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 //------------------------------------------------------------------------------ 20 // Get Kernel Index 21 //------------------------------------------------------------------------------ 22 int CeedGetXsmmInd_Tensor(CeedInt nelem, CeedInt add, CeedTransposeMode tmode, 23 CeedInt B, CeedInt C, CeedInt J, CeedInt currdim, 24 CeedInt dim) { 25 return (nelem == 8 ? 1:0)*4*2*dim + (add ? 1:0)*4*dim + 26 (tmode ? 1:0)*2*dim + (B == J ? 1:0)*dim + currdim; 27 } 28 29 int CeedGetXsmmInd_NonTensor(CeedInt add, CeedInt P, CeedInt Q, CeedInt B, 30 CeedInt C, CeedInt J) { 31 return (C == 8 ? 1:0)*4*2 + (add ? 1:0)*4 + 32 (B == P ? (J == Q ? 0:1) : (B == Q ? 2:3)); 33 } 34 35 //------------------------------------------------------------------------------ 36 // Tensor Contract C=1 37 //------------------------------------------------------------------------------ 38 static int CeedTensorContract_Xsmm_C1(CeedTensorContract contract, 39 CeedInt A, CeedInt B, CeedInt C, 40 CeedInt J, const CeedScalar *restrict t, 41 CeedTransposeMode tmode, 42 const CeedInt Add, 43 const CeedScalar *restrict u, 44 CeedScalar *restrict v) { 45 CeedScalar alpha = 1.0, beta = 1.0; 46 char transu = 'N', transt = 'N'; 47 if ((tmode == CEED_TRANSPOSE && C != 1) 48 || (tmode == CEED_NOTRANSPOSE && C == 1)) 49 transt = 'T'; 50 51 if (!Add) 52 beta = 0.0; 53 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 //------------------------------------------------------------------------------ 63 // Tensor Contract Apply 64 //------------------------------------------------------------------------------ 65 static int CeedTensorContractApply_Xsmm(CeedTensorContract contract, CeedInt A, 66 CeedInt B, CeedInt C, CeedInt J, 67 const CeedScalar *restrict t, 68 CeedTransposeMode tmode, 69 const CeedInt add, 70 const CeedScalar *restrict u, 71 CeedScalar *restrict v) { 72 int ierr; 73 CeedInt blksize = 8, ind, nelem; 74 CeedTensorContract_Xsmm *impl; 75 ierr = CeedTensorContractGetData(contract, (void *)&impl); CeedChk(ierr); 76 77 // Get nelem and current dim 78 CeedScalar currdim = log(C/blksize) / log(J); 79 if (!(C % blksize) && currdim - (int)currdim < 1e-15) { 80 nelem = blksize; 81 } else { 82 nelem = 1; 83 currdim = log(C) / log(J); 84 } 85 86 // Get kernel index 87 if (impl->tensorbasis) 88 ind = CeedGetXsmmInd_Tensor(nelem, add, tmode==CEED_TRANSPOSE?1:0, B, C, J, 89 (CeedInt)currdim, impl->dim); 90 else 91 ind = CeedGetXsmmInd_NonTensor(add, impl->P, impl->Q, B, C, J); 92 93 // Run kernel or fallback to default implementation 94 if (C != 1) 95 for (CeedInt a=0; a<A; a++) 96 impl->kernels[ind](&u[a*B*C], &t[0], &v[a*J*C], NULL, NULL, NULL); 97 else 98 CeedTensorContract_Xsmm_C1(contract, A, B, C, J, t, tmode, add, u, v); 99 100 return 0; 101 } 102 103 //------------------------------------------------------------------------------ 104 // Tensor Contract Destroy 105 //------------------------------------------------------------------------------ 106 static int CeedTensorContractDestroy_Xsmm(CeedTensorContract contract) { 107 int ierr; 108 CeedTensorContract_Xsmm *impl; 109 ierr = CeedTensorContractGetData(contract, (void *)&impl); CeedChk(ierr); 110 ierr = CeedFree(&impl->kernels); CeedChk(ierr); 111 ierr = CeedFree(&impl); CeedChk(ierr); 112 113 return 0; 114 } 115 116 //------------------------------------------------------------------------------ 117 // Tensor Contract Create 118 //------------------------------------------------------------------------------ 119 int CeedTensorContractCreate_Xsmm(CeedBasis basis, 120 CeedTensorContract contract) { 121 int ierr; 122 Ceed ceed; 123 ierr = CeedTensorContractGetCeed(contract, &ceed); CeedChk(ierr); 124 CeedTensorContract_Xsmm *impl; 125 ierr = CeedCalloc(1, &impl); CeedChk(ierr); 126 127 // Set up pointers to kernels 128 ierr = CeedBasisGetTensorStatus(basis, &impl->tensorbasis); CeedChk(ierr); 129 if (impl->tensorbasis) { 130 ierr = CeedBasisGetNumNodes1D(basis, &impl->P); CeedChk(ierr); 131 ierr = CeedBasisGetNumQuadraturePoints1D(basis, &impl->Q); CeedChk(ierr); 132 ierr = CeedBasisGetDimension(basis, &impl->dim); CeedChk(ierr); 133 // Set up kernel pointer array 134 impl->numkernels = 2*2*4*impl->dim; 135 ierr = CeedCalloc(impl->numkernels, &impl->kernels); CeedChk(ierr); 136 for (CeedInt nelem = 1; nelem <= 8; nelem+=7) 137 for (CeedInt add = 0; add <= 1; add++) 138 for (CeedInt tmode = 0; tmode <= 1; tmode++) 139 for (CeedInt grad = 0; grad <=1; grad++) 140 for (CeedInt dim = 0; dim < impl->dim; dim++) { 141 const int flags = LIBXSMM_GEMM_FLAGS('N', tmode ? 'T' : 'N'); 142 CeedInt B = grad ? impl->Q : (tmode ? impl->Q : impl->P), 143 J = grad ? impl->Q : (tmode ? impl->P : impl->Q), 144 C = nelem*CeedIntPow(J, dim); 145 int ind = CeedGetXsmmInd_Tensor(nelem, add, tmode, B, C, J, dim, 146 impl->dim); 147 CeedScalar alpha = 1.0, beta = 1.0; 148 if (!add) beta = 0.0; 149 impl->kernels[ind] = libxsmm_dmmdispatch(C, J, B, 150 NULL, NULL, NULL, &alpha, 151 &beta, &flags, NULL); 152 if (!impl->kernels[ind]) 153 // LCOV_EXCL_START 154 return CeedError(ceed, 1, "LIBXSMM kernel failed to build."); 155 // LCOV_EXCL_STOP 156 } 157 } else { 158 ierr = CeedBasisGetNumNodes(basis, &impl->P); CeedChk(ierr); 159 ierr = CeedBasisGetNumQuadraturePoints(basis, &impl->Q); CeedChk(ierr); 160 ierr = CeedBasisGetDimension(basis, &impl->dim); CeedChk(ierr); 161 // Set up kernel pointer array 162 impl->numkernels = 4*2*2; 163 ierr = CeedCalloc(impl->numkernels, &impl->kernels); CeedChk(ierr); 164 for (CeedInt nelem = 1; nelem <= 8; nelem+=7) 165 for (CeedInt add = 0; add <= 1; add++) 166 for (CeedInt tmode = 0; tmode <= 1; tmode++) 167 for (CeedInt grad = 1; grad <= impl->dim; grad+=impl->dim-1) { 168 const int flags = LIBXSMM_GEMM_FLAGS('N', tmode ? 'T' : 'N'); 169 CeedInt B = tmode ? grad*impl->Q : impl->P, 170 J = tmode ? impl->P : grad*impl->Q; 171 int ind = CeedGetXsmmInd_NonTensor(add, impl->P, impl->Q, B, nelem, 172 J); 173 CeedScalar alpha = 1.0, beta = 1.0; 174 if (!add) beta = 0.0; 175 impl->kernels[ind] = libxsmm_dmmdispatch(nelem, J, B, 176 NULL, NULL, NULL, &alpha, 177 &beta, &flags, NULL); 178 if (!impl->kernels[ind]) 179 // LCOV_EXCL_START 180 return CeedError(ceed, 1, "LIBXSMM kernel failed to build."); 181 // LCOV_EXCL_STOP 182 } 183 } 184 ierr = CeedTensorContractSetData(contract, (void *)&impl); CeedChk(ierr); 185 186 ierr = CeedSetBackendFunction(ceed, "TensorContract", contract, "Apply", 187 CeedTensorContractApply_Xsmm); CeedChk(ierr); 188 ierr = CeedSetBackendFunction(ceed, "TensorContract", contract, "Destroy", 189 CeedTensorContractDestroy_Xsmm); CeedChk(ierr); 190 191 return 0; 192 } 193 //------------------------------------------------------------------------------ 194