Band Pass: FFT or conv?

erano111 — Mon, 14 Feb 2011 06:31:04 GMT

Hello,

I'm working with set of images, each image is 320*240 pixels, divide to 17X15 blocks (each block 16X16 pixels).

I wish to do give a "focus grade" to each block.

For now I used IPP FFT with 16X16 LP filter. Does IPP conv will give faster run time ?

Thanks,

Eran

Chao_Y_Intel — Fri, 18 Feb 2011 06:31:51 GMT

Eran,

For the Conv function, it internally may use the FFT functions depending on the image size, data type, etc.

For example, in the ippiConvFuLL,
It takes some the condition like the followings:

#define _FFTUSE 63 * 47

#define MAX_DIR_KERNEL 11 * 11

if(( dstSize.width * dstSize.height < _FFTUSE )|| ( kernel.width * kernel.height < MAX_DIR_KERNEL )){/* Direct implementation here */

} else {

FFT implementation.

}

For your specific size, I suggestion you run the some demo \demo\ippiDemo.exe, or some benchmark tools at

\tools\perfsys\ to check one is good.

Thanks,
Chao

topic Band Pass: FFT or conv? in Intel® Integrated Performance Primitives