1 // Copyright (c) 2017-2024, 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-hip-compile.h" 9 10 #include <ceed.h> 11 #include <ceed/backend.h> 12 #include <ceed/jit-tools.h> 13 #include <stdarg.h> 14 #include <string.h> 15 #include <hip/hiprtc.h> 16 17 #include <sstream> 18 19 #include "ceed-hip-common.h" 20 21 #define CeedChk_hiprtc(ceed, x) \ 22 do { \ 23 hiprtcResult result = static_cast<hiprtcResult>(x); \ 24 if (result != HIPRTC_SUCCESS) return CeedError((ceed), CEED_ERROR_BACKEND, hiprtcGetErrorString(result)); \ 25 } while (0) 26 27 #define CeedCallHiprtc(ceed, ...) \ 28 do { \ 29 int ierr_q_ = __VA_ARGS__; \ 30 CeedChk_hiprtc(ceed, ierr_q_); \ 31 } while (0) 32 33 //------------------------------------------------------------------------------ 34 // Compile HIP kernel 35 //------------------------------------------------------------------------------ 36 int CeedCompile_Hip(Ceed ceed, const char *source, hipModule_t *module, const CeedInt num_defines, ...) { 37 size_t ptx_size; 38 char *jit_defs_source, *ptx; 39 const char *jit_defs_path; 40 const int num_opts = 3; 41 CeedInt num_jit_source_dirs = 0; 42 const char **opts; 43 int runtime_version; 44 hiprtcProgram prog; 45 struct hipDeviceProp_t prop; 46 Ceed_Hip *ceed_data; 47 48 hipFree(0); // Make sure a Context exists for hiprtc 49 50 std::ostringstream code; 51 52 // Add hip runtime include statement for generation if runtime < 40400000 (implies ROCm < 4.5) 53 CeedCallHip(ceed, hipRuntimeGetVersion(&runtime_version)); 54 if (runtime_version < 40400000) { 55 code << "\n#include <hip/hip_runtime.h>\n"; 56 } 57 // With ROCm 4.5, need to include these definitions specifically for hiprtc (but cannot include the runtime header) 58 else { 59 code << "#include <stddef.h>\n"; 60 code << "#define __forceinline__ inline __attribute__((always_inline))\n"; 61 code << "#define HIP_DYNAMIC_SHARED(type, var) extern __shared__ type var[];\n"; 62 } 63 64 // Kernel specific options, such as kernel constants 65 if (num_defines > 0) { 66 va_list args; 67 va_start(args, num_defines); 68 char *name; 69 int val; 70 71 for (int i = 0; i < num_defines; i++) { 72 name = va_arg(args, char *); 73 val = va_arg(args, int); 74 code << "#define " << name << " " << val << "\n"; 75 } 76 va_end(args); 77 } 78 79 // Standard libCEED definitions for HIP backends 80 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/hip/hip-jit.h", &jit_defs_path)); 81 CeedCallBackend(CeedLoadSourceToBuffer(ceed, jit_defs_path, &jit_defs_source)); 82 code << jit_defs_source; 83 code << "\n\n"; 84 CeedCallBackend(CeedFree(&jit_defs_path)); 85 CeedCallBackend(CeedFree(&jit_defs_source)); 86 87 // Non-macro options 88 CeedCallBackend(CeedCalloc(num_opts, &opts)); 89 opts[0] = "-default-device"; 90 CeedCallBackend(CeedGetData(ceed, (void **)&ceed_data)); 91 CeedCallHip(ceed, hipGetDeviceProperties(&prop, ceed_data->device_id)); 92 std::string arch_arg = "--gpu-architecture=" + std::string(prop.gcnArchName); 93 opts[1] = arch_arg.c_str(); 94 opts[2] = "-munsafe-fp-atomics"; 95 { 96 const char **jit_source_dirs; 97 98 CeedCallBackend(CeedGetJitSourceRoots(ceed, &num_jit_source_dirs, &jit_source_dirs)); 99 CeedCallBackend(CeedRealloc(num_opts + num_jit_source_dirs, &opts)); 100 for (CeedInt i = 0; i < num_jit_source_dirs; i++) { 101 std::ostringstream include_dirs_arg; 102 103 include_dirs_arg << "-I" << jit_source_dirs[i]; 104 CeedCallBackend(CeedStringAllocCopy(include_dirs_arg.str().c_str(), (char **)&opts[num_opts + i])); 105 } 106 CeedCallBackend(CeedRestoreJitSourceRoots(ceed, &jit_source_dirs)); 107 } 108 109 // Add string source argument provided in call 110 code << source; 111 112 // Create Program 113 CeedCallHiprtc(ceed, hiprtcCreateProgram(&prog, code.str().c_str(), NULL, 0, NULL, NULL)); 114 115 // Compile kernel 116 hiprtcResult result = hiprtcCompileProgram(prog, num_opts + num_jit_source_dirs, opts); 117 118 for (CeedInt i = 0; i < num_jit_source_dirs; i++) { 119 CeedCallBackend(CeedFree(&opts[num_opts + i])); 120 } 121 CeedCallBackend(CeedFree(&opts)); 122 if (result != HIPRTC_SUCCESS) { 123 size_t log_size; 124 char *log; 125 126 CeedDebug256(ceed, CEED_DEBUG_COLOR_ERROR, "---------- CEED JIT SOURCE FAILED TO COMPILE ----------\n"); 127 CeedDebug(ceed, "Source:\n%s\n", code.str().c_str()); 128 CeedDebug256(ceed, CEED_DEBUG_COLOR_ERROR, "---------- CEED JIT SOURCE FAILED TO COMPILE ----------\n"); 129 CeedChk_hiprtc(ceed, hiprtcGetProgramLogSize(prog, &log_size)); 130 CeedCallBackend(CeedMalloc(log_size, &log)); 131 CeedCallHiprtc(ceed, hiprtcGetProgramLog(prog, log)); 132 return CeedError(ceed, CEED_ERROR_BACKEND, "%s\n%s", hiprtcGetErrorString(result), log); 133 } 134 135 CeedCallHiprtc(ceed, hiprtcGetCodeSize(prog, &ptx_size)); 136 CeedCallBackend(CeedMalloc(ptx_size, &ptx)); 137 CeedCallHiprtc(ceed, hiprtcGetCode(prog, ptx)); 138 CeedCallHiprtc(ceed, hiprtcDestroyProgram(&prog)); 139 140 CeedCallHip(ceed, hipModuleLoadData(module, ptx)); 141 CeedCallBackend(CeedFree(&ptx)); 142 return CEED_ERROR_SUCCESS; 143 } 144 145 //------------------------------------------------------------------------------ 146 // Get HIP kernel 147 //------------------------------------------------------------------------------ 148 int CeedGetKernel_Hip(Ceed ceed, hipModule_t module, const char *name, hipFunction_t *kernel) { 149 CeedCallHip(ceed, hipModuleGetFunction(kernel, module, name)); 150 return CEED_ERROR_SUCCESS; 151 } 152 153 //------------------------------------------------------------------------------ 154 // Run HIP kernel 155 //------------------------------------------------------------------------------ 156 int CeedRunKernel_Hip(Ceed ceed, hipFunction_t kernel, const int grid_size, const int block_size, void **args) { 157 CeedCallHip(ceed, hipModuleLaunchKernel(kernel, grid_size, 1, 1, block_size, 1, 1, 0, NULL, args, NULL)); 158 return CEED_ERROR_SUCCESS; 159 } 160 161 //------------------------------------------------------------------------------ 162 // Run HIP kernel for spatial dimension 163 //------------------------------------------------------------------------------ 164 int CeedRunKernelDim_Hip(Ceed ceed, hipFunction_t kernel, const int grid_size, const int block_size_x, const int block_size_y, const int block_size_z, 165 void **args) { 166 CeedCallHip(ceed, hipModuleLaunchKernel(kernel, grid_size, 1, 1, block_size_x, block_size_y, block_size_z, 0, NULL, args, NULL)); 167 return CEED_ERROR_SUCCESS; 168 } 169 170 //------------------------------------------------------------------------------ 171 // Run HIP kernel for spatial dimension with shared memory 172 //------------------------------------------------------------------------------ 173 int CeedRunKernelDimShared_Hip(Ceed ceed, hipFunction_t kernel, const int grid_size, const int block_size_x, const int block_size_y, 174 const int block_size_z, const int shared_mem_size, void **args) { 175 CeedCallHip(ceed, hipModuleLaunchKernel(kernel, grid_size, 1, 1, block_size_x, block_size_y, block_size_z, shared_mem_size, NULL, args, NULL)); 176 return CEED_ERROR_SUCCESS; 177 } 178 179 //------------------------------------------------------------------------------ 180