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 13*b2165e7aSSebastian Grimberg #include "../hip/ceed-hip-common.h" 147d8d0e25Snbeams #include "../hip/ceed-hip-compile.h" 152b730f8bSJeremy L Thompson #include "ceed-hip-gen-operator-build.h" 162b730f8bSJeremy L Thompson #include "ceed-hip-gen.h" 177d8d0e25Snbeams 187d8d0e25Snbeams //------------------------------------------------------------------------------ 197d8d0e25Snbeams // Destroy operator 207d8d0e25Snbeams //------------------------------------------------------------------------------ 217d8d0e25Snbeams static int CeedOperatorDestroy_Hip_gen(CeedOperator op) { 227d8d0e25Snbeams CeedOperator_Hip_gen *impl; 232b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetData(op, &impl)); 242b730f8bSJeremy L Thompson CeedCallBackend(CeedFree(&impl)); 25e15f9bd0SJeremy L Thompson return CEED_ERROR_SUCCESS; 267d8d0e25Snbeams } 277d8d0e25Snbeams 287d8d0e25Snbeams //------------------------------------------------------------------------------ 297d8d0e25Snbeams // Apply and add to output 307d8d0e25Snbeams //------------------------------------------------------------------------------ 312b730f8bSJeremy L Thompson static int CeedOperatorApplyAdd_Hip_gen(CeedOperator op, CeedVector input_vec, CeedVector output_vec, CeedRequest *request) { 327d8d0e25Snbeams Ceed ceed; 332b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 347d8d0e25Snbeams CeedOperator_Hip_gen *data; 352b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetData(op, &data)); 367d8d0e25Snbeams CeedQFunction qf; 377d8d0e25Snbeams CeedQFunction_Hip_gen *qf_data; 382b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 392b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetData(qf, &qf_data)); 409e201c85SYohann CeedInt num_elem, num_input_fields, num_output_fields; 412b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 429e201c85SYohann CeedOperatorField *op_input_fields, *op_output_fields; 432b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 449e201c85SYohann CeedQFunctionField *qf_input_fields, *qf_output_fields; 452b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 469e201c85SYohann CeedEvalMode eval_mode; 472c2ea1dbSJeremy L Thompson CeedVector vec, output_vecs[CEED_FIELD_MAX] = {NULL}; 487d8d0e25Snbeams 497d8d0e25Snbeams // Creation of the operator 50eb7e6cafSJeremy L Thompson CeedCallBackend(CeedOperatorBuildKernel_Hip_gen(op)); 517d8d0e25Snbeams 527d8d0e25Snbeams // Input vectors 539e201c85SYohann for (CeedInt i = 0; i < num_input_fields; i++) { 542b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 559e201c85SYohann if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 569e201c85SYohann data->fields.inputs[i] = NULL; 577d8d0e25Snbeams } else { 587d8d0e25Snbeams // Get input vector 592b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 609e201c85SYohann if (vec == CEED_VECTOR_ACTIVE) vec = input_vec; 612b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorGetArrayRead(vec, CEED_MEM_DEVICE, &data->fields.inputs[i])); 627d8d0e25Snbeams } 637d8d0e25Snbeams } 647d8d0e25Snbeams 657d8d0e25Snbeams // Output vectors 669e201c85SYohann for (CeedInt i = 0; i < num_output_fields; i++) { 672b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 689e201c85SYohann if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 699e201c85SYohann data->fields.outputs[i] = NULL; 707d8d0e25Snbeams } else { 717d8d0e25Snbeams // Get output vector 722b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 739e201c85SYohann if (vec == CEED_VECTOR_ACTIVE) vec = output_vec; 749e201c85SYohann output_vecs[i] = vec; 757d8d0e25Snbeams // Check for multiple output modes 767d8d0e25Snbeams CeedInt index = -1; 777d8d0e25Snbeams for (CeedInt j = 0; j < i; j++) { 789e201c85SYohann if (vec == output_vecs[j]) { 797d8d0e25Snbeams index = j; 807d8d0e25Snbeams break; 817d8d0e25Snbeams } 827d8d0e25Snbeams } 837d8d0e25Snbeams if (index == -1) { 842b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorGetArray(vec, CEED_MEM_DEVICE, &data->fields.outputs[i])); 857d8d0e25Snbeams } else { 869e201c85SYohann data->fields.outputs[i] = data->fields.outputs[index]; 877d8d0e25Snbeams } 887d8d0e25Snbeams } 897d8d0e25Snbeams } 907d8d0e25Snbeams 917d8d0e25Snbeams // Get context data 922b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetInnerContextData(qf, CEED_MEM_DEVICE, &qf_data->d_c)); 937d8d0e25Snbeams 947d8d0e25Snbeams // Apply operator 952b730f8bSJeremy L Thompson void *opargs[] = {(void *)&num_elem, &qf_data->d_c, &data->indices, &data->fields, &data->B, &data->G, &data->W}; 967d8d0e25Snbeams const CeedInt dim = data->dim; 979e201c85SYohann const CeedInt Q_1d = data->Q_1d; 989e201c85SYohann const CeedInt P_1d = data->max_P_1d; 999e201c85SYohann const CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 100b3e1519bSnbeams CeedInt block_sizes[3]; 1012b730f8bSJeremy L Thompson CeedCallBackend(BlockGridCalculate_Hip_gen(dim, num_elem, P_1d, Q_1d, block_sizes)); 1027d8d0e25Snbeams if (dim == 1) { 1032b730f8bSJeremy L Thompson CeedInt grid = num_elem / block_sizes[2] + ((num_elem / block_sizes[2] * block_sizes[2] < num_elem) ? 1 : 0); 1049e201c85SYohann CeedInt sharedMem = block_sizes[2] * thread_1d * sizeof(CeedScalar); 105eb7e6cafSJeremy L Thompson CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 1067d8d0e25Snbeams } else if (dim == 2) { 1072b730f8bSJeremy L Thompson CeedInt grid = num_elem / block_sizes[2] + ((num_elem / block_sizes[2] * block_sizes[2] < num_elem) ? 1 : 0); 1089e201c85SYohann CeedInt sharedMem = block_sizes[2] * thread_1d * thread_1d * sizeof(CeedScalar); 109eb7e6cafSJeremy L Thompson CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 1107d8d0e25Snbeams } else if (dim == 3) { 1112b730f8bSJeremy L Thompson CeedInt grid = num_elem / block_sizes[2] + ((num_elem / block_sizes[2] * block_sizes[2] < num_elem) ? 1 : 0); 1129e201c85SYohann CeedInt sharedMem = block_sizes[2] * thread_1d * thread_1d * sizeof(CeedScalar); 113eb7e6cafSJeremy L Thompson CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 1147d8d0e25Snbeams } 1157d8d0e25Snbeams 1167d8d0e25Snbeams // Restore input arrays 1179e201c85SYohann for (CeedInt i = 0; i < num_input_fields; i++) { 1182b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 1199e201c85SYohann if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 1207d8d0e25Snbeams } else { 1212b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 1229e201c85SYohann if (vec == CEED_VECTOR_ACTIVE) vec = input_vec; 1232b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorRestoreArrayRead(vec, &data->fields.inputs[i])); 1247d8d0e25Snbeams } 1257d8d0e25Snbeams } 1267d8d0e25Snbeams 1277d8d0e25Snbeams // Restore output arrays 1289e201c85SYohann for (CeedInt i = 0; i < num_output_fields; i++) { 1292b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 1309e201c85SYohann if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 1317d8d0e25Snbeams } else { 1322b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 1339e201c85SYohann if (vec == CEED_VECTOR_ACTIVE) vec = output_vec; 1347d8d0e25Snbeams // Check for multiple output modes 1357d8d0e25Snbeams CeedInt index = -1; 1367d8d0e25Snbeams for (CeedInt j = 0; j < i; j++) { 1379e201c85SYohann if (vec == output_vecs[j]) { 1387d8d0e25Snbeams index = j; 1397d8d0e25Snbeams break; 1407d8d0e25Snbeams } 1417d8d0e25Snbeams } 1427d8d0e25Snbeams if (index == -1) { 1432b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorRestoreArray(vec, &data->fields.outputs[i])); 1447d8d0e25Snbeams } 1457d8d0e25Snbeams } 1467d8d0e25Snbeams } 1477d8d0e25Snbeams 1487d8d0e25Snbeams // Restore context data 1492b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionRestoreInnerContextData(qf, &qf_data->d_c)); 150441428dfSJeremy L Thompson 151e15f9bd0SJeremy L Thompson return CEED_ERROR_SUCCESS; 1527d8d0e25Snbeams } 1537d8d0e25Snbeams 1547d8d0e25Snbeams //------------------------------------------------------------------------------ 1557d8d0e25Snbeams // Create operator 1567d8d0e25Snbeams //------------------------------------------------------------------------------ 1577d8d0e25Snbeams int CeedOperatorCreate_Hip_gen(CeedOperator op) { 1587d8d0e25Snbeams Ceed ceed; 1592b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1607d8d0e25Snbeams CeedOperator_Hip_gen *impl; 1617d8d0e25Snbeams 1622b730f8bSJeremy L Thompson CeedCallBackend(CeedCalloc(1, &impl)); 1632b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorSetData(op, impl)); 1647d8d0e25Snbeams 1652b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAdd_Hip_gen)); 1662b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Hip_gen)); 167e15f9bd0SJeremy L Thompson return CEED_ERROR_SUCCESS; 1687d8d0e25Snbeams } 1692a86cc9dSSebastian Grimberg 1707d8d0e25Snbeams //------------------------------------------------------------------------------ 171