malloc

[Albert Netymk]

Is it possible to use malloc on Epiphany architecture? If so, is there any small examples?

[ysapir]

Albert
You can definitely use malloc(), like any other standard library function. However, using the malloc() function, *printf() and a few more that rely on dynamic memory management, will most probably not fit into the code space in the core. It is then necessary to use the fast.ldf linker script for placing libc calls in external mem, or legacy.ldf, for placing all code in external mem.

Also make sure that the heap, defined in the LDF, is where you think it is.

[Albert Netymk]

Thank ysapir for the quick reply. Inspired by your explanation, I reviewed the sdk manual to confirm that all three linker files (legacy, fast and internal) don't place heap into the 32k SRAM inside each other. For high performance applications, it's tricky to manage memory efficiently. Therefore, it's beter just stick with static variables.

One reminder, though, is the address operator '&' only returns the local address if this variable is in internal SRAM, and I had to call e_xxx apis to convert it to global ones manually. Not sure, if there's any overhead using global address in the own core.

[ysapir]

Using local space pointers should be considered a preferred practice where possible, b/c global pointers go through the mesh node router on their way back to the local memory. In some abnormal and very rare situations, this may cause program instability.

[Albert Netymk]

It's not very surprising that there's indeed some overhead using global addresses. In chapter 5 eMeshTM Network-On-Chip of Epiphay architecture, it mentioned something like routing. Where could I find any documented proof that the overhead exists? (Need it for my scientific paper.)

[notzed]

It's possible to write a very small dynamic memory implementation that would fit quite easily. With such a limited space it may be difficult to make use of it though.

[CIB]

*casts thread necromancy*

Thank you, exactly the info I needed. With such limited memory, but high in-chip bandwidth, maybe a more naive memory management approach, e.g. using bitmaps, would be efficient?

[Gravis]

[CIB]

[LamsonNguyen]

How to be creative when faced with limitations......now where have I heard of that before........?