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.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 const char *opts[num_opts]; 42 int runtime_version; 43 hiprtcProgram prog; 44 struct hipDeviceProp_t prop; 45 Ceed_Hip *ceed_data; 46 47 hipFree(0); // Make sure a Context exists for hiprtc 48 49 std::ostringstream code; 50 51 // Add hip runtime include statement for generation if runtime < 40400000 (implies ROCm < 4.5) 52 CeedCallHip(ceed, hipRuntimeGetVersion(&runtime_version)); 53 if (runtime_version < 40400000) { 54 code << "\n#include <hip/hip_runtime.h>\n"; 55 } 56 // With ROCm 4.5, need to include these definitions specifically for hiprtc (but cannot include the runtime header) 57 else { 58 code << "#include <stddef.h>\n"; 59 code << "#define __forceinline__ inline __attribute__((always_inline))\n"; 60 code << "#define HIP_DYNAMIC_SHARED(type, var) extern __shared__ type var[];\n"; 61 } 62 63 // Kernel specific options, such as kernel constants 64 if (num_defines > 0) { 65 va_list args; 66 va_start(args, num_defines); 67 char *name; 68 int val; 69 70 for (int i = 0; i < num_defines; i++) { 71 name = va_arg(args, char *); 72 val = va_arg(args, int); 73 code << "#define " << name << " " << val << "\n"; 74 } 75 va_end(args); 76 } 77 78 // Standard libCEED definitions for HIP backends 79 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/hip/hip-jit.h", &jit_defs_path)); 80 CeedCallBackend(CeedLoadSourceToBuffer(ceed, jit_defs_path, &jit_defs_source)); 81 code << jit_defs_source; 82 code << "\n\n"; 83 CeedCallBackend(CeedFree(&jit_defs_path)); 84 CeedCallBackend(CeedFree(&jit_defs_source)); 85 86 // Non-macro options 87 opts[0] = "-default-device"; 88 CeedCallBackend(CeedGetData(ceed, (void **)&ceed_data)); 89 CeedCallHip(ceed, hipGetDeviceProperties(&prop, ceed_data->device_id)); 90 std::string arch_arg = "--gpu-architecture=" + std::string(prop.gcnArchName); 91 opts[1] = arch_arg.c_str(); 92 opts[2] = "-munsafe-fp-atomics"; 93 94 // Add string source argument provided in call 95 code << source; 96 97 // Create Program 98 CeedCallHiprtc(ceed, hiprtcCreateProgram(&prog, code.str().c_str(), NULL, 0, NULL, NULL)); 99 100 // Compile kernel 101 hiprtcResult result = hiprtcCompileProgram(prog, num_opts, opts); 102 103 if (result != HIPRTC_SUCCESS) { 104 size_t log_size; 105 char *log; 106 107 CeedChk_hiprtc(ceed, hiprtcGetProgramLogSize(prog, &log_size)); 108 CeedCallBackend(CeedMalloc(log_size, &log)); 109 CeedCallHiprtc(ceed, hiprtcGetProgramLog(prog, log)); 110 return CeedError(ceed, CEED_ERROR_BACKEND, "%s\n%s", hiprtcGetErrorString(result), log); 111 } 112 113 CeedCallHiprtc(ceed, hiprtcGetCodeSize(prog, &ptx_size)); 114 CeedCallBackend(CeedMalloc(ptx_size, &ptx)); 115 CeedCallHiprtc(ceed, hiprtcGetCode(prog, ptx)); 116 CeedCallHiprtc(ceed, hiprtcDestroyProgram(&prog)); 117 118 CeedCallHip(ceed, hipModuleLoadData(module, ptx)); 119 CeedCallBackend(CeedFree(&ptx)); 120 return CEED_ERROR_SUCCESS; 121 } 122 123 //------------------------------------------------------------------------------ 124 // Get HIP kernel 125 //------------------------------------------------------------------------------ 126 int CeedGetKernel_Hip(Ceed ceed, hipModule_t module, const char *name, hipFunction_t *kernel) { 127 CeedCallHip(ceed, hipModuleGetFunction(kernel, module, name)); 128 return CEED_ERROR_SUCCESS; 129 } 130 131 //------------------------------------------------------------------------------ 132 // Run HIP kernel 133 //------------------------------------------------------------------------------ 134 int CeedRunKernel_Hip(Ceed ceed, hipFunction_t kernel, const int grid_size, const int block_size, void **args) { 135 CeedCallHip(ceed, hipModuleLaunchKernel(kernel, grid_size, 1, 1, block_size, 1, 1, 0, NULL, args, NULL)); 136 return CEED_ERROR_SUCCESS; 137 } 138 139 //------------------------------------------------------------------------------ 140 // Run HIP kernel for spatial dimension 141 //------------------------------------------------------------------------------ 142 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, 143 void **args) { 144 CeedCallHip(ceed, hipModuleLaunchKernel(kernel, grid_size, 1, 1, block_size_x, block_size_y, block_size_z, 0, NULL, args, NULL)); 145 return CEED_ERROR_SUCCESS; 146 } 147 148 //------------------------------------------------------------------------------ 149 // Run HIP kernel for spatial dimension with shared memory 150 //------------------------------------------------------------------------------ 151 int CeedRunKernelDimShared_Hip(Ceed ceed, hipFunction_t 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 CeedCallHip(ceed, hipModuleLaunchKernel(kernel, grid_size, 1, 1, block_size_x, block_size_y, block_size_z, shared_mem_size, NULL, args, NULL)); 154 return CEED_ERROR_SUCCESS; 155 } 156 157 //------------------------------------------------------------------------------ 158