That means each core has its own I guess, great

Does someone know if the ESP32-P4 has a single FPU shared among the two HS CPUs or if each HS CPU has its own FPU, so 2 FPUs?
Thx.

This flag is another interesting one https://docs.espressif.com/projects/esp-idf/en/latest/esp32/api-reference/kconfig.html#config-spiram-malloc-reserve-internal

When SPIRAM allocation with malloc are enabled, it is possible to reserve a block of RAM for "specific needs" such as DMA etc.
If this ...

I think there is this flag, which may influence the behaviour:
CONFIG_ESP32_IRAM_AS_8BIT_ACCESSIBLE_MEMORY
https://docs.espressif.com/projects/esp-idf/en/latest/esp32/api-reference/kconfig.html#config-esp32-iram-as-8bit-accessible-memory

But I can't find this option any more in idf.py menuconfig ...

If you really don't want allocations to happen in a certain area, have you tried "reserving" it by doing the allocation before the other (smaller) allocations? If you need that space temporarily, you can still try to assign memory from the "reserved" space temporarily, effectively implementing ...

heap_caps_malloc_prefer

only allows to separate internal from external (PSRAM) memory, but not DRAM from D/IRAM as far as I know.

For my application it is even more difficult as I use std::unique_ptr and would possible have to reimplement the allocator of std::unique_ptr.

Thanks for your help!

heap fragmentation and how to best avoid it

General recommendation: Reduce the number of allocations from the heap.
- Use the stack/local variables instead of heap when possible
- Use statically allocated variables
- Avoid reallocations/resizing of allocations (C++ string concatenation, std ...

That's a good resource for heap analysis.

Unfortunately nothing is said about heap fragmentation and how to best avoid it in order to get a large as possible contiguous block.

Any more ideas?

Which entries in menuconfig? I don't want to place code in the PSRAM, there is an option to allow malloc to use PSRAM, but that is RAM is much slower.

Any hints? Thx

Hi,

On ESP32:

Is there any way to tell malloc to first use memory fragments for heap allocation from the D/IRAM region instead of the DRAM region?
It seems to first consume the DRAM which in turn reduces the size of the largest available contiguous block.
This disallows some critical allocations ...