Deferred Lighting - How get light-space positions?

Unity Engine
1

I’m trying to do some custom light attenuation in the deferred pipeline. Generally it involves reading a spotlight’s cookie in a unique way.

In order to do this, I need to get the light-space coordinates of a fragment. Trouble is, it appears _LightMatrix0 doesn’t mean the same thing as it does in the forward path (despite what the documentation says).

In other words, this line…

mul(_LightMatrix0, float4(wpos, 1));

…doesn’t look like it’s producing light-space coordinates for wpos.

Any thoughts?

EDIT:
As of Unity 5.4: LightMatrix0 was renamed to unity_WorldToLight

2

It works for me. Can you give more details? Where are you using _LightMatrix0, what light type, how you compute cookie coordinates?

3

Deferred Pipeline, Spotlight.

Well, for starters, I need to multiply the output of

mul(_LightMatrix0, float4(wpos, 1));

…by -1 to get close to what I’m looking for.

The z-coordinate is in unity units (great!), but the x, y coordinates are normalized to the spot angle. So [1, 0, 0] maps to a position at the limit of the spot-lights angle. which doesn’t help me. I need light-space coordinates.

For cookie coordinates, I need to figure out the angle from the local x-axis ( [1, 0, 0] in light space ) to the xy vector ( [x, y, 0] in light space ). To get a “horizontal” angle, this will map to my u coordinate in the light’s cookie. Easy enough with the result.

But, for the vertical angle ( Angle between [0, 0, 1] and [x, y, z] in light space ). This will map to my v coordinate in the light’s cookie.

The end goal is I need to be able to compare the forward distance to lateral (x, y) distance, since the units are not consistent, I can’t use them.

4

Ok, for spot light, _LightMatrix0 is a projection matrix that is used to project 2d texture inside spot light’s cone (it has different meaning for other light types). x,y coordinates are from 0 to 1 inside the cone. You can do something like this to get the cookie color:

float4 cookieUv = mul(_LightMatrix0, float4(wpos, 1));
cookieUv.xy /= cookieUv.w;
float4 cookieColor = tex2D(_Cookie, cookieUv.xy);

This should be the same for forward and deferred. It looks like you want a matrix that will transform from world space to light space. Without the projection? I don’t think there is one.

5

Projection is fine, as long coordinates in x, y, z are all relative to each other.

mul(_LightMatrix0,float4( worldPos, 1)) in the forward pipeline gets me coordinates that are all proportional to the range of the spotlight.
Forward Shader logic (In ForwardAdd pass of the fragment shader)

#ifdef SPOT
    return mul(_LightMatrix0,float4( worldPos, 1));
#endif

(This I can use, since the x, y and x values are all normalized and can be compared to each other to get angles)

mul(_LightMatrix0, float4(wpos, 1)) * -1 in the deferred pipeline gets me coordinates that are all proportional to the angle of the spotlight, or, a projection in other words. In addition, I have to multiply the result by -1.
Deferred Shader logic (In CalculateLight() of Internal-DeferredShading)

float4 lightSpacePos = mul(_LightMatrix0, float4(wpos, 1)); 
    return half4(lightSpacePos.x, lightSpacePos.y, 0, 0) * -1;

(This I can’t use, I can use the x, y coordinates to get angle yes, but without the spot angle of the light I have no way of comparing z against the x, y coordinates)

See the attached image. Left quad = lit w/ a deferred pipeline, Right quad = lit w/ a forward pipeline. Both of them use lighting shaders that simply return the bolded code snippets as colours (light-space coordinates) if the matrix was identical, I’d expect there to be no seam/differences between the quads.

2624043--184311--Capture.PNG.jpg
2624043--184311--Capture.PNG.jpg1255×836 133 KB

6

Btw, thanks for your help so far. I’m self-taught on a lot of this stuff so forgive me if I’m getting lost on concepts like projection.

7

You are right. For reasons that are beyond me, the projection matrix is different in forward and deferred rendering. I don’t have time to reverse engineer why and I guess it doesn’t really matter in the end since you can’t change it anyway.

For deferred pipeline, you shouldn’t multiply the result by -1, but divide it by w component to get out of the homogeneous space. After that, x and y coordinates are in range [0, 1] inside the spot angle. Near and far plane of the projection is 0 and 1, so z coordinate has no relation to range. I’m afraid you can’t get what you want without knowing spot angle and/or range.

For reference, here’s how Unity handles cookie in both forward and deferred pipeline:

// FORWARD -----------------------------------------------------------------------------
uniform sampler2D _LightTexture0;
uniform unityShadowCoord4x4 _LightMatrix0;
uniform sampler2D _LightTextureB0;
inline fixed UnitySpotCookie(unityShadowCoord4 LightCoord)
{
    return tex2D(_LightTexture0, LightCoord.xy / LightCoord.w + 0.5).w;
}
inline fixed UnitySpotAttenuate(unityShadowCoord3 LightCoord)
{
    return tex2D(_LightTextureB0, dot(LightCoord, LightCoord).xx).UNITY_ATTEN_CHANNEL;
}
#define UNITY_LIGHT_ATTENUATION(destName, input, worldPos) \
    unityShadowCoord4 lightCoord = mul(_LightMatrix0, unityShadowCoord4(worldPos, 1)); \
    fixed destName = (lightCoord.z > 0) * UnitySpotCookie(lightCoord) * UnitySpotAttenuate(lightCoord.xyz) * SHADOW_ATTENUATION(input);

// DEFERRED ------------------------------------------------------------------------
float3 tolight = _LightPos.xyz - wpos;
half3 lightDir = normalize (tolight);
       
float4 uvCookie = mul (_LightMatrix0, float4(wpos,1));
// negative bias because http://aras-p.info/blog/2010/01/07/screenspace-vs-mip-mapping/
float atten = tex2Dbias (_LightTexture0, float4(uvCookie.xy / uvCookie.w, 0, -8)).w;
atten *= uvCookie.w < 0;
float att = dot(tolight, tolight) * _LightPos.w;
atten *= tex2D (_LightTextureB0, att.rr).UNITY_ATTEN_CHANNEL;
8

Understood on the dividing by w.

I can force a global range and angle on the lights given how the project works and ensure the shader is using the same value, not ideal, but it will work. Hoping I wouldn’t have to go there, but thems the breaks.

Thanks again for the help.

9

No problem. I wish I could help you more. At least we both learned something today…