Performance of floating point operations of F429

[JWilk.1]

Hello,

I am measuring execution time of a function that operates with floating point numbers and noticed that it is more than I expect. I am using Nucleo F429ZI, clock is configured to the 50 MHz. For measuring the execution time I use TIM6 configured in the following way:

  TIM6->PSC = 49;
  TIM6->ARR = 0xFFFF;
  TIM6->CNT = 0;

For the test, I am measuring the time for which the following statement is processed:

float a = 1.3;
float b = 0.0007;
float c = 1.0;
float d = 12.25;
 
TIM6->CNT = 0;
TIM6->CR1 |= 1;
c += (a - b) * d;
TIM6->CR1 &= ~1;

TIM6->CNT shows 28 microseconds. Disassembly is:

100       	c += (a - b) * d;
08000548:   vldr    s14, [r7, #20]
0800054c:   vldr    s15, [r7, #16]
08000550:   vsub.f32        s14, s14, s15
08000554:   vldr    s15, [r7, #8]
08000558:   vmul.f32        s15, s14, s15
0800055c:   vldr    s14, [r7, #12]
08000560:   vadd.f32        s15, s14, s15

According to ARM, the total execution time for this instructions should be 11 cycles (according to CYCCNT they are 15).

The point that I am confused is why 15 cycles correspond to nearly 30 microseconds. Is way I am measuring it incorrect? Or there are some waits that I am not aware of?

Best regards.

John

[waclawek.jan · Accepted solution]

TIM6->PSC = 49;

TIMx_is preloaded - it means, that this value does not get "active" (loaded into the prescaler counter itself) until Update event occurs - either as overflow, or as a "forced" event through writing 1 to TIMx_EGR.UG.

> TIM6->CNT shows 28 microseconds.

Provided I'm correct in my guess, i.e. that the prescaler counter is still at its default 0 (i.e. no prescaling), this would mean 28 cycles, which is nearly a match, given there are some cycles to stop the timer, and then things like FLASH latency, RAM latency etc.

JW

[Tesla DeLorean]

Make sure the chip is clocking at the rate expected. For short intervals I can't seen the need to a prescaler. A 32-bit TIM like TIM2/TIM5 would give broader coverage, but DWT_CYCCNT is the usual method.

30uS sounds a bit long

[gregstm]

.. an alternative method is to use the BSRR register to set/reset an IO pin and measure with an oscilloscope - crude I know, but removes the uncertainty of whether the timer is behaving as you hoped..

[JWilk.1]

Thanks to all of you for the quick responses.

I missed the fact that the prescaler value is preloaded. So after configuring the prescaler I added:

TIM6->EGR = TIM_EGR_UG;

And the counter was showing 0. After removing the prescaler at all, I got CNT of 30 (cycles) which looks much more reasonable.

Thanks a lot one more time.

Best regards,

John

[waclawek.jan]

What's crude in this?

Au contraire, I find this to be the finest method of all.

JW

[JWilk.1]

I also agree that this is the finest method. However, due to the current situation (COVID, quarantine) I don't have an oscilloscope near at hand.

[gregstm]

yes, "crude" is a poor word choice - "simple" probably better. The BSRR register is one of my favourite features/tool of the chip. I hope you and your oscilloscope are reunited soon.