AngelScript 1.9.1 WIP 2 (2004/09/14)
After finding out that Lua was about 2 times as fast as AngelScript I started thinking more about how to improve the performance. I already implemented the easiest optimizations which has improved the performance with about 30-50%. I have plenty of ideas for other changes that might improve the performance even more, and I'm certain that I will be able to make AS at least as fast as Lua. AngelScript has the advantage of being strongly typed and the ability to call C/C++ functions directly without wrapper functions. The ultimate optimization would be to compile directly to machine-code, but that is a step I'm not ready to take yet. That part will probably be done through a plug-in-able JIT compiler or something like that.
I'm planning on making these optimizations gradually over time, as I would otherwise have to stop further development of the library features, which is not something that I want to do.
Here are the changes for the latest WIP:
- bug fix: Switch-case didn't work correctly if the cases were not ordered in ascending order (thanks Dmitry "RCL" Rekman)
- bug fix: The ternary operator ?: didn't release temporary variables as it should, causing an assert failure in the debug version (thanks Dmitry "RCL" Rekman)
- bug fix: Passing string constants to condition operator would cause an assertion failure. (thanks Joe "Desdemona" Wright)
- Added support for asCALL_CDECL_OBJFIRST (thanks Tristan "Kurioes" Hartskeerl)
- Improved performance of byte code execution with about 30%
Regards,
Andreas Jönsson
Author of AngelScript
[Edited by - WitchLord on September 14, 2004 7:44:05 PM]
Hi,
I tried 1.9.1-WIP1 and I am pretty glad with the results. Performance gain is not just 30%-50% as you state. Right now the results are:
Lua: 20.499574 sec
AngelScript: 23.772902 sec
So it is nearly the same!!
I dunno why there is such a great performance boost. Older version performed nearly 2 times slower.
By the way, I did a mistake and AS version of performance test was not 100% identical to Lua version (there is one 'multiply' operation more each iteration, and the loop is one iteration less than Lua version). For tests to be identical, '*=' should be replaced with '=' in func3() for AS version and condition in 'for' loop should be changed from '<' to '<=' to match Lua's 'for'. I thought that could be the reason affecting performance, but even with the 100% identical tests, older version performs nearly 2 times as slow as Lua:
(Results for 1.9.0, 'tweaked' by me (ExecuteNext() rearranged, bcSize[] changed to enumerants), identical script )
Lua: 20.489861 sec
AngelScript: 36.730612 sec
Great work!
Cheers,
RCL
I tried 1.9.1-WIP1 and I am pretty glad with the results. Performance gain is not just 30%-50% as you state. Right now the results are:
Lua: 20.499574 sec
AngelScript: 23.772902 sec
So it is nearly the same!!
I dunno why there is such a great performance boost. Older version performed nearly 2 times slower.
By the way, I did a mistake and AS version of performance test was not 100% identical to Lua version (there is one 'multiply' operation more each iteration, and the loop is one iteration less than Lua version). For tests to be identical, '*=' should be replaced with '=' in func3() for AS version and condition in 'for' loop should be changed from '<' to '<=' to match Lua's 'for'. I thought that could be the reason affecting performance, but even with the 100% identical tests, older version performs nearly 2 times as slow as Lua:
(Results for 1.9.0, 'tweaked' by me (ExecuteNext() rearranged, bcSize[] changed to enumerants), identical script )
Lua: 20.489861 sec
AngelScript: 36.730612 sec
Great work!
Cheers,
RCL
Well, that is certainly great news. [smile] Already almost as fast as Lua, and I haven't even applied all the optimizations that I thought of. That was just the most simplest of them.
I don't know either why you get such an incredible performance improvement. When I measured the performance difference between 1.9.0 and 1.9.1 WIP1 it was only about 30%-50% faster. Could it be that you changed the compiler flags?
Wait a minute...
It's all about calculations:
1 - (23.77 / 36.73) = 0.35 (35% performance increase)
36.73 / 23.77 = 1.55 (1.9.0 is 54% slower than 1.9.1WIP1, or 65% as fast as 1.9.1WIP1)
23.77 / 20.50 = 1.16 (AS1.9.1WIP1 is 16% slower than Lua, or 86% as fast as Lua)
I'm not sure if the meaning of these calculations are actually what I think but I believe they are.
I don't know either why you get such an incredible performance improvement. When I measured the performance difference between 1.9.0 and 1.9.1 WIP1 it was only about 30%-50% faster. Could it be that you changed the compiler flags?
Wait a minute...
It's all about calculations:
1 - (23.77 / 36.73) = 0.35 (35% performance increase)
36.73 / 23.77 = 1.55 (1.9.0 is 54% slower than 1.9.1WIP1, or 65% as fast as 1.9.1WIP1)
23.77 / 20.50 = 1.16 (AS1.9.1WIP1 is 16% slower than Lua, or 86% as fast as Lua)
I'm not sure if the meaning of these calculations are actually what I think but I believe they are.
I've just uploaded the second WIP. The changes are:
No optimizations were made in this update.
- String constants are now concatenated by the compiler if they are only separated with whitespace or comments, i.e. "Hello " "Andreas" becomes "Hello Andreas"
- asCALL_CDECL_OBJFIRST should now work, which allows application to register global functions as object methods where the first parameter is the object pointer
- The ?: operator has been corrected once more due to bug reports that I received from Joe Wright
No optimizations were made in this update.
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement
