Author |
Topic: Letting GCC do the hard work (Read 1774 times) |
|
David Williams
Developer
member is offline
meh
Gender: 
Posts: 452
|
 |
Re: Letting GCC do the hard work
« Reply #7 on: Aug 17th, 2014, 08:13am » |
|
Probably my last post for a while...
1000 depth-sorted 'vector balls' based on graphics routines written in C (including the Shell Sort code I borrowed from Rosetta Code). I get 60 fps on my laptop, which is quite impressive, considering (well, considering that I don't yet really know what I'm doing with C):
www.bb4wgames.com/temp/vector_balls.zip [EXE; 142 Kb]
Update: Two versions of glib.dll compiled using different GCC optimisation settings (-O2 and -O3 plus -ffast-math). Also edited this post to include the image link below.
Screenshot:
www.bb4wgames.com/temp/vecballs.jpg
This kind of performance is convincing me that hybridizing BB4W and C/C++ code is the way to go (for me personally).
I'll include the BB4W part of the source below for curious people.
David.
---
Code:
*FLOAT 64
*ESC OFF
ON ERROR PROCerror
HIMEM = PAGE + 2*&100000
PROCFixWndSz : MODE 8 : OFF
INSTALL @lib$ + "GLIB"
PROCInitGLIB( @lib$ + "glib.dll", g{} )
ON ERROR PROCCleanup : PROCerror
ON CLOSE PROCCleanup : QUIT
GetTickCount = FN`s("GetTickCount")
LerpClr = FNImport("LerpClr")
Plot = FNImport("Plot")
InitExPoints = FNImport("InitExPointList")
ShellSort = FNImport("ShellSortExPointListZValues")
Rotate = FNImport("RotateExPoints")
MakeBallBitmap = FNImport("MakeBallBitmap1")
REM Create a 48x48 ball bitmap:
DIM ball% 4*(48*48 + 1)
ball%=ball% + 3 AND -4
SYS MakeBallBitmap, ball%, 48, &40AA20, &FF0020, 0.99*&10000, &10000
N% = 1000
DIM list{(N%-1) x#, y#, z#, x2#, y2#, z2#, key%, d0%, d1%, d2% }
listBaseAddr% = ^list{(0)}.x#
IF (listBaseAddr% AND 3) <> 0 THEN
PRINT '" The coordinates list base address is not DWORD-aligned!"
PRINT '" This may affect performance. Continuing in 5 seconds..."
WAIT 500
ENDIF
REM Define objects (balls) 3D (x,y,z) coordinates:
FOR I% = 0 TO N%-1
list{(I%)}.x# = (RND(1)-0.5) * 800.0
list{(I%)}.y# = (RND(1)-0.5) * 800.0
list{(I%)}.z# = (RND(1)-0.5) * 800.0
list{(I%)}.x2# = 0.0
list{(I%)}.y2# = 0.0
list{(I%)}.z2# = 0.0
NEXT
REM Initialise rotation angles:
t1 = 2*PI*RND(1)
t2 = 2*PI*RND(1)
t3 = 2*PI*RND(1)
F% = &10000 : REM Frequently used constant
frames% = 0
*REFRESH OFF
SYS GetTickCount TO time0%
REPEAT
REM Draw background:
SYS LerpClr, g{}, &102030, &805090
REM Init key values for Z-sort
REM (this could and perhaps should be moved into the rotation routine):
SYS InitExPoints, listBaseAddr%, N%
REM Rotate coordinates:
SYS Rotate, listBaseAddr%, 1+2+4, N%, F%*t1, F%*t2, F%*t3, \
\ F%*-100, F%*-200, F%*50, \
\ F%*320, F%*256, F%*0, \
\ 1, F%*300, F%*800
REM Shell sort code (in C) courtesy of Rosetta Code, many thanks:
SYS ShellSort, listBaseAddr%, N%, -1
REM Draw the depth-sorted balls:
FOR I% = 0 TO N%-1
J% = list{(I%)}.key%
SYS Plot, g{}, ball%, 48, 48, list{(J%)}.x2#-16, list{(J%)}.y2#-16
NEXT I%
PROCDisplay( TRUE )
frames% += 1
SYS GetTickCount TO time1%
IF time1%-time0%>=1000 THEN
SYS GetTickCount TO time0%
SYS "SetWindowText", @hwnd%, STR$frames% + " fps"
frames% = 0
ENDIF
REM Bump rotation angles:
t1 += 0.02001
t2 += 0.01905
t3 += 0.00598
UNTIL FALSE
DEF PROCFixWndSz
LOCAL W%
SYS"GetWindowLong",@hwnd%,-16 TO W%
SYS"SetWindowLong",@hwnd%,-16,W% ANDNOT&40000 ANDNOT&10000
ENDPROC
DEF PROCerror
OSCLI "REFRESH ON" : CLS : ON : PRINT '" ";
REPORT : PRINT " at line "; ERL;
REPEAT UNTIL INKEY(1)=0
ENDPROC
|
|
|
|
rtr
Guest
|
|
Re: Letting GCC do the hard work
« Reply #8 on: Aug 17th, 2014, 10:32am » |
|
on Aug 17th, 2014, 08:13am, David Williams wrote:| This kind of performance is convincing me that hybridizing BB4W and C/C++ code is the way to go (for me personally). |
|
Performance (if by that you mean speed) is not a good reason to go down the GCC route. Even though the code generators in modern compilers are very good, you will almost always be able to do better with hand-crafted assembler.
That is especially true when you are not targeting a particular CPU architecture, but want the code to run on a wide range of machines. In that case much of the clever code-optimising for a specific architecture, that GCC can do very well, will benefit some machines at the expense of others.
If you are compiling with the -march=native switch and then testing your code on the same machine you are getting a misleading impression of performance (unless of course you want to go down the route of including machine code for a range of different architectures and choosing the best one at run time).
Where using C does admittedly have advantages is in speed and ease of coding, and especially in time taken debugging. If those are the issues that most concern you, then fine.
Richard.
|
|
 Logged
|
|
|
|
David Williams
Developer
member is offline
meh
Gender:
Posts: 452
|
|
Re: Letting GCC do the hard work
« Reply #9 on: Aug 17th, 2014, 11:38am » |
|
on Aug 17th, 2014, 10:32am, Richard Russell wrote:| If you are compiling with the -march=native switch and then testing your code on the same machine you are getting a misleading impression of performance [...] |
|
Yes, I did use that switch and when after uploading the EXE for public consumption, I discovered that it crashed my 32-bit XP-based laptop (which I hardly use now!). I suspect my use of -march=native caused GCC to generate 64-bit code since the laptop the code was compiled on is a 64-bit machine. Lesson learned.
Re-compiling the vector balls demo without the aforementioned switch results in the code working on the 32-bit laptop, although the frame rate isn't as high as on the compilation machine (which is a little faster anyway, I think).
Quote:| Where using C does admittedly have advantages is in speed and ease of coding, and especially in time taken debugging. If those are the issues that most concern you, then fine. |
|
For the vast majority of my 'applications', the speed of GCC's generated ASM code suffices (and in some cases, has exceeded the speed of my hand-written ASM code, which isn't too surprising!). I won't be touching -- or rather, writing -- assembler code again unless my life depends on it.
David.
--
|
|
Logged
|
|
|
|
|
