Digital Mars - c++ - Counting Clock cycles

↑ ↓ ← → "Walter" <walter digitalmars.com> writes:

The Intel x86 RDTSC instruction is what you're looking for.

"jim p" <x y.com> wrote in message news:blvp83$22tm$1 digitaldaemon.com...
 I hope this isn't too much of a stupid question.

 I've just developed a silly little real time network routing simulator.

 I need to determine the minimum processor speed that will perform all the
 necessary calculations to keep it operating in real time.

 Forgetting memory access times, to do this I need to know how much CPU


 the application is doing.

 Is there any way of counting CPU clock cycles for example, for a


 function call ???

 Jim

Oct 08 2003

↑ ↓ ← → "jim p" <x y.com> writes:

Any pointers to some documentation on the x86 RDTSC instruction ??




"Walter" <walter digitalmars.com> wrote in message
news:bm0cve$2tkb$1 digitaldaemon.com...
 The Intel x86 RDTSC instruction is what you're looking for.

 "jim p" <x y.com> wrote in message news:blvp83$22tm$1 digitaldaemon.com...
 I hope this isn't too much of a stupid question.

 I've just developed a silly little real time network routing simulator.

 I need to determine the minimum processor speed that will perform all




 necessary calculations to keep it operating in real time.

 Forgetting memory access times, to do this I need to know how much CPU


 the application is doing.

 Is there any way of counting CPU clock cycles for example, for a


 function call ???

 Jim

Oct 08 2003

↑ ↓ ← → "jim p" <x y.com> writes:

Thanks.

Another question for you Walter.

Is it possible to determine the number of CPU instructions used by a
function ?
If so, with help from the RDTSC code below, I could the time taken per
instruction.
And then.....would it be possible to say that a 1700 MHz processor processes
this function in X amount of time, therefore a 1200 MHz processor will take
Y amount time to perform the same function.

Am I talking bollocks ??



"Walter" <walter digitalmars.com> wrote in message
news:bm1ul0$iag$1 digitaldaemon.com...
 "jim p" <x y.com> wrote in message news:bm0md1$9sk$1 digitaldaemon.com...
 Any pointers to some documentation on the x86 RDTSC instruction ??


 long long rdtsc()
 {
     __asm
     {
         RDTSC
     }
 }

 should do it.

Oct 08 2003

↑ ↓ ← → "Walter" <walter digitalmars.com> writes:

"jim p" <x y.com> wrote in message news:bm24t5$sk1$1 digitaldaemon.com...
 Is it possible to determine the number of CPU instructions used by a
 function ?
 If so, with help from the RDTSC code below, I could the time taken per
 instruction.
 And then.....would it be possible to say that a 1700 MHz processor


 this function in X amount of time, therefore a 1200 MHz processor will


 Y amount time to perform the same function.

 Am I talking bollocks ??


Each instruction takes widely varying lengths of time, so you won't be
getting a very meaningful result. You're also reaching a level where it's
time to download the Intel CPU processor manuals; I suggest that
www.intel.com should be your next stop!

Oct 08 2003

↑ ↓ ← → "_" <mjoyl lineone.netX> writes:

"jim p" <x y.com> wrote in message news:bm24t5$sk1$1 digitaldaemon.com...
 Thanks.

 Another question for you Walter.

 Is it possible to determine the number of CPU instructions used by a
 function ?
 If so, with help from the RDTSC code below, I could the time taken per
 instruction.
 And then.....would it be possible to say that a 1700 MHz processor


 this function in X amount of time, therefore a 1200 MHz processor will


 Y amount time to perform the same function.

 Am I talking bollocks ??


As I read it, you are thinking single threaded single tasked CPU loading and
that is just not the case in real world installations, while feasibility
studies on whether C++ could possibly produce fast enough executable code is
a valid proposition, recommended minimum and ideal systems really should be
determined by real world trials as different manufacturers processors and
their individual renditions of their processor types all perform differently
per CPU core clock cycle, not to mention real-time processes hogging the CPU
time and all the myriad of threads running on even a home PC let alone a
networked router. If you have a functional prototype that works on all the
machines you have.. maybe it's time to put it out for limited trials to
people with alpha/beta test experience and various system configurations. If
there are performance problems, determine the bottlenecks and maybe try
re-writing those bottlenecks in inline assembler if the product is to be
specific to x86 class processors or #ifdef inline assembler to replace the C
functions for x86 platforms... that's just my opinion and I've yet to create
anything meaningful with DMC++ yet but I've been debugging/hacking/hand
optimising, alpha/beta testing software for several years.

But if you must benchtest your functions... why not just use them in a
simple benchtest program that reads the current time to highest resolution
then calls an empty function say a million times or what ever's needed and
reads time after the sequence is done to determine benchtests overheads then
re-run again using each of your functions with simulated data and and from
that determine whets eating processing time and go from there... i.e.
produce calibrated test environment for the functions of the network routing
simulator ;-)
HTH :-)

Dec 17 2003

↑ ↓ ← → "jim p" <x y.com> writes:

OK its an assembler instruction.
But, I've never used assembler.
Has anyone written a C function to perform this task ??



"Walter" <walter digitalmars.com> wrote in message
news:bm0cve$2tkb$1 digitaldaemon.com...
 The Intel x86 RDTSC instruction is what you're looking for.

 "jim p" <x y.com> wrote in message news:blvp83$22tm$1 digitaldaemon.com...
 I hope this isn't too much of a stupid question.

 I've just developed a silly little real time network routing simulator.

 I need to determine the minimum processor speed that will perform all




 necessary calculations to keep it operating in real time.

 Forgetting memory access times, to do this I need to know how much CPU


 the application is doing.

 Is there any way of counting CPU clock cycles for example, for a


 function call ???

 Jim

Oct 08 2003

↑ ↓ ← → roland <--rv ronetech.com> writes:

jim p a �crit :
 OK its an assembler instruction.
 But, I've never used assembler.
 Has anyone written a C function to perform this task ??
 


------------------------------------------------------------------------------------------------------
------------------- 

Written by Roland VARICHON for RONE Technologies - 254 rue de Pressense 
69100 Villeurbane - France - www.ronetech.com   //does not hurt
                                         Free use
//of course if you make a nuclear plan blow because of this code we are 
not responsible
------------------------------------------------------------------------------------------------------
------------------- 

                             #endif
#define _TICK_CTR_PORT 0x43     //read only
#define _TICK_PORT	0x40

long getcpufreq();
long long rdtsc_get();
long cputick2ns(const long cputick);
long ns2cputick(const long ns);

static long _bus_freq = 1193180ul; //bus clock
static int _gettick();
static long _cpu_freq = getcpufreq();



long getcpufreq() {
     int dummy[1]; dummy[0] = 0; //force stack frame, may be not 
necessary, just a bad habit
     long long rdtsci,rdtscf;
     unsigned ticki,tickf;

     asm {
     	    pushf           //save if flag stat
             cli             //lock interrupt
     //cpuid is faster after it had been executed 3 time ???
             cpuid
             cpuid
             cpuid
     //warm the cache
     		mov  DWORD PTR rdtsci,eax
     		mov  DWORD PTR rdtsci[4],edx
     		mov  ticki,eax
     		mov  tickf,eax
     //loop until timer >= 0x110
         lp0:
             call _gettick
     		cmp  eax,110h
     		jb lp0
     //now timer >= 0x110		
     //now going to loop until timer decrements one for precision
     		dec  eax
             mov  ticki,eax
             sub  eax,100h
             mov  tickf,eax
         lp1:
             call _gettick
     		cmp  eax,ticki
     		ja   lp1
             cpuid       //flush pipeline
     		rdtsc
     		mov  DWORD PTR rdtsci,eax
     		mov  DWORD PTR rdtsci[4],edx
     //ok we are now synchronized, initial rdtsc value saved
     //now wait until timer decrements 0x100     		
         lp2:
             call _gettick
     		cmp  eax,tickf
     		ja   lp2
             cpuid
     		rdtsc
     		mov  DWORD PTR rdtscf,eax
     		mov  DWORD PTR rdtscf[4],edx
     //restaure if stat    		
     	    popf
	    }
	    return ((rdtscf-rdtsci)*_bus_freq)/0x100;
     //bus_T = 1/bus_Freq
     //cpu_T = 1/cpu_Freq
     //0x100*bus_T = delta_rdtsc*cpu_T
     //=> cpu_Freq = bus_Freq*delta_rdtsc/0x100
}


long long rdtsc_get() {
//return current value of rdtsc (in cpu tick)
     long long ret;
     asm cpuid;
     asm rdtsc;
//    asm mov DWORD PTR ret,eax;
//    asm mov DWORD PTR ret[4],eax;
     return;
}


long cputick2ns(const long cputick) {
//cpu tick to nanosecond
//convert cpu ticks in nanosecond
//excess convertion
     return ( (((long long)cputick) * 1000000000LL) + ((long 
long)(cpufreq()-1)) ) / ((long long)cpufreq());
}

long ns2cputick(const long ns) {
//nanosecond to cpu tisk
//convert nanoseconds in cpu ticks
//excess convertion
     return ( (((long long)ns) * ((long long)cpufreq())) + 
(1000000000LL-1) ) / 1000000000LL;
}


//------------------------------------------------------------------------
static int _gettick() {
//return the current value of the 8254's timer 0
#if (_INTSIZE==4)
     asm {
     		xor eax,eax
     		out _TICK_CTR_PORT,al	//counter 0 latch
     		in al,_TICK_PORT	//lo cp0
     		mov ah,al
     		in al,_TICK_PORT	//hi cp0
     		xchg ah,al
     }
     return _EAX;
#else
     int ret = 0;
     outp(_TICK_CTR_PORT,0);
     ret |= inp(_TICK_PORT);
     ret |= (((unsigned)inp(_TICK_PORT))<<8);
     return ret;
#endif
}

Oct 16 2003

↑ ↓ ← → "jim p" <x y.com> writes:

No problem, I found a c++ class to do the job.

CCPUTicker v1.22
Written by J. M.McGuiness
source code can be downloaded from
http://www.codeproject.com/datetime/ccputicker.asp?print=true if anyone is
interested



"Walter" <walter digitalmars.com> wrote in message
news:bm0cve$2tkb$1 digitaldaemon.com...
 The Intel x86 RDTSC instruction is what you're looking for.

 "jim p" <x y.com> wrote in message news:blvp83$22tm$1 digitaldaemon.com...
 I hope this isn't too much of a stupid question.

 I've just developed a silly little real time network routing simulator.

 I need to determine the minimum processor speed that will perform all




 necessary calculations to keep it operating in real time.

 Forgetting memory access times, to do this I need to know how much CPU


 the application is doing.

 Is there any way of counting CPU clock cycles for example, for a


 function call ???

 Jim

Oct 08 2003

↑ ↓ ← → "Charles Sanders" <sanders-consulting comcast.net> writes:

Very cool, what are you doing all this profiling for ?  Id like to see some
results of whatever your testing.

C
"jim p" <x y.com> wrote in message news:bm1a7i$1e2d$1 digitaldaemon.com...
 No problem, I found a c++ class to do the job.

 CCPUTicker v1.22
 Written by J. M.McGuiness
 source code can be downloaded from
 http://www.codeproject.com/datetime/ccputicker.asp?print=true if anyone is
 interested



 "Walter" <walter digitalmars.com> wrote in message
 news:bm0cve$2tkb$1 digitaldaemon.com...
 The Intel x86 RDTSC instruction is what you're looking for.

 "jim p" <x y.com> wrote in message




 I hope this isn't too much of a stupid question.

 I've just developed a silly little real time network routing






 I need to determine the minimum processor speed that will perform all




 necessary calculations to keep it operating in real time.

 Forgetting memory access times, to do this I need to know how much CPU


 the application is doing.

 Is there any way of counting CPU clock cycles for example, for a


 function call ???

 Jim

Oct 08 2003

↑ ↓ ← → "Christian Kaiser" <chk online.de> writes:

Please remember that you might be in a multitasking environment, so a
function can take a picosecond or a second, depends on the system load,
process priority etc.

Christian

"Charles Sanders" <sanders-consulting comcast.net> wrote in message
news:bm2g79$1c9o$1 digitaldaemon.com...
 Very cool, what are you doing all this profiling for ?  Id like to see


 results of whatever your testing.

 C
 "jim p" <x y.com> wrote in message news:bm1a7i$1e2d$1 digitaldaemon.com...
 No problem, I found a c++ class to do the job.

 CCPUTicker v1.22
 Written by J. M.McGuiness
 source code can be downloaded from
 http://www.codeproject.com/datetime/ccputicker.asp?print=true if anyone




 interested



 "Walter" <walter digitalmars.com> wrote in message
 news:bm0cve$2tkb$1 digitaldaemon.com...
 The Intel x86 RDTSC instruction is what you're looking for.

 "jim p" <x y.com> wrote in message




 I hope this isn't too much of a stupid question.

 I've just developed a silly little real time network routing






 I need to determine the minimum processor speed that will perform








 the
 necessary calculations to keep it operating in real time.

 Forgetting memory access times, to do this I need to know how much








 work
 the application is doing.

 Is there any way of counting CPU clock cycles for example, for a


 function call ???

 Jim

Oct 09 2003

↑ ↓ ← → "jim p" <x y.com> writes:

I was trying to gather some meaningful results for my masters dissertation
that has to be handed in tomorrow.
I've run out of time and haven't been able to get the exact results I
wanted.

But hey, not to worry.
There goes my distinction!!

I'm finally free from this damned course!!!


"jim p" <x y.com> wrote in message news:bm1a7i$1e2d$1 digitaldaemon.com...
 No problem, I found a c++ class to do the job.

 CCPUTicker v1.22
 Written by J. M.McGuiness
 source code can be downloaded from
 http://www.codeproject.com/datetime/ccputicker.asp?print=true if anyone is
 interested



 "Walter" <walter digitalmars.com> wrote in message
 news:bm0cve$2tkb$1 digitaldaemon.com...
 The Intel x86 RDTSC instruction is what you're looking for.

 "jim p" <x y.com> wrote in message




 I hope this isn't too much of a stupid question.

 I've just developed a silly little real time network routing






 I need to determine the minimum processor speed that will perform all




 necessary calculations to keep it operating in real time.

 Forgetting memory access times, to do this I need to know how much CPU


 the application is doing.

 Is there any way of counting CPU clock cycles for example, for a


 function call ???

 Jim

Oct 09 2003