Re: CUDA version of RANDOMN? [message #61998] |
Fri, 15 August 2008 17:58  |
hotplainrice@gmail.co
Messages: 15 Registered: July 2008
|
Junior Member |
|
|
On Aug 16, 3:15 am, wlandsman <wlands...@gmail.com> wrote:
> On Aug 15, 11:28 am, wlandsman <wlands...@gmail.com> wrote:
>
>
>
>> On Aug 15, 11:16 am, "hotplainr...@gmail.com" <hotplainr...@gmail.com>
>> wrote:
>
>>> On Aug 16, 12:28 am, wlandsman <wlands...@gmail.com> wrote:
>
>>>> On Aug 15, 10:11 am, "hotplainr...@gmail.com" <hotplainr...@gmail.com>
>>>> wrote:
>
>>>> > Hey guys,
>
>>>> > I need to write a kernel to replace the IDL RANDOMN POISSON
>
>>>> > for loop
>>>> > for loop
>>>> > for loop
>>>> > c = data[x,y,b]
>>>> > if c gt 0.0 then begin
>>>> > n = RANDOMN( seedP, POISSON=c )
>>>> > endif else begin
>>>> > n = 0
>>>> > endelse
>>>> > data[x,y,b] = n
>>>> > endfor
>>>> > endfor
>>>> > endfor
>
>>>> > Could someone point out an example code of how RANDOMN POISSON so that
>>>> > I can implement it in CUDA?
>
>>>> Your best bet is to probably look at the Poisson generating algorithm
>>>> in "Numerical Recipes in C" if you are going to implement it CUDA.
>
>>>> I have implemented the "Numerical Recipes in C" algorithm into the IDL
>>>> procedure poidev.pro at http://idlastro.gsfc.nasa.gov/ftp/pro/math/poidev.pro.
>>>> Although poidev.pro is normally slower than calling randomn(POISSON=),
>>>> it has advantages for just the problem you describe, which can be
>>>> written as simply
>
>>>> data = poidev(data)
>
>>>> rather than using a triple FOR loop. --Wayne
>
>>> Thanks for the reply. I was about to use your code until I discovered
>>> the problem of achieving this.
>
>>> c = data[x,y,b]
>>> if c gt 0.0 then begin
>>> n = RANDOMN( seedP, POISSON=c )
>>> endif else begin
>>> n = 0
>>> endelse
>
>>> I guess the only way is to code a poisson kernel and then do tiling on
>>> the data.
>
>> Yes, that does mean the code becomes 3 lines instead of 1
>
>> g = where( data GT 0, Ng ,complement=g1, Ncomplement=Ng1)
>> if Ng GT 0 then data[g] = poidev(data[g])
>> if Ng1 GT 0 then data[g1] = 0
>
>> --Wayne
>
> I still made it too complicated. poidev automatically sets any
> negative numbers to zero (since the Poisson distribution is not
> defined for negative numbers). So the original code
>
> data = poidev(data)
>
> should be fine. --Wayne
Thanks Wayne, that function helped me decrease times by half and
provided me with some code if I want to implement it in GPUs.
|
|
|