3 files changed, 247 insertions, 477 deletions

diff --git a/source/Noise.cpp b/source/Noise.cpp
index 5261f967c..93a710970 100644
--- a/source/Noise.cpp
+++ b/source/Noise.cpp
@@ -17,20 +17,59 @@
 
 
 
-NOISE_DATATYPE CubicInterpolate(NOISE_DATATYPE a_A, NOISE_DATATYPE a_B, NOISE_DATATYPE a_C, NOISE_DATATYPE a_D, NOISE_DATATYPE a_Pct)
-{
-	NOISE_DATATYPE P = (a_D - a_C) - (a_A - a_B);
-	NOISE_DATATYPE Q = (a_A - a_B) - P;
-	NOISE_DATATYPE R = a_C - a_A;
-	NOISE_DATATYPE S = a_B;
+///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Globals:
 
-	return ((P * a_Pct + Q) * a_Pct + R) * a_Pct + S;
+void IntArrayLinearInterpolate2D(
+	int * a_Array, 
+	int a_SizeX, int a_SizeY,  // Dimensions of the array
+	int a_AnchorStepX, int a_AnchorStepY  // Distances between the anchor points in each direction
+)
+{
+	// First interpolate columns where the anchor points are:
+	int LastYCell = a_SizeY - a_AnchorStepY;
+	for (int y = 0; y < LastYCell; y += a_AnchorStepY)
+	{
+		int Idx = a_SizeX * y;
+		for (int x = 0; x < a_SizeX; x += a_AnchorStepX)
+		{
+			int StartValue = a_Array[Idx];
+			int EndValue   = a_Array[Idx + a_SizeX * a_AnchorStepY];
+			int Diff = EndValue - StartValue;
+			for (int CellY = 1; CellY < a_AnchorStepY; CellY++)
+			{
+				a_Array[Idx + a_SizeX * CellY] = StartValue + CellY * Diff / a_AnchorStepY;
+			}  // for CellY
+			Idx += a_AnchorStepX;
+		}  // for x
+	}  // for y
+	
+	// Now interpolate in rows, each row has values in the anchor columns
+	int LastXCell = a_SizeX - a_AnchorStepX;
+	for (int y = 0; y < a_SizeY; y++)
+	{
+		int Idx = a_SizeX * y;
+		for (int x = 0; x < LastXCell; x += a_AnchorStepX)
+		{
+			int StartValue = a_Array[Idx];
+			int EndValue   = a_Array[Idx + a_AnchorStepX];
+			int Diff = EndValue - StartValue;
+			for (int CellX = 1; CellX < a_AnchorStepX; CellX++)
+			{
+				a_Array[Idx + CellX] = StartValue + CellX * Diff / a_AnchorStepX;
+			}  // for CellY
+			Idx += a_AnchorStepX;
+		}
+	}
 }
 
 
 
 
 
+///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// cCubicCell2D:
+
 class cCubicCell2D
 {
 public:
@@ -105,14 +144,14 @@ void cCubicCell2D::Generate(
 	{
 		NOISE_DATATYPE Interp[4];
 		NOISE_DATATYPE FracY = m_FracY[y];
-		Interp[0] = CubicInterpolate((*m_WorkRnds)[0][0], (*m_WorkRnds)[0][1], (*m_WorkRnds)[0][2], (*m_WorkRnds)[0][3], FracY);
-		Interp[1] = CubicInterpolate((*m_WorkRnds)[1][0], (*m_WorkRnds)[1][1], (*m_WorkRnds)[1][2], (*m_WorkRnds)[1][3], FracY);
-		Interp[2] = CubicInterpolate((*m_WorkRnds)[2][0], (*m_WorkRnds)[2][1], (*m_WorkRnds)[2][2], (*m_WorkRnds)[2][3], FracY);
-		Interp[3] = CubicInterpolate((*m_WorkRnds)[3][0], (*m_WorkRnds)[3][1], (*m_WorkRnds)[3][2], (*m_WorkRnds)[3][3], FracY);
+		Interp[0] = cNoise::CubicInterpolate((*m_WorkRnds)[0][0], (*m_WorkRnds)[0][1], (*m_WorkRnds)[0][2], (*m_WorkRnds)[0][3], FracY);
+		Interp[1] = cNoise::CubicInterpolate((*m_WorkRnds)[1][0], (*m_WorkRnds)[1][1], (*m_WorkRnds)[1][2], (*m_WorkRnds)[1][3], FracY);
+		Interp[2] = cNoise::CubicInterpolate((*m_WorkRnds)[2][0], (*m_WorkRnds)[2][1], (*m_WorkRnds)[2][2], (*m_WorkRnds)[2][3], FracY);
+		Interp[3] = cNoise::CubicInterpolate((*m_WorkRnds)[3][0], (*m_WorkRnds)[3][1], (*m_WorkRnds)[3][2], (*m_WorkRnds)[3][3], FracY);
 		int idx = y * m_SizeX + a_FromX;
 		for (int x = a_FromX; x < a_ToX; x++)
 		{
-			m_Array[idx++] = CubicInterpolate(Interp[0], Interp[1], Interp[2], Interp[3], m_FracX[x]);
+			m_Array[idx++] = cNoise::CubicInterpolate(Interp[0], Interp[1], Interp[2], Interp[3], m_FracX[x]);
 		}  // for x
 	}  // for y
 }
@@ -189,291 +228,138 @@ cNoise::cNoise(unsigned int a_Seed) :
 
 
 
-#if NOISE_USE_SSE
-/****************
- * SSE Random value generator
- **/
-__m128 cNoise::SSE_IntNoise2D( int a_X1, int a_Y1, int a_X2, int a_Y2, int a_X3, int a_Y3, int a_X4, int a_Y4 ) const
+NOISE_DATATYPE cNoise::LinearNoise1D(NOISE_DATATYPE a_X) const
 {
-	const __m128i X4 = _mm_set_epi32(a_X4, a_X3, a_X2, a_X1);
-	const __m128i Y4 = _mm_set_epi32(a_Y4, a_Y3, a_Y2, a_Y1);
-
-	const __m128 One4 = _mm_set_ps1( 1.f );
-	const __m128i YScale4 = _mm_set1_epi32( 57 );
-
-	const __m128i i15731 = _mm_set1_epi32( 15731 );
-	const __m128i i789221 = _mm_set1_epi32( 789221 );
-	const __m128i i1376312589 = _mm_set1_epi32(1376312589);
-	const __m128i MaskValue4 = _mm_set1_epi32(0x7fffffff);
-	const __m128 f1073741824 = _mm_set_ps1( 1073741824.0f );
-
-	const __m128i Seed4 = _mm_mullo_epi32( _mm_mullo_epi32( _mm_set1_epi32( m_Seed ), YScale4 ), YScale4 );
-
-	const __m128i ScaledY4 = _mm_mullo_epi32( Y4, YScale4 );
-	const __m128i n4 = _mm_add_epi32( _mm_add_epi32( X4, ScaledY4 ), Seed4 );
-	const __m128i nn4 = _mm_slli_epi32( n4, 13 );
-	const __m128i nnn4 = _mm_xor_si128( nn4, n4 );
-
-	const __m128i StepA4 = _mm_mullo_epi32( nnn4, nnn4 );
-	const __m128i StepAA4 = _mm_add_epi32( _mm_mullo_epi32( StepA4, i15731 ), i789221 );
-	const __m128i StepB4 = _mm_add_epi32( _mm_mullo_epi32( nnn4, StepAA4 ), i1376312589 );
-	const __m128i StepC4 = _mm_and_si128( StepB4, MaskValue4 );
-	const __m128 StepD4 = _mm_div_ps( _mm_cvtepi32_ps( StepC4 ), f1073741824 );
-	const __m128 Result4 = _mm_sub_ps( One4, StepD4 );
-
-	return Result4;
-}
-#endif
-
-
-
-
-
-/***************
- * Interpolated (and 1 smoothed) noise in 1-dimension
- **/
-float cNoise::LinearNoise1D( float a_X ) const
-{
-	int BaseX = FAST_FLOOR( a_X );
-	float FracX = (a_X) - BaseX;
-	return LinearInterpolate( IntNoise( BaseX ), IntNoise( BaseX+1 ), FracX);
+	int BaseX = FAST_FLOOR(a_X);
+	NOISE_DATATYPE FracX = a_X - BaseX;
+	return LinearInterpolate(IntNoise1D(BaseX), IntNoise1D(BaseX + 1), FracX);
 }
 
 
 
 
 
-float cNoise::CosineNoise1D( float a_X ) const
+NOISE_DATATYPE cNoise::CosineNoise1D(NOISE_DATATYPE a_X) const
 {
-	int BaseX = FAST_FLOOR( a_X );
-	float FracX = (a_X) - BaseX;
-	return CosineInterpolate( IntNoise( BaseX ), IntNoise( BaseX+1 ), FracX);
+	int BaseX = FAST_FLOOR(a_X);
+	NOISE_DATATYPE FracX = a_X - BaseX;
+	return CosineInterpolate(IntNoise1D(BaseX), IntNoise1D(BaseX + 1), FracX);
 }
 
 
 
 
 
-float cNoise::CubicNoise1D( float a_X ) const
+NOISE_DATATYPE cNoise::CubicNoise1D(NOISE_DATATYPE a_X) const
 {
-	int BaseX = FAST_FLOOR( a_X );
-	float FracX = (a_X) - BaseX;
-	return CubicInterpolate( IntNoise( BaseX-1 ), IntNoise( BaseX ), IntNoise( BaseX+1 ), IntNoise( BaseX+2 ), FracX);
+	int BaseX = FAST_FLOOR(a_X);
+	NOISE_DATATYPE FracX = a_X - BaseX;
+	return CubicInterpolate(IntNoise1D(BaseX - 1), IntNoise1D(BaseX), IntNoise1D(BaseX + 1), IntNoise1D(BaseX + 2), FracX);
 }
 
 
 
 
 
-float cNoise::SmoothNoise1D( int a_X ) const
+NOISE_DATATYPE cNoise::SmoothNoise1D(int a_X) const
 {
-	return IntNoise(a_X)/2  +  IntNoise(a_X-1)/4  +  IntNoise(a_X+1)/4;
+	return IntNoise1D(a_X) / 2 + IntNoise1D(a_X - 1) / 4 + IntNoise1D(a_X + 1) / 4;
 }
 
 
 
 
 
-/******************
- * Interpolated (and 1 smoothed) noise in 2-dimensions
- **/
-float cNoise::LinearNoise2D( float a_X, float a_Y ) const
+NOISE_DATATYPE cNoise::CubicNoise2D(NOISE_DATATYPE a_X, NOISE_DATATYPE a_Y) const
 {
-	const int	BaseX = FAST_FLOOR( a_X );
-	const int	BaseY = FAST_FLOOR( a_Y );
-
-	const float	tl = IntNoise2D( BaseX,		BaseY );
-	const float	tr = IntNoise2D( BaseX+1,	BaseY );
-	const float	bl = IntNoise2D( BaseX,		BaseY+1 );
-	const float	br = IntNoise2D( BaseX+1,		BaseY+1 );
-
-	const float FracX = (a_X) - BaseX;
-	const float interp1 = LinearInterpolate( tl, tr, FracX );
-	const float interp2 = LinearInterpolate( bl, br, FracX );
-
-	const float	FracY = (a_Y) - BaseY;
-	return LinearInterpolate( interp1, interp2, FracY );
-}
-
-
-
-
-
-float cNoise::CosineNoise2D( float a_X, float a_Y ) const
-{
-	const int	BaseX = FAST_FLOOR( a_X );
-	const int	BaseY = FAST_FLOOR( a_Y );
-
-	const float	tl = IntNoise2D( BaseX,		BaseY );
-	const float	tr = IntNoise2D( BaseX+1,	BaseY );
-	const float	bl = IntNoise2D( BaseX,		BaseY+1 );
-	const float	br = IntNoise2D( BaseX+1,		BaseY+1 );
-
-	const float FracX = (a_X) - BaseX;
-	const float interp1 = CosineInterpolate( tl, tr, FracX );
-	const float interp2 = CosineInterpolate( bl, br, FracX );
-
-	const float	FracY = (a_Y) - BaseY;
-	return CosineInterpolate( interp1, interp2, FracY );
-}
-
-
-
-
-
-float cNoise::CubicNoise2D( float a_X, float a_Y ) const
-{
-	const int	BaseX = FAST_FLOOR( a_X );
-	const int	BaseY = FAST_FLOOR( a_Y );
+	const int	BaseX = FAST_FLOOR(a_X);
+	const int	BaseY = FAST_FLOOR(a_Y);
 	
-	const float points[4][4] =
+	const NOISE_DATATYPE points[4][4] =
 	{
-		IntNoise2D( BaseX-1, BaseY-1 ), IntNoise2D( BaseX, BaseY-1 ), IntNoise2D( BaseX+1, BaseY-1 ), IntNoise2D( BaseX+2, BaseY-1 ),
-		IntNoise2D( BaseX-1, BaseY ),   IntNoise2D( BaseX, BaseY ),   IntNoise2D( BaseX+1, BaseY ),   IntNoise2D( BaseX+2, BaseY ),
-		IntNoise2D( BaseX-1, BaseY+1 ), IntNoise2D( BaseX, BaseY+1 ), IntNoise2D( BaseX+1, BaseY+1 ), IntNoise2D( BaseX+2, BaseY+1 ),
-		IntNoise2D( BaseX-1, BaseY+2 ), IntNoise2D( BaseX, BaseY+2 ), IntNoise2D( BaseX+1, BaseY+2 ), IntNoise2D( BaseX+2, BaseY+2 ),
-	};
-
-	const float	FracX = (a_X) - BaseX;
-	const float interp1 = CubicInterpolate( points[0][0], points[0][1], points[0][2], points[0][3], FracX );
-	const float interp2 = CubicInterpolate( points[1][0], points[1][1], points[1][2], points[1][3], FracX );
-	const float interp3 = CubicInterpolate( points[2][0], points[2][1], points[2][2], points[2][3], FracX );
-	const float interp4 = CubicInterpolate( points[3][0], points[3][1], points[3][2], points[3][3], FracX );
-
-
-	const float	FracY = (a_Y) - BaseY;
-	return CubicInterpolate( interp1, interp2, interp3, interp4, FracY );
-}
-
-
-
-
-
-#if NOISE_USE_SSE
-float cNoise::SSE_CubicNoise2D( float a_X, float a_Y ) const
-{
-	const int	BaseX = FAST_FLOOR( a_X );
-	const int	BaseY = FAST_FLOOR( a_Y );
-
-	__m128 points4[4] = {
-		SSE_IntNoise2D( BaseX-1, BaseY-1,	BaseX-1, BaseY, BaseX-1, BaseY+1,	BaseX-1, BaseY+2 ),
-		SSE_IntNoise2D( BaseX,	 BaseY-1,	BaseX, BaseY,	BaseX, BaseY+1,		BaseX, BaseY+2 ),
-		SSE_IntNoise2D( BaseX+1, BaseY-1,	BaseX+1, BaseY, BaseX+1, BaseY+1,	BaseX+1, BaseY+2 ),
-		SSE_IntNoise2D( BaseX+2, BaseY-1,	BaseX+2, BaseY,	BaseX+2, BaseY+1,	BaseX+2, BaseY+2 ),
+		IntNoise2D(BaseX - 1, BaseY - 1), IntNoise2D(BaseX, BaseY - 1), IntNoise2D(BaseX + 1, BaseY - 1), IntNoise2D(BaseX + 2, BaseY - 1),
+		IntNoise2D(BaseX - 1, BaseY),     IntNoise2D(BaseX, BaseY),     IntNoise2D(BaseX + 1, BaseY),     IntNoise2D(BaseX + 2, BaseY),
+		IntNoise2D(BaseX - 1, BaseY + 1), IntNoise2D(BaseX, BaseY + 1), IntNoise2D(BaseX + 1, BaseY + 1), IntNoise2D(BaseX + 2, BaseY + 1),
+		IntNoise2D(BaseX - 1, BaseY + 2), IntNoise2D(BaseX, BaseY + 2), IntNoise2D(BaseX + 1, BaseY + 2), IntNoise2D(BaseX + 2, BaseY + 2),
 	};
 
-	const float	FracX = (a_X) - BaseX;
-	union { __m128 p4; float p[4]; } 
-	AllInterp = { CubicInterpolate4( points4[0], points4[1], points4[2], points4[3], FracX ) };
-
-	const float	FracY = (a_Y) - BaseY;
-	return CubicInterpolate( AllInterp.p[0], AllInterp.p[1], AllInterp.p[2], AllInterp.p[3], FracY );
-}
-#endif
-
-
-
+	const NOISE_DATATYPE FracX = a_X - BaseX;
+	const NOISE_DATATYPE interp1 = CubicInterpolate(points[0][0], points[0][1], points[0][2], points[0][3], FracX);
+	const NOISE_DATATYPE interp2 = CubicInterpolate(points[1][0], points[1][1], points[1][2], points[1][3], FracX);
+	const NOISE_DATATYPE interp3 = CubicInterpolate(points[2][0], points[2][1], points[2][2], points[2][3], FracX);
+	const NOISE_DATATYPE interp4 = CubicInterpolate(points[3][0], points[3][1], points[3][2], points[3][3], FracX);
 
 
-/******************
- * Interpolated (and 1 smoothed) noise in 3-dimensions
- **/
-float cNoise::CosineNoise3D( float a_X, float a_Y, float a_Z ) const
-{
-	const int	BaseX = FAST_FLOOR( a_X );
-	const int	BaseY = FAST_FLOOR( a_Y );
-	const int	BaseZ = FAST_FLOOR( a_Z );
-
-	const float	ftl = IntNoise3D( BaseX,	BaseY,		BaseZ );
-	const float	ftr = IntNoise3D( BaseX+1,	BaseY,		BaseZ  );
-	const float	fbl = IntNoise3D( BaseX,	BaseY+1,	BaseZ  );
-	const float	fbr = IntNoise3D( BaseX+1,	BaseY+1,	BaseZ  );
-
-	const float	btl = IntNoise3D( BaseX,	BaseY,		BaseZ+1 );
-	const float	btr = IntNoise3D( BaseX+1,	BaseY,		BaseZ+1 );
-	const float	bbl = IntNoise3D( BaseX,	BaseY+1,	BaseZ+1 );
-	const float	bbr = IntNoise3D( BaseX+1,	BaseY+1,	BaseZ+1 );
-
-	const float FracX = (a_X) - BaseX;
-	const float finterp1 = CosineInterpolate( ftl, ftr, FracX );
-	const float finterp2 = CosineInterpolate( fbl, fbr, FracX );
-	const float binterp1 = CosineInterpolate( btl, btr, FracX );
-	const float binterp2 = CosineInterpolate( bbl, bbr, FracX );
-
-	const float	FracY = (a_Y) - BaseY;
-	const float interp1 = CosineInterpolate( finterp1, finterp2, FracY );
-	const float interp2 = CosineInterpolate( binterp1, binterp2, FracY );
-
-	const float	FracZ = (a_Z) - BaseZ;
-	return CosineInterpolate( interp1, interp2, FracZ );
+	const NOISE_DATATYPE FracY = a_Y - BaseY;
+	return CubicInterpolate(interp1, interp2, interp3, interp4, FracY);
 }
 
 
 
 
 
-float cNoise::CubicNoise3D( float a_X, float a_Y, float a_Z ) const
+NOISE_DATATYPE cNoise::CubicNoise3D(NOISE_DATATYPE a_X, NOISE_DATATYPE a_Y, NOISE_DATATYPE a_Z) const
 {
-	const int	BaseX = FAST_FLOOR( a_X );
-	const int	BaseY = FAST_FLOOR( a_Y );
-	const int	BaseZ = FAST_FLOOR( a_Z );
+	const int	BaseX = FAST_FLOOR(a_X);
+	const int	BaseY = FAST_FLOOR(a_Y);
+	const int	BaseZ = FAST_FLOOR(a_Z);
 	
-	const float points1[4][4] = { 
-		IntNoise3D( BaseX-1, BaseY-1, BaseZ-1 ), IntNoise3D( BaseX, BaseY-1, BaseZ-1 ),	IntNoise3D( BaseX+1, BaseY-1, BaseZ-1 ), IntNoise3D( BaseX+2, BaseY-1, BaseZ-1 ),
-		IntNoise3D( BaseX-1, BaseY,	  BaseZ-1 ), IntNoise3D( BaseX, BaseY,   BaseZ-1 ),	IntNoise3D( BaseX+1, BaseY,   BaseZ-1 ), IntNoise3D( BaseX+2, BaseY,   BaseZ-1 ),
-		IntNoise3D( BaseX-1, BaseY+1, BaseZ-1 ), IntNoise3D( BaseX, BaseY+1, BaseZ-1 ),	IntNoise3D( BaseX+1, BaseY+1, BaseZ-1 ), IntNoise3D( BaseX+2, BaseY+1, BaseZ-1 ),
-		IntNoise3D( BaseX-1, BaseY+2, BaseZ-1 ), IntNoise3D( BaseX, BaseY+2, BaseZ-1 ),	IntNoise3D( BaseX+1, BaseY+2, BaseZ-1 ), IntNoise3D( BaseX+2, BaseY+2, BaseZ-1 ),
+	const NOISE_DATATYPE points1[4][4] = { 
+		IntNoise3D(BaseX - 1, BaseY - 1, BaseZ - 1), IntNoise3D(BaseX, BaseY - 1, BaseZ - 1), IntNoise3D(BaseX + 1, BaseY - 1, BaseZ - 1), IntNoise3D(BaseX + 2, BaseY - 1, BaseZ - 1),
+		IntNoise3D(BaseX - 1, BaseY,     BaseZ - 1), IntNoise3D(BaseX, BaseY,     BaseZ - 1), IntNoise3D(BaseX + 1, BaseY,     BaseZ - 1), IntNoise3D(BaseX + 2, BaseY,     BaseZ - 1),
+		IntNoise3D(BaseX - 1, BaseY + 1, BaseZ - 1), IntNoise3D(BaseX, BaseY + 1, BaseZ - 1), IntNoise3D(BaseX + 1, BaseY + 1, BaseZ - 1), IntNoise3D(BaseX + 2, BaseY + 1, BaseZ - 1),
+		IntNoise3D(BaseX - 1, BaseY + 2, BaseZ - 1), IntNoise3D(BaseX, BaseY + 2, BaseZ - 1), IntNoise3D(BaseX + 1, BaseY + 2, BaseZ - 1), IntNoise3D(BaseX + 2, BaseY + 2, BaseZ - 1),
 	};
 
-	const float	FracX = (a_X) - BaseX;
-	const float x1interp1 = CubicInterpolate( points1[0][0], points1[0][1], points1[0][2], points1[0][3], FracX );
-	const float x1interp2 = CubicInterpolate( points1[1][0], points1[1][1], points1[1][2], points1[1][3], FracX );
-	const float x1interp3 = CubicInterpolate( points1[2][0], points1[2][1], points1[2][2], points1[2][3], FracX );
-	const float x1interp4 = CubicInterpolate( points1[3][0], points1[3][1], points1[3][2], points1[3][3], FracX );
+	const NOISE_DATATYPE	FracX = (a_X) - BaseX;
+	const NOISE_DATATYPE x1interp1 = CubicInterpolate( points1[0][0], points1[0][1], points1[0][2], points1[0][3], FracX );
+	const NOISE_DATATYPE x1interp2 = CubicInterpolate( points1[1][0], points1[1][1], points1[1][2], points1[1][3], FracX );
+	const NOISE_DATATYPE x1interp3 = CubicInterpolate( points1[2][0], points1[2][1], points1[2][2], points1[2][3], FracX );
+	const NOISE_DATATYPE x1interp4 = CubicInterpolate( points1[3][0], points1[3][1], points1[3][2], points1[3][3], FracX );
 
-	const float points2[4][4] = { 
+	const NOISE_DATATYPE points2[4][4] = { 
 		IntNoise3D( BaseX-1, BaseY-1, BaseZ ), IntNoise3D( BaseX, BaseY-1, BaseZ ),	IntNoise3D( BaseX+1, BaseY-1, BaseZ ), IntNoise3D( BaseX+2, BaseY-1, BaseZ ),
 		IntNoise3D( BaseX-1, BaseY,	  BaseZ ), IntNoise3D( BaseX, BaseY,   BaseZ ),	IntNoise3D( BaseX+1, BaseY,   BaseZ ), IntNoise3D( BaseX+2, BaseY,   BaseZ ),
 		IntNoise3D( BaseX-1, BaseY+1, BaseZ ), IntNoise3D( BaseX, BaseY+1, BaseZ ),	IntNoise3D( BaseX+1, BaseY+1, BaseZ ), IntNoise3D( BaseX+2, BaseY+1, BaseZ ),
 		IntNoise3D( BaseX-1, BaseY+2, BaseZ ), IntNoise3D( BaseX, BaseY+2, BaseZ ),	IntNoise3D( BaseX+1, BaseY+2, BaseZ ), IntNoise3D( BaseX+2, BaseY+2, BaseZ ),
 	};
 
-	const float x2interp1 = CubicInterpolate( points2[0][0], points2[0][1], points2[0][2], points2[0][3], FracX );
-	const float x2interp2 = CubicInterpolate( points2[1][0], points2[1][1], points2[1][2], points2[1][3], FracX );
-	const float x2interp3 = CubicInterpolate( points2[2][0], points2[2][1], points2[2][2], points2[2][3], FracX );
-	const float x2interp4 = CubicInterpolate( points2[3][0], points2[3][1], points2[3][2], points2[3][3], FracX );
+	const NOISE_DATATYPE x2interp1 = CubicInterpolate( points2[0][0], points2[0][1], points2[0][2], points2[0][3], FracX );
+	const NOISE_DATATYPE x2interp2 = CubicInterpolate( points2[1][0], points2[1][1], points2[1][2], points2[1][3], FracX );
+	const NOISE_DATATYPE x2interp3 = CubicInterpolate( points2[2][0], points2[2][1], points2[2][2], points2[2][3], FracX );
+	const NOISE_DATATYPE x2interp4 = CubicInterpolate( points2[3][0], points2[3][1], points2[3][2], points2[3][3], FracX );
 
-	const float points3[4][4] = { 
+	const NOISE_DATATYPE points3[4][4] = { 
 		IntNoise3D( BaseX-1, BaseY-1, BaseZ+1 ), IntNoise3D( BaseX, BaseY-1, BaseZ+1 ),	IntNoise3D( BaseX+1, BaseY-1, BaseZ+1 ), IntNoise3D( BaseX+2, BaseY-1, BaseZ+1 ),
 		IntNoise3D( BaseX-1, BaseY,	  BaseZ+1 ), IntNoise3D( BaseX, BaseY,   BaseZ+1 ),	IntNoise3D( BaseX+1, BaseY,   BaseZ+1 ), IntNoise3D( BaseX+2, BaseY,   BaseZ+1 ),
 		IntNoise3D( BaseX-1, BaseY+1, BaseZ+1 ), IntNoise3D( BaseX, BaseY+1, BaseZ+1 ),	IntNoise3D( BaseX+1, BaseY+1, BaseZ+1 ), IntNoise3D( BaseX+2, BaseY+1, BaseZ+1 ),
 		IntNoise3D( BaseX-1, BaseY+2, BaseZ+1 ), IntNoise3D( BaseX, BaseY+2, BaseZ+1 ),	IntNoise3D( BaseX+1, BaseY+2, BaseZ+1 ), IntNoise3D( BaseX+2, BaseY+2, BaseZ+1 ),
 	};
 
-	const float x3interp1 = CubicInterpolate( points3[0][0], points3[0][1], points3[0][2], points3[0][3], FracX );
-	const float x3interp2 = CubicInterpolate( points3[1][0], points3[1][1], points3[1][2], points3[1][3], FracX );
-	const float x3interp3 = CubicInterpolate( points3[2][0], points3[2][1], points3[2][2], points3[2][3], FracX );
-	const float x3interp4 = CubicInterpolate( points3[3][0], points3[3][1], points3[3][2], points3[3][3], FracX );
+	const NOISE_DATATYPE x3interp1 = CubicInterpolate( points3[0][0], points3[0][1], points3[0][2], points3[0][3], FracX );
+	const NOISE_DATATYPE x3interp2 = CubicInterpolate( points3[1][0], points3[1][1], points3[1][2], points3[1][3], FracX );
+	const NOISE_DATATYPE x3interp3 = CubicInterpolate( points3[2][0], points3[2][1], points3[2][2], points3[2][3], FracX );
+	const NOISE_DATATYPE x3interp4 = CubicInterpolate( points3[3][0], points3[3][1], points3[3][2], points3[3][3], FracX );
 
-	const float points4[4][4] = { 
+	const NOISE_DATATYPE points4[4][4] = { 
 		IntNoise3D( BaseX-1, BaseY-1, BaseZ+2 ), IntNoise3D( BaseX, BaseY-1, BaseZ+2 ),	IntNoise3D( BaseX+1, BaseY-1, BaseZ+2 ), IntNoise3D( BaseX+2, BaseY-1, BaseZ+2 ),
 		IntNoise3D( BaseX-1, BaseY,	  BaseZ+2 ), IntNoise3D( BaseX, BaseY,   BaseZ+2 ),	IntNoise3D( BaseX+1, BaseY,   BaseZ+2 ), IntNoise3D( BaseX+2, BaseY,   BaseZ+2 ),
 		IntNoise3D( BaseX-1, BaseY+1, BaseZ+2 ), IntNoise3D( BaseX, BaseY+1, BaseZ+2 ),	IntNoise3D( BaseX+1, BaseY+1, BaseZ+2 ), IntNoise3D( BaseX+2, BaseY+1, BaseZ+2 ),
 		IntNoise3D( BaseX-1, BaseY+2, BaseZ+2 ), IntNoise3D( BaseX, BaseY+2, BaseZ+2 ),	IntNoise3D( BaseX+1, BaseY+2, BaseZ+2 ), IntNoise3D( BaseX+2, BaseY+2, BaseZ+2 ),
 	};
 
-	const float x4interp1 = CubicInterpolate( points4[0][0], points4[0][1], points4[0][2], points4[0][3], FracX );
-	const float x4interp2 = CubicInterpolate( points4[1][0], points4[1][1], points4[1][2], points4[1][3], FracX );
-	const float x4interp3 = CubicInterpolate( points4[2][0], points4[2][1], points4[2][2], points4[2][3], FracX );
-	const float x4interp4 = CubicInterpolate( points4[3][0], points4[3][1], points4[3][2], points4[3][3], FracX );
+	const NOISE_DATATYPE x4interp1 = CubicInterpolate( points4[0][0], points4[0][1], points4[0][2], points4[0][3], FracX );
+	const NOISE_DATATYPE x4interp2 = CubicInterpolate( points4[1][0], points4[1][1], points4[1][2], points4[1][3], FracX );
+	const NOISE_DATATYPE x4interp3 = CubicInterpolate( points4[2][0], points4[2][1], points4[2][2], points4[2][3], FracX );
+	const NOISE_DATATYPE x4interp4 = CubicInterpolate( points4[3][0], points4[3][1], points4[3][2], points4[3][3], FracX );
 
-	const float	FracY = (a_Y) - BaseY;
-	const float yinterp1 = CubicInterpolate( x1interp1, x1interp2, x1interp3, x1interp4, FracY );
-	const float yinterp2 = CubicInterpolate( x2interp1, x2interp2, x2interp3, x2interp4, FracY );
-	const float yinterp3 = CubicInterpolate( x3interp1, x3interp2, x3interp3, x3interp4, FracY );
-	const float yinterp4 = CubicInterpolate( x4interp1, x4interp2, x4interp3, x4interp4, FracY );
+	const NOISE_DATATYPE	FracY = (a_Y) - BaseY;
+	const NOISE_DATATYPE yinterp1 = CubicInterpolate( x1interp1, x1interp2, x1interp3, x1interp4, FracY );
+	const NOISE_DATATYPE yinterp2 = CubicInterpolate( x2interp1, x2interp2, x2interp3, x2interp4, FracY );
+	const NOISE_DATATYPE yinterp3 = CubicInterpolate( x3interp1, x3interp2, x3interp3, x3interp4, FracY );
+	const NOISE_DATATYPE yinterp4 = CubicInterpolate( x4interp1, x4interp2, x4interp3, x4interp4, FracY );
 
-	const float	FracZ = (a_Z) - BaseZ;
+	const NOISE_DATATYPE	FracZ = (a_Z) - BaseZ;
 	return CubicInterpolate( yinterp1, yinterp2, yinterp3, yinterp4, FracZ );
 }
 
@@ -481,84 +367,6 @@ float cNoise::CubicNoise3D( float a_X, float a_Y, float a_Z ) const
 
 
 
-/******************
- * Private
- **/
-
-#if NOISE_USE_SSE
-__m128 cNoise::CubicInterpolate4( const __m128 & a_A, const __m128 & a_B, const __m128 & a_C, const __m128 & a_D, float a_Pct ) const
-{
-	const __m128 P = _mm_sub_ps( _mm_sub_ps( a_D, a_C ), _mm_sub_ps( a_A, a_B ) );
-	const __m128 Q = _mm_sub_ps( _mm_sub_ps( a_A, a_B ), P );
-	const __m128 R = _mm_sub_ps( a_C, a_A );
-
-	const __m128 Pct = _mm_set_ps1( a_Pct );
-	const __m128 Pct2 = _mm_mul_ps( Pct, Pct );
-	const __m128 Pct3 = _mm_mul_ps( Pct2, Pct );
-
-	return _mm_add_ps( _mm_add_ps( _mm_add_ps( _mm_mul_ps(P, Pct3), _mm_mul_ps( Q, Pct2 ) ), _mm_mul_ps( R, Pct ) ), a_B );
-}
-#endif
-
-
-
-
-
-void IntArrayLinearInterpolate2D(
-	int * a_Array, 
-	int a_SizeX, int a_SizeY,  // Dimensions of the array
-	int a_AnchorStepX, int a_AnchorStepY  // Distances between the anchor points in each direction
-)
-{
-	// First interpolate columns where the anchor points are:
-	int LastYCell = a_SizeY - a_AnchorStepY;
-	for (int y = 0; y < LastYCell; y += a_AnchorStepY)
-	{
-		int Idx = a_SizeX * y;
-		for (int x = 0; x < a_SizeX; x += a_AnchorStepX)
-		{
-			int StartValue = a_Array[Idx];
-			int EndValue   = a_Array[Idx + a_SizeX * a_AnchorStepY];
-			int Diff = EndValue - StartValue;
-			for (int CellY = 1; CellY < a_AnchorStepY; CellY++)
-			{
-				a_Array[Idx + a_SizeX * CellY] = StartValue + CellY * Diff / a_AnchorStepY;
-			}  // for CellY
-			Idx += a_AnchorStepX;
-		}  // for x
-	}  // for y
-	
-	// Now interpolate in rows, each row has values in the anchor columns
-	int LastXCell = a_SizeX - a_AnchorStepX;
-	for (int y = 0; y < a_SizeY; y++)
-	{
-		int Idx = a_SizeX * y;
-		for (int x = 0; x < LastXCell; x += a_AnchorStepX)
-		{
-			int StartValue = a_Array[Idx];
-			int EndValue   = a_Array[Idx + a_AnchorStepX];
-			int Diff = EndValue - StartValue;
-			for (int CellX = 1; CellX < a_AnchorStepX; CellX++)
-			{
-				a_Array[Idx + CellX] = StartValue + CellX * Diff / a_AnchorStepX;
-			}  // for CellY
-			Idx += a_AnchorStepX;
-		}
-	}
-}
-
-
-
-
-
-#if NOISE_USE_INLINE
-	#include "Noise.inc"
-#endif
-
-
-
-
-
 ///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
 // cCubicNoise:
 
diff --git a/source/Noise.h b/source/Noise.h
index d1098d7f8..2829ac827 100644
--- a/source/Noise.h
+++ b/source/Noise.h
@@ -1,28 +1,21 @@
+
+// Noise.h
+
+// Declares the cNoise, cCubicNoise and cPerlinNoise classes for generating noise
+
 #pragma once
 
 // Some settings
-#define NOISE_USE_INLINE 1
-#define NOISE_USE_SSE 0
-
 #define NOISE_DATATYPE float
 
 
 
 
 
-// Do not touch
-#if NOISE_USE_INLINE
-	#ifdef _MSC_VER
-		#define __NOISE_INLINE__ __forceinline
-	#else
-		#define __NOISE_INLINE__ inline
-	#endif  // _MSC_VER
+#ifdef _MSC_VER
+	#define INLINE __forceinline
 #else
-	#define __NOISE_INLINE__
-#endif
-
-#if NOISE_USE_SSE
-	#include <emmintrin.h>
+	#define INLINE inline
 #endif
 
 
@@ -32,48 +25,36 @@
 class cNoise
 {
 public:
-	cNoise( unsigned int a_Seed );
+	cNoise(unsigned int a_Seed);
 
-#if NOISE_USE_SSE
-	__m128 SSE_IntNoise2D( int a_X1, int a_Y1, int a_X2, int a_Y2, int a_X3, int a_Y3, int a_X4, int a_Y4 ) const;
-#endif
-
-	__NOISE_INLINE__ float IntNoise( int a_X ) const;
-	__NOISE_INLINE__ float IntNoise2D( int a_X, int a_Y ) const;
-	__NOISE_INLINE__ float IntNoise3D( int a_X, int a_Y, int a_Z ) const;
+	// The following functions, if not marked INLINE, are about 20 % slower
+	INLINE NOISE_DATATYPE IntNoise1D(int a_X) const;
+	INLINE NOISE_DATATYPE IntNoise2D(int a_X, int a_Y) const;
+	INLINE NOISE_DATATYPE IntNoise3D(int a_X, int a_Y, int a_Z) const;
 
 	// Note: These functions have a mod8-irregular chance - each of the mod8 remainders has different chance of occurrence. Divide by 8 to rectify.
-	__NOISE_INLINE__ int IntNoise1DInt( int a_X ) const;
-	__NOISE_INLINE__ int IntNoise2DInt( int a_X, int a_Y ) const;
-	__NOISE_INLINE__ int IntNoise3DInt( int a_X, int a_Y, int a_Z ) const;
+	INLINE int IntNoise1DInt(int a_X) const;
+	INLINE int IntNoise2DInt(int a_X, int a_Y) const;
+	INLINE int IntNoise3DInt(int a_X, int a_Y, int a_Z) const;
 
-	float LinearNoise1D( float a_X ) const;
-	float CosineNoise1D( float a_X ) const;
-	float CubicNoise1D( float a_X ) const;
-	float SmoothNoise1D( int a_X ) const;
+	NOISE_DATATYPE LinearNoise1D(NOISE_DATATYPE a_X) const;
+	NOISE_DATATYPE CosineNoise1D(NOISE_DATATYPE a_X) const;
+	NOISE_DATATYPE CubicNoise1D (NOISE_DATATYPE a_X) const;
+	NOISE_DATATYPE SmoothNoise1D(int a_X) const;
 
-	float LinearNoise2D( float a_X, float a_Y ) const;
-	float CosineNoise2D( float a_X, float a_Y ) const;
-	float CubicNoise2D( float a_X, float a_Y ) const;
-	float SSE_CubicNoise2D( float a_X, float a_Y ) const;
+	NOISE_DATATYPE CubicNoise2D (NOISE_DATATYPE a_X, NOISE_DATATYPE a_Y) const;
 
-	float CosineNoise3D( float a_X, float a_Y, float a_Z ) const;
-	float CubicNoise3D( float a_X, float a_Y, float a_Z ) const;
+	NOISE_DATATYPE CubicNoise3D (NOISE_DATATYPE a_X, NOISE_DATATYPE a_Y, NOISE_DATATYPE a_Z) const;
 
-	void SetSeed( unsigned int a_Seed ) { m_Seed = a_Seed; }
+	void SetSeed(unsigned int a_Seed) { m_Seed = a_Seed; }
 
-	__NOISE_INLINE__ static float CubicInterpolate (float a_A, float a_B, float a_C, float a_D, float a_Pct);
-	__NOISE_INLINE__ static float CosineInterpolate(float a_A, float a_B, float a_Pct);
-	__NOISE_INLINE__ static float LinearInterpolate(float a_A, float a_B, float a_Pct);
+	INLINE static NOISE_DATATYPE CubicInterpolate (NOISE_DATATYPE a_A, NOISE_DATATYPE a_B, NOISE_DATATYPE a_C, NOISE_DATATYPE a_D, NOISE_DATATYPE a_Pct);
+	INLINE static NOISE_DATATYPE CosineInterpolate(NOISE_DATATYPE a_A, NOISE_DATATYPE a_B, NOISE_DATATYPE a_Pct);
+	INLINE static NOISE_DATATYPE LinearInterpolate(NOISE_DATATYPE a_A, NOISE_DATATYPE a_B, NOISE_DATATYPE a_Pct);
 
 private:
-
-#if NOISE_USE_SSE
-	__m128 CubicInterpolate4( const __m128 & a_A, const __m128 & a_B, const __m128 & a_C, const __m128 & a_D, float a_Pct ) const;
-#endif
-
 	unsigned int m_Seed;
-};
+} ;
 
 
 
@@ -241,3 +222,108 @@ public:
 
 
 
+
+///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Inline function definitions:
+// These need to be in the header, otherwise linker error occur in MSVC
+
+NOISE_DATATYPE cNoise::IntNoise1D(int a_X) const
+{
+	int x = ((a_X * m_Seed) << 13) ^ a_X;
+	return (1 - (NOISE_DATATYPE)((x * (x * x * 15731 + 789221) + 1376312589) & 0x7fffffff) / 1073741824);
+	// returns a float number in the range of [-1, 1]
+}
+
+
+
+
+
+NOISE_DATATYPE cNoise::IntNoise2D(int a_X, int a_Y) const
+{
+	int n = a_X + a_Y * 57 + m_Seed * 57 * 57;
+	n = (n << 13) ^ n;
+	return (1 - (NOISE_DATATYPE)((n * (n * n * 15731 + 789221) + 1376312589) & 0x7fffffff) / 1073741824);
+	// returns a float number in the range of [-1, 1]
+}
+
+
+
+
+
+NOISE_DATATYPE cNoise::IntNoise3D(int a_X, int a_Y, int a_Z) const
+{
+	int n = a_X + a_Y * 57 + a_Z * 57 * 57 + m_Seed * 57 * 57 * 57;
+	n = (n << 13) ^ n;
+	return ((NOISE_DATATYPE)1 - (NOISE_DATATYPE)((n * (n * n * 15731 + 789221) + 1376312589) & 0x7fffffff) / 1073741824.0f);
+	// returns a float number in the range of [-1, 1]
+}
+
+
+
+
+
+int cNoise::IntNoise1DInt(int a_X) const
+{
+	int x = ((a_X * m_Seed) << 13) ^ a_X;
+	return ((x * (x * x * 15731 + 789221) + 1376312589) & 0x7fffffff);
+}
+
+
+
+
+
+int cNoise::IntNoise2DInt(int a_X, int a_Y) const
+{
+	int n = a_X + a_Y * 57 + m_Seed * 57 * 57;
+	n = (n << 13) ^ n;
+	return ((n * (n * n * 15731 + 789221) + 1376312589) & 0x7fffffff);
+}
+
+
+
+
+
+int cNoise::IntNoise3DInt(int a_X, int a_Y, int a_Z) const
+{
+	int n = a_X + a_Y * 57 + a_Z * 57 * 57 + m_Seed * 57 * 57 * 57;
+	n = (n << 13) ^ n;
+	return ((n * (n * n * 15731 + 789221) + 1376312589) & 0x7fffffff);
+}
+
+
+
+
+
+NOISE_DATATYPE cNoise::CubicInterpolate(NOISE_DATATYPE a_A, NOISE_DATATYPE a_B, NOISE_DATATYPE a_C, NOISE_DATATYPE a_D, NOISE_DATATYPE a_Pct)
+{
+	NOISE_DATATYPE P = (a_D - a_C) - (a_A - a_B);
+	NOISE_DATATYPE Q = (a_A - a_B) - P;
+	NOISE_DATATYPE R = a_C - a_A;
+	NOISE_DATATYPE S = a_B;
+
+	return ((P * a_Pct + Q) * a_Pct + R) * a_Pct + S;
+}
+
+
+
+
+
+NOISE_DATATYPE cNoise::CosineInterpolate(NOISE_DATATYPE a_A, NOISE_DATATYPE a_B, NOISE_DATATYPE a_Pct)
+{
+	const NOISE_DATATYPE ft = a_Pct * (NOISE_DATATYPE)3.1415927;
+	const NOISE_DATATYPE f = (1 - cos(ft)) * (NOISE_DATATYPE)0.5;
+	return  a_A * (1 - f) + a_B * f;
+}
+
+
+
+
+
+NOISE_DATATYPE cNoise::LinearInterpolate(NOISE_DATATYPE a_A, NOISE_DATATYPE a_B, NOISE_DATATYPE a_Pct)
+{
+	return  a_A * (1 - a_Pct) + a_B * a_Pct;
+}
+
+
+
+
diff --git a/source/Noise.inc b/source/Noise.inc
deleted file mode 100644
index 60fd2f394..000000000
--- a/source/Noise.inc
+++ /dev/null
@@ -1,124 +0,0 @@
-

-#ifndef __C_NOISE_INC__

-#define __C_NOISE_INC__

-

-#include <math.h>

-

-

-

-

-

-/****************

- * Random value generator

- **/

- 

-float cNoise::IntNoise( int a_X ) const

-{

-	int x = ((a_X*m_Seed)<<13) ^ a_X;

-	return ( 1.0f - ( (x * (x * x * 15731 + 789221) + 1376312589) & 0x7fffffff) / 1073741824.0f); 

-	// returns a float number in the range of [-1, 1]

-}

-

-

-

-

-

-float cNoise::IntNoise2D( int a_X, int a_Y ) const

-{

-	int n = a_X + a_Y * 57 + m_Seed*57*57;

-	n = (n<<13) ^ n;

-	return ( 1.0f - ( (n * (n * n * 15731 + 789221) + 1376312589) & 0x7fffffff) / 1073741824.0f);  

-	// returns a float number in the range of [-1, 1]

-}

-

-

-

-

-

-float cNoise::IntNoise3D( int a_X, int a_Y, int a_Z ) const

-{

-	int n = a_X + a_Y * 57 + a_Z * 57*57 + m_Seed*57*57*57;

-	n = (n<<13) ^ n;

-	return ( 1.0f - ( (n * (n * n * 15731 + 789221) + 1376312589) & 0x7fffffff) / 1073741824.0f);  

-	// returns a float number in the range of [-1, 1]

-}

-

-

-

-

-

-int cNoise::IntNoise1DInt( int a_X ) const

-{

-	int x = ((a_X*m_Seed)<<13) ^ a_X;

-	return ( (x * (x * x * 15731 + 789221) + 1376312589) & 0x7fffffff);

-}

-

-

-

-

-

-int cNoise::IntNoise2DInt( int a_X, int a_Y ) const

-{

-	int n = a_X + a_Y * 57 + m_Seed*57*57;

-	n = (n<<13) ^ n;

-	return ( (n * (n * n * 15731 + 789221) + 1376312589) & 0x7fffffff);

-}

-

-

-

-

-

-int cNoise::IntNoise3DInt( int a_X, int a_Y, int a_Z ) const

-{

-	int n = a_X + a_Y * 57 + a_Z * 57*57 + m_Seed*57*57*57;

-	n = (n<<13) ^ n;

-	return ( (n * (n * n * 15731 + 789221) + 1376312589) & 0x7fffffff);

-}

-

-

-

-

-

-/****************

- * Interpolation functions

- **/

- 

-float cNoise::CubicInterpolate( float a_A, float a_B, float a_C, float a_D, float a_Pct )

-{

-	float P = (a_D - a_C) - (a_A - a_B);

-	float Q = (a_A - a_B) - P;

-	float R = a_C - a_A;

-	float S = a_B;

-

-	return ((P * a_Pct + Q) * a_Pct + R) * a_Pct + S;

-}

-

-

-

-

-

-float cNoise::CosineInterpolate( float a_A, float a_B, float a_Pct )

-{

-	const float ft = a_Pct * 3.1415927f;

-	const float f = (1.f - cosf(ft)) * 0.5f;

-	return  a_A*(1-f) + a_B*f;

-}

-

-

-

-

-

-float cNoise::LinearInterpolate( float a_A, float a_B, float a_Pct )

-{

-	return  a_A*(1.f-a_Pct) + a_B*a_Pct;

-}

-

-

-

-

-

-#endif

-

-

-

-