If possible, it would be nice to support periodic padding as I always manually pad periodic image before convolve. It It is slow and memory consuming. I think my application would benefit from this kind of ability.

@sebastienleclaire That shouldn't be hard I think. I wouldn't be able to say much about performance unless we run nvprof on it.

For information, this is the function I use for periodic (circularly) padding (it could be use to write and debug new code):

array paddingArrayCircularly(const array& arrayToPad,
bool isPaddingDim0, bool isPaddingDim1, bool isPaddingDim2, bool isPaddingDim3)
{
int dim0 = arrayToPad.dims(0);
int dim1 = arrayToPad.dims(1);
int dim2 = arrayToPad.dims(2);
int dim3 = arrayToPad.dims(3);

	array id0 = isPaddingDim0 ? join(0, seq(dim0 - 1, dim0 - 1), seq(0, dim0 - 1), seq(0, 0)) : seq(0, dim0 - 1);
	array id1 = isPaddingDim1 ? join(0, seq(dim1 - 1, dim1 - 1), seq(0, dim1 - 1), seq(0, 0)) : seq(0, dim1 - 1);
	array id2 = isPaddingDim2 ? join(0, seq(dim2 - 1, dim2 - 1), seq(0, dim2 - 1), seq(0, 0)) : seq(0, dim2 - 1);
	array id3 = isPaddingDim3 ? join(0, seq(dim3 - 1, dim3 - 1), seq(0, dim3 - 1), seq(0, 0)) : seq(0, dim3 - 1);

	return  arrayToPad(id0, id1, id2, id3);
}

I'm pretty sure that whatever you come up with will be faster and will consume less memory. As it stand right now, This function use approximately twice the memory inside the array "arrayToPad" to work.

@sebastienleclaire For the same sizes as you tried, periodic padding has the following global read/write efficiency.

    Kernel: void cuda::padBorders<float, af_border_type=3>(cuda::Param<float>, cuda::CParam<cuda::Param>, int, int, int, int, unsigned int, unsigned int)
         25                            gld_efficiency             Global Memory Load Efficiency      79.64%      79.64%      79.64%
         25                            gst_efficiency            Global Memory Store Efficiency      88.75%      88.75%      88.75%

I am not sure what the above mean (gld_efficiency and gst_efficiency), but thank you for the periodic padding!

That is global memory read and write efficiency metric names.

👍 👍
Need this for a client project. Adding 3.7 milestone.