Author Topic: foperf  (Read 4730 times)

JovankaB

  • Guest
foperf
« on: August 15, 2014, 02:48:25 pm »
Foperf is a simple script allowing you to compare performance of snippets of FOnline code.

Download

Get it from foperf repository on GitHub.

How to use

You can check examples in the tests directory of the repository. Below there is an explanation step by step how to create your own test:

1. First define FOPERF_TESTCASE - a list of pointers to snippet functions:
Code: [Select]
#define FOPERF_TESTCASE @test1, @test2
2. Then include the foperf.fos script:
Code: [Select]
#include "foperf.fos"
3. Write the snippet functions (functions with the snippets of the code which you want to test). For example to compare string::toFloat and StrToFloat:
Code: [Select]
   string test1()
   {
      string s = "3.1415";
      float f = 0;
      FOPERF_START;
      f = s.toFloat();
      FOPERF_STOP;
      return "string::toFloat";
   }
   
    string test2()
   {
      string s = "3.1415";
      float f = 0;
      FOPERF_START;
      StrToFloat(s, f);
      FOPERF_STOP;
      return "StrToFloat";
   }

Inside the snippet functions, place the tested code between the macros FOPERF_START and FOPERF_STOP. You should use the macros only once per snippet function. If you need some preparation code, put it before FOPERF_START. Finally return a string with the name of the snippet (the name can be anything, it's used to display the results in a friendly manner).

4. Use ASCompiler with `-run foperf::main` option to run the test. For example let's say the testcase script is saved as "tofloat.fos":
Code: [Select]
ASCompiler tofloat.fos -run foperf::main   
Here is the result part of the script output:
Code: [Select]
Result:
string::toFloat: 940 (BEST)
StrToFloat: 858 (91%)

The higher score the better.  According to foperf, StrToFloat is slower by *at least* 9% than string::toFloat.
 
Accuracy

The script is supposed to tell you which code is faster, and as long as the speed difference is more than 2-3%, it should tell the truth most of the time.

For snippets which take very little time to execute (under 2ms) the test is not very accurate, because within the macros there is a hidden loop code that is also measured. This generally benefits slower snippets by making them look less slow compared to the fastest one. For example if you compare following snippet functions:

Code: [Select]
string test3()
{
  FOPERF_START;
  int a = 1;
  FOPERF_STOP;
  return "one int";
}

string test4()
{
  FOPERF_START;
  int a = 1;
  int b = 2;
  FOPERF_STOP;
  return "two ints";
}

You may be surprised to see that the second snippet is only 5% slower, and the logic tells us that it should be around 50%. That's because the hidden loop code distorts the results for very small snippets. However, the result is still useful - it tells which code is the fastest. For code that takes more than 2ms to execute, the distortion of the result should be negligible.

« Last Edit: August 15, 2014, 05:01:49 pm by JovankaB »

JovankaB

  • Guest
Re: foperf
« Reply #1 on: August 15, 2014, 07:39:35 pm »
Here is a list of some of the things that I found about thanks to foperf, with links to the test scripts.

Tests done with ASCompiler revision-452. On Windows you can easily repeat the tests by drag&dropping the test scripts on foperf.bat (after you get the whole thing from repository).

  • string::toFloat is faster than StrToFloat by at least 9%.
    tofloat.fos
  • Operations on Unicode characters is much slower than on raw 8-bit characters. Even just using the index operator is slower than rawGet, and the larger the index, the slower it gets (so for example str[100] is slower than str[0]). This could get particularly nasty if a string was very long, for example if you load a 50KB text file and do a lot of basic string operations on it.
    unicode.fos
  • I thought it would be a silly test to do, but turns out ++i and i++ is at least 13% faster than i += 1.
    increment.fos
« Last Edit: August 16, 2014, 07:26:31 am by JovankaB »