1 // Copyright (c) 2017-2022, Lawrence Livermore National Security, LLC and other CEED contributors. 2 // All Rights Reserved. See the top-level LICENSE and NOTICE files for details. 3 // 4 // SPDX-License-Identifier: BSD-2-Clause 5 // 6 // This file is part of CEED: http://github.com/ceed 7 8 #include "ceed-cuda-compile.h" 9 10 #include <ceed.h> 11 #include <ceed/backend.h> 12 #include <ceed/jit-tools.h> 13 #include <cuda_runtime.h> 14 #include <nvrtc.h> 15 #include <stdarg.h> 16 #include <string.h> 17 18 #include <sstream> 19 20 #include "ceed-cuda-common.h" 21 22 #define CeedChk_Nvrtc(ceed, x) \ 23 do { \ 24 nvrtcResult result = static_cast<nvrtcResult>(x); \ 25 if (result != NVRTC_SUCCESS) return CeedError((ceed), CEED_ERROR_BACKEND, nvrtcGetErrorString(result)); \ 26 } while (0) 27 28 #define CeedCallNvrtc(ceed, ...) \ 29 do { \ 30 int ierr_q_ = __VA_ARGS__; \ 31 CeedChk_Nvrtc(ceed, ierr_q_); \ 32 } while (0) 33 34 //------------------------------------------------------------------------------ 35 // Compile CUDA kernel 36 //------------------------------------------------------------------------------ 37 int CeedCompile_Cuda(Ceed ceed, const char *source, CUmodule *module, const CeedInt num_defines, ...) { 38 size_t ptx_size; 39 char *jit_defs_path, *jit_defs_source, *ptx; 40 const int num_opts = 3; 41 const char *opts[num_opts]; 42 nvrtcProgram prog; 43 struct cudaDeviceProp prop; 44 Ceed_Cuda *ceed_data; 45 46 cudaFree(0); // Make sure a Context exists for nvrtc 47 48 std::ostringstream code; 49 50 // Get kernel specific options, such as kernel constants 51 if (num_defines > 0) { 52 va_list args; 53 va_start(args, num_defines); 54 char *name; 55 int val; 56 57 for (int i = 0; i < num_defines; i++) { 58 name = va_arg(args, char *); 59 val = va_arg(args, int); 60 code << "#define " << name << " " << val << "\n"; 61 } 62 va_end(args); 63 } 64 65 // Standard libCEED definitions for CUDA backends 66 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/cuda/cuda-jit.h", &jit_defs_path)); 67 CeedCallBackend(CeedLoadSourceToBuffer(ceed, jit_defs_path, &jit_defs_source)); 68 code << jit_defs_source; 69 code << "\n\n"; 70 CeedCallBackend(CeedFree(&jit_defs_path)); 71 CeedCallBackend(CeedFree(&jit_defs_source)); 72 73 // Non-macro options 74 opts[0] = "-default-device"; 75 CeedCallBackend(CeedGetData(ceed, &ceed_data)); 76 CeedCallCuda(ceed, cudaGetDeviceProperties(&prop, ceed_data->device_id)); 77 std::string arch_arg = "-arch=compute_" + std::to_string(prop.major) + std::to_string(prop.minor); 78 opts[1] = arch_arg.c_str(); 79 opts[2] = "-Dint32_t=int"; 80 81 // Add string source argument provided in call 82 code << source; 83 84 // Create Program 85 CeedCallNvrtc(ceed, nvrtcCreateProgram(&prog, code.str().c_str(), NULL, 0, NULL, NULL)); 86 87 // Compile kernel 88 nvrtcResult result = nvrtcCompileProgram(prog, num_opts, opts); 89 90 if (result != NVRTC_SUCCESS) { 91 char *log; 92 size_t log_size; 93 94 CeedCallNvrtc(ceed, nvrtcGetProgramLogSize(prog, &log_size)); 95 CeedCallBackend(CeedMalloc(log_size, &log)); 96 CeedCallNvrtc(ceed, nvrtcGetProgramLog(prog, log)); 97 return CeedError(ceed, CEED_ERROR_BACKEND, "%s\n%s", nvrtcGetErrorString(result), log); 98 } 99 100 CeedCallNvrtc(ceed, nvrtcGetPTXSize(prog, &ptx_size)); 101 CeedCallBackend(CeedMalloc(ptx_size, &ptx)); 102 CeedCallNvrtc(ceed, nvrtcGetPTX(prog, ptx)); 103 CeedCallNvrtc(ceed, nvrtcDestroyProgram(&prog)); 104 105 CeedCallCuda(ceed, cuModuleLoadData(module, ptx)); 106 CeedCallBackend(CeedFree(&ptx)); 107 return CEED_ERROR_SUCCESS; 108 } 109 110 //------------------------------------------------------------------------------ 111 // Get CUDA kernel 112 //------------------------------------------------------------------------------ 113 int CeedGetKernel_Cuda(Ceed ceed, CUmodule module, const char *name, CUfunction *kernel) { 114 CeedCallCuda(ceed, cuModuleGetFunction(kernel, module, name)); 115 return CEED_ERROR_SUCCESS; 116 } 117 118 //------------------------------------------------------------------------------ 119 // Run CUDA kernel with block size selected automatically based on the kernel 120 // (which may use enough registers to require a smaller block size than the 121 // hardware is capable) 122 //------------------------------------------------------------------------------ 123 int CeedRunKernelAutoblockCuda(Ceed ceed, CUfunction kernel, size_t points, void **args) { 124 int min_grid_size, max_block_size; 125 126 CeedCallCuda(ceed, cuOccupancyMaxPotentialBlockSize(&min_grid_size, &max_block_size, kernel, NULL, 0, 0x10000)); 127 CeedCallBackend(CeedRunKernel_Cuda(ceed, kernel, CeedDivUpInt(points, max_block_size), max_block_size, args)); 128 return CEED_ERROR_SUCCESS; 129 } 130 131 //------------------------------------------------------------------------------ 132 // Run CUDA kernel 133 //------------------------------------------------------------------------------ 134 int CeedRunKernel_Cuda(Ceed ceed, CUfunction kernel, const int grid_size, const int block_size, void **args) { 135 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, kernel, grid_size, block_size, 1, 1, 0, args)); 136 return CEED_ERROR_SUCCESS; 137 } 138 139 //------------------------------------------------------------------------------ 140 // Run CUDA kernel for spatial dimension 141 //------------------------------------------------------------------------------ 142 int CeedRunKernelDim_Cuda(Ceed ceed, CUfunction kernel, const int grid_size, const int block_size_x, const int block_size_y, const int block_size_z, 143 void **args) { 144 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, kernel, grid_size, block_size_x, block_size_y, block_size_z, 0, args)); 145 return CEED_ERROR_SUCCESS; 146 } 147 148 //------------------------------------------------------------------------------ 149 // Run CUDA kernel for spatial dimension with shared memory 150 //------------------------------------------------------------------------------ 151 int CeedRunKernelDimShared_Cuda(Ceed ceed, CUfunction kernel, const int grid_size, const int block_size_x, const int block_size_y, 152 const int block_size_z, const int shared_mem_size, void **args) { 153 #if CUDA_VERSION >= 9000 154 cuFuncSetAttribute(kernel, CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES, shared_mem_size); 155 #endif 156 CUresult result = cuLaunchKernel(kernel, grid_size, 1, 1, block_size_x, block_size_y, block_size_z, shared_mem_size, NULL, args, NULL); 157 158 if (result == CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES) { 159 int max_threads_per_block, shared_size_bytes, num_regs; 160 161 cuFuncGetAttribute(&max_threads_per_block, CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK, kernel); 162 cuFuncGetAttribute(&shared_size_bytes, CU_FUNC_ATTRIBUTE_SHARED_SIZE_BYTES, kernel); 163 cuFuncGetAttribute(&num_regs, CU_FUNC_ATTRIBUTE_NUM_REGS, kernel); 164 return CeedError(ceed, CEED_ERROR_BACKEND, 165 "CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES: max_threads_per_block %d on block size (%d,%d,%d), shared_size %d, num_regs %d", 166 max_threads_per_block, block_size_x, block_size_y, block_size_z, shared_size_bytes, num_regs); 167 } else CeedChk_Cu(ceed, result); 168 return CEED_ERROR_SUCCESS; 169 } 170 171 //------------------------------------------------------------------------------ 172