🚀 go-pugleaf

In article <C1Cnq5.1zt@news.cso.uiuc.edu> ercolessi@uimrl3.mrl.uiuc.edu (furio ercolessi) writes:
...comments suggesting that Livermore Loops do not contain interesting
...representatives of his type of applications (classical molecular
...dynamics), partly due to lack of use of the indirect addressing
...coding technique.

I consider the Livermore Loops a necessary but not sufficient test of
compiler code generation quality.  I suspect one reason why it does not
contain much indirect addressing is that most codes from which the loops
were selected were developed in the 1970's and before.  Architectures
were different, problem sizes were smaller, etc, etc.  At time passed
we (the computing community) evolve different approaches, so our commonly
used benchmarks need to evolve and expand also.  I would welcome some
meaningful kernel benchmarks of indirect addressing.

Ideally, such a benchmark would contain:
1) Source code in reasonably portable Fortran (so we are all looking at the
same thing)
2) Information about whether the different elements of the indirectly
addressed arrays are independent (very important for adding compiler
directives to allow vector or parallel execution.
3) Several data sets for evaluating the benchmark
  The data sets should have the following characteristics
  a) A small data set which executes quickly on the slowest machine the
	application is interested in (say a second or two on a current 20
	Mips, 20 Mflops box)
    This data set is useful to the compiler developers for basic correctness
    testing, and for the architectural simulator people for studying new
    architectures.
  b) A medium data set which executes in not more than a few hours on the
	slowest machine of interest.
  c) A large data set which is large enough to properly stress architectural
	features such as cache, large numbers of processors, communication
	latencies, etc, etc.
For an indirect addressing benchmark particularly, the pattern of memory
accesses should fairly represent production environment data sets so future
systems will be tuning to maximize real data sets, not some artifically
simplified data set.  An extreme example of a bad data set would be one
where each indirect address was at a successive location in memory.

I understand the Perfect Group is actively seeking such benchmarks.
If they adopt a benchmark and put it in their suite, it will gain instant
credibility with a number of vendors.  For contact info on the Perfect
Group, see the FAQ in comp.benchmarks.