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 cudaFree(0); // Make sure a Context exists for nvrtc 39 nvrtcProgram prog; 40 41 std::ostringstream code; 42 43 // Get kernel specific options, such as kernel constants 44 if (num_defines > 0) { 45 va_list args; 46 va_start(args, num_defines); 47 char *name; 48 int val; 49 for (int i = 0; i < num_defines; i++) { 50 name = va_arg(args, char *); 51 val = va_arg(args, int); 52 code << "#define " << name << " " << val << "\n"; 53 } 54 va_end(args); 55 } 56 57 // Standard libCEED definitions for CUDA backends 58 char *jit_defs_path, *jit_defs_source; 59 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/cuda/cuda-jit.h", &jit_defs_path)); 60 CeedCallBackend(CeedLoadSourceToBuffer(ceed, jit_defs_path, &jit_defs_source)); 61 code << jit_defs_source; 62 code << "\n\n"; 63 CeedCallBackend(CeedFree(&jit_defs_path)); 64 CeedCallBackend(CeedFree(&jit_defs_source)); 65 66 // Non-macro options 67 const int num_opts = 3; 68 const char *opts[num_opts]; 69 opts[0] = "-default-device"; 70 struct cudaDeviceProp prop; 71 Ceed_Cuda *ceed_data; 72 CeedCallBackend(CeedGetData(ceed, &ceed_data)); 73 CeedCallCuda(ceed, cudaGetDeviceProperties(&prop, ceed_data->device_id)); 74 std::string arch_arg = "-arch=compute_" + std::to_string(prop.major) + std::to_string(prop.minor); 75 opts[1] = arch_arg.c_str(); 76 opts[2] = "-Dint32_t=int"; 77 78 // Add string source argument provided in call 79 code << source; 80 81 // Create Program 82 CeedCallNvrtc(ceed, nvrtcCreateProgram(&prog, code.str().c_str(), NULL, 0, NULL, NULL)); 83 84 // Compile kernel 85 nvrtcResult result = nvrtcCompileProgram(prog, num_opts, opts); 86 if (result != NVRTC_SUCCESS) { 87 size_t log_size; 88 CeedCallNvrtc(ceed, nvrtcGetProgramLogSize(prog, &log_size)); 89 char *log; 90 CeedCallBackend(CeedMalloc(log_size, &log)); 91 CeedCallNvrtc(ceed, nvrtcGetProgramLog(prog, log)); 92 return CeedError(ceed, CEED_ERROR_BACKEND, "%s\n%s", nvrtcGetErrorString(result), log); 93 } 94 95 size_t ptx_size; 96 CeedCallNvrtc(ceed, nvrtcGetPTXSize(prog, &ptx_size)); 97 char *ptx; 98 CeedCallBackend(CeedMalloc(ptx_size, &ptx)); 99 CeedCallNvrtc(ceed, nvrtcGetPTX(prog, ptx)); 100 CeedCallNvrtc(ceed, nvrtcDestroyProgram(&prog)); 101 102 CeedCallCuda(ceed, cuModuleLoadData(module, ptx)); 103 CeedCallBackend(CeedFree(&ptx)); 104 return CEED_ERROR_SUCCESS; 105 } 106 107 //------------------------------------------------------------------------------ 108 // Get CUDA kernel 109 //------------------------------------------------------------------------------ 110 int CeedGetKernel_Cuda(Ceed ceed, CUmodule module, const char *name, CUfunction *kernel) { 111 CeedCallCuda(ceed, cuModuleGetFunction(kernel, module, name)); 112 return CEED_ERROR_SUCCESS; 113 } 114 115 //------------------------------------------------------------------------------ 116 // Run CUDA kernel with block size selected automatically based on the kernel 117 // (which may use enough registers to require a smaller block size than the 118 // hardware is capable) 119 //------------------------------------------------------------------------------ 120 int CeedRunKernelAutoblockCuda(Ceed ceed, CUfunction kernel, size_t points, void **args) { 121 int min_grid_size, max_block_size; 122 CeedCallCuda(ceed, cuOccupancyMaxPotentialBlockSize(&min_grid_size, &max_block_size, kernel, NULL, 0, 0x10000)); 123 CeedCallBackend(CeedRunKernel_Cuda(ceed, kernel, CeedDivUpInt(points, max_block_size), max_block_size, args)); 124 return 0; 125 } 126 127 //------------------------------------------------------------------------------ 128 // Run CUDA kernel 129 //------------------------------------------------------------------------------ 130 int CeedRunKernel_Cuda(Ceed ceed, CUfunction kernel, const int grid_size, const int block_size, void **args) { 131 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, kernel, grid_size, block_size, 1, 1, 0, args)); 132 return CEED_ERROR_SUCCESS; 133 } 134 135 //------------------------------------------------------------------------------ 136 // Run CUDA kernel for spatial dimension 137 //------------------------------------------------------------------------------ 138 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, 139 void **args) { 140 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, kernel, grid_size, block_size_x, block_size_y, block_size_z, 0, args)); 141 return CEED_ERROR_SUCCESS; 142 } 143 144 //------------------------------------------------------------------------------ 145 // Run CUDA kernel for spatial dimension with shared memory 146 //------------------------------------------------------------------------------ 147 int CeedRunKernelDimShared_Cuda(Ceed ceed, CUfunction kernel, const int grid_size, const int block_size_x, const int block_size_y, 148 const int block_size_z, const int shared_mem_size, void **args) { 149 #if CUDA_VERSION >= 9000 150 cuFuncSetAttribute(kernel, CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES, shared_mem_size); 151 #endif 152 CUresult result = cuLaunchKernel(kernel, grid_size, 1, 1, block_size_x, block_size_y, block_size_z, shared_mem_size, NULL, args, NULL); 153 if (result == CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES) { 154 int max_threads_per_block, shared_size_bytes, num_regs; 155 cuFuncGetAttribute(&max_threads_per_block, CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK, kernel); 156 cuFuncGetAttribute(&shared_size_bytes, CU_FUNC_ATTRIBUTE_SHARED_SIZE_BYTES, kernel); 157 cuFuncGetAttribute(&num_regs, CU_FUNC_ATTRIBUTE_NUM_REGS, kernel); 158 return CeedError(ceed, CEED_ERROR_BACKEND, 159 "CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES: max_threads_per_block %d on block size (%d,%d,%d), shared_size %d, num_regs %d", 160 max_threads_per_block, block_size_x, block_size_y, block_size_z, shared_size_bytes, num_regs); 161 } else CeedChk_Cu(ceed, result); 162 return CEED_ERROR_SUCCESS; 163 } 164 165 //------------------------------------------------------------------------------ 166