Intel® Software Network

Hi,
I've never used ippiRotate_... for 90, 180, 270 degree rotations, instead I use a combination of mirror and/or transpose functions. I never had time to check if it was faster than ippiRotate but at least I'm sure to get an exact rotation (no rounding...).

For a 90 degree rotation :
ippiMirror_8u_C1R(pSrc, srcStep, pTempBuffer, srcStep, srcSize, ippAxsHorizontal);
ippiTranspose_8u_C1R(pTempBuffer, srcStep, pDst, dstStep, dstSize);
with dstSize.width = srcSize.height...

For a 270 degree rotation :
ippiTranspose_8u_C1R(pSrc, srcStep, pTempBuffer, dstStep, srcSize);
ippiMirror_8u_C1R(pTempBuffer, dstStep, pDst, dstStep, dstSize, ippAxsHorizontal);

For a 180 degree rotation :
ippiMirror_8u_C1R(pSrc, srcStep, pDst, dstStep, roiSize, ippAxsBoth);


Matthieu

Hi Matthieu,

Thank you for your suggestion. Yes, we recomment to use ippiMirror to do special rotation like 90,180,270.

To mahantesh,

if use ippiRotate, the no-intergate interpolation, round issue are unavaidable.  ippiRotate provide one parameter IPPI_SMOOTH_EDGE to make edge look better.  But the black strips may still visiable. how width of the strips in your test? does it affect your result?

Here is small piece of code for your reference.

 IppiSize srcTempSize = {w, h};
 IppiSize dstTempSize = {h, w};

IppiRect srcTempRect = { 0, 0,w, h };
IppiRect dstTempRect = { 0, 0,h, w };
 double xTempShift = 0.0;
 double yTempShift = 1.0*w; // conterwise rotate 90 degree, origin (0, 0) will be (0, w) after rotation, so the y shift is w

/* Use this function if you need to rotate an image about an arbitrary center (xCenter, yCenter) rather than the origin (0,0), we will rotate the image by 0, 0 first, so it is not necessary to call it here
/*
//ippiGetRotateShift(0.0, 0.0, 90.0, &xTempShift, &yTempShift);

ippiSet_8u_C4R ( 0, (Ipp8u*) dstTemp->imageData, dstTemp->widthStep, dstTempSize );
status=ippiRotate_8u_C4R ( (Ipp8u*) srcTemp->imageData, srcTempSize, srcTemp->widthStep, srcTempRect,
       (Ipp8u*) dstTemp->imageData, dstTemp->widthStep, dstTempRect, 90.0,
    xTempShift, yTempShift, IPPI_INTER_LINEAR|IPPI_SMOOTH_EDGE);
//my test: the rotate operation will show 1 pixel width black line on upper of the rotated image)

Best Regards,
Ying
 

Hello Mahantesh,

Which IPP vesion are you using?  There are some known issue regarding the ippiResize() at
http://software.intel.com/en-us/forums/intel-integrated-performance-primitives/topic/66171/

Maybe same. What is symptom when resize the 160x120?

Best Regards,
Ying

Hi Ying,

Hi Mahantesh, 
Thank you for your reply. I don't rememeber there is known issue for rotating 160x120 specially.

could you send me a example?

Best Regards,
Ying

Hi Mahantesh,

Did you notice which kind of system (CPU type and OS) have the problem?  You may check your cpu type by the way I list in http://software.intel.com/en-us/forums/intel-integrated-performance-primitives/topic/68333  =>item 2

Thanks
Ying

We have found that Transpose is faster than Rotate at the 90-degree rotations, and yes, also avoid any potential rounding artefacts. I presume most would want 90-degree rotations to be exact. You can actually improve slightly on the speed by avoiding the Mirroring-calls. This can be done by effectively flipping either the source or the destination image "during" the transposing. This can be done by setting a negative stride and also adjust the associated pointer to point at the start of the last row/line of the image.
Otherwise, you may consider cutting down on memory use by skipping the temporary buffer and do the mirroring in-place in the 'dst' buffer.


Regards,

- Jay

Hi mahantesh,

I'm looking into the ippiRotate issue and get back to you if any findings.

About the ippiTranspose function, right, you should use the size same as the source input, see the description of the function' parameter in ippiman.pdf,
roiSize  - Size of the source ROI in pixels.

So for 90 degeree rotation,
the pTempBuffer is a source with (srcSize.width, srcSize.heigh),  so you will enter srcSize there.

for 270 degree rotation, you call ippiTranspose first,
the pSrc is the input source of the function, so it also needs srcSize.

Regards,
Ying

 double xTempShift = 0.0;
 double yTempShift = 1.0*w;
 IppiRect srcTempRect = { 0, 0,w, h };
 IppiRect dstTempRect = { 0, 0,h, w };

 status=ippiRotate_8u_C3R ( (Ipp8u*) srcTemp->imageData, srcTempSize, srcTemp->widthStep, srcTempRect,
       (Ipp8u*) dstTemp->imageData, dstTemp->widthStep, dstTempRect, 90.0,
    xTempShift, yTempShift, IPPI_INTER_NN|IPPI_SMOOTH_EDGE);

We rotate a 227x149 image to 149x227 image.  
I add black edge to dst image, so you can see there is the white line on the upper of image clearly. It is the line I mentioned last time. 
It is common rounding artefacts because pixel in IPP affine transforms (include rotate operations) is not square but is a point.

For workaround you can just simple to remove it by subtract 1 from yshift or xShift. for example,

status=ippiRotate_8u_C3R ( (Ipp8u*) srcTemp->imageData, srcTempSize, srcTemp->widthStep, srcTempRect,
       (Ipp8u*) dstTemp->imageData, dstTemp->widthStep, dstTempRect, 90.0,
    xTempShift, yTempShift-1, IPPI_INTER_NN|IPPI_SMOOTH_EDGE);

The line is gone.
(It is just sample, for who is using ippiRotate and has same problem, please consider such kind of  modification based on your input image.)

 
Regards,
Ying