Reduce code duplication between the SIMD and non-SIMD contact solve and warmstart.

8 files changed, 577 insertions, 626 deletions

diff --git a/src/dynamics/solver/delta_vel.rs b/src/dynamics/solver/delta_vel.rs
index 378d302..b50cb76 100644
--- a/src/dynamics/solver/delta_vel.rs
+++ b/src/dynamics/solver/delta_vel.rs
@@ -1,5 +1,6 @@
 use crate::math::{AngVector, Vector};
 use na::{Scalar, SimdRealField};
+use std::ops::AddAssign;
 
 #[derive(Copy, Clone, Debug)]
 //#[repr(align(64))]
@@ -16,3 +17,10 @@ impl<N: SimdRealField> DeltaVel<N> {
         }
     }
 }
+
+impl<N: SimdRealField> AddAssign for DeltaVel<N> {
+    fn add_assign(&mut self, rhs: Self) {
+        self.linear += rhs.linear;
+        self.angular += rhs.angular;
+    }
+}
diff --git a/src/dynamics/solver/mod.rs b/src/dynamics/solver/mod.rs
index 090d0f3..33b6f9a 100644
--- a/src/dynamics/solver/mod.rs
+++ b/src/dynamics/solver/mod.rs
@@ -24,9 +24,11 @@ pub(self) use position_ground_constraint::*;
 #[cfg(feature = "simd-is-enabled")]
 pub(self) use position_ground_constraint_wide::*;
 pub(self) use velocity_constraint::*;
+pub(self) use velocity_constraint_element::*;
 #[cfg(feature = "simd-is-enabled")]
 pub(self) use velocity_constraint_wide::*;
 pub(self) use velocity_ground_constraint::*;
+pub(self) use velocity_ground_constraint_element::*;
 #[cfg(feature = "simd-is-enabled")]
 pub(self) use velocity_ground_constraint_wide::*;
 
@@ -55,9 +57,11 @@ mod position_solver;
 #[cfg(not(feature = "parallel"))]
 mod solver_constraints;
 mod velocity_constraint;
+mod velocity_constraint_element;
 #[cfg(feature = "simd-is-enabled")]
 mod velocity_constraint_wide;
 mod velocity_ground_constraint;
+mod velocity_ground_constraint_element;
 #[cfg(feature = "simd-is-enabled")]
 mod velocity_ground_constraint_wide;
 #[cfg(not(feature = "parallel"))]
diff --git a/src/dynamics/solver/velocity_constraint.rs b/src/dynamics/solver/velocity_constraint.rs
index 169f280..434e6fb 100644
--- a/src/dynamics/solver/velocity_constraint.rs
+++ b/src/dynamics/solver/velocity_constraint.rs
@@ -1,13 +1,12 @@
-use super::DeltaVel;
 use crate::dynamics::solver::VelocityGroundConstraint;
 #[cfg(feature = "simd-is-enabled")]
 use crate::dynamics::solver::{WVelocityConstraint, WVelocityGroundConstraint};
 use crate::dynamics::{IntegrationParameters, RigidBodySet};
 use crate::geometry::{ContactManifold, ContactManifoldIndex};
-use crate::math::{AngVector, Real, Vector, DIM, MAX_MANIFOLD_POINTS};
+use crate::math::{Real, Vector, DIM, MAX_MANIFOLD_POINTS};
 use crate::utils::{WAngularInertia, WBasis, WCross, WDot};
-#[cfg(feature = "dim2")]
-use na::SimdPartialOrd;
+
+use super::{DeltaVel, VelocityConstraintElement, VelocityConstraintNormalPart};
 
 //#[repr(align(64))]
 #[derive(Copy, Clone, Debug)]
@@ -79,72 +78,6 @@ impl AnyVelocityConstraint {
 }
 
 #[derive(Copy, Clone, Debug)]
-pub(crate) struct VelocityConstraintTangentPart {
-    pub gcross1: [AngVector<Real>; DIM - 1],
-    pub gcross2: [AngVector<Real>; DIM - 1],
-    pub rhs: [Real; DIM - 1],
-    #[cfg(feature = "dim2")]
-    pub impulse: [Real; DIM - 1],
-    #[cfg(feature = "dim3")]
-    pub impulse: na::Vector2<Real>,
-    pub r: [Real; DIM - 1],
-}
-
-#[cfg(not(target_arch = "wasm32"))]
-impl VelocityConstraintTangentPart {
-    fn zero() -> Self {
-        Self {
-            gcross1: [na::zero(); DIM - 1],
-            gcross2: [na::zero(); DIM - 1],
-            rhs: [0.0; DIM - 1],
-            #[cfg(feature = "dim2")]
-            impulse: [0.0; DIM - 1],
-            #[cfg(feature = "dim3")]
-            impulse: na::zero(),
-            r: [0.0; DIM - 1],
-        }
-    }
-}
-
-#[derive(Copy, Clone, Debug)]
-pub(crate) struct VelocityConstraintNormalPart {
-    pub gcross1: AngVector<Real>,
-    pub gcross2: AngVector<Real>,
-    pub rhs: Real,
-    pub impulse: Real,
-    pub r: Real,
-}
-
-#[cfg(not(target_arch = "wasm32"))]
-impl VelocityConstraintNormalPart {
-    fn zero() -> Self {
-        Self {
-            gcross1: na::zero(),
-            gcross2: na::zero(),
-            rhs: 0.0,
-            impulse: 0.0,
-            r: 0.0,
-        }
-    }
-}
-
-#[derive(Copy, Clone, Debug)]
-pub(crate) struct VelocityConstraintElement {
-    pub normal_part: VelocityConstraintNormalPart,
-    pub tangent_part: VelocityConstraintTangentPart,
-}
-
-#[cfg(not(target_arch = "wasm32"))]
-impl VelocityConstraintElement {
-    pub fn zero() -> Self {
-        Self {
-            normal_part: VelocityConstraintNormalPart::zero(),
-            tangent_part: VelocityConstraintTangentPart::zero(),
-        }
-    }
-}
-
-#[derive(Copy, Clone, Debug)]
 pub(crate) struct VelocityConstraint {
     pub dir1: Vector<Real>, // Non-penetration force direction for the first body.
     #[cfg(feature = "dim3")]
@@ -159,7 +92,7 @@ pub(crate) struct VelocityConstraint {
     pub manifold_id: ContactManifoldIndex,
     pub manifold_contact_id: [u8; MAX_MANIFOLD_POINTS],
     pub num_contacts: u8,
-    pub elements: [VelocityConstraintElement; MAX_MANIFOLD_POINTS],
+    pub elements: [VelocityConstraintElement<Real>; MAX_MANIFOLD_POINTS],
 }
 
 impl VelocityConstraint {
@@ -352,116 +285,36 @@ impl VelocityConstraint {
         let mut mj_lambda1 = DeltaVel::zero();
         let mut mj_lambda2 = DeltaVel::zero();
 
-        #[cfg(feature = "dim3")]
-        let tangents1 = [self.tangent1, self.dir1.cross(&self.tangent1)];
-        #[cfg(feature = "dim2")]
-        let tangents1 = self.dir1.orthonormal_basis();
-
-        for i in 0..self.num_contacts as usize {
-            let elt = &self.elements[i].normal_part;
-            mj_lambda1.linear += self.dir1 * (self.im1 * elt.impulse);
-            mj_lambda1.angular += elt.gcross1 * elt.impulse;
-
-            mj_lambda2.linear += self.dir1 * (-self.im2 * elt.impulse);
-            mj_lambda2.angular += elt.gcross2 * elt.impulse;
-
-            for j in 0..DIM - 1 {
-                let elt = &self.elements[i].tangent_part;
-                mj_lambda1.linear += tangents1[j] * (self.im1 * elt.impulse[j]);
-                mj_lambda1.angular += elt.gcross1[j] * elt.impulse[j];
-
-                mj_lambda2.linear += tangents1[j] * (-self.im2 * elt.impulse[j]);
-                mj_lambda2.angular += elt.gcross2[j] * elt.impulse[j];
-            }
-        }
-
-        mj_lambdas[self.mj_lambda1 as usize].linear += mj_lambda1.linear;
-        mj_lambdas[self.mj_lambda1 as usize].angular += mj_lambda1.angular;
-        mj_lambdas[self.mj_lambda2 as usize].linear += mj_lambda2.linear;
-        mj_lambdas[self.mj_lambda2 as usize].angular += mj_lambda2.angular;
+        VelocityConstraintElement::warmstart_group(
+            &self.elements,
+            &self.dir1,
+            #[cfg(feature = "dim3")]
+            &self.tangent1,
+            self.im1,
+            self.im2,
+            &mut mj_lambda1,
+            &mut mj_lambda2,
+        );
+
+        mj_lambdas[self.mj_lambda1 as usize] += mj_lambda1;
+        mj_lambdas[self.mj_lambda2 as usize] += mj_lambda2;
     }
 
     pub fn solve(&mut self, mj_lambdas: &mut [DeltaVel<Real>]) {
         let mut mj_lambda1 = mj_lambdas[self.mj_lambda1 as usize];
         let mut mj_lambda2 = mj_lambdas[self.mj_lambda2 as usize];
 
-        // Solve friction.
-        #[cfg(feature = "dim3")]
-        let bitangent1 = self.dir1.cross(&self.tangent1);
-        #[cfg(feature = "dim2")]
-        let tangents1 = self.dir1.orthonormal_basis();
-
-        #[cfg(feature = "dim2")]
-        for i in 0..self.num_contacts as usize {
-            let normal_elt = &self.elements[i].normal_part;
-            let elt = &mut self.elements[i].tangent_part;
-            let dimpulse = tangents1[0].dot(&mj_lambda1.linear)
-                + elt.gcross1[0].gdot(mj_lambda1.angular)
-                - tangents1[0].dot(&mj_lambda2.linear)
-                + elt.gcross2[0].gdot(mj_lambda2.angular)
-                + elt.rhs[0];
-            let limit = self.limit * normal_elt.impulse;
-            let new_impulse = (elt.impulse[0] - elt.r[0] * dimpulse).simd_clamp(-limit, limit);
-            let dlambda = new_impulse - elt.impulse[0];
-            elt.impulse[0] = new_impulse;
-
-            mj_lambda1.linear += tangents1[0] * (self.im1 * dlambda);
-            mj_lambda1.angular += elt.gcross1[0] * dlambda;
-
-            mj_lambda2.linear += tangents1[0] * (-self.im2 * dlambda);
-            mj_lambda2.angular += elt.gcross2[0] * dlambda;
-        }
-
-        #[cfg(feature = "dim3")]
-        for i in 0..self.num_contacts as usize {
-            let limit = self.limit * self.elements[i].normal_part.impulse;
-            let elt = &mut self.elements[i].tangent_part;
-
-            let dimpulse_0 = self.tangent1.dot(&mj_lambda1.linear)
-                + elt.gcross1[0].gdot(mj_lambda1.angular)
-                - self.tangent1.dot(&mj_lambda2.linear)
-                + elt.gcross2[0].gdot(mj_lambda2.angular)
-                + elt.rhs[0];
-            let dimpulse_1 = bitangent1.dot(&mj_lambda1.linear)
-                + elt.gcross1[1].gdot(mj_lambda1.angular)
-                - bitangent1.dot(&mj_lambda2.linear)
-                + elt.gcross2[1].gdot(mj_lambda2.angular)
-                + elt.rhs[1];
-
-            let new_impulse = na::Vector2::new(
-                elt.impulse[0] - elt.r[0] * dimpulse_0,
-                elt.impulse[1] - elt.r[1] * dimpulse_1,
-            );
-            let new_impulse = new_impulse.cap_magnitude(limit);
-            let dlambda = new_impulse - elt.impulse;
-            elt.impulse = new_impulse;
-
-            mj_lambda1.linear +=
-                self.tangent1 * (self.im1 * dlambda[0]) + bitangent1 * (self.im1 * dlambda[1]);
-            mj_lambda1.angular += elt.gcross1[0] * dlambda[0] + elt.gcross1[1] * dlambda[1];
-
-            mj_lambda2.linear +=
-                self.tangent1 * (-self.im2 * dlambda[0]) + bitangent1 * (-self.im2 * dlambda[1]);
-            mj_lambda2.angular += elt.gcross2[0] * dlambda[0] + elt.gcross2[1] * dlambda[1];
-        }
-
-        // Solve non-penetration.
-        for i in 0..self.num_contacts as usize {
-            let elt = &mut self.elements[i].normal_part;
-            let dimpulse = self.dir1.dot(&mj_lambda1.linear) + elt.gcross1.gdot(mj_lambda1.angular)
-                - self.dir1.dot(&mj_lambda2.linear)
-                + elt.gcross2.gdot(mj_lambda2.angular)
-                + elt.rhs;
-            let new_impulse = (elt.impulse - elt.r * dimpulse).max(0.0);
-            let dlambda = new_impulse - elt.impulse;
-            elt.impulse = new_impulse;
-
-            mj_lambda1.linear += self.dir1 * (self.im1 * dlambda);
-            mj_lambda1.angular += elt.gcross1 * dlambda;
-
-            mj_lambda2.linear += self.dir1 * (-self.im2 * dlambda);
-            mj_lambda2.angular += elt.gcross2 * dlambda;
-        }
+        VelocityConstraintElement::solve_group(
+            &mut self.elements,
+            &self.dir1,
+            #[cfg(feature = "dim3")]
+            &self.tangent1,
+            self.im1,
+            self.im2,
+            self.limit,
+            &mut mj_lambda1,
+            &mut mj_lambda2,
+        );
 
         mj_lambdas[self.mj_lambda1 as usize] = mj_lambda1;
         mj_lambdas[self.mj_lambda2 as usize] = mj_lambda2;
diff --git a/src/dynamics/solver/velocity_constraint_element.rs b/src/dynamics/solver/velocity_constraint_element.rs
new file mode 100644
index 0000000..e80d3dd
--- /dev/null
+++ b/src/dynamics/solver/velocity_constraint_element.rs
@@ -0,0 +1,257 @@
+use super::DeltaVel;
+use crate::math::{AngVector, Vector, DIM};
+use crate::utils::{WBasis, WDot};
+use na::SimdRealField;
+
+#[derive(Copy, Clone, Debug)]
+pub(crate) struct VelocityConstraintTangentPart<N: SimdRealField> {
+    pub gcross1: [AngVector<N>; DIM - 1],
+    pub gcross2: [AngVector<N>; DIM - 1],
+    pub rhs: [N; DIM - 1],
+    #[cfg(feature = "dim2")]
+    pub impulse: [N; DIM - 1],
+    #[cfg(feature = "dim3")]
+    pub impulse: na::Vector2<N>,
+    pub r: [N; DIM - 1],
+}
+
+impl<N: SimdRealField> VelocityConstraintTangentPart<N> {
+    #[cfg(not(target_arch = "wasm32"))]
+    fn zero() -> Self {
+        Self {
+            gcross1: [na::zero(); DIM - 1],
+            gcross2: [na::zero(); DIM - 1],
+            rhs: [na::zero(); DIM - 1],
+            #[cfg(feature = "dim2")]
+            impulse: [na::zero(); DIM - 1],
+            #[cfg(feature = "dim3")]
+            impulse: na::zero(),
+            r: [na::zero(); DIM - 1],
+        }
+    }
+
+    pub fn warmstart(
+        &self,
+        tangents1: [&Vector<N>; DIM - 1],
+        im1: N,
+        im2: N,
+        mj_lambda1: &mut DeltaVel<N>,
+        mj_lambda2: &mut DeltaVel<N>,
+    ) where
+        AngVector<N>: WDot<AngVector<N>, Result = N>,
+        N::Element: SimdRealField,
+    {
+        for j in 0..DIM - 1 {
+            mj_lambda1.linear += tangents1[j] * (im1 * self.impulse[j]);
+            mj_lambda1.angular += self.gcross1[j] * self.impulse[j];
+
+            mj_lambda2.linear += tangents1[j] * (-im2 * self.impulse[j]);
+            mj_lambda2.angular += self.gcross2[j] * self.impulse[j];
+        }
+    }
+
+    pub fn solve(
+        &mut self,
+        tangents1: [&Vector<N>; DIM - 1],
+        im1: N,
+        im2: N,
+        limit: N,
+        mj_lambda1: &mut DeltaVel<N>,
+        mj_lambda2: &mut DeltaVel<N>,
+    ) where
+        AngVector<N>: WDot<AngVector<N>, Result = N>,
+        N::Element: SimdRealField,
+    {
+        #[cfg(feature = "dim2")]
+        {
+            let dimpulse = tangents1[0].dot(&mj_lambda1.linear)
+                + self.gcross1[0].gdot(mj_lambda1.angular)
+                - tangents1[0].dot(&mj_lambda2.linear)
+                + self.gcross2[0].gdot(mj_lambda2.angular)
+                + self.rhs[0];
+            let new_impulse = (self.impulse[0] - self.r[0] * dimpulse).simd_clamp(-limit, limit);
+            let dlambda = new_impulse - self.impulse[0];
+            self.impulse[0] = new_impulse;
+
+            mj_lambda1.linear += tangents1[0] * (im1 * dlambda);
+            mj_lambda1.angular += self.gcross1[0] * dlambda;
+
+            mj_lambda2.linear += tangents1[0] * (-im2 * dlambda);
+            mj_lambda2.angular += self.gcross2[0] * dlambda;
+        }
+
+        #[cfg(feature = "dim3")]
+        {
+            let dimpulse_0 = tangents1[0].dot(&mj_lambda1.linear)
+                + self.gcross1[0].gdot(mj_lambda1.angular)
+                - tangents1[0].dot(&mj_lambda2.linear)
+                + self.gcross2[0].gdot(mj_lambda2.angular)
+                + self.rhs[0];
+            let dimpulse_1 = tangents1[1].dot(&mj_lambda1.linear)
+                + self.gcross1[1].gdot(mj_lambda1.angular)
+                - tangents1[1].dot(&mj_lambda2.linear)
+                + self.gcross2[1].gdot(mj_lambda2.angular)
+                + self.rhs[1];
+
+            let new_impulse = na::Vector2::new(
+                self.impulse[0] - self.r[0] * dimpulse_0,
+                self.impulse[1] - self.r[1] * dimpulse_1,
+            );
+            let new_impulse = new_impulse.simd_cap_magnitude(limit);
+            let dlambda = new_impulse - self.impulse;
+            self.impulse = new_impulse;
+
+            mj_lambda1.linear +=
+                tangents1[0] * (im1 * dlambda[0]) + tangents1[1] * (im1 * dlambda[1]);
+            mj_lambda1.angular += self.gcross1[0] * dlambda[0] + self.gcross1[1] * dlambda[1];
+
+            mj_lambda2.linear +=
+                tangents1[0] * (-im2 * dlambda[0]) + tangents1[1] * (-im2 * dlambda[1]);
+            mj_lambda2.angular += self.gcross2[0] * dlambda[0] + self.gcross2[1] * dlambda[1];
+        }
+    }
+}
+
+#[derive(Copy, Clone, Debug)]
+pub(crate) struct VelocityConstraintNormalPart<N: SimdRealField> {
+    pub gcross1: AngVector<N>,
+    pub gcross2: AngVector<N>,
+    pub rhs: N,
+    pub impulse: N,
+    pub r: N,
+}
+
+impl<N: SimdRealField> VelocityConstraintNormalPart<N> {
+    #[cfg(not(target_arch = "wasm32"))]
+    fn zero() -> Self {
+        Self {
+            gcross1: na::zero(),
+            gcross2: na::zero(),
+            rhs: na::zero(),
+            impulse: na::zero(),
+            r: na::zero(),
+        }
+    }
+
+    #[inline]
+    pub fn warmstart(
+        &self,
+        dir1: &Vector<N>,
+        im1: N,
+        im2: N,
+        mj_lambda1: &mut DeltaVel<N>,
+        mj_lambda2: &mut DeltaVel<N>,
+    ) where
+        AngVector<N>: WDot<AngVector<N>, Result = N>,
+    {
+        mj_lambda1.linear += dir1 * (im1 * self.impulse);
+        mj_lambda1.angular += self.gcross1 * self.impulse;
+
+        mj_lambda2.linear += dir1 * (-im2 * self.impulse);
+        mj_lambda2.angular += self.gcross2 * self.impulse;
+    }
+
+    #[inline]
+    pub fn solve(
+        &mut self,
+        dir1: &Vector<N>,
+        im1: N,
+        im2: N,
+        mj_lambda1: &mut DeltaVel<N>,
+        mj_lambda2: &mut DeltaVel<N>,
+    ) where
+        AngVector<N>: WDot<AngVector<N>, Result = N>,
+    {
+        let dimpulse = dir1.dot(&mj_lambda1.linear) + self.gcross1.gdot(mj_lambda1.angular)
+            - dir1.dot(&mj_lambda2.linear)
+            + self.gcross2.gdot(mj_lambda2.angular)
+            + self.rhs;
+        let new_impulse = (self.impulse - self.r * dimpulse).simd_max(N::zero());
+        let dlambda = new_impulse - self.impulse;
+        self.impulse = new_impulse;
+
+        mj_lambda1.linear += dir1 * (im1 * dlambda);
+        mj_lambda1.angular += self.gcross1 * dlambda;
+
+        mj_lambda2.linear += dir1 * (-im2 * dlambda);
+        mj_lambda2.angular += self.gcross2 * dlambda;
+    }
+}
+
+#[derive(Copy, Clone, Debug)]
+pub(crate) struct VelocityConstraintElement<N: SimdRealField> {
+    pub normal_part: VelocityConstraintNormalPart<N>,
+    pub tangent_part: VelocityConstraintTangentPart<N>,
+}
+
+impl<N: SimdRealField> VelocityConstraintElement<N> {
+    #[cfg(not(target_arch = "wasm32"))]
+    pub fn zero() -> Self {
+        Self {
+            normal_part: VelocityConstraintNormalPart::zero(),
+            tangent_part: VelocityConstraintTangentPart::zero(),
+        }
+    }
+
+    pub fn warmstart_group(
+        elements: &[Self],
+        dir1: &Vector<N>,
+        #[cfg(feature = "dim3")] tangent1: &Vector<N>,
+        im1: N,
+        im2: N,
+        mj_lambda1: &mut DeltaVel<N>,
+        mj_lambda2: &mut DeltaVel<N>,
+    ) where
+        Vector<N>: WBasis,
+        AngVector<N>: WDot<AngVector<N>, Result = N>,
+        N::Element: SimdRealField,
+    {
+        #[cfg(feature = "dim3")]
+        let tangents1 = [tangent1, &dir1.cross(&tangent1)];
+        #[cfg(feature = "dim2")]
+        let tangents1 = [&dir1.orthonormal_vector()];
+
+        for element in elements {
+            element
+                .tangent_part
+                .warmstart(tangents1, im1, im2, mj_lambda1, mj_lambda2);
+            element
+                .normal_part
+                .warmstart(dir1, im1, im2, mj_lambda1, mj_lambda2);
+        }
+    }
+
+    pub fn solve_group(
+        elements: &mut [Self],
+        dir1: &Vector<N>,
+        #[cfg(feature = "dim3")] tangent1: &Vector<N>,
+        im1: N,
+        im2: N,
+        limit: N,
+        mj_lambda1: &mut DeltaVel<N>,
+        mj_lambda2: &mut DeltaVel<N>,
+    ) where
+        Vector<N>: WBasis,
+        AngVector<N>: WDot<AngVector<N>, Result = N>,
+        N::Element: SimdRealField,
+    {
+        // Solve friction.
+        #[cfg(feature = "dim3")]
+        let tangents1 = [tangent1, &dir1.cross(&tangent1)];
+        #[cfg(feature = "dim2")]
+        let tangents1 = [&dir1.orthonormal_vector()];
+
+        for element in elements.iter_mut() {
+            let limit = limit * element.normal_part.impulse;
+            let part = &mut element.tangent_part;
+            part.solve(tangents1, im1, im2, limit, mj_lambda1, mj_lambda2);
+        }
+
+        // Solve penetration.
+        for element in elements.iter_mut() {
+            element
+                .normal_part
+                .solve(&dir1, im1, im2, mj_lambda1, mj_lambda2);
+        }
+    }
+}
diff --git a/src/dynamics/solver/velocity_constraint_wide.rs b/src/dynamics/solver/velocity_constraint_wide.rs
index ec24e56..691a983 100644
--- a/src/dynamics/solver/velocity_constraint_wide.rs
+++ b/src/dynamics/solver/velocity_constraint_wide.rs
@@ -1,4 +1,6 @@
-use super::{AnyVelocityConstraint, DeltaVel};
+use super::{
+    AnyVelocityConstraint, DeltaVel, VelocityConstraintElement, VelocityConstraintNormalPart,
+};
 use crate::dynamics::{IntegrationParameters, RigidBodySet};
 use crate::geometry::{ContactManifold, ContactManifoldIndex};
 use crate::math::{
@@ -11,76 +13,13 @@ use num::Zero;
 use simba::simd::{SimdPartialOrd, SimdValue};
 
 #[derive(Copy, Clone, Debug)]
-pub(crate) struct WVelocityConstraintTangentPart {
-    pub gcross1: [AngVector<SimdReal>; DIM - 1],
-    pub gcross2: [AngVector<SimdReal>; DIM - 1],
-    pub rhs: [SimdReal; DIM - 1],
-    #[cfg(feature = "dim2")]
-    pub impulse: [SimdReal; DIM - 1],
-    #[cfg(feature = "dim3")]
-    pub impulse: na::Vector2<SimdReal>,
-    pub r: [SimdReal; DIM - 1],
-}
-
-impl WVelocityConstraintTangentPart {
-    pub fn zero() -> Self {
-        Self {
-            gcross1: [AngVector::zero(); DIM - 1],
-            gcross2: [AngVector::zero(); DIM - 1],
-            rhs: [SimdReal::zero(); DIM - 1],
-            #[cfg(feature = "dim2")]
-            impulse: [SimdReal::zero(); DIM - 1],
-            #[cfg(feature = "dim3")]
-            impulse: na::Vector2::zeros(),
-            r: [SimdReal::zero(); DIM - 1],
-        }
-    }
-}
-
-#[derive(Copy, Clone, Debug)]
-pub(crate) struct WVelocityConstraintNormalPart {
-    pub gcross1: AngVector<SimdReal>,
-    pub gcross2: AngVector<SimdReal>,
-    pub rhs: SimdReal,
-    pub impulse: SimdReal,
-    pub r: SimdReal,
-}
-
-impl WVelocityConstraintNormalPart {
-    pub fn zero() -> Self {
-        Self {
-            gcross1: AngVector::zero(),
-            gcross2: AngVector::zero(),
-            rhs: SimdReal::zero(),
-            impulse: SimdReal::zero(),
-            r: SimdReal::zero(),
-        }
-    }
-}
-
-#[derive(Copy, Clone, Debug)]
-pub(crate) struct WVelocityConstraintElement {
-    pub normal_part: WVelocityConstraintNormalPart,
-    pub tangent_part: WVelocityConstraintTangentPart,
-}
-
-impl WVelocityConstraintElement {
-    pub fn zero() -> Self {
-        Self {
-            normal_part: WVelocityConstraintNormalPart::zero(),
-            tangent_part: WVelocityConstraintTangentPart::zero(),
-        }
-    }
-}
-
-#[derive(Copy, Clone, Debug)]
 pub(crate) struct WVelocityConstraint {
     pub dir1: Vector<SimdReal>, // Non-penetration force direction for the first body.
     #[cfg(feature = "dim3")]
     pub tangent1: Vector<SimdReal>, // One of the friction force directions.
     #[cfg(feature = "dim3")]
     pub tangent_rot1: na::UnitComplex<SimdReal>, // Orientation of the tangent basis wrt. the reference basis.
-    pub elements: [WVelocityConstraintElement; MAX_MANIFOLD_POINTS],
+    pub elements: [VelocityConstraintElement<SimdReal>; MAX_MANIFOLD_POINTS],
     pub num_contacts: u8,
     pub im1: SimdReal,
     pub im2: SimdReal,
@@ -159,7 +98,7 @@ impl WVelocityConstraint {
                 tangent1: tangents1[0],
                 #[cfg(feature = "dim3")]
                 tangent_rot1,
-                elements: [WVelocityConstraintElement::zero(); MAX_MANIFOLD_POINTS],
+                elements: [VelocityConstraintElement::zero(); MAX_MANIFOLD_POINTS],
                 im1,
                 im2,
                 limit: SimdReal::splat(0.0),
@@ -212,7 +151,7 @@ impl WVelocityConstraint {
                     rhs +=
                         dist.simd_min(SimdReal::zero()) * (velocity_based_erp_inv_dt * is_resting);
 
-                    constraint.elements[k].normal_part = WVelocityConstraintNormalPart {
+                    constraint.elements[k].normal_part = VelocityConstraintNormalPart {
                         gcross1,
                         gcross2,
                         rhs,
@@ -277,28 +216,16 @@ impl WVelocityConstraint {
             ),
         };
 
-        #[cfg(feature = "dim3")]
-        let tangents1 = [self.tangent1, self.dir1.cross(&self.tangent1)];
-        #[cfg(feature = "dim2")]
-        let tangents1 = self.dir1.orthonormal_basis();
-
-        for i in 0..self.num_contacts as usize {
-            let elt = &self.elements[i].normal_part;
-            mj_lambda1.linear += self.dir1 * (self.im1 * elt.impulse);
-            mj_lambda1.angular += elt.gcross1 * elt.impulse;
-
-            mj_lambda2.linear += self.dir1 * (-self.im2 * elt.impulse);
-            mj_lambda2.angular += elt.gcross2 * elt.impulse;
-
-            for j in 0..DIM - 1 {
-                let elt = &self.elements[i].tangent_part;
-                mj_lambda1.linear += tangents1[j] * (self.im1 * elt.impulse[j]);
-                mj_lambda1.angular += elt.gcross1[j] * elt.impulse[j];
-
-                mj_lambda2.linear += tangents1[j] * (-self.im2 * elt.impulse[j]);
-                mj_lambda2.angular += elt.gcross2[j] * elt.impulse[j];
-            }
-        }
+        VelocityConstraintElement::warmstart_group(
+            &self.elements,
+            &self.dir1,
+            #[cfg(feature = "dim3")]
+            &self.tangent1,
+            self.im1,
+            self.im2,
+            &mut mj_lambda1,
+            &mut mj_lambda2,
+        );
 
         for ii in 0..SIMD_WIDTH {
             mj_lambdas[self.mj_lambda1[ii] as usize].linear = mj_lambda1.linear.extract(ii);
@@ -329,82 +256,17 @@ impl WVelocityConstraint {
             ),
         };
 
-        // Solve friction.
-        #[cfg(feature = "dim3")]
-        let bitangent1 = self.dir1.cross(&self.tangent1);
-        #[cfg(feature = "dim2")]
-        let tangents1 = self.dir1.orthonormal_basis();
-
-        #[cfg(feature = "dim2")]
-        for i in 0..self.num_contacts as usize {
-            let normal_elt = &self.elements[i].normal_part;
-
-            let elt = &mut self.elements[i].tangent_part;
-            let dimpulse = tangents1[0].dot(&mj_lambda1.linear)
-                + elt.gcross1[0].gdot(mj_lambda1.angular)
-                - tangents1[0].dot(&mj_lambda2.linear)
-                + elt.gcross2[0].gdot(mj_lambda2.angular)
-                + elt.rhs[0];
-            let limit = self.limit * normal_elt.impulse;
-            let new_impulse = (elt.impulse[0] - elt.r[0] * dimpulse).simd_clamp(-limit, limit);
-            let dlambda = new_impulse - elt.impulse[0];
-            elt.impulse[0] = new_impulse;
-
-            mj_lambda1.linear += tangents1[0] * (self.im1 * dlambda);
-            mj_lambda1.angular += elt.gcross1[0] * dlambda;
-            mj_lambda2.linear += tangents1[0] * (-self.im2 * dlambda);
-            mj_lambda2.angular += elt.gcross2[0] * dlambda;
-        }
-
-        #[cfg(feature = "dim3")]
-        for i in 0..self.num_contacts as usize {
-            let limit = self.limit * self.elements[i].normal_part.impulse;
-            let elts = &mut self.elements[i].tangent_part;
-
-            let dimpulse_0 = self.tangent1.dot(&mj_lambda1.linear)
-                + elts.gcross1[0].gdot(mj_lambda1.angular)
-                - self.tangent1.dot(&mj_lambda2.linear)
-                + elts.gcross2[0].gdot(mj_lambda2.angular)
-                + elts.rhs[0];
-            let dimpulse_1 = bitangent1.dot(&mj_lambda1.linear)
-                + elts.gcross1[1].gdot(mj_lambda1.angular)
-                - bitangent1.dot(&mj_lambda2.linear)
-                + elts.gcross2[1].gdot(mj_lambda2.angular)
-                + elts.rhs[1];
-
-            let new_impulse = na::Vector2::new(
-                elts.impulse[0] - elts.r[0] * dimpulse_0,
-                elts.impulse[1] - elts.r[1] * dimpulse_1,
-            );
-            let new_impulse = new_impulse.simd_cap_magnitude(limit);
-            let dlambda = new_impulse - elts.impulse;
-            elts.impulse = new_impulse;
-
-            mj_lambda1.linear +=
-                self.tangent1 * (self.im1 * dlambda[0]) + bitangent1 * (self.im1 * dlambda[1]);
-            mj_lambda1.angular += elts.gcross1[0] * dlambda[0] + elts.gcross1[1] * dlambda[1];
-
-            mj_lambda2.linear +=
-                self.tangent1 * (-self.im2 * dlambda[0]) + bitangent1 * (-self.im2 * dlambda[1]);
-            mj_lambda2.angular += elts.gcross2[0] * dlambda[0] + elts.gcross2[1] * dlambda[1];
-        }
-
-        // Solve non-penetration after friction.
-        for i in 0..self.num_contacts as usize {
-            let elt = &mut self.elements[i].normal_part;
-            let dimpulse = self.dir1.dot(&mj_lambda1.linear) + elt.gcross1.gdot(mj_lambda1.angular)
-                - self.dir1.dot(&mj_lambda2.linear)
-                + elt.gcross2.gdot(mj_lambda2.angular)
-                + elt.rhs;
-            let new_impulse = (elt.impulse - elt.r * dimpulse).simd_max(SimdReal::zero());
-            let dlambda = new_impulse - elt.impulse;
-            elt.impulse = new_impulse;
-
-            mj_lambda1.linear += self.dir1 * (self.im1 * dlambda);
-            mj_lambda1.angular += elt.gcross1 * dlambda;
-            mj_lambda2.linear += self.dir1 * (-self.im2 * dlambda);
-            mj_lambda2.angular += elt.gcross2 * dlambda;
-        }
+        VelocityConstraintElement::solve_group(
+            &mut self.elements,
+            &self.dir1,
+            #[cfg(feature = "dim3")]
+            &self.tangent1,
+            self.im1,
+            self.im2,
+            self.limit,
+            &mut mj_lambda1,
+            &mut mj_lambda2,
+        );
 
         for ii in 0..SIMD_WIDTH {
             mj_lambdas[self.mj_lambda1[ii] as usize].linear = mj_lambda1.linear.extract(ii);
diff --git a/src/dynamics/solver/velocity_ground_constraint.rs b/src/dynamics/solver/velocity_ground_constraint.rs
index 78afd39..3509dd5 100644
--- a/src/dynamics/solver/velocity_ground_constraint.rs
+++ b/src/dynamics/solver/velocity_ground_constraint.rs
@@ -1,90 +1,30 @@
-use super::{AnyVelocityConstraint, DeltaVel};
-use crate::math::{AngVector, Real, Vector, DIM, MAX_MANIFOLD_POINTS};
+use super::{
+    AnyVelocityConstraint, DeltaVel, VelocityGroundConstraintElement,
+    VelocityGroundConstraintNormalPart,
+};
+use crate::math::{Real, Vector, DIM, MAX_MANIFOLD_POINTS};
 #[cfg(feature = "dim2")]
 use crate::utils::WBasis;
 use crate::utils::{WAngularInertia, WCross, WDot};
 
 use crate::dynamics::{IntegrationParameters, RigidBodySet};
 use crate::geometry::{ContactManifold, ContactManifoldIndex};
-#[cfg(feature = "dim2")]
-use na::SimdPartialOrd;
-
-#[derive(Copy, Clone, Debug)]
-pub(crate) struct VelocityGroundConstraintTangentPart {
-    pub gcross2: [AngVector<Real>; DIM - 1],
-    pub rhs: [Real; DIM - 1],
-    #[cfg(feature = "dim2")]
-    pub impulse: [Real; DIM - 1],
-    #[cfg(feature = "dim3")]
-    pub impulse: na::Vector2<Real>,
-    pub r: [Real; DIM - 1],
-}
-
-#[cfg(not(target_arch = "wasm32"))]
-impl VelocityGroundConstraintTangentPart {
-    fn zero() -> Self {
-        Self {
-            gcross2: [na::zero(); DIM - 1],
-            rhs: [0.0; DIM - 1],
-            #[cfg(feature = "dim2")]
-            impulse: [0.0; DIM - 1],
-            #[cfg(feature = "dim3")]
-            impulse: na::zero(),
-            r: [0.0; DIM - 1],
-        }
-    }
-}
-
-#[derive(Copy, Clone, Debug)]
-pub(crate) struct VelocityGroundConstraintNormalPart {
-    pub gcross2: AngVector<Real>,
-    pub rhs: Real,
-    pub impulse: Real,
-    pub r: Real,
-}
-
-#[cfg(not(target_arch = "wasm32"))]
-impl VelocityGroundConstraintNormalPart {
-    fn zero() -> Self {
-        Self {
-            gcross2: na::zero(),
-            rhs: 0.0,
-            impulse: 0.0,
-            r: 0.0,
-        }
-    }
-}
-
-#[derive(Copy, Clone, Debug)]
-pub(crate) struct VelocityGroundConstraintElement {
-    pub normal_part: VelocityGroundConstraintNormalPart,
-    pub tangent_part: VelocityGroundConstraintTangentPart,
-}
-
-#[cfg(not(target_arch = "wasm32"))]
-impl VelocityGroundConstraintElement {
-    pub fn zero() -> Self {
-        Self {
-            normal_part: VelocityGroundConstraintNormalPart::zero(),
-            tangent_part: VelocityGroundConstraintTangentPart::zero(),
-        }
-    }
-}
 
 #[derive(Copy, Clone, Debug)]
 pub(crate) struct VelocityGroundConstraint {
+    pub mj_lambda2: usize,
     pub dir1: Vector<Real>, // Non-penetration force direction for the first body.
     #[cfg(feature = "dim3")]
     pub tangent1: Vector<Real>, // One of the friction force directions.
-    #[cfg(feature = "dim3")]
-    pub tangent_rot1: na::UnitComplex<Real>, // Orientation of the tangent basis wrt. the reference basis.
     pub im2: Real,
     pub limit: Real,
-    pub mj_lambda2: usize,
+    pub elements: [VelocityGroundConstraintElement<Real>; MAX_MANIFOLD_POINTS],
+
+    #[cfg(feature = "dim3")]
+    pub tangent_rot1: na::UnitComplex<Real>, // Orientation of the tangent basis wrt. the reference basis.
     pub manifold_id: ContactManifoldIndex,
     pub manifold_contact_id: [u8; MAX_MANIFOLD_POINTS],
     pub num_contacts: u8,
-    pub elements: [VelocityGroundConstraintElement; MAX_MANIFOLD_POINTS],
 }
 
 impl VelocityGroundConstraint {
@@ -254,22 +194,15 @@ impl Veloc