Instruction and Memory statistics (Performance counters)

[vanchiramani]

Hello Forum members!

I have written applications using coprthr-2 API for parallella. I am interested in knowing the number and type of instructions executed per core.

I would also like to know the number of accesses per program variable and how many times a given e-core is trying to access its neighbours.

Is there any way to get these performance counter values?

Any help is really appreciated. Thanks a lot in advance!
-V Vanchinathan

[nickoppen]

Hi,

I covered some benchmarking on my latest blog post: . I hope it gives you some pointers for what you want to measure.

There is also where you can see the instructions for a given chuck of source code. Choose Epiphany gcc 4.8 on the right hand panel and that should give you an idea of what your asm code looks like.

Good luck.

nick

[vanchiramani]

Hi Nick

Thanks a lot for your prompt reply.

I am actually interested in finding the dynamic instruction count to see if an application has lot of computation instructions. The INT_ALU and FP_ALU timers can give this value. However, i cannot find any flags, that can give me the number of loads / stores executed in a program. The EXT_LOAD_STALL timer flag does not capture this.

It will be great if you can point me to the mesh timer document as I can find this information in the manual.

Thank you!
V Vanchinathan

[sebraa]

You can't. The timers are designed as debugging tools, not as perfect benchmarking tools.

Loads and stores are IALU instructions (see arch ref, p.35), so they appear in the IALU Instruction counter.
Remote loads produce external stalls and can be seen with the Ext Fetch/Data Stall counters.
Remote stores access the cMesh and should appear in the Mesh Traffic counter.