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-xsmm.h" 18 19 // Utility functions for index in pointer array 20 int CeedGetXsmmInd_Tensor(CeedInt nelem, CeedInt add, CeedTransposeMode tmode, 21 CeedInt B, CeedInt C, CeedInt J, CeedInt currdim, 22 CeedInt dim) { 23 return (nelem == 8 ? 1:0)*4*2*dim + (add ? 1:0)*4*dim + 24 (tmode ? 1:0)*2*dim + (B == J ? 1:0)*dim + currdim; 25 } 26 27 int CeedGetXsmmInd_NonTensor(CeedInt add, CeedInt P, CeedInt Q, CeedInt B, 28 CeedInt C, CeedInt J) { 29 return (C == 8 ? 1:0)*4*2 + (add ? 1:0)*4 + 30 (B == P ? (J == Q ? 0:1) : (B == Q ? 2:3)); 31 } 32 33 // Default Tensor Contact 34 static int CeedTensorContract_Xsmm_Default(CeedTensorContract contract, 35 CeedInt A, CeedInt B, CeedInt C, CeedInt J, const CeedScalar *restrict t, 36 CeedTransposeMode tmode, const CeedInt Add, const CeedScalar *restrict u, 37 CeedScalar *restrict v) { 38 CeedScalar alpha = 1.0, beta = 1.0; 39 char transu = 'N', transt = 'N'; 40 if ((tmode == CEED_TRANSPOSE && C != 1) 41 || (tmode == CEED_NOTRANSPOSE && C == 1)) 42 transt = 'T'; 43 44 if (!Add) 45 beta = 0.0; 46 47 if (C != 1) 48 for (CeedInt a=0; a<A; a++) 49 // libXSMM GEMM 50 libxsmm_dgemm(&transu, &transt, &C, &J, &B, 51 &alpha, &u[a*B*C], NULL, &t[0], NULL, 52 &beta, &v[a*J*C], NULL); 53 else 54 // libXSMM GEMM 55 libxsmm_dgemm(&transt, &transu, &J, &A, &B, 56 &alpha, &t[0], NULL, &u[0], NULL, 57 &beta, &v[0], NULL); 58 59 return 0; 60 } 61 62 // Switch for Tensor Contract 63 static int CeedTensorContractApply_Xsmm(CeedTensorContract contract, CeedInt A, 64 CeedInt B, CeedInt C, CeedInt J, 65 const CeedScalar *restrict t, 66 CeedTransposeMode tmode, 67 const CeedInt add, 68 const CeedScalar *restrict u, 69 CeedScalar *restrict v) { 70 int ierr; 71 CeedInt blksize = 8, ind, nelem; 72 CeedTensorContract_Xsmm *impl; 73 ierr = CeedTensorContractGetData(contract, (void *)&impl); CeedChk(ierr); 74 75 // Get nelem and current dim 76 CeedScalar currdim = log(C/blksize) / log(J); 77 if (!(C % blksize) && currdim - (int)currdim < 1e-15) 78 nelem = blksize; 79 else { 80 nelem = 1; 81 currdim = log(C) / log(J); 82 } 83 84 // Get kernel index 85 if (impl->tensorbasis) 86 ind = CeedGetXsmmInd_Tensor(nelem, add, tmode==CEED_TRANSPOSE?1:0, B, C, 87 J, (CeedInt)currdim, impl->dim); 88 else 89 ind = CeedGetXsmmInd_NonTensor(add, impl->P, impl->Q, B, C, J); 90 91 // Run kernel or fallback to default implementation 92 if (C != 1 && impl->kernels[ind]) 93 for (CeedInt a=0; a<A; a++) 94 impl->kernels[ind](&u[a*B*C], &t[0], &v[a*J*C], NULL, NULL, NULL); 95 else 96 CeedTensorContract_Xsmm_Default(contract, A, B, C, J, t, tmode, add, u, v); 97 98 return 0; 99 } 100 101 static int CeedTensorContractDestroy_Xsmm(CeedTensorContract contract) { 102 int ierr; 103 CeedTensorContract_Xsmm *impl; 104 ierr = CeedTensorContractGetData(contract, (void *)&impl); CeedChk(ierr); 105 ierr = CeedFree(&impl->kernels); CeedChk(ierr); 106 ierr = CeedFree(&impl); CeedChk(ierr); 107 108 return 0; 109 } 110 111 int CeedTensorContractCreate_Xsmm(CeedBasis basis, 112 CeedTensorContract contract) { 113 int ierr; 114 Ceed ceed; 115 ierr = CeedTensorContractGetCeed(contract, &ceed); CeedChk(ierr); 116 CeedTensorContract_Xsmm *impl; 117 ierr = CeedCalloc(1, &impl); CeedChk(ierr); 118 119 // Set up pointers to kernels 120 ierr = CeedBasisGetTensorStatus(basis, &impl->tensorbasis); CeedChk(ierr); 121 if (impl->tensorbasis) { 122 ierr = CeedBasisGetNumNodes1D(basis, &impl->P); CeedChk(ierr); 123 ierr = CeedBasisGetNumQuadraturePoints1D(basis, &impl->Q); CeedChk(ierr); 124 ierr = CeedBasisGetDimension(basis, &impl->dim); CeedChk(ierr); 125 // Set up kernel pointer array 126 impl->numkernels = 2*2*4*impl->dim; 127 ierr = CeedCalloc(impl->numkernels, &impl->kernels); CeedChk(ierr); 128 for (CeedInt nelem = 1; nelem <= 8; nelem+=7) { 129 for (CeedInt add = 0; add <= 1; add++) { 130 for (CeedInt tmode = 0; tmode <= 1; tmode++) { 131 for (CeedInt grad = 0; grad <=1; grad++) { 132 for (CeedInt dim = 0; dim < impl->dim; dim++) { 133 const int flags = LIBXSMM_GEMM_FLAGS('N', tmode ? 'T' : 'N'); 134 CeedInt B = grad ? impl->Q : (tmode ? impl->Q : impl->P), 135 J = grad ? impl->Q : (tmode ? impl->P : impl->Q), 136 C = nelem*CeedIntPow(J, dim); 137 int ind = CeedGetXsmmInd_Tensor(nelem, add, tmode, B, C, J, dim, 138 impl->dim); 139 CeedScalar alpha = 1.0, beta = 1.0; 140 if (!add) beta = 0.0; 141 impl->kernels[ind] = libxsmm_dmmdispatch(C, J, B, 142 NULL, NULL, NULL, &alpha, 143 &beta, &flags, NULL); 144 if (!impl->kernels[ind]) 145 return CeedError(ceed, 1, 146 "LIBXSMM kernel failed to build."); 147 } 148 } 149 } 150 } 151 } 152 } else { 153 ierr = CeedBasisGetNumNodes(basis, &impl->P); CeedChk(ierr); 154 ierr = CeedBasisGetNumQuadraturePoints(basis, &impl->Q); CeedChk(ierr); 155 ierr = CeedBasisGetDimension(basis, &impl->dim); CeedChk(ierr); 156 // Set up kernel pointer array 157 impl->numkernels = 4*2*2; 158 ierr = CeedCalloc(impl->numkernels, &impl->kernels); CeedChk(ierr); 159 for (CeedInt nelem = 1; nelem <= 8; nelem+=7) { 160 for (CeedInt add = 0; add <= 1; add++) { 161 for (CeedInt tmode = 0; tmode <= 1; tmode++) { 162 for (CeedInt grad = 1; grad <= impl->dim; grad+=impl->dim-1) { 163 const int flags = LIBXSMM_GEMM_FLAGS('N', tmode ? 'T' : 'N'); 164 CeedInt B = tmode ? grad*impl->Q : impl->P, 165 J = tmode ? impl->P : grad*impl->Q; 166 int ind = CeedGetXsmmInd_NonTensor(add, impl->P, impl->Q, B, nelem, 167 J); 168 CeedScalar alpha = 1.0, beta = 1.0; 169 if (!add) beta = 0.0; 170 impl->kernels[ind] = libxsmm_dmmdispatch(nelem, J, B, 171 NULL, NULL, NULL, &alpha, 172 &beta, &flags, NULL); 173 if (!impl->kernels[ind]) 174 return CeedError(ceed, 1, 175 "LIBXSMM kernel failed to build."); 176 } 177 } 178 } 179 } 180 } 181 ierr = CeedTensorContractSetData(contract, (void *)&impl); CeedChk(ierr); 182 183 ierr = CeedSetBackendFunction(ceed, "TensorContract", contract, "Apply", 184 CeedTensorContractApply_Xsmm); CeedChk(ierr); 185 ierr = CeedSetBackendFunction(ceed, "TensorContract", contract, "Destroy", 186 CeedTensorContractDestroy_Xsmm); CeedChk(ierr); 187 188 return 0; 189 } 190