Execution Performance Advice

Hakan Moray

Hi;

I have a TMDSEVL6670LE. I am using the built-in XDSv100 emulator.

What is the best way to get the execution time of a release mode compiled ".out" file ? (using TSC?)

If I connect to target DSP core from CCS and load the release mode program, does the execution time I printed in the console give me the exact time that I will see in production ? Does the emulator effect the potential execution time ?

Also, what is the best way to measure the execution time of a code segment in microseconds ?

Thanks.

Hakan.

over 10 years ago

0 Pubesh over 10 years ago

TI__Mastermind 20790 points

Hi,

The clock is the measurement option (Profiler-> Clock Setup->Count).
You will get the actual clock cycles taken to execute your range of code.
Find this wiki for more details,
http://processors.wiki.ti.com/index.php/Profile_clock_in_CCS

0 Raja over 10 years ago in reply to Pubesh

TI__Guru* 81335 points

Please use TSC (TSCL+TSCH) register for high resolution profiling. This register works identically on both c64x+ and c66x.

The following wiki has the example code:http://processors.wiki.ti.com/index.php/Porting_GPP_code_to_DSP_and_Codec_Engine#Benchmark_Code_on_DSP_using_the_Time_Stamp_Counter_.28TSC.29

Since TSCL rolls over every ~4 billion cycles, most code does not require using TSCH so you can save some overhead by not reading TSCH.

If you have a larger body of code, the compiler can help instrument the code using the entry/exit hook compiler flags. Please see section 2.16 of the compiler documentation spru187S.

Please refer below thread for more information.

http://e2e.ti.com/support/dsp/c6000_multi-core_dsps/f/639/t/130875.aspx

Thanks.

0 Pubesh over 10 years ago in reply to Raja

TI__Mastermind 20790 points

Hakan,
Did you get the execution time of a code segment in microseconds.
However this discussion of thread can helpful for future readers, If it answers your question.

0 Hakan Moray over 10 years ago in reply to Pubesh

Prodigy 130 points

Hi,

I was able to profile clock cycles successfully, thanks.

But in my FFT example, 1024 complex numbers were executed in 17 us.

In the documentation it states that :

Single precision floating-point FFT, 2048 pt. radix 4 is 14.60 us

Fixed-point FFT, 2048 pt, radix 4 is 4.64 us

How can I assure that I use fixed-point FFT ? There is only single FFT function in DSPLIB library.

Thanks in advance,

Hakan.

Processors

Processors forum

Execution Performance Advice