| Re: strange behaviour of bytscl by large arrays [message #80023 is a reply to message #79963] |
Tue, 24 April 2012 13:03   |
lecacheux.alain
Messages: 325
Registered: January 2008
|
Senior Member |
|
|
On 24 avr, 19:30, fawltylangu...@gmail.com wrote:
> On Monday, April 23, 2012 10:22:08 PM UTC+2, Chris Torrence wrote:
>> Well, wrong is perhaps too strong of a word. The real word is "fast". I just did a test where I changed the internal implementation of FINDGEN to use an integer counter. The "float" counter is 4 times faster than using an integer counter and converting it to floats.
>
>> However, perhaps we could look at the size of the input array, and switch to using the slower integer counter if it was absolutely necessary. I'll give it a thought.
>
>> Thanks for reporting this!
>
>> Cheers,
>> Chris
>> Exelis VIS
>
> I could not reproduce this 4x slowdown. The integer counter + conversion method is only 30% slower in the following C test program (Intel Core i5-2500, 64 bit Linux):
>
> #include <time.h>
> #include <stdio.h>
> #include <stdlib.h>
>
> double timediff(struct timeval* tv1, struct timeval* tv2)
> {
> return tv2->tv_sec-tv1->tv_sec+(tv2->tv_usec-tv1->tv_usec)*1e-6;
>
> }
>
> int main()
> {
> int n=1000000000, j;
> float* x=malloc(n*sizeof(float));
> float f;
> struct timeval tv1, tv2;
>
> gettimeofday(&tv1, NULL);
> for (j=0; j<n; j++) x[j]=j;
> gettimeofday(&tv2, NULL);
> printf("integer counter: %lf %f\n", timediff(&tv1, &tv2), x[n-1]);
>
> gettimeofday(&tv1, NULL);
> f=0.0;
> for (j=0; j<n; j++) x[j]=f++;
> gettimeofday(&tv2, NULL);
> printf("float counter: %lf %f\n", timediff(&tv1, &tv2), x[n-1]);
>
> }
>
> Also, IDL help says:
>
> The FINDGEN function creates a floating-point array of the specified dimensions. Each element of the array is set to the value of its one-dimensional subscript.
>
> So it should be equivalent to float(lindgen()), as one-dimensional subscript is an integer.
>
> But I don't want to convince you, I can accept that it is a feature :-)
>
> regards,
> Lajos
>
>
By using the IDL profiler with :
l = lindgen(100000)
f = findgen(100000)
fl = float(l)
I get:
findgen -> 0.805 s.
lindgen -> 0.894 s.
float -> 0.209 s.
showing that FPU addition is faster than CPU's one, and type
conversion is a relatively slow process.
alain.
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
