ippiCrossCorrValid_NormLevel_8u32f performance issue

[cio7]

Hi,

While evaluating Ipp6.1 version (x86, 32bit) I have noticed that the
ippiCrossCorrValid_NormLevel_8u32f performs significantly slower (2-3 times) than
ippiCrossCorrValid_Norm_8u32f function on identical input.
Furthermore, it does perform slower than the ippiCrossCorrValid_NormLevel_8u32f from version IPP5.1.

The behaviour was identical on both Intel Core2Duo and Intel Atom processors.
The difference in performance is somewhat dependent on the actual size parameters for source and template.

Any information on this topic would be wellcome.

[PaulF_IntelCorp]

I suspect this has to do with differences in the nature of the calculation. The two are described below, note that the significant difference is the "cross-correlation value" that is calculated. The formulae are nearly identical, but one involves substantially more calculations than the other.

Go to this page:

http://software.intel.com/sites/products/documentation/hpc/ipp/ippi/ippi_ch11/ch11_image_proximity_measures.html#ch11_image_proximity_measures

in the documentation to see a detailed description of the two different correlation values:
_tx(r,c) and_tx(r,c)

Description

The function ippiCrossCorrValid_Norm is declared in the ippi.h file. It operates with ROI (see Regions of Interest in Intel IPP). This function computes the normalized cross-correlation values _tx(r,c) for the pixels of the source image without zero-padded edges. The size of the resulting matrix with normalized cross-correlation values is

(W_s - W_t + 1) * (H_s - H_t + 1),

where W_s , H_s denote the width and height of the source image, and W_t , H_t denote the width and height of the template. The template anchor for matching the image pixel is always at the geometric center of the template. (See the formula for _tx(r,c) in section Image Proximity Measures).

Description

The function ippiCrossCorrValid_NormLevel is declared in the ippi.h file. It operates with ROI (see Regions of Interest in Intel IPP). This function computes the normalized correlation coefficients _tx(r,c) for the pixels of the source image without zero-padded edges. The size of the resulting matrix with normalized correlation coefficients is

(W_s - W_t + 1) * (H_s - H_t + 1),

where W_s , H_s denote the width and height of the source image, and W_t , H_t denote the width and height of the template. The template anchor for matching the image pixel is always at the geometric center of the template. (See the formula for _tx(r,c) in section Image Proximity Measures).