The states for premultiplied alpha blending are:
[font=courier new,courier,monospace] RenderState.SourceBlend = Blend.One; [/font]
[font=courier new,courier,monospace] RenderState.DestinationBlend = Blend.InverseSourceAlpha; [/font]

It is not correct to use SourceAlpha, One when using pre-multiplied alpha in order to do an additive blend. The thing with pre-multiplied alpha is that an additive blend is controlled purely by the alpha channel of the source.

Here is the formula for the typical SrcAlpha/OneMinusSrcAlpha blend:


OutColor=Src.rgb*Src.alpha + Dest.rgb*(1-Src.alpha)


Premultiplication merely performs the Src.rgb*Src.alpha step beforehand, so that component of the function can be replaced as so:


OutColor=Src.rgb+Dest.rgb*(1-Src.alpha)


Say you have a typical alpha-blended sprite with alpha of 1 where the colors are non-zero (ie not black, or not supposed to be transparent) and alpha of 0 with rgb=0 otherwise. The alpha is premultiplied. Wherever the color is black and the alpha is 0, the result of the operation is this:

OutColor=0+Dest.rgb*(1-0)
OutColor=Dest.rgb


Meaning that the final pixel is solely what is already in the destination, no blend. Conversely, wherever there is color and an alpha of 1:


OutColor=Src.rgb + Dest.rgba*(1-1)
OutColor=Src.rgb


Meaning that the final pixel is solely what is in the source, none of the destination affects it. Values in between 0 and 1 properly affect it.

Now this assumes that the Src.rgb*src.alpha operation was done as a pre-process pass, so that places where the alpha=0 are set to rgb=0,0,0. What if this is not the case, though? What if you have a pixel in the source with an alpha of 0 and a non-zero rgb?


OutColor=Src.rgb+Dest.rgb*(1-0)
OutColor=Src.rgb+Dest.rgb


The result is an additive blend of Src.rgb and Dest.rgb. No change of blend state is required. By varying the alpha value, you can reduce the power of the additive effect until at 1 it becomes just the standard replacement blend. So by keeping your alpha values close to 0, you can have varying strengths of additive blend, without having to overload the source alpha to 0 and losing your details.

I said close to 0. If you give your source alpha values some low range to vary within, you can get the additive blend to greater or lesser degree. It's all just how you play with alpha. But yes, if you just naively set your alpha to 0, you're going to lose whatever variance exists within the alpha channel of the texture.

Also, why are you doing the premultiplication in the shader? That's not premultiplication. That's just multiplication. The typical use-case of pre-multiplication is that you perform the Source.rgb*Source.a during asset creation, when the texture is created (via Photoshop, Gimp, or whatever); or, conversely, as some sort of asset-crunching process pass done, say, when you do a rebuild of your asset tree in preparation for a release build of your game. Or maybe at game load, if you don't have to do a whole bunch of them. However you construct your asset pipeline. The thing is, if you do the multiply in the shader, then you don't take advantage of the additive use-case of premultiplied alpha, which doesn't actually multiply Source.rgb*Source.a. You're just defaulting everything to a regular Source.rgb*Source.a blend.

However, before you go with a solution such as premultiplied alpha, why don't you figure out exactly what it is you are trying to achieve, and figure out exactly what configuration of blend modes and alpha setups you need to achieve the effect?