Port fix from 2.0 to trunk. Problem discovered thanks to Jern-Kuan.

src/nvtt/cuda/CudaMath.h

@@ -194,14 +194,26 @@ inline __device__ __host__ float lengthSquared(float3 a)
 }
 
 
-
 // Use power method to find the first eigenvector.
 // http://www.miislita.com/information-retrieval-tutorial/matrix-tutorial-3-eigenvalues-eigenvectors.html
 inline __device__ __host__ float3 firstEigenVector( float matrix[6] )
 {
     // 8 iterations seems to be more than enough.
 
-	float3 v = make_float3(1.0f, 1.0f, 1.0f);
+    float3 row0 = make_float3(matrix[0], matrix[1], matrix[2]);
+    float3 row1 = make_float3(matrix[1], matrix[3], matrix[4]);
+    float3 row2 = make_float3(matrix[2], matrix[4], matrix[5]);
+
+    float r0 = dot(row0, row0);
+    float r1 = dot(row1, row1);
+    float r2 = dot(row2, row2);
+
+    float3 v;
+    if (r0 > r1 && r0 > r2) v = row0;
+    else if (r1 > r2) v = row1;
+    else v = row2;
+
+    //float3 v = make_float3(1.0f, 1.0f, 1.0f);
     for(int i = 0; i < 8; i++) {
         float x = v.x * matrix[0] + v.y * matrix[1] + v.z * matrix[2];
         float y = v.x * matrix[1] + v.y * matrix[3] + v.z * matrix[4];
@@ -215,6 +227,7 @@ inline __device__ __host__ float3 firstEigenVector( float matrix[6] )
     return v;
 }
 
+
 inline __device__ bool singleColor(const float3 * colors)
 {
 #if __DEVICE_EMULATION__
@@ -316,12 +329,23 @@ inline __device__ float3 bestFitLine(const float3 * colors, float3 color_sum, fl
     return firstEigenVector(covariance);
 }
 
+
 // @@ For 2D this may not be the most efficient method. It's a quadratic equation, right?
 inline __device__ __host__ float2 firstEigenVector2D( float matrix[3] )
 {
     // @@ 8 iterations is probably more than enough.
 
-	float2 v = make_float2(1.0f, 1.0f);
+    const float3 row0 = make_float2(matrix[0], matrix[1]);
+    const float3 row1 = make_float2(matrix[1], matrix[2]);
+
+    float r0 = lengthSquared(row0);
+    float r1 = lengthSquared(row1);
+
+    float2 v;
+    if (r0 > r1) v = row0;
+    v = row1;
+
+    //float2 v = make_float2(1.0f, 1.0f);
     for(int i = 0; i < 8; i++) {
         float x = v.x * matrix[0] + v.y * matrix[1];
         float y = v.x * matrix[1] + v.y * matrix[2];