16ca0f394SUmesh Unnikrishnan // Copyright (c) 2017-2022, Lawrence Livermore National Security, LLC and other CEED contributors. 26ca0f394SUmesh Unnikrishnan // All Rights Reserved. See the top-level LICENSE and NOTICE files for details. 36ca0f394SUmesh Unnikrishnan // 46ca0f394SUmesh Unnikrishnan // SPDX-License-Identifier: BSD-2-Clause 56ca0f394SUmesh Unnikrishnan // 66ca0f394SUmesh Unnikrishnan // This file is part of CEED: http://github.com/ceed 76ca0f394SUmesh Unnikrishnan 86ca0f394SUmesh Unnikrishnan #include <ceed/backend.h> 96ca0f394SUmesh Unnikrishnan #include <ceed/ceed.h> 106ca0f394SUmesh Unnikrishnan #include <stddef.h> 116ca0f394SUmesh Unnikrishnan 126ca0f394SUmesh Unnikrishnan #include "../sycl/ceed-sycl-compile.hpp" 136ca0f394SUmesh Unnikrishnan #include "ceed-sycl-gen-operator-build.hpp" 146ca0f394SUmesh Unnikrishnan #include "ceed-sycl-gen.hpp" 156ca0f394SUmesh Unnikrishnan 166ca0f394SUmesh Unnikrishnan //------------------------------------------------------------------------------ 176ca0f394SUmesh Unnikrishnan // Destroy operator 186ca0f394SUmesh Unnikrishnan //------------------------------------------------------------------------------ 196ca0f394SUmesh Unnikrishnan static int CeedOperatorDestroy_Sycl_gen(CeedOperator op) { 206ca0f394SUmesh Unnikrishnan CeedOperator_Sycl_gen *impl; 21*dd64fc84SJeremy L Thompson 226ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedOperatorGetData(op, &impl)); 236ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedFree(&impl)); 246ca0f394SUmesh Unnikrishnan return CEED_ERROR_SUCCESS; 256ca0f394SUmesh Unnikrishnan } 266ca0f394SUmesh Unnikrishnan 276ca0f394SUmesh Unnikrishnan //------------------------------------------------------------------------------ 286ca0f394SUmesh Unnikrishnan // Apply and add to output 296ca0f394SUmesh Unnikrishnan //------------------------------------------------------------------------------ 306ca0f394SUmesh Unnikrishnan static int CeedOperatorApplyAdd_Sycl_gen(CeedOperator op, CeedVector input_vec, CeedVector output_vec, CeedRequest *request) { 316ca0f394SUmesh Unnikrishnan Ceed ceed; 326ca0f394SUmesh Unnikrishnan Ceed_Sycl *ceed_Sycl; 33*dd64fc84SJeremy L Thompson CeedInt num_elem, num_input_fields, num_output_fields; 34*dd64fc84SJeremy L Thompson CeedEvalMode eval_mode; 35*dd64fc84SJeremy L Thompson CeedVector output_vecs[CEED_FIELD_MAX] = {}; 36*dd64fc84SJeremy L Thompson CeedQFunctionField *qf_input_fields, *qf_output_fields; 376ca0f394SUmesh Unnikrishnan CeedQFunction_Sycl_gen *qf_impl; 38*dd64fc84SJeremy L Thompson CeedQFunction qf; 39*dd64fc84SJeremy L Thompson CeedOperatorField *op_input_fields, *op_output_fields; 40*dd64fc84SJeremy L Thompson CeedOperator_Sycl_gen *impl; 41*dd64fc84SJeremy L Thompson 42*dd64fc84SJeremy L Thompson CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 43*dd64fc84SJeremy L Thompson CeedCallBackend(CeedGetData(ceed, &ceed_Sycl)); 44*dd64fc84SJeremy L Thompson CeedCallBackend(CeedOperatorGetData(op, &impl)); 456ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 466ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedQFunctionGetData(qf, &qf_impl)); 476ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 486ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 496ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 506ca0f394SUmesh Unnikrishnan 516ca0f394SUmesh Unnikrishnan // Creation of the operator 526ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedOperatorBuildKernel_Sycl_gen(op)); 536ca0f394SUmesh Unnikrishnan 546ca0f394SUmesh Unnikrishnan // Input vectors 556ca0f394SUmesh Unnikrishnan for (CeedInt i = 0; i < num_input_fields; i++) { 566ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 576ca0f394SUmesh Unnikrishnan if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 586ca0f394SUmesh Unnikrishnan impl->fields->inputs[i] = NULL; 596ca0f394SUmesh Unnikrishnan } else { 60*dd64fc84SJeremy L Thompson CeedVector vec; 61*dd64fc84SJeremy L Thompson 626ca0f394SUmesh Unnikrishnan // Get input vector 636ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 646ca0f394SUmesh Unnikrishnan if (vec == CEED_VECTOR_ACTIVE) vec = input_vec; 656ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedVectorGetArrayRead(vec, CEED_MEM_DEVICE, &impl->fields->inputs[i])); 666ca0f394SUmesh Unnikrishnan } 676ca0f394SUmesh Unnikrishnan } 686ca0f394SUmesh Unnikrishnan 696ca0f394SUmesh Unnikrishnan // Output vectors 706ca0f394SUmesh Unnikrishnan for (CeedInt i = 0; i < num_output_fields; i++) { 716ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 726ca0f394SUmesh Unnikrishnan if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 736ca0f394SUmesh Unnikrishnan impl->fields->outputs[i] = NULL; 746ca0f394SUmesh Unnikrishnan } else { 75*dd64fc84SJeremy L Thompson CeedVector vec; 76*dd64fc84SJeremy L Thompson 776ca0f394SUmesh Unnikrishnan // Get output vector 786ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 796ca0f394SUmesh Unnikrishnan if (vec == CEED_VECTOR_ACTIVE) vec = output_vec; 806ca0f394SUmesh Unnikrishnan output_vecs[i] = vec; 816ca0f394SUmesh Unnikrishnan // Check for multiple output modes 826ca0f394SUmesh Unnikrishnan CeedInt index = -1; 836ca0f394SUmesh Unnikrishnan for (CeedInt j = 0; j < i; j++) { 846ca0f394SUmesh Unnikrishnan if (vec == output_vecs[j]) { 856ca0f394SUmesh Unnikrishnan index = j; 866ca0f394SUmesh Unnikrishnan break; 876ca0f394SUmesh Unnikrishnan } 886ca0f394SUmesh Unnikrishnan } 896ca0f394SUmesh Unnikrishnan if (index == -1) { 906ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedVectorGetArray(vec, CEED_MEM_DEVICE, &impl->fields->outputs[i])); 916ca0f394SUmesh Unnikrishnan } else { 926ca0f394SUmesh Unnikrishnan impl->fields->outputs[i] = impl->fields->outputs[index]; 936ca0f394SUmesh Unnikrishnan } 946ca0f394SUmesh Unnikrishnan } 956ca0f394SUmesh Unnikrishnan } 966ca0f394SUmesh Unnikrishnan 976ca0f394SUmesh Unnikrishnan // Get context data 986ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedQFunctionGetInnerContextData(qf, CEED_MEM_DEVICE, &qf_impl->d_c)); 996ca0f394SUmesh Unnikrishnan 1006ca0f394SUmesh Unnikrishnan // Apply operator 1016ca0f394SUmesh Unnikrishnan const CeedInt dim = impl->dim; 1026ca0f394SUmesh Unnikrishnan const CeedInt Q_1d = impl->Q_1d; 1036ca0f394SUmesh Unnikrishnan const CeedInt P_1d = impl->max_P_1d; 1046ca0f394SUmesh Unnikrishnan CeedInt block_sizes[3], grid = 0; 105*dd64fc84SJeremy L Thompson 1066ca0f394SUmesh Unnikrishnan CeedCallBackend(BlockGridCalculate_Sycl_gen(dim, P_1d, Q_1d, block_sizes)); 1076ca0f394SUmesh Unnikrishnan if (dim == 1) { 1086ca0f394SUmesh Unnikrishnan grid = num_elem / block_sizes[2] + ((num_elem / block_sizes[2] * block_sizes[2] < num_elem) ? 1 : 0); 1096ca0f394SUmesh Unnikrishnan // CeedCallBackend(CeedRunKernelDimSharedSycl(ceed, impl->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 1106ca0f394SUmesh Unnikrishnan } else if (dim == 2) { 1116ca0f394SUmesh Unnikrishnan grid = num_elem / block_sizes[2] + ((num_elem / block_sizes[2] * block_sizes[2] < num_elem) ? 1 : 0); 1126ca0f394SUmesh Unnikrishnan // CeedCallBackend(CeedRunKernelDimSharedSycl(ceed, impl->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 1136ca0f394SUmesh Unnikrishnan } else if (dim == 3) { 1146ca0f394SUmesh Unnikrishnan grid = num_elem / block_sizes[2] + ((num_elem / block_sizes[2] * block_sizes[2] < num_elem) ? 1 : 0); 1156ca0f394SUmesh Unnikrishnan // CeedCallBackend(CeedRunKernelDimSharedSycl(ceed, impl->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 1166ca0f394SUmesh Unnikrishnan } 1176ca0f394SUmesh Unnikrishnan 1186ca0f394SUmesh Unnikrishnan sycl::range<3> local_range(block_sizes[2], block_sizes[1], block_sizes[0]); 1196ca0f394SUmesh Unnikrishnan sycl::range<3> global_range(grid * block_sizes[2], block_sizes[1], block_sizes[0]); 1206ca0f394SUmesh Unnikrishnan sycl::nd_range<3> kernel_range(global_range, local_range); 1216ca0f394SUmesh Unnikrishnan 1226ca0f394SUmesh Unnikrishnan //----------- 1236ca0f394SUmesh Unnikrishnan // Order queue 1246ca0f394SUmesh Unnikrishnan sycl::event e = ceed_Sycl->sycl_queue.ext_oneapi_submit_barrier(); 1256ca0f394SUmesh Unnikrishnan 1266ca0f394SUmesh Unnikrishnan CeedCallSycl(ceed, ceed_Sycl->sycl_queue.submit([&](sycl::handler &cgh) { 1276ca0f394SUmesh Unnikrishnan cgh.depends_on(e); 1286ca0f394SUmesh Unnikrishnan cgh.set_args(num_elem, qf_impl->d_c, impl->indices, impl->fields, impl->B, impl->G, impl->W); 1296ca0f394SUmesh Unnikrishnan cgh.parallel_for(kernel_range, *(impl->op)); 1306ca0f394SUmesh Unnikrishnan })); 1316ca0f394SUmesh Unnikrishnan CeedCallSycl(ceed, ceed_Sycl->sycl_queue.wait_and_throw()); 1326ca0f394SUmesh Unnikrishnan 1336ca0f394SUmesh Unnikrishnan // Restore input arrays 1346ca0f394SUmesh Unnikrishnan for (CeedInt i = 0; i < num_input_fields; i++) { 1356ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 1366ca0f394SUmesh Unnikrishnan if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 1376ca0f394SUmesh Unnikrishnan } else { 138*dd64fc84SJeremy L Thompson CeedVector vec; 139*dd64fc84SJeremy L Thompson 1406ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 1416ca0f394SUmesh Unnikrishnan if (vec == CEED_VECTOR_ACTIVE) vec = input_vec; 1426ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedVectorRestoreArrayRead(vec, &impl->fields->inputs[i])); 1436ca0f394SUmesh Unnikrishnan } 1446ca0f394SUmesh Unnikrishnan } 1456ca0f394SUmesh Unnikrishnan 1466ca0f394SUmesh Unnikrishnan // Restore output arrays 1476ca0f394SUmesh Unnikrishnan for (CeedInt i = 0; i < num_output_fields; i++) { 1486ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 1496ca0f394SUmesh Unnikrishnan if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 1506ca0f394SUmesh Unnikrishnan } else { 151*dd64fc84SJeremy L Thompson CeedVector vec; 152*dd64fc84SJeremy L Thompson 1536ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 1546ca0f394SUmesh Unnikrishnan if (vec == CEED_VECTOR_ACTIVE) vec = output_vec; 1556ca0f394SUmesh Unnikrishnan // Check for multiple output modes 1566ca0f394SUmesh Unnikrishnan CeedInt index = -1; 157*dd64fc84SJeremy L Thompson 1586ca0f394SUmesh Unnikrishnan for (CeedInt j = 0; j < i; j++) { 1596ca0f394SUmesh Unnikrishnan if (vec == output_vecs[j]) { 1606ca0f394SUmesh Unnikrishnan index = j; 1616ca0f394SUmesh Unnikrishnan break; 1626ca0f394SUmesh Unnikrishnan } 1636ca0f394SUmesh Unnikrishnan } 1646ca0f394SUmesh Unnikrishnan if (index == -1) { 1656ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedVectorRestoreArray(vec, &impl->fields->outputs[i])); 1666ca0f394SUmesh Unnikrishnan } 1676ca0f394SUmesh Unnikrishnan } 1686ca0f394SUmesh Unnikrishnan } 1696ca0f394SUmesh Unnikrishnan 1706ca0f394SUmesh Unnikrishnan // Restore context data 1716ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedQFunctionRestoreInnerContextData(qf, &qf_impl->d_c)); 1726ca0f394SUmesh Unnikrishnan return CEED_ERROR_SUCCESS; 1736ca0f394SUmesh Unnikrishnan } 1746ca0f394SUmesh Unnikrishnan 1756ca0f394SUmesh Unnikrishnan //------------------------------------------------------------------------------ 1766ca0f394SUmesh Unnikrishnan // Create operator 1776ca0f394SUmesh Unnikrishnan //------------------------------------------------------------------------------ 1786ca0f394SUmesh Unnikrishnan int CeedOperatorCreate_Sycl_gen(CeedOperator op) { 1796ca0f394SUmesh Unnikrishnan Ceed ceed; 1806ca0f394SUmesh Unnikrishnan Ceed_Sycl *sycl_data; 181*dd64fc84SJeremy L Thompson CeedOperator_Sycl_gen *impl; 182*dd64fc84SJeremy L Thompson 183*dd64fc84SJeremy L Thompson CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1846ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedGetData(ceed, &sycl_data)); 1856ca0f394SUmesh Unnikrishnan 1866ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedCalloc(1, &impl)); 1876ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedOperatorSetData(op, impl)); 1886ca0f394SUmesh Unnikrishnan 1896ca0f394SUmesh Unnikrishnan impl->indices = sycl::malloc_device<FieldsInt_Sycl>(1, sycl_data->sycl_device, sycl_data->sycl_context); 1906ca0f394SUmesh Unnikrishnan impl->fields = sycl::malloc_host<Fields_Sycl>(1, sycl_data->sycl_context); 1916ca0f394SUmesh Unnikrishnan impl->B = sycl::malloc_device<Fields_Sycl>(1, sycl_data->sycl_device, sycl_data->sycl_context); 1926ca0f394SUmesh Unnikrishnan impl->G = sycl::malloc_device<Fields_Sycl>(1, sycl_data->sycl_device, sycl_data->sycl_context); 1936ca0f394SUmesh Unnikrishnan impl->W = sycl::malloc_device<CeedScalar>(1, sycl_data->sycl_device, sycl_data->sycl_context); 1946ca0f394SUmesh Unnikrishnan 1956ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAdd_Sycl_gen)); 1966ca0f394SUmesh Unnikrishnan CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Sycl_gen)); 1976ca0f394SUmesh Unnikrishnan return CEED_ERROR_SUCCESS; 1986ca0f394SUmesh Unnikrishnan } 1996ca0f394SUmesh Unnikrishnan 2006ca0f394SUmesh Unnikrishnan //------------------------------------------------------------------------------ 201