Hi Yann,
On 05/27, Yann Orlarey wrote:
>
> Since it is a typical "reduce"
> operation, I suggest having "reduce" in the name.
Agreed!
> We can imagine having
> three different versions depending if we start the reduction from the top
> outputs, the bottom outputs, or in parallel.
Agreed,
> // parallel reduce: starts reductions in parallel on both halves
> parReduce(op,1) = si.bus(outputs(op));
> parReduce(op,2) = op;
> parReduce(op,n) = parReduce(op,int(n/2)), parReduce(op,n-int(n/2)) : op;
Yann, sorry for the spam, I am sure this is obvious to you, but let me
comment this change anyway. Lets ignore the fact that this assumes that
inputs(op) == outputs(op) * 2.
IMHO this change makes sense anyway, even if inputs == 2 and outputs == 1,
at least for "n-ary" functions like max/min/etc. Simply because this is
more compiler-friendly.
I mean... I don't know how to explain this in English, so let me provide
2 trivial examples,
test1.dsp:
// this is what we currently have
parallelOp(op,1) = _;
parallelOp(op,2) = op;
parallelOp(op,n) = op(parallelOp(op,n-1));
process = parallelOp(max,1024);
test2.dsp:
// this is what you suggest
parallelOp(op,1) = _;
parallelOp(op,2) = op;
parallelOp(op,n) = parallelOp(op,int(n/2)), parallelOp(op,n-int(n/2)) :
op;
process = parallelOp(max,1024);
Now, on my laptop,
$ time faust -o /dev/null -t 1000000 /tmp/test1.dsp
real 3m20.498s
user 3m20.433s
sys 0m0.167s
$ time faust -o /dev/null -t 1000000 /tmp/test2.dsp
real 0m0.128s
user 0m0.117s
sys 0m0.010s
as you can see there is a difference ;)
Oleg.
_______________________________________________
Faudiostream-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/faudiostream-users