• Advertisement
Sign in to follow this  

Stable Cascaded Shadow Maps have made me lose all my hair

This topic is 1639 days old which is more than the 365 day threshold we allow for new replies. Please post a new topic.

If you intended to correct an error in the post then please contact us.

Recommended Posts

I've been working furiously and so far stable cascaded shadow maps are mocking me..... I can't seem to get this working for the life of me.....


So I will run through my algorithm and perhaps some kind soul will take pity on me and show me the error of my ways......


[1] I calculate the split frustum values for FOUR cascades.


     I use this equation:

    // get the currently active camera
    ICamera *pcCamera = getCurrentCamera ();
    vsAssert (pcCamera != NULL);

    if (m_fSplitNearZ < pcCamera->getNearZ ())
        m_fSplitNearZ = pcCamera->getNearZ ();


    for (vsInt32 iSplit = 1; iSplit < NUM_CASCADE_SPLITS; iSplit++)
        vsFloat32 fStep = iSplit / (vsFloat32) NUM_CASCADE_SPLITS;

        vsFloat32 fLogSplit = m_fSplitNearZ * powf (m_fSplitFarZ / m_fSplitNearZ, fStep);

        vsFloat32 fLinearSplit = m_fSplitNearZ + (m_fSplitFarZ - m_fSplitNearZ) * fStep;

        // linearly interpolate between the linear and logarithmic split values
        m_fSplitPositions [iSplit] = fLogSplit * m_fLogSplitWeight + fLinearSplit * (1.0f - m_fLogSplitWeight);

    // ensure the border split values are accurate
    m_fSplitPositions [0] = m_fSplitNearZ;
    m_fSplitPositions [NUM_CASCADE_SPLITS] = m_fSplitFarZ;

So now I have my frustum splits.....


I use these splits to calculate the corners of the cascade frustum that is being considered. So if I am going to update the first cascade shadow map I use the first and second frustum split positions. ie. m_fSplitPositions [0] and m_fSplitPositions [1].


So now I have 8 view space corners for the split frustum in view space. I generate the bounding box of these points and get the center. Again this is in View space. I take the max AABB point, subtract from the centroid I calculated and get it's length. This is what I use as the radius of the sphere that encloses this split frustum.


I then multiply the View Space frustum center by the inverse view to get the center in world space. I do the above for all FOUR cascade frustums when I am done I have FOUR spheres in TOTAL. One for each split sub frustum. Since I have 4 cascades I have 4 spheres.

CalculateCascadeBoundingSphere (void)
    // find the minimum enclosing sphere for the split frustum, this is done in
    // local space to avoid precision variation between frames

    ICamera *pcCurrentCamera = getCurrentCamera ();
    vsAssert (pcCurrentCamera != NULL);

    vsFloat32 fTanHalfFOVY = pcCurrentCamera->getTanHalfFOVY ();
    vsFloat32 fTanHalfFOVX = fTanHalfFOVY * pcCurrentCamera->getAspectRatio ();

    vsFloat32 fNearX = fTanHalfFOVX * m_fCurrentNearSplit;
    vsFloat32 fNearY = fTanHalfFOVY * m_fCurrentNearSplit;
    vsFloat32 fNearZ = m_fCurrentNearSplit;

     vsFloat32 fFarX = fTanHalfFOVX  * m_fCurrentFarSplit;
     vsFloat32 fFarY = fTanHalfFOVY * m_fCurrentFarSplit;
     vsFloat32 fFarZ = m_fCurrentFarSplit;

    // calculate the frustum AABB in view space
    CAABB cFrustumAABB;

    cFrustumAABB.m_cMin.setVector (FLT_MAX, FLT_MAX, FLT_MAX);
    cFrustumAABB.m_cMax.setVector (-FLT_MAX, -FLT_MAX, -FLT_MAX);

    cFrustumAABB.Add (CVector3 (fNearX, fNearY, fNearZ));
    cFrustumAABB.Add (CVector3 (fNearX, -fNearY, fNearZ));
    cFrustumAABB.Add (CVector3 (-fNearX, -fNearY, fNearZ));
    cFrustumAABB.Add (CVector3 (-fNearX, fNearY, fNearZ));

    cFrustumAABB.Add (CVector3 (fFarX, fFarY, fFarZ));
    cFrustumAABB.Add (CVector3 (fFarX, -fFarY, fFarZ));
    cFrustumAABB.Add (CVector3 (-fFarX, -fFarY, fFarZ));
    cFrustumAABB.Add (CVector3 (-fFarX, fFarY, fFarZ));

    CVector3 cCenter, cRadius;

    cCenter.Add (cFrustumAABB.m_cMin, cFrustumAABB.m_cMax);
    cCenter.MulScalar (0.5f);

    cRadius.Sub (cFrustumAABB.m_cMax, cCenter);

    vsFloat32 fRadius = cRadius.getLength ();

These spheres I use to generate FOUR light view and FOUR light view projection matrices. I do that as follows:


Remember I do this FOUR times, so at the end I have FOUR light view matrices and FOUR light projection matrices:

CalculateCascadeViewMatrix (vsUInt32 iCascade)
    vsAssert (m_pcDirectionalLightEditorObject != NULL);

    // calculate the cascade view matrix

    const CVector3 &rcLightDirection = m_pcDirectionalLightEditorObject->getDirection ();

    CVector3 cEye, cTarget;

    // target is center of view split frustum in world space
    cTarget.setVector (m_cCurrentCascadeBoundingSphere [iCascade].m_cCenter);

    // scale the light direction based on the shadow range
    CVector3 cScaledLightDirection (rcLightDirection);
    cScaledLightDirection.MulScalar (m_fShadowRange);

    // eye = target - (light_direction * shadow_range)
    cEye.Sub (cTarget, cScaledLightDirection);

    // generate the light view look at matrix
    LookAt (m_cViewMatrix [m_iCascade], cEye, cTarget);

LookAt (CMatrix4 &rcViewMatrix, const CVector3 &rcEye, const CVector3 &rcTarget)
    CVector3 cZAxis;

    // we look down -z axis this is why vector is reversed
    cZAxis.Sub (rcEye, rcTarget);
    cZAxis.Normalize ();

    CVector3 cYAxis (CVector3::Y_AXIS);

	// check it's not coincident with direction
    if (CMath::FAbs (cYAxis.DotProduct (cZAxis)) >= 1.0f)
	  // use camera up
          cYAxis.setVector (CVector3::Z_AXIS);

    // cross twice to rederive, only direction is unaltered
    CVector3 cXAxis;
    cXAxis.CrossProduct (cYAxis, cZAxis);
    cXAxis.Normalize ();

    cYAxis.CrossProduct (cZAxis, cXAxis);
    cYAxis.Normalize ();

    // setup the light view matrix (this places objects in light view space)
    rcViewMatrix.m [m00] = cXAxis.x;
    rcViewMatrix.m [m01] = cXAxis.y;
    rcViewMatrix.m [m02] = cXAxis.z;
    rcViewMatrix.m [m03] = -cXAxis.DotProduct (rcEye);

    rcViewMatrix.m [m10] = cYAxis.x;
    rcViewMatrix.m [m11] = cYAxis.y;
    rcViewMatrix.m [m12] = cYAxis.z;
    rcViewMatrix.m [m13] = -cYAxis.DotProduct (rcEye);

    rcViewMatrix.m [m20] = cZAxis.x;
    rcViewMatrix.m [m21] = cZAxis.y;
    rcViewMatrix.m [m22] = cZAxis.z;
    rcViewMatrix.m [m23] = -cZAxis.DotProduct (rcEye);

    rcViewMatrix.m [m30] = 0.0f;
    rcViewMatrix.m [m31] = 0.0f;
    rcViewMatrix.m [m32] = 0.0f;
    rcViewMatrix.m [m33] = 1.0f;

Ok now I generate FOUR orthographic projection matrices, using the radius values of the four frustum split bounding spheres I generated up above:

CalculateCascadeProjectionMatrix (vsUInt32 iCascade)
    vsFloat32 fWidth = m_cCascadeBoundingSphere [iCascade].m_fRadius * 2.0f;
    vsFloat32 fHeight = m_cCascadeBoundingSphere [iCascade].m_fRadius * 2.0f;

    // calculate the cascade orthographic projection matrix
    m_cProjectionMatrix [iCascade].m [m00] = 2.0f / fWidth;
    m_cProjectionMatrix [iCascade].m [m11] = 2.0f / fHeight;

    m_cProjectionMatrix [iCascade].m [m22] = -2.0f / (m_fProjectionFarZ - m_fProjectionNearZ);
    m_cProjectionMatrix [iCascade].m [m23] = -(m_fProjectionFarZ + m_fProjectionNearZ) / (m_fProjectionFarZ - m_fProjectionNearZ);
    m_cProjectionMatrix [iCascade].m [m33] = 1.0f;

    // engine specific projection matrix with render system depth range
    m_pcRenderSystem->ConvertProjectionMatrix (m_cProjectionMatrix [iCascade], m_cProjectionRSDepthMatrix [iCascade], true);

I also generate a rounding matrix to prevent the shadow map cascades from shimmering as the camera translates:

CalculateCascadeRoundMatrix (vsUInt32 iCascade)
    CVector3 cOriginShadow;
    CVector3 cOrigin (0.0f, 0.0f, 0.0f);

    // transform origin to light view projection space
    m_cLightViewProjectionMatrix [iCascade].Transform (cOriginShadow, cOrigin);

    vsFloat32 fShadowMapSize = getShadowMapSize ();

    // convert clip space to texture coordinates
    vsFloat32 fTexCoordX = cOriginShadow.x * fShadowMapSize * 0.5f;
    vsFloat32 fTexCoordY = cOriginShadow.y * fShadowMapSize * 0.5f;

    // round to the nearest whole texel
    vsFloat32 fTexCoordRoundedX = CMath::Round (fTexCoordX);
    vsFloat32 fTexCoordRoundedY = CMath::Round (fTexCoordY);

       the difference between the rounded and actual tex coordinate is the
       amount by which we need to translate the shadow matrix in order to
       cancel sub-texel movement
    vsFloat32 fDX = fTexCoordRoundedX - fTexCoordX;
    vsFloat32 fDY = fTexCoordRoundedY - fTexCoordY;

    // transform fDX, fDY back to homogenous light space
    fDX /= fShadowMapSize * 0.5f;
    fDY /= fShadowMapSize * 0.5f;

    // set the rounding matrix
    m_cRoundMatrix [iCascade].m [m03] = fDX;
    m_cRoundMatrix [iCascade].m [m13] = fDY;
    m_cRoundMatrix [iCascade].m [m23] = 0.0f;
    m_cRoundMatrix [iCascade].m [m33] = 1.0f;

So now I have CMatrix4x4 m_cLightViewMatrix [4],  CMatrix4x4 m_cLightProjectionMatrix [4] and m_cRoundMatrix [4];


I loop like so and generate the shadow matrix:


for (iCascade = 0; iCascade < 4; iCascade++)


       m_cShadowMatrix [iCascade] = m_cRoundMatrix [iCascade] * m_cLightProjectionMatrix [iCascade] * m_cLightViewMatrix [iCascade];



I use m_cShadowMatrix [4] to generate the shadow maps in my shadow atlas..... So In the shader I generate the vertices of the shadow caster into word space and then multiply like so:


         m_cShadowMatrix * m_cWorldMatrix * vsInput.vPosition;


In the pixel shader I save out the z value.


Is my algortihm sound? Or perhaps I missed something in my understanding of the algorithm. I would appreciate any help anyone can give me as to where I could be going wrong. Also if anyone knows of any demos with source that I could run to test this algorithm I would be ever so greatful.


Many thanks to anyone who can help me!

Share this post

Link to post
Share on other sites

I am working on figuring this out...... Right now I get a blank white screen..... So it won't be very helpful for me to post this..... :)

Share this post

Link to post
Share on other sites

It seems doubtful that misunderstanding the overall algorithm is the source of a totally blank screen; presumably, if the algorithm as you're implementing it seems to make sense to you, you'll at least get some kind of output, even if it's not correct. That said, what you're describing (with words) seems reasonable; with the code/math, it's hard to say (at least for me) whether you've overlooked something.


In fact, it seems like a bit too much code for it to be practical to find a bug just by reading the whole thing, particularly since the bug may well stem from the execution of one or more seemingly-trivial steps. Since you've already got the code broken down into smaller, more manageable chunks, have you already tried testing some of those smaller parts on some fixed input to see if you get results that are consistent with what you expect?


My prediction is that you'll find that some small part of your implementation just doesn't do what you think it does; my experience has always been that once I get done fixing problems like that, I've already gotten enough feedback that any fundamental logic errors become immediately apparent (either that, or things just work).

Share this post

Link to post
Share on other sites

I find it instructive to start at the end of the pipeline and use comment codes to shut off everything.  Comment out everything in the shaders, only have what is necessary to produce some form of output.  For instance, have your vertex shader output only what is necessary to get it working, which is only position. 

Now make sure that the fragment shader is outputting only color.  If you get this far then you know that they are both setup and activated properly.  Now start moving the comment codes and re-compile as you re-add a line or two at a time. 


Also, you can setup a screen aligned quad and use this to display the shadow texture(s). 

If this is the part where you see nothing but a white screen, then maybe try and adjust your camera's near and far settings.  If you are using a very big spread between these two then maybe you are stretching out the z-axis enough to blank out the shadows.  For example, if you are using near = 0.001, far 10000, then you might see nothing but a white screen for the depth texture.  Try setting up some keyboard controls and adjust them starting at something like, near = .1, far = 100.   Now if you actually can see the shadows, adjust them as needed.

Then again, I would suppose that some camera setups may not exhibit this issue.

Share this post

Link to post
Share on other sites



When I did this, I validated each step before moving on.  The first thing I did was draw a visualization of a frustum from an object's point of view, then I could move my camera around and look at it.  Then I split it into 4 cascades and again draw each frustum split on screen.


Once you have that, then you know that's working and can move on to the orthographic projection step (which I rendered into textures and drew those as quads on the screen to validate those).


And so on...



Share this post

Link to post
Share on other sites

One thing I have noticed in the shader is that when I multiply by the combined light view and orthographic projection matrix for a specific frustum, the generated point is NOT in the range -1.0f to 1.0f in both the x and y.


This seems to be the problem. Isn't the projection matrix supposed to get the coordinates in these ranges?   Has anyone else encountered this? Also if you do not think this is an issue won't it be an issue when the shadow map is sampled using these coordinates?


Thanks again for anyone's help

Edited by gcard28

Share this post

Link to post
Share on other sites

Well as expected, it was a small error on my part that prevented this from working. Once I followed the advice of breaking everything down piece by piece I was able to discover what I did wrong.


Thanks again for everyone's help, I now have stable cascaded shadow maps working!

Edited by gcard28

Share this post

Link to post
Share on other sites
Sign in to follow this  

  • Advertisement