xref: /libCEED/rust/libceed-sys/c-src/backends/hip-ref/ceed-hip-ref-qfunction.c (revision 46dc07349c44057b9efae59cd6a2d41f419237bd)
10d0321e0SJeremy L Thompson // Copyright (c) 2017-2018, Lawrence Livermore National Security, LLC.
20d0321e0SJeremy L Thompson // Produced at the Lawrence Livermore National Laboratory. LLNL-CODE-734707.
30d0321e0SJeremy L Thompson // All Rights reserved. See files LICENSE and NOTICE for details.
40d0321e0SJeremy L Thompson //
50d0321e0SJeremy L Thompson // This file is part of CEED, a collection of benchmarks, miniapps, software
60d0321e0SJeremy L Thompson // libraries and APIs for efficient high-order finite element and spectral
70d0321e0SJeremy L Thompson // element discretizations for exascale applications. For more information and
80d0321e0SJeremy L Thompson // source code availability see http://github.com/ceed.
90d0321e0SJeremy L Thompson //
100d0321e0SJeremy L Thompson // The CEED research is supported by the Exascale Computing Project 17-SC-20-SC,
110d0321e0SJeremy L Thompson // a collaborative effort of two U.S. Department of Energy organizations (Office
120d0321e0SJeremy L Thompson // of Science and the National Nuclear Security Administration) responsible for
130d0321e0SJeremy L Thompson // the planning and preparation of a capable exascale ecosystem, including
140d0321e0SJeremy L Thompson // software, applications, hardware, advanced system engineering and early
150d0321e0SJeremy L Thompson // testbed platforms, in support of the nation's exascale computing imperative.
160d0321e0SJeremy L Thompson 
170d0321e0SJeremy L Thompson #include <ceed/ceed.h>
180d0321e0SJeremy L Thompson #include <ceed/backend.h>
190d0321e0SJeremy L Thompson #include <hip/hip_runtime.h>
200d0321e0SJeremy L Thompson #include <stdio.h>
210d0321e0SJeremy L Thompson #include <string.h>
220d0321e0SJeremy L Thompson #include "ceed-hip-ref.h"
230d0321e0SJeremy L Thompson #include "ceed-hip-ref-qfunction-load.h"
240d0321e0SJeremy L Thompson #include "../hip/ceed-hip-compile.h"
250d0321e0SJeremy L Thompson 
260d0321e0SJeremy L Thompson //------------------------------------------------------------------------------
270d0321e0SJeremy L Thompson // Apply QFunction
280d0321e0SJeremy L Thompson //------------------------------------------------------------------------------
290d0321e0SJeremy L Thompson static int CeedQFunctionApply_Hip(CeedQFunction qf, CeedInt Q,
300d0321e0SJeremy L Thompson                                   CeedVector *U, CeedVector *V) {
310d0321e0SJeremy L Thompson   int ierr;
320d0321e0SJeremy L Thompson   Ceed ceed;
330d0321e0SJeremy L Thompson   ierr = CeedQFunctionGetCeed(qf, &ceed); CeedChkBackend(ierr);
340d0321e0SJeremy L Thompson 
350d0321e0SJeremy L Thompson   // Build and compile kernel, if not done
360d0321e0SJeremy L Thompson   ierr = CeedHipBuildQFunction(qf); CeedChkBackend(ierr);
370d0321e0SJeremy L Thompson 
380d0321e0SJeremy L Thompson   CeedQFunction_Hip *data;
390d0321e0SJeremy L Thompson   ierr = CeedQFunctionGetData(qf, &data); CeedChkBackend(ierr);
400d0321e0SJeremy L Thompson   Ceed_Hip *ceed_Hip;
410d0321e0SJeremy L Thompson   ierr = CeedGetData(ceed, &ceed_Hip); CeedChkBackend(ierr);
42437930d1SJeremy L Thompson   CeedInt num_input_fields, num_output_fields;
43437930d1SJeremy L Thompson   ierr = CeedQFunctionGetNumArgs(qf, &num_input_fields, &num_output_fields);
440d0321e0SJeremy L Thompson   CeedChkBackend(ierr);
450d0321e0SJeremy L Thompson   const int blocksize = ceed_Hip->opt_block_size;
460d0321e0SJeremy L Thompson 
470d0321e0SJeremy L Thompson   // Read vectors
48437930d1SJeremy L Thompson   for (CeedInt i = 0; i < num_input_fields; i++) {
490d0321e0SJeremy L Thompson     ierr = CeedVectorGetArrayRead(U[i], CEED_MEM_DEVICE, &data->fields.inputs[i]);
500d0321e0SJeremy L Thompson     CeedChkBackend(ierr);
510d0321e0SJeremy L Thompson   }
52437930d1SJeremy L Thompson   for (CeedInt i = 0; i < num_output_fields; i++) {
530d0321e0SJeremy L Thompson     ierr = CeedVectorGetArrayWrite(V[i], CEED_MEM_DEVICE, &data->fields.outputs[i]);
540d0321e0SJeremy L Thompson     CeedChkBackend(ierr);
550d0321e0SJeremy L Thompson   }
560d0321e0SJeremy L Thompson 
570d0321e0SJeremy L Thompson   // Get context data
580d0321e0SJeremy L Thompson   CeedQFunctionContext ctx;
590d0321e0SJeremy L Thompson   ierr = CeedQFunctionGetInnerContext(qf, &ctx); CeedChkBackend(ierr);
600d0321e0SJeremy L Thompson   if (ctx) {
610d0321e0SJeremy L Thompson     ierr = CeedQFunctionContextGetData(ctx, CEED_MEM_DEVICE, &data->d_c);
620d0321e0SJeremy L Thompson     CeedChkBackend(ierr);
630d0321e0SJeremy L Thompson   }
640d0321e0SJeremy L Thompson 
650d0321e0SJeremy L Thompson   // Run kernel
660d0321e0SJeremy L Thompson   void *args[] = {&data->d_c, (void *) &Q, &data->fields};
67437930d1SJeremy L Thompson   ierr = CeedRunKernelHip(ceed, data->QFunction, CeedDivUpInt(Q, blocksize),
680d0321e0SJeremy L Thompson                           blocksize, args); CeedChkBackend(ierr);
690d0321e0SJeremy L Thompson 
700d0321e0SJeremy L Thompson   // Restore vectors
71437930d1SJeremy L Thompson   for (CeedInt i = 0; i < num_input_fields; i++) {
720d0321e0SJeremy L Thompson     ierr = CeedVectorRestoreArrayRead(U[i], &data->fields.inputs[i]);
730d0321e0SJeremy L Thompson     CeedChkBackend(ierr);
740d0321e0SJeremy L Thompson   }
75437930d1SJeremy L Thompson   for (CeedInt i = 0; i < num_output_fields; i++) {
760d0321e0SJeremy L Thompson     ierr = CeedVectorRestoreArray(V[i], &data->fields.outputs[i]);
770d0321e0SJeremy L Thompson     CeedChkBackend(ierr);
780d0321e0SJeremy L Thompson   }
790d0321e0SJeremy L Thompson 
800d0321e0SJeremy L Thompson   // Restore context
810d0321e0SJeremy L Thompson   if (ctx) {
820d0321e0SJeremy L Thompson     ierr = CeedQFunctionContextRestoreData(ctx, &data->d_c);
830d0321e0SJeremy L Thompson     CeedChkBackend(ierr);
840d0321e0SJeremy L Thompson   }
850d0321e0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
860d0321e0SJeremy L Thompson }
870d0321e0SJeremy L Thompson 
880d0321e0SJeremy L Thompson //------------------------------------------------------------------------------
890d0321e0SJeremy L Thompson // Destroy QFunction
900d0321e0SJeremy L Thompson //------------------------------------------------------------------------------
910d0321e0SJeremy L Thompson static int CeedQFunctionDestroy_Hip(CeedQFunction qf) {
920d0321e0SJeremy L Thompson   int ierr;
930d0321e0SJeremy L Thompson   CeedQFunction_Hip *data;
940d0321e0SJeremy L Thompson   ierr = CeedQFunctionGetData(qf, &data); CeedChkBackend(ierr);
950d0321e0SJeremy L Thompson   Ceed ceed;
960d0321e0SJeremy L Thompson   ierr = CeedQFunctionGetCeed(qf, &ceed); CeedChkBackend(ierr);
970d0321e0SJeremy L Thompson   if  (data->module)
980d0321e0SJeremy L Thompson     CeedChk_Hip(ceed, hipModuleUnload(data->module));
990d0321e0SJeremy L Thompson   ierr = CeedFree(&data); CeedChkBackend(ierr);
100437930d1SJeremy L Thompson 
1010d0321e0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
1020d0321e0SJeremy L Thompson }
1030d0321e0SJeremy L Thompson 
1040d0321e0SJeremy L Thompson //------------------------------------------------------------------------------
1050d0321e0SJeremy L Thompson // Create QFunction
1060d0321e0SJeremy L Thompson //------------------------------------------------------------------------------
1070d0321e0SJeremy L Thompson int CeedQFunctionCreate_Hip(CeedQFunction qf) {
1080d0321e0SJeremy L Thompson   int ierr;
1090d0321e0SJeremy L Thompson   Ceed ceed;
1100d0321e0SJeremy L Thompson   CeedQFunctionGetCeed(qf, &ceed);
1110d0321e0SJeremy L Thompson   CeedQFunction_Hip *data;
1120d0321e0SJeremy L Thompson   ierr = CeedCalloc(1,&data); CeedChkBackend(ierr);
1130d0321e0SJeremy L Thompson   ierr = CeedQFunctionSetData(qf, data); CeedChkBackend(ierr);
114437930d1SJeremy L Thompson   CeedInt num_input_fields, num_output_fields;
115437930d1SJeremy L Thompson   ierr = CeedQFunctionGetNumArgs(qf, &num_input_fields, &num_output_fields);
1160d0321e0SJeremy L Thompson   CeedChkBackend(ierr);
1170d0321e0SJeremy L Thompson 
1180d0321e0SJeremy L Thompson   // Read QFunction source
119437930d1SJeremy L Thompson   ierr = CeedQFunctionGetKernelName(qf, &data->qfunction_name);
1200d0321e0SJeremy L Thompson   CeedChkBackend(ierr);
121*46dc0734SJeremy L Thompson   CeedDebug256(ceed, 2, "----- Loading QFunction User Source -----\n");
122437930d1SJeremy L Thompson   ierr = CeedQFunctionLoadSourceToBuffer(qf, &data->qfunction_source);
1230d0321e0SJeremy L Thompson   CeedChkBackend(ierr);
124*46dc0734SJeremy L Thompson   CeedDebug256(ceed, 2, "----- Loading QFunction User Source Complete! -----\n");
1250d0321e0SJeremy L Thompson 
1260d0321e0SJeremy L Thompson   // Register backend functions
1270d0321e0SJeremy L Thompson   ierr = CeedSetBackendFunction(ceed, "QFunction", qf, "Apply",
1280d0321e0SJeremy L Thompson                                 CeedQFunctionApply_Hip); CeedChkBackend(ierr);
1290d0321e0SJeremy L Thompson   ierr = CeedSetBackendFunction(ceed, "QFunction", qf, "Destroy",
1300d0321e0SJeremy L Thompson                                 CeedQFunctionDestroy_Hip); CeedChkBackend(ierr);
1310d0321e0SJeremy L Thompson   return CEED_ERROR_SUCCESS;
1320d0321e0SJeremy L Thompson }
1330d0321e0SJeremy L Thompson //------------------------------------------------------------------------------
134