Re: Extend slices for n dimensional arrays for Ada 202X

[Robert Eachus <rieachus@comcast.net>]

On Thursday, January 26, 2017 at 1:29:07 PM UTC-5, Randy Brukardt wrote:

> So I don't see this happening - it doesn't make sense in the context of Ada 
> implementation strategies.

Randy, I hope you are confused.  First, if you are copying one array to another with correct (possibly sliding) bounds, you should check the bounds, then do whatever copy is fastest on the current hardware.  (This is moving from 64 to 128 bits at a time, and current hardware supports single instruction 256-bit and even 512-bit moves.)  Yes, technically, if in strict mode and the floating-point type you are using does not use non-signalling infinities, you might need to discover where an exception occurs and do a partial copy.  But if this is an object to object assignment, the best code with signalling infinities would be to scan the data for signalling entries, then do the copy.  Having to back out half the copy is messy. 

For decades I've been doing linear algebra in Ada where sometimes it helps to have some arrays in row-major order and others in column-major order.  Take simple matrix multiplication for example A * B --> C.  If I declare B as  "with Convention => Fortran;" now I get significantly faster results from:
      for I in A'Range(1) loop
        for J in B'Range(2) loop
          Temp := 0.0;
          for K in A'Range(2) loop
            Temp := Temp + A(I,K) * B(K,J);
          end loop;
          C(I,J) := Temp;
        end loop;
      end loop;

If you try it for reasonably large I, J, and K, you should find that the payoff is large compared to the cost of transposing B once. (It is possible to transpose B in place (the J=K case is a lot easier than the general case) but I'm used to having way more (virtual and real) memory than I need 

In this case, B is accessed in the proper fashion.  But the "ugly" computations you complained about would be needed to iterate over B in row major order.  I just assumed that all compilers generate the correct code for the hard case, which I don't use anyway. ;-)