Fast IO access

[elral]

Description of changes

For each device with multiple accesses to I/O's the arduino functions digitalread() and digitalWrite()is replaced by direct port access to reduce the time each changed device needs.
This will allow to increase the number of multiplexers while keeping the functionality of encoders and will give the possibility to implement encoders connected to multiplexers.

As this change will exceed the flash for the ProMicro, it will be only implemented for:

• Mega
• Uno/Nano
  per compile flag and could be implemented for the Pico (not changed so far).

More description is available in issue #239

Fixes #239

[github-actions]

Firmware for this pull request:
Firmware.zip

[GioCC]

A few first-look comments:

• Storing two bytes for every "fast" pin (port + mask) wastes a lot of RAM.
  Having these values ready probably only makes sense for driver pins (shift registers, multiplexers, LED drivers), whose number is limited (and could be even more limited if they were shared) and whose usage is much more intensive. It's not as important e.g. for buttons.
  If a better performance than the standard digitalRead is still desired, you could have a second overload of the "fast" versions, with the simple pin number as argument, to be used in "regular" cases.
  This overload would incorporate the digitalPinToBitMask() and portOutputRegister(digitalPinToPort()) calls; these are simple macro lookups and therefore still remarkably faster than the standard functions.
  It could be evaluated if these "less optimized" functions have sufficient performance to be used also for driver pins.

• I see that in some cases, both the pin number and the pin port+mask are stored; I wonder if it's possible to drop pin number (not examined the code to this extent yet).

• In digitalWriteFast(), the read->update->write operation is supposed to be atomic, therefore it's necessary to add the "Save SREG / cli" sequence as used in the default function.

[elral]

A few first-look comments:

• Storing two bytes for every "fast" pin (port + mask) wastes a lot of RAM.
  Having these values ready probably only makes sense for driver pins (shift registers, multiplexers, LED drivers), whose number is limited (and could be even more limited if they were shared) and whose usage is much more intensive. It's not as important e.g. for buttons.
  If a better performance than the standard digitalRead is still desired, you could have a second overload of the "fast" versions, with the simple pin number as argument, to be used in "regular" cases.
  This overload would incorporate the digitalPinToBitMask() and portOutputRegister(digitalPinToPort()) calls; these are simple macro lookups and therefore still remarkably faster than the standard functions.
  It could be evaluated if these "less optimized" functions have sufficient performance to be used also for driver pins.

Give me some time to think about (and understand ;) ) this

• I see that in some cases, both the pin number and the pin port+mask are stored; I wonder if it's possible to drop pin number (not examined the code to this extent yet).

Yeah, I thought about this a longer time. Idea's are welcome.
But this opens another discussion. The pin numbes are required in some detach() functions, but not in all. The newer ones switches the pins back to input where the "older" ones leave the state as they are. A clean solution is the first appoach (switching back the pin to input), but is it really required? If so, shouldn't we change it for the detach()functions where it is not done? But maybe better to discuss this in a separate topic.

• In digitalWriteFast(), the read->update->write operation is supposed to be atomic, therefore it's necessary to add the "Save SREG / cli" sequence as used in the default function.

Good point!

Thanks and regards

Ralf

[GioCC]

Thanks to you for trying to find even more improvements!

OK, hoping to spare you some time by trying to be clearer:
let's say we add another function in MFFastIO.h...

inline uint8_t digitalReadFast(uint8_t pinNr)
{
    return (*portOutputRegister(digitalPinToPort(pinNr) & digitalPinToBitMask(pinNr)) ? HIGH : LOW;
}

this way, e.g. Buttons would continue to use pin numbers without the need to store pin port and pin mask.

About the detach() functions: even there, I would imagine that a "fast" version of the pinMode() could be introduced to reset the pin to input. I haven't looked at the code yet though.

Switching unused pins back to input is definitely a thing to recommend though.

[github-actions]

Firmware for this pull request:
Firmware.zip

[github-actions]

Firmware for this pull request:
Firmware.zip

[github-actions]

Firmware for this pull request:
Firmware.zip

[github-actions]

Firmware for this pull request:
Firmware.zip

[github-actions]

Firmware for this pull request:
Firmware.zip

[github-actions]

Firmware for this pull request:
Firmware.zip

[github-actions]

Firmware for this pull request:
Firmware.zip

[github-actions]

Firmware for this pull request:
Firmware.zip

[github-actions]

Firmware for this pull request:
Firmware.zip

[github-actions]

Firmware for this pull request:
Firmware.zip

[github-actions]

Firmware for this pull request:
Firmware.zip

[github-actions]

Firmware for this pull request:
Firmware.zip

[github-actions]

Firmware for this pull request:
Firmware.zip

[github-actions]

Firmware for this pull request:
Firmware.zip

[github-actions]

Firmware for this pull request:
Firmware.zip

[github-actions]

Board and firmware folder for this pull request:
Mobiflight-Connector.zip

[github-actions]

Board and firmware folder for this pull request:
Mobiflight-Connector.zip