Mega Code Archive
Performanceone
<![CDATA[
Program Performance
*******************************************************************
[my stuff]
re 'performance'
or more specifically, 'performance monitoring',
here's a tip using vital info from Bob Swart's excellent page
on 'Delphi Optimising' at:
http://www.tietovayla.fi/BORLAND/techlib/dl390/dl390.html
let's say you're worried about the performance of your app,
and in the absence of a 'profiler', you want to try and track
down performance bottlenecks if possible...
here's a quick way:
include MMSYSTEM in your uses clause
then find the TIME taken to execute particular blocks of code
by enclosing those blocks with calls to the timeGetTime function
(which is in MMSYSTEM.DLL, and which returns 'milliseconds
since Windows started as a LongInt) -note that a LongInt can
hold 47 days worth of milliseconds
like this...
var
startTime: LongInt;
endTime: LongInt;
selectTime: LongInt;
redrawTime: LongInt;
tempString: String;
begin
startTime := timeGetTime;
//select the step that we want to select
ListsManager.SetStepSelection(clickedElem);
endTime := timeGetTime;
selectTime := (endTime - startTime);
startTime := timeGetTime;
//redraw everything
DrawObj.ReDrawAll(DrawArea.Canvas, DrawArea.Width, DrawArea.Height,
ListsManager.stepObjList, ListsManager.linkObjList);
endTime := timeGetTime;
redrawTime := (endTime - startTime);
tempString := 'The selection process takes ' + IntToStr(selectTime) +
' milliseconds' + Chr(13) +
'The redraw process takes ' + IntToStr(redrawTime) +
' milliseconds';
Application.MessageBox(PChar(tempString), ' Programmers Debug Window', mb_OK);
end;
for instance...
*******************************************************************************
[this next is similar to the above but more in-depth]...
Here is a contribution to the Optimization debate. Comments will be welcome.
Irwin Scollar:
{TestTime, a program to Test the timing of a routine, using the RDSTC (Read
Time Stamp Counter) for the number of cycles since a Pentium processor was
started or reset, based on:
Jon Shemitz, Using RDTSC for Pentium Benchmarking, in Don Taylor et al,
Delphi 3 Programming, Coriolis Group Press, ISBN 1-57610-179-7, 1997}
{Repeatedly run it a number of times and take lowest observed value to
account for Windows doing other things with higher priority.
Under WinNT/2000 try:
start /high testtime to force high running priority
or even perhaps
start /realtime testime to force still higher priority}
{$APPTYPE CONSOLE}
{try various complier options for dcc32.exe}
Program TestTime;
uses
SysUtils;
const
D32 = $66;
Function RDTSC : Comp;
Var
TimeStamp : Record
case byte of
1 : (Whole:comp);
2 : (Lo, Hi : Longint);
end;
begin
asm
db $0F;
db $31;
{$ifdef Cpu386}
mov [Timestamp.Lo],eax
mov [Timestamp.Hi],edx
{$else}
db D32
mov word ptr TimeStamp.Lo,AX
db D32
mov word ptr TimeStamp.Hi,DX
{$endif}
end;
Result := TimeStamp.Whole;
end;
function Comp2Str(N : Comp): string;
begin
Result := Format('%.0n',[N]);
end;
procedure TimeIt;
var
StartTime, StopTime : Comp;
i : integer;
Overhead : Comp;
begin
i := 0;
{get Timer overhead}
StartTime := RDTSC;
Overhead := RDTSC - StartTime;
StartTime := RDTSC;
{code to time goes here, e.g.}
While (i < 10000) do
inc(i);
StopTime := RDTSC - StartTime - Overhead;
{display result in processor cycles}
Write('Code takes ',Comp2Str(StopTime), ' cycles');
ReadLn;
end;
begin
TimeIt;
end.
_______________________________________________
Delphi mailing list -> Delphi@elists.org
http://elists.org/mailman/listinfo/delphi
********************************************************************
SEE BOB SWART'S DELPHI OPTIMISATION PAGE AT:
http://www.tietovayla.fi/BORLAND/techlib/dl390/dl390.html
IT'S VERY GOOD
to pinpoint your performance problems get yourself a decent profiler. I used
SpeedDaemon and GProfiler. GProfiler is free. Get it here:
www.eccentrica.org/gabr/gpprofile/gpprofile.htm (includes source)
The improve your application's design have a look here 8)
http://oop.com/white_papers/delphi/business-objects.htm
**********************************************************************
I would have to agree that the use of profiling tools can help a programmer
to drastically improve the performance of his/her code, and that many of the
major bottlenecks in modern applications come from unnecessary recursion or
looping, network traffic or poorly designed SQL Queries.
In my experience, many programmers today (myself included to some extent),
consistently write code that greatly underperforms, we continue to use the
same badly thoughtout, and inefficient, solutions to the same problems over
and over again. How many of you have seen or written code that scans a list
for the data you are looking for, and once you have found the data continue
to search the rest of the list? Use of the Continue, Break and Exit
commands are always handy to bear in mind in such circumstances....
Many programmers remain ignorant of good algorithm design, and as such,
rarely consider the use of indexing, binary chop functions, hash tables and
the like (me included), and even if they do, use the least effective
solution for the problem in question. All too often have I seen programmers
"fix" their performance problems, by simply upping the specification of the
machines that their users will use, that is fine if you are in a position to
control that, but many of us do not have such a luxury.
I often think back to the days of my good old C64 (for those new to such
stuff an old 8-bit 64KB god of a machine you plugged into your TV), and
wonder what the programmers of such machines (who knew exactly how to
squeeze every last pound of flesh from those beasties) could do with the
monolithic specification of todays desktop PC's, there is a definate
connection between the speed of the target machine, and the laziness of the
programmer. Call it unrealistic in todays marketplace if you must, but the
modern programmer is seriously underperforming, and the specification of
modern machines mean that the laziest and least tallented programmers can
still compete...
Many of you have suggested ensuring that optimization is turned on, but in
my experience, and the experience of many others, is that turning
optimisation on can change the meaning of code, and give undesired and often
unexplainable behaviour, and as such I never use this feature...
What is lacking is knowledge of how best to optimise algorithms, queries,
and reduce network traffic, I think this thread could be both usefull and
interesting to all who read it if we stop debating over David's initial
tips, and start to submit new ideas and sources of information to each other
regarding the above.
Remember, many of us do not trust optimisation, it has bitten us too many
times :-(
regards,
Paul Jackson
******************************************************************************************
> > This last one is an old trick for swapping two elementes without using a
> > temporary variable.
> > Swap(a,b):
> > a = 0010
> > b = 0101
> > b := a xor b (=0111)
> > a := a xor b (=0101)
> > b := a xor b (=0010)
>I've seen this befoer, and I recall that it isn't really any faster.
>Firstly, you have to declare a and b with the var modifier. This means
>that each reference to a or b requires a memory access. Your version
>requires three MOVs and an XOR for each line, and they're all on memory
>locations instead of registers. Unoptimized, the standard temporary
>variable method produces similar code, all with memory access since the
>Intel instruction set has no way of transferring directly between two
>memory addesses.
>
>Optimized, apparently, neither routine changes. In this case, I'd still
>choose the standard method. It is easier to read--more obvious that it
>swaps the contents of two variables. It also yields ten MOVs rather than
>nine MOVs and three XORs. When put into the context of another
>subroutine, I suspect the benefits would be even greater. Takse Swap out
>of its own procedure--inline it--and it should work even better since the
>compiler can better utilize the registers.
If this really needs to be optimized, the following will do it in 4 MOVs, a
PUSH, and a POP:
procedure swap (var a, b: integer);
asm
push edi
mov ecx, [a] // pointer to "a" is in EAX, so this is really mov ecx,
[eax]
mov edi, [b] // pointer to "b" is in EDX, so this is really mov edi,
[edx]
mov [a], edi
mov [b], ecx
pop edi
end;
If you're absolutely certain the caller is not using EDI, then the PUSH/POP
could be removed, but as a general-purpose procedure EDI needs to be preserved.
]]>