Re: Discussion summary regarding optimisations in 2D convolution during GSoC 2021.

From: Mateusz Loskot (mateusz_at_[hidden])
Date: 2021-05-27 08:21:24

On Sun, 23 May 2021 at 08:52, prathamesh tagore
<prathameshtagore_at_[hidden]> wrote:
>
> Summary of my discussion/meet with @Olzhas on slack is as follows :
>
> 1. Determining whether a kernel can be separated as required for separable
> 2D convolution will need extra Linear algebra calculations and might
> require support from an external BLAS library. Instead, we can give option
> to user whether they want to perform separable 2D convolution and then let
> them provide required 1D kernels. We may choose to implement it directly in
> some known cases.

Generally, unless a new feature is aimed to be included in GIL as an extension,
introduction of a new dependency on an external library should be avoided.

If implementing functionality required by a feature is non-trivial,
given time and scope of your project, then we need to consider
splitting functionality of that feature into core part and extension part.
Then, the extension part can rely on external library.
Then, users will have choice, depending on their needs.

Best regards,
-- 
Mateusz Loskot, http://mateusz.loskot.net