So I'm using a fairly traditional method for depth reconstruction through a Vertex/Pixel Shader. First, here's the relevant code:


VS_FARCLIPVEC VS_RenderScreen( float4 vPos : POSITION,
float2 vTexCoord0 : TEXCOORD0)
{
VS_FARCLIPVEC OUT;
OUT.Position = vPos;
float fTexX = (vTexCoord0.x * 2 - 1);
float fTexY = ((1- vTexCoord0.y) * 2 - 1);
float4 fProjPoint = float4(fTexX,fTexY,1,1);
float4 fFarClipVector = mul(fProjPoint, g_mProjectionInverse);
fFarClipVector.xyz /= fFarClipVector.w;
OUT.vTexCoord0 = vTexCoord0;
OUT.fFarClipVector = fFarClipVector;

float4 fLightPosVS = mul(float4(g_singleLightPos,1), g_mView);

OUT.fLightPosVS = fLightPosVS.xyz / fLightPosVS.w;
return OUT;
}

float4 PS_CreateShadowBufferPlanets( VS_FARCLIPVEC IN ) : COLOR0
{
float fDepth = tex2D(Tex0Sampler, IN.vTexCoord0).r;
if(fDepth == 1.0)
discard;

float3 vPosVS = IN.fFarClipVector.xyz * fDepth;

float dist = length(vPosVS - g_vLightPosVS.xyz);

for(int i = 0; i < g_iPlanetCount; i++)
{
float3 p1 = vPosVS;
float3 p2 = g_vLightPosVS.xyz;
float3 p3 = g_vPlanetPositions.xyz;
float r = g_fPlanetRadii;

float3 d = normalize(p2 - p1);
float a = dot(d, d);
float b = 2.0 * dot(d, p1 - p3);
float c = dot(p3, p3) + dot(p1, p1) - 2.0 * dot(p3, p1) - r*r;

float discriminant = b*b - 4*a*c;
if(discriminant >= 0)
{
float smoothing = sqrt(discriminant) / (4*a);
float dist2 = -b / (4*a) - smoothing;
if(dist2 > 0 && dist2 < dist)
return smoothing / r * 7;
}
}

discard;
return 0;
}


This whole code is used to render planet shadows regardless of distance and without using shadow maps. Performance isn't ideal, but I've selected this code because it has a minimal amount of inputs. This is largely working as expected, except for one thing: the shadows that are calculated shake, badly. Whenever I rotate the camera, even if just a little bit, the shadows tend to slide a bit in one direction, then snap back to their original location. Rinse and repeat as the camera rotates. If the camera moves, the effect still appears, but it takes a very, very large amount of movement before just one "snapping" happens.

If the camera doesn't move, the shadows are pixel-accurate and don't have any visible defect. The g_vPlanetPositions and g_vLightPosVS use world-space XYZ coords which are projected into the camera's view space prior to the shader running. I do not believe that those are problematic, because I also have another shadow algorithm (this one using cascade shadow maps) that does the same thing.

Therefore, I am suspecting one of two things: either my view matrix or my projection matrix is wrong. I do not know which would be the case. In order to calculate my inverse projection matrix, I just use D3DXMatrixInverse. The initial projection matrix is a simple D3DXMatrixPerspectiveFovLH call with a field of view between 70 and 90 degrees.

The view matrices are calculated from the camera position and a rotation matrix.

It must be said that the scale of the scene is fairly large. The distance between the light and the casters can be in the 500,000 units range, same for the receivers. I am uncertain whether the entire thing is caused by floating point precision issues, because my shadow maps have a much higher accuracy than the steps I notice and the rest of the scene renders fine.

Does anyone have pointers as for what could cause this? I can paste more code as necessary, just ask.

What are your near and far clip values for the scene and the shadow map? If it is in magnitude of 500000 units, you'll be having problems, especially if the near clip is rather small.

Cheers!

The far clip values can go really high, above 400,000 units (depends on the cascade and the viewer's orientation), while the near clip is determined by scene geometry and usually is around 100,000 (since the source is always a star, there tends to be very little close by).

My problem is that while I'd understand this being an issue for shadow mapping, the algorithm I've posted doesn't actually use shadow maps. It only reconstructs a position from the depth buffer, without using any shadow map. Could it be that the far clip for the main depth buffer is too large, too? The near clip for that is set at 0.5 units as a constant, while the far clip can go to 600,000 or more, depending on scene geometry and user settings.

The position reconstruction will get pretty inaccurate too at those precisions, especially if you are reading the depth value from 24-bit z/w depth buffer.

Cheers!

We're using a 32-bit floating point buffer for the depth.

EDIT: Figured it out. We were reconstructing our view matrix from an eye position and a rotation matrix by passing the required vectors to D3DXMatrixLookAtLH (which we'd calculate from the rotation matrix and eye position). Doing the calculations manually the long way seems to have corrected the problem with no other changes required.