Parallel algorithms POC

CommonCompilerConfig.cmake

@@ -100,8 +100,9 @@ set(boost_required ON)
 if(NOT FORCE_CPP11)
   if(${CMAKE_CXX_COMPILER_ID} STREQUAL "GNU")
     if(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL "9")
-      set(CMAKE_CXX_STANDARD 17)
-      set(boost_required OFF)
+	 set(extra_lib_for_filesystem "tbb") #TODO hijacked for parallel TS, link with tbb for g++-9
+         set(CMAKE_CXX_STANDARD 17)
+	 set(boost_required OFF)
     elseif(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL "8")
       set(CMAKE_CXX_STANDARD 17)
       set(extra_lib_for_filesystem "stdc++fs")

src/CMakeLists.txt

@@ -171,6 +171,7 @@ set(regions_files
 
 set(types_files
     htm/types/Exception.hpp
+    htm/types/Parallelizable.hpp
     htm/types/Types.hpp
     htm/types/Serializable.hpp
     htm/types/Sdr.hpp

src/htm/algorithms/SpatialPooler.cpp

@@ -27,6 +27,7 @@
 #include <htm/algorithms/SpatialPooler.hpp>
 #include <htm/utils/Topology.hpp>
 #include <htm/utils/VectorHelpers.hpp>
+#include <htm/types/Parallelizable.hpp>
 
 using namespace std;
 using namespace htm;
@@ -863,19 +864,25 @@ void SpatialPooler::inhibitColumnsGlobal_(const vector<Real> &overlaps,
   // faster than a regular sort because it stops after it partitions the
   // elements about the Nth element, with all elements on their correct side of
   // the Nth element.
-  std::nth_element(
+  tNth.start();
+  std::nth_element(htm::parallel::mode,
     activeColumns.begin(),
     activeColumns.begin() + numDesired,
     activeColumns.end(),
     compare);
   // Remove the columns which lost the competition.
   activeColumns.resize(numDesired);
+  tNth.stop();
   // Finish sorting the winner columns by their overlap.
+  tSort.start();
   std::sort(activeColumns.begin(), activeColumns.end(), compare);
+  tSort.stop();
   // Remove sub-threshold winners
+  tWhile.start();
   while( !activeColumns.empty() &&
          overlaps[activeColumns.back()] < stimulusThreshold_)
       activeColumns.pop_back();
+  tWhile.stop();
 }
 
 

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 {
 
@@ -1212,6 +1212,7 @@ class SpatialPooler : public Serializable
   Random rng_;
 
 public:
+  mutable Timer tSort, tNth, tWhile;
   const Connections& connections = connections_; //for inspection of details in connections. Const, so users cannot break the SP internals.
   const Connections& getConnections() const { return connections_; } // as above, but for use in pybind11
 };

src/htm/types/Parallelizable.hpp

@@ -0,0 +1,29 @@
+/** 
+ * Parallelizable.hpp
+ *
+ * include this header to files where you want to run blocks of code in parallel. 
+ *  We use [C++17 standard Parallel TS](https://en.cppreference.com/w/cpp/experimental/parallelism). 
+ *
+ * Requirements: 
+ *  - c++17
+ *  - The Building Blocks (tbb) linked to the library
+ *  - [supported compiler](https://en.cppreference.com/w/cpp/compiler_support#cpp17): currently GCC-9+, MSVC 2019
+ * //TODO: switch to c++17 by default, or implement `transform()` temporarily as a custom method? 
+ *
+ * Functionality:
+ *  - include all needed headers for given platform, compiler, ... 
+ *  - handle define `NUM_PARALLEL=n`
+ *    - to run in single thread, set NUM_PARALLEL=1
+ *
+ */  
+
+//includes for TS Parallel
+#include <execution> //std::execution::par, seq, par_unseq
+// #include <tbb/parallel_for.h> 
+#include <mutex>
+
+namespace htm {
+namespace parallel {
+  const constexpr auto mode = std::execution::par_unseq; //TODO ifdef NUM_PARALLEL=1 -> seq
+}
+}