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-hip-compile.h" 9 10 #include <ceed/backend.h> 11 #include <ceed/ceed.h> 12 #include <ceed/jit-tools.h> 13 #include <hip/hiprtc.h> 14 #include <stdarg.h> 15 #include <string.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 CeedCompileHip(Ceed ceed, const char *source, hipModule_t *module, const CeedInt num_defines, ...) { 37 hipFree(0); // Make sure a Context exists for hiprtc 38 hiprtcProgram prog; 39 40 std::ostringstream code; 41 42 // Add hip runtime include statement for generation if runtime < 40400000 43 // (implies ROCm < 4.5) 44 int runtime_version; 45 CeedCallHip(ceed, hipRuntimeGetVersion(&runtime_version)); 46 if (runtime_version < 40400000) { 47 code << "\n#include <hip/hip_runtime.h>\n"; 48 } 49 // With ROCm 4.5, need to include these definitions specifically for hiprtc 50 // (but cannot include the runtime header) 51 else { 52 code << "#include <stddef.h>\n"; 53 code << "#define __forceinline__ inline __attribute__((always_inline))\n"; 54 code << "#define HIP_DYNAMIC_SHARED(type, var) extern __shared__ type var[];\n"; 55 } 56 57 // Kernel specific options, such as kernel constants 58 if (num_defines > 0) { 59 va_list args; 60 va_start(args, num_defines); 61 char *name; 62 int val; 63 for (int i = 0; i < num_defines; i++) { 64 name = va_arg(args, char *); 65 val = va_arg(args, int); 66 code << "#define " << name << " " << val << "\n"; 67 } 68 va_end(args); 69 } 70 71 // Standard libCEED definitions for HIP backends 72 char *jit_defs_path, *jit_defs_source; 73 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/hip/hip-jit.h", &jit_defs_path)); 74 CeedCallBackend(CeedLoadSourceToBuffer(ceed, jit_defs_path, &jit_defs_source)); 75 code << jit_defs_source; 76 code << "\n\n"; 77 CeedCallBackend(CeedFree(&jit_defs_path)); 78 CeedCallBackend(CeedFree(&jit_defs_source)); 79 80 // Non-macro options 81 const int num_opts = 3; 82 const char *opts[num_opts]; 83 opts[0] = "-default-device"; 84 struct hipDeviceProp_t prop; 85 Ceed_Hip *ceed_data; 86 CeedCallBackend(CeedGetData(ceed, (void **)&ceed_data)); 87 CeedCallHip(ceed, hipGetDeviceProperties(&prop, ceed_data->device_id)); 88 std::string arch_arg = "--gpu-architecture=" + std::string(prop.gcnArchName); 89 opts[1] = arch_arg.c_str(); 90 opts[2] = "-munsafe-fp-atomics"; 91 92 // Add string source argument provided in call 93 code << source; 94 95 // Create Program 96 CeedCallHiprtc(ceed, hiprtcCreateProgram(&prog, code.str().c_str(), NULL, 0, NULL, NULL)); 97 98 // Compile kernel 99 hiprtcResult result = hiprtcCompileProgram(prog, num_opts, opts); 100 if (result != HIPRTC_SUCCESS) { 101 size_t log_size; 102 CeedChk_hiprtc(ceed, hiprtcGetProgramLogSize(prog, &log_size)); 103 char *log; 104 CeedCallBackend(CeedMalloc(log_size, &log)); 105 CeedCallHiprtc(ceed, hiprtcGetProgramLog(prog, log)); 106 return CeedError(ceed, CEED_ERROR_BACKEND, "%s\n%s", hiprtcGetErrorString(result), log); 107 } 108 109 size_t ptx_size; 110 CeedCallHiprtc(ceed, hiprtcGetCodeSize(prog, &ptx_size)); 111 char *ptx; 112 CeedCallBackend(CeedMalloc(ptx_size, &ptx)); 113 CeedCallHiprtc(ceed, hiprtcGetCode(prog, ptx)); 114 CeedCallHiprtc(ceed, hiprtcDestroyProgram(&prog)); 115 116 CeedCallHip(ceed, hipModuleLoadData(module, ptx)); 117 CeedCallBackend(CeedFree(&ptx)); 118 119 return CEED_ERROR_SUCCESS; 120 } 121 122 //------------------------------------------------------------------------------ 123 // Get HIP kernel 124 //------------------------------------------------------------------------------ 125 int CeedGetKernelHip(Ceed ceed, hipModule_t module, const char *name, hipFunction_t *kernel) { 126 CeedCallHip(ceed, hipModuleGetFunction(kernel, module, name)); 127 return CEED_ERROR_SUCCESS; 128 } 129 130 //------------------------------------------------------------------------------ 131 // Run HIP kernel 132 //------------------------------------------------------------------------------ 133 int CeedRunKernelHip(Ceed ceed, hipFunction_t kernel, const int grid_size, const int block_size, void **args) { 134 CeedCallHip(ceed, hipModuleLaunchKernel(kernel, grid_size, 1, 1, block_size, 1, 1, 0, NULL, args, NULL)); 135 return CEED_ERROR_SUCCESS; 136 } 137 138 //------------------------------------------------------------------------------ 139 // Run HIP kernel for spatial dimension 140 //------------------------------------------------------------------------------ 141 int CeedRunKernelDimHip(Ceed ceed, hipFunction_t kernel, const int grid_size, const int block_size_x, const int block_size_y, const int block_size_z, 142 void **args) { 143 CeedCallHip(ceed, hipModuleLaunchKernel(kernel, grid_size, 1, 1, block_size_x, block_size_y, block_size_z, 0, NULL, args, NULL)); 144 return CEED_ERROR_SUCCESS; 145 } 146 147 //------------------------------------------------------------------------------ 148 // Run HIP kernel for spatial dimension with shared memory 149 //------------------------------------------------------------------------------ 150 int CeedRunKernelDimSharedHip(Ceed ceed, hipFunction_t kernel, const int grid_size, const int block_size_x, const int block_size_y, 151 const int block_size_z, const int shared_mem_size, void **args) { 152 CeedCallHip(ceed, hipModuleLaunchKernel(kernel, grid_size, 1, 1, block_size_x, block_size_y, block_size_z, shared_mem_size, NULL, args, NULL)); 153 return CEED_ERROR_SUCCESS; 154 } 155