diff --git a/desmume/src/gfx3d.cpp b/desmume/src/gfx3d.cpp
index 7e80aab53..48384cfac 100644
--- a/desmume/src/gfx3d.cpp
+++ b/desmume/src/gfx3d.cpp
@@ -99,7 +99,13 @@ static float normalTable[1024];
 #define fix10_2float(v) (((float)((s32)(v))) / (float)(1<<9))
 
 // Matrix stack handling
-static CACHE_ALIGN MatrixStack	mtxStack[4];
+static CACHE_ALIGN MatrixStack	mtxStack[4] = {
+	MatrixStack(1), // Projection stack
+	MatrixStack(31), // Coordinate stack
+	MatrixStack(31), // Directional stack
+	MatrixStack(1), // Texture stack
+};
+
 static CACHE_ALIGN float		mtxCurrent [4][16];
 static CACHE_ALIGN float		mtxTemporal[16];
 static u32 mode = 0;
@@ -252,11 +258,6 @@ void gfx3d_reset()
 	listTwiddle = 1;
 	twiddleLists();
 
-	MatrixStackSetMaxSize(&mtxStack[0], 1);		// Projection stack
-	MatrixStackSetMaxSize(&mtxStack[1], 31);	// Coordinate stack
-	MatrixStackSetMaxSize(&mtxStack[2], 31);	// Directional stack
-	MatrixStackSetMaxSize(&mtxStack[3], 1);		// Texture stack
-
 	MatrixInit (mtxCurrent[0]);
 	MatrixInit (mtxCurrent[1]);
 	MatrixInit (mtxCurrent[2]);
@@ -910,6 +911,9 @@ void gfx3d_glNormal(unsigned long v)
 						normalTable[(v>>10)&1023],
 						normalTable[(v>>20)&1023]};
 
+	//use the current normal transform matrix
+	MatrixMultVec3x3 (mtxCurrent[2], normal);
+
 	if (texCoordinateTransform == 2)
 	{
 		last_s =(	(normal[0] *mtxCurrent[3][0] + normal[1] *mtxCurrent[3][4] +
@@ -918,8 +922,6 @@ void gfx3d_glNormal(unsigned long v)
 					 normal[2] *mtxCurrent[3][9]) + _t);
 	}
 
-	//use the current normal transform matrix
-	MatrixMultVec3x3 (mtxCurrent[2], normal);
 
 	//apply lighting model
 	{
diff --git a/desmume/src/matrix.cpp b/desmume/src/matrix.cpp
index de4c31c37..4d74a8581 100644
--- a/desmume/src/matrix.cpp
+++ b/desmume/src/matrix.cpp
@@ -26,6 +26,7 @@
 
 extern "C" {
 
+
 void MatrixInit  (float *matrix)
 {
 	memset (matrix, 0, sizeof(float)*16);
@@ -161,13 +162,6 @@ int MATRIXFASTCALL MatrixCompare (const float* matrixDST, const float* matrixSRC
 	return memcmp((void*)matrixDST, matrixSRC, sizeof(float)*16);
 }
 
-void MatrixStackInit (MatrixStack *stack)
-{
-	stack->matrix	= NULL;
-	stack->position	= 0;
-	stack->size		= 0;
-}
-
 void MatrixStackSetMaxSize (MatrixStack *stack, int size)
 {
 	int i;
@@ -188,6 +182,11 @@ void MatrixStackSetMaxSize (MatrixStack *stack, int size)
 }
 
 
+MatrixStack::MatrixStack(int size)
+{
+	MatrixStackSetMaxSize(this,size);
+}
+
 void MatrixStackSetStackPosition (MatrixStack *stack, int pos)
 {
 	stack->position += pos;
diff --git a/desmume/src/matrix.h b/desmume/src/matrix.h
index 38d37de16..bd8d729d1 100644
--- a/desmume/src/matrix.h
+++ b/desmume/src/matrix.h
@@ -25,12 +25,13 @@
 
 extern "C" {
 
-typedef struct MatrixStack
+struct MatrixStack
 {
+	MatrixStack(int size);
 	float	*matrix;
 	int		position;
 	int		size;
-} MatrixStack;
+};
 
 void	MatrixInit				(float *matrix);
 
diff --git a/desmume/src/saves.cpp b/desmume/src/saves.cpp
index 49c45497c..e22a0cc7c 100644
--- a/desmume/src/saves.cpp
+++ b/desmume/src/saves.cpp
@@ -163,6 +163,9 @@ SFORMAT SF_NDS[]={
 	{ 0 }
 };
 
+extern u32 DMASrc[2][4];
+extern u32 DMADst[2][4];
+
 SFORMAT SF_MMU[]={
 	{ "M7BI", 1, 0x4000,  MMU.ARM7_BIOS},
 	{ "M7ER", 1, 0x10000, MMU.ARM7_ERAM},
@@ -184,10 +187,13 @@ SFORMAT SF_MMU[]={
 	{ "MIME", 4, 2,       MMU.reg_IME},
 	{ "MIE_", 4, 2,       MMU.reg_IE},
 	{ "MIF_", 4, 2,       MMU.reg_IF},
+
 	{ "MDST", 4, 8,       MMU.DMAStartTime},
 	{ "MDCY", 4, 8,       MMU.DMACycle},
 	{ "MDCR", 4, 8,       MMU.DMACrt},
 	{ "MDMA", 4, 8,       MMU.DMAing},
+	{ "MDSR", 4, 8,       DMASrc},
+	{ "MDDS", 4, 8,       DMADst},
 
 	{ "MDV1", 4, 1,       &MMU.divRunning},
 	{ "MDV2", 8, 1,       &MMU.divResult},