Fine-tuning guidance for S32K396 (like AN5191 for MPC5777C)

[ricardofranca]

Hello,

Feel free to jump to the questions in the end of this post, as my experiment description is quite verbose and may be superfluous...

I have been doing some benchmarking with the S32K396-BGA-DC1 running in S32DS for S32. My reference for timing behavior expectations is a MPC5777CEVB running the same program in S32DS for Power Architecture - I usually run it at 264MHz but also tried it a couple of times at 300MHz (not often because my chip is the old PN that is not supposed to run at 300MHz).

I believe I am running the MPC5777C at its optimal configuration (stack in cache, optimized Flash read/prefetch) as recommended by AN5191. My program takes some 750us to run when I let the processor run at 264MHz and some 650us when the processor runs at 300MHz (it is a sort of sequential code which does not rely very heavily on caches).

As I am a beginner on S32K3, I inserted my code in the Dio_Example_S32K396 example available in "S32K396 AUTOSAR R21-11 RTD 4.0.0 P14 D2403 Example Projects" - its linker directives allocate code/constants in Flash, data in SRAM and stack in DTCM. I also used the configuration tool to set the following clocks (using the "Mode A+" values presented in Table 122 of the S32K39/37 reference manual):

PLL_PHI0 = 320MHz

PLL_PHI1 = 480MHz

CORE_CLK = 160MHz

AIPS_PLAT_CLK = 80MHz

AIPS_SLOW_CLK = 40MHz

HSE_CLK = 80MHz

DCM_CLK = 40MHz

LBIST_CLK = 40MHz

QSPI_MEM_CLK = 320MHz

CM7_CORE_CLK = 320MHz

While I expected the new MCU to beat the old one due to its faster and more efficient core, I was surprised to see that it ran my program faster than the 264MHz MPC5777C but slower than the 300MHz MPC5777C. Both processors had very similar performance when I allocated all my data in the CPU0 DTCM, which is not something that would work for a larger application that would use most of the SRAM.

I did not configure the Flash by hand but it seems the configuration tool (or the example startup code) had set it wait states to 5 (which seems to match the 160MHz CORE_CLK), although it did not enable prefetching. I suppose this is a bit critical for my code, as disabling prefetching in the MPC5777C caused a 100us increase in its execution time.

Having said this all, my questions are:

- Can I enable prefetching using the S32DS configuration tool? If this is not possible, what would be the cleanest way of doing this? I do not know where my _start function comes from...

- Given the 5 wait states of the Flash (vs. 3 of the MPC5777C when its peripheral clock runs at 132MHz or even 150MHz), is it correct to assume the MPC5777C Flash memory performs better than the S32K396 one?

- Is there any document like the AN5191 that provides guidance for extracting the best performance of the S32K396?

Thanks!

Ricardo

[petervlna · Accepted solution]

Hello,

- Can I enable prefetching using the S32DS configuration tool? If this is not possible, what would be the cleanest way of doing this? I do not know where my _start function comes from...

No — prefetching is not configurable via the S32DS Configuration Tool (as of the latest RTD and SDK versions).

You’ll need to manually configure the Flash prefetch settings in your startup code.

- Given the 5 wait states of the Flash (vs. 3 of the MPC5777C when its peripheral clock runs at 132MHz or even 150MHz), is it correct to assume the MPC5777C Flash memory performs better than the S32K396 one?

• MPC5777C:

  • Based on Power Architecture, with dual e200z7 cores.
  • Has a more advanced Flash controller with better pipelining and prefetch capabilities.
  • Often used in high-performance automotive and industrial applications.

• S32K396:

  • Based on Arm Cortex-M7, optimized for flexibility and integration.
  • Flash controller is simpler, and performance is more sensitive to wait states and cache configuration.

- Is there any document like the AN5191 that provides guidance for extracting the best performance of the S32K396?

This is the closest equivalent to AN5191 for the S32K3 family. It provides:

https://www.google.com/url?sa=t&source=web&rct=j&opi=89978449&url=https://community.nxp.com/pwmxy876...

Best regards,

Peter