Decreasing Build times

[dark_sylinc]

I just pushed a new branch v2-2-build-time-opt.

The goal is to try to reduce compilation times on all platforms.

I've been using -ftime-trace; and although it says I've performed impressive optimizations; overall build times have not drastically decreased (includes buildings samples):

Code: Select all

Before Clang9 C++14
real	2m8,344s
user	15m53,879s
sys	0m50,416s

After Clang9 C++14
real	1m58,393s
user	14m40,097s
sys	0m47,608s

Before GCC C++98
real	2m2,504s
user	14m32,773s
sys	1m2,664s

After GCC C++98
real	1m57,145s
user	13m54,947s
sys	1m1,055s

Before GCC C++14
real	3m59,728s
user	29m34,520s
sys	1m39,011s

After GCC C++14
real	3m40,869s
user	27m11,995s
sys	1m30,205s

Note however OgrePrerequisites.h is much lighter (and so is OgreVector3.h etc); so it is quite possible this branch improves build for users of Ogre; and not just building the library.

I wasn't able to benchmark Visual Studio yet (but it compiles fine).

According to time trace, parsing <string> takes a lot of time, so that may benefit from using precompiled headers on all platforms (right now we're only using precompiled headers in Windows)

OK so what?
If you've got further ideas to improve build times, please bring them on here or in the ticket

If you can try the new branch, then please do! It should be stable as I only moved headers around and used a few forward declaration tricks.
But if you have Ogre as a dependency it may break with silly missing header errors.

Let me know what you encounter!

[TaaTT4]

Everything good on my side. I've tested the latest commit (c3752d1) in v2-2-build-time-opt branch and the only drawbacks I encountered are:

• Include <sstream> where needed
• Replace snippets of code similar to the one here below:

  Code: Select all

  const NameValuePairList windowParameters
  {
  	{"externalWindowHandle", to_string(reinterpret_cast<uintptr_t>(windowInfo.info.win.window))},
  	{"gamma", renderSystem->getConfigOptions()["sRGB Gamma Conversion"].currentValue},
  	{"vsync", renderSystem->getConfigOptions()["VSync"].currentValue},
  	{"vsyncInterval", renderSystem->getConfigOptions()["VSync Interval"].currentValue}
  };
  

  With:

  Code: Select all

  NameValuePairList windowParameters;
  windowParameters.insert_or_assign("externalWindowHandle", to_string(reinterpret_cast<uintptr_t>(windowInfo.info.win.window)));
  windowParameters.insert_or_assign("gamma", renderSystem->getConfigOptions()["sRGB Gamma Conversion"].currentValue);
  windowParameters.insert_or_assign("vsync", renderSystem->getConfigOptions()["VSync"].currentValue);
  windowParameters.insert_or_assign("vsyncInterval", renderSystem->getConfigOptions()["VSync Interval"].currentValue);
  

  Otherwise I get the following error: 'initializing': cannot convert from 'initializer list' to 'std::map<K,V,P,A>'

Platform Toolset: Visual Studio 2019 (v142)
C++ Language Standard: ISO C++17 Standard (std:c++17)

[dark_sylinc]

[*]Replace snippets of code similar to the one here below:

Code: Select all

const NameValuePairList windowParameters
{
	{"externalWindowHandle", to_string(reinterpret_cast<uintptr_t>(windowInfo.info.win.window))},
	{"gamma", renderSystem->getConfigOptions()["sRGB Gamma Conversion"].currentValue},
	{"vsync", renderSystem->getConfigOptions()["VSync"].currentValue},
	{"vsyncInterval", renderSystem->getConfigOptions()["VSync Interval"].currentValue}
};

With:

Code: Select all

NameValuePairList windowParameters;
windowParameters.insert_or_assign("externalWindowHandle", to_string(reinterpret_cast<uintptr_t>(windowInfo.info.win.window)));
windowParameters.insert_or_assign("gamma", renderSystem->getConfigOptions()["sRGB Gamma Conversion"].currentValue);
windowParameters.insert_or_assign("vsync", renderSystem->getConfigOptions()["VSync"].currentValue);
windowParameters.insert_or_assign("vsyncInterval", renderSystem->getConfigOptions()["VSync Interval"].currentValue);

That last part is definitely not intended behavior and hopefully fixable.
I'll check what went wrong.

[TaaTT4]

That last part is definitely not intended behavior and hopefully fixable.
I'll check what went wrong.

Commit 7e43c27 fixed the issue, thanks.

[dark_sylinc]

New two commits after new data appeared.

These two changes weren't as intrusive, hence I'm not making a separate branch this time. New timings:

Code: Select all

Clang 9 C++14
real	1m55,995s
user	14m27,704s
sys	0m46,568s

GCC C++14
real	3m33,953s
user	26m35,735s
sys	1m27,774s

GCC C++98
real	1m49,469s
user	13m22,307s
sys	0m57,774s

Btw Clang C++98 is the fastest at only 1m16,497s (1 minute and 16 seconds!!! including samples!) but there's a small issue in Viewport.cpp that prevents compiling with Clang c++98 (I workarounded the issue to do the testing)

[chchwy]

May I ask what "real", "user", "sys" actually mean?

[dark_sylinc]

Real: the one that matters. Real world times
User: time spent in user space, in all cores (threads?), hence why it's larger than 'real' time. I have a 4 core with 8 threads.
Sys: time spent in kernel space (e.g. IO, thread scheduling, etc)

[xrgo]

its been a while since I haven't updated Ogre in my engine, and I just did a clean built like I always do and its definitely faster, maybe 1/3 of what I remembered, great job!! (win10, VS2017)

Saludos!!