Is it possible to use VEX for accumulated calculations?

[magneto]

Basically I am wondering if it's possible to calculate accumulated values using VEX?

Let's say a basic example of calculating the average or sum of all points.

Something like:

vector sum;

sop vexop()
{
    sum += P;
}

AFAIK VEX doesn't allow this kind of shared/static variables outside functions, right?

How would one do something like this? Can you do it using a detail attribute? I couldn't find a way to access them in VEX.

So I thought of doing this:

sop vexop()
{
    if (ptnum == 0)
    {
        vector sum;
        for (int i=0; i<Npt; i++)
        {
            vector p;
	    import ( "P", p, 0, i );
            sum += p;
        }
    }
}

Is this reasonable? Also how can I get this value out reliably? Right now I just print it. If I stored it in an attribute, the last point would have the correct value (if VEX loops points in order)?

Thanks

Edited March 20, 2012 by magneto

[symek]

Basically it won't work, unless you're in POPs, or inside SolverSOP. VEX virtual machine has a very specific architecture:

1) Copy attributes from geometry into Vex memory arrays.

2) Split array into chunks.

3) Run vex code in one thread per chunk, one time per point.

4) Copy all modified arrays back onto geometry.

Accumulating values, recursion etc, are possible only outside VEX.

As to attributes, you can access detail's in the same way as points' attributes. VEX will find it by name.

[magneto]

Thanks Symek. I understand it now.

So you wouldn't use it even for 1 time accumulation-like calculations? By this I mean things like sum/average but only once, but not time/history dependent like solver SOP.

I wanted to use VEX because Python version way very slow, so thought VEX might be much faster, and didn't want to dip into HDK yet

[Guest mantragora]

...and didn't want to dip into HDK yet

Heh, well, HDK is not a one day trip for sure . But you can try InlineCPP. That's like one-two week trip to get to now each other a little.

Edited March 21, 2012 by mantragora

[symek]

Heh, well, HDK is not a one day trip for sure . But you can try InlineCPP. That's like one-two week trip to get to now each other a little.

+1

Not to mention, you can use VEX in Python also!

[magneto]

Yeah I need to make some solid time before I cross the HDK waters

@Symek, I knew I could call VEX in Python, but if I can't use VEX for this code, then it doesn't matter for me for this case.

Inlinecpp looks like something I might try

[tjeeds]

Sorry if I'm stating the obvious here but using Attribute Promote sum attributes to Detail class is pretty handily the speed and ease of use champion. If you can work that in as a precursor to whatever you're putting together you'll be well ahead of any other solution I can imagine.

[magneto]

Thanks for mentioning that trick. I know it but here I said sum/average for simplification. The actual calculation is way more complicated (500 lines)

[edward]

Is this reasonable? Also how can I get this value out reliably? Right now I just print it. If I stored it in an attribute, the last point would have the correct value (if VEX loops points in order)?

Offhand, the hack should work in theory to me. Have you tried some simple examples?Of course, it's not clear to me if this will be fast at all.

[magneto]

Thanks Edward, I tried. It seems to work In fact it was pretty fast, still have to measure because I haven't finished it completely.

But one thing I would appreciate if you can clear for me is if VEX loops through points in order? Like is it guaranteed for 0 to be executed before 1, and 1 before 2, etc?

This is not relevant for my case because I use my own for loop, but still thought I would ask

[edward]

No, points in VEX are not looped in order. If you think about it, there's no way for it to do that for multi-threading. Of course, you can always get lucky. The points are processed in "chunks" at a time so for a small number of points less than the "chunk" size, it won't be threaded and you may get it processed in ascending/descending order.

[magneto]

Thanks Eward, good info

Would it be just like you said, if you set your VEX node to "No Threading"? I thought maybe that one might be handled differently.

[anim]

VEX is based on SIMD architecture

each instruction is run on all points before the next one

so there is no way to access result value of any points to compute another, it simply doesn't exist yet, the only data you have access to are from input geometry

and this is regardless of thread settings

[edward]

Would it be just like you said, if you set your VEX node to "No Threading"? I thought maybe that one might be handled differently.

<shrugs> It may work today (or not) but expect the behaviour to change at any time. Basically, the order is *undefined*, so don't rely on it.

[magneto]

Thanks guys, appreciate the extra info