StellarPad: sbi() + sbi() 4.6x Faster than digitalWrite()

[Rei Vilo 695]

sbi() and sbi() are 4.6x faster than digitalWrite() on the LaunchPad Stellaris.

 

Is there a way to improve the implementation of  digitalWrite()?

 

Results are:

 

Stellaris on embedXcode

 

digitalWrdigitalWrite() -() sbi() Comparison

Times in ms

digitalWrite() 1500

cbi() sbi()    325

 

Stellaris on Energia

 

digitalWrite() - cbi() sbi() Comparison

Times in ms

digitalWrite() 1500

cbi() sbi()    400

 

sbi() and sbi()require functions that are not implemented with the MSP430, namely portBASERegister and GPIOPinWrite.

 

The measures were provided by the following basic sketch, compiled with Energia or embedXcode:

// Core library for code-sense

#include "Energia.h"

 

#define LOOPS 1000000

#define _pin RED_LED

#define portOutputRegister(x) (regtype)portBASERegister(x)

#define cbi(reg, mask) GPIOPinWrite(reg, mask, 0)

#define sbi(reg, mask) GPIOPinWrite(reg, mask, mask)

typedef volatile uint32_t regtype;

typedef uint8_t regsize;

 

uint32_t chrono;

regtype _port;

regsize _bit;

 

void setup()

{

  Serial.begin(9600);

  delay(10);

  

  // put your setup code here, to run once:

  Serial.println("digitalWrite() - cbi() sbi() Comparison");

  Serial.println("Times in ms");

 

  Serial.print("digitalWrite() ");

  pinMode(_pin, OUTPUT);

 

  chrono = millis();

  for (uint32_t i=0; i<LOOPS; i++) {

    digitalWrite(_pin, HIGH);

    digitalWrite(_pin, LOW);

  }

  chrono = millis() - chrono;

 

  Serial.println(chrono, DEC);

  delay(1000);

 

  Serial.print("cbi() sbi()    ");

  _port   = portOutputRegister(digitalPinToPort(_pin));

  _bit    = digitalPinToBitMask(_pin);

 

  chrono = millis();

  for (uint32_t i=0; i<LOOPS; i++) {

    sbi(_port, _bit);

    cbi(_port, _bit);

  }

  chrono = millis() - chrono;

    

  Serial.println(chrono, DEC);

 

}

 

void loop()

{

    ;

}

 

[chicken 630]

Well, you moved most of the overhead of digitalWrite/Read outside of the loop with sbi/cbi. Include the calculation of _port and _mask into the second loop and we're probably getting closer.

 

When you look at the source of StellarisWare / driverlib, you will see that GPIOPinWrite/Read are a single command:

return(HWREG(ulPort + (GPIO_O_DATA + (ucPins << 2))));

and

HWREG(ulPort + (GPIO_O_DATA + (ucPins << 2))) = ucVal;

respectively.

 

I'd venture to guess that moving those lines directly into wiring_digital.c instead of calling ROM_GPIO already makes a big difference. Given that we save a call, the size penalty should be small.

 

Another thing I notice is, that we first create a bitmask out of the single pin and that mask then get's shifted by 2 in the code above. Maybe the two operations can be combined. Though I guess creating another look-up table just for this minor improvement is not worth the trade-off.