Method WriteDeviceLinkStatements

Source/cmNinjaNormalTargetGenerator.cxx:891–974

}

void cmNinjaNormalTargetGenerator::WriteDeviceLinkStatements(
  std::string const& config, std::vector<std::string> const& architectures,
  std::string const& output)
{
  // Ensure there are no duplicates.
  cmNinjaDeps const explicitDeps = [&]() -> std::vector<std::string> {
    std::unordered_set<std::string> depsSet;
    cmNinjaDeps const linkDeps =
      this->ComputeLinkDeps(this->TargetLinkLanguage(config), config, true);
    cmNinjaDeps const objects = this->GetObjects(config);
    depsSet.insert(linkDeps.begin(), linkDeps.end());
    depsSet.insert(objects.begin(), objects.end());

    std::vector<std::string> deps;
    std::copy(depsSet.begin(), depsSet.end(), std::back_inserter(deps));
    return deps;
  }();

  cmGlobalNinjaGenerator* globalGen{ this->GetGlobalGenerator() };
  std::string const objectDir =
    cmStrCat(this->GeneratorTarget->GetSupportDirectory(),
             globalGen->ConfigDirectory(config));
  std::string const ninjaOutputDir = this->ConvertToNinjaPath(objectDir);

  cmNinjaBuild fatbinary(this->LanguageLinkerCudaFatbinaryRule(config));

  // Link device code for each architecture.
  for (std::string const& architectureKind : architectures) {
    // Clang always generates real code, so strip the specifier.
    std::string const architecture =
      architectureKind.substr(0, architectureKind.find('-'));
    std::string const cubin =
      cmStrCat(ninjaOutputDir, "/sm_", architecture, ".cubin");

    cmNinjaBuild dlink(this->LanguageLinkerCudaDeviceRule(config));
    dlink.ExplicitDeps = explicitDeps;
    dlink.Outputs = { cubin };
    dlink.Variables["ARCH"] = cmStrCat("sm_", architecture);

    // The generated register file contains macros that when expanded register
    // the device routines. Because the routines are the same for all
    // architectures the register file will be the same too. Thus generate it
    // only on the first invocation to reduce overhead.
    if (fatbinary.ExplicitDeps.empty()) {
      dlink.Variables["REGISTER"] = cmStrCat(
        "--register-link-binaries=", ninjaOutputDir, "/cmake_cuda_register.h");
    }

    fatbinary.Variables["PROFILES"] +=
      cmStrCat(" -im=profile=sm_", architecture, ",file=", cubin);
    fatbinary.ExplicitDeps.emplace_back(cubin);

    globalGen->WriteBuild(this->GetCommonFileStream(), dlink);
  }

  // Combine all architectures into a single fatbinary.
  fatbinary.Outputs = { cmStrCat(ninjaOutputDir, "/cmake_cuda_fatbin.h") };
  globalGen->WriteBuild(this->GetCommonFileStream(), fatbinary);