Hi Giorgio,
>Again, this "all at once" is not only a problem of causality or not, it is,
>more concretely,
>the problem of what happens if some instances are running in parallel e.g.
>in a SIMD
>architecture. All the instances running in parallel will read the global
>variables at the same
>time and will exports at the same time (more or less).
Yes, you are correct here. I thought that the standard specified
that k-rate executions are atomic, but now I see that this is not
the case. I think that this is the right answer: to make it
clear that k-rate executions cannot be parallelized.
I think the major benefit of SIMD implementation comes in the
a-rate computations (since there's much more of it in typical
instruments). So it is not such a loss to force the k-rate
instances to be executed sequentially. Would you agree?
If so, it requires only a small corrigendum. A good place for
it is list item 10 of Subclause 7.3.3.6 in my copy of the FDIS
("Decoder execution while streaming" -- clause numbers are still
changing, so it's hard to refer to them properly at the moment).
Just to say the following:
The executions of the various instrument instances at the k-rate
shall be performed as atomic operations; that is, one must
complete strictly before the next begins. This restriction is
not placed on a-rate executions of instruments.
fixes the problem, I think. It costs a small bit of efficiency,
but allows the proper semantics of inter-instance communication
with global variables.
Please tell me your opinion on this.
Best,
-- Eric
This archive was generated by hypermail 2b29 : Mon Jan 28 2002 - 11:46:33 EST