improvements in Haar CascadeClassifier: 1) use CV_32S instead of CV_32F for the integral of squares (which is more accurate and more efficient); 2) skip the window if its contrast is too low

1. Use built-in mad() instead of += and *.
2. For stump stages, if weight.z == 0, avoid unnecessary calculation
because some features only have 2 rectangles..

It could improve
OCL_Cascade_Image_MinSize_CascadeClassifier.CascadeClassifier/*
about 10% mean values.
Signed-off-by: Yan Wang <yan.wang@linux.intel.com>

Signed-off-by: Yan Wang <yan.wang@linux.intel.com>

Signed-off-by: Yan Wang <yan.wang@linux.intel.com>

and unsynchronized.
Signed-off-by: Yan Wang <yan.wang@linux.intel.com>

Signed-off-by: Yan Wang <yan.wang@linux.intel.com>

Signed-off-by: Yan Wang <yan.wang@linux.intel.com>

the parameter variable.

It could improve the performance of
OCL_Cascade_Image_MinSize_CascadeClassifier.CascadeClassifier/*.
Especially,
OCL_Cascade_Image_MinSize_CascadeClassifier.CascadeClassifier/15
OCL_Cascade_Image_MinSize_CascadeClassifier.CascadeClassifier/16
could be improved about 2% in Intel platform.
Signed-off-by: Yan Wang <yan.wang@linux.intel.com>