performance degrading bug fixes and further speed optimization

CHANGES

@@ -5,3 +5,10 @@
 0.3 (Dec 2016)
   Fixed several bugs that affect prediction performance (especially for small datasets).
 
+0.4 (Aug 2017)
+  Fixed bug which truncated negative float values to numeric_limits<float>::min(),  causing degration in prediction performance for datasets with negative values; changed truncation to numeric_limits<float>::lowest()
+
+0.5 (Sept 2017)
+  Added openmp support and multi-threading for discretization
+  Added loop unrolling and compilation option for simd optimization
+

CMakeLists.txt

@@ -4,13 +4,37 @@ cmake_minimum_required (VERSION 2.8.0)
 
 project (FastRGF) 
 
-set(CMAKE_CXX_FLAGS "-O3 -std=c++11")
+# whether to use openmp (default is on)
+option(OPENMP "Use openmp for multi-threads" ON)
 
-# you may need to use the following for g++-4.8
-#set(CMAKE_CXX_FLAGS "-O3 -std=c++11 -pthread")
 
+set(CMAKE_CXX_FLAGS "-O3 -std=c++11")
 #set(CMAKE_CXX_FLAGS "-g -std=c++11 -Wall")
 
+if(OPENMP)
+ message("use openmp for multi-threads")
+ # use openmp
+ add_definitions("-DUSE_OMP")		
+
+  if(${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
+    #  mac os x: clang does not support openmp, need to install a g++  with openmp support
+    set(CMAKE_CXX_COMPILER "/usr/local/bin/g++-7")
+    set (CMAKE_CXX_COMPILER_ID "GNU")
+  endif()
+
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fopenmp")
+else()
+ message("use standard c++11 thread library")
+endif()
+
+if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
+   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread -ftree-vectorize -ffast-math")
+endif()		
+message("c++  compiler= " ${CMAKE_CXX_COMPILER})
+message("c++  options= " ${CMAKE_CXX_FLAGS})
+get_directory_property( cDirDefs DIRECTORY ${CMAKE_SOURCE_DIR} COMPILE_DEFINITIONS )
+message("c++  definitions= " ${cDirDefs})
+
 include_directories(include)
 
 

README.md

@@ -25,12 +25,6 @@ The implementation of greedy tree node optimization employs second order Newton
 Please see the file [CHANGES](CHANGES) for version information.
 The software is written in c++11, and it has been tested under linux and macos, and it may require g++ version 4.8 or above and cmake version 2.8 or above. 
 
- If you use *g++-4.8*, after running the exmaples, you may get error messages similar to the following:
-
-    terminate called after throwing an instance of 'std::system_error'
-    what():  Enable multithreading to use std::thread: Operation not permitted
-
- If this occurs, you need to add the **-pthread** flag in [CMakeLists.txt](CMakeLists.txt) to the variable CMAKE_CXX_FLAGS in order to enable multi-threading. This problem seems to be a bug in the g++ compiler. There may be variations of this problem specific to your system that require different fixes.
 
  To install the binaries, unpackage the software into a directory.
 

include/header.h

@@ -30,15 +30,22 @@
 
 #include <cstring>
 
+#ifdef USE_OMP
+#include "omp.h"
+#endif
 
 namespace rgf {
 
 
-#define VER "version 0.2 (August 2016) by Tong Zhang"
+#define VER "version 0.4 (Aug 2017) by Tong Zhang"
+
 
   const int max_thrds=128;
 
 
+  const int vect_width=8;
+
+
   using int_t=long;
 
 

include/utils.h

@@ -12,6 +12,7 @@
 
 #include "header.h"
 
+
 namespace rgf {
 
   class MyIO {
@@ -241,13 +242,26 @@ class Timer {
 	mr.reduce(0);
 	return;
       }
+      static const bool use_omp=true;
+
+#ifndef USE_OMP
       for (tid=0; tid<nthreads; tid++) {
 	_th[tid]= thread(& MapReduceRunner::single_thread_map_reduce<T>, this,
 			 std::ref(mr),begin, end, tid, nthreads,run_range);
       }
+#else
+      omp_set_num_threads(nthreads);
+#pragma omp parallel for
+      for (tid=0; tid<nthreads; tid++) {
+	single_thread_map_reduce<T>(std::ref(mr),begin,end,tid,nthreads,run_range);
+      }
+#endif
+
       mr.master();
       for (tid=0; tid<nthreads; tid++) {
+#ifndef USE_OMP
 	_th[tid].join();
+#endif
 	mr.reduce(tid);
       }
     }

src/base/data.cpp

@@ -55,7 +55,7 @@ class MyDataInputLineParseResult {
 
     long tmp=atol(token_str);
     if (tmp >= numeric_limits<i_t>::max() ||
-	tmp< numeric_limits<i_t>::min()) {
+	tmp< numeric_limits<i_t>::lowest()) {
       throw MyDataInputException(" : index out of range",lno);      
     }
     result.index=tmp;
@@ -64,13 +64,14 @@ class MyDataInputLineParseResult {
     if (is_same<v_t,float>::value ||is_same<v_t,double>::value) {
       double tmp= atof(token_str + (sep+1));
       if (tmp>= numeric_limits<float>::max()) tmp=numeric_limits<float>::max();
-      if (tmp<= numeric_limits<float>::min()) tmp=numeric_limits<float>::min();
+      if (tmp<= numeric_limits<float>::lowest()) tmp=numeric_limits<float>::lowest();
+      if (!(tmp == tmp)) tmp=numeric_limits<float>::lowest(); 
       result.value=(float)tmp;
     }
     else {
       tmp=atol(token_str+ (sep+1));
       if (tmp >= numeric_limits<v_t>::max() ||
-	  tmp< numeric_limits<v_t>::min()) {
+	  tmp< numeric_limits<v_t>::lowest()) {
       throw MyDataInputException(" : value out of range",lno);      
       }
       result.value=tmp;
@@ -93,7 +94,8 @@ class MyDataInputLineParseResult {
 
 	  double tmp=stod(token);
 	  if (tmp>= numeric_limits<float>::max()) tmp=numeric_limits<float>::max();
-	  if (tmp<= numeric_limits<float>::min()) tmp=numeric_limits<float>::min();
+	  if (tmp<= numeric_limits<float>::lowest()) tmp=numeric_limits<float>::lowest();
+	  if (!(tmp == tmp)) tmp=numeric_limits<float>::lowest(); 
 	  feats_dense.push_back((float)tmp);
 	  continue;
 	}