Limits of profiler counters

Ki-Soo Lee said:

Hello,

Please see the attached presentation, especially slide 4.

Thanks

ki

Thanks for the slides. Between posting my question and getting a reply, I did some profiling that shows there are at least some parts that use >32 bits. Below I have put a portion of the results

Address Range, Symbol Name, SLR, Symbol Type, Access Count, CPU Cycles: Incl. Total

0x3e8a53-0x3e8b74, cusum_mode_update, 406-492:cusum01.c, function, 41451, 6017339474
0x3e8a36-0x3e8a53, cusum_update, 328-399:cusum01.c, function, 49874, 6022759589
0x3e88ab-0x3e89f4, main, 64-308:moog_cusum01.c, function, 1, 2282310293

The last line is for main(), which was called once with an inclusive total of 2282310293 cycles. But both cusum_mode_update and cusum_update, called either directly or indirectly from main(), have cycle counts above 2^32. Adding 2^32 to the cycle count for main(), gives 6577277589, which looks about right. It appears that each individual timing must be < 2^32, but the totals are accumulated with more bits. At least 33 bits must be used to store 6022759589, but how large of a number can be stored is still not known. I did not see anything in the slides that mentioned more than 32 bits for storing values.

The profiler clock also appears to wrap at 2^32 since it showed a total cycle count of 2,746,485,085.

The last slide of known profiling limitations was also interesting

• CCS Profiling IS:
  • Intrusive (BP based) on HW and older simulators

The slides were last updated in 2005, so "older" is a bit relative. I am currently using CCS 3.3 and the cycle-accurate simulator for the 2812. Would real hardware be more accurate, less accurate, or about the same?

Jim

Note that the above presentation I posted was geared towards profiling on HW. Simulators work differently. There is a special 64 bit counter with simulators which is used by the CCS profiler. CCS Profiling on simulators is also non-intrusive since it does not rely on breakpoints.

Simulation is certainly easier for profiling since it has less limitations. Especially when profiling code in flash.

Ki-Soo Lee said:

Note that the above presentation I posted was geared towards profiling on HW. Simulators work differently. There is a special 64 bit counter with simulators which is used by the CCS profiler. CCS Profiling on simulators is also non-intrusive since it does not rely on breakpoints.

Simulation is certainly easier for profiling since it has less limitations. Especially when profiling code in flash.

If the simulators have 64-bit counters, they are not fully used by the profiler, at least not in CCS 3.3. Whenever I have seen loop or function cycle counts that should be greater than 2^32 cycles, wrapping occurs. Perhaps that has changed in later versions. It does appear that there are variables larger than 32 bits to store the cumulative totals within the profiler software since I have seen these exceed 2^32. How much larger and the response of the profiler on overflow were the original question, and I still do not know. I am certain that at least 33 bits are available. Thankfully my simulations have become much faster lately, so I have not had the opportunity to do more experimentation on the limits.

Jim