[Autotuner] Add support for multi-step search

The previous implementation assumed that a search strategy's
iteration consist of only one step, however it is possible that
multiple steps that depend on each other (and are not full
"iterations") are necessary.

Author: Theodoros Theodoridis

tc/autotuner/autotuner-inl.h

@@ -79,16 +79,16 @@ TuningHarness<Backend>::bestMappingOptions() const {
   }
 
 template <typename Backend>
-template <typename SearchStrategy>
-void TuningHarness<Backend>::doCompile(SearchStrategy& searchStrategy) {
+template <typename Candidates>
+void TuningHarness<Backend>::doCompile(Candidates& candidates) {
   // Atomically fetch and add the next job until there are no jobs left
   while (true) {
     auto current = currentCompilationJob_.fetch_add(1);
-    if (current >= searchStrategy.population.size()) {
+    if (current >= candidates.size()) {
       break;
     }
     std::unique_ptr<typename Backend::ExecutorType> pExecutor(nullptr);
-    auto pConf = searchStrategy.population.at(current).get();
+    auto pConf = candidates.at(current).get();
     auto options = makeOptions<Backend>(baseMapping_, *pConf);
     try {
       if (FLAGS_debug_tuner) {
@@ -245,14 +245,15 @@ void TuningHarness<Backend>::runOneIteration(
   auto devices = detail::parseDevices<Backend>(FLAGS_tuner_devices);
   CHECK(executors_.empty());
   CHECK(configurations_.empty());
-
-  {
+  for (uint64_t step = 0; step < searchStrategy.stepsPerIteration; ++step) {
+    auto& candidates = searchStrategy.candidatesOfStep(step);
     // Initialize for this round
     currentCompilationJob_.store(0);
     numEvaluations_.store(0);
     Printer printer(
         iteration,
-        searchStrategy.population.size(),
+        step,
+        candidates.size(),
         currentCompilationJob_,
         numEvaluations_);
     auto logIterations = FLAGS_tuner_gen_log_generations;
@@ -273,7 +274,7 @@ void TuningHarness<Backend>::runOneIteration(
     });
     for (size_t i = 0; i < FLAGS_tuner_threads; ++i) {
       cpuCompilationThreads.emplace_back(
-          [this, &searchStrategy]() { this->doCompile(searchStrategy); });
+          [this, &candidates]() { this->doCompile(candidates); });
     }
 
     // Just spawn and join new threads for each device
@@ -287,12 +288,13 @@ void TuningHarness<Backend>::runOneIteration(
         workerThread.join();
       }
     });
-    auto populationSize = searchStrategy.population.size();
+    auto populationSize = candidates.size();
     for (auto device : devices) {
       workerThreads.emplace_back([this, device, populationSize, &printer]() {
         this->doEvaluate(device, populationSize, printer);
       });
     }
+    searchStrategy.finishStep(step);
   }
 
   // At this point everything is synchronized because out of scope, done

tc/autotuner/genetic_search.cc

@@ -307,6 +307,13 @@ void GeneticSearch::updateParameters() {
   }
 }
 
+GeneticSearch::Population& GeneticSearch::candidatesOfStep(uint64_t step) {
+  if (step != 0) {
+    throw std::invalid_argument("GeneticSearch has only one step");
+  }
+  return population;
+}
+
 } // namespace autotune
 } // namespace tc
 

tc/autotuner/genetic_search.h

@@ -92,13 +92,17 @@ class GeneticSearch {
 
   using Population = std::vector<std::unique_ptr<CandidateConfiguration>>;
 
+  Population& candidatesOfStep(uint64_t);
+  void finishStep(uint64_t) {}
+
   Population population;
   TuningConfiguration lastBestConf;
   const size_t numGenerations;
   const size_t maxPopulationSize;
   const uint8_t crossOverRate;
   const uint8_t mutationRate;
   const size_t numberElites;
+  const size_t stepsPerIteration = 1;
 
   /*
    * c++11 seeding is (apparently) not of the highest quality:

tc/autotuner/utils.cc

@@ -64,7 +64,7 @@ void Printer::printLoop() {
     std::this_thread::sleep_for(std::chrono::seconds(1));
 
     std::stringstream ss;
-    ss << "Iteration " << iteration_;
+    ss << "Iteration.Step " << iteration_ << '.' << step_;
     ss << "\tJobs(Compiled, Evaluated)/total  ("
        << std::min(total_, currentCompilationJob_.load()) << ", "
        << std::min(total_, numEvaluations_.load()) << ")/" << total_;
@@ -100,10 +100,12 @@ void Printer::printLoop() {
 
 Printer::Printer(
     size_t iteration,
+    size_t step,
     size_t total,
     const std::atomic_size_t& currentCompilationJob,
     const std::atomic_size_t& numEvaluations)
     : iteration_(iteration),
+      step_(step),
       printerThread_([this]() { printLoop(); }),
       total_(total),
       currentCompilationJob_(currentCompilationJob),

tc/autotuner/utils.h

@@ -47,6 +47,7 @@ class Printer {
  public:
   Printer(
       size_t iteration,
+      size_t step,
       size_t total,
       const std::atomic_size_t& currentCompilationJob,
       const std::atomic_size_t& numEvaluations);
@@ -61,6 +62,7 @@ class Printer {
   void printLoop();
 
   size_t iteration_;
+  size_t step_;
   std::vector<Duration> runtimes_;
   mutable std::mutex runtimesMtx_;