Method CompileCudaModule

paddle/cinn/backends/compiler.cc:673–711

}

void Compiler::CompileCudaModule(const Module& module,
                                 const std::string& code) {
#ifdef CINN_WITH_CUDA
  auto _host_module_device_module_ =
      SplitDeviceAndHostModule(module);  // NOLINT
  auto& host_module = std::get<0>(_host_module_device_module_);
  auto& device_module = std::get<1>(_host_module_device_module_);
  VLOG(3) << "[CUDA] host module:\n" << host_module;

  VLOG(3) << "[CUDA] device module:\n" << device_module;
  std::string source_code;

  if (!FLAGS_cinn_debug_custom_code_path.empty()) {
    std::string file_path = FLAGS_cinn_debug_custom_code_path;
    source_code = GetFileContent(file_path);
  } else if (code.empty()) {
    CodeGenCudaDev codegen(target_);
    source_code = codegen.Compile(device_module);
  } else {
    source_code = code;
  }

  PADDLE_ENFORCE_EQ(!source_code.empty(),
                    true,
                    ::common::errors::InvalidArgument(
                        "Compile CUDA C code failed from device module"));
  VLOG(3) << "[CUDA] C:\n" << source_code;
  SourceCodePrint::GetInstance()->write(source_code);
  device_fn_code_ += source_code;

  for (auto& fn : device_module.functions()) {
    std::string kernel_fn_name = fn->name;
    device_fn_name_.emplace_back(kernel_fn_name);
  }
  engine_->Link<CodeGenGpuHost>(host_module);
#else
  CINN_NOT_IMPLEMENTED
#endif
}

void Compiler::CompileCustomDeviceModule(const Module& module,
                                         const std::string& code) {