Use External Flash (QSPI) for CubeAI weights

[Raggio]

Hi all,

I would like to ask if there is a suggested procedure to move the network weights, the one generated inside network_data.h, to external memory-mapped QSPI FLASH.

I've already moved both input/output data and activation to external SRAM, but I'm facing with a compile time error when I do the same with weights (initializer element is not constant).

Before start changing things to solve it, maybe someone of you can help!

I've analysed the example FP-AI-VISION1, but it seems to move only input/output data and activations, not weights.

Thank you and kind regards.

[Raggio]

Ok, sorry, I've missed the definition at the head of FP-AI-VISION1, network_data.h, file:

#if WEIGHT_QSPI == 1 && WEIGHT_QSPI_PROGED == 0

#if defined ( __ICCARM__ )

#pragma location="ExtQSPIFlashSection"

__root /* force linker to keep variable */

#elif defined ( __CC_ARM ) || ( __GNUC__ )

__attribute__((section(".ExtQSPIFlashSection"), used))

#endif

#endif

I will add the section to the linker file and keep you posted.

If you have any other suggestion I thank you in advance.

[Tesla DeLorean · Accepted solution]

Swear I saw a similar question the other week, search here is totally useless

Some related

https://community.st.com/s/question/0D50X0000BB1CCWSQ3/how-to-make-xcubeai-use-extended-flash-and-ram-memory

https://community.st.com/s/question/0D53W00000e0TSBSA2/xcubeai-how-could-i-use-the-external-qspi-sdram-in-order-to-speed-up-the-inference-

https://community.st.com/s/question/0D53W00000903gISAQ/i-am-using-xcubeai-to-generate-c-code-for-my-ml-model-but-the-resulting-constant-data-is-larger-than-my-flash-im-using-f7-with-2mb-flash

https://community.st.com/s/question/0D53W00000pReAPSA0/validation-model-failed-to-build-on-cube-ide

[Raggio]

Thank you so much for your reply Tesla.

The solution proposed by the FP-AI-VISION1 works very well.