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. 37d8d0e25Snbeams // 43d8e8822SJeremy L Thompson // SPDX-License-Identifier: BSD-2-Clause 57d8d0e25Snbeams // 63d8e8822SJeremy L Thompson // This file is part of CEED: http://github.com/ceed 77d8d0e25Snbeams 8ec3da8bcSJed Brown #include <ceed/backend.h> 9*2b730f8bSJeremy L Thompson #include <ceed/ceed.h> 103d576824SJeremy L Thompson #include <stddef.h> 11*2b730f8bSJeremy L Thompson 127d8d0e25Snbeams #include "../hip/ceed-hip-compile.h" 13*2b730f8bSJeremy L Thompson #include "ceed-hip-gen-operator-build.h" 14*2b730f8bSJeremy L Thompson #include "ceed-hip-gen.h" 157d8d0e25Snbeams 167d8d0e25Snbeams //------------------------------------------------------------------------------ 177d8d0e25Snbeams // Destroy operator 187d8d0e25Snbeams //------------------------------------------------------------------------------ 197d8d0e25Snbeams static int CeedOperatorDestroy_Hip_gen(CeedOperator op) { 207d8d0e25Snbeams CeedOperator_Hip_gen *impl; 21*2b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetData(op, &impl)); 22*2b730f8bSJeremy L Thompson CeedCallBackend(CeedFree(&impl)); 23e15f9bd0SJeremy L Thompson return CEED_ERROR_SUCCESS; 247d8d0e25Snbeams } 257d8d0e25Snbeams 267d8d0e25Snbeams //------------------------------------------------------------------------------ 277d8d0e25Snbeams // Apply and add to output 287d8d0e25Snbeams //------------------------------------------------------------------------------ 29*2b730f8bSJeremy L Thompson static int CeedOperatorApplyAdd_Hip_gen(CeedOperator op, CeedVector input_vec, CeedVector output_vec, CeedRequest *request) { 307d8d0e25Snbeams Ceed ceed; 31*2b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 327d8d0e25Snbeams CeedOperator_Hip_gen *data; 33*2b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetData(op, &data)); 347d8d0e25Snbeams CeedQFunction qf; 357d8d0e25Snbeams CeedQFunction_Hip_gen *qf_data; 36*2b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 37*2b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetData(qf, &qf_data)); 389e201c85SYohann CeedInt num_elem, num_input_fields, num_output_fields; 39*2b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 409e201c85SYohann CeedOperatorField *op_input_fields, *op_output_fields; 41*2b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 429e201c85SYohann CeedQFunctionField *qf_input_fields, *qf_output_fields; 43*2b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 449e201c85SYohann CeedEvalMode eval_mode; 459e201c85SYohann CeedVector vec, output_vecs[CEED_FIELD_MAX] = {}; 467d8d0e25Snbeams 477d8d0e25Snbeams // Creation of the operator 48*2b730f8bSJeremy L Thompson CeedCallBackend(CeedHipGenOperatorBuild(op)); 497d8d0e25Snbeams 507d8d0e25Snbeams // Input vectors 519e201c85SYohann for (CeedInt i = 0; i < num_input_fields; i++) { 52*2b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 539e201c85SYohann if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 549e201c85SYohann data->fields.inputs[i] = NULL; 557d8d0e25Snbeams } else { 567d8d0e25Snbeams // Get input vector 57*2b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 589e201c85SYohann if (vec == CEED_VECTOR_ACTIVE) vec = input_vec; 59*2b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorGetArrayRead(vec, CEED_MEM_DEVICE, &data->fields.inputs[i])); 607d8d0e25Snbeams } 617d8d0e25Snbeams } 627d8d0e25Snbeams 637d8d0e25Snbeams // Output vectors 649e201c85SYohann for (CeedInt i = 0; i < num_output_fields; i++) { 65*2b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 669e201c85SYohann if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 679e201c85SYohann data->fields.outputs[i] = NULL; 687d8d0e25Snbeams } else { 697d8d0e25Snbeams // Get output vector 70*2b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 719e201c85SYohann if (vec == CEED_VECTOR_ACTIVE) vec = output_vec; 729e201c85SYohann output_vecs[i] = vec; 737d8d0e25Snbeams // Check for multiple output modes 747d8d0e25Snbeams CeedInt index = -1; 757d8d0e25Snbeams for (CeedInt j = 0; j < i; j++) { 769e201c85SYohann if (vec == output_vecs[j]) { 777d8d0e25Snbeams index = j; 787d8d0e25Snbeams break; 797d8d0e25Snbeams } 807d8d0e25Snbeams } 817d8d0e25Snbeams if (index == -1) { 82*2b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorGetArray(vec, CEED_MEM_DEVICE, &data->fields.outputs[i])); 837d8d0e25Snbeams } else { 849e201c85SYohann data->fields.outputs[i] = data->fields.outputs[index]; 857d8d0e25Snbeams } 867d8d0e25Snbeams } 877d8d0e25Snbeams } 887d8d0e25Snbeams 897d8d0e25Snbeams // Get context data 90*2b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetInnerContextData(qf, CEED_MEM_DEVICE, &qf_data->d_c)); 917d8d0e25Snbeams 927d8d0e25Snbeams // Apply operator 93*2b730f8bSJeremy L Thompson void *opargs[] = {(void *)&num_elem, &qf_data->d_c, &data->indices, &data->fields, &data->B, &data->G, &data->W}; 947d8d0e25Snbeams const CeedInt dim = data->dim; 959e201c85SYohann const CeedInt Q_1d = data->Q_1d; 969e201c85SYohann const CeedInt P_1d = data->max_P_1d; 979e201c85SYohann const CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 98b3e1519bSnbeams CeedInt block_sizes[3]; 99*2b730f8bSJeremy L Thompson CeedCallBackend(BlockGridCalculate_Hip_gen(dim, num_elem, P_1d, Q_1d, block_sizes)); 1007d8d0e25Snbeams if (dim == 1) { 101*2b730f8bSJeremy L Thompson CeedInt grid = num_elem / block_sizes[2] + ((num_elem / block_sizes[2] * block_sizes[2] < num_elem) ? 1 : 0); 1029e201c85SYohann CeedInt sharedMem = block_sizes[2] * thread_1d * sizeof(CeedScalar); 103*2b730f8bSJeremy L Thompson CeedCallBackend(CeedRunKernelDimSharedHip(ceed, data->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 1047d8d0e25Snbeams } else if (dim == 2) { 105*2b730f8bSJeremy L Thompson CeedInt grid = num_elem / block_sizes[2] + ((num_elem / block_sizes[2] * block_sizes[2] < num_elem) ? 1 : 0); 1069e201c85SYohann CeedInt sharedMem = block_sizes[2] * thread_1d * thread_1d * sizeof(CeedScalar); 107*2b730f8bSJeremy L Thompson CeedCallBackend(CeedRunKernelDimSharedHip(ceed, data->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 1087d8d0e25Snbeams } else if (dim == 3) { 109*2b730f8bSJeremy L Thompson CeedInt grid = num_elem / block_sizes[2] + ((num_elem / block_sizes[2] * block_sizes[2] < num_elem) ? 1 : 0); 1109e201c85SYohann CeedInt sharedMem = block_sizes[2] * thread_1d * thread_1d * sizeof(CeedScalar); 111*2b730f8bSJeremy L Thompson CeedCallBackend(CeedRunKernelDimSharedHip(ceed, data->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 1127d8d0e25Snbeams } 1137d8d0e25Snbeams 1147d8d0e25Snbeams // Restore input arrays 1159e201c85SYohann for (CeedInt i = 0; i < num_input_fields; i++) { 116*2b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 1179e201c85SYohann if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 1187d8d0e25Snbeams } else { 119*2b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 1209e201c85SYohann if (vec == CEED_VECTOR_ACTIVE) vec = input_vec; 121*2b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorRestoreArrayRead(vec, &data->fields.inputs[i])); 1227d8d0e25Snbeams } 1237d8d0e25Snbeams } 1247d8d0e25Snbeams 1257d8d0e25Snbeams // Restore output arrays 1269e201c85SYohann for (CeedInt i = 0; i < num_output_fields; i++) { 127*2b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 1289e201c85SYohann if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 1297d8d0e25Snbeams } else { 130*2b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 1319e201c85SYohann if (vec == CEED_VECTOR_ACTIVE) vec = output_vec; 1327d8d0e25Snbeams // Check for multiple output modes 1337d8d0e25Snbeams CeedInt index = -1; 1347d8d0e25Snbeams for (CeedInt j = 0; j < i; j++) { 1359e201c85SYohann if (vec == output_vecs[j]) { 1367d8d0e25Snbeams index = j; 1377d8d0e25Snbeams break; 1387d8d0e25Snbeams } 1397d8d0e25Snbeams } 1407d8d0e25Snbeams if (index == -1) { 141*2b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorRestoreArray(vec, &data->fields.outputs[i])); 1427d8d0e25Snbeams } 1437d8d0e25Snbeams } 1447d8d0e25Snbeams } 1457d8d0e25Snbeams 1467d8d0e25Snbeams // Restore context data 147*2b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionRestoreInnerContextData(qf, &qf_data->d_c)); 148441428dfSJeremy L Thompson 149e15f9bd0SJeremy L Thompson return CEED_ERROR_SUCCESS; 1507d8d0e25Snbeams } 1517d8d0e25Snbeams 1527d8d0e25Snbeams //------------------------------------------------------------------------------ 1537d8d0e25Snbeams // Create operator 1547d8d0e25Snbeams //------------------------------------------------------------------------------ 1557d8d0e25Snbeams int CeedOperatorCreate_Hip_gen(CeedOperator op) { 1567d8d0e25Snbeams Ceed ceed; 157*2b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1587d8d0e25Snbeams CeedOperator_Hip_gen *impl; 1597d8d0e25Snbeams 160*2b730f8bSJeremy L Thompson CeedCallBackend(CeedCalloc(1, &impl)); 161*2b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorSetData(op, impl)); 1627d8d0e25Snbeams 163*2b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAdd_Hip_gen)); 164*2b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Hip_gen)); 165e15f9bd0SJeremy L Thompson return CEED_ERROR_SUCCESS; 1667d8d0e25Snbeams } 1677d8d0e25Snbeams //------------------------------------------------------------------------------ 168