diff --git a/Aquaria/AnimationEditor.cpp b/Aquaria/AnimationEditor.cpp
index 1404525..514e7a5 100644
--- a/Aquaria/AnimationEditor.cpp
+++ b/Aquaria/AnimationEditor.cpp
@@ -472,7 +472,7 @@ void AnimationEditor::applyState()
 	splinegrid->setTexture("mithalas-house-bg-0001");
 	splinegrid->setWidthHeight(200, 200);
 	splinegrid->position = Vector(400, 300);
-	splinegrid->resize(3,3,10,10);
+	splinegrid->resize(3,3,10,10,3);
 	splinegrid->resetControlPoints();
 	addRenderObject(splinegrid, LR_PARTICLES_TOP);
 
@@ -647,12 +647,12 @@ void AnimationEditor::moveBoneStripPoint(const Vector &mov)
 		{
 			if (!sel->changeStrip.empty())
 			{
-				if (b->strip.size() < sel->changeStrip.size())
+				if (b->grid.size() < sel->changeStrip.size())
 				{
-					b->strip.resize(sel->changeStrip.size());
+					b->grid.resize(sel->changeStrip.size());
 				}
 
-				b->strip[selectedStripPoint] = sel->changeStrip[selectedStripPoint] += mov*0.006f;
+				b->grid[selectedStripPoint] = sel->changeStrip[selectedStripPoint] += mov*0.006f;
 			}
 		}
 	}
@@ -1266,7 +1266,7 @@ void AnimationEditor::cloneBoneAhead()
 			b2->x = b1->x;
 			b2->y = b1->y;
 			b2->rot = b1->rot;
-			b2->strip = b1->strip;
+			b2->grid = b1->grid;
 		}
 	}
 }
diff --git a/BBGE/Interpolators.cpp b/BBGE/Interpolators.cpp
index ed89a83..a071607 100644
--- a/BBGE/Interpolators.cpp
+++ b/BBGE/Interpolators.cpp
@@ -1,5 +1,6 @@
 #include "Interpolators.h"
 #include <math.h>
+#include "tbsp.hh"
 
 CosineInterpolator::CosineInterpolator()
 {
@@ -66,3 +67,56 @@ tail:
     for( ; k < n; ++k)
         ys[k] = v;
 }
+
+BSpline2D::BSpline2D()
+    : _cpx(0), _cpy(0), _degx(0), _degy(0), _tmin(0), _tmax(0)
+{
+}
+
+void BSpline2D::resize(size_t cx, size_t cy, unsigned degx, unsigned degy, float tmin, float tmax)
+{
+    controlpoints.resize(cx * cy);
+    knotsX.resize(tbsp__getNumKnots(cx, degx));
+    knotsY.resize(tbsp__getNumKnots(cy, degy));
+    tbsp::fillKnotVector<float>(&knotsX[0], cx, degx, tmin, tmax);
+    tbsp::fillKnotVector<float>(&knotsY[0], cy, degy, tmin, tmax);
+    _cpx = cx;
+    _cpy = cy;
+    _degx = degx;
+    _degy = degy;
+    _tmin = tmin;
+    _tmax = tmax;
+}
+
+void BSpline2D::recalc(Vector* dst, size_t xres, size_t yres)
+{
+    std::vector<Vector> tmpv;
+    size_t degn = std::max(_degx, _degy);
+    size_t tmpn = (yres * _cpx) + degn;
+    size_t tmpsz = tmpn * sizeof(Vector);
+    Vector *tmp;
+    if(tmpsz < 17*1024)
+        tmp = (Vector*)alloca(tmpsz);
+    else
+    {
+        tmpv.resize(tmpn);
+        tmp = &tmpv[0];
+    }
+    Vector *work = tmp + (tmpn - degn);
+
+    // Each column -> Y-axis interpolation
+    for(size_t x = 0; x < _cpx; ++x)
+    {
+        const Vector *srccol = &controlpoints[x];
+        Vector *dstcol = &tmp[x];
+        tbsp::evalRange(dstcol, yres, &work[0], &knotsY[0], srccol, _cpy, _degy, _tmin, _tmax, _cpx, _cpx);
+    }
+
+    // Each row -> X-axis interpolation
+    for(size_t y = 0; y < yres; ++y)
+    {
+        const Vector *srcrow = &tmp[y * _cpx];
+        tbsp::evalRange(dst, xres, &work[0], &knotsX[0], srcrow, _cpx, _degx, _tmin, _tmax);
+        dst += xres;
+    }
+}
diff --git a/BBGE/Interpolators.h b/BBGE/Interpolators.h
index 5d47d80..7d6d34c 100644
--- a/BBGE/Interpolators.h
+++ b/BBGE/Interpolators.h
@@ -21,3 +21,32 @@ private:
 
 #endif
 
+class BSpline2D
+{
+public:
+    BSpline2D();
+
+    // # of control points on each axis
+    void resize(size_t cx, size_t cy, unsigned degx, unsigned degy, float tmin, float tmax);
+    void recalc(Vector *dst, size_t xres, size_t yres);
+
+    std::vector<Vector> controlpoints;
+
+    inline Vector& controlpoint(size_t x, size_t y)
+    {
+        return controlpoints[y * _cpx + x];
+    }
+
+
+    inline size_t ctrlX() const { return _cpx; }
+    inline size_t ctrlY() const { return _cpy; }
+
+    inline unsigned degX() const { return _degx; }
+    inline unsigned degY() const { return _degy; }
+
+private:
+    size_t _cpx, _cpy; // # of control points
+    unsigned _degx, _degy;
+    float _tmin, _tmax;
+    std::vector<float> knotsX, knotsY;
+};
diff --git a/BBGE/Quad.h b/BBGE/Quad.h
index 49771b4..9baf55e 100644
--- a/BBGE/Quad.h
+++ b/BBGE/Quad.h
@@ -77,7 +77,8 @@ public:
 	enum GridType
 	{
 		GRID_WAVY	= 0,
-		GRID_SET	= 1
+		GRID_STRIP	= 1, // quad is in strip mode
+		GRID_INTERP = 2, // quad is in grid mode
 	};
 	unsigned char gridType;  // unsigned char to save space
 
diff --git a/BBGE/SkeletalSprite.cpp b/BBGE/SkeletalSprite.cpp
index b07e27c..ad56016 100644
--- a/BBGE/SkeletalSprite.cpp
+++ b/BBGE/SkeletalSprite.cpp
@@ -25,6 +25,7 @@ Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.
 #include "SimpleIStringStream.h"
 #include "ReadXML.h"
 #include "RenderBase.h"
+#include "SplineGrid.h"
 
 #include <tinyxml2.h>
 using namespace tinyxml2;
@@ -137,7 +138,7 @@ void Bone::createStrip(bool vert, int num)
 		createGrid(2, num);
 	}
 	stripVert = vert;
-	gridType = GRID_SET;
+	gridType = GRID_STRIP;
 	changeStrip.resize(num);
 }
 
@@ -1083,10 +1084,10 @@ bool SkeletalSprite::saveSkeletal(const std::string &fn)
 			{
 				BoneKeyframe *b = &a->keyframes[j].keyframes[k];
 				os << b->idx << " " << b->x << " " << b->y << " " << b->rot << " ";
-				os << b->strip.size() << " ";
-				for (size_t i = 0; i < b->strip.size(); i++)
+				os << b->grid.size() << " ";
+				for (size_t i = 0; i < b->grid.size(); i++)
 				{
-					os << b->strip[i].x << " " << b->strip[i].y << " ";
+					os << b->grid[i].x << " " << b->grid[i].y << " ";
 				}
 				if (b->doScale)
 				{
@@ -1652,10 +1653,10 @@ void SkeletalSprite::loadSkeletal(const std::string &fn)
 						b.rot = rot;
 						if (strip>0)
 						{
-							b.strip.resize(strip);
-							for (size_t i = 0; i < b.strip.size(); i++)
+							b.grid.resize(strip);
+							for (size_t i = 0; i < b.grid.size(); i++)
 							{
-								is >> b.strip[i].x >> b.strip[i].y;
+								is >> b.grid[i].x >> b.grid[i].y;
 
 							}
 						}
@@ -1746,9 +1747,38 @@ void SkeletalSprite::loadSkeletal(const std::string &fn)
 			XMLElement *interp = animation->FirstChildElement("Interpolator");
 			while(interp)
 			{
+				Bone *bi = NULL;
+				if(const char *sbone = interp->Attribute("bone"))
+					bi = getBoneByIdx(atoi(sbone));
+				if(!bi)
+					continue;
+				//SplineType spline = SPLINE_BSPLINE;
+				int cx = 3, cy = 3, degx = 3, degy = 3;
+				if(const char *stype = interp->Attribute("type"))
+				{
+					SimpleIStringStream is(stype);
+					std::string ty;
+					is >> ty;
+					BoneGridInterpolator bgip;
+					if(ty == "bspline")
+					{
+						//spline = SPLINE_BSPLINE;
+						is >> cx >> cy >> degx >> degy;
+					}
+				}
 
+				BoneGridInterpolator bgip;
+				//bgip.type = spline;
+				bgip.idx = bi->boneIdx;
+				bgip.pointsX = cx;
+				bgip.pointsY = cy;
+				bgip.degreeX = degx;
+				bgip.degreeY = degy;
+
+				// TODO
 
 				interp = interp->NextSiblingElement("Interpolator");
+
 			}
 
 			animation = animation->NextSiblingElement("Animation");
@@ -1885,13 +1915,13 @@ void AnimationLayer::updateBones()
 						}
 						if (b->animated==Bone::ANIM_ALL && !b->changeStrip.empty())
 						{
-							if (bkey2->strip.size() < b->changeStrip.size())
-								bkey2->strip.resize(b->changeStrip.size());
-							if (bkey1->strip.size() < b->changeStrip.size())
-								bkey1->strip.resize(b->changeStrip.size());
+							if (bkey2->grid.size() < b->changeStrip.size())
+								bkey2->grid.resize(b->changeStrip.size());
+							if (bkey1->grid.size() < b->changeStrip.size())
+								bkey1->grid.resize(b->changeStrip.size());
 							for (size_t i = 0; i < b->changeStrip.size(); i++)
 							{
-								b->changeStrip[i] = Vector(lerp(bkey1->strip[i].x, bkey2->strip[i].x, dt, lerpType), lerp(bkey1->strip[i].y, bkey2->strip[i].y, dt, lerpType));
+								b->changeStrip[i] = Vector(lerp(bkey1->grid[i].x, bkey2->grid[i].x, dt, lerpType), lerp(bkey1->grid[i].y, bkey2->grid[i].y, dt, lerpType));
 							}
 							b->setGridPoints(b->stripVert, b->changeStrip);
 						}
diff --git a/BBGE/SkeletalSprite.h b/BBGE/SkeletalSprite.h
index fe21e6f..d99c635 100644
--- a/BBGE/SkeletalSprite.h
+++ b/BBGE/SkeletalSprite.h
@@ -128,7 +128,8 @@ public:
 	int x, y, rot;
 	float sx, sy;
 	bool doScale;
-	std::vector<Vector> strip;
+	std::vector<Vector> grid;
+	std::vector<Vector> controlpoints;
 };
 
 class SkeletalKeyframe
@@ -150,6 +151,15 @@ public:
 	void copyAllButTime(SkeletalKeyframe *copy);
 };
 
+class BoneGridInterpolator
+{
+public:
+	size_t idx;
+	unsigned pointsX, pointsY;
+	unsigned degreeX, degreeY;
+	void updateGrid(BoneKeyframe& bk);
+};
+
 class Animation
 {
 public:
@@ -170,6 +180,9 @@ public:
 	size_t getNumKeyframes();
 	void reverse();
 	bool resetPassOnEnd;
+
+	typedef std::vector <BoneGridInterpolator> Interpolators;
+	Interpolators interpolators;
 };
 
 class SkeletalSprite;
diff --git a/BBGE/SplineGrid.cpp b/BBGE/SplineGrid.cpp
index af08848..a0538ae 100644
--- a/BBGE/SplineGrid.cpp
+++ b/BBGE/SplineGrid.cpp
@@ -1,8 +1,6 @@
 #include "SplineGrid.h"
-#include "tbsp.hh"
 #include "RenderBase.h"
 #include "Core.h"
-#include "Interpolators.h"
 
 SplineGridCtrlPoint *SplineGridCtrlPoint::movingPoint;
 
@@ -15,7 +13,8 @@ SplineGridCtrlPoint::SplineGridCtrlPoint()
 Vector SplineGridCtrlPoint::getSplinePosition() const
 {
     SplineGridCtrlPoint *p = (SplineGridCtrlPoint*)getParent();
-    return Vector(2 * position.x / p->width, 2 * position.y / p->height);
+    // always return alpha == 1
+    return Vector(position.x / p->width, position.y / p->height, 1.0f);
 }
 
 void SplineGridCtrlPoint::onUpdate(float dt)
@@ -40,7 +39,7 @@ void SplineGridCtrlPoint::onUpdate(float dt)
 }
 
 SplineGrid::SplineGrid()
-    : degreeX(3), degreeY(3), _cpx(0), _cpy(0), _xres(0), _yres(0), splinetype(SPLINE_BSPLINE)
+    : _xres(0), _yres(0)
 {
     setWidthHeight(128, 128);
     renderQuad = true;
@@ -51,16 +50,13 @@ SplineGrid::~SplineGrid()
 {
 }
 
-void SplineGrid::resize(size_t w, size_t h, size_t xres, size_t yres)
+void SplineGrid::resize(size_t w, size_t h, size_t xres, size_t yres, unsigned deg)
 {
-
-    knotsX.resize(tbsp__getNumKnots(w, degreeX));
-    knotsY.resize(tbsp__getNumKnots(h, degreeY));
-    tmp.resize(xres * h);
-    tbsp::fillKnotVector<float>(&knotsX[0], w, degreeX, -1.0f, 1.0f);
-    tbsp::fillKnotVector<float>(&knotsY[0], h, degreeY, -1.0f, 1.0f);
+    size_t oldcpx = bsp.ctrlX();
+    size_t oldcpy = bsp.ctrlY();
 
     this->createGrid(xres, yres);
+    gridpoints.resize(xres * yres);
 
     std::vector<SplineGridCtrlPoint*> oldp;
     ctrlp.swap(oldp);
@@ -68,21 +64,20 @@ void SplineGrid::resize(size_t w, size_t h, size_t xres, size_t yres)
 
     // move any old points over that fit within the new size
     {
-        const size_t cw = std::min(_cpx, w);
-        const size_t ch = std::min(_cpy, h);
+        const size_t cw = std::min(oldcpx, w);
+        const size_t ch = std::min(oldcpy, h);
         for(size_t y = 0; y < ch; ++y)
             for(size_t x = 0; x < cw; ++x)
             {
-                SplineGridCtrlPoint *& ref = oldp[y * _cpx + x];
+                SplineGridCtrlPoint *& ref = oldp[y * oldcpx + x];
                 ctrlp[y * w + x] = ref;
                 ref = NULL;
             }
     }
 
-    _cpx = w;
-    _cpy = h;
     _xres = xres;
     _yres = yres;
+    bsp.resize(w, h, deg, deg, -1.0f, 1.0f);
 
     // kill any excess points
     for(size_t i = 0; i < oldp.size(); ++i)
@@ -97,102 +92,46 @@ void SplineGrid::resize(size_t w, size_t h, size_t xres, size_t yres)
             if(!ref)
                 ref = createControlPoint(x, y);
         }
+
+    recalc();
 }
 
 void SplineGrid::recalc()
 {
-    switch(splinetype)
-    {
-        case SPLINE_BSPLINE:
-            recalcBSpline();
-            break;
-
-        case SPLINE_COSINE:
-            recalcCosine();
-            break;
-    }
-}
-
-void SplineGrid::recalcBSpline()
-{
-    std::vector<Vector> px(std::max(_cpx, _xres));
-    std::vector<Vector> work(std::max(degreeX, degreeY));
-
-    // Each row -> X-axis interpolation
-    for(size_t y = 0; y < _cpy; ++y)
-    {
-
-        for(size_t x = 0; x < _cpx; ++x)
-            px[x] = ctrlp[y * _cpx + x]->getSplinePosition();
-
-        Vector *dst = &tmp[y * _xres];
-        tbsp::evalRange(dst, _xres, &work[0], &knotsX[0], &px[0], _cpx, degreeX, -1.0f, 1.0f);
-    }
-
-    this->resetGrid();
+    for(size_t i = 0; i < ctrlp.size(); ++i)
+        bsp.controlpoints[i] = ctrlp[i]->getSplinePosition();
+    Vector *gp = &gridpoints[0];
+    bsp.recalc(gp, _xres, _yres);
     Vector **dg = this->getDrawGrid();
-
-    // Each column -> Y-axis interpolation
-    std::vector<Vector> out(_yres);
-    for(size_t x = 0; x < _xres; ++x)
-    {
-        for(size_t y = 0; y < _cpy; ++y)
-            px[y] = tmp[y * _xres + x];
-
-        tbsp::evalRange(&out[0], _yres, &work[0], &knotsY[0], &px[0], _cpy, degreeY, -1.0f, 1.0f);
-
-        for(size_t y = 0; y < _yres; ++y)
-        {
-            // output values are in [-1,+1] while drawgrid normally goes from [-0.5,+0.5]
-            Vector gp = out[y] * 0.5f;
-            gp.z = 1; // used as alpha
-            dg[x][y] = gp;
-        }
-    }
-
+    for(size_t y = 0; y < _yres; ++y)
+        for(size_t x = 0; x < _xres; ++x)
+            dg[x][y] = *gp++;
 }
 
-void SplineGrid::recalcCosine()
-{
-    // TODO
-    /*std::vector<Vector> px(std::max(_cpx, _xres));
-    std::vector<float> out(std::max(_xres, _yres));
-
-    CosineInterpolator top;
-    // Each row -> X-axis interpolation
-    for(size_t y = 0; y < _cpy; ++y)
-    {
-
-        for(size_t x = 0; x < _cpx; ++x)
-            px[x] = ctrlp[y * _cpx + x]->getSplinePosition();
-        std::sort(px.begin(), px.end());
-
-        top.setPoints(&px[0], _cpx);
-
-        Vector *dst = &tmp[y * _xres];
-        top.interpolateRange(&out[0,
-    */
-}
 
 void SplineGrid::resetControlPoints()
 {
-    const float dx = width / float(_cpx - 1);
-    const float dy = height / float(_cpy - 1);
+    const size_t cpx = bsp.ctrlX();
+    const size_t cpy = bsp.ctrlY();
+    const float dx = width / float(cpx - 1);
+    const float dy = height / float(cpy - 1);
     float yy = height * -0.5f;
-    for(size_t y = 0; y < _cpy; ++y, yy += dy)
+    for(size_t y = 0; y < cpy; ++y, yy += dy)
     {
         float xx = width * -0.5f;
-        SplineGridCtrlPoint **row = &ctrlp[y * _cpx];
-        for(size_t x = 0; x < _cpx; ++x, xx += dx)
+        SplineGridCtrlPoint **row = &ctrlp[y * cpx];
+        for(size_t x = 0; x < cpx; ++x, xx += dx)
             row[x]->position = Vector(xx, yy);
     }
 }
 
 SplineGridCtrlPoint* SplineGrid::createControlPoint(size_t x, size_t y)
 {
-    assert(x < _cpx && y < _cpy);
+    const size_t cpx = bsp.ctrlX();
+    const size_t cpy = bsp.ctrlY();
+    assert(x < cpx && y < cpy);
     const Vector wh(width, height);
-    const Vector pos01(float(x) / float(_cpx-1), float(y) / float(_cpy-1));
+    const Vector pos01(float(x) / float(cpx-1), float(y) / float(cpy-1));
     SplineGridCtrlPoint *cp = new SplineGridCtrlPoint();
     cp->position = (pos01 - Vector(0.5f, 0.5f)) * wh;
     this->addChild(cp, PM_POINTER);
@@ -215,12 +154,15 @@ void SplineGrid::onRender(const RenderState& rs) const
     glLineWidth(2);
     glColor4f(0.0f, 1.0f, 0.3f, 0.3f);
 
+    const size_t cpx = bsp.ctrlX();
+    const size_t cpy = bsp.ctrlY();
+
     // X axis
-    for(size_t y = 0; y < _cpy; ++y)
+    for(size_t y = 0; y < cpy; ++y)
     {
         glBegin(GL_LINE_STRIP);
-        const SplineGridCtrlPoint * const *row = &ctrlp[y * _cpx];
-        for(size_t x = 0; x < _cpx; ++x)
+        const SplineGridCtrlPoint * const *row = &ctrlp[y * cpx];
+        for(size_t x = 0; x < cpx; ++x)
         {
             const Vector p = row[x]->position;
             glVertex2f(p.x, p.y);
@@ -229,31 +171,16 @@ void SplineGrid::onRender(const RenderState& rs) const
     }
 
     // Y axis
-    for(size_t x = 0; x < _cpx; ++x)
+    for(size_t x = 0; x < cpx; ++x)
     {
         glBegin(GL_LINE_STRIP);
-        for(size_t y = 0; y < _cpy; ++y)
+        for(size_t y = 0; y < cpy; ++y)
         {
-            const Vector p = ctrlp[y * _cpx + x]->position;
+            const Vector p = ctrlp[y * cpx + x]->position;
             glVertex2f(p.x, p.y);
         }
         glEnd();
     }
-
-    /*
-    glColor4f(0.0f, 0.2f, 1.0f, 1.0f);
-    for(size_t y = 0; y < _cpy; ++y)
-    {
-        glBegin(GL_LINE_STRIP);
-        const value_type *line = &tmp[y * _xres];
-        for(size_t x = 0; x < _xres; ++x)
-        {
-            Vector pos = line[x] * wh2;
-            glVertex2f(pos.x, pos.y);
-        }
-        glEnd();
-    }
-    */
 }
 
 
diff --git a/BBGE/SplineGrid.h b/BBGE/SplineGrid.h
index 5a3a671..d2a8e82 100644
--- a/BBGE/SplineGrid.h
+++ b/BBGE/SplineGrid.h
@@ -5,6 +5,7 @@
 #include "Vector.h"
 #include "glm/glm.hpp"
 #include "Quad.h"
+#include "Interpolators.h"
 
 enum SplineType
 {
@@ -31,7 +32,7 @@ public:
 	~SplineGrid();
 
 	// # of control points on each axis
-	void resize(size_t w, size_t h, size_t xres, size_t yres);
+	void resize(size_t w, size_t h, size_t xres, size_t yres, unsigned deg);
 	void recalc();
 
 	void resetControlPoints();
@@ -41,22 +42,13 @@ public:
 
 private:
 
-	void recalcBSpline();
-	void recalcCosine();
-
 	SplineGridCtrlPoint *createControlPoint(size_t x, size_t y);
 
-	unsigned degreeX, degreeY;
-
-	std::vector<float> knotsX, knotsY;
-	std::vector<value_type> tmp; // X-axis temporary values
-
-	size_t _cpx, _cpy; // # of control points
-	size_t _xres; // resolution for x-axis eval before doing 2D expansion
-	size_t _yres;
-
 	std::vector<SplineGridCtrlPoint*> ctrlp;
-	SplineType splinetype;
+	std::vector<Vector> gridpoints;
+	size_t _xres, _yres;
+	unsigned deg;
+	BSpline2D bsp;
 };
 
 
diff --git a/ExternalLibs/tbsp.hh b/ExternalLibs/tbsp.hh
index 8855389..bacec09 100644
--- a/ExternalLibs/tbsp.hh
+++ b/ExternalLibs/tbsp.hh
@@ -95,14 +95,14 @@ static void genKnotsUniform(K *knots, size_t nn, K mink, K maxk)
 }
 
 template<typename K, typename T>
-static T deBoor(T *work, const T *src, const K *knots, const size_t r, const size_t k, const K t)
+static T deBoor(T *work, const T *src, const K *knots, const size_t r, const size_t k, const K t, size_t inputStride)
 {
     T last = src[0]; // init so that it works correctly even with degree == 0
     for(size_t worksize = k; worksize > 1; --worksize)
     {
         const size_t j = k - worksize + 1; // iteration number, starting with 1, going up to k
         const size_t tmp = r - k + 1 + j;
-        for(size_t w = 0; w < worksize - 1; ++w)
+        for(size_t w = 0, wr = 0; w < worksize - 1; ++w, wr += inputStride)
         {
             const size_t i = w + tmp;
             const K ki = knots[i];
@@ -111,9 +111,10 @@ static T deBoor(T *work, const T *src, const K *knots, const size_t r, const siz
             TBSP_ASSERT(div > 0);
             const K a = (t - ki) / div;
             const K a1 = K(1) - a;
-            work[w] = last = (src[w] * a1) + (src[w+1] * a); // lerp
+            work[w] = last = (src[wr] * a1) + (src[wr + inputStride] * a); // lerp
         }
         src = work; // done writing the initial data to work, now use that as input for further iterations
+        inputStride = 1;
     }
     return last;
 }
@@ -175,7 +176,7 @@ static T evalOne(T *work, const K *knots, const T *points, size_t numpoints, siz
 
 // evaluate numdst points in range [tmin..tmax], equally spaced
 template<typename K, typename T>
-static void evalRange(T *dst, size_t numdst, T *work, const K *knots, const T *points, size_t numpoints, size_t degree, K tmin, K tmax)
+static void evalRange(T *dst, size_t numdst, T *work, const K *knots, const T *points, size_t numpoints, size_t degree, K tmin, K tmax, size_t inputStride = 1, size_t outputStride = 1)
 {
     TBSP_ASSERT(tmin <= tmax);
     if(numpoints - 1 < degree)
@@ -196,7 +197,10 @@ static void evalRange(T *dst, size_t numdst, T *work, const K *knots, const T *p
 
     // left out-of-bounds
     for( ; i < numdst && t < knots[0]; ++i, t += step)
-        dst[i] = points[0];
+    {
+        *dst = points[0];
+        dst += outputStride;
+    }
 
     // actually interpolated points
     const K maxknot = knots[numknots - 1];
@@ -205,13 +209,21 @@ static void evalRange(T *dst, size_t numdst, T *work, const K *knots, const T *p
         while(r < maxidx && knots[r+1] < t) // find new index; don't need to do binary search again
             ++r;
 
-        const T* const src = &points[r - degree];
-        dst[i] = detail::deBoor(work, src, knots, r, k, t);
+        const T* const src = &points[(r - degree) * inputStride];
+        *dst = detail::deBoor(work, src, knots, r, k, t, inputStride);
+        dst += outputStride;
     }
 
     // right out-of-bounds
-    for( ; i < numdst; ++i)
-        dst[i] = points[numpoints - 1];
+    if(i < numdst)
+    {
+        T last =  points[(numpoints - 1) * inputStride];
+        for( ; i < numdst; ++i)
+        {
+            *dst = last;
+            dst += outputStride;
+        }
+    }
 }