Minor optimizations.

Fix DXT1a support.

src/nvtt/squish/colourfit.cpp

@@ -43,10 +43,11 @@ void ColourFit::Compress( void* block )
 	bool isDxt1 = ( ( m_flags & kDxt1 ) != 0 );
 	if( isDxt1 )
 	{
-		Compress4( block );
+		Compress3( block );
+	
 		if( !m_colours->IsTransparent() )
 		{		
-			Compress3( block );
+			Compress4( block );
 		}
 	}
 	else

src/nvtt/squish/colourset.cpp

@@ -43,16 +43,16 @@ ColourSet::ColourSet( u8 const* rgba, int flags, bool createMinimalSet/*=false*/
 	// create the minimal set
 	for( int i = 0; i < 16; ++i )
 	{
-		// check for transparent pixels when using dxt1
-		if( isDxt1 && rgba[4*i + 3] == 0 )
-		{
-			m_remap[i] = -1;
-			m_transparent = true;
-			if (createMinimalSet) continue;
-		}
-		
 		if (createMinimalSet)
 		{
+			// check for transparent pixels when using dxt1
+			if( isDxt1 && rgba[4*i + 3] == 0 )
+			{
+				m_remap[i] = -1;
+				m_transparent = true;
+				continue;
+			}
+
 			// loop over previous points for a match
 			for( int j = 0;; ++j )
 			{
@@ -100,6 +100,17 @@ ColourSet::ColourSet( u8 const* rgba, int flags, bool createMinimalSet/*=false*/
 		}
 		else
 		{
+			// check for transparent pixels when using dxt1
+			if( isDxt1 && rgba[4*i + 3] == 0 )
+			{
+				m_remap[i] = -1;
+				m_transparent = true;
+			}
+			else
+			{
+				m_remap[i] = m_count;
+			}
+
 			// normalise coordinates to [0,1]
 			float x = ( float )rgba[4*i + 2] / 255.0f;
 			float y = ( float )rgba[4*i + 1] / 255.0f;
@@ -111,7 +122,6 @@ ColourSet::ColourSet( u8 const* rgba, int flags, bool createMinimalSet/*=false*/
 			// add the point
 			m_points[m_count] = Vec3( x, y, z );
 			m_weights[m_count] = ( weightByAlpha ? w : 1.0f );
-			m_remap[i] = m_count;
 			
 			// advance
 			++m_count;

src/nvtt/squish/fastclusterfit.cpp

@@ -104,8 +104,8 @@ struct Precomp {
 	float factor;
 };
 
-static Precomp s_threeElement[153];
-static Precomp s_fourElement[969];
+static SQUISH_ALIGN_16 Precomp s_threeElement[153];
+static SQUISH_ALIGN_16 Precomp s_fourElement[969];
 
 void FastClusterFit::DoPrecomputation()
 {
@@ -229,11 +229,12 @@ void FastClusterFit::Compress3( void* block )
 		x1 = zero;
 		
 		for( int c1 = 0; c1 <= 16-c0; c1++)
-		{	
-			Vec4 const alpha2_sum = VEC4_CONST(s_threeElement[i].alpha2_sum);
-			Vec4 const beta2_sum = VEC4_CONST(s_threeElement[i].beta2_sum);
-			Vec4 const alphabeta_sum = VEC4_CONST(s_threeElement[i].alphabeta_sum);
-			Vec4 const factor = VEC4_CONST(s_threeElement[i].factor);
+		{
+			Vec4 const constants = Vec4((const float *)&s_threeElement[i]);
+			Vec4 const alpha2_sum = constants.SplatX();
+			Vec4 const beta2_sum = constants.SplatY();
+			Vec4 const alphabeta_sum = constants.SplatZ();
+			Vec4 const factor = constants.SplatW();
 			i++;
 			
 			Vec4 const alphax_sum = MultiplyAdd(half, x1, x0);
@@ -310,6 +311,8 @@ void FastClusterFit::Compress3( void* block )
 		for( int i = 0; i < 16; ++i )
 			ordered[m_order[i]] = bestindices[i];
 		
+		m_colours->RemapIndices( ordered, bestindices ); // Set alpha indices.
+
 		// save the block
 		WriteColourBlock3( beststart.GetVec3(), bestend.GetVec3(), ordered, block );
 		
@@ -347,10 +350,11 @@ void FastClusterFit::Compress4( void* block )
 			
 			for( int c2 = 0; c2 <= 16-c0-c1; c2++)
 			{
-				Vec4 const alpha2_sum = VEC4_CONST(s_fourElement[i].alpha2_sum);
-				Vec4 const beta2_sum = VEC4_CONST(s_fourElement[i].beta2_sum);
-				Vec4 const alphabeta_sum = VEC4_CONST(s_fourElement[i].alphabeta_sum);
-				Vec4 const factor = VEC4_CONST(s_fourElement[i].factor);
+				Vec4 const constants = Vec4((const float *)&s_fourElement[i]);
+				Vec4 const alpha2_sum = constants.SplatX();
+				Vec4 const beta2_sum = constants.SplatY();
+				Vec4 const alphabeta_sum = constants.SplatZ();
+				Vec4 const factor = constants.SplatW();
 				i++;
 				
 				Vec4 const alphax_sum = x0 + MultiplyAdd(x1, twothirds, x2 * onethird);

src/nvtt/squish/simd_sse.h

@@ -35,6 +35,12 @@
 #define SQUISH_SSE_SPLAT( a )										\
 	( ( a ) | ( ( a ) << 2 ) | ( ( a ) << 4 ) | ( ( a ) << 6 ) )
 
+#ifdef __GNUC__
+#	define SQUISH_ALIGN_16 __attribute__ ((__aligned__ (16)))
+#else
+#	define SQUISH_ALIGN_16 __declspec(align(16))
+#endif
+
 namespace squish {
 
 #define VEC4_CONST( X ) Vec4( _mm_set1_ps( X ) )
@@ -55,7 +61,12 @@ public:
 		m_v = arg.m_v;
 		return *this;
 	}
-	
+
+	Vec4( const float * v )
+	{
+		m_v = _mm_load_ps( v );
+	}
+
 	Vec4( float x, float y, float z, float w )
 	{
 		m_v = _mm_setr_ps( x, y, z, w );
@@ -63,11 +74,7 @@ public:
 	
 	Vec3 GetVec3() const
 	{
-#ifdef __GNUC__
-		__attribute__ ((__aligned__ (16))) float c[4];
-#else
-		__declspec(align(16)) float c[4];
-#endif
+		SQUISH_ALIGN_16 float c[4];
 		_mm_store_ps( c, m_v );
 		return Vec3( c[0], c[1], c[2] );
 	}

src/nvtt/squish/simd_ve.h

@@ -29,6 +29,8 @@
 #include <altivec.h>
 #undef bool
 
+#define SQUISH_ALIGN_16 __attribute__ ((__aligned__ (16)))
+
 namespace squish {
 
 #define VEC4_CONST( X ) Vec4( ( vector float )( X ) )

src/nvtt/squish/weightedclusterfit.cpp

@@ -231,8 +231,11 @@ void WeightedClusterFit::Compress3( void* block )
 		for( int i = 0; i < 16; ++i )
 			ordered[m_order[i]] = bestindices[i];
 		
+		m_colours->RemapIndices( ordered, bestindices ); // Set alpha indices.
+
+
 		// save the block
-		WriteColourBlock3( beststart.GetVec3(), bestend.GetVec3(), ordered, block );
+		WriteColourBlock3( beststart.GetVec3(), bestend.GetVec3(), bestindices, block );
 		
 		// save the error
 		m_besterror = besterror;