Implement new Vector methods correctly to take advantage of performance improvements about imagesharp HOT 6 CLOSED

From @voidstar69 on January 23, 2016 14:27

@JimBobSquarePants is this the sort of change that you are looking for?

Changing Color.Lerp from this:

    public static Color Lerp(Color from, Color to, float amount)
    {
        amount = amount.Clamp(0f, 1f);

        if (Math.Abs(from.A - 1) < Epsilon && Math.Abs(to.A - 1) < Epsilon)
        {
            return from + (to - from) * amount;
        }

        // Premultiplied.
        return from * (1 - amount) + to;
    }

to this:

    public static Color Lerp(Color from, Color to, float amount)
    {
        amount = amount.Clamp(0f, 1f);

        if (Math.Abs(from.A - 1) < Epsilon && Math.Abs(to.A - 1) < Epsilon)
        {
            return Vector4.Lerp(from, to, amount);
        }

        // Premultiplied.
        // TODO: not sure how to use Vector4.Lerp here
        return from * (1 - amount) + to;
    }

might improve performance by making use of the Vector4.Lerp method (need to measure performance to be certain).

from imagesharp.

You're absolutely correct. I can use the built in Vector4.Lerp in that situation. The Color minus operator utilizes the backing vector but we might as well cut out the middle man.

The code, I believe, would be:

return Vector4.Lerp(from.backingVector, to.backingVector, amount);

For premultiplied alpha we'd probably have to fall back to the non SIMD approach.

I'd imagine it would be a very tiny performance improvement but every little most certainly helps.

from imagesharp.

I see usage of Vector4 in a bunch of places now, are there any remaining areas which we need to convert? 😄

from imagesharp.

I think it might be possible to vectorize bits and bobs throughout, maybe in colorspace conversions or running through lists. I know the subsampling scaling in jpeg decoder is vectorised in libjpeg turbo also.

I just don't know the Numerics API's well enough though to figure out what could benefit.

from imagesharp.

Subsampling is not an issue, the only test input where it goes over 1% is the cmyk watch:

from imagesharp.

Closing this. @antonfirsov has done incredible work to implement vectorization throughout the library.

from imagesharp.