 
555f10650afSjeremylt //------------------------------------------------------------------------------
55670a7ffb3SJeremy L Thompson // Assemble Linear QFunction
55770a7ffb3SJeremy L Thompson //------------------------------------------------------------------------------
5582b730f8bSJeremy L Thompson static int CeedOperatorLinearAssembleQFunction_Blocked(CeedOperator op, CeedVector *assembled, CeedElemRestriction *rstr, CeedRequest *request) {
5592b730f8bSJeremy L Thompson   return CeedOperatorLinearAssembleQFunctionCore_Blocked(op, true, assembled, rstr, request);
56070a7ffb3SJeremy L Thompson }
56170a7ffb3SJeremy L Thompson 
56270a7ffb3SJeremy L Thompson //------------------------------------------------------------------------------
56370a7ffb3SJeremy L Thompson // Update Assembled Linear QFunction
56470a7ffb3SJeremy L Thompson //------------------------------------------------------------------------------
5652b730f8bSJeremy L Thompson static int CeedOperatorLinearAssembleQFunctionUpdate_Blocked(CeedOperator op, CeedVector assembled, CeedElemRestriction rstr, CeedRequest *request) {
5662b730f8bSJeremy L Thompson   return CeedOperatorLinearAssembleQFunctionCore_Blocked(op, false, &assembled, &rstr, request);
56770a7ffb3SJeremy L Thompson }
56870a7ffb3SJeremy L Thompson 
56970a7ffb3SJeremy L Thompson //------------------------------------------------------------------------------
570f10650afSjeremylt // Operator Destroy
571f10650afSjeremylt //------------------------------------------------------------------------------
572f10650afSjeremylt static int CeedOperatorDestroy_Blocked(CeedOperator op) {
573f10650afSjeremylt   CeedOperator_Blocked *impl;
5742b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetData(op, &impl));
575f10650afSjeremylt 
5764fc1f125SJeremy L Thompson   for (CeedInt i = 0; i < impl->num_inputs + impl->num_outputs; i++) {
5772b730f8bSJeremy L Thompson     CeedCallBackend(CeedElemRestrictionDestroy(&impl->blk_restr[i]));
5782b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorDestroy(&impl->e_vecs_full[i]));
579f10650afSjeremylt   }
5802b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl->blk_restr));
5812b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl->e_vecs_full));
5822b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl->input_states));
583f10650afSjeremylt 
5844fc1f125SJeremy L Thompson   for (CeedInt i = 0; i < impl->num_inputs; i++) {
5852b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorDestroy(&impl->e_vecs_in[i]));
5862b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorDestroy(&impl->q_vecs_in[i]));
587f10650afSjeremylt   }
5882b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl->e_vecs_in));
5892b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl->q_vecs_in));
590f10650afSjeremylt 
5914fc1f125SJeremy L Thompson   for (CeedInt i = 0; i < impl->num_outputs; i++) {
5922b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorDestroy(&impl->e_vecs_out[i]));
5932b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorDestroy(&impl->q_vecs_out[i]));
594f10650afSjeremylt   }
5952b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl->e_vecs_out));
5962b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl->q_vecs_out));
597f10650afSjeremylt 
598bb219a0fSJeremy L Thompson   // QFunction assembly data
5994fc1f125SJeremy L Thompson   for (CeedInt i = 0; i < impl->num_active_in; i++) {
6002b730f8bSJeremy L Thompson     CeedCallBackend(CeedVectorDestroy(&impl->qf_active_in[i]));
601bb219a0fSJeremy L Thompson   }
6022b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl->qf_active_in));
6032b730f8bSJeremy L Thompson   CeedCallBackend(CeedVectorDestroy(&impl->qf_l_vec));
6042b730f8bSJeremy L Thompson   CeedCallBackend(CeedElemRestrictionDestroy(&impl->qf_blk_rstr));
605bb219a0fSJeremy L Thompson 
6062b730f8bSJeremy L Thompson   CeedCallBackend(CeedFree(&impl));
607e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
608f10650afSjeremylt }
609f10650afSjeremylt 
610f10650afSjeremylt //------------------------------------------------------------------------------
611f10650afSjeremylt // Operator Create
612f10650afSjeremylt //------------------------------------------------------------------------------
6134a2e7687Sjeremylt int CeedOperatorCreate_Blocked(CeedOperator op) {
614fe2413ffSjeremylt   Ceed ceed;
6152b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorGetCeed(op, &ceed));
6164ce2993fSjeremylt   CeedOperator_Blocked *impl;
6174a2e7687Sjeremylt 
6182b730f8bSJeremy L Thompson   CeedCallBackend(CeedCalloc(1, &impl));
6192b730f8bSJeremy L Thompson   CeedCallBackend(CeedOperatorSetData(op, impl));
620fe2413ffSjeremylt 
6212b730f8bSJeremy L Thompson   CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunction", CeedOperatorLinearAssembleQFunction_Blocked));
6222b730f8bSJeremy L Thompson   CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "LinearAssembleQFunctionUpdate", CeedOperatorLinearAssembleQFunctionUpdate_Blocked));
6232b730f8bSJeremy L Thompson   CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAdd_Blocked));
6242b730f8bSJeremy L Thompson   CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Blocked));
625e15f9bd0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
6264a2e7687Sjeremylt }
6272a86cc9dSSebastian Grimberg 
628f10650afSjeremylt //------------------------------------------------------------------------------
629