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.h> 18 #include <ceed-backend.h> 19 #include <string.h> 20 #include "ceed-magma.h" 21 22 static int CeedDestroy_Magma(Ceed ceed) { 23 int ierr; 24 Ceed_Magma *data; 25 ierr = CeedGetData(ceed, &data); CeedChk(ierr); 26 magma_queue_destroy( data->queue ); 27 ierr = CeedFree(&data); CeedChk(ierr); 28 return 0; 29 } 30 31 static int CeedInit_Magma(const char *resource, Ceed ceed) { 32 int ierr; 33 if (strcmp(resource, "/gpu/cuda/magma") && strcmp(resource, "/gpu/hip/magma")) 34 // LCOV_EXCL_START 35 return CeedError(ceed, 1, "Magma backend cannot use resource: %s", resource); 36 // LCOV_EXCL_STOP 37 38 // Create reference CEED that implementation will be dispatched 39 // through unless overridden 40 Ceed ceedref; 41 #ifdef HAVE_HIP 42 CeedInit("/gpu/hip/ref", &ceedref); 43 #else 44 CeedInit("/gpu/cuda/ref", &ceedref); 45 #endif 46 ierr = CeedSetDelegate(ceed, ceedref); CeedChk(ierr); 47 48 ierr = magma_init(); 49 if (ierr) 50 // LCOV_EXCL_START 51 return CeedError(ceed, 1, "error in magma_init(): %d\n", ierr); 52 // LCOV_EXCL_STOP 53 54 Ceed_Magma *data; 55 ierr = CeedCalloc(sizeof(Ceed_Magma), &data); CeedChk(ierr); 56 ierr = CeedSetData(ceed, data); CeedChk(ierr); 57 58 // kernel selection 59 data->basis_kernel_mode = MAGMA_KERNEL_DIM_SPECIFIC; 60 61 // kernel max threads per thread-block 62 data->maxthreads[0] = 128; // for 1D kernels 63 data->maxthreads[1] = 128; // for 2D kernels 64 data->maxthreads[2] = 64; // for 3D kernels 65 66 // create a queue that uses the null stream 67 magma_getdevice( &(data->device) ); 68 #ifdef HAVE_HIP 69 magma_queue_create_from_hip(data->device, NULL, NULL, NULL, &(data->queue)); 70 #else 71 magma_queue_create_from_cuda(data->device, NULL, NULL, NULL, &(data->queue)); 72 #endif 73 74 ierr = CeedSetBackendFunction(ceed, "Ceed", ceed, "ElemRestrictionCreate", 75 CeedElemRestrictionCreate_Magma); CeedChk(ierr); 76 ierr = CeedSetBackendFunction(ceed, "Ceed", ceed, 77 "ElemRestrictionCreateBlocked", 78 CeedElemRestrictionCreateBlocked_Magma); CeedChk(ierr); 79 ierr = CeedSetBackendFunction(ceed, "Ceed", ceed, "BasisCreateTensorH1", 80 CeedBasisCreateTensorH1_Magma); CeedChk(ierr); 81 ierr = CeedSetBackendFunction(ceed, "Ceed", ceed, "BasisCreateH1", 82 CeedBasisCreateH1_Magma); CeedChk(ierr); 83 ierr = CeedSetBackendFunction(ceed, "Ceed", ceed, "Destroy", 84 CeedDestroy_Magma); CeedChk(ierr); 85 return 0; 86 } 87 88 CEED_INTERN int CeedRegister_Magma(void) { 89 #ifdef HAVE_HIP 90 return CeedRegister("/gpu/hip/magma", CeedInit_Magma, 120); 91 #else 92 return CeedRegister("/gpu/cuda/magma", CeedInit_Magma, 120); 93 #endif 94 } 95