ramakarl
commented
4 years ago Ah ha, nice work!
I was investigating this as well and came to the same conclusion. The SetTransform was not computing those correctly. Here is my updated code for comparison:
void VolumeGVDB::SetTransform(Vector3DF pretrans, Vector3DF scal, Vector3DF angs, Vector3DF trans )
{
mPretrans = pretrans;
mScale = scal;
mAngs = angs;
mTrans = trans;
// rotation only
mXrot.RotateTZYX(mAngs, Vector3DF(0, 0, 0));
mInvXrot = mXrot;
mInvXrot.InvertTRS();
// p' = T S R PT p APPLY_ORDER: pretrans -> rotate -> scale -> translate
mXform.Identity();
mXform.Translate ( mTrans );
mXform.Scale ( mScale );
mXform.Rotate ( mXrot );
mXform.Translate ( mPretrans );
// p = PT^-1 R^-1 S^-1 T^-1 p' APPLY_ORDER: inv trans -> inv scale -> inv rot -> inv pretrans
mInvXform.Identity();
mInvXform.InvTranslate ( mPretrans );
mInvXform.InvRotate ( mXrot );
mInvXform.InvScale ( mScale );
mInvXform.InvTranslate ( mTrans );
}This should be a lot cleaner/clearer. One thing that bothered me for a while was that the Matrix/Vector classes didn't have composing operators. I have added those and used them here. They apply the correct matrix multiplications to clean up the code a lot.
//------------------------------------------- Composing operations
// M' = M*T
Matrix4F& Matrix4F::Translate ( const Vector3DF& t )
{
data[12] += (VTYPE) t.x;
data[13] += (VTYPE) t.y;
data[14] += (VTYPE) t.z;
return *this;
}
// M' = M*R
Matrix4F& Matrix4F::Rotate ( const Matrix4F& rot )
{
Multiply ( rot );
return *this;
}
// M' = M*S -- actual a matrix-vector multiplication
Matrix4F& Matrix4F::Scale ( const Vector3DF& s )
{
data[0] *= (VTYPE) s.x; data[1] *= (VTYPE) s.y; data[2] *= (VTYPE) s.z;
data[4] *= (VTYPE) s.x; data[5] *= (VTYPE) s.y; data[6] *= (VTYPE) s.z;
data[8] *= (VTYPE) s.x; data[9] *= (VTYPE) s.y; data[10] *= (VTYPE) s.z;
data[12] *= (VTYPE) s.x; data[13] *= (VTYPE) s.y; data[14] *= (VTYPE) s.z;
return *this;
}
// M' = M * T^-1
Matrix4F& Matrix4F::InvTranslate ( const Vector3DF& t )
{
data[12] -= (VTYPE) t.x;
data[13] -= (VTYPE) t.y;
data[14] -= (VTYPE) t.z;
return *this;
}
Matrix4F& Matrix4F::InvRotate ( Matrix4F mtx )
{
mtx.InvertTRS();
Multiply ( mtx );
return *this;
}
Matrix4F& Matrix4F::InvScale ( const Vector3DF& s )
{
data[0] /= (VTYPE) s.x; data[1] /= (VTYPE) s.y; data[2] /= (VTYPE) s.z;
data[4] /= (VTYPE) s.x; data[5] /= (VTYPE) s.y; data[6] /= (VTYPE) s.z;
data[8] /= (VTYPE) s.x; data[9] /= (VTYPE) s.y; data[10] /= (VTYPE) s.z;
data[12] /= (VTYPE) s.x; data[13] /= (VTYPE) s.y; data[14] /= (VTYPE) s.z;
return *this;
}Regarding the OptiX rendering, what I have now is:
RT_PROGRAM void vol_intersect( int primIdx )
{
float3 hit = make_float3(NOHIT,NOHIT,NOHIT);
float3 norm = make_float3(0,0,0);
float4 hclr = make_float4(1,1,1,1);
float t;
if (ray.ray_type == MESH_RAY ) return;
//-- Ray march
float3 orig = mmult(ray.origin, SCN_INVXFORM);
float3 dir = mmult(normalize(ray.direction), SCN_INVXROT);
rayCast ( &gvdbObj, gvdbChan, orig, dir, hit, norm, hclr, raySurfaceTrilinearBrickF4 );
if ( hit.z >= NOHIT ) return;
t = length ( hit - ray.origin );
// report intersection to optix
if ( rtPotentialIntersection( t ) ) {
shading_normal = norm;
geometric_normal = shading_normal;
front_hit_point = mmult(hit, SCN_XFORM);
back_hit_point = mmult(hit, SCN_XFORM);
deep_color = hclr;
rtReportIntersection( mat_id );
}
}This works correctly now, including shadows, with OptiX.
Nice find on SetTransform! Glad we found that independently as confirmation.
digbeta
NBickford-NV
I was debugging some issues getting a volume to spin around its center of mass and noted that when rendering in GVDB, the volume was orbiting the origin instead of spinning. As a sanity check, I passed the same computed mXform to a mesh object in the scene and it spun correctly and it did not orbit.
The orbiting behavior of the volume looked like what you might see if rotation was happening after the translation (assuming in the inverse transform because the mXform appeared to be correct). As I began to look into the transform more closely, I wanted to double check the mInvXform that was being calculated in SetTransform(), so I added this to the end of the function:
And found that I was not getting an identity matrix from the result. So, I modified SetTransform as follows. Note I left the mInvXrot unchanged as there were issues I couldn't immediately diagnose, but I changed the mInvXform to calculate the inverse matrix using the RotateTZYXS function and a new tmp2 variable to make things readable below. As a result, I was now getting an identity matrix from the newly calculated mInvXform and the spinning behavior works correctly for volumes:
This seems to have broken some things in the Optix rendering that I am still investigating.
Comments and thoughts are welcomed here. Thanks!