CPP : Parsing String Stream is too slow

Question

My cpp code needs to read a 7 MB text file of space separated float values. It's taking about 6 seconds to parse the string values into a float array, which is too much for my use case.

I've been checking online and people say it is usually the physical IO that takes time. To eliminate this, I'm reading the file into a stringstream in one shot and using this for the float parsing. Still no improvement in code speed. Any ideas how to get it to run faster ?

Here's my code (replaced the array entries with dummy_f for simplicity):

    #include "stdafx.h"
    #include <iostream>
    #include <fstream>
    #include "time.h"
    #include <sstream>
    using namespace std;

    int main()
    {
      ifstream testfile;
      string filename = "test_file.txt";
      testfile.open(filename.c_str());

      stringstream string_stream;
      string_stream << testfile.rdbuf();

      testfile.close();

      clock_t begin = clock();
      float dummy_f;

      cout<<"started stream at time "<<(double) (clock() - begin) /(double) CLOCKS_PER_SEC<<endl;

      for(int t = 0; t < 6375; t++)
      {

           string_stream >> dummy_f;

           for(int t1 = 0; t1 < 120; t1++)
           {
               string_stream >> dummy_f;
           }
      }

      cout<<"finished stream at time "<<(double) (clock() - begin) /(double) CLOCKS_PER_SEC<<endl;

      string_stream.str("");

      return 0;
     } 

Edit:

Here's a link to the test_cases.txt file https://drive.google.com/file/d/0BzHKbgLzf282N0NBamZ1VW5QeFE/view?usp=sharing

Please change the inner loop dimension to 128 when running with this file (made a typo)

Edit: Found a way to make it work. Declared dummy_f as string and read from the stringstream as a string word. Then used atof to convert the string into float. Time taken is 0.4 seconds which is good enough for me.

  string dummy_f;
  vector<float> my_vector;
  for(int t = 0; t < 6375; t++)
  {

       string_stream >> dummy_f;
       my_vector.push_back(atof(dummy_f.c_str()));
       for(int t1 = 0; t1 < 128; t1++)
       {
           string_stream >> dummy_f;
            my_vector.push_back(atof(dummy_f.c_str()));
       }
  }

Answer 1

update: Discussion in comments with @Mats concluded that locking overhead is unlikely to have anything to do with this, so we're back at square one as far as explaining why Visual C++'s library is so slow at parsing floats. Your sample test file looked like it was mostly numbers with magnitude not too far from 1.0, with nothing weird going on. (Intel's FPU in Sandybridge & later doesn't have a perf penalty for denormals anyway, according to Agner Fog's tables.)

As others have said, it's time to profile your code and find out which function is taking all the CPU time. Also, performance counters could tell you if branch mispredictions or cache misses are causing problems.

---

Every call to cin >> dummy_f requires locking to make sure another thread doesn't modify the input buffer at the same time. Reading 4 or 8 floats at once with scanf("%f%f%f%f", &dummy_array[0], &dummy_array[1], ...) would be a little bit more efficient, if that's where the bottleneck is. (scanf is not a great API for this either, since it needs the address of every array element as a function argument. Unrolling by using multiple conversions in one scanf is still a small performance win, though.)

You're trying to work around this with a stringstream, which may or may not be effective. It's a local variable in a function, so if the compiler can see all the functions and inline them, it can not bother with the locking. There can't be any other threads with access to this variable.

Answer 2

On my Linux machine it only takes <0.3 seconds, so if OP didn't make mistake on Debug/Release build, then it should be a problem unique to Windows:

hidden$ cat read-float.cpp 
#include <fstream>
#include <iostream>
#include <vector>
using namespace std;

int main() {
  ifstream fs("/tmp/xx.txt");
  vector<float> v;
  for (int i = 0; i < 6375; i++) {
    for (int j = 0; j < 129; j++) {
      float f;
      fs >> f;
      v.emplace_back(f);
    }
  }
  cout << "Read " << v.size() << " floats" << endl;
}
hidden$ g++ -std=c++11 read-float.cpp -O3
hidden$ time ./a.out 
Read 822375 floats

real    0m0.287s
user    0m0.279s
sys 0m0.008s

hidden$ g++ -v
Using built-in specs.
COLLECT_GCC=g++
COLLECT_LTO_WRAPPER=/usr/lib/gcc/x86_64-linux-gnu/4.8/lto-wrapper
Target: x86_64-linux-gnu
Configured with: ../src/configure -v --with-pkgversion='Ubuntu 4.8.4-2ubuntu1~14.04' --with-bugurl=file:///usr/share/doc/gcc-4.8/README.Bugs --enable-languages=c,c++,java,go,d,fortran,objc,obj-c++ --prefix=/usr --program-suffix=-4.8 --enable-shared --enable-linker-build-id --libexecdir=/usr/lib --without-included-gettext --enable-threads=posix --with-gxx-include-dir=/usr/include/c++/4.8 --libdir=/usr/lib --enable-nls --with-sysroot=/ --enable-clocale=gnu --enable-libstdcxx-debug --enable-libstdcxx-time=yes --enable-gnu-unique-object --disable-libmudflap --enable-plugin --with-system-zlib --disable-browser-plugin --enable-java-awt=gtk --enable-gtk-cairo --with-java-home=/usr/lib/jvm/java-1.5.0-gcj-4.8-amd64/jre --enable-java-home --with-jvm-root-dir=/usr/lib/jvm/java-1.5.0-gcj-4.8-amd64 --with-jvm-jar-dir=/usr/lib/jvm-exports/java-1.5.0-gcj-4.8-amd64 --with-arch-directory=amd64 --with-ecj-jar=/usr/share/java/eclipse-ecj.jar --enable-objc-gc --enable-multiarch --disable-werror --with-arch-32=i686 --with-abi=m64 --with-multilib-list=m32,m64,mx32 --with-tune=generic --enable-checking=release --build=x86_64-linux-gnu --host=x86_64-linux-gnu --target=x86_64-linux-gnu
Thread model: posix
gcc version 4.8.4 (Ubuntu 4.8.4-2ubuntu1~14.04) 

Answer 3

An alternate implementation using atof that runs 3 times faster is pasted below. On my laptop the original stringstream based one takes 2.3 seconds to complete whereas this one completes in under 0.8 seconds for same number of floats.

static char filecontents[10*1024*1024];

int testfun2()
{
  ifstream testfile;
  string filename = "test_file.txt";
  testfile.open(filename.c_str());
  int numfloats=0;
  testfile.read(filecontents,10*1024*1024);
  size_t numBytesRead = testfile.gcount();
  filecontents[numBytesRead]='\0';
  testfile.close();

  clock_t begin = clock();
  float dummy_f;

  cout<<endl<<"started at time "<<(double) (clock() - begin) /(double) CLOCKS_PER_SEC<<endl;

  char* p= filecontents;
  char* pend = p + numBytesRead;
  while(p<pend)
  {
      while(*p && (*p <= ' '))
      {
         ++p; //skip leading white space ,\r, \n
      }
      char* pvar = p;
      while(*p > ' ')
      {
        ++p; //skip over numbers
      }
      if(*p)
      {  *p = '\0';// shorter input makes atof faster.
        ++p;
      }
      if(*pvar)
      {
         dummy_f = atof(pvar);
         ++numfloats;
      }
      //cout << endl << dummy_f;
  }

  cout<<endl<< "finished at time "<<(double) (clock() - begin) /(double) CLOCKS_PER_SEC<<endl;

  cout << endl << "numfloats= " << numfloats;
  return numfloats;
 }