Achieving Nanosecond Precise Timing using ASM

[Michael Zeng]

I'm attempting to achieve nano-second precision delays by injecting ASM code like so:

__ASM ( \
" NOP\n\t NOP\n\t NOP\n\t NOP\n\t NOP\n\t NOP\n\t NOP\n\t NOP\n\t" \
" NOP\n\t NOP\n\t NOP\n\t NOP\n\t NOP\n\t NOP\n\t NOP\n\t NOP\n\t" \
" NOP\n\t NOP\n\t NOP\n\t" \
); \

This works relatively consistently, however, the delay time doesn't amount to what I expect.

This particular code is supposed to delay 296.875 ns; I am running on an nRF5340 running at 64 MHz (therefore, the clock period is 15.625 ns), and there are 19 "NOP" commands, which should take 19 cycles, amounting to the 296.875 ns delay. However, measuring the delay on a logic analyzer, I'm seeing closer to 220-250 ns. For larger delays, the discrepancy is even larger. I would just like to see if anyone knows why this is happening.

[Vidar Berg]

Hello, 

The NOP instruction is not guaranteed to produce a deterministic delay: https://developer.arm.com/documentation/100235/0004/the-cortex-m33-instruction-set/miscellaneous-instructions/nop. Instead, I would recommend you try using k_busy_wait(). To further improve the timing, you may also consider requesting the 64M crystal oscillator (HFXO) as done in the radio test sample here: https://github.com/nrfconnect/sdk-nrf/blob/72e0f9a63481181994f35f4d2646117ed56d8778/samples/peripheral/radio_test/src/main.c#L11.  

Delay function used by k_busy_wait(): https://github.com/zephyrproject-rtos/hal_nordic/blob/56e0b052dff311c2f8eb08c6804e60fc79feb56f/nrfx/soc/nrfx_coredep.h#L140 

Best regards,

Vidar

[Michael Zeng]

Sorry, I'm not totally sure what you're suggesting. Form the first link you send, I agree that NOP does not produce a deterministic delay. However, the functions you linked all delay an integer number of microseconds, but I need another order of magnitude of precision--down to a hundred nanoseconds of 0.1 microseconds. Is there any way to achieve that without, for example modifying the functions in hal_nordic/nrfx/soc/nrfx_coredep.h at 56e0b052dff311c2f8eb08c6804e60fc79feb56f · zephyrproject-rtos/hal_nordic (github.com) (which I'm not sure would even work)?

Also, could you elaborate on what you mean by "requesting the 64M crystal oscillator (HFXO)"? Is the nRF52840 not already using its 64M crystal?

[Vidar Berg]

Sorry, I must have overlooked the fact that you wanted nanosecond precision when I wrote my reply. For nanosecond accuracy, you will need to create your own delay implementation as you initially tried. However, instead of using NOPs, you might get better results by basing your implementation on one of the delay functions in nrfx_coredep_delay_us.

Michael Zeng said:

Also, could you elaborate on what you mean by "requesting the 64M crystal oscillator (HFXO)"? Is the nRF52840 not already using its 64M crystal?

It is running of the less accurate 64 MHz internal oscillator (HFINT) clock source by default.