diff --git a/Extras/BulletMultiThreaded/SpuBatchRaycaster.cpp b/Extras/BulletMultiThreaded/SpuBatchRaycaster.cpp
index 79bd97833..64d073685 100644
--- a/Extras/BulletMultiThreaded/SpuBatchRaycaster.cpp
+++ b/Extras/BulletMultiThreaded/SpuBatchRaycaster.cpp
@@ -14,6 +14,7 @@ subject to the following restrictions:
 */
 
 #include <new>
+#include "BulletCollision/CollisionShapes/btCollisionShape.h"
 #include "LinearMath/btAlignedAllocator.h"
 #include "SpuBatchRaycaster.h"
 
@@ -55,10 +56,57 @@ SpuBatchRaycaster::setCollisionObjects (btCollisionObjectArray& castUponObjects,
 }
 
 void
-SpuBatchRaycaster::addRay (const btVector3& rayFrom, const btVector3& rayTo)
+SpuBatchRaycaster::setCollisionObjectsSkipPE (btCollisionObjectArray& castUponObjects, int numCastUponObjects)
+{
+	if (castUponObjectWrappers)
+	{
+		btAlignedFree (castUponObjectWrappers);
+		castUponObjectWrappers = NULL;
+	}
+
+	int numNonPEShapes = 0;
+	for (int i = 0; i < numCastUponObjects; i++)
+	{
+		const btCollisionShape* shape = castUponObjects[i]->getCollisionShape();
+
+		if (shape->getShapeType () == BOX_SHAPE_PROXYTYPE ||
+			shape->getShapeType () == SPHERE_SHAPE_PROXYTYPE ||
+			shape->getShapeType () == CAPSULE_SHAPE_PROXYTYPE)
+		{
+			continue;
+		}
+
+		numNonPEShapes++;
+	}
+
+	castUponObjectWrappers = (SpuCollisionObjectWrapper*)btAlignedAlloc (sizeof(SpuCollisionObjectWrapper) * numNonPEShapes,16);
+	numCastUponObjectWrappers = numNonPEShapes;
+
+	int index = 0;
+	for (int i = 0; i < numCastUponObjects; i++)
+	{
+		const btCollisionShape* shape = castUponObjects[i]->getCollisionShape();
+
+		if (shape->getShapeType () == BOX_SHAPE_PROXYTYPE ||
+			shape->getShapeType () == SPHERE_SHAPE_PROXYTYPE ||
+			shape->getShapeType () == CAPSULE_SHAPE_PROXYTYPE)
+		{
+			continue;
+		}
+
+		castUponObjectWrappers[index] = SpuCollisionObjectWrapper(castUponObjects[i]);
+		index++;
+	}
+
+	printf("Number of shapes bullet is casting against: %d\n", numNonPEShapes);
+	btAssert (index == numNonPEShapes);
+}
+
+void
+SpuBatchRaycaster::addRay (const btVector3& rayFrom, const btVector3& rayTo, const btScalar hitFraction)
 {
 	SpuRaycastTaskWorkUnitOut workUnitOut;
-	workUnitOut.hitFraction = 1.0;
+	workUnitOut.hitFraction = hitFraction;
 	workUnitOut.hitNormal = btVector3(0.0, 1.0, 0.0);
 
 	rayBatchOutput.push_back (workUnitOut);
diff --git a/Extras/BulletMultiThreaded/SpuBatchRaycaster.h b/Extras/BulletMultiThreaded/SpuBatchRaycaster.h
index da2f65de6..88271fbd7 100644
--- a/Extras/BulletMultiThreaded/SpuBatchRaycaster.h
+++ b/Extras/BulletMultiThreaded/SpuBatchRaycaster.h
@@ -38,8 +38,8 @@ public:
 	SpuBatchRaycaster (class btThreadSupportInterface* threadInterface, int maxNumOutstandingTasks);
 	~SpuBatchRaycaster ();
 	void setCollisionObjects (btCollisionObjectArray& castUponObjects, int numCastUponObjects);
-
-	void addRay (const btVector3& rayFrom, const btVector3& rayTo);
+	void setCollisionObjectsSkipPE (btCollisionObjectArray& castUponObjects, int numCastUponObjects);
+	void addRay (const btVector3& rayFrom, const btVector3& rayTo, const btScalar hitFraction = 1.0);
 	void clearRays ();
 	void performBatchRaycast ();
 	const SpuRaycastTaskWorkUnitOut& operator [] (int i) const;
diff --git a/Extras/BulletMultiThreaded/SpuNarrowPhaseCollisionTask/SpuCollisionShapes.cpp b/Extras/BulletMultiThreaded/SpuNarrowPhaseCollisionTask/SpuCollisionShapes.cpp
index 5dec4d4ff..8d987165c 100644
--- a/Extras/BulletMultiThreaded/SpuNarrowPhaseCollisionTask/SpuCollisionShapes.cpp
+++ b/Extras/BulletMultiThreaded/SpuNarrowPhaseCollisionTask/SpuCollisionShapes.cpp
@@ -298,7 +298,11 @@ void dmaBvhShapeData (bvhMeshShape_LocalStoreMemory* bvhMeshShape, btBvhTriangle
 	dmaSize = sizeof(btTriangleIndexVertexArray);
 	dmaPpuAddress2 = reinterpret_cast<ppu_address_t>(triMeshShape->getMeshInterface());
 	//	spu_printf("trimeshShape->getMeshInterface() == %llx\n",dmaPpuAddress2);
+#ifdef DISABLE_CELL_DMA_GET_READ_ONLY // needed for PE ray caster
+	cellDmaGet(&bvhMeshShape->gTriangleMeshInterfaceStorage, dmaPpuAddress2  , dmaSize, DMA_TAG(1), 0, 0);
+#else
 	bvhMeshShape->gTriangleMeshInterfacePtr = (btTriangleIndexVertexArray*)cellDmaGetReadOnly(&bvhMeshShape->gTriangleMeshInterfaceStorage, dmaPpuAddress2  , dmaSize, DMA_TAG(1), 0, 0);
+#endif
 	//cellDmaWaitTagStatusAll(DMA_MASK(1));
 	
 	///now DMA over the BVH