rocm-systems/hipamd/api/hip/hip_platform.cpp

/*
Copyright (c) 2015 - present Advanced Micro Devices, Inc. All rights reserved.

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
*/

#include <hip/hip_runtime.h>

#include "hip_internal.hpp"
#include "platform/program.hpp"
#include "platform/runtime.hpp"

#include <unordered_map>
#include "elfio.hpp"

constexpr unsigned __hipFatMAGIC2 = 0x48495046; // "HIPF"

thread_local std::stack<ihipExec_t> execStack_;
PlatformState* PlatformState::platform_ = new PlatformState();

struct __CudaFatBinaryWrapper {
  unsigned int magic;
  unsigned int version;
  void*        binary;
  void*        dummy1;
};

#define CLANG_OFFLOAD_BUNDLER_MAGIC_STR "__CLANG_OFFLOAD_BUNDLE__"
#define HIP_AMDGCN_AMDHSA_TRIPLE "hip-amdgcn-amd-amdhsa"
#define HCC_AMDGCN_AMDHSA_TRIPLE "hcc-amdgcn-amd-amdhsa-"

struct __ClangOffloadBundleDesc {
  uint64_t offset;
  uint64_t size;
  uint64_t tripleSize;
  const char triple[1];
};

struct __ClangOffloadBundleHeader {
  const char magic[sizeof(CLANG_OFFLOAD_BUNDLER_MAGIC_STR) - 1];
  uint64_t numBundles;
  __ClangOffloadBundleDesc desc[1];
};

hipError_t hipModuleGetGlobal(hipDeviceptr_t* dptr, size_t* bytes,
    hipModule_t hmod, const char* name);

hipError_t ihipCreateGlobalVarObj(const char* name, hipModule_t hmod, amd::Memory** amd_mem_obj,
                                  hipDeviceptr_t* dptr, size_t* bytes);

static bool isCompatibleCodeObject(const std::string& codeobj_target_id,
                                   const char* device_name) {
  // Workaround for gfx906 device name mismatch.
  // If bundle target id starts with gfx906 and device name starts with
  // gfx906, treat them as match.
  return codeobj_target_id.compare(device_name) == 0 ||
         (codeobj_target_id.find("gfx906") == 0 &&
          std::string(device_name).find("gfx906") == 0);
}

// Extracts code objects from fat binary in data for device names given in devices.
// Returns true if code objects are extracted successfully.
bool __hipExtractCodeObjectFromFatBinary(const void* data,
                                         const std::vector<const char*>& devices,
                                         std::vector<std::pair<const void*, size_t>>& code_objs)
{
  HIP_INIT();

  std::string magic((const char*)data, sizeof(CLANG_OFFLOAD_BUNDLER_MAGIC_STR) - 1);
  if (magic.compare(CLANG_OFFLOAD_BUNDLER_MAGIC_STR)) {
    return false;
  }

  code_objs.resize(devices.size());
  const auto obheader = reinterpret_cast<const __ClangOffloadBundleHeader*>(data);
  const auto* desc = &obheader->desc[0];
  unsigned num_code_objs = 0;
  for (uint64_t i = 0; i < obheader->numBundles; ++i,
       desc = reinterpret_cast<const __ClangOffloadBundleDesc*>(
           reinterpret_cast<uintptr_t>(&desc->triple[0]) + desc->tripleSize)) {

    std::string triple(desc->triple, sizeof(HIP_AMDGCN_AMDHSA_TRIPLE) - 1);
    if (triple.compare(HIP_AMDGCN_AMDHSA_TRIPLE))
      continue;

    std::string target(desc->triple + sizeof(HIP_AMDGCN_AMDHSA_TRIPLE),
                       desc->tripleSize - sizeof(HIP_AMDGCN_AMDHSA_TRIPLE));

    const void *image = reinterpret_cast<const void*>(
        reinterpret_cast<uintptr_t>(obheader) + desc->offset);
    size_t size = desc->size;

    for (size_t dev = 0; dev < devices.size(); ++dev) {
      const char* name = devices[dev];

      if (!isCompatibleCodeObject(target, name)) {
          continue;
      }
      code_objs[dev] = std::make_pair(image, size);
      num_code_objs++;
    }
  }
  if (num_code_objs == devices.size())
    return true;
  else
    return false;
}

extern "C" std::vector< std::pair<hipModule_t, bool> >* __hipRegisterFatBinary(const void* data)
{
  HIP_INIT();

  if(g_devices.empty()) {
    return nullptr;
  }
  const __CudaFatBinaryWrapper* fbwrapper = reinterpret_cast<const __CudaFatBinaryWrapper*>(data);
  if (fbwrapper->magic != __hipFatMAGIC2 || fbwrapper->version != 1) {
    return nullptr;
  }

  std::vector<const char*> devices;
  std::vector<std::pair<const void*, size_t>> code_objs;
  for (size_t dev = 0; dev < g_devices.size(); ++dev) {
    amd::Context* ctx = g_devices[dev];
    devices.push_back(ctx->devices()[0]->info().name_);
  }

  if (!__hipExtractCodeObjectFromFatBinary((char*)fbwrapper->binary, devices, code_objs)) {
    return nullptr;
  }

  auto programs = new std::vector< std::pair<hipModule_t, bool> >{g_devices.size()};
  for (size_t dev = 0; dev < g_devices.size(); ++dev) {
    amd::Context* ctx = g_devices[dev];
    amd::Program* program = new amd::Program(*ctx);
    if (program == nullptr) {
      return nullptr;
    }
    if (CL_SUCCESS == program->addDeviceProgram(*ctx->devices()[0], code_objs[dev].first, code_objs[dev].second)) {
      programs->at(dev) = std::make_pair(reinterpret_cast<hipModule_t>(as_cl(program)) , false);
    }
  }

  return programs;
}

void PlatformState::registerVar(const void* hostvar,
                                const DeviceVar& rvar) {
  amd::ScopedLock lock(lock_);
  vars_.insert(std::make_pair(hostvar, rvar));
}

void PlatformState::registerFunction(const void* hostFunction,
                                     const DeviceFunction& func) {
  amd::ScopedLock lock(lock_);
  functions_.insert(std::make_pair(hostFunction, func));
}

bool ihipGetFuncAttributes(const char* func_name, amd::Program* program, hipFuncAttributes* func_attr) {
  device::Program* dev_program
    = program->getDeviceProgram(*hip::getCurrentContext()->devices()[0]);

  const auto it = dev_program->kernels().find(std::string(func_name));
  if (it == dev_program->kernels().cend()) {
    return false;
  }

  const device::Kernel::WorkGroupInfo* wginfo = it->second->workGroupInfo();
  func_attr->localSizeBytes = wginfo->localMemSize_;
  func_attr->sharedSizeBytes = wginfo->size_;
  func_attr->maxThreadsPerBlock = wginfo->wavefrontSize_;
  func_attr->numRegs = wginfo->usedVGPRs_;

  return true;
}

hipFunction_t PlatformState::getFunc(const void* hostFunction, int deviceId) {
  amd::ScopedLock lock(lock_);
  const auto it = functions_.find(hostFunction);
  if (it != functions_.cend()) {
    PlatformState::DeviceFunction& devFunc = it->second;
    if (devFunc.functions[deviceId] == 0) {
      hipModule_t module = (*devFunc.modules)[deviceId].first;
      if (!(*devFunc.modules)[deviceId].second) {
        amd::Program* program = as_amd(reinterpret_cast<cl_program>(module));
        if (CL_SUCCESS != program->build(g_devices[deviceId]->devices(), nullptr, nullptr, nullptr)) {
          return nullptr;
        }
        (*devFunc.modules)[deviceId].second = true;
      }
      hipFunction_t function = nullptr;
      if (hipSuccess == hipModuleGetFunction(&function, module, devFunc.deviceName.c_str()) &&
          function != nullptr) {
        devFunc.functions[deviceId] = function;
      }
      else {
   //     tprintf(DB_FB, "__hipRegisterFunction cannot find kernel %s for"
   //         " device %d\n", deviceName, deviceId);
      }
    }
    return devFunc.functions[deviceId];
  }
  return nullptr;
}

bool PlatformState::getFuncAttr(const void* hostFunction,
                                hipFuncAttributes* func_attr) {

  if (func_attr == nullptr) {
    return false;
  }

  const auto it = functions_.find(hostFunction);
  if (it == functions_.cend()) {
    return false;
  }

  PlatformState::DeviceFunction& devFunc = it->second;
  int deviceId = ihipGetDevice();

  /* If module has not been initialized yet, build the kernel now*/
  if (!(*devFunc.modules)[deviceId].second) {
    if (nullptr == PlatformState::instance().getFunc(hostFunction, deviceId)) {
      return false;
    }
  }

  amd::Program* program = as_amd(reinterpret_cast<cl_program>((*devFunc.modules)[deviceId].first));
  if (!ihipGetFuncAttributes(devFunc.deviceName.c_str(), program, func_attr)) {
    return false;
  }
  return true;
}


bool PlatformState::getGlobalVar(const void* hostVar, int deviceId,
                                 hipDeviceptr_t* dev_ptr, size_t* size_ptr) {
  amd::ScopedLock lock(lock_);
  const auto it = vars_.find(hostVar);
  if (it != vars_.cend()) {
    DeviceVar& dvar = it->second;
    if (dvar.rvars[deviceId].getdeviceptr() == nullptr) {
      size_t sym_size = 0;
      hipDeviceptr_t device_ptr = nullptr;
      amd::Memory* amd_mem_obj = nullptr;

      if (!(*dvar.modules)[deviceId].second) {
        amd::Program* program = as_amd(reinterpret_cast<cl_program>((*dvar.modules)[deviceId].first));
        if (CL_SUCCESS != program->build(g_devices[deviceId]->devices(), nullptr, nullptr, nullptr)) {
          return false;
        }
        (*dvar.modules)[deviceId].second = true;
      }
      if((hipSuccess == ihipCreateGlobalVarObj(dvar.hostVar.c_str(), (*dvar.modules)[deviceId].first,
                                               &amd_mem_obj, &device_ptr, &sym_size))
           && (device_ptr != nullptr)) {
        dvar.rvars[deviceId].size_ = sym_size;
        dvar.rvars[deviceId].devicePtr_ = device_ptr;
        dvar.rvars[deviceId].amd_mem_obj_ = amd_mem_obj;
        amd::MemObjMap::AddMemObj(device_ptr, amd_mem_obj);
      } else {
        LogError("[HIP] __hipRegisterVar cannot find kernel for device \n");
      }
    }
    *size_ptr = dvar.rvars[deviceId].getvarsize();
    *dev_ptr = dvar.rvars[deviceId].getdeviceptr();
    return true;
  } else {
    return false;
  }
}

void PlatformState::setupArgument(const void *arg, size_t size, size_t offset) {
  auto& arguments = execStack_.top().arguments_;

  if (arguments.size() < offset + size) {
    arguments.resize(offset + size);
  }

  ::memcpy(&arguments[offset], arg, size);
}

void PlatformState::configureCall(dim3 gridDim, dim3 blockDim, size_t sharedMem,
                                  hipStream_t stream) {
  execStack_.push(ihipExec_t{gridDim, blockDim, sharedMem, stream});
}

void PlatformState::popExec(ihipExec_t& exec) {
  exec = std::move(execStack_.top());
  execStack_.pop();
}

extern "C" void __hipRegisterFunction(
  std::vector<std::pair<hipModule_t,bool> >* modules,
  const void*  hostFunction,
  char*        deviceFunction,
  const char*  deviceName,
  unsigned int threadLimit,
  uint3*       tid,
  uint3*       bid,
  dim3*        blockDim,
  dim3*        gridDim,
  int*         wSize)
{
  HIP_INIT();
  PlatformState::DeviceFunction func{ std::string{deviceName}, modules, std::vector<hipFunction_t>{ g_devices.size() }};
  PlatformState::instance().registerFunction(hostFunction, func);
}

// Registers a device-side global variable.
// For each global variable in device code, there is a corresponding shadow
// global variable in host code. The shadow host variable is used to keep
// track of the value of the device side global variable between kernel
// executions.
extern "C" void __hipRegisterVar(
  std::vector<std::pair<hipModule_t,bool> >* modules,   // The device modules containing code object
  char*       var,       // The shadow variable in host code
  char*       hostVar,   // Variable name in host code
  char*       deviceVar, // Variable name in device code
  int         ext,       // Whether this variable is external
  int         size,      // Size of the variable
  int         constant,  // Whether this variable is constant
  int         global)    // Unknown, always 0
{
  HIP_INIT();

  PlatformState::DeviceVar dvar{ std::string{ hostVar }, modules,
    std::vector<PlatformState::RegisteredVar>{ g_devices.size() } };

  PlatformState::instance().registerVar(hostVar, dvar);
}

extern "C" void __hipUnregisterFatBinary(std::vector< std::pair<hipModule_t, bool> >* modules)
{
  HIP_INIT();

  std::for_each(modules->begin(), modules->end(), [](std::pair<hipModule_t, bool> module){
    if (module.first != nullptr) {
      as_amd(reinterpret_cast<cl_program>(module.first))->release();
    }
  });
  delete modules;
}

extern "C" hipError_t hipConfigureCall(
  dim3 gridDim,
  dim3 blockDim,
  size_t sharedMem,
  hipStream_t stream)
{
  HIP_INIT_API(gridDim, blockDim, sharedMem, stream);

  PlatformState::instance().configureCall(gridDim, blockDim, sharedMem, stream);

  HIP_RETURN(hipSuccess);
}

extern "C" hipError_t hipSetupArgument(
  const void *arg,
  size_t size,
  size_t offset)
{
  HIP_INIT_API(arg, size, offset);

  PlatformState::instance().setupArgument(arg, size, offset);

  HIP_RETURN(hipSuccess);
}

extern "C" hipError_t hipLaunchByPtr(const void *hostFunction)
{
  HIP_INIT_API(hostFunction);

  int deviceId = ihipGetDevice();
  hipFunction_t func = PlatformState::instance().getFunc(hostFunction, deviceId);
  if (func == nullptr) {
    HIP_RETURN(hipErrorUnknown);
  }

  ihipExec_t exec;
  PlatformState::instance().popExec(exec);

  size_t size = exec.arguments_.size();
  void *extra[] = {
      HIP_LAUNCH_PARAM_BUFFER_POINTER, &exec.arguments_[0],
      HIP_LAUNCH_PARAM_BUFFER_SIZE, &size,
      HIP_LAUNCH_PARAM_END
    };

  HIP_RETURN(hipModuleLaunchKernel(func,
    exec.gridDim_.x, exec.gridDim_.y, exec.gridDim_.z,
    exec.blockDim_.x, exec.blockDim_.y, exec.blockDim_.z,
    exec.sharedMem_, exec.hStream_, nullptr, extra));
}

hipError_t hipGetSymbolAddress(void** devPtr, const void* symbolName) {
  size_t size = 0;
  if(!PlatformState::instance().getGlobalVar(symbolName, ihipGetDevice(), devPtr, &size)) {
    HIP_RETURN(hipErrorUnknown);
  }
  HIP_RETURN(hipSuccess);
}

hipError_t hipGetSymbolSize(size_t* sizePtr, const void* symbolName) {
  hipDeviceptr_t devPtr = nullptr;
  if (!PlatformState::instance().getGlobalVar(symbolName, ihipGetDevice(), &devPtr, sizePtr)) {
    HIP_RETURN(hipErrorUnknown);
  }
  HIP_RETURN(hipSuccess);
}

hipError_t ihipCreateGlobalVarObj(const char* name, hipModule_t hmod, amd::Memory** amd_mem_obj, hipDeviceptr_t* dptr, size_t* bytes)
{
  HIP_INIT();

  amd::Program* program = nullptr;
  device::Program* dev_program = nullptr;

  /* Get Device Program pointer*/
  program = as_amd(reinterpret_cast<cl_program>(hmod));
  dev_program = program->getDeviceProgram(*hip::getCurrentContext()->devices()[0]);

  if (dev_program == nullptr) {
    HIP_RETURN(hipErrorUnknown);
  }
  /* Find the global Symbols */
  if(!dev_program->createGlobalVarObj(amd_mem_obj, dptr, bytes, name)) {
    HIP_RETURN(hipErrorUnknown);
  }

  HIP_RETURN(hipSuccess);
}


#if defined(ATI_OS_LINUX)

namespace hip_impl {

struct dl_phdr_info {
  ELFIO::Elf64_Addr        dlpi_addr;
  const char       *dlpi_name;
  const ELFIO::Elf64_Phdr *dlpi_phdr;
  ELFIO::Elf64_Half        dlpi_phnum;
};

extern "C" int dl_iterate_phdr(
  int (*callback) (struct dl_phdr_info *info, size_t size, void *data), void *data
);

struct Symbol {
  std::string name;
  ELFIO::Elf64_Addr value = 0;
  ELFIO::Elf_Xword size = 0;
  ELFIO::Elf_Half sect_idx = 0;
  uint8_t bind = 0;
  uint8_t type = 0;
  uint8_t other = 0;
};

inline Symbol read_symbol(const ELFIO::symbol_section_accessor& section, unsigned int idx) {
  assert(idx < section.get_symbols_num());

  Symbol r;
  section.get_symbol(idx, r.name, r.value, r.size, r.bind, r.type, r.sect_idx, r.other);

  return r;
}

template <typename P>
inline ELFIO::section* find_section_if(ELFIO::elfio& reader, P p) {
    const auto it = find_if(reader.sections.begin(), reader.sections.end(), std::move(p));

    return it != reader.sections.end() ? *it : nullptr;
}

std::vector<std::pair<uintptr_t, std::string>> function_names_for(const ELFIO::elfio& reader,
                                                                  ELFIO::section* symtab) {
  std::vector<std::pair<uintptr_t, std::string>> r;
  ELFIO::symbol_section_accessor symbols{reader, symtab};

  for (auto i = 0u; i != symbols.get_symbols_num(); ++i) {
    auto tmp = read_symbol(symbols, i);

    if (tmp.type == STT_FUNC && tmp.sect_idx != SHN_UNDEF && !tmp.name.empty()) {
      r.emplace_back(tmp.value, tmp.name);
    }
  }

  return r;
}

const std::vector<std::pair<uintptr_t, std::string>>& function_names_for_process() {
  static constexpr const char self[] = "/proc/self/exe";

  static std::vector<std::pair<uintptr_t, std::string>> r;
  static std::once_flag f;

  std::call_once(f, []() {
    ELFIO::elfio reader;

    if (reader.load(self)) {
      const auto it = find_section_if(
          reader, [](const ELFIO::section* x) { return x->get_type() == SHT_SYMTAB; });

      if (it) r = function_names_for(reader, it);
    }
  });

  return r;
}


const std::unordered_map<uintptr_t, std::string>& function_names()
{
  static std::unordered_map<uintptr_t, std::string> r{
    function_names_for_process().cbegin(),
    function_names_for_process().cend()};
  static std::once_flag f;

  std::call_once(f, []() {
    dl_iterate_phdr([](dl_phdr_info* info, size_t, void*) {
      ELFIO::elfio reader;

      if (reader.load(info->dlpi_name)) {
        const auto it = find_section_if(
            reader, [](const ELFIO::section* x) { return x->get_type() == SHT_SYMTAB; });

        if (it) {
          auto n = function_names_for(reader, it);

          for (auto&& f : n) f.first += info->dlpi_addr;

          r.insert(make_move_iterator(n.begin()), make_move_iterator(n.end()));
        }
      }
      return 0;
    },
    nullptr);
  });

  return r;
}

std::vector<char> bundles_for_process() {
  static constexpr const char self[] = "/proc/self/exe";
  static constexpr const char kernel_section[] = ".kernel";
  std::vector<char> r;

  ELFIO::elfio reader;

  if (reader.load(self)) {
    auto it = find_section_if(
        reader, [](const ELFIO::section* x) { return x->get_name() == kernel_section; });

    if (it) r.insert(r.end(), it->get_data(), it->get_data() + it->get_size());
  }

  return r;
}

const std::vector<hipModule_t>& modules() {
    static std::vector<hipModule_t> r;
    static std::once_flag f;

    std::call_once(f, []() {
      static std::vector<std::vector<char>> bundles{bundles_for_process()};

      dl_iterate_phdr(
          [](dl_phdr_info* info, std::size_t, void*) {
        ELFIO::elfio tmp;
        if (tmp.load(info->dlpi_name)) {
          const auto it = find_section_if(
              tmp, [](const ELFIO::section* x) { return x->get_name() == ".kernel"; });

          if (it) bundles.emplace_back(it->get_data(), it->get_data() + it->get_size());
        }
        return 0;
      },
      nullptr);

      for (auto&& bundle : bundles) {
        if (bundle.empty()) {
          continue;
        }
        std::string magic(&bundle[0], sizeof(CLANG_OFFLOAD_BUNDLER_MAGIC_STR) - 1);
        if (magic.compare(CLANG_OFFLOAD_BUNDLER_MAGIC_STR))
          continue;

        const auto obheader = reinterpret_cast<const __ClangOffloadBundleHeader*>(&bundle[0]);
        const auto* desc = &obheader->desc[0];
        for (uint64_t i = 0; i < obheader->numBundles; ++i,
             desc = reinterpret_cast<const __ClangOffloadBundleDesc*>(
                 reinterpret_cast<uintptr_t>(&desc->triple[0]) + desc->tripleSize)) {

          std::string triple(desc->triple, sizeof(HCC_AMDGCN_AMDHSA_TRIPLE) - 1);
          if (triple.compare(HCC_AMDGCN_AMDHSA_TRIPLE))
            continue;

          std::string target(desc->triple + sizeof(HCC_AMDGCN_AMDHSA_TRIPLE),
                             desc->tripleSize - sizeof(HCC_AMDGCN_AMDHSA_TRIPLE));

          if (isCompatibleCodeObject(target, hip::getCurrentContext()->devices()[0]->info().name_)) {
            hipModule_t module;
            if (hipSuccess == hipModuleLoadData(&module, reinterpret_cast<const void*>(
                reinterpret_cast<uintptr_t>(obheader) + desc->offset)))
              r.push_back(module);
              break;
          }
        }
      }
    });

    return r;
}

const std::unordered_map<uintptr_t, hipFunction_t>& functions()
{
  static std::unordered_map<uintptr_t, hipFunction_t> r;
  static std::once_flag f;

  std::call_once(f, []() {
    for (auto&& function : function_names()) {
      for (auto&& module : modules()) {
        hipFunction_t f;
        if (hipSuccess == hipModuleGetFunction(&f, module, function.second.c_str()))
          r[function.first] = f;
      }
    }
  });

  return r;
}


void hipLaunchKernelGGLImpl(
  uintptr_t function_address,
  const dim3& numBlocks,
  const dim3& dimBlocks,
  uint32_t sharedMemBytes,
  hipStream_t stream,
  void** kernarg)
{
  HIP_INIT();

  const auto it = functions().find(function_address);
  if (it == functions().cend())
    assert(0);

  hipModuleLaunchKernel(it->second,
    numBlocks.x, numBlocks.y, numBlocks.z,
    dimBlocks.x, dimBlocks.y, dimBlocks.z,
    sharedMemBytes, stream, nullptr, kernarg);
}

}

// conversion routines between float and half precision
static inline std::uint32_t f32_as_u32(float f) { union { float f; std::uint32_t u; } v; v.f = f; return v.u; }
static inline float u32_as_f32(std::uint32_t u) { union { float f; std::uint32_t u; } v; v.u = u; return v.f; }
static inline int clamp_int(int i, int l, int h) { return std::min(std::max(i, l), h); }

// half float, the f16 is in the low 16 bits of the input argument
static inline float __convert_half_to_float(std::uint32_t a) noexcept {
  std::uint32_t u = ((a << 13) + 0x70000000U) & 0x8fffe000U;
  std::uint32_t v = f32_as_u32(u32_as_f32(u) * u32_as_f32(0x77800000U)/*0x1.0p+112f*/) + 0x38000000U;
  u = (a & 0x7fff) != 0 ? v : u;
  return u32_as_f32(u) * u32_as_f32(0x07800000U)/*0x1.0p-112f*/;
}

// float half with nearest even rounding
// The lower 16 bits of the result is the bit pattern for the f16
static inline std::uint32_t __convert_float_to_half(float a) noexcept {
  std::uint32_t u = f32_as_u32(a);
  int e = static_cast<int>((u >> 23) & 0xff) - 127 + 15;
  std::uint32_t m = ((u >> 11) & 0xffe) | ((u & 0xfff) != 0);
  std::uint32_t i = 0x7c00 | (m != 0 ? 0x0200 : 0);
  std::uint32_t n = ((std::uint32_t)e << 12) | m;
  std::uint32_t s = (u >> 16) & 0x8000;
  int b = clamp_int(1-e, 0, 13);
  std::uint32_t d = (0x1000 | m) >> b;
  d |= (d << b) != (0x1000 | m);
  std::uint32_t v = e < 1 ? d : n;
  v = (v >> 2) + (((v & 0x7) == 3) | ((v & 0x7) > 5));
  v = e > 30 ? 0x7c00 : v;
  v = e == 143 ? i : v;
  return s | v;
}

extern "C" float __gnu_h2f_ieee(unsigned short h){
  return __convert_half_to_float((std::uint32_t) h);
}

extern "C" unsigned short __gnu_f2h_ieee(float f){
  return (unsigned short)__convert_float_to_half(f);
}

#endif // defined(ATI_OS_LINUX)