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_C1(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 // libXSMM GEMM 48 libxsmm_dgemm(&transt, &transu, &J, &A, &B, 49 &alpha, &t[0], NULL, &u[0], NULL, 50 &beta, &v[0], NULL); 51 52 return 0; 53 } 54 55 // Switch for Tensor Contract 56 static int CeedTensorContractApply_Xsmm(CeedTensorContract contract, CeedInt A, 57 CeedInt B, CeedInt C, CeedInt J, 58 const CeedScalar *restrict t, 59 CeedTransposeMode tmode, 60 const CeedInt add, 61 const CeedScalar *restrict u, 62 CeedScalar *restrict v) { 63 int ierr; 64 CeedInt blksize = 8, ind, nelem; 65 CeedTensorContract_Xsmm *impl; 66 ierr = CeedTensorContractGetData(contract, (void *)&impl); CeedChk(ierr); 67 68 // Get nelem and current dim 69 CeedScalar currdim = log(C/blksize) / log(J); 70 if (!(C % blksize) && currdim - (int)currdim < 1e-15) { 71 nelem = blksize; 72 } else { 73 nelem = 1; 74 currdim = log(C) / log(J); 75 } 76 77 // Get kernel index 78 if (impl->tensorbasis) 79 ind = CeedGetXsmmInd_Tensor(nelem, add, tmode==CEED_TRANSPOSE?1:0, B, C, J, 80 (CeedInt)currdim, impl->dim); 81 else 82 ind = CeedGetXsmmInd_NonTensor(add, impl->P, impl->Q, B, C, J); 83 84 // Run kernel or fallback to default implementation 85 if (C != 1) 86 for (CeedInt a=0; a<A; a++) 87 impl->kernels[ind](&u[a*B*C], &t[0], &v[a*J*C], NULL, NULL, NULL); 88 else 89 CeedTensorContract_Xsmm_C1(contract, A, B, C, J, t, tmode, add, u, v); 90 91 return 0; 92 } 93 94 static int CeedTensorContractDestroy_Xsmm(CeedTensorContract contract) { 95 int ierr; 96 CeedTensorContract_Xsmm *impl; 97 ierr = CeedTensorContractGetData(contract, (void *)&impl); CeedChk(ierr); 98 ierr = CeedFree(&impl->kernels); CeedChk(ierr); 99 ierr = CeedFree(&impl); CeedChk(ierr); 100 101 return 0; 102 } 103 104 int CeedTensorContractCreate_Xsmm(CeedBasis basis, 105 CeedTensorContract contract) { 106 int ierr; 107 Ceed ceed; 108 ierr = CeedTensorContractGetCeed(contract, &ceed); CeedChk(ierr); 109 CeedTensorContract_Xsmm *impl; 110 ierr = CeedCalloc(1, &impl); CeedChk(ierr); 111 112 // Set up pointers to kernels 113 ierr = CeedBasisGetTensorStatus(basis, &impl->tensorbasis); CeedChk(ierr); 114 if (impl->tensorbasis) { 115 ierr = CeedBasisGetNumNodes1D(basis, &impl->P); CeedChk(ierr); 116 ierr = CeedBasisGetNumQuadraturePoints1D(basis, &impl->Q); CeedChk(ierr); 117 ierr = CeedBasisGetDimension(basis, &impl->dim); CeedChk(ierr); 118 // Set up kernel pointer array 119 impl->numkernels = 2*2*4*impl->dim; 120 ierr = CeedCalloc(impl->numkernels, &impl->kernels); CeedChk(ierr); 121 for (CeedInt nelem = 1; nelem <= 8; nelem+=7) 122 for (CeedInt add = 0; add <= 1; add++) 123 for (CeedInt tmode = 0; tmode <= 1; tmode++) 124 for (CeedInt grad = 0; grad <=1; grad++) 125 for (CeedInt dim = 0; dim < impl->dim; dim++) { 126 const int flags = LIBXSMM_GEMM_FLAGS('N', tmode ? 'T' : 'N'); 127 CeedInt B = grad ? impl->Q : (tmode ? impl->Q : impl->P), 128 J = grad ? impl->Q : (tmode ? impl->P : impl->Q), 129 C = nelem*CeedIntPow(J, dim); 130 int ind = CeedGetXsmmInd_Tensor(nelem, add, tmode, B, C, J, dim, 131 impl->dim); 132 CeedScalar alpha = 1.0, beta = 1.0; 133 if (!add) beta = 0.0; 134 impl->kernels[ind] = libxsmm_dmmdispatch(C, J, B, 135 NULL, NULL, NULL, &alpha, 136 &beta, &flags, NULL); 137 if (!impl->kernels[ind]) 138 // LCOV_EXCL_START 139 return CeedError(ceed, 1, "LIBXSMM kernel failed to build."); 140 // LCOV_EXCL_STOP 141 } 142 } else { 143 ierr = CeedBasisGetNumNodes(basis, &impl->P); CeedChk(ierr); 144 ierr = CeedBasisGetNumQuadraturePoints(basis, &impl->Q); CeedChk(ierr); 145 ierr = CeedBasisGetDimension(basis, &impl->dim); CeedChk(ierr); 146 // Set up kernel pointer array 147 impl->numkernels = 4*2*2; 148 ierr = CeedCalloc(impl->numkernels, &impl->kernels); CeedChk(ierr); 149 for (CeedInt nelem = 1; nelem <= 8; nelem+=7) 150 for (CeedInt add = 0; add <= 1; add++) 151 for (CeedInt tmode = 0; tmode <= 1; tmode++) 152 for (CeedInt grad = 1; grad <= impl->dim; grad+=impl->dim-1) { 153 const int flags = LIBXSMM_GEMM_FLAGS('N', tmode ? 'T' : 'N'); 154 CeedInt B = tmode ? grad*impl->Q : impl->P, 155 J = tmode ? impl->P : grad*impl->Q; 156 int ind = CeedGetXsmmInd_NonTensor(add, impl->P, impl->Q, B, nelem, 157 J); 158 CeedScalar alpha = 1.0, beta = 1.0; 159 if (!add) beta = 0.0; 160 impl->kernels[ind] = libxsmm_dmmdispatch(nelem, J, B, 161 NULL, NULL, NULL, &alpha, 162 &beta, &flags, NULL); 163 if (!impl->kernels[ind]) 164 // LCOV_EXCL_START 165 return CeedError(ceed, 1, "LIBXSMM kernel failed to build."); 166 // LCOV_EXCL_STOP 167 } 168 } 169 ierr = CeedTensorContractSetData(contract, (void *)&impl); CeedChk(ierr); 170 171 ierr = CeedSetBackendFunction(ceed, "TensorContract", contract, "Apply", 172 CeedTensorContractApply_Xsmm); CeedChk(ierr); 173 ierr = CeedSetBackendFunction(ceed, "TensorContract", contract, "Destroy", 174 CeedTensorContractDestroy_Xsmm); CeedChk(ierr); 175 176 return 0; 177 } 178