Backout tangent assert experiment

2 files changed, 66 insertions, 139 deletions

diff --git a/indra/llmath/llvector4a.inl b/indra/llmath/llvector4a.inl
index 6860252a75..7c52ffef21 100755
--- a/indra/llmath/llvector4a.inl
+++ b/indra/llmath/llvector4a.inl
@@ -331,9 +331,6 @@ inline LLSimdScalar LLVector4a::dot4(const LLVector4a& b) const
 // Note that this does not consider zero length vectors!
 inline void LLVector4a::normalize3()
 {
-	// find out about bad math before it takes two man-days to track down
-	llassert(isFinite3() && !equals3(getZero()));
-
 	// lenSqrd = a dot a
 	LLVector4a lenSqrd; lenSqrd.setAllDot3( *this, *this );
 	// rsqrt = approximate reciprocal square (i.e., { ~1/len(a)^2, ~1/len(a)^2, ~1/len(a)^2, ~1/len(a)^2 }
@@ -382,9 +379,6 @@ inline void LLVector4a::normalize4()
 // Note that this does not consider zero length vectors!
 inline LLSimdScalar LLVector4a::normalize3withLength()
 {
-	// find out about bad math before it takes two man-days to track down
-	llassert(isFinite3() && !equals3(getZero()));
-
 	// lenSqrd = a dot a
 	LLVector4a lenSqrd; lenSqrd.setAllDot3( *this, *this );
 	// rsqrt = approximate reciprocal square (i.e., { ~1/len(a)^2, ~1/len(a)^2, ~1/len(a)^2, ~1/len(a)^2 }
@@ -415,27 +409,6 @@ inline void LLVector4a::normalize3fast()
 	mQ = _mm_mul_ps( mQ, approxRsqrt );
 }
 
-// Normalize this vector with respect to the x, y, and z components only. Accurate only to 10-12 bits of precision. W component is destroyed
-// Note that this does not consider zero length vectors!
-inline void LLVector4a::normalize3fast_checked(LLVector4a* default)
-{
-	// handle bogus inputs before NaNs are generated below
-	//
-	if (!isFinite3() || (dot3(*this).getF32() < F_APPROXIMATELY_ZERO))
-	{
-		if (default)
-			*this = *default;
-		else
-			set(0,1,0,1);
-
-		return;
-	}
-
-	LLVector4a lenSqrd; lenSqrd.setAllDot3( *this, *this );
-	const LLQuad approxRsqrt = _mm_rsqrt_ps(lenSqrd.mQ);
-	mQ = _mm_mul_ps( mQ, approxRsqrt );
-}
-
 // Return true if this vector is normalized with respect to x,y,z up to tolerance
 inline LLBool32 LLVector4a::isNormalized3( F32 tolerance ) const
 {
diff --git a/indra/llmath/llvolume.cpp b/indra/llmath/llvolume.cpp
index 15621c2625..1932272afb 100755
--- a/indra/llmath/llvolume.cpp
+++ b/indra/llmath/llvolume.cpp
@@ -4472,9 +4472,6 @@ void LLVolume::generateSilhouetteVertices(std::vector<LLVector3> &vertices,
 					continue; //skip degenerate face
 				}
 
-				LLVector4a default_norm;
-				default_norm.set(0,1,0,1);
-
 				//for each edge
 				for (S32 k = 0; k < 3; k++) {
 					S32 index = face.mEdge[j*3+k];
@@ -4496,14 +4493,14 @@ void LLVolume::generateSilhouetteVertices(std::vector<LLVector3> &vertices,
 
 						norm_mat.rotate(n[v1], t);
 
-						t.normalize3fast_checked(&default_norm);
+						t.normalize3fast();
 						normals.push_back(LLVector3(t[0], t[1], t[2]));
 
 						mat.affineTransform(v[v2], t);
 						vertices.push_back(LLVector3(t[0], t[1], t[2]));
 						
 						norm_mat.rotate(n[v2], t);
-						t.normalize3fast_checked(&default_norm);
+						t.normalize3fast();
 						normals.push_back(LLVector3(t[0], t[1], t[2]));
 					}
 				}		
@@ -6099,9 +6096,6 @@ BOOL LLVolumeFace::createUnCutCubeCap(LLVolume* volume, BOOL partial_build)
 	{
 		VertexData	corners[4];
 		VertexData baseVert;
-		LLVector4a default_norm;
-		default_norm.set(0,1,0,1);
-
 		for(S32 t = 0; t < 4; t++)
 		{
 			corners[t].getPosition().load3( mesh[offset + (grid_size*t)].mPos.mV);
@@ -6114,8 +6108,8 @@ BOOL LLVolumeFace::createUnCutCubeCap(LLVolume* volume, BOOL partial_build)
 			lhs.setSub(corners[1].getPosition(), corners[0].getPosition());
 			LLVector4a rhs;
 			rhs.setSub(corners[2].getPosition(), corners[1].getPosition());
-			baseVert.getNormal().setCross3(lhs, rhs);
-			baseVert.getNormal().normalize3fast_checked(&default_norm);
+			baseVert.getNormal().setCross3(lhs, rhs); 
+			baseVert.getNormal().normalize3fast();
 		}
 
 		if(!(mTypeMask & TOP_MASK))
@@ -6565,12 +6559,17 @@ BOOL LLVolumeFace::createCap(LLVolume* volume, BOOL partial_build)
 	d1.setSub(mPositions[mIndices[2]], mPositions[mIndices[0]]);
 
 	LLVector4a normal;
-	LLVector4a default_norm;
-	default_norm.set(0,1,0,1);
-
 	normal.setCross3(d0,d1);
-	normal.normalize3fast_checked(&default_norm);
-	
+
+	if (normal.dot3(normal).getF32() > F_APPROXIMATELY_ZERO)
+	{
+		normal.normalize3fast();
+	}
+	else
+	{ //degenerate, make up a value
+		normal.set(0,0,1);
+	}
+
 	llassert(llfinite(normal.getF32ptr()[0]));
 	llassert(llfinite(normal.getF32ptr()[1]));
 	llassert(llfinite(normal.getF32ptr()[2]));
@@ -6612,13 +6611,11 @@ void LLVolumeFace::createTangents()
 		CalculateTangentArray(mNumVertices, mPositions, mNormals, mTexCoords, mNumIndices/3, mIndices, mTangents);
 
 		//normalize tangents
-		LLVector4a default_norm;
-		default_norm.set(0,1,0,1);
 		for (U32 i = 0; i < mNumVertices; i++) 
 		{
 			//binorm[i].normalize3fast();
 			//bump map/planar projection code requires normals to be normalized
-			mNormals[i].normalize3fast_checked(&default_norm);
+			mNormals[i].normalize3fast();
 		}
 	}
 }
@@ -6796,9 +6793,6 @@ void LLVolumeFace::appendFace(const LLVolumeFace& face, LLMatrix4& mat_in, LLMat
 	mat.loadu(mat_in);
 	norm_mat.loadu(norm_mat_in);
 
-	LLVector4a default_norm;
-	default_norm.set(0,1,0,1);
-
 	for (U32 i = 0; i < face.mNumVertices; ++i)
 	{
 		//transform appended face position and store
@@ -6806,7 +6800,7 @@ void LLVolumeFace::appendFace(const LLVolumeFace& face, LLMatrix4& mat_in, LLMat
 
 		//transform appended face normal and store
 		norm_mat.rotate(src_norm[i], dst_norm[i]);
-		dst_norm[i].normalize3fast_checked(&default_norm);
+		dst_norm[i].normalize3fast();
 
 		//copy appended face texture coordinate
 		dst_tc[i] = src_tc[i];
@@ -7215,53 +7209,46 @@ BOOL LLVolumeFace::createSide(LLVolume* volume, BOOL partial_build)
 	return TRUE;
 }
 
-#define TANGENTIAL_PARANOIA_ASSERTS 0
-
-#if TANGENTIAL_PARANOIA_ASSERTS
-	#define tangential_paranoia(a) llassert(a)
-#else
-	#define tangential_paranoia(a)
-#endif
-
 //adapted from Lengyel, Eric. “Computing Tangent Space Basis Vectors for an Arbitrary Mesh”. Terathon Software 3D Graphics Library, 2001. http://www.terathon.com/code/tangent.html
 void CalculateTangentArray(U32 vertexCount, const LLVector4a *vertex, const LLVector4a *normal,
         const LLVector2 *texcoord, U32 triangleCount, const U16* index_array, LLVector4a *tangent)
 {
+    //LLVector4a *tan1 = new LLVector4a[vertexCount * 2];
 	LLVector4a* tan1 = (LLVector4a*) ll_aligned_malloc_16(vertexCount*2*sizeof(LLVector4a));
 
-   LLVector4a* tan2 = tan1 + vertexCount;
+    LLVector4a* tan2 = tan1 + vertexCount;
 
 	memset(tan1, 0, vertexCount*2*sizeof(LLVector4a));
         
-   for (U32 a = 0; a < triangleCount; a++)
-   {
-      U32 i1 = *index_array++;
-      U32 i2 = *index_array++;
-      U32 i3 = *index_array++;
+    for (U32 a = 0; a < triangleCount; a++)
+    {
+        U32 i1 = *index_array++;
+        U32 i2 = *index_array++;
+        U32 i3 = *index_array++;
         
-      const LLVector4a& v1 = vertex[i1];
-      const LLVector4a& v2 = vertex[i2];
-      const LLVector4a& v3 = vertex[i3];
+        const LLVector4a& v1 = vertex[i1];
+        const LLVector4a& v2 = vertex[i2];
+        const LLVector4a& v3 = vertex[i3];
         
-      const LLVector2& w1 = texcoord[i1];
-      const LLVector2& w2 = texcoord[i2];
-      const LLVector2& w3 = texcoord[i3];
+        const LLVector2& w1 = texcoord[i1];
+        const LLVector2& w2 = texcoord[i2];
+        const LLVector2& w3 = texcoord[i3];
         
 		const F32* v1ptr = v1.getF32ptr();
 		const F32* v2ptr = v2.getF32ptr();
 		const F32* v3ptr = v3.getF32ptr();
 		
-      float x1 = v2ptr[0] - v1ptr[0];
-      float x2 = v3ptr[0] - v1ptr[0];
-      float y1 = v2ptr[1] - v1ptr[1];
-      float y2 = v3ptr[1] - v1ptr[1];
-      float z1 = v2ptr[2] - v1ptr[2];
-      float z2 = v3ptr[2] - v1ptr[2];
+        float x1 = v2ptr[0] - v1ptr[0];
+        float x2 = v3ptr[0] - v1ptr[0];
+        float y1 = v2ptr[1] - v1ptr[1];
+        float y2 = v3ptr[1] - v1ptr[1];
+        float z1 = v2ptr[2] - v1ptr[2];
+        float z2 = v3ptr[2] - v1ptr[2];
         
-      float s1 = w2.mV[0] - w1.mV[0];
-      float s2 = w3.mV[0] - w1.mV[0];
-      float t1 = w2.mV[1] - w1.mV[1];
-      float t2 = w3.mV[1] - w1.mV[1];
+        float s1 = w2.mV[0] - w1.mV[0];
+        float s2 = w3.mV[0] - w1.mV[0];
+        float t1 = w2.mV[1] - w1.mV[1];
+        float t2 = w3.mV[1] - w1.mV[1];
         
 		F32 rd = s1*t2-s2*t1;
 
@@ -7275,48 +7262,18 @@ void CalculateTangentArray(U32 vertexCount, const LLVector4a *vertex, const LLVe
 		LLVector4a tdir((s1 * x2 - s2 * x1) * r, (s1 * y2 - s2 * y1) * r,
 				(s1 * z2 - s2 * z1) * r);
         
-		
 		tan1[i1].add(sdir);
 		tan1[i2].add(sdir);
 		tan1[i3].add(sdir);
-      
-		tangential_paranoia(tan1[i1].isFinite3());
-		tangential_paranoia(tan1[i2].isFinite3());
-		tangential_paranoia(tan1[i3].isFinite3());
-
+        
 		tan2[i1].add(tdir);
 		tan2[i2].add(tdir);
 		tan2[i3].add(tdir);
-
-		tangential_paranoia(tan2[i1].isFinite3());
-		tangential_paranoia(tan2[i2].isFinite3());
-		tangential_paranoia(tan2[i3].isFinite3());
-   }
-
-	// These appear to come out of the summing above distinctly non-unit-length
-	//
-	LLVector4a default_norm;
-	default_norm.set(0,1,0,1);
-
-	for (U32 a = 0; a < vertexCount; a++)
-	{
-		tan1[a].normalize3fast_checked(&default_norm);
-		tan2[a].normalize3fast_checked(&default_norm);
-
-		tangential_paranoia(tan1[a].isFinite3());
-		tangential_paranoia(tan2[a].isFinite3());		
-		tangential_paranoia(tan1[a].isNormalized3(0.03f));
-		tangential_paranoia(tan2[a].isNormalized3(0.03f));	
-	}
-
-	LLVector4a default_tangent;
-	default_tangent.set(0,0,1,1);
-
-   for (U32 a = 0; a < vertexCount; a++)
-	{
-		LLVector4a n = normal[a];
-		
-		n.normalize3fast_checked(&default_norm);
+    }
+    
+    for (U32 a = 0; a < vertexCount; a++)
+    {
+        LLVector4a n = normal[a];
 
 		const LLVector4a& t = tan1[a];
 
@@ -7326,41 +7283,38 @@ void CalculateTangentArray(U32 vertexCount, const LLVector4a *vertex, const LLVe
 		LLVector4a ncrosst;
 		ncrosst.setCross3(n,t);
 
-		F32 n_dot_t = n.dot3(t).getF32();
-
-		tangential_paranoia(llfinite(n_dot_t) && !llisnan(n_dot_t));
-
-		// Gram-Schmidt orthogonalize
-      n.mul(n_dot_t);
-
-		tangential_paranoia(n.isFinite3());
+        // Gram-Schmidt orthogonalize
+        n.mul(n.dot3(t).getF32());
 
 		LLVector4a tsubn;
 		tsubn.setSub(t,n);
 
-		tangential_paranoia(tsubn.isFinite3());
-
-		tsubn.normalize3fast_checked(&default_tangent);
+		if (tsubn.dot3(tsubn).getF32() > F_APPROXIMATELY_ZERO)
+		{
+			tsubn.normalize3fast();
 		
-		// Calculate handedness
-		F32 handedness = ncrosst.dot3(tan2[a]).getF32() < 0.f ? -1.f : 1.f;
+			// Calculate handedness
+			F32 handedness = ncrosst.dot3(tan2[a]).getF32() < 0.f ? -1.f : 1.f;
 		
-		tsubn.getF32ptr()[3] = handedness;
+			tsubn.getF32ptr()[3] = handedness;
 
-		tangent[a] = tsubn;
+			tangent[a] = tsubn;
 
-		tangential_paranoia(tangent[a].isNormalized3(0.1f));
+			llassert(llfinite(tangent[a].getF32ptr()[0]));
+			llassert(llfinite(tangent[a].getF32ptr()[1]));
+			llassert(llfinite(tangent[a].getF32ptr()[2]));
 
-		llassert(llfinite(tangent[a].getF32ptr()[0]));
-		llassert(llfinite(tangent[a].getF32ptr()[1]));
-		llassert(llfinite(tangent[a].getF32ptr()[2]));
-
-		llassert(!llisnan(tangent[a].getF32ptr()[0]));
-		llassert(!llisnan(tangent[a].getF32ptr()[1]));
-		llassert(!llisnan(tangent[a].getF32ptr()[2]));
+			llassert(!llisnan(tangent[a].getF32ptr()[0]));
+			llassert(!llisnan(tangent[a].getF32ptr()[1]));
+			llassert(!llisnan(tangent[a].getF32ptr()[2]));
+		}
+		else
+		{ //degenerate, make up a value
+			tangent[a].set(0,0,1,1);
+		}
     }
     
-	 ll_aligned_free_16(tan1);
+	ll_aligned_free_16(tan1);
 }