Retrieve multiple ArrayFire subarrays from min/max data points

Question

I have an array with sections of touching values in it. For example:

0 0 1 0 0 0 0 0 0 0
0 1 1 1 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 2 2 2 0 0
0 0 0 0 0 0 0 2 0 0
0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0
0 0 3 0 0 0 0 0 0 0
0 0 3 0 0 0 0 0 0 0
0 0 3 0 0 0 0 0 0 0

from this, I created a set of af::arrays: minX, maxX, minY, maxY. These define the box that encloses each group. so for this example:

minX would be: [1,5,2] // 1 for label(1), 5 for label(2) and 2 for label(3)
maxX would be: [3,7,2] // 3 for label(1), 7 for label(2) and 2 for label(3)
minY would be: [0,3,7] // 0 for label(1), 3 for label(2) and 7 for label(3)
maxY would be: [1,4,9] // 1 for label(1), 4 for label(2) and 9 for label(3)

So if you take the i'th element from each of those arrays, you can get the upperleft/lowerright bounds of a box that encloses the corresponding label.

I would like use these values to pull out subarrays from this larger array. My goal is to put these values enclosed in the boxes into a flat list. In GPU memory, I also have calculated how many entries I would need for each box using the max/min X/Y values. So in this example - the result of the flat list should be:

result=[0 1 0 1 1 1 2 2 2 0 0 2 3 3 3]

where the first 6 entries are from the box

 ______
|0 1 0 |
|1 1 1 |
 ------

the second 6 entries are from the box

 ______
|2 2 2 |
|0 0 2 |
 ------

and the final three entries are from the box

 ___
| 3 |
| 3 |
| 3 |
 ---

I cannot figure out how to index into this af::array with min/max values in memory that resides on the GPU (and do not want to transfer them to the CPU). I was trying to see if gfor/seq would work for me, but it appears that af::seq cannot use array data, and everything I have tried with using af::index i could not get to work for me either.

I am able to change how I represent min/max (I could store indices for upper left/lower right) but my main goal is to do this efficiently on the GPU without moving data back and forth between the GPU and CPU.

How can this be achieved efficiently with ArrayFire?

Thank you for your help

Answer 1

How did you get there so far? which language are you using?

I guess you could be tiling the results to 3rd dimensions to handle each regions separately and end up with min/max vectors in GPU memory.