1*3d8e8822SJeremy L Thompson // Copyright (c) 2017-2022, Lawrence Livermore National Security, LLC and other CEED contributors. 2*3d8e8822SJeremy L Thompson // All Rights Reserved. See the top-level LICENSE and NOTICE files for details. 330f4f45fSnbeams // 4*3d8e8822SJeremy L Thompson // SPDX-License-Identifier: BSD-2-Clause 530f4f45fSnbeams // 6*3d8e8822SJeremy L Thompson // This file is part of CEED: http://github.com/ceed 730f4f45fSnbeams 8ec3da8bcSJed Brown #include <ceed/ceed.h> 9ec3da8bcSJed Brown #include <ceed/backend.h> 1030f4f45fSnbeams #include <sstream> 1130f4f45fSnbeams #include <stdarg.h> 123d576824SJeremy L Thompson #include <string.h> 1330f4f45fSnbeams #include <hip/hiprtc.h> 147fcac036SJeremy L Thompson #include "ceed-hip-common.h" 1530f4f45fSnbeams #include "ceed-hip-compile.h" 1630f4f45fSnbeams 1730f4f45fSnbeams #define CeedChk_hiprtc(ceed, x) \ 1830f4f45fSnbeams do { \ 1930f4f45fSnbeams hiprtcResult result = static_cast<hiprtcResult>(x); \ 2030f4f45fSnbeams if (result != HIPRTC_SUCCESS) \ 21e15f9bd0SJeremy L Thompson return CeedError((ceed), CEED_ERROR_BACKEND, hiprtcGetErrorString(result)); \ 2230f4f45fSnbeams } while (0) 2330f4f45fSnbeams 2430f4f45fSnbeams //------------------------------------------------------------------------------ 2530f4f45fSnbeams // Compile HIP kernel 2630f4f45fSnbeams //------------------------------------------------------------------------------ 2730f4f45fSnbeams int CeedCompileHip(Ceed ceed, const char *source, hipModule_t *module, 280d0321e0SJeremy L Thompson const CeedInt num_opts, ...) { 2930f4f45fSnbeams int ierr; 3030f4f45fSnbeams hipFree(0); // Make sure a Context exists for hiprtc 3130f4f45fSnbeams hiprtcProgram prog; 3230f4f45fSnbeams 3330f4f45fSnbeams std::ostringstream code; 349faa5937SNatalie Beams // Add hip runtime include statement for generation if runtime < 40400000 359faa5937SNatalie Beams // (implies ROCm < 4.5) 369faa5937SNatalie Beams int runtime_version; 379faa5937SNatalie Beams CeedChk_Hip(ceed, hipRuntimeGetVersion(&runtime_version)); 389faa5937SNatalie Beams if (runtime_version < 40400000) { 3930f4f45fSnbeams code << "\n#include <hip/hip_runtime.h>\n"; 409faa5937SNatalie Beams } 419faa5937SNatalie Beams // With ROCm 4.5, need to include these definitions specifically for hiprtc 429faa5937SNatalie Beams // (but cannot include the runtime header) 439faa5937SNatalie Beams else { 449faa5937SNatalie Beams code << "#include <stddef.h>\n"; 459faa5937SNatalie Beams code << "#define __forceinline__ inline __attribute__((always_inline))\n"; 469faa5937SNatalie Beams code << "#define HIP_DYNAMIC_SHARED(type, var) extern __shared__ type var[];\n"; 479faa5937SNatalie Beams } 4830f4f45fSnbeams 4930f4f45fSnbeams // Macro definitions 5030f4f45fSnbeams // Get kernel specific options, such as kernel constants 51b3c5430cSnbeams const int opts_size = 3; 520d0321e0SJeremy L Thompson const char *opts[opts_size]; 530d0321e0SJeremy L Thompson if (num_opts > 0) { 5430f4f45fSnbeams va_list args; 550d0321e0SJeremy L Thompson va_start(args, num_opts); 5630f4f45fSnbeams char *name; 5730f4f45fSnbeams int val; 580d0321e0SJeremy L Thompson for (int i = 0; i < num_opts; i++) { 5930f4f45fSnbeams name = va_arg(args, char *); 6030f4f45fSnbeams val = va_arg(args, int); 6130f4f45fSnbeams code << "#define " << name << " " << val << "\n"; 6230f4f45fSnbeams } 6330f4f45fSnbeams va_end(args); 6430f4f45fSnbeams } 6530f4f45fSnbeams 6630f4f45fSnbeams // Standard backend options 6780a9ef05SNatalie Beams if (CEED_SCALAR_TYPE == CEED_SCALAR_FP32) { 6880a9ef05SNatalie Beams code << "#define CeedScalar float\n"; 6980a9ef05SNatalie Beams } 7080a9ef05SNatalie Beams else { 71e15f9bd0SJeremy L Thompson code << "#define CeedScalar double\n"; 7280a9ef05SNatalie Beams } 73e15f9bd0SJeremy L Thompson code << "#define CeedInt int\n"; 74e15f9bd0SJeremy L Thompson code << "#define CEED_ERROR_SUCCESS 0\n\n"; 7530f4f45fSnbeams 7630f4f45fSnbeams // Non-macro options 7730f4f45fSnbeams opts[0] = "-default-device"; 7830f4f45fSnbeams struct hipDeviceProp_t prop; 7930f4f45fSnbeams Ceed_Hip *ceed_data; 80e15f9bd0SJeremy L Thompson ierr = CeedGetData(ceed, (void **)&ceed_data); CeedChkBackend(ierr); 810d0321e0SJeremy L Thompson CeedChk_Hip(ceed, hipGetDeviceProperties(&prop, ceed_data->device_id)); 820d0321e0SJeremy L Thompson std::string arch_arg = "--gpu-architecture=" + std::string(prop.gcnArchName); 830d0321e0SJeremy L Thompson opts[1] = arch_arg.c_str(); 84b3c5430cSnbeams opts[2] = "-munsafe-fp-atomics"; 8530f4f45fSnbeams 8630f4f45fSnbeams // Add string source argument provided in call 8730f4f45fSnbeams code << source; 8830f4f45fSnbeams 8930f4f45fSnbeams // Create Program 9030f4f45fSnbeams CeedChk_hiprtc(ceed, hiprtcCreateProgram(&prog, code.str().c_str(), NULL, 0, NULL, NULL)); 9130f4f45fSnbeams 9230f4f45fSnbeams // Compile kernel 930d0321e0SJeremy L Thompson hiprtcResult result = hiprtcCompileProgram(prog, opts_size, opts); 9430f4f45fSnbeams if (result != HIPRTC_SUCCESS) { 950d0321e0SJeremy L Thompson size_t log_size; 960d0321e0SJeremy L Thompson CeedChk_hiprtc(ceed, hiprtcGetProgramLogSize(prog, &log_size)); 9730f4f45fSnbeams char *log; 980d0321e0SJeremy L Thompson ierr = CeedMalloc(log_size, &log); CeedChkBackend(ierr); 9930f4f45fSnbeams CeedChk_hiprtc(ceed, hiprtcGetProgramLog(prog, log)); 1000d0321e0SJeremy L Thompson return CeedError(ceed, CEED_ERROR_BACKEND, "%s\n%s", 1010d0321e0SJeremy L Thompson hiprtcGetErrorString(result), log); 10230f4f45fSnbeams } 10330f4f45fSnbeams 1040d0321e0SJeremy L Thompson size_t ptx_size; 1050d0321e0SJeremy L Thompson CeedChk_hiprtc(ceed, hiprtcGetCodeSize(prog, &ptx_size)); 10630f4f45fSnbeams char *ptx; 1070d0321e0SJeremy L Thompson ierr = CeedMalloc(ptx_size, &ptx); CeedChkBackend(ierr); 10830f4f45fSnbeams CeedChk_hiprtc(ceed, hiprtcGetCode(prog, ptx)); 10917fed040Snbeams CeedChk_hiprtc(ceed, hiprtcDestroyProgram(&prog)); 11030f4f45fSnbeams 11130f4f45fSnbeams CeedChk_Hip(ceed, hipModuleLoadData(module, ptx)); 112e15f9bd0SJeremy L Thompson ierr = CeedFree(&ptx); CeedChkBackend(ierr); 11330f4f45fSnbeams 114e15f9bd0SJeremy L Thompson return CEED_ERROR_SUCCESS; 11530f4f45fSnbeams } 11630f4f45fSnbeams 11730f4f45fSnbeams //------------------------------------------------------------------------------ 11830f4f45fSnbeams // Get HIP kernel 11930f4f45fSnbeams //------------------------------------------------------------------------------ 12030f4f45fSnbeams int CeedGetKernelHip(Ceed ceed, hipModule_t module, const char *name, 12130f4f45fSnbeams hipFunction_t *kernel) { 12230f4f45fSnbeams 12330f4f45fSnbeams CeedChk_Hip(ceed, hipModuleGetFunction(kernel, module, name)); 124e15f9bd0SJeremy L Thompson return CEED_ERROR_SUCCESS; 12530f4f45fSnbeams } 12630f4f45fSnbeams 12730f4f45fSnbeams //------------------------------------------------------------------------------ 12830f4f45fSnbeams // Run HIP kernel 12930f4f45fSnbeams //------------------------------------------------------------------------------ 1300d0321e0SJeremy L Thompson int CeedRunKernelHip(Ceed ceed, hipFunction_t kernel, const int grid_size, 1310d0321e0SJeremy L Thompson const int block_size, void **args) { 1320d0321e0SJeremy L Thompson CeedChk_Hip(ceed, hipModuleLaunchKernel(kernel, grid_size, 1, 1, block_size, 1, 13330f4f45fSnbeams 1, 0, NULL, args, NULL)); 134e15f9bd0SJeremy L Thompson return CEED_ERROR_SUCCESS; 13530f4f45fSnbeams } 13630f4f45fSnbeams 13730f4f45fSnbeams //------------------------------------------------------------------------------ 13830f4f45fSnbeams // Run HIP kernel for spatial dimension 13930f4f45fSnbeams //------------------------------------------------------------------------------ 1400d0321e0SJeremy L Thompson int CeedRunKernelDimHip(Ceed ceed, hipFunction_t kernel, const int grid_size, 1410d0321e0SJeremy L Thompson const int block_size_x, const int block_size_y, 1420d0321e0SJeremy L Thompson const int block_size_z, void **args) { 1430d0321e0SJeremy L Thompson CeedChk_Hip(ceed, hipModuleLaunchKernel(kernel, grid_size, 1, 1, 1440d0321e0SJeremy L Thompson block_size_x, block_size_y, block_size_z, 14530f4f45fSnbeams 0, NULL, args, NULL)); 146e15f9bd0SJeremy L Thompson return CEED_ERROR_SUCCESS; 14730f4f45fSnbeams } 14830f4f45fSnbeams 14930f4f45fSnbeams //------------------------------------------------------------------------------ 150e15f9bd0SJeremy L Thompson // Run HIP kernel for spatial dimension with shared memory 15130f4f45fSnbeams //------------------------------------------------------------------------------ 1520d0321e0SJeremy L Thompson int CeedRunKernelDimSharedHip(Ceed ceed, hipFunction_t kernel, const int grid_size, 1530d0321e0SJeremy L Thompson const int block_size_x, const int block_size_y, 1540d0321e0SJeremy L Thompson const int block_size_z, const int shared_mem_size, 15530f4f45fSnbeams void **args) { 1560d0321e0SJeremy L Thompson CeedChk_Hip(ceed, hipModuleLaunchKernel(kernel, grid_size, 1, 1, 1570d0321e0SJeremy L Thompson block_size_x, block_size_y, block_size_z, 1580d0321e0SJeremy L Thompson shared_mem_size, NULL, args, NULL)); 159e15f9bd0SJeremy L Thompson return CEED_ERROR_SUCCESS; 16030f4f45fSnbeams } 161