[glob2-devel] glob2 autovectorization using g++ 4.0.1

[Stephane Magnenat]

Hi,

Glob2 got vectorized when using g++- 4.0.1 and the following command line 
options on x86:
-ftree-vectorize -march=pentium4 -O3 -ftree-vectorizer-verbose=1


Result of autovectorization:

Cmd line to produce this output was:
grep "LOOP VECTORIZED\|vectorized .* loops" g2dump | grep -v "vectorized 0 
loops"
where g2dump is generated by:
make >& /tmp/g2dump
with tcsh in glob2's directory

/usr/lib/gcc/i486-linux-gnu/4.0.1/../../../../include/c++/4.0.1/bits/valarray_array.h:127:
 
note: LOOP VECTORIZED.
GraphicContext.cpp:273: note: vectorized 1 loops in function.
GraphicContext.cpp:556: note: LOOP VECTORIZED.
GraphicContext.cpp:524: note: vectorized 1 loops in function.
GraphicContext.cpp:707: note: LOOP VECTORIZED.
GraphicContext.cpp:670: note: vectorized 1 loops in function.
/usr/lib/gcc/i486-linux-gnu/4.0.1/../../../../include/c++/4.0.1/bits/stl_algobase.h:542:
 
note: LOOP VECTORIZED.
StreamBackend.cpp:40: note: vectorized 1 loops in function.
AINumbi.cpp:50: note: LOOP VECTORIZED.
AINumbi.cpp:41: note: vectorized 1 loops in function.
AICastor.cpp:135: note: LOOP VECTORIZED.
AICastor.cpp:122: note: vectorized 1 loops in function.
AICastor.cpp:1903: note: LOOP VECTORIZED.
AICastor.cpp:1899: note: vectorized 1 loops in function.
AICastor.cpp:135: note: LOOP VECTORIZED.
AICastor.cpp:149: note: vectorized 1 loops in function.
AICastor.cpp:135: note: LOOP VECTORIZED.
AICastor.cpp:149: note: vectorized 1 loops in function.
AICastor.cpp:135: note: LOOP VECTORIZED.
AICastor.cpp:158: note: vectorized 1 loops in function.
AICastor.cpp:135: note: LOOP VECTORIZED.
AICastor.cpp:158: note: vectorized 1 loops in function.
AICastor.cpp:1903: note: LOOP VECTORIZED.
AICastor.cpp:1903: note: LOOP VECTORIZED.
AICastor.cpp:403: note: vectorized 2 loops in function.
Building.cpp:263: note: LOOP VECTORIZED.
Building.cpp:157: note: vectorized 1 loops in function.
Building.cpp:95: note: LOOP VECTORIZED.
Building.cpp:138: note: LOOP VECTORIZED.
Building.cpp:46: note: vectorized 2 loops in function.
Building.cpp:95: note: LOOP VECTORIZED.
Building.cpp:138: note: LOOP VECTORIZED.
Building.cpp:46: note: vectorized 2 loops in function.
Engine.cpp:51: note: LOOP VECTORIZED.
Engine.cpp:48: note: vectorized 1 loops in function.
Engine.cpp:51: note: LOOP VECTORIZED.
Engine.cpp:48: note: vectorized 1 loops in function.
GUIMapPreview.cpp:152: note: LOOP VECTORIZED.
GUIMapPreview.cpp:63: note: vectorized 1 loops in function.
/usr/lib/gcc/i486-linux-gnu/4.0.1/../../../../include/c++/4.0.1/bits/stl_algobase.h:542:
 
note: LOOP VECTORIZED.
Game.cpp:1150: note: vectorized 1 loops in function.
Game.cpp:135: note: LOOP VECTORIZED.
Game.cpp:97: note: vectorized 1 loops in function.
GameGUI.cpp:239: note: LOOP VECTORIZED.
GameGUI.cpp:174: note: vectorized 1 loops in function.
NetGame.cpp:176: note: LOOP VECTORIZED.
NetGame.cpp:179: note: LOOP VECTORIZED.
NetGame.cpp:167: note: vectorized 2 loops in function.
NetGame.cpp:176: note: LOOP VECTORIZED.
NetGame.cpp:179: note: LOOP VECTORIZED.
NetGame.cpp:111: note: LOOP VECTORIZED.
NetGame.cpp:120: note: LOOP VECTORIZED.
NetGame.cpp:160: note: LOOP VECTORIZED.
NetGame.cpp:99: note: vectorized 5 loops in function.
NetGame.cpp:1435: note: LOOP VECTORIZED.
NetGame.cpp:1448: note: LOOP VECTORIZED.
NetGame.cpp:1333: note: vectorized 2 loops in function.
NetGame.cpp:68: note: LOOP VECTORIZED.
NetGame.cpp:176: note: LOOP VECTORIZED.
NetGame.cpp:179: note: LOOP VECTORIZED.
NetGame.cpp:35: note: vectorized 3 loops in function.
NetGame.cpp:68: note: LOOP VECTORIZED.
NetGame.cpp:176: note: LOOP VECTORIZED.
NetGame.cpp:179: note: LOOP VECTORIZED.
NetGame.cpp:35: note: vectorized 3 loops in function.
SessionConnection.cpp:25: note: LOOP VECTORIZED.
SessionConnection.cpp:22: note: vectorized 1 loops in function.
SessionConnection.cpp:25: note: LOOP VECTORIZED.
SessionConnection.cpp:22: note: vectorized 1 loops in function.
Team.cpp:175: note: LOOP VECTORIZED.
Team.cpp:178: note: LOOP VECTORIZED.
Team.cpp:173: note: vectorized 2 loops in function.
Team.cpp:175: note: LOOP VECTORIZED.
Team.cpp:178: note: LOOP VECTORIZED.
Team.cpp:143: note: vectorized 2 loops in function.
Team.cpp:175: note: LOOP VECTORIZED.
Team.cpp:178: note: LOOP VECTORIZED.
Team.cpp:143: note: vectorized 2 loops in function.
Team.cpp:175: note: LOOP VECTORIZED.
Team.cpp:178: note: LOOP VECTORIZED.
Team.cpp:153: note: vectorized 2 loops in function.
Team.cpp:175: note: LOOP VECTORIZED.
Team.cpp:178: note: LOOP VECTORIZED.
Team.cpp:153: note: vectorized 2 loops in function.
UnitType.cpp:25: note: LOOP VECTORIZED.
UnitType.cpp:23: note: vectorized 1 loops in function.
UnitType.cpp:93: note: LOOP VECTORIZED.
UnitType.cpp:91: note: vectorized 1 loops in function.
UnitType.cpp:107: note: LOOP VECTORIZED.
UnitType.cpp:105: note: vectorized 1 loops in function.
UnitType.cpp:25: note: LOOP VECTORIZED.
UnitType.cpp:36: note: vectorized 1 loops in function.
MultiplayersHost.cpp:240: note: LOOP VECTORIZED.
MultiplayersHost.cpp:238: note: vectorized 1 loops in function.
MultiplayersHost.cpp:240: note: LOOP VECTORIZED.
MultiplayersHost.cpp:632: note: vectorized 1 loops in function.
MultiplayersHost.cpp:240: note: LOOP VECTORIZED.
MultiplayersHost.cpp:381: note: vectorized 1 loops in function.
MultiplayersHost.cpp:240: note: LOOP VECTORIZED.
MultiplayersHost.cpp:1323: note: vectorized 1 loops in function.
MultiplayersHost.cpp:240: note: LOOP VECTORIZED.
MultiplayersHost.cpp:868: note: vectorized 1 loops in function.
MultiplayersHost.cpp:240: note: LOOP VECTORIZED.
MultiplayersHost.cpp:35: note: vectorized 1 loops in function.
MultiplayersHost.cpp:240: note: LOOP VECTORIZED.
MultiplayersHost.cpp:35: note: vectorized 1 loops in function.
MultiplayersCrossConnectable.cpp:382: note: LOOP VECTORIZED.
MultiplayersCrossConnectable.cpp:354: note: LOOP VECTORIZED.
MultiplayersCrossConnectable.cpp:330: note: LOOP VECTORIZED.
MultiplayersCrossConnectable.cpp:304: note: vectorized 3 loops in function.
MapGenerationDescriptor.cpp:42: note: LOOP VECTORIZED.
MapGenerationDescriptor.cpp:27: note: vectorized 1 loops in function.
MapGenerationDescriptor.cpp:42: note: LOOP VECTORIZED.
MapGenerationDescriptor.cpp:27: note: vectorized 1 loops in function.
EndGameScreen.cpp:183: note: LOOP VECTORIZED.
EndGameScreen.cpp:107: note: vectorized 1 loops in function.
EndGameScreen.cpp:183: note: LOOP VECTORIZED.
EndGameScreen.cpp:107: note: vectorized 1 loops in function.
/usr/lib/gcc/i486-linux-gnu/4.0.1/../../../../include/c++/4.0.1/bits/valarray_array.h:127:
 
note: LOOP VECTORIZED.
BitArray.cpp:88: note: vectorized 1 loops in function.
/usr/lib/gcc/i486-linux-gnu/4.0.1/../../../../include/c++/4.0.1/bits/valarray_array.h:127:
 
note: LOOP VECTORIZED.
/usr/lib/gcc/i486-linux-gnu/4.0.1/../../../../include/c++/4.0.1/bits/valarray_array.h:127:
 
note: LOOP VECTORIZED.
BitArray.cpp:42: note: vectorized 2 loops in function.


Statistics:

The number of loops vectorized has been counted by:
grep "LOOP VECTORIZED" g2dump | wc -l
The number of loop not vectorized has been counted by:
grep "vectorized 0 loops" g2dump | wc -l

72 loops vectorized
879 loops that could not have been vectorized for some reason


Analysis

As you can see, there is a lot of loops that could be, but have not been 
vectorized because of missalignements and so on. g++ gives very detailed 
explanations why loops have not been vectorized (providing a debug level 
-ftree-vectorizer-verbose greater than one or the use of -fdump-tree-vect to 
dump all vectorized infos in SOURCEFILE.vect). So there is plenty of room for 
optimisation there, I suggest anyway to wait for g++ 4.1 and test on 
different plateforms before modifying any code to optimise the 
autovectorization.


Conclusion

Right now g++ 4 is able to autovectorized array filling with constant value 
and array copy, if alignement matches. It doesn't seem to be yet able to 
vectorize loops where complex operations take place (i.e. image blitting with 
alpha blending), but the structure is in place and we can expect much better 
result with g++ 4.1. Anyway, the autovectorizer output can help you to 
optimize your programs.

Have fun,

Steph

P.S.            Compilation went faster than with g++-3.X, but I don't have 
numbers

-- 
http://nct.ysagoon.com