diff --git a/Extras/Serialize/BulletFileLoader/bFile.cpp b/Extras/Serialize/BulletFileLoader/bFile.cpp
index 0f5b7bcb0..3531dd59c 100644
--- a/Extras/Serialize/BulletFileLoader/bFile.cpp
+++ b/Extras/Serialize/BulletFileLoader/bFile.cpp
@@ -183,7 +183,8 @@ void bFile::parseInternal(bool verboseDumpAllTypes, char* memDna,int memDnaLengt
 		return;
 
 	char *blenderData = mFileBuffer;
-	int sdnaPos=0;
+	bChunkInd dna;
+	dna.oldPtr = 0;
 
 	char *tempBuffer = blenderData;
 	for (int i=0; i<mFileLen; i++)
@@ -193,15 +194,33 @@ void bFile::parseInternal(bool verboseDumpAllTypes, char* memDna,int memDnaLengt
 
 		if (!mDataStart && strncmp(tempBuffer, "REND", 4)==0)
 			mDataStart = i;
-		if (!sdnaPos && strncmp(tempBuffer, "SDNANAME", 8)==0)
-			sdnaPos = i;
 
-		if (mDataStart && sdnaPos) break;
+		if (strncmp(tempBuffer, "DNA1", 4)==0)
+		{
+			// read the DNA1 block and extract SDNA
+			if (getNextBlock(&dna, tempBuffer, mFlags) > 0)
+			{
+				if (strncmp((tempBuffer + ChunkUtils::getOffset(mFlags)), "SDNANAME", 8) ==0) 
+					dna.oldPtr = (tempBuffer + ChunkUtils::getOffset(mFlags));
+				else dna.oldPtr = 0;
+			}
+			else 
+				dna.oldPtr = 0;
+		}
+
+		if (mDataStart && dna.oldPtr) break;
 		tempBuffer++;
 	}
+	if (!dna.oldPtr || !dna.len)
+	{
+		printf("Failed to find DNA1+SDNA pair\n");
+		mFlags &= ~FD_OK;
+		return;
+	}
+
 
 	mFileDNA = new bDNA();
-	mFileDNA->init(blenderData+sdnaPos, mFileLen-sdnaPos, (mFlags & FD_ENDIAN_SWAP)!=0);
+	mFileDNA->init((char*)dna.oldPtr, dna.len, (mFlags & FD_ENDIAN_SWAP)!=0);
 
 	if (mVersion==276)
 	{
@@ -348,9 +367,10 @@ char* bFile::readStruct(char *head, bChunkInd&  dataChunk)
 
 				numallocs++;
 				// numBlocks * length
-                int allocLen = curLen > oldLen ? curLen : oldLen;
+
+                int allocLen = (curLen);
     			char *dataAlloc = new char[(dataChunk.nr*allocLen)+1];
-				memset(dataAlloc, 0, (dataChunk.nr*allocLen)+1);
+				memset(dataAlloc, 0, (dataChunk.nr*allocLen));
 
 				// track allocated
 				addDataBlock(dataAlloc);
diff --git a/src/BulletMultiThreaded/btParallelConstraintSolver.cpp b/src/BulletMultiThreaded/btParallelConstraintSolver.cpp
index 8b83a1146..9f2767c71 100644
--- a/src/BulletMultiThreaded/btParallelConstraintSolver.cpp
+++ b/src/BulletMultiThreaded/btParallelConstraintSolver.cpp
@@ -99,11 +99,11 @@ void pfxSolveLinearConstraintRow(PfxConstraintRow &constraint,
 	float massInvB,const vmMatrix3 &inertiaInvB,const vmVector3 &rB)
 {
 	const vmVector3 normal(btReadVector3(constraint.mNormal));
-	float deltaImpulse = constraint.mRhs;
+	btScalar deltaImpulse = constraint.mRhs;
 	vmVector3 dVA = deltaLinearVelocityA + cross(deltaAngularVelocityA,rA);
 	vmVector3 dVB = deltaLinearVelocityB + cross(deltaAngularVelocityB,rB);
 	deltaImpulse -= constraint.mJacDiagInv * dot(normal,dVA-dVB);
-	float oldImpulse = constraint.mAccumImpulse;
+	btScalar oldImpulse = constraint.mAccumImpulse;
 	constraint.mAccumImpulse = btClamped(oldImpulse + deltaImpulse,constraint.mLowerLimit,constraint.mUpperLimit);
 	deltaImpulse = constraint.mAccumImpulse - oldImpulse;
 	deltaLinearVelocityA += deltaImpulse * massInvA * normal;
diff --git a/src/BulletMultiThreaded/btParallelConstraintSolver.h b/src/BulletMultiThreaded/btParallelConstraintSolver.h
index 2e8fbaf13..0ff820a4c 100644
--- a/src/BulletMultiThreaded/btParallelConstraintSolver.h
+++ b/src/BulletMultiThreaded/btParallelConstraintSolver.h
@@ -133,6 +133,31 @@ ATTRIBUTE_ALIGNED16(struct) PfxSolverBody {
 #include "SpuDispatch/BulletPE2ConstraintSolverSpursSupport.h"
 #endif
 
+static SIMD_FORCE_INLINE vmVector3 btReadVector3(const double* p)
+{
+	float tmp[3] = {p[0],p[1],p[2]};
+	vmVector3 v;
+	loadXYZ(v, tmp);
+	return v;
+}
+
+static SIMD_FORCE_INLINE vmQuat btReadQuat(const double* p)
+{
+	float tmp[4] = {p[0],p[1],p[2],p[4]};
+	vmQuat vq;
+	loadXYZW(vq, tmp);
+	return vq;
+}
+
+static SIMD_FORCE_INLINE void btStoreVector3(const vmVector3 &src, double* p)
+{
+	float tmp[3];
+	vmVector3 v = src;
+	storeXYZ(v, tmp);
+	p[0] = tmp[0];
+	p[1] = tmp[1];
+	p[2] = tmp[2];
+}
 
 
 static SIMD_FORCE_INLINE vmVector3 btReadVector3(const float* p)