CCS/TMS320F280049C: CLA Results are not Consistent

Hi,

Which compiler version are you using? Please make sure you are using 18.12.1.

There is a hardware bug with the MCMP and MSUB instructions which would be used for your integer comparison in your code.

We have a compiler work around for this. Your issue may be due to this hardware bug and your integer comparison.

Have you noticed if your comparison is not working proerly?

sal

Sal!

First, The compiler version that I am using is TI v18.1.2.LTS
Second,
By comparison, you mean, the calculation that I am doing in my code, in an if condition.
I am doing two comparisons i.e.

if(row >= Decimation_Factor)
window = (window < (h_poly_len-1)) ? (window + 1) : 0;

I have checked the results and both are working fine

Thanks,

Bilal

Sal!

Apologies for late reply. Can you tell me, where can I find the MSTF Register?
I want to add more to the above question: I have been running CLA with level 4 optimization, 100 MHz Clock, and speed vs size trade off at maximum (5). With the above mentioned setting I was able to achieve 2.5us execution time on CLA and 1us on CPU. Both run in parallel. The problems/facts that I have noticed are as follow:

-- When I run code with full optimization, I do not receive the exact answers each time.
-- When I run the code without optimization, I achieve almost the exact results with a bit of consistency. the floating point is not that much precise but it is acceptable. I have already verified my algorithm on Visual Studio and the results are exact. But the code runs slower.

-- Is there any collective problem with the debug mode and the optimization setting?

-- I know that the debug capacity decreases when we increase the optimization level. But does it decrease below a level where it starts to produce wrong results?

-- I can afford 5% variations in the results, but current results are not acceptable for me.

-- Further that, With full optimization, when I try to pass information from one processor to another in debug mode( with manual start and stop on both processors i.e. breakpoints), the information does not flows. I am referring to Cla1ToCpuMsgRAM and CpuToCla1MsgRAM. The variables that I pass, are not received correctly. They remain at their initial values. Sometimes they jump out of range. I tried to synchronize the cpu with cla based on the message/flag from the cla. I wrote a simple code. My test code was as follow:


In CLA task 1(): a variable flag is declared in Cla1ToCpuMsgRAM. When the CLA fires, triggered by the ADC interrupt (ADC is triggered by ePWM), is sets the
///////////////////////////////////////////////////////////////////////////
CLA task 1()
{
flag = 1;
processing....
processing....
processing....
processing....
processing....
processing....
processing....
processing....
flag = 0;
}
//////////////////////////////////////////////////////////////////////////
In the main cpu:

main()
{
initialization
....
...
...

for(;;)
{
if(flag)
{
led on
}
else
{
led off
}

}
///////////////////////////////////////////////////////////////////////////

Note: my entire code is based on example. Also that I have disabled the CLA_ISR(). I have also tried to store the value of flag in a variable and then check the value of that variable but it too did not work.

Although I found a better way to synchronize the CLA and CPU i.e. the ADC interrupt that fires the cla. The same interrupt can also be written in the if condition in main CPU. This synchronization was working properly.
But my concern in the above algorithm is, the information is not being shared between both the processors, with full optimization.I have no choice to turn off the optimization. the code is already too slow.

I hope I was able to project the problem more clearly.
I will also share my project fill in this thread, soon. Hopefully, if I am able to run the CLA, I am planing to make a video on my experience with CLA so others may benifit from it as well.

Thanks,
Bilal

Hi Sal!

1) With full optimization, I was still unable to debug the code. I even disabled the CLA and just collected the samples from ADC in the CPU and the stored those in an array. After a while (1024 samples), When, I tried to read the array's content in debug mode, I found all zeros. So the error is no longer related to CLA and the shared variables.

2) Later, I decided to use the Serial Communication Interface (SCI) and received data on serial monitor (Docklight) in computer. I integrated the SCI code at the bottom of my algorithm to send the final results to serial port on PC. I also turned ON the full optimization. The entire data was accurate (with 5% variation as you said, "the accuracy will decrease").

So, I conclude that the debug mode does not support optimization (completely) and it was all debugger 's fault.

3) About using the volatile keywork. I was using the volatile key word for the respective variables. But I found out that when I declare those variables as volatile, the execution times jumps to 10 - 20 folds (2.5us to 40us). So I eliminated the volatile declaration. Can you further comment on this behaviour? But never the less, without declaring the variables as volatile, I received the exact data at serial port.

4) About using stdint.h, I will look into it. Thanks for sharing that information.

Regards,
Bilal