topic Separable convolution in Intel® oneAPI Math Kernel Library

Separable convolution

rahul_modi — Fri, 06 Apr 2012 15:24:02 GMT

Hi,

I need to know if the current implementation of multi-dimensional convolution allows convolution by seperability (for faster computation), if the input kernel is separable (http://www.dspguide.com/ch24/3.htm).

I know I can implement this myself using repeated application of one-dimensional convolution, but this is just to avoid reinventing the wheel in case MKL already offers such a method.

Thanks,

Rahul Modi

Separable convolution

Victor_Gladkikh — Mon, 09 Apr 2012 07:49:36 GMT

MKL provides 2 modes for compiting convolution. MKL can compute convolution directly or by using fast Fourier transform.Youcould use the second way to compute convolution becausein general it requires lessfloat pointer operations thencomputing convolution directly.

Separable convolution

rahul_modi — Mon, 09 Apr 2012 12:53:27 GMT

Hey Victor,

Thanks for replying back. I am aware of MKL allowing two modes for convolution and as of nowI am using the FFT method. My question is more toward a specific case, of separable kernels used for convolution. (http://blogs.mathworks.com/steve/2006/10/04/separable-convolution/). I need to know if MKL incorporates any such optimization for N-D convolutions.

Thanks,
Rahul modi.

Separable convolution

Victor_Gladkikh — Tue, 10 Apr 2012 03:28:48 GMT

No, MKL doesn't provide such functionality, alsothis optimizationdoesn't help for FFT mode because separate kernels doesn't significantly reduce nuber of operation in this mode.

Separable convolution

rahul_modi — Tue, 10 Apr 2012 03:31:45 GMT

Thanks Victor. That helps.
-Rahul