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/ceed.h> 18 #include <ceed/backend.h> 19 #include <ceed/jit-tools.h> 20 #include <iostream> 21 #include <sstream> 22 #include <string.h> 23 #include "ceed-cuda-ref.h" 24 #include "../cuda/ceed-cuda-compile.h" 25 26 //------------------------------------------------------------------------------ 27 // Build QFunction kernel 28 //------------------------------------------------------------------------------ 29 extern "C" int CeedCudaBuildQFunction(CeedQFunction qf) { 30 CeedInt ierr; 31 using std::ostringstream; 32 using std::string; 33 Ceed ceed; 34 CeedQFunctionGetCeed(qf, &ceed); 35 CeedQFunction_Cuda *data; 36 ierr = CeedQFunctionGetData(qf, (void **)&data); CeedChkBackend(ierr); 37 38 // QFunction is built 39 if (data->QFunction) 40 return CEED_ERROR_SUCCESS; 41 42 if (!data->qfunction_source) 43 // LCOV_EXCL_START 44 return CeedError(ceed, CEED_ERROR_BACKEND, 45 "No QFunction source or CUfunction provided."); 46 // LCOV_EXCL_STOP 47 48 // QFunction kernel generation 49 CeedInt num_input_fields, num_output_fields, size; 50 CeedQFunctionField *input_fields, *output_fields; 51 ierr = CeedQFunctionGetFields(qf, &num_input_fields, &input_fields, 52 &num_output_fields, &output_fields); 53 CeedChkBackend(ierr); 54 55 // Build strings for final kernel 56 char *read_write_kernel_path, *read_write_kernel_source; 57 ierr = CeedPathConcatenate(ceed, __FILE__, "kernels/cuda-ref-qfunction.h", 58 &read_write_kernel_path); CeedChkBackend(ierr); 59 ierr = CeedLoadSourceToBuffer(ceed, read_write_kernel_path, &read_write_kernel_source); 60 CeedChkBackend(ierr); 61 string qfunction_source(data->qfunction_source); 62 string qfunction_name(data->qfunction_name); 63 string read_write(read_write_kernel_source); 64 string kernel_name = "CeedKernel_Cuda_ref_" + qfunction_name; 65 ostringstream code; 66 67 // Defintions 68 code << "\n#define CEED_QFUNCTION(name) inline __device__ int name\n"; 69 code << "#define CEED_QFUNCTION_HELPER inline __device__\n"; 70 code << "#define CeedPragmaSIMD\n"; 71 code << "#define CEED_ERROR_SUCCESS 0\n"; 72 code << "#define CEED_Q_VLA 1\n\n"; 73 code << "typedef struct { const CeedScalar* inputs[16]; CeedScalar* outputs[16]; } Fields_Cuda;\n"; 74 code << read_write; 75 code << qfunction_source; 76 code << "extern \"C\" __global__ void " << kernel_name << "(void *ctx, CeedInt Q, Fields_Cuda fields) {\n"; 77 78 // Inputs 79 for (CeedInt i = 0; i < num_input_fields; i++) { 80 code << " // Input field " << i << "\n"; 81 ierr = CeedQFunctionFieldGetSize(input_fields[i], &size); CeedChkBackend(ierr); 82 code << " const CeedInt size_in_" << i << " = "<<size<<";\n"; 83 code << " CeedScalar r_q" << i << "[size_in_" << i << "];\n"; 84 } 85 code << "\n"; 86 87 // Outputs 88 for (CeedInt i = 0; i < num_output_fields; i++) { 89 code << " // Output field " << i << "\n"; 90 ierr = CeedQFunctionFieldGetSize(output_fields[i], &size); CeedChkBackend(ierr); 91 code << " const CeedInt size_out_" << i << " = " << size << ";\n"; 92 code << " CeedScalar r_qq" << i << "[size_out_" << i << "];\n"; 93 } 94 code << "\n"; 95 96 // Setup input/output arrays 97 code << " const CeedScalar* in[" << num_input_fields << "];\n"; 98 for (CeedInt i = 0; i < num_input_fields; i++) { 99 code << " in[" << i << "] = r_q" << i << ";\n"; 100 } 101 code << " CeedScalar* out[" << num_output_fields << "];\n"; 102 for (CeedInt i = 0; i < num_output_fields; i++) { 103 code << " out[" << i << "] = r_qq" << i << ";\n"; 104 } 105 code << "\n"; 106 107 // Loop over quadrature points 108 code << " for (CeedInt q = blockIdx.x * blockDim.x + threadIdx.x; q < Q; q += blockDim.x * gridDim.x) {\n"; 109 110 // Load inputs 111 for (CeedInt i = 0; i < num_input_fields; i++) { 112 code << " // Input field " << i << "\n"; 113 code << " readQuads<size_in_" << i << ">(q, Q, fields.inputs[" << i << "], r_q" << i << ");\n"; 114 } 115 // QFunction 116 code << " // QFunction\n"; 117 code << " " << qfunction_name << "(ctx, 1, in, out);\n"; 118 119 // Write outputs 120 for (CeedInt i = 0; i < num_output_fields; i++) { 121 code << " // Output field " << i << "\n"; 122 code << " writeQuads<size_out_" << i << ">(q, Q, r_qq" << i << ", fields.outputs[" << i << "]);\n"; 123 } 124 code << " }\n"; 125 code << "}\n"; 126 127 // View kernel for debugging 128 CeedDebug256(ceed, 1, "Generated QFunction Kernels:\n"); 129 CeedDebug(ceed, code.str().c_str()); 130 131 // Compile kernel 132 ierr = CeedCompileCuda(ceed, code.str().c_str(), &data->module, 0); 133 CeedChkBackend(ierr); 134 ierr = CeedGetKernelCuda(ceed, data->module, kernel_name.c_str(), &data->QFunction); 135 CeedChkBackend(ierr); 136 137 // Cleanup 138 ierr = CeedFree(&data->qfunction_source); CeedChkBackend(ierr); 139 ierr = CeedFree(&read_write_kernel_path); CeedChkBackend(ierr); 140 ierr = CeedFree(&read_write_kernel_source); CeedChkBackend(ierr); 141 142 return CEED_ERROR_SUCCESS; 143 } 144 //------------------------------------------------------------------------------ 145