Arduino Nano 33 BLE (nRF52840) seems not to run on max. clock speed

RE: Arduino Nano 33 BLE (nRF52840) seems not to run on max. clock speed

bjorn-spockeli — Mon, 27 Apr 2020 10:28:26 GMT

I found this ticket: https://devzone.nordicsemi.com/f/nordic-q-a/45241/nrf52-fpu-single-precision-floating-point---gcc-compiler-flags

So you could try to experiment with the mentioned compiler flags.

Could you also try to run the matrix code from flash with the cache enabled, https://devzone.nordicsemi.com/f/nordic-q-a/14420/nrf52-cache.

RE: Arduino Nano 33 BLE (nRF52840) seems not to run on max. clock speed

janjongboom — Mon, 27 Apr 2020 08:00:50 GMT

Thanks Bjorn, Arduino was pointing me in the direction of the ART as well. I've ordered a nRF52-DK and will do some experiments on that with the Nordic SDK in the next few days.

RE: Arduino Nano 33 BLE (nRF52840) seems not to run on max. clock speed

bjorn-spockeli — Mon, 27 Apr 2020 06:33:39 GMT

Hmm, coremark-wise the STM32L475xx (3.42 CoreMark/MHz @ 80 MHz) is faster than the nRF52840 (3.3 CoreMark per MHz, running CoreMark from flash, cache enabled), which is expected given that it runs at 80MHz vs the 64MHz on the nRF52, but 2 times slower execution on the nRF52840 seems much.

I was thinking that it could be that the cache wasnt enabled, but if you are running the code from RAM then it cant be that. I see that the STM32L475xx has an Adaptive real-time memory accelerator (ART Accelerator), which may improve the execution time on the ST compared to the nRF52840.

RE: Arduino Nano 33 BLE (nRF52840) seems not to run on max. clock speed

janjongboom — Sun, 26 Apr 2020 14:10:21 GMT

Hi Bjorn, thanks a lot for your reply. I indeed realized that `-mfpu=fpv4-sp-d16` was not declared in the compile options for my project (I assumed this would have been set automatically, but apparently not), and this significantly speeded up the compilation. With GCC9, `-O3` `mfloat-abi=hard` (over softfp which I had before) and by placing the functions in RAM this yields execution time of 786 ms. Still significantly higher than on the ST target unfortunately.

RE: Arduino Nano 33 BLE (nRF52840) seems not to run on max. clock speed

bjorn-spockeli — Thu, 23 Apr 2020 11:14:29 GMT

Hi Jan,

the CPU frequency of the nRF52840 is not configurable so there must be another explanation for the slower execution. Could it be that the FPU on the nRF52840 isnt enabled in the compilation settings?How are you compiling the source code? GCC?

I copy pasted this from another DevZone case:

You can enable the FPU in your compiler setting.

For Keil and IAR, you can find the "hardware floating point" option in the project settings, while in GCC you set this in the makefile:

CFLAGS += -mfloat-abi=hard -mfpu=fpv4-sp-d16

LDFLAGS += -mfloat-abi=hard -mfpu=fpv4-sp-d16

All the above methods will define __FPU_USED, which will enable the FPU in system_nrf52.c:

/* Enable the FPU if the compiler used floating point unit instructions.

* __FPU_USED is a MACRO defined by the

* compiler. Since the FPU consumes energy, remember to disable FPU use in the

* compiler if floating point unit

* operations are not used in your code. */

#if (__FPU_USED == 1)

SCB->CPACR |= (3UL << 20) | (3UL << 22);

__DSB();

__ISB();

#endif

__DSB and __ISB are intrinsics, which are C-wrapped assembly-calls.

Their function is to provide "Data synchronization barrier" and "Instruction synchronization barrier", to ensure that the write of the memory/instruction is performed before moving along in the execution of code.

The ARM infocenter website didn't want to load at my end, but I found this stackoverflow thread that explains the functionality:

http://stackoverflow.com/questions/15491751/real-life-use-cases-of-barriers-dsb-dmb-isb-in-arm