STM32N6 AI Model Memory Sections Wrong

[rphii]

Hello everyone,

 

I can run my model through the stedgeai / STM32Cube AI Studio and see the following:

 

But when I now go into the resulting application and build it myself, I see these memory sections, which tell a totally different story:

 

How can I fix this? I added the AXISRAMx sections (they are not provided per default) which results in them being used by 0%, so that also can’t be right.

 

Best regards, rphii

[rphii]

I guess this is what I’ll do for now, it doesn’t seem to break the inference, at least.

 

#define KB_IN_BYTES(N)    (1024UL * (N))

/* set these KB according the STM32Cube AI Studio */
#define AI_NPURAM5_SIZE     KB_IN_BYTES(400)
#define AI_NPURAM4_SIZE     KB_IN_BYTES(192)
__attribute__((section(".AI_NPURAM5")))
static uint8_t ai_npuram5[AI_NPURAM5_SIZE];
__attribute__((section(".AI_NPURAM4")))
static uint8_t ai_npuram4[AI_NPURAM4_SIZE];


/****in linker script, sections *****


  AXISRAM4 	      : ORIGIN = 0x34270000, LENGTH = 0x70000
  AXISRAM5 	      : ORIGIN = 0x342E0000, LENGTH = 0x70000


************** after .text **********

  .AI_NPURAM4 (NOLOAD): {
    . = ALIGN(4);
    KEEP(*(.AI_NPURAM4))
    . = ALIGN(4);
  } >AXISRAM4

  .AI_NPURAM5 (NOLOAD): {
    . = ALIGN(4);
    KEEP(*(.AI_NPURAM5))
    . = ALIGN(4);
  } >AXISRAM5


*************************************/

Since my axisram4 is not filled a lot, I used the rest of that space for something else, and (again) the inference still works. I therefore take this as a confirmation to my assumption that the AI studio/model/exported code simply uses the beginning of that memory… (hopefully)