Optimized code for subtracting two arrays from each other

What processor are you using?

DM6437 EVM board with Code Composer Studio 3.3.

And got both the IMGLIB and VLIB libraries working.

no ideas???

I'm sure there are many pre-build libraries that do exactly what you want, and I was hoping someone who knows more about them would have answered by now. You are probably better off looking for one of those libraries.

If you want to write the code yourself, the compiler should be able to make a highly optimized loop by using the SUB4 instruction if you make sure to let the compiler know the arrays are aligned favorably, are big enough, and don't alias one another.  For instance:

cl6x -mv6400+ -O2 char_array_diff.c

#define PTR_IS_64BIT_ALIGNED(x) (_nassert(((unsigned)(x) & 0x7) == 0))

void char_array_diff(char * restrict x, char * restrict y, char * restrict a)
{
    int i;
    PTR_IS_64BIT_ALIGNED(x);
    PTR_IS_64BIT_ALIGNED(y);
    PTR_IS_64BIT_ALIGNED(a);
    for (i = 0; i < 400; i++)
        a[i] = x[i] - y[i];
}

This gives a software pipelined loop that computes 16 output chars every 3 cycles, which I believe is the fastest possible for char vector difference on this architecture.  You could also look at the _sub4() intrinsic, which would also result in the SUB4 instruction.

[Edit: make clear what the performance is --Archaeologist]

John Kofod said:

If there are optimized libraries for this, they are well hidden.

The suggestion "Archaeologist" wrote, wasn't that far from what I already had, and didn't speed up the process, damn. Though it was a good shot. Thank you.

I've tried the different compiler instructions, but it still takes around 100 ms to process those 150.000 subtractions.

The CPU can run the software pipelined loop in about 28k cycles, which is well under 1 ms.  Your bottleneck must be memory access; if that's the case, no matter how tight you make the loop, you will not see speedup.  Can you move some of these arrays to faster memory?

I'm guessing you're right. And I'm trying to do that with slicing using EDMA. This must be the right approach?

At the moment, all the data and output arrays are placed in DDR2 extern memory.

But the DM6437 doesn't have that much intern memory. And its some big arrays that needs to be processed.

Thanks anyway...

If your data is in external memory it's critical to make sure that both the data cache has been enabled and the MAR bits correctly set.  More details here:

http://processors.wiki.ti.com/index.php?title=Enabling_64x%2B_Cache

I don't believe we have any libraries for this purpose.  I've seen some 3rd party libraries (e.g. Kane Computing, etc) that do vector math operations.  Ours focus on things like convolutions, fir filters, iir filters, median filters, etc.

Best regards,

Brad

That solved it...