Problems to resize a frame...

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi everyone,

I would like developping a program which computes two frames: one large and one small. I try to resize the small one for fit with the large one by a simple proportionnal transform (and then add them but the problem is not there).

The code is very short :

Variables ended by HD mean it’s the large frame

Variables ended by LD mean it’s the small frame

__global__ void d_framemix(float *outputImage, float *inputDataHD, float *inputDataLD,	unsigned short widthHD, unsigned short heightHD ,	unsigned short widthLD,unsigned short heightLD)

{

    const	int xIndex = threadIdx.x + blockDim.x * blockIdx.x;

    const	int yIndexHD = threadIdx.y * widthHD + widthHD * blockIdx.y * blockDim.x;

   const float ratiowidth = (float)widthLD/(float)widthHD;

    const float ratioheight = (float)heightLD/(float)heightHD;

   const int Index = (int)((float)(xIndex*ratiowidth)+(int)((float)(threadIdx.y * widthHD * ratioheight) + (float)(widthHD * blockIdx.y * blockDim.x * ratioheight));

   

 outputImage[xIndex+yIndexHD]=inputDataLD[Index] + inputDataHD[xIndex+yIndexHD];

}

result :

819×361 27.7 KB

My problem is the small frame is well resized along x axis but not with y axis (which keeps the same size)…

Thanks for your help :)

2 
const int yIndexHD = threadIdx.y * widthHD + widthHD * blockIdx.y * blockDim.x;

works for squared images, will ghive problems if blockDim.x != blockDim.y, so better:

const int yIndexHD = threadIdx.y * widthHD + widthHD * blockIdx.y * blockDim.y;

The y part of your index calculation is wrong. Try this:

const int Index = (int)((float)(xIndex*ratiowidth)

                       +(int)((float)((threadIdx.y + (blockIdx.y * blockDim.y))*ratioheight*widthLD));

  • Mainly exchanged withHD by widthLD else your scaling will do nothing at all

  • again exchanged blockDim.x by blockDim.y

Vrah

3

Hi VrahoK !

The y axis is well resized and the x axis too but one line over two are shifted by a half width…

I keep trying to solve this problem :)

Thanks for your help !

4

Okay, just for understanding this correctly:

You have a small image with lets say width and height are half of the big one. You want to resize the small image to the size of the big one (or at least pick the pixels in this manner) and then add the values.

lets say those images are square, so widthLD = heightLD and widthHD = heightHD
What I thought the problem was:

you have the y position in your output image:

Iy = threadIdx.y + (blockIdx.y*blockDim.y)

you want to know the position in the small source image:

Iy_small = Iy * ratioheight

And now you have to multiply by the width of the small source to get the correct memory position:

y_offset = Iy_small * widthLD

Please correct me if I’m wrong but thats how I understood this thing should work.

5

You are still missing a cast to int. Two strong recommendations:

  1. use tex2D if you can, then you do not need any of that stuff

  2. Don’t put all the index calculation in one line, this is so confusing that you managed to confuse even yourself, not to speak of anyone else who might have to read your code.

6

yes you have well understand the problem I should try to compute it it through texture memory as suggest Reimar. It would be more stable…

Thank you for your helps :happy:

7

Yes, that’s exactly right, but thanks to putting everything in a single line that cast to int that would happen in the Iysmall assignment got lost.

So e.g. line 3 in the large image might become line 1.5 in the small one, which makes no sense.

So something like that might work:

const int Index = (int)((float)(xIndex*ratiowidth)

                      +(int)((float)((int)((threadIdx.y + (blockIdx.y * blockDim.y))*ratioheight)*widthLD));

But I’d like to empahsize once more: When a programmer finds obfuscated code like this in production code you better keep him from finding the person who wrote it…