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