1 // Copyright (c) 2017-2018, Lawrence Livermore National Security, LLC. 2 // Produced at the Lawrence Livermore National Laboratory. LLNL-CODE-734707. 3 // All Rights reserved. See files LICENSE and NOTICE for details. 4 // 5 // This file is part of CEED, a collection of benchmarks, miniapps, software 6 // libraries and APIs for efficient high-order finite element and spectral 7 // element discretizations for exascale applications. For more information and 8 // source code availability see http://github.com/ceed. 9 // 10 // The CEED research is supported by the Exascale Computing Project 17-SC-20-SC, 11 // a collaborative effort of two U.S. Department of Energy organizations (Office 12 // of Science and the National Nuclear Security Administration) responsible for 13 // the planning and preparation of a capable exascale ecosystem, including 14 // software, applications, hardware, advanced system engineering and early 15 // testbed platforms, in support of the nation's exascale computing imperative. 16 17 #include <ceed/ceed.h> 18 #include <ceed/backend.h> 19 #include <stddef.h> 20 #include "ceed-hip-gen.h" 21 #include "ceed-hip-gen-operator-build.h" 22 #include "../hip/ceed-hip-compile.h" 23 24 //------------------------------------------------------------------------------ 25 // Destroy operator 26 //------------------------------------------------------------------------------ 27 static int CeedOperatorDestroy_Hip_gen(CeedOperator op) { 28 int ierr; 29 CeedOperator_Hip_gen *impl; 30 ierr = CeedOperatorGetData(op, &impl); CeedChkBackend(ierr); 31 ierr = CeedFree(&impl); CeedChkBackend(ierr); 32 return CEED_ERROR_SUCCESS; 33 } 34 35 //------------------------------------------------------------------------------ 36 // Apply and add to output 37 //------------------------------------------------------------------------------ 38 static int CeedOperatorApplyAdd_Hip_gen(CeedOperator op, CeedVector invec, 39 CeedVector outvec, CeedRequest *request) { 40 int ierr; 41 Ceed ceed; 42 ierr = CeedOperatorGetCeed(op, &ceed); CeedChkBackend(ierr); 43 CeedOperator_Hip_gen *data; 44 ierr = CeedOperatorGetData(op, &data); CeedChkBackend(ierr); 45 CeedQFunction qf; 46 CeedQFunction_Hip_gen *qf_data; 47 ierr = CeedOperatorGetQFunction(op, &qf); CeedChkBackend(ierr); 48 ierr = CeedQFunctionGetData(qf, &qf_data); CeedChkBackend(ierr); 49 CeedInt nelem, numinputfields, numoutputfields; 50 ierr = CeedOperatorGetNumElements(op, &nelem); CeedChkBackend(ierr); 51 CeedOperatorField *opinputfields, *opoutputfields; 52 ierr = CeedOperatorGetFields(op, &numinputfields, &opinputfields, 53 &numoutputfields, &opoutputfields); 54 CeedChkBackend(ierr); 55 CeedQFunctionField *qfinputfields, *qfoutputfields; 56 ierr = CeedQFunctionGetFields(qf, NULL, &qfinputfields, NULL, &qfoutputfields); 57 CeedChkBackend(ierr); 58 CeedEvalMode emode; 59 CeedVector vec, outvecs[16] = {}; 60 61 //Creation of the operator 62 ierr = CeedHipGenOperatorBuild(op); CeedChkBackend(ierr); 63 64 // Input vectors 65 for (CeedInt i = 0; i < numinputfields; i++) { 66 ierr = CeedQFunctionFieldGetEvalMode(qfinputfields[i], &emode); 67 CeedChkBackend(ierr); 68 if (emode == CEED_EVAL_WEIGHT) { // Skip 69 data->fields.in[i] = NULL; 70 } else { 71 // Get input vector 72 ierr = CeedOperatorFieldGetVector(opinputfields[i], &vec); CeedChkBackend(ierr); 73 if (vec == CEED_VECTOR_ACTIVE) vec = invec; 74 ierr = CeedVectorGetArrayRead(vec, CEED_MEM_DEVICE, &data->fields.in[i]); 75 CeedChkBackend(ierr); 76 } 77 } 78 79 // Output vectors 80 for (CeedInt i = 0; i < numoutputfields; i++) { 81 ierr = CeedQFunctionFieldGetEvalMode(qfoutputfields[i], &emode); 82 CeedChkBackend(ierr); 83 if (emode == CEED_EVAL_WEIGHT) { // Skip 84 data->fields.out[i] = NULL; 85 } else { 86 // Get output vector 87 ierr = CeedOperatorFieldGetVector(opoutputfields[i], &vec); 88 CeedChkBackend(ierr); 89 if (vec == CEED_VECTOR_ACTIVE) vec = outvec; 90 outvecs[i] = vec; 91 // Check for multiple output modes 92 CeedInt index = -1; 93 for (CeedInt j = 0; j < i; j++) { 94 if (vec == outvecs[j]) { 95 index = j; 96 break; 97 } 98 } 99 if (index == -1) { 100 ierr = CeedVectorGetArray(vec, CEED_MEM_DEVICE, &data->fields.out[i]); 101 CeedChkBackend(ierr); 102 } else { 103 data->fields.out[i] = data->fields.out[index]; 104 } 105 } 106 } 107 108 // Get context data 109 ierr = CeedQFunctionGetInnerContextData(qf, CEED_MEM_DEVICE, &qf_data->d_c); 110 CeedChkBackend(ierr); 111 112 // Apply operator 113 void *opargs[] = {(void *) &nelem, &qf_data->d_c, &data->indices, 114 &data->fields, &data->B, &data->G, &data->W 115 }; 116 const CeedInt dim = data->dim; 117 const CeedInt Q1d = data->Q1d; 118 const CeedInt P1d = data->maxP1d; 119 const CeedInt thread1d = CeedIntMax(Q1d, P1d); 120 if (dim==1) { 121 CeedInt elemsPerBlock = 64*thread1d > 256? 256/thread1d : 64; 122 elemsPerBlock = elemsPerBlock>0?elemsPerBlock:1; 123 CeedInt grid = nelem/elemsPerBlock + ( (nelem/elemsPerBlock*elemsPerBlock<nelem) 124 ? 1 : 0 ); 125 CeedInt sharedMem = elemsPerBlock*thread1d*sizeof(CeedScalar); 126 ierr = CeedRunKernelDimSharedHip(ceed, data->op, grid, thread1d, 1, 127 elemsPerBlock, sharedMem, opargs); 128 } else if (dim==2) { 129 const CeedInt elemsPerBlock = thread1d<4? 16 : 2; 130 CeedInt grid = nelem/elemsPerBlock + ( (nelem/elemsPerBlock*elemsPerBlock<nelem) 131 ? 1 : 0 ); 132 CeedInt sharedMem = elemsPerBlock*thread1d*thread1d*sizeof(CeedScalar); 133 ierr = CeedRunKernelDimSharedHip(ceed, data->op, grid, thread1d, thread1d, 134 elemsPerBlock, sharedMem, opargs); 135 } else if (dim==3) { 136 const CeedInt elemsPerBlock = thread1d<6? 4 : (thread1d<8? 2 : 1); 137 CeedInt grid = nelem/elemsPerBlock + ( (nelem/elemsPerBlock*elemsPerBlock<nelem) 138 ? 1 : 0 ); 139 CeedInt sharedMem = elemsPerBlock*thread1d*thread1d*sizeof(CeedScalar); 140 ierr = CeedRunKernelDimSharedHip(ceed, data->op, grid, thread1d, thread1d, 141 elemsPerBlock, sharedMem, opargs); 142 } 143 CeedChkBackend(ierr); 144 145 // Restore input arrays 146 for (CeedInt i = 0; i < numinputfields; i++) { 147 ierr = CeedQFunctionFieldGetEvalMode(qfinputfields[i], &emode); 148 CeedChkBackend(ierr); 149 if (emode == CEED_EVAL_WEIGHT) { // Skip 150 } else { 151 ierr = CeedOperatorFieldGetVector(opinputfields[i], &vec); CeedChkBackend(ierr); 152 if (vec == CEED_VECTOR_ACTIVE) vec = invec; 153 ierr = CeedVectorRestoreArrayRead(vec, &data->fields.in[i]); 154 CeedChkBackend(ierr); 155 } 156 } 157 158 // Restore output arrays 159 for (CeedInt i = 0; i < numoutputfields; i++) { 160 ierr = CeedQFunctionFieldGetEvalMode(qfoutputfields[i], &emode); 161 CeedChkBackend(ierr); 162 if (emode == CEED_EVAL_WEIGHT) { // Skip 163 } else { 164 ierr = CeedOperatorFieldGetVector(opoutputfields[i], &vec); 165 CeedChkBackend(ierr); 166 if (vec == CEED_VECTOR_ACTIVE) vec = outvec; 167 // Check for multiple output modes 168 CeedInt index = -1; 169 for (CeedInt j = 0; j < i; j++) { 170 if (vec == outvecs[j]) { 171 index = j; 172 break; 173 } 174 } 175 if (index == -1) { 176 ierr = CeedVectorRestoreArray(vec, &data->fields.out[i]); 177 CeedChkBackend(ierr); 178 } 179 } 180 } 181 182 // Restore context data 183 ierr = CeedQFunctionRestoreInnerContextData(qf, &qf_data->d_c); 184 CeedChkBackend(ierr); 185 186 return CEED_ERROR_SUCCESS; 187 } 188 189 //------------------------------------------------------------------------------ 190 // Create operator 191 //------------------------------------------------------------------------------ 192 int CeedOperatorCreate_Hip_gen(CeedOperator op) { 193 int ierr; 194 Ceed ceed; 195 ierr = CeedOperatorGetCeed(op, &ceed); CeedChkBackend(ierr); 196 CeedOperator_Hip_gen *impl; 197 198 ierr = CeedCalloc(1, &impl); CeedChkBackend(ierr); 199 ierr = CeedOperatorSetData(op, impl); CeedChkBackend(ierr); 200 201 ierr = CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", 202 CeedOperatorApplyAdd_Hip_gen); CeedChkBackend(ierr); 203 ierr = CeedSetBackendFunction(ceed, "Operator", op, "Destroy", 204 CeedOperatorDestroy_Hip_gen); CeedChkBackend(ierr); 205 return CEED_ERROR_SUCCESS; 206 } 207 //------------------------------------------------------------------------------ 208