AM6442: DhryStone benchmark results

All A53's score about 3DMIPS/MHz with a modern compiler. We run the score in Linux and documented in the performance guide https://software-dl.ti.com/processor-sdk-linux/esd/AM64X/latest/exports/docs/devices/AM64X/RT_Linux_Performance_Guide.html#dhrystone (looks like this release it is 2.9DMIPS/MHz, I see it vary from 2.90 to 3.05). This is part of the Linux filesystem, in /usr/bin/runDhrystone. This is what we run in nightly tests and is the base for what we document in the performance guide.

Mustafa Filiz said:

I was trying to benchmark A53 SMP with dhrystones example.

There is table of result at the bottom of the link 

From the above it looks like you are trying to run A53 freeRTOS SMP? The link (https://software-dl.ti.com/mcu-plus-sdk/esd/AM64X/09_00_00_35/exports/docs/api_guide_am64x/SMP_FREERTOS_GUIDE.html ) is to an experimental demo of running freeRTOS SMP. The score looks like is not as good as one would expect for A53.

Mustafa Filiz said:

i think there is overflow at  ( numThreads * numIterations * Mic_secs_Per_Second).

Mustafa Filiz said:

the result of multiply should be stored in uint64_t.

Using 32-bit unsigned integer looks like a bug to to me as well. Most implementations, including what you can see at https://github.com/Keith-S-Thompson/dhrystone and TI SDKs in Linux use a float type for this to avoid overflow. As a reminder the The Dhrystone benchmark itself is very old (from the 1980s) and without any official or certification body, so there is no golden reference.

I'll file a bug on the A53 demo Dhrystone in MCU+ SDK.

  Pekka