diff --git a/src/java/org/lwjgl/test/opengl/Grass.java b/src/java/org/lwjgl/test/opengl/Grass.java
new file mode 100644
index 00000000..105e3e8a
--- /dev/null
+++ b/src/java/org/lwjgl/test/opengl/Grass.java
@@ -0,0 +1,320 @@
+import org.lwjgl.input.*;
+import org.lwjgl.opengl.*;
+import org.lwjgl.*;
+
+import java.net.*;
+import java.io.*;
+import java.nio.*;
+import java.util.*;
+
+	class Aslod {	
+		float angle;
+		float value;
+		float ripple;
+		float count;
+	}
+
+public class Grass {
+
+        private static boolean finished = false;
+        private static Random rand = new Random();
+
+	static {
+		try {
+			DisplayMode[] modes = Display.getAvailableDisplayModes();
+			System.out.println("Available display modes:");
+			for (int i = 0; i < modes.length; i ++)
+				System.out.println(modes[i]);
+			// For now let's just pick a mode we're certain to have
+			Display.create(new DisplayMode(800, 600, 16, 60), false);
+			System.out.println("Created display.");
+		} catch (Exception e) {
+			System.err.println("Failed to create display due to "+e);
+			System.exit(1);
+		}
+	}
+
+	public static final GL gl = new GL(16, 0, 16, 8);
+	public static final GLU glu = new GLU(gl);
+	static {
+		try {
+			gl.create();
+                        Keyboard.create();
+                        Keyboard.enableBuffer();
+                        Mouse.create();
+			System.out.println("Created OpenGL.");
+		} catch (Exception e) {
+			System.err.println("Failed to create OpenGL due to "+e);
+			System.exit(1);
+		}
+
+	}
+
+        private static Aslod aslod = new Aslod();
+
+        private static int mesh;
+        private static int program_handle;
+
+	private static byte[] loadFile(String file) {
+		int next;
+		java.util.Vector bytes = new java.util.Vector();
+		try {
+                        ClassLoader loader = ClassLoader.getSystemClassLoader();
+                        URL url = loader.getResource(file);
+			InputStream stream = new BufferedInputStream(url.openStream());
+			while ((next = (stream.read())) != -1)
+				bytes.add(new Byte((byte)next));
+			stream.close();
+			byte[] result = new byte[bytes.size()];
+			for (int i = 0; i < result.length; i++)
+				result[i] = ((Byte)bytes.get(i)).byteValue();
+			return result;
+		} catch (Exception e) {
+			e.printStackTrace();
+		}
+		return null;
+	}
+
+
+        public static void main(String[] args) {
+                
+		ByteBuffer byte_buf = ByteBuffer.allocateDirect(4);
+                byte_buf.order(ByteOrder.nativeOrder());
+                System.out.println("Vertex program supported: " + gl.NV_vertex_program);
+	        gl.genProgramsNV(1, Sys.getDirectBufferAddress(byte_buf));
+                IntBuffer int_buf = byte_buf.asIntBuffer();
+        	if (int_buf.get(0) == 0)
+	        	throw new RuntimeException("Could not allocate new vertex program id!");
+
+                program_handle = int_buf.get(0);
+		byte[] program = loadFile("cg_grass2.vp");
+                ByteBuffer program_buf = ByteBuffer.allocateDirect(program.length);
+                program_buf.order(ByteOrder.nativeOrder());
+                program_buf.rewind();
+                program_buf.put(program);
+                program_buf.rewind();
+		gl.loadProgramNV(GL.VERTEX_PROGRAM_NV, program_handle, program_buf.remaining(), Sys.getDirectBufferAddress(program_buf));
+		/*gl.getIntegerv(GL.PROGRAM_ERROR_POSITION_NV, Sys.getDirectBufferAddress(int_buf));
+		System.out.println("error position: " + int_buf.get(0));*/
+
+		genMesh();
+
+		float[] LightDiffuse = {1.0f, 0.0f, 0.0f, 1.0f};
+		float[] LightPosition = {1.0f, 1.0f, 1.0f, 0.0f};
+                ByteBuffer light_buf = ByteBuffer.allocateDirect(4*4);
+                light_buf.order(ByteOrder.nativeOrder());
+                FloatBuffer light_buf_f = light_buf.asFloatBuffer();
+                light_buf_f.rewind();
+                light_buf_f.put(LightDiffuse);
+
+		gl.lightfv(GL.LIGHT0, GL.DIFFUSE, Sys.getDirectBufferAddress(light_buf_f));
+                light_buf_f.rewind();
+                light_buf_f.put(LightPosition);
+		gl.lightfv(GL.LIGHT0, GL.POSITION, Sys.getDirectBufferAddress(light_buf_f));
+		gl.enable(GL.LIGHT0);
+
+		gl.enable(GL.LIGHTING);
+
+		gl.enable(GL.DEPTH_TEST);
+
+		gl.blendFunc(GL.SRC_ALPHA, GL.ONE_MINUS_SRC_ALPHA);
+		gl.enable(GL.BLEND);
+
+		gl.matrixMode(GL.PROJECTION);
+		glu.perspective(40.0, 1.0, 1.0, 50.0);
+
+		gl.matrixMode(GL.MODELVIEW);
+
+		glu.lookAt(14.0, 10.0, -16.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0);
+
+		aslod.angle = 2.6179935f;
+		aslod.value = 0.2f;
+		aslod.ripple = 0.0f;
+		aslod.count = 0.0f;
+
+		while (!finished) {
+                        keyPoll();
+			float degree = (1.0f + (aslod.value * 20.0f)) * 0.01745329f;
+
+			degree *= (0.5 + myrand()) ;
+
+			ptrAnimate(degree);
+			gl.clear(GL.COLOR_BUFFER_BIT | GL.DEPTH_BUFFER_BIT);
+
+			//ptrDraw();
+
+			grsDraw();
+
+			gl.swapBuffers();
+		}
+                Mouse.destroy();
+                Keyboard.destroy();
+                gl.destroy();
+                Display.destroy();
+        }
+
+	private static float myrand() {
+		// returns a value between 0 and 1
+		return rand.nextFloat(); 
+	}
+
+	private static void genGrass( float fFaceHeight, float fFaceWidth, float fX, float fZ ) {
+		int cFaces;
+		int numFaces;
+		float cWidth;
+		float fDecWidth, frndWidth, frndHeight;
+		float fRotate;
+		float fRigid;
+
+		numFaces = 5;
+		frndHeight = fFaceHeight + ((fFaceHeight / 1.5f) * (float)java.lang.Math.cos(java.lang.Math.abs(rand.nextInt())));
+		frndWidth = fFaceWidth + ((fFaceWidth / 4.0f) * (float)java.lang.Math.cos(java.lang.Math.abs(rand.nextInt())));
+
+		fDecWidth = frndWidth / 5.0f;
+
+		fRotate = myrand() * 3.1415f;
+
+		fRigid = ((fRigid = myrand()) < 0.2f) ? 0.2f : fRigid;
+
+		if (myrand() < 0.3)
+			gl.begin(GL.LINE_STRIP);
+
+		else
+			gl.begin(GL.QUAD_STRIP);
+
+		for (cFaces = 0; cFaces < numFaces; cFaces++)
+		{
+			for (cWidth = frndWidth; cWidth >=-frndWidth; cWidth-=(frndWidth * 2.0f))
+			{
+				gl.color4f(fX, fRigid, fZ, (float)cFaces/(float)numFaces);
+
+				gl.vertex3f((float)(((cFaces-2)*0.1f)*java.lang.Math.cos(fRotate)+(cWidth)*java.lang.Math.sin(fRotate)), cFaces*frndHeight, 
+						-(float)(((cFaces-2)*0.1f)*java.lang.Math.sin(fRotate) + (cWidth)*java.lang.Math.cos(fRotate)));
+			}
+
+			frndWidth -= fDecWidth;
+		}
+
+		gl.end();
+
+	}
+
+	private static void genMesh() {
+		float cI, cJ, fArea;
+
+		fArea = 20.0f;
+                mesh = gl.genLists(1);
+		gl.newList(mesh, GL.COMPILE);
+
+		for (cI = -fArea/2; cI < fArea/2; cI+=0.25f)
+		{
+			for (cJ = -fArea/2; cJ < fArea/2; cJ+=0.25f)
+			{
+				genGrass(0.5f, 0.1f, cI, cJ);
+			}
+		}
+
+		gl.endList();
+
+	}
+
+	private static void grsDraw()
+	{
+		gl.enable(GL.VERTEX_PROGRAM_NV);
+		gl.bindProgramNV(GL.VERTEX_PROGRAM_NV, program_handle);
+		gl.trackMatrixNV(GL.VERTEX_PROGRAM_NV, 0, GL.MODELVIEW_PROJECTION_NV, GL.IDENTITY_NV);
+
+		gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 4, 0.0f, 0.0f, 0.0f, 0.0f);
+		gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 5, 0.0f, 0.0f, 0.0f, 0.0f);
+		gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 6, 1.763609f, 0.496495f, 0.0f, 0.0f);
+		gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 7, -0.943599f, 3.203737f, 0.0f, 0.0f);
+		gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 8, 4.101107f, 0.943413f, 0.0f, 0.0f);  
+		gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 9, -1.218603f, 6.259399f, 0.0f, 0.0f);  
+		gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 10, 7.214299f, 1.352961f, 0.0f, 0.0f);
+		gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 11, -1.540748f, 10.080958f, 0.0f, 0.0f);
+		gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 12, 10.880035f, 1.759046f, 0.0f, 0.0f);
+		gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 13, -1.852705f, 14.468674f, 0.0f, 0.0f);
+		gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 14, 14.292879f, 1.973329f, 0.0f, 0.0f);
+		gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 15, -1.973387f, 18.506531f, 0.0f, 0.0f);
+		gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 16, (float)(java.lang.Math.sin(aslod.angle) * (aslod.value + aslod.ripple)), 0.0f, (float)(java.lang.Math.cos(aslod.angle) * (aslod.value + aslod.ripple)), 0.0f);
+
+                gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 17, 1.7f, 5f, 2f, 0f);
+                gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 18, -0.0187293f, 0.074261f, 0.2121144f, 1.570729f);
+                gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 20, 0f, 0.5f, 1f, 0f);
+                gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 21, 0.25f, -9f, 0.75f, 0.1591549f);
+                gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 22, 24.9808f, -24.9808f, -60.14581f, 60.14581f);
+                gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 23, 85.45379f, -85.45379f, -64.93935f, 64.93935f);
+                gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 24, 19.73921f, -19.73921f, -1f, 1f);
+                gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 25, 0f, 4f, 0f, 0f);
+                gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 19, 1f, 3.141593f, 0.5f, 1f);
+                gl.programParameter4fNV( GL.VERTEX_PROGRAM_NV, 26, 0.7f, 0.4f, 0f, 0f);
+		gl.callList(mesh);
+		gl.disable(GL.VERTEX_PROGRAM_NV);
+
+	}
+
+/*	private static void ptrDraw()
+	{
+		glRotatef((aslod.angle * 180.0f) / 3.1415f, 0, 1, 0);
+		glTranslatef(0, 4.5, -7.5);
+		glRotatef(-90, 0, 1, 0);
+		glRotatef(-45, 0, 0, 1);
+
+		glMaterialfv(GL.FRONT, GL.AMBIENT, vec4f(.1f,.1f,0,1).v);
+		glMaterialfv(GL.FRONT, GL.DIFFUSE, vec4f(.6f,.6f,.1f,1).v);
+		glMaterialfv(GL.FRONT, GL.SPECULAR, vec4f(1,1,.75f,1).v);
+		glMaterialf(GL.FRONT, GL.SHININESS, 128.f);
+
+		glutSolidTeapot(aslod.value*5);
+
+		gl.rotatef(45, 0, 0, 1);
+		gl.totatef(90, 0, 1, 0);
+		gl.translatef(0, -4.5, 7.5);
+		gl.rotatef(-(aslod.angle * 180.0f) / 3.1415f, 0f, 1f, 0f);
+
+	}
+*/
+	private static void ptrAnimate(float degree)
+	{
+		aslod.count += degree;
+
+		aslod.ripple = (float)(java.lang.Math.cos(aslod.count) / 80.0);
+
+	}
+
+	private static void keyPoll() {
+		Keyboard.read();
+		for (int i = 0; i < Keyboard.getNumKeyboardEvents(); i++) {
+			Keyboard.next();
+			if (!Keyboard.state)
+				continue;
+			switch(Keyboard.key)
+			{
+				case Keyboard.KEY_A:
+
+					aslod.angle += 0.1;
+					break;
+
+				case Keyboard.KEY_D:
+
+					aslod.angle -= 0.1;
+					break;
+
+				case Keyboard.KEY_W:
+
+					aslod.value += (aslod.value >= 0.15) ? 0.0 : 0.0025;
+					break;
+
+				case Keyboard.KEY_S:
+
+					aslod.value -= (aslod.value <= 0.005) ? 0.0 : 0.0025;
+					break;
+				case Keyboard.KEY_ESCAPE:
+					finished = true;
+					break;
+			}
+		}
+	}
+
+}
+
diff --git a/src/java/org/lwjgl/test/opengl/cg_grass2.cg b/src/java/org/lwjgl/test/opengl/cg_grass2.cg
new file mode 100644
index 00000000..4a743e9e
--- /dev/null
+++ b/src/java/org/lwjgl/test/opengl/cg_grass2.cg
@@ -0,0 +1,72 @@
+#pragma bind app2vert.position = ATTR0
+#pragma bind app2vert.normal = ATTR2
+#pragma bind app2vert.color = ATTR3
+
+
+struct app2vert : application2vertex
+{
+  float4 position; //in object space
+  float3 normal; //in object space
+  float4 color;
+};
+
+#pragma bind vert2frag.HPOS = HPOS
+#pragma bind vert2frag.COL0 = COL0
+
+struct vert2frag : vertex2fragment 
+{
+  float4 HPOS;
+  float4 COL0;
+};
+
+vert2frag main(app2vert IN,
+               uniform float4x4 ModelViewProj,
+               uniform float4 Stiff[12],
+               uniform float4 Load)
+{
+	vert2frag OUT;
+	float4 position;
+	float3 normal;
+	float3 disloc;
+    float4 dif;
+    float index;
+    float teta;
+	float alpha;
+
+	alpha = IN.position.y / 1.7;
+
+	// matrix line index
+    index = IN.color.a * 5 * 2;
+
+	// straw stiffness (randomic)
+	Load = Load * IN.color.y;
+
+    disloc.x = IN.position.x + dot(Stiff[index].xy, Load.xz);
+    disloc.y = IN.position.y;
+    disloc.z = IN.position.z + dot(Stiff[index + 1].xy, Load.xz);
+
+	// find the normal
+    teta = acos(IN.position.y / length(disloc));
+
+    normal = IN.position.xyz - disloc.xyz;
+    normal.y = sin(teta) * length(disloc);
+    normal = normalize(normal);
+
+    position.x = disloc.x + IN.color.x;
+    position.y = disloc.y;
+    position.z = disloc.z + IN.color.z;
+    position.w = 1;
+
+    OUT.HPOS = mul(ModelViewProj, position);
+
+	// calculate diffuse lighting off the normal that was just calculated
+    float4 lightPos = float4(0,4,0,0);
+    float4 lightVec = normalize(lightPos - position);
+    float diffuseInten = dot(lightVec.xyz, normal);
+
+	diffuseInten = (diffuseInten < 0.5) ? 0.5 : diffuseInten;
+
+    OUT.COL0 = float4(IN.color.y * 0.7, 0.7, 0.4, alpha) * diffuseInten;
+
+    return OUT;
+}
diff --git a/src/java/org/lwjgl/test/opengl/cg_grass2.vp b/src/java/org/lwjgl/test/opengl/cg_grass2.vp
new file mode 100644
index 00000000..892a6430
--- /dev/null
+++ b/src/java/org/lwjgl/test/opengl/cg_grass2.vp
@@ -0,0 +1,116 @@
+!!VP1.0
+# NV_vertex_program generated by NVIDIA Cg compiler
+# cgc version 1.5.0001, build date Sep 12 2002  15:49:58
+# command line args: -profile vp20 cg_grass2.cg
+# nv30vp backend compiling 'main' program
+#vendor NVIDIA Corporation
+#version 1.0.1
+#profile vp20
+#program main
+#semantic main.ModelViewProj
+#semantic main.Stiff
+#semantic main.Load
+#var float4 IN.position : $vin.ATTR0 : ATTR0 : 0 : 1
+#var float3 IN.normal : $vin.ATTR2 : ATTR2 : 0 : 1
+#var float4 IN.color : $vin.ATTR3 : ATTR3 : 0 : 1
+#var float4x4 ModelViewProj :  : c[0], 4 : 1 : 1
+#var float4 Stiff[0] :  : c[4] : 2 : 1
+#var float4 Stiff[1] :  : c[5] : 2 : 1
+#var float4 Stiff[2] :  : c[6] : 2 : 1
+#var float4 Stiff[3] :  : c[7] : 2 : 1
+#var float4 Stiff[4] :  : c[8] : 2 : 1
+#var float4 Stiff[5] :  : c[9] : 2 : 1
+#var float4 Stiff[6] :  : c[10] : 2 : 1
+#var float4 Stiff[7] :  : c[11] : 2 : 1
+#var float4 Stiff[8] :  : c[12] : 2 : 1
+#var float4 Stiff[9] :  : c[13] : 2 : 1
+#var float4 Stiff[10] :  : c[14] : 2 : 1
+#var float4 Stiff[11] :  : c[15] : 2 : 1
+#var float4 Load :  : c[16] : 3 : 1
+#var float4 HPOS : $vout.HPOS : HPOS : -1 : 1
+#var float4 COL0 : $vout.COL0 : COL0 : -1 : 1
+#const c[17] = 1.7 5 2 0
+#const c[18] = -0.0187293 0.074261 0.2121144 1.570729
+#const c[20] = 0 0.5 1 0
+#const c[21] = 0.25 -9 0.75 0.1591549
+#const c[22] = 24.9808 -24.9808 -60.14581 60.14581
+#const c[23] = 85.45379 -85.45379 -64.93935 64.93935
+#const c[24] = 19.73921 -19.73921 -1 1
+#const c[25] = 0 4 0 0
+#const c[19] = 1 3.141593 0.5 1
+#const c[26] = 0.7 0.4 0 0
+	MUL R2, c[16], v[3].y;
+	MUL R0, v[3].w, c[17].y;
+	MUL R1, R0.x, c[17].z;
+	ARL A0.x, R1.x;
+	MUL R0, c[A0.x + 4].xyxx, R2.xzxx;
+	ADD R0, R0.x, R0.y;
+	ADD R3.x, v[0].x, R0.x;
+	MOV R3.y, v[0].y;
+	MUL R0, c[A0.x + 5].xyxx, R2.xzxx;
+	ADD R0, R0.x, R0.y;
+	ADD R3.z, v[0].z, R0.x;
+	ADD R2.x, R3.x, v[3].x;
+	MOV R2.y, R3.y;
+	ADD R2.z, R3.z, v[3].z;
+	MOV R2.w, c[19].x;
+	DP4 R0.x, c[0], R2;
+	DP4 R0.y, c[1], R2;
+	DP4 R0.z, c[2], R2;
+	DP4 R0.w, c[3], R2;
+	MOV o[HPOS], R0;
+	ADD R1.xz, v[0].xyzx, -R3.xyzx;
+	DP3 R0, R3.xyzx, R3.xyzx;
+	RSQ R0, R0.x;
+	RCP R0, R0.x;
+	RCP R3, R0.x;
+	MUL R6, v[0].y, R3.x;
+	MAX R5, R6.x, -R6.x;
+	MAD R3, c[18].x, R5.x, c[18].y;
+	MAD R3, R3.x, R5.x, -c[18].z;
+	MAD R4, R3.x, R5.x, c[18].w;
+	ADD R3, c[19].x, -R5.x;
+	RSQ R3, R3.x;
+	RCP R3, R3.x;
+	MUL R5, R4.x, R3.x;
+	SLT R4, R6.x, c[17].w;
+	MUL R3, c[17].z, R4.x;
+	MUL R3, R3.x, R5.x;
+	ADD R3, R5.x, -R3.x;
+	MAD R3, R4.x, c[19].y, R3.x;
+	MAD R6, c[21].w, R3.x, -c[21].x;
+	EXP R6.y, R6.x;
+	ADD R3, c[20].xyzx, -R6.yyyy;
+	MUL R5, R3.xyzx, R3.xyzx;
+	SLT R4.x, R6.y, c[21].x;
+	SGE R3, R6.yyyy, c[21].yzyy;
+	MOV R4.yz, R3.xxyw;
+	DP3 R4.y, R4.xyzx, c[24].zwzz;
+	MAD R3, c[22].xyxx, R5.xyzx, c[22].zwzz;
+	MAD R3, R3.xyzx, R5.xyzx, c[23].xyxx;
+	MAD R3, R3.xyzx, R5.xyzx, c[23].zwzz;
+	MAD R3, R3.xyzx, R5.xyzx, c[24].xyxx;
+	MAD R3, R3.xyzx, R5.xyzx, c[24].zwzz;
+	DP3 R5.x, R3.xyzx, -R4.xyzx;
+	MUL R1.y, R5.x, R0.x;
+	DP3 R0, R1.xyzx, R1.xyzx;
+	RSQ R0, R0.x;
+	MUL R3, R0.x, R1.xyzx;
+	ADD R1, c[25], -R2;
+	DP4 R0, R1, R1;
+	RSQ R0, R0.x;
+	MUL R0, R0.x, R1;
+	DP3 R2, R0.xyzx, R3.xyzx;
+	SLT R1, R2.x, c[19].z;
+	ADD R0, c[19].w, -R1.x;
+	MUL R0, R0.x, R2.x;
+	MAD R2, R1.x, c[19].z, R0.x;
+	MUL R1.x, v[3].y, c[26].x;
+	MOV R1.y, c[26].x;
+	MOV R1.z, c[26].y;
+	RCP R0, c[17].x;
+	MUL R1.w, v[0].y, R0.x;
+	MUL o[COL0], R1, R2.x;
+END
+# 72 instructions
+# End of program