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-backend.h> 18 #include "ceed-cuda-gen.h" 19 #include "ceed-cuda-gen-operator-build.h" 20 #include "../cuda/ceed-cuda.h" 21 22 static int CeedOperatorDestroy_Cuda_gen(CeedOperator op) { 23 int ierr; 24 CeedOperator_Cuda_gen *impl; 25 ierr = CeedOperatorGetData(op, (void *)&impl); CeedChk(ierr); 26 27 ierr = CeedFree(&impl); CeedChk(ierr); 28 return 0; 29 } 30 31 static int CeedOperatorApplyAdd_Cuda_gen(CeedOperator op, CeedVector invec, 32 CeedVector outvec, CeedRequest *request) { 33 int ierr; 34 Ceed ceed; 35 ierr = CeedOperatorGetCeed(op, &ceed); CeedChk(ierr); 36 CeedOperator_Cuda_gen *data; 37 ierr = CeedOperatorGetData(op, (void *)&data); CeedChk(ierr); 38 CeedQFunction qf; 39 CeedQFunction_Cuda_gen *qf_data; 40 ierr = CeedOperatorGetQFunction(op, &qf); CeedChk(ierr); 41 ierr = CeedQFunctionGetData(qf, (void **)&qf_data); CeedChk(ierr); 42 CeedInt nelem, numinputfields, numoutputfields; 43 ierr = CeedOperatorGetNumElements(op, &nelem); CeedChk(ierr); 44 ierr = CeedQFunctionGetNumArgs(qf, &numinputfields, &numoutputfields); 45 CeedChk(ierr); 46 CeedOperatorField *opinputfields, *opoutputfields; 47 ierr = CeedOperatorGetFields(op, &opinputfields, &opoutputfields); 48 CeedChk(ierr); 49 CeedQFunctionField *qfinputfields, *qfoutputfields; 50 ierr = CeedQFunctionGetFields(qf, &qfinputfields, &qfoutputfields); 51 CeedChk(ierr); 52 CeedEvalMode emode; 53 CeedVector vec; 54 55 //Creation of the operator 56 ierr = CeedCudaGenOperatorBuild(op); CeedChk(ierr); 57 58 // Input vectors 59 for (CeedInt i = 0; i < numinputfields; i++) { 60 ierr = CeedQFunctionFieldGetEvalMode(qfinputfields[i], &emode); 61 CeedChk(ierr); 62 if (emode == CEED_EVAL_WEIGHT) { // Skip 63 data->fields.in[i] = NULL; 64 } else { 65 // Get input vector 66 ierr = CeedOperatorFieldGetVector(opinputfields[i], &vec); CeedChk(ierr); 67 if (vec == CEED_VECTOR_ACTIVE) vec = invec; 68 ierr = CeedVectorGetArrayRead(vec, CEED_MEM_DEVICE, &data->fields.in[i]); 69 CeedChk(ierr); 70 } 71 } 72 73 // Output vectors 74 for (CeedInt i = 0; i < numoutputfields; i++) { 75 ierr = CeedQFunctionFieldGetEvalMode(qfoutputfields[i], &emode); 76 CeedChk(ierr); 77 if (emode == CEED_EVAL_WEIGHT) { // Skip 78 data->fields.out[i] = NULL; 79 } else { 80 // Get output vector 81 ierr = CeedOperatorFieldGetVector(opoutputfields[i], &vec); CeedChk(ierr); 82 if (vec == CEED_VECTOR_ACTIVE) vec = outvec; 83 ierr = CeedVectorGetArray(vec, CEED_MEM_DEVICE, &data->fields.out[i]); 84 CeedChk(ierr); 85 } 86 } 87 88 // Copy the context 89 size_t ctxsize; 90 ierr = CeedQFunctionGetContextSize(qf, &ctxsize); CeedChk(ierr); 91 if (ctxsize > 0) { 92 if (!qf_data->d_c) { 93 ierr = cudaMalloc(&qf_data->d_c, ctxsize); CeedChk_Cu(ceed, ierr); 94 } 95 void *ctx; 96 ierr = CeedQFunctionGetInnerContext(qf, &ctx); CeedChk(ierr); 97 ierr = cudaMemcpy(qf_data->d_c, ctx, ctxsize, cudaMemcpyHostToDevice); 98 CeedChk_Cu(ceed, ierr); 99 } 100 101 // Apply operator 102 void *opargs[] = {(void *) &nelem, &qf_data->d_c, &data->indices, 103 &data->fields, &data->B, &data->G, &data->W 104 }; 105 const CeedInt dim = data->dim; 106 const CeedInt Q1d = data->Q1d; 107 if (dim==1) { 108 const CeedInt elemsPerBlock = 32; 109 CeedInt grid = nelem/elemsPerBlock + ( (nelem/elemsPerBlock*elemsPerBlock<nelem) 110 ? 1 : 0 ); 111 CeedInt sharedMem = elemsPerBlock*Q1d*sizeof(CeedScalar); 112 ierr = CeedRunKernelDimSharedCuda(ceed, data->op, grid, Q1d, 1, elemsPerBlock, 113 sharedMem, opargs); 114 } else if (dim==2) { 115 const CeedInt elemsPerBlock = Q1d<4? 16 : 2; 116 CeedInt grid = nelem/elemsPerBlock + ( (nelem/elemsPerBlock*elemsPerBlock<nelem) 117 ? 1 : 0 ); 118 CeedInt sharedMem = elemsPerBlock*Q1d*Q1d*sizeof(CeedScalar); 119 ierr = CeedRunKernelDimSharedCuda(ceed, data->op, grid, Q1d, Q1d, 120 elemsPerBlock, sharedMem, opargs); 121 } else if (dim==3) { 122 const CeedInt elemsPerBlock = Q1d<6? 4 : (Q1d<8? 2 : 1); 123 CeedInt grid = nelem/elemsPerBlock + ( (nelem/elemsPerBlock*elemsPerBlock<nelem) 124 ? 1 : 0 ); 125 CeedInt sharedMem = elemsPerBlock*Q1d*Q1d*sizeof(CeedScalar); 126 ierr = CeedRunKernelDimSharedCuda(ceed, data->op, grid, Q1d, Q1d, 127 elemsPerBlock, sharedMem, opargs); 128 } 129 CeedChk(ierr); 130 131 // Restore input arrays 132 for (CeedInt i = 0; i < numinputfields; i++) { 133 ierr = CeedQFunctionFieldGetEvalMode(qfinputfields[i], &emode); 134 CeedChk(ierr); 135 if (emode == CEED_EVAL_WEIGHT) { // Skip 136 } else { 137 ierr = CeedOperatorFieldGetVector(opinputfields[i], &vec); CeedChk(ierr); 138 if (vec == CEED_VECTOR_ACTIVE) vec = invec; 139 ierr = CeedVectorRestoreArrayRead(vec, &data->fields.in[i]); 140 CeedChk(ierr); 141 } 142 } 143 144 // Restore output arrays 145 for (CeedInt i = 0; i < numoutputfields; i++) { 146 ierr = CeedQFunctionFieldGetEvalMode(qfoutputfields[i], &emode); 147 CeedChk(ierr); 148 if (emode == CEED_EVAL_WEIGHT) { // Skip 149 } else { 150 ierr = CeedOperatorFieldGetVector(opoutputfields[i], &vec); CeedChk(ierr); 151 if (vec == CEED_VECTOR_ACTIVE) vec = outvec; 152 ierr = CeedVectorRestoreArray(vec, &data->fields.out[i]); 153 CeedChk(ierr); 154 } 155 } 156 157 return 0; 158 } 159 160 static int CeedOperatorAssembleLinearQFunction_Cuda(CeedOperator op) { 161 int ierr; 162 Ceed ceed; 163 ierr = CeedOperatorGetCeed(op, &ceed); CeedChk(ierr); 164 return CeedError(ceed, 1, "Backend does not implement QFunction assembly"); 165 } 166 167 int CeedOperatorCreate_Cuda_gen(CeedOperator op) { 168 int ierr; 169 Ceed ceed; 170 ierr = CeedOperatorGetCeed(op, &ceed); CeedChk(ierr); 171 CeedOperator_Cuda_gen *impl; 172 173 ierr = CeedCalloc(1, &impl); CeedChk(ierr); 174 ierr = CeedOperatorSetData(op, (void *)&impl); 175 176 ierr = CeedSetBackendFunction(ceed, "Operator", op, "AssembleLinearQFunction", 177 CeedOperatorAssembleLinearQFunction_Cuda); 178 CeedChk(ierr); 179 ierr = CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", 180 CeedOperatorApplyAdd_Cuda_gen); CeedChk(ierr); 181 ierr = CeedSetBackendFunction(ceed, "Operator", op, "Destroy", 182 CeedOperatorDestroy_Cuda_gen); CeedChk(ierr); 183 return 0; 184 } 185