Standardize eigen threading and batch size logic

cpp/command/benchmark.cpp

@@ -249,7 +249,7 @@ int MainCmds::benchmark(const vector<string>& args) {
       cout << "WARNING: Your nnMaxBatchSize is hardcoded to " + cfg.getString("nnMaxBatchSize") + ", recommend deleting it and using the default (which this benchmark assumes)" << endl;
 #ifdef USE_EIGEN_BACKEND
     if(cfg.contains("numEigenThreadsPerModel")) {
-      cout << "Note: Your numEigenThreadsPerModel is hardcoded to " + cfg.getString("numEigenThreadsPerModel") + ", consider deleting it and using the default (which this benchmark assumes when computing its performance stats)" << endl;
+      cout << "Note: Your numEigenThreadsPerModel is hardcoded to " + cfg.getString("numEigenThreadsPerModel") + ", this benchmark ignores it assumes that it is always set equal to the smaller of the number of search threads and the number of CPU cores on your computer when computing its performance stats." << endl;
     }
 #endif
 
@@ -328,7 +328,7 @@ static void setNumThreads(SearchParams& params, NNEvaluator* nnEval, Logger& log
   //Also, disable the logger to suppress the kill and respawn messages.
   logger.setDisabled(true);
   nnEval->killServerThreads();
-  nnEval->setNumThreads(vector<int>(numThreads,-1));
+  nnEval->setNumThreads(vector<int>(Setup::computeDefaultEigenBackendThreads(numThreads,logger),-1));
   nnEval->spawnServerThreads();
   //Also since we killed and respawned all the threads, re-warm them
   Rand seedRand;

cpp/program/setup.cpp

@@ -174,22 +174,9 @@ vector<NNEvaluator*> Setup::initializeNNEvaluators(
       cfg.contains("numNNServerThreadsPerModel") ? cfg.getInt("numNNServerThreadsPerModel",1,1024) : 1;
 #else
     cfg.markAllKeysUsedWithPrefix("numNNServerThreadsPerModel");
-    auto getNumCores = [&logger]() {
-      int numCores = (int)std::thread::hardware_concurrency();
-      if(numCores <= 0) {
-        logger.write("Could not determine number of cores on this machine, choosing default parameters as if it were 8");
-        numCores = 8;
-      }
-      return numCores;
-    };
     int numNNServerThreadsPerModel =
       cfg.contains("numEigenThreadsPerModel") ? cfg.getInt("numEigenThreadsPerModel",1,1024) :
-      setupFor == SETUP_FOR_DISTRIBUTED ? std::min(expectedConcurrentEvals,getNumCores()) :
-      setupFor == SETUP_FOR_MATCH ? std::min(expectedConcurrentEvals,getNumCores()) :
-      setupFor == SETUP_FOR_ANALYSIS ? std::min(expectedConcurrentEvals,getNumCores()) :
-      setupFor == SETUP_FOR_GTP ? expectedConcurrentEvals :
-      setupFor == SETUP_FOR_BENCHMARK ? expectedConcurrentEvals :
-      cfg.getInt("numEigenThreadsPerModel",1,1024);
+      computeDefaultEigenBackendThreads(expectedConcurrentEvals,logger);
 #endif
 
     vector<int> gpuIdxByServerThread;
@@ -306,7 +293,7 @@ vector<NNEvaluator*> Setup::initializeNNEvaluators(
     //and doesn't greatly benefit from having a bigger chunk of parallelizable work to do on the large scale.
     //So we just fix a size here that isn't crazy and saves memory, completely ignore what the user would have
     //specified for GPUs.
-    int nnMaxBatchSize = 4;
+    int nnMaxBatchSize = 2;
     cfg.markAllKeysUsedWithPrefix("nnMaxBatchSize");
     (void)defaultMaxBatchSize;
 #endif
@@ -349,6 +336,18 @@ vector<NNEvaluator*> Setup::initializeNNEvaluators(
   return nnEvals;
 }
 
+int Setup::computeDefaultEigenBackendThreads(int expectedConcurrentEvals, Logger& logger) {
+  auto getNumCores = [&logger]() {
+    int numCores = (int)std::thread::hardware_concurrency();
+    if(numCores <= 0) {
+      logger.write("Could not determine number of cores on this machine, choosing eigen backend threads as if it were 8");
+      numCores = 8;
+    }
+    return numCores;
+  };
+  return std::min(expectedConcurrentEvals,getNumCores());
+}
+
 string Setup::loadHomeDataDirOverride(
   ConfigParser& cfg
 ){

cpp/program/setup.h

@@ -61,6 +61,8 @@ namespace Setup {
   constexpr double DEFAULT_ANALYSIS_WIDE_ROOT_NOISE = 0.04;
   constexpr bool DEFAULT_ANALYSIS_IGNORE_PRE_ROOT_HISTORY = true;
 
+  int computeDefaultEigenBackendThreads(int expectedConcurrentEvals, Logger& logger);
+
   //Loads search parameters for bot from config, by bot idx.
   //Fails if no parameters are found.
   std::vector<SearchParams> loadParams(