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. 30d0321e0SJeremy L Thompson // 43d8e8822SJeremy L Thompson // SPDX-License-Identifier: BSD-2-Clause 50d0321e0SJeremy L Thompson // 63d8e8822SJeremy L Thompson // This file is part of CEED: http://github.com/ceed 70d0321e0SJeremy L Thompson 849aac155SJeremy L Thompson #include <ceed.h> 90d0321e0SJeremy L Thompson #include <ceed/backend.h> 1049aac155SJeremy L Thompson #include <ceed/jit-source/hip/hip-types.h> 110d0321e0SJeremy L Thompson #include <hip/hip_runtime.h> 122b730f8bSJeremy L Thompson 1349aac155SJeremy L Thompson #include "../hip/ceed-hip-common.h" 140d0321e0SJeremy L Thompson #include "../hip/ceed-hip-compile.h" 152b730f8bSJeremy L Thompson #include "ceed-hip-ref-qfunction-load.h" 162b730f8bSJeremy L Thompson #include "ceed-hip-ref.h" 170d0321e0SJeremy L Thompson 180d0321e0SJeremy L Thompson //------------------------------------------------------------------------------ 190d0321e0SJeremy L Thompson // Apply QFunction 200d0321e0SJeremy L Thompson //------------------------------------------------------------------------------ 212b730f8bSJeremy L Thompson static int CeedQFunctionApply_Hip(CeedQFunction qf, CeedInt Q, CeedVector *U, CeedVector *V) { 220d0321e0SJeremy L Thompson Ceed ceed; 23*b7453713SJeremy L Thompson Ceed_Hip *ceed_Hip; 24*b7453713SJeremy L Thompson CeedInt num_input_fields, num_output_fields; 25*b7453713SJeremy L Thompson CeedQFunction_Hip *data; 26*b7453713SJeremy L Thompson 272b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetCeed(qf, &ceed)); 280d0321e0SJeremy L Thompson 290d0321e0SJeremy L Thompson // Build and compile kernel, if not done 30eb7e6cafSJeremy L Thompson CeedCallBackend(CeedQFunctionBuildKernel_Hip_ref(qf)); 310d0321e0SJeremy L Thompson 322b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetData(qf, &data)); 332b730f8bSJeremy L Thompson CeedCallBackend(CeedGetData(ceed, &ceed_Hip)); 342b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetNumArgs(qf, &num_input_fields, &num_output_fields)); 35b2165e7aSSebastian Grimberg const int block_size = ceed_Hip->opt_block_size; 360d0321e0SJeremy L Thompson 370d0321e0SJeremy L Thompson // Read vectors 38437930d1SJeremy L Thompson for (CeedInt i = 0; i < num_input_fields; i++) { 392b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorGetArrayRead(U[i], CEED_MEM_DEVICE, &data->fields.inputs[i])); 400d0321e0SJeremy L Thompson } 41437930d1SJeremy L Thompson for (CeedInt i = 0; i < num_output_fields; i++) { 422b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorGetArrayWrite(V[i], CEED_MEM_DEVICE, &data->fields.outputs[i])); 430d0321e0SJeremy L Thompson } 440d0321e0SJeremy L Thompson 450d0321e0SJeremy L Thompson // Get context data 462b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetInnerContextData(qf, CEED_MEM_DEVICE, &data->d_c)); 470d0321e0SJeremy L Thompson 480d0321e0SJeremy L Thompson // Run kernel 490d0321e0SJeremy L Thompson void *args[] = {&data->d_c, (void *)&Q, &data->fields}; 50*b7453713SJeremy L Thompson 51b2165e7aSSebastian Grimberg CeedCallBackend(CeedRunKernel_Hip(ceed, data->QFunction, CeedDivUpInt(Q, block_size), block_size, args)); 520d0321e0SJeremy L Thompson 530d0321e0SJeremy L Thompson // Restore vectors 54437930d1SJeremy L Thompson for (CeedInt i = 0; i < num_input_fields; i++) { 552b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorRestoreArrayRead(U[i], &data->fields.inputs[i])); 560d0321e0SJeremy L Thompson } 57437930d1SJeremy L Thompson for (CeedInt i = 0; i < num_output_fields; i++) { 582b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorRestoreArray(V[i], &data->fields.outputs[i])); 590d0321e0SJeremy L Thompson } 600d0321e0SJeremy L Thompson 610d0321e0SJeremy L Thompson // Restore context 622b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionRestoreInnerContextData(qf, &data->d_c)); 630d0321e0SJeremy L Thompson return CEED_ERROR_SUCCESS; 640d0321e0SJeremy L Thompson } 650d0321e0SJeremy L Thompson 660d0321e0SJeremy L Thompson //------------------------------------------------------------------------------ 670d0321e0SJeremy L Thompson // Destroy QFunction 680d0321e0SJeremy L Thompson //------------------------------------------------------------------------------ 690d0321e0SJeremy L Thompson static int CeedQFunctionDestroy_Hip(CeedQFunction qf) { 700d0321e0SJeremy L Thompson Ceed ceed; 71*b7453713SJeremy L Thompson CeedQFunction_Hip *data; 72*b7453713SJeremy L Thompson 73*b7453713SJeremy L Thompson CeedCallBackend(CeedQFunctionGetData(qf, &data)); 742b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetCeed(qf, &ceed)); 752b730f8bSJeremy L Thompson if (data->module) CeedCallHip(ceed, hipModuleUnload(data->module)); 762b730f8bSJeremy L Thompson CeedCallBackend(CeedFree(&data)); 770d0321e0SJeremy L Thompson return CEED_ERROR_SUCCESS; 780d0321e0SJeremy L Thompson } 790d0321e0SJeremy L Thompson 800d0321e0SJeremy L Thompson //------------------------------------------------------------------------------ 810d0321e0SJeremy L Thompson // Create QFunction 820d0321e0SJeremy L Thompson //------------------------------------------------------------------------------ 830d0321e0SJeremy L Thompson int CeedQFunctionCreate_Hip(CeedQFunction qf) { 840d0321e0SJeremy L Thompson Ceed ceed; 85*b7453713SJeremy L Thompson CeedInt num_input_fields, num_output_fields; 860d0321e0SJeremy L Thompson CeedQFunction_Hip *data; 87*b7453713SJeremy L Thompson 88*b7453713SJeremy L Thompson CeedQFunctionGetCeed(qf, &ceed); 892b730f8bSJeremy L Thompson CeedCallBackend(CeedCalloc(1, &data)); 902b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionSetData(qf, data)); 912b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetNumArgs(qf, &num_input_fields, &num_output_fields)); 920d0321e0SJeremy L Thompson 930d0321e0SJeremy L Thompson // Read QFunction source 942b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetKernelName(qf, &data->qfunction_name)); 9523d4529eSJeremy L Thompson CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading QFunction User Source -----\n"); 962b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionLoadSourceToBuffer(qf, &data->qfunction_source)); 9723d4529eSJeremy L Thompson CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading QFunction User Source Complete! -----\n"); 980d0321e0SJeremy L Thompson 990d0321e0SJeremy L Thompson // Register backend functions 1002b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "QFunction", qf, "Apply", CeedQFunctionApply_Hip)); 1012b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "QFunction", qf, "Destroy", CeedQFunctionDestroy_Hip)); 1020d0321e0SJeremy L Thompson return CEED_ERROR_SUCCESS; 1030d0321e0SJeremy L Thompson } 1042a86cc9dSSebastian Grimberg 1050d0321e0SJeremy L Thompson //------------------------------------------------------------------------------ 106