Merge pull request #564 from htm-community/sp_opt

Connections: microoptimizations

Author: ctrl-z-9000-times

src/examples/hotgym/HelloSPTP.cpp

@@ -158,6 +158,9 @@ Real64 BenchmarkHotgym::run(UInt EPOCHS, bool useSPlocal, bool useSPglobal, bool
       cout << "Init:\t" << tInit.getElapsed() << endl;
       cout << "Random:\t" << tRng.getElapsed() << endl;
       cout << "Encode:\t" << tEnc.getElapsed() << endl;
+      cout << "Inh " << spGlobal.tInh.getElapsed() << endl;
+      cout << "Overlaps " << spGlobal.tOverlap.getElapsed() << endl; 
+
       if(useSPlocal)  cout << "SP (l):\t" << tSPloc.getElapsed()*1.0f  << endl;
       if(useSPglobal) cout << "SP (g):\t" << tSPglob.getElapsed() << endl;
       if(useTM) cout << "TM:\t" << tTM.getElapsed() << endl;

src/examples/mnist/MNIST_SP.cpp

@@ -155,7 +155,9 @@ void train(const bool skipSP=false) {
   }
 
   tTrain.stop();
-  cout << "MNIST train time: " << tTrain.getElapsed() << endl; 
+  cout << "MNIST train time: " << tTrain.getElapsed() << endl;
+  cout << "inh " << sp.tInh.getElapsed() << endl;
+  cout << "over " << sp.tOverlap.getElapsed() << endl; 
 
   // Save the connections to file for postmortem analysis.
   ofstream dump("mnist_sp_learned.connections", ofstream::binary | ofstream::trunc | ofstream::out);

src/htm/algorithms/Connections.cpp

@@ -268,9 +268,11 @@ void Connections::updateSynapsePermanence(Synapse synapse,
 
   const bool before = synData.permanence >= connectedThreshold_;
   const bool after  = permanence         >= connectedThreshold_;
+
+  // update the permanence
   synData.permanence = permanence;
 
-  if( before == after ) { //no change
+  if( before == after ) { //no change in dis/connected status
       return;
   }
     const auto &presyn    = synData.presynapticCell;
@@ -371,17 +373,17 @@ bool Connections::compareSegments(const Segment a, const Segment b) const {
   }
 }
 
-vector<Synapse>
-Connections::synapsesForPresynapticCell(CellIdx presynapticCell) const {
-  vector<Synapse> all(
-      potentialSynapsesForPresynapticCell_.at(presynapticCell).begin(),
-      potentialSynapsesForPresynapticCell_.at(presynapticCell).end());
-  all.insert( all.end(),
-      connectedSynapsesForPresynapticCell_.at(presynapticCell).begin(),
-      connectedSynapsesForPresynapticCell_.at(presynapticCell).end());
+
+vector<Synapse> Connections::synapsesForPresynapticCell(const CellIdx presynapticCell) const {
+  const auto& potential = potentialSynapsesForPresynapticCell_.at(presynapticCell);
+  const auto& connected = connectedSynapsesForPresynapticCell_.at(presynapticCell);
+
+  vector<Synapse> all( potential.cbegin(), potential.cend());
+  all.insert( all.cend(), connected.cbegin(), connected.cend());
   return all;
 }
 
+
 void Connections::reset()
 {
   if( not timeseries_ ) {

src/htm/algorithms/Connections.hpp

@@ -42,8 +42,8 @@ using SynapseIdx= UInt16; /** Index of synapse in segment. */
 using Segment   = UInt32;    /** Index of segment's data. */
 using Synapse   = UInt32;    /** Index of synapse's data. */
 using Permanence= Real32; //TODO experiment with half aka float16
-const Permanence minPermanence = 0.0f;
-const Permanence maxPermanence = 1.0f;
+constexpr const Permanence minPermanence = 0.0f;
+constexpr const Permanence maxPermanence = 1.0f;
 
 
 /**
@@ -651,10 +651,13 @@ class Connections : public Serializable
   Permanence               connectedThreshold_; //TODO make const
 
   // Extra bookkeeping for faster computing of segment activity.
-  std::unordered_map<CellIdx, std::vector<Synapse>> potentialSynapsesForPresynapticCell_;
-  std::unordered_map<CellIdx, std::vector<Synapse>> connectedSynapsesForPresynapticCell_;
-  std::map<CellIdx, std::vector<Segment>> potentialSegmentsForPresynapticCell_;
-  std::map<CellIdx, std::vector<Segment>> connectedSegmentsForPresynapticCell_;
+ 
+  struct identity { constexpr size_t operator()( const CellIdx t ) const noexcept { return t; };   };	//TODO in c++20 use std::identity 
+
+  std::unordered_map<CellIdx, std::vector<Synapse>, identity> potentialSynapsesForPresynapticCell_;
+  std::unordered_map<CellIdx, std::vector<Synapse>, identity> connectedSynapsesForPresynapticCell_;
+  std::unordered_map<CellIdx, std::vector<Segment>, identity> potentialSegmentsForPresynapticCell_;
+  std::unordered_map<CellIdx, std::vector<Segment>, identity> connectedSegmentsForPresynapticCell_;
 
   std::vector<Segment> segmentOrdinals_;
   std::vector<Synapse> synapseOrdinals_;

src/htm/algorithms/SpatialPooler.cpp

@@ -463,17 +463,22 @@ void SpatialPooler::compute(const SDR &input, const bool learn, SDR &active) {
   input.reshape(  inputDimensions_ );
   active.reshape( columnDimensions_ );
   updateBookeepingVars_(learn);
+
+  tOverlap.start();
   calculateOverlap_(input, overlaps_);
+  tOverlap.stop();
 
   boostOverlaps_(overlaps_, boostedOverlaps_);
 
+  tInh.start();
   auto &activeVector = active.getSparse();
   inhibitColumns_(boostedOverlaps_, activeVector);
   // Notify the active SDR that its internal data vector has changed.  Always
   // call SDR's setter methods even if when modifying the SDR's own data
   // inplace.
   sort( activeVector.begin(), activeVector.end() );
   active.setSparse( activeVector );
+  tInh.stop();
 
   if (learn) {
     adaptSynapses_(input, active);

src/htm/algorithms/SpatialPooler.hpp

@@ -29,7 +29,7 @@
 #include <htm/types/Types.hpp>
 #include <htm/types/Serializable.hpp>
 #include <htm/types/Sdr.hpp>
-
+#include <htm/os/Timer.hpp> 
 
 namespace htm {
 
@@ -61,6 +61,7 @@ using namespace std;
 class SpatialPooler : public Serializable
 {
 public:
+  Timer tInh, tOverlap;
 
   const Real MAX_LOCALAREADENSITY = 0.5f; //require atleast 2 areas