GPIO Performance

[skywave]

I found a wonderful article about GPIO performance here: http://naberius.de/2015/05/14/esp8266-g ... rformance/

and tried to include as asm on my project to set GPIO fast:

__asm__ volatile ("movi a2, 0x60000304\n"
"movi a4, %0\n"
"memw\n"
"s32i a4, a2, 0\n"
::"r" (setbits):"a2", "a4");
setbits is a static uint16;

However the linker gives me: undefined reference to `a5' (that I'm not using). Could you please advice on what can be wrong? Thank's in advance.

[dkinzer]

I haven't attempted to discover what's going on with the inline assembler code but I say why bother? The eagle_soc.h file contains definitions that allow direct port manipulation that results in similar code. For example:

Code: Select all

40211000:   60000304
40211004:   8000
[...]
   GPIO_REG_WRITE(GPIO_OUT_W1TS_ADDRESS, 0x8000);
40211022:   fff831           l32r   a3, 40211004 <_irom0_text_start+0x4>
40211025:   fff621           l32r   a2, 40211000 <_irom0_text_start>
40211028:   0020c0           memw
4021102b:   006232           s32i   a3, a2, 0


My hypothesis is that the W1TS in the name of the register means "write 1 to set". There are other definitions in eagle_soc.h containing W1TC meaning "write 1 to clear".
You could create some macros that are less verbose and, perhaps, more suggestive of their function, e.g.

Code: Select all

#define GPIO_OUT_SET(m)  GPIO_REG_WRITE(GPIO_OUT_W1TS_ADDRESS, m)

[eriksl]

Also the assembly code won't give you any speed increase, because the rate at which the GPIO's can be toggled is limited and is way lower than you can ever achieve with compiled C code, so there is no point.

I keep wondering why all these people want to toggle GPIO pins at Mhz rate???? If it's for pwm, just use the supplied pwm functionality, no need to re-invent it yourself.

[skywave]

If you want to do PWM at 1000Hz with 1000 steps -> the minimum step/stage must have 1us ON and 999us OFF. So it really matters to toggle the pins at Mhz rate. I have also find out that the interrupt routine (on PWM file) takes around 3us to execute, even if we don't do a thing inside (only reload the timer). So that code won't work for the smallest steps 1us, 2us, 3us.

[eriksl]

I know. That's why other microcontroller have this stuff in hardware, it must act incredibly fast (and do nothing most of the time as well).

The atmega328p has a 16 bits counter/timer/pwm that can run directly on the system clock which is 20 MHz at most. That will give you a 16 bit resolution at 305 Hz AT MOST. I've found that actually the required resolution is about 12 bits (for tuning leds right down to almost invisible in the dark), then you could achieve 5 KHz rate. And then we're talking dedicated hardware on the not-low-end-atmel series.

I think you should reconsider if your goals are realistic. The SDK PWM functionality already is quite efficient, I doubt you can do very much better than that (and still keep the SDK code happy).