(j3.2006) an intrinsic for SORT()

[Robert Corbett]

It depends on the compiler.  Oracle Fortran uses *gemm and *gemv to implement MATMUL unless it thinks it can do better (such as for small constant shape arrays).  In simple cases, it will use the transpose flag.

Bob Corbett

> On Feb 6, 2018, at 11:59 AM, Van Snyder <van.snyder at jpl.nasa.gov> wrote:
> 
>> On Tue, 2018-02-06 at 14:46 -0500, Steve Lionel wrote:
>>> On 2/6/2018 2:45 PM, Van Snyder wrote:
>>> 
>>> If I write MATMUL(TRANSPOSE(A),B) will most processors create a temp for
>>> TRANSPOSE(A), or will they do what DGEMM does and handle the TRANSPOSE
>>> as a flag that changes how the multiply is done?
>>> 
>> That depends on whether or not such usage is in a benchmark...
> 
> So I should be using DGEMM, just to be sure.  TRANSPOSE is only O(n^2),
> seemingly small compared to the O(n^3) for the multiply, but the
> coefficient is large because it has to visit the storage allocator.
> 
> In a language intended for science and engineering, not accounting,
> adding a "transpose" flag for either argument would seem to be a higher
> priority than a SORT intrinsic.  But, yes, I do sort stuff in my
> satellite data analysis codes.  It wasn't hard to find sort routines.
> It isn't hard to write correct ones.  I wrote several for our math
> library decades ago.  Who among us hasn't?  I chose insertion sort
> because the array is almost sorted.  A runtime algorithm selector would
> just add overhead.
> 
>> Steve
> 
> _______________________________________________
> J3 mailing list
> J3 at mailman.j3-fortran.org
> http://mailman.j3-fortran.org/mailman/listinfo/j3