Parallella Community

Hi there,

I am trying to build a custom barrier for the Epiphany PEs. In order to test the performance of my barrier I run some tests and compared it with the default e_barrier. I visited the site with the implementation for the e_barrier:



and copied the code of 'e_barrier' so I would have a starting point for my implementation. I named this implementation 'my_barrier'.
I noticed that the e_barrier function which exists in libe-lib.a (e_mutex_barrier.o) has 5 times better performance (number of cycles) compared to my implementation taken from the link above. Notice that I use the same barrier vars and the default e_barrier_init function.
Where does this difference in performance come from? Are special flags used when compiling the e_mutex_barrier.c file in order to create libe-lib.a?
In my Makefile I use the follwing flags: -O3 -funroll-loops -ffast-math.

Try looking at the assembly language output. Or better yet, just post your whole example. I've written my own working barrier implementation but i didn't do any timing as i was focusing on code-size (it's in ezesdk).

Make sure all the code is on-core. And verify your timing code has compiled as you expect.

FWIW I would highly recommend against using either -O3 or -funroll-loops except in very specific and verified circumstances (i.e. you have to check the generated assembly is sane every time). There just isn't enough on-core space for what horrible things it often does to the code generator, and quite likely it will produce poorly performing code anyway. The compiler usually only unrolls loops when it has a constant loop bound anyway.

Thank you for answer,

I think I found the answer I was looking for. I changed the flags of e-gcc from "-O3 -funroll-loops -ffast-math" to "-O3 -mfp-mode=round-nearest". This change seems to speedup the barrier execution. See



for more informations about this particular flag.