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-magma.h" 18 19 static int CeedDestroy_Magma(Ceed ceed) { 20 int ierr; 21 Ceed_Magma *data; 22 ierr = CeedGetData(ceed, &data); CeedChk(ierr); 23 magma_queue_destroy( data->queue ); 24 ierr = CeedFree(&data); CeedChk(ierr); 25 return 0; 26 } 27 28 static int CeedInit_Magma(const char *resource, Ceed ceed) { 29 int ierr; 30 if (strcmp(resource, "/gpu/cuda/magma") && strcmp(resource, "/gpu/hip/magma")) 31 // LCOV_EXCL_START 32 return CeedError(ceed, 1, "Magma backend cannot use resource: %s", resource); 33 // LCOV_EXCL_STOP 34 35 // Create reference CEED that implementation will be dispatched 36 // through unless overridden 37 Ceed ceedref; 38 #ifdef HAVE_HIP 39 CeedInit("/gpu/hip/ref", &ceedref); 40 #else 41 CeedInit("/gpu/cuda/ref", &ceedref); 42 #endif 43 ierr = CeedSetDelegate(ceed, ceedref); CeedChk(ierr); 44 45 ierr = magma_init(); 46 if (ierr) 47 // LCOV_EXCL_START 48 return CeedError(ceed, 1, "error in magma_init(): %d\n", ierr); 49 // LCOV_EXCL_STOP 50 51 Ceed_Magma *data; 52 ierr = CeedCalloc(sizeof(Ceed_Magma), &data); CeedChk(ierr); 53 ierr = CeedSetData(ceed, data); CeedChk(ierr); 54 55 // kernel selection 56 data->basis_kernel_mode = MAGMA_KERNEL_DIM_SPECIFIC; 57 58 // kernel max threads per thread-block 59 data->maxthreads[0] = 128; // for 1D kernels 60 data->maxthreads[1] = 128; // for 2D kernels 61 data->maxthreads[2] = 64; // for 3D kernels 62 63 // create a queue that uses the null stream 64 magma_getdevice( &(data->device) ); 65 #ifdef HAVE_HIP 66 magma_queue_create_from_hip(data->device, NULL, NULL, NULL, &(data->queue)); 67 #else 68 magma_queue_create_from_cuda(data->device, NULL, NULL, NULL, &(data->queue)); 69 #endif 70 71 ierr = CeedSetBackendFunction(ceed, "Ceed", ceed, "ElemRestrictionCreate", 72 CeedElemRestrictionCreate_Magma); CeedChk(ierr); 73 ierr = CeedSetBackendFunction(ceed, "Ceed", ceed, 74 "ElemRestrictionCreateBlocked", 75 CeedElemRestrictionCreateBlocked_Magma); CeedChk(ierr); 76 ierr = CeedSetBackendFunction(ceed, "Ceed", ceed, "BasisCreateTensorH1", 77 CeedBasisCreateTensorH1_Magma); CeedChk(ierr); 78 ierr = CeedSetBackendFunction(ceed, "Ceed", ceed, "BasisCreateH1", 79 CeedBasisCreateH1_Magma); CeedChk(ierr); 80 ierr = CeedSetBackendFunction(ceed, "Ceed", ceed, "Destroy", 81 CeedDestroy_Magma); CeedChk(ierr); 82 return 0; 83 } 84 85 __attribute__((constructor)) 86 static void Register(void) { 87 #ifdef HAVE_HIP 88 CeedRegister("/gpu/hip/magma", CeedInit_Magma, 20); 89 #else 90 CeedRegister("/gpu/cuda/magma", CeedInit_Magma, 20); 91 #endif 92 } 93