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 849aac155SJeremy L Thompson #include <ceed.h> 9ec3da8bcSJed Brown #include <ceed/backend.h> 1049aac155SJeremy L Thompson #include <ceed/jit-source/hip/hip-types.h> 113d576824SJeremy L Thompson #include <stddef.h> 122b730f8bSJeremy L Thompson 137d8d0e25Snbeams #include "../hip/ceed-hip-compile.h" 142b730f8bSJeremy L Thompson #include "ceed-hip-gen-operator-build.h" 152b730f8bSJeremy L Thompson #include "ceed-hip-gen.h" 167d8d0e25Snbeams 177d8d0e25Snbeams //------------------------------------------------------------------------------ 187d8d0e25Snbeams // Destroy operator 197d8d0e25Snbeams //------------------------------------------------------------------------------ 207d8d0e25Snbeams static int CeedOperatorDestroy_Hip_gen(CeedOperator op) { 217d8d0e25Snbeams CeedOperator_Hip_gen *impl; 222b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetData(op, &impl)); 232b730f8bSJeremy L Thompson CeedCallBackend(CeedFree(&impl)); 24e15f9bd0SJeremy L Thompson return CEED_ERROR_SUCCESS; 257d8d0e25Snbeams } 267d8d0e25Snbeams 277d8d0e25Snbeams //------------------------------------------------------------------------------ 287d8d0e25Snbeams // Apply and add to output 297d8d0e25Snbeams //------------------------------------------------------------------------------ 302b730f8bSJeremy L Thompson static int CeedOperatorApplyAdd_Hip_gen(CeedOperator op, CeedVector input_vec, CeedVector output_vec, CeedRequest *request) { 317d8d0e25Snbeams Ceed ceed; 322b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 337d8d0e25Snbeams CeedOperator_Hip_gen *data; 342b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetData(op, &data)); 357d8d0e25Snbeams CeedQFunction qf; 367d8d0e25Snbeams CeedQFunction_Hip_gen *qf_data; 372b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 382b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetData(qf, &qf_data)); 399e201c85SYohann CeedInt num_elem, num_input_fields, num_output_fields; 402b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 419e201c85SYohann CeedOperatorField *op_input_fields, *op_output_fields; 422b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 439e201c85SYohann CeedQFunctionField *qf_input_fields, *qf_output_fields; 442b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 459e201c85SYohann CeedEvalMode eval_mode; 462c2ea1dbSJeremy L Thompson CeedVector vec, output_vecs[CEED_FIELD_MAX] = {NULL}; 477d8d0e25Snbeams 487d8d0e25Snbeams // Creation of the operator 492b730f8bSJeremy L Thompson CeedCallBackend(CeedHipGenOperatorBuild(op)); 507d8d0e25Snbeams 517d8d0e25Snbeams // Input vectors 529e201c85SYohann for (CeedInt i = 0; i < num_input_fields; i++) { 532b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 549e201c85SYohann if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 559e201c85SYohann data->fields.inputs[i] = NULL; 567d8d0e25Snbeams } else { 577d8d0e25Snbeams // Get input vector 582b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 599e201c85SYohann if (vec == CEED_VECTOR_ACTIVE) vec = input_vec; 602b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorGetArrayRead(vec, CEED_MEM_DEVICE, &data->fields.inputs[i])); 617d8d0e25Snbeams } 627d8d0e25Snbeams } 637d8d0e25Snbeams 647d8d0e25Snbeams // Output vectors 659e201c85SYohann for (CeedInt i = 0; i < num_output_fields; i++) { 662b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 679e201c85SYohann if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 689e201c85SYohann data->fields.outputs[i] = NULL; 697d8d0e25Snbeams } else { 707d8d0e25Snbeams // Get output vector 712b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 729e201c85SYohann if (vec == CEED_VECTOR_ACTIVE) vec = output_vec; 739e201c85SYohann output_vecs[i] = vec; 747d8d0e25Snbeams // Check for multiple output modes 757d8d0e25Snbeams CeedInt index = -1; 767d8d0e25Snbeams for (CeedInt j = 0; j < i; j++) { 779e201c85SYohann if (vec == output_vecs[j]) { 787d8d0e25Snbeams index = j; 797d8d0e25Snbeams break; 807d8d0e25Snbeams } 817d8d0e25Snbeams } 827d8d0e25Snbeams if (index == -1) { 832b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorGetArray(vec, CEED_MEM_DEVICE, &data->fields.outputs[i])); 847d8d0e25Snbeams } else { 859e201c85SYohann data->fields.outputs[i] = data->fields.outputs[index]; 867d8d0e25Snbeams } 877d8d0e25Snbeams } 887d8d0e25Snbeams } 897d8d0e25Snbeams 907d8d0e25Snbeams // Get context data 912b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetInnerContextData(qf, CEED_MEM_DEVICE, &qf_data->d_c)); 927d8d0e25Snbeams 937d8d0e25Snbeams // Apply operator 942b730f8bSJeremy L Thompson void *opargs[] = {(void *)&num_elem, &qf_data->d_c, &data->indices, &data->fields, &data->B, &data->G, &data->W}; 957d8d0e25Snbeams const CeedInt dim = data->dim; 969e201c85SYohann const CeedInt Q_1d = data->Q_1d; 979e201c85SYohann const CeedInt P_1d = data->max_P_1d; 989e201c85SYohann const CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 99b3e1519bSnbeams CeedInt block_sizes[3]; 1002b730f8bSJeremy L Thompson CeedCallBackend(BlockGridCalculate_Hip_gen(dim, num_elem, P_1d, Q_1d, block_sizes)); 1017d8d0e25Snbeams if (dim == 1) { 1022b730f8bSJeremy L Thompson CeedInt grid = num_elem / block_sizes[2] + ((num_elem / block_sizes[2] * block_sizes[2] < num_elem) ? 1 : 0); 1039e201c85SYohann CeedInt sharedMem = block_sizes[2] * thread_1d * sizeof(CeedScalar); 1042b730f8bSJeremy L Thompson CeedCallBackend(CeedRunKernelDimSharedHip(ceed, data->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 1057d8d0e25Snbeams } else if (dim == 2) { 1062b730f8bSJeremy L Thompson CeedInt grid = num_elem / block_sizes[2] + ((num_elem / block_sizes[2] * block_sizes[2] < num_elem) ? 1 : 0); 1079e201c85SYohann CeedInt sharedMem = block_sizes[2] * thread_1d * thread_1d * sizeof(CeedScalar); 1082b730f8bSJeremy L Thompson CeedCallBackend(CeedRunKernelDimSharedHip(ceed, data->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 1097d8d0e25Snbeams } else if (dim == 3) { 1102b730f8bSJeremy L Thompson CeedInt grid = num_elem / block_sizes[2] + ((num_elem / block_sizes[2] * block_sizes[2] < num_elem) ? 1 : 0); 1119e201c85SYohann CeedInt sharedMem = block_sizes[2] * thread_1d * thread_1d * sizeof(CeedScalar); 1122b730f8bSJeremy L Thompson CeedCallBackend(CeedRunKernelDimSharedHip(ceed, data->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 1137d8d0e25Snbeams } 1147d8d0e25Snbeams 1157d8d0e25Snbeams // Restore input arrays 1169e201c85SYohann for (CeedInt i = 0; i < num_input_fields; i++) { 1172b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 1189e201c85SYohann if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 1197d8d0e25Snbeams } else { 1202b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 1219e201c85SYohann if (vec == CEED_VECTOR_ACTIVE) vec = input_vec; 1222b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorRestoreArrayRead(vec, &data->fields.inputs[i])); 1237d8d0e25Snbeams } 1247d8d0e25Snbeams } 1257d8d0e25Snbeams 1267d8d0e25Snbeams // Restore output arrays 1279e201c85SYohann for (CeedInt i = 0; i < num_output_fields; i++) { 1282b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 1299e201c85SYohann if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 1307d8d0e25Snbeams } else { 1312b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 1329e201c85SYohann if (vec == CEED_VECTOR_ACTIVE) vec = output_vec; 1337d8d0e25Snbeams // Check for multiple output modes 1347d8d0e25Snbeams CeedInt index = -1; 1357d8d0e25Snbeams for (CeedInt j = 0; j < i; j++) { 1369e201c85SYohann if (vec == output_vecs[j]) { 1377d8d0e25Snbeams index = j; 1387d8d0e25Snbeams break; 1397d8d0e25Snbeams } 1407d8d0e25Snbeams } 1417d8d0e25Snbeams if (index == -1) { 1422b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorRestoreArray(vec, &data->fields.outputs[i])); 1437d8d0e25Snbeams } 1447d8d0e25Snbeams } 1457d8d0e25Snbeams } 1467d8d0e25Snbeams 1477d8d0e25Snbeams // Restore context data 1482b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionRestoreInnerContextData(qf, &qf_data->d_c)); 149441428dfSJeremy L Thompson 150e15f9bd0SJeremy L Thompson return CEED_ERROR_SUCCESS; 1517d8d0e25Snbeams } 1527d8d0e25Snbeams 1537d8d0e25Snbeams //------------------------------------------------------------------------------ 1547d8d0e25Snbeams // Create operator 1557d8d0e25Snbeams //------------------------------------------------------------------------------ 1567d8d0e25Snbeams int CeedOperatorCreate_Hip_gen(CeedOperator op) { 1577d8d0e25Snbeams Ceed ceed; 1582b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1597d8d0e25Snbeams CeedOperator_Hip_gen *impl; 1607d8d0e25Snbeams 1612b730f8bSJeremy L Thompson CeedCallBackend(CeedCalloc(1, &impl)); 1622b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorSetData(op, impl)); 1637d8d0e25Snbeams 1642b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAdd_Hip_gen)); 1652b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Hip_gen)); 166e15f9bd0SJeremy L Thompson return CEED_ERROR_SUCCESS; 1677d8d0e25Snbeams } 168*2a86cc9dSSebastian Grimberg 1697d8d0e25Snbeams //------------------------------------------------------------------------------ 170