C6678, L2RAM performance issue

Hi Fabio,
I have assigned this thread to our expert. Thank you for your patience.

Before we continue let's do the following:
Disable L1D cache and L1P cache and repeat the experiment. Report back what are the performances with disabled caches

Ran

Hi
disabled L1D and L1P cache

time are equal for both buffer , 2400 us

Fabio

OK

The second buffer the code is already in the L1P cache, so it takes less time.

And you may have other variables in L2 that may cause cache trashing for L1D


To check for the first item do the following:

Enable only L1P - not L1D

Now run the first routine twice (I mean with the vector in the first part of L2) and then the second one (the vector in the second half of L2) and compare the numbers

If they are the same, this is l1P issue

Report to me again

Thanks

Ran

Sorry but  I forgot  to say one important thing

I did 3 different tests

first (original pseudocode source)

main()

{

int32 buffer1[ size]

int32 buffer2[ size]

//fill buffer1 and buffer2

myFunction( buffer1)    //800 us

myFunction( buffer2)   //700 us

}

second

main()

{

int32 buffer1[ size]

int32 buffer2[ size]

//fill buffer1 and buffer2

myFunction( buffer2)    //700us

myFunction( buffer2)   //700 us

}

third

main()

{

int32 buffer1[ size]

int32 buffer2[ size]

//fill buffer1 and buffer2

myFunction( buffer1)    //800 us

myFunction( buffer1)   //800 us

}

So buffer2 takes less time even if I never call myFunction( buffer1)

Sorry but I think I badly explained the problem, so let's start from scratch

My original code actually looks like this

int32 buffer1[ SIZE]
int32 buffer2[ SIZE]
int pingpong = 1

//This function is called periodically after an IRQ
void execute()
{
   if(pingpong ) myFunction( buffer1) //800 us
   else myFunction( buffer2) //700 us

   pingpong != pingpong
}

So I use alternately buffer1 and buffer2, the time sequence is 800,700,800,700,800,700,800,700........

I modified the code as follows ( using always buffer1)

if(pingpong ) myFunction( buffer1)
else          myFunction( buffer1) //always buffer1

and the time sequence is 800,800,800,800,800,800........

I modified the code as follows ( using always buffer2)

if(pingpong ) myFunction( buffer2)
else          myFunction( buffer2) //always buffer2

and the time sequence is 700,700,700,700,700,700,700,700......

so as you can see, buffer1 takes always 800 us while buffer2 always 700.

Then You suggested to Disable L1D cache and L1P cache and I always have 2400,2400,2400,2400,2400,2400,2400,2400,2400.... in the 3 samples above

Additional info : I tried the following

instead of using buffer1 @0x800000 , and buffer2 @0x80E100
I set buffer1 @0x811000 , and buffer2 @0x81F100

In this case I get 700,700,700,700,700 ......