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 13b2165e7aSSebastian 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; 23*b7453713SJeremy L Thompson 242b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetData(op, &impl)); 252b730f8bSJeremy L Thompson CeedCallBackend(CeedFree(&impl)); 26e15f9bd0SJeremy L Thompson return CEED_ERROR_SUCCESS; 277d8d0e25Snbeams } 287d8d0e25Snbeams 297d8d0e25Snbeams //------------------------------------------------------------------------------ 307d8d0e25Snbeams // Apply and add to output 317d8d0e25Snbeams //------------------------------------------------------------------------------ 322b730f8bSJeremy L Thompson static int CeedOperatorApplyAdd_Hip_gen(CeedOperator op, CeedVector input_vec, CeedVector output_vec, CeedRequest *request) { 337d8d0e25Snbeams Ceed ceed; 34*b7453713SJeremy L Thompson CeedInt num_elem, num_input_fields, num_output_fields; 35*b7453713SJeremy L Thompson CeedEvalMode eval_mode; 36*b7453713SJeremy L Thompson CeedVector output_vecs[CEED_FIELD_MAX] = {NULL}; 37*b7453713SJeremy L Thompson CeedQFunctionField *qf_input_fields, *qf_output_fields; 387d8d0e25Snbeams CeedQFunction_Hip_gen *qf_data; 39*b7453713SJeremy L Thompson CeedQFunction qf; 40*b7453713SJeremy L Thompson CeedOperatorField *op_input_fields, *op_output_fields; 41*b7453713SJeremy L Thompson CeedOperator_Hip_gen *data; 42*b7453713SJeremy L Thompson 43*b7453713SJeremy L Thompson CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 44*b7453713SJeremy L Thompson CeedCallBackend(CeedOperatorGetData(op, &data)); 452b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 462b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetData(qf, &qf_data)); 472b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 482b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 492b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 507d8d0e25Snbeams 517d8d0e25Snbeams // Creation of the operator 52eb7e6cafSJeremy L Thompson CeedCallBackend(CeedOperatorBuildKernel_Hip_gen(op)); 537d8d0e25Snbeams 547d8d0e25Snbeams // Input vectors 559e201c85SYohann for (CeedInt i = 0; i < num_input_fields; i++) { 562b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 579e201c85SYohann if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 589e201c85SYohann data->fields.inputs[i] = NULL; 597d8d0e25Snbeams } else { 60*b7453713SJeremy L Thompson CeedVector vec; 61*b7453713SJeremy L Thompson 627d8d0e25Snbeams // Get input vector 632b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 649e201c85SYohann if (vec == CEED_VECTOR_ACTIVE) vec = input_vec; 652b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorGetArrayRead(vec, CEED_MEM_DEVICE, &data->fields.inputs[i])); 667d8d0e25Snbeams } 677d8d0e25Snbeams } 687d8d0e25Snbeams 697d8d0e25Snbeams // Output vectors 709e201c85SYohann for (CeedInt i = 0; i < num_output_fields; i++) { 712b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 729e201c85SYohann if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 739e201c85SYohann data->fields.outputs[i] = NULL; 747d8d0e25Snbeams } else { 75*b7453713SJeremy L Thompson CeedVector vec; 76*b7453713SJeremy L Thompson 777d8d0e25Snbeams // Get output vector 782b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 799e201c85SYohann if (vec == CEED_VECTOR_ACTIVE) vec = output_vec; 809e201c85SYohann output_vecs[i] = vec; 817d8d0e25Snbeams // Check for multiple output modes 827d8d0e25Snbeams CeedInt index = -1; 837d8d0e25Snbeams for (CeedInt j = 0; j < i; j++) { 849e201c85SYohann if (vec == output_vecs[j]) { 857d8d0e25Snbeams index = j; 867d8d0e25Snbeams break; 877d8d0e25Snbeams } 887d8d0e25Snbeams } 897d8d0e25Snbeams if (index == -1) { 902b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorGetArray(vec, CEED_MEM_DEVICE, &data->fields.outputs[i])); 917d8d0e25Snbeams } else { 929e201c85SYohann data->fields.outputs[i] = data->fields.outputs[index]; 937d8d0e25Snbeams } 947d8d0e25Snbeams } 957d8d0e25Snbeams } 967d8d0e25Snbeams 977d8d0e25Snbeams // Get context data 982b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionGetInnerContextData(qf, CEED_MEM_DEVICE, &qf_data->d_c)); 997d8d0e25Snbeams 1007d8d0e25Snbeams // Apply operator 1012b730f8bSJeremy L Thompson void *opargs[] = {(void *)&num_elem, &qf_data->d_c, &data->indices, &data->fields, &data->B, &data->G, &data->W}; 1027d8d0e25Snbeams const CeedInt dim = data->dim; 1039e201c85SYohann const CeedInt Q_1d = data->Q_1d; 1049e201c85SYohann const CeedInt P_1d = data->max_P_1d; 1059e201c85SYohann const CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 106b3e1519bSnbeams CeedInt block_sizes[3]; 107*b7453713SJeremy L Thompson 1082b730f8bSJeremy L Thompson CeedCallBackend(BlockGridCalculate_Hip_gen(dim, num_elem, P_1d, Q_1d, block_sizes)); 1097d8d0e25Snbeams if (dim == 1) { 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 * sizeof(CeedScalar); 112*b7453713SJeremy L Thompson 113eb7e6cafSJeremy L Thompson CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 1147d8d0e25Snbeams } else if (dim == 2) { 1152b730f8bSJeremy L Thompson CeedInt grid = num_elem / block_sizes[2] + ((num_elem / block_sizes[2] * block_sizes[2] < num_elem) ? 1 : 0); 1169e201c85SYohann CeedInt sharedMem = block_sizes[2] * thread_1d * thread_1d * sizeof(CeedScalar); 117*b7453713SJeremy L Thompson 118eb7e6cafSJeremy L Thompson CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 1197d8d0e25Snbeams } else if (dim == 3) { 1202b730f8bSJeremy L Thompson CeedInt grid = num_elem / block_sizes[2] + ((num_elem / block_sizes[2] * block_sizes[2] < num_elem) ? 1 : 0); 1219e201c85SYohann CeedInt sharedMem = block_sizes[2] * thread_1d * thread_1d * sizeof(CeedScalar); 122*b7453713SJeremy L Thompson 123eb7e6cafSJeremy L Thompson CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 1247d8d0e25Snbeams } 1257d8d0e25Snbeams 1267d8d0e25Snbeams // Restore input arrays 1279e201c85SYohann for (CeedInt i = 0; i < num_input_fields; i++) { 1282b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 1299e201c85SYohann if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 1307d8d0e25Snbeams } else { 131*b7453713SJeremy L Thompson CeedVector vec; 132*b7453713SJeremy L Thompson 1332b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 1349e201c85SYohann if (vec == CEED_VECTOR_ACTIVE) vec = input_vec; 1352b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorRestoreArrayRead(vec, &data->fields.inputs[i])); 1367d8d0e25Snbeams } 1377d8d0e25Snbeams } 1387d8d0e25Snbeams 1397d8d0e25Snbeams // Restore output arrays 1409e201c85SYohann for (CeedInt i = 0; i < num_output_fields; i++) { 1412b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 1429e201c85SYohann if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 1437d8d0e25Snbeams } else { 144*b7453713SJeremy L Thompson CeedVector vec; 145*b7453713SJeremy L Thompson 1462b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 1479e201c85SYohann if (vec == CEED_VECTOR_ACTIVE) vec = output_vec; 1487d8d0e25Snbeams // Check for multiple output modes 1497d8d0e25Snbeams CeedInt index = -1; 1507d8d0e25Snbeams for (CeedInt j = 0; j < i; j++) { 1519e201c85SYohann if (vec == output_vecs[j]) { 1527d8d0e25Snbeams index = j; 1537d8d0e25Snbeams break; 1547d8d0e25Snbeams } 1557d8d0e25Snbeams } 1567d8d0e25Snbeams if (index == -1) { 1572b730f8bSJeremy L Thompson CeedCallBackend(CeedVectorRestoreArray(vec, &data->fields.outputs[i])); 1587d8d0e25Snbeams } 1597d8d0e25Snbeams } 1607d8d0e25Snbeams } 1617d8d0e25Snbeams 1627d8d0e25Snbeams // Restore context data 1632b730f8bSJeremy L Thompson CeedCallBackend(CeedQFunctionRestoreInnerContextData(qf, &qf_data->d_c)); 164e15f9bd0SJeremy L Thompson return CEED_ERROR_SUCCESS; 1657d8d0e25Snbeams } 1667d8d0e25Snbeams 1677d8d0e25Snbeams //------------------------------------------------------------------------------ 1687d8d0e25Snbeams // Create operator 1697d8d0e25Snbeams //------------------------------------------------------------------------------ 1707d8d0e25Snbeams int CeedOperatorCreate_Hip_gen(CeedOperator op) { 1717d8d0e25Snbeams Ceed ceed; 1727d8d0e25Snbeams CeedOperator_Hip_gen *impl; 1737d8d0e25Snbeams 174*b7453713SJeremy L Thompson CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1752b730f8bSJeremy L Thompson CeedCallBackend(CeedCalloc(1, &impl)); 1762b730f8bSJeremy L Thompson CeedCallBackend(CeedOperatorSetData(op, impl)); 1772b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAdd_Hip_gen)); 1782b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Hip_gen)); 179e15f9bd0SJeremy L Thompson return CEED_ERROR_SUCCESS; 1807d8d0e25Snbeams } 1812a86cc9dSSebastian Grimberg 1827d8d0e25Snbeams //------------------------------------------------------------------------------ 183