示例#1
0
void Routine::InitProgram(std::initializer_list<const char *> source) {

  // Determines the identifier for this particular routine call
  auto routine_info = routine_name_;
  for (const auto &kernel_name : kernel_names_) {
    routine_info += "_" + kernel_name + db_(kernel_name).GetValuesString();
  }
  log_debug(routine_info);

  // Queries the cache to see whether or not the program (context-specific) is already there
  bool has_program;
  program_ = ProgramCache::Instance().Get(ProgramKeyRef{ context_(), device_(), precision_, routine_info },
                                          &has_program);
  if (has_program) { return; }

  // Sets the build options from an environmental variable (if set)
  auto options = std::vector<std::string>();
  const auto environment_variable = std::getenv("CLBLAST_BUILD_OPTIONS");
  if (environment_variable != nullptr) {
    options.push_back(std::string(environment_variable));
  }

  // Queries the cache to see whether or not the binary (device-specific) is already there. If it
  // is, a program is created and stored in the cache
  const auto device_name = GetDeviceName(device_);
  const auto platform_id = device_.PlatformID();
  bool has_binary;
  auto binary = BinaryCache::Instance().Get(BinaryKeyRef{platform_id,  precision_, routine_info, device_name },
                                            &has_binary);
  if (has_binary) {
    program_ = std::make_shared<Program>(device_, context_, binary);
    program_->Build(device_, options);
    ProgramCache::Instance().Store(ProgramKey{ context_(), device_(), precision_, routine_info },
                                    std::shared_ptr<Program>{program_});
    return;
  }

  // Otherwise, the kernel will be compiled and program will be built. Both the binary and the
  // program will be added to the cache.

  // Inspects whether or not FP64 is supported in case of double precision
  if ((precision_ == Precision::kDouble && !PrecisionSupported<double>(device_)) ||
      (precision_ == Precision::kComplexDouble && !PrecisionSupported<double2>(device_))) {
    throw RuntimeErrorCode(StatusCode::kNoDoublePrecision);
  }

  // As above, but for FP16 (half precision)
  if (precision_ == Precision::kHalf && !PrecisionSupported<half>(device_)) {
    throw RuntimeErrorCode(StatusCode::kNoHalfPrecision);
  }

  // Collects the parameters for this device in the form of defines
  auto source_string = std::string{""};
  for (const auto &kernel_name : kernel_names_) {
    source_string += db_(kernel_name).GetDefines();
  }

  // Adds routine-specific code to the constructed source string
  for (const char *s: source) {
    source_string += s;
  }

  // Completes the source and compiles the kernel
  program_ = CompileFromSource(source_string, precision_, routine_name_,
                               device_, context_, options, 0);


  // Store the compiled binary and program in the cache
  BinaryCache::Instance().Store(BinaryKey{platform_id, precision_, routine_info, device_name},
                                program_->GetIR());

  ProgramCache::Instance().Store(ProgramKey{context_(), device_(), precision_, routine_info},
                                 std::shared_ptr<Program>{program_});
}
示例#2
0
文件: sync.hpp 项目: planaria/kumori
		void resume()
		{
			std::lock_guard<std::mutex> lock(mutex_);
			current_impl() = this;
			context_ = std::get<0>(context_(this));
		}