Go Down

Topic: Problem on Due number of clock cycles  (Read 1 time) previous topic - next topic

gok

Nov 30, 2017, 10:17 am Last Edit: Nov 30, 2017, 10:46 am by gok
 I have measured on off time of digital pin of due by using oscilloscope .
Later I have tested with number ticks.
first line of code 5 ticks others 2 ticks so I ignored first 5 ticks.
I have seen on oscilloscope screen and serial monitor different  things.
Every for setting  logic 1 and 0 frq=10.5Mhz which means 8 ticks.
On serial monitor for setting  logic 1 and 0 number of ticks=4 tick

Oscilloscope says every line of code takes 4ticks
Serial monitor says every line of code 2 ticks.

If I couldn't trust number ticks.
How do I measure frequency correctly ?

I am taking 8000 samples 888micro second.
Matlap calculate fft according to the 888 micro second for 8000 samples.
Below I added how I take 8000 samples.
Code: [Select]
void readValues() {
  EthernetClient client = server.available();
  byte X[8000];
  uint32_t i, t0, t1,t[8000];
  byte *a = &X[0];
  
  elapsed_Time = 0;
  noInterrupts();
  //PIO_PDSR&0B00000000000000000000000011111111  ARDUINO DUE  (MSB)D11 D14 D15 D25 D26 D27 D28 D29(LSB) PORTLARININ STATUS REGISTER

  t0 = SysTick->VAL; //

  // first reading  7clk laters 3clk

  T(T(T(*a++ = PIOD->PIO_PDSR & 0B00000000000000000000000011111111))) //1) 3008tick //with this method we can read port 1000 times with one line of code.
number of tick is same with write 1000 line of code from A[0] to A[1000]
  T(T(T(*a++ = PIOD->PIO_PDSR & 0B00000000000000000000000011111111))) //2) 3000tick
  T(T(T(*a++ = PIOD->PIO_PDSR & 0B00000000000000000000000011111111))) //3) 3000tick //PIO_PDSR= Pin Data Status Register
  T(T(T(*a++ = PIOD->PIO_PDSR & 0B00000000000000000000000011111111))) //4) 3000tick
  T(T(T(*a++ = PIOD->PIO_PDSR & 0B00000000000000000000000011111111))) //5) 3000tick
  T(T(T(*a++ = PIOD->PIO_PDSR & 0B00000000000000000000000011111111))) //6) 3000tick
  T(T(T(*a++ = PIOD->PIO_PDSR & 0B00000000000000000000000011111111))) //7) 3000tick
  T(T(T(*a++ = PIOD->PIO_PDSR & 0B00000000000000000000000011111111))) //8) 3000tick



  t1 = SysTick->VAL; //  to-t1=24008ticks
   interrupts();
    a=&X[0];//pointer goes to the beginning of array .
    T(T(T( client.write(*a++))))//1)
    T(T(T( client.write(*a++))))//2)
    T(T(T( client.write(*a++))))//3)
    T(T(T( client.write(*a++))))//4)
    T(T(T( client.write(*a++))))//5)
    T(T(T( client.write(*a++))))//6)
    T(T(T( client.write(*a++))))//7)
    T(T(T( client.write(*a++))))//8)


  Serial.print("n_ticks: ");
  Serial.println( ((t0 < t1) ? 84000 + t0 : t0) - t1 );
  Serial.println();
  
  delay(30);


  
}

and the most maddening thing is every reading of a port takes 8 ticks with sending its values to an array .
But reading a port  takes 4 ticks for reading a port without sending array.
???
Before I asked for number of ticks But I couldn't find solutions in assembly language.
if solution is in assembly file.

if the problem have a solution in assambly
(Frankly I haven't experienced in assembly field.)
please send clearly understandable definition.

weird_dave

You do know that a line of code does not necessarily translate to 1 instruction and that 1 instruction may not necessarily take only 1 clock cycle?

gok

Yes You are right
one of  my  main problem was taking samples as soon as fast.
 Uno reads 1600 times in 304micro seconds.

https://digibird1.wordpress.com/arduino-as-a-5m-sample-osciloscope/

So I readed digital port of arduino uno at the speed of 5M samples per second  and every reading 0.19 micro second which means nearly 3 clocks cycles.

while uno is able to  reading a port in 3 clocks cycles why reading a port for due takes 8 clock cycles.
if I solve this problem my speed of samples per second  will increases Then I can show on Matlab frequency via fft as correct as possible.

westfw

Quote
why reading a port for due takes 8 clock cycles.
The short answer is that the flash memory and GPIO ports in the Due are not as much faster than the AVR as the CPU is.  In fact, the flash memory is about the same speed (~50ns) as on an AVR  (however, there are other factors that make it "effectively" faster.)  It looks like PIO access takes at least two clocks (assuming that it is configured to be as fast as possible.)
Newer ARM Microcontrollers were more "aware" of engineers' desire to read/write GPIO as quickly as possible, and moved the GPIO controller to higher-speed bus interfaces, as well as making "flash acelleration" more complicated.  For example, the SAMD21 on Arduino Zero actually has an instruction cache, AND has GPIO on a "low-latency CPU local bus", so it would be interesting to see if the Zero (48MHz) runs your code faster than a Due (84MHz.)   *I* can't tell just from looking at the datasheet; ARMs are complex and have all sorts of features that interfere with figuring out execution speed just from looking at the code :-(


ard_newbie


Sam3x8e reads 32 input pins in 6 to 8 clock cycles. In your case, if I understand what your are actually doing, you need to read the first 8 input pins of PIOD.

8 input pins reading / 8 clock cycles should give you something close to 1 clock cycle for 1 input pin reading, right ?

gok

PIOD->PIO_PDSR & 0B00000000000000000000000011111111
I have modified upper line to
PIOD->PIO_PDSR
So I have removed  and& operation .

Result
Before with and operation every reading takes 8 clock cycles
After removing and operation it takes 6 clock cycles which means
Speed of reading increases from 10.5MSP to 14MPS.

Thanks so much @ard_newbie  !!!

I want Due to read a port in 3 clocks cycles
but as  @westfw  said
latency is due to with CPU and Sram communication.

westfw

Quote
it would be interesting to see if the Zero (48MHz) runs your code faster than a Due (84MHz.)
I think I've shown that the Zero will read or write a port at essentially once per clock cycle if you use the (relatively undocumented, even in the datasheet) IOBUS high speed access.
That means that a string of 

Code: [Select]
    p->OUTTGL.reg = (1 << 21);
    x = p->IN.reg;   // x is volatile, so the store always happens

produces a ~5.3MHz square wave, which means that the sample/store is happening at better than 10MHz.
The instruction sequence produced is:

Code: [Select]
    2138:       61d9            str     r1, [r3, #28]  ;; do pin toggle
    213a:       6a18            ldr     r0, [r3, #32]  ;; load port value
    213c:       6010            str     r0, [r2, #0]   ;; store into RAM

See additional info here: https://community.atmel.com/forum/d21-faster-or-non-blocking-io-access

gok

Hi everybody
A colleague of mine advice me that
 to detect elapsed time for any function in the codes
set a pin high start your reading process
end of reading set pin off
and measure the On time of the signal  via oscilloscope.
I have added the image of oscilloscope.

Setting-pin13-ON =2 ticks
Reading X
  • =13 ticks

Reading X [1]=10 ticks
Reading X [2]= 6 ticks
.
.
.
Reading [1023]=6 ticks

Elapsed_T_ON=1022×6 + 25 =6157 ticks
Elapsed_T_ON=6157×11.9=73.3ns

So I have seen 73.32ns on oscilloscope screen.
It's Ok to say after two reading Due reads port in 6 clock cycles .

But When I tried to read 8000 times
Something went wrong. It takes to much time to finish reading.
I don't know why Due behave like that But of course there is an answer in depth of MCU structure.




Go Up