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-cuda-gen.h" 18 #include "ceed-cuda-gen-operator-build.h" 19 20 //------------------------------------------------------------------------------ 21 // Destroy operator 22 //------------------------------------------------------------------------------ 23 static int CeedOperatorDestroy_Cuda_gen(CeedOperator op) { 24 int ierr; 25 CeedOperator_Cuda_gen *impl; 26 ierr = CeedOperatorGetData(op, &impl); CeedChk(ierr); 27 ierr = CeedFree(&impl); CeedChk(ierr); 28 return 0; 29 } 30 31 //------------------------------------------------------------------------------ 32 // Apply and add to output 33 //------------------------------------------------------------------------------ 34 static int CeedOperatorApplyAdd_Cuda_gen(CeedOperator op, CeedVector invec, 35 CeedVector outvec, CeedRequest *request) { 36 int ierr; 37 Ceed ceed; 38 ierr = CeedOperatorGetCeed(op, &ceed); CeedChk(ierr); 39 CeedOperator_Cuda_gen *data; 40 ierr = CeedOperatorGetData(op, &data); CeedChk(ierr); 41 CeedQFunction qf; 42 CeedQFunction_Cuda_gen *qf_data; 43 ierr = CeedOperatorGetQFunction(op, &qf); CeedChk(ierr); 44 ierr = CeedQFunctionGetData(qf, &qf_data); CeedChk(ierr); 45 CeedInt nelem, numinputfields, numoutputfields; 46 ierr = CeedOperatorGetNumElements(op, &nelem); CeedChk(ierr); 47 ierr = CeedQFunctionGetNumArgs(qf, &numinputfields, &numoutputfields); 48 CeedChk(ierr); 49 CeedOperatorField *opinputfields, *opoutputfields; 50 ierr = CeedOperatorGetFields(op, &opinputfields, &opoutputfields); 51 CeedChk(ierr); 52 CeedQFunctionField *qfinputfields, *qfoutputfields; 53 ierr = CeedQFunctionGetFields(qf, &qfinputfields, &qfoutputfields); 54 CeedChk(ierr); 55 CeedEvalMode emode; 56 CeedVector vec, outvecs[16] = {}; 57 58 //Creation of the operator 59 ierr = CeedCudaGenOperatorBuild(op); CeedChk(ierr); 60 61 // Input vectors 62 for (CeedInt i = 0; i < numinputfields; i++) { 63 ierr = CeedQFunctionFieldGetEvalMode(qfinputfields[i], &emode); 64 CeedChk(ierr); 65 if (emode == CEED_EVAL_WEIGHT) { // Skip 66 data->fields.in[i] = NULL; 67 } else { 68 // Get input vector 69 ierr = CeedOperatorFieldGetVector(opinputfields[i], &vec); CeedChk(ierr); 70 if (vec == CEED_VECTOR_ACTIVE) vec = invec; 71 ierr = CeedVectorGetArrayRead(vec, CEED_MEM_DEVICE, &data->fields.in[i]); 72 CeedChk(ierr); 73 } 74 } 75 76 // Output vectors 77 for (CeedInt i = 0; i < numoutputfields; i++) { 78 ierr = CeedQFunctionFieldGetEvalMode(qfoutputfields[i], &emode); 79 CeedChk(ierr); 80 if (emode == CEED_EVAL_WEIGHT) { // Skip 81 data->fields.out[i] = NULL; 82 } else { 83 // Get output vector 84 ierr = CeedOperatorFieldGetVector(opoutputfields[i], &vec); CeedChk(ierr); 85 if (vec == CEED_VECTOR_ACTIVE) vec = outvec; 86 outvecs[i] = vec; 87 // Check for multiple output modes 88 CeedInt index = -1; 89 for (CeedInt j = 0; j < i; j++) { 90 if (vec == outvecs[j]) { 91 index = j; 92 break; 93 } 94 } 95 if (index == -1) { 96 ierr = CeedVectorGetArray(vec, CEED_MEM_DEVICE, &data->fields.out[i]); 97 CeedChk(ierr); 98 } else { 99 data->fields.out[i] = data->fields.out[index]; 100 } 101 } 102 } 103 104 // Get context data 105 CeedQFunctionContext ctx; 106 ierr = CeedQFunctionGetInnerContext(qf, &ctx); CeedChk(ierr); 107 if (ctx) { 108 ierr = CeedQFunctionContextGetData(ctx, CEED_MEM_DEVICE, &qf_data->d_c); 109 CeedChk(ierr); 110 } 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 if (dim==1) { 119 const CeedInt elemsPerBlock = 32; 120 CeedInt grid = nelem/elemsPerBlock + ( (nelem/elemsPerBlock*elemsPerBlock<nelem) 121 ? 1 : 0 ); 122 CeedInt sharedMem = elemsPerBlock*Q1d*sizeof(CeedScalar); 123 ierr = CeedRunKernelDimSharedCuda(ceed, data->op, grid, Q1d, 1, elemsPerBlock, 124 sharedMem, opargs); 125 } else if (dim==2) { 126 const CeedInt elemsPerBlock = Q1d<4? 16 : 2; 127 CeedInt grid = nelem/elemsPerBlock + ( (nelem/elemsPerBlock*elemsPerBlock<nelem) 128 ? 1 : 0 ); 129 CeedInt sharedMem = elemsPerBlock*Q1d*Q1d*sizeof(CeedScalar); 130 ierr = CeedRunKernelDimSharedCuda(ceed, data->op, grid, Q1d, Q1d, 131 elemsPerBlock, sharedMem, opargs); 132 } else if (dim==3) { 133 const CeedInt elemsPerBlock = Q1d<6? 4 : (Q1d<8? 2 : 1); 134 CeedInt grid = nelem/elemsPerBlock + ( (nelem/elemsPerBlock*elemsPerBlock<nelem) 135 ? 1 : 0 ); 136 CeedInt sharedMem = elemsPerBlock*Q1d*Q1d*sizeof(CeedScalar); 137 ierr = CeedRunKernelDimSharedCuda(ceed, data->op, grid, Q1d, Q1d, 138 elemsPerBlock, sharedMem, opargs); 139 } 140 CeedChk(ierr); 141 142 // Restore input arrays 143 for (CeedInt i = 0; i < numinputfields; i++) { 144 ierr = CeedQFunctionFieldGetEvalMode(qfinputfields[i], &emode); 145 CeedChk(ierr); 146 if (emode == CEED_EVAL_WEIGHT) { // Skip 147 } else { 148 ierr = CeedOperatorFieldGetVector(opinputfields[i], &vec); CeedChk(ierr); 149 if (vec == CEED_VECTOR_ACTIVE) vec = invec; 150 ierr = CeedVectorRestoreArrayRead(vec, &data->fields.in[i]); 151 CeedChk(ierr); 152 } 153 } 154 155 // Restore output arrays 156 for (CeedInt i = 0; i < numoutputfields; i++) { 157 ierr = CeedQFunctionFieldGetEvalMode(qfoutputfields[i], &emode); 158 CeedChk(ierr); 159 if (emode == CEED_EVAL_WEIGHT) { // Skip 160 } else { 161 ierr = CeedOperatorFieldGetVector(opoutputfields[i], &vec); CeedChk(ierr); 162 if (vec == CEED_VECTOR_ACTIVE) vec = outvec; 163 // Check for multiple output modes 164 CeedInt index = -1; 165 for (CeedInt j = 0; j < i; j++) { 166 if (vec == outvecs[j]) { 167 index = j; 168 break; 169 } 170 } 171 if (index == -1) { 172 ierr = CeedVectorRestoreArray(vec, &data->fields.out[i]); 173 CeedChk(ierr); 174 } 175 } 176 } 177 178 // Restore context data 179 if (ctx) { 180 ierr = CeedQFunctionContextRestoreData(ctx, &qf_data->d_c); 181 CeedChk(ierr); 182 } 183 return 0; 184 } 185 186 //------------------------------------------------------------------------------ 187 // Create FDM element inverse not supported 188 //------------------------------------------------------------------------------ 189 static int CeedOperatorCreateFDMElementInverse_Cuda(CeedOperator op) { 190 // LCOV_EXCL_START 191 int ierr; 192 Ceed ceed; 193 ierr = CeedOperatorGetCeed(op, &ceed); CeedChk(ierr); 194 return CeedError(ceed, 1, "Backend does not implement FDM inverse creation"); 195 // LCOV_EXCL_STOP 196 } 197 198 //------------------------------------------------------------------------------ 199 // Create operator 200 //------------------------------------------------------------------------------ 201 int CeedOperatorCreate_Cuda_gen(CeedOperator op) { 202 int ierr; 203 Ceed ceed; 204 ierr = CeedOperatorGetCeed(op, &ceed); CeedChk(ierr); 205 CeedOperator_Cuda_gen *impl; 206 207 ierr = CeedCalloc(1, &impl); CeedChk(ierr); 208 ierr = CeedOperatorSetData(op, impl); CeedChk(ierr); 209 210 ierr = CeedSetBackendFunction(ceed, "Operator", op, "CreateFDMElementInverse", 211 CeedOperatorCreateFDMElementInverse_Cuda); 212 CeedChk(ierr); 213 ierr = CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", 214 CeedOperatorApplyAdd_Cuda_gen); CeedChk(ierr); 215 ierr = CeedSetBackendFunction(ceed, "Operator", op, "Destroy", 216 CeedOperatorDestroy_Cuda_gen); CeedChk(ierr); 217 return 0; 218 } 219 //------------------------------------------------------------------------------ 220