engine: optimize traces

Author: nillerusr

engine/cmodel.cpp

@@ -873,9 +873,9 @@ bool IntersectRayWithBoxBrush( TraceInfo_t *pTraceInfo, const cbrush_t *pBrush,
 	FPExceptionDisabler hideExceptions;
 
 	// Load the unaligned ray/box parameters into SIMD registers
-	fltx4 start = LoadUnaligned3SIMD(pTraceInfo->m_start.Base());
-	fltx4 extents = LoadUnaligned3SIMD(pTraceInfo->m_extents.Base());
-	fltx4 delta = LoadUnaligned3SIMD(pTraceInfo->m_delta.Base());
+	fltx4 start = LoadAlignedSIMD(pTraceInfo->m_start.Base());
+	fltx4 extents = LoadAlignedSIMD(pTraceInfo->m_extents.Base());
+	fltx4 delta = LoadAlignedSIMD(pTraceInfo->m_delta.Base());
 	fltx4 boxMins = LoadAlignedSIMD( pBox->mins.Base() );
 	fltx4 boxMaxs = LoadAlignedSIMD( pBox->maxs.Base() );
 
@@ -899,7 +899,7 @@ bool IntersectRayWithBoxBrush( TraceInfo_t *pTraceInfo, const cbrush_t *pBrush,
 
 	fltx4 crossPlane = OrSIMD(XorSIMD(startOutMins,endOutMins), XorSIMD(startOutMaxs,endOutMaxs));
 	// now build the per-axis interval of t for intersections
-	fltx4 invDelta = LoadUnaligned3SIMD(pTraceInfo->m_invDelta.Base());
+	fltx4 invDelta = LoadAlignedSIMD(pTraceInfo->m_invDelta.Base());
 	fltx4 tmins = MulSIMD( offsetMinsExpanded, invDelta );
 	fltx4 tmaxs = MulSIMD( offsetMaxsExpanded, invDelta );
 	// now sort the interval per axis
@@ -1037,9 +1037,9 @@ bool IntersectRayWithBox( const Ray_t &ray, const VectorAligned &inInvDelta, con
 	pTrace->fraction = 1.0f;
 
 	// Load the unaligned ray/box parameters into SIMD registers
-	fltx4 start = LoadUnaligned3SIMD(ray.m_Start.Base());
-	fltx4 extents = LoadUnaligned3SIMD(ray.m_Extents.Base());
-	fltx4 delta = LoadUnaligned3SIMD(ray.m_Delta.Base());
+	fltx4 start = LoadAlignedSIMD(ray.m_Start.Base());
+	fltx4 extents = LoadAlignedSIMD(ray.m_Extents.Base());
+	fltx4 delta = LoadAlignedSIMD(ray.m_Delta.Base());
 	fltx4 boxMins = LoadAlignedSIMD( inBoxMins.Base() );
 	fltx4 boxMaxs = LoadAlignedSIMD( inBoxMaxs.Base() );
 
@@ -1372,9 +1372,9 @@ void FASTCALL CM_ClipBoxToBrush( TraceInfo_t * RESTRICT pTraceInfo, const cbrush
 
 inline bool IsTraceBoxIntersectingBoxBrush( TraceInfo_t *pTraceInfo, cboxbrush_t *pBox )
 {
-	fltx4 start = LoadUnaligned3SIMD(pTraceInfo->m_start.Base());
-	fltx4 mins = LoadUnaligned3SIMD(pTraceInfo->m_mins.Base());
-	fltx4 maxs = LoadUnaligned3SIMD(pTraceInfo->m_maxs.Base());
+	fltx4 start = LoadAlignedSIMD(pTraceInfo->m_start.Base());
+	fltx4 mins = LoadAlignedSIMD(pTraceInfo->m_mins.Base());
+	fltx4 maxs = LoadAlignedSIMD(pTraceInfo->m_maxs.Base());
 
 	fltx4 boxMins = LoadAlignedSIMD( pBox->mins.Base() );
 	fltx4 boxMaxs = LoadAlignedSIMD( pBox->maxs.Base() );
@@ -1569,15 +1569,15 @@ void FASTCALL CM_TraceToLeaf( TraceInfo_t * RESTRICT pTraceInfo, int ndxLeaf, fl
 		if (IsX360())
 		{
 			// set up some relatively constant variables we'll use in the loop below
-			fltx4 traceStart = LoadUnaligned3SIMD(pTraceInfo->m_start.Base());
-			fltx4 traceDelta = LoadUnaligned3SIMD(pTraceInfo->m_delta.Base());
-			fltx4 traceInvDelta = LoadUnaligned3SIMD(pTraceInfo->m_invDelta.Base());
+			fltx4 traceStart = LoadAlignedSIMD(pTraceInfo->m_start.Base());
+			fltx4 traceDelta = LoadAlignedSIMD(pTraceInfo->m_delta.Base());
+			fltx4 traceInvDelta = LoadAlignedSIMD(pTraceInfo->m_invDelta.Base());
 			static const fltx4 vecEpsilon = {DISPCOLL_DIST_EPSILON,DISPCOLL_DIST_EPSILON,DISPCOLL_DIST_EPSILON,DISPCOLL_DIST_EPSILON};
 			// only used in !IS_POINT version:
 			fltx4 extents;
 			if (!IS_POINT)
 			{
-				extents = LoadUnaligned3SIMD(pTraceInfo->m_extents.Base());
+				extents = LoadAlignedSIMD(pTraceInfo->m_extents.Base());
 			}
 
 			// TODO: this loop probably ought to be unrolled so that we can make a more efficient

engine/cmodel_private.h

@@ -42,13 +42,13 @@ struct TraceInfo_t
 		m_nCheckDepth = -1;
 	}
 
-	Vector m_start;
-	Vector m_end;
-	Vector m_mins;
-	Vector m_maxs;
-	Vector m_extents;
-	Vector m_delta;
-	Vector m_invDelta;
+	VectorAligned m_start;
+	VectorAligned m_end;
+	VectorAligned m_mins;
+	VectorAligned m_maxs;
+	VectorAligned m_extents;
+	VectorAligned m_delta;
+	VectorAligned m_invDelta;
 
 	trace_t m_trace;
 	trace_t m_stabTrace;

engine/gl_rsurf.cpp

@@ -4934,7 +4934,7 @@ static bool EnumerateLeafInBox_R(mnode_t * RESTRICT node, const EnumLeafBoxInfo_
 	*/
 
 	// take advantage of high throughput/high latency
-	fltx4 planeNormal = LoadUnaligned3SIMD( plane->normal.Base() );
+	fltx4 planeNormal = LoadAlignedSIMD( plane->normal.Base() );
 	fltx4 vecBoxMin = LoadAlignedSIMD(pInfo->m_vecBoxMin);
 	fltx4 vecBoxMax = LoadAlignedSIMD(pInfo->m_vecBoxMax);
 	fltx4 cornermin, cornermax;

engine/spatialpartition.cpp

@@ -987,7 +987,7 @@ class CPartitionVisitor
 	int m_iTree;
 };
 
-
+/*
 class CIntersectPoint : public CPartitionVisitor
 {
 public:
@@ -1009,7 +1009,7 @@ class CIntersectPoint : public CPartitionVisitor
 private:
 	fltx4 m_f4Point;
 };
-
+*/
 
 class CIntersectBox : public CPartitionVisitor
 {
@@ -1040,8 +1040,8 @@ class CIntersectRay : public CPartitionVisitor
 public:
 	CIntersectRay( CVoxelTree *pPartition, const Ray_t &ray, const Vector &vecInvDelta ) : CPartitionVisitor( pPartition )
 	{
-		m_f4Start = LoadUnaligned3SIMD( ray.m_Start.Base() );
-		m_f4Delta = LoadUnaligned3SIMD( ray.m_Delta.Base() );
+		m_f4Start = LoadAlignedSIMD( ray.m_Start.Base() );
+		m_f4Delta = LoadAlignedSIMD( ray.m_Delta.Base() );
 		m_f4InvDelta = LoadUnaligned3SIMD( vecInvDelta.Base() );
 	}
 
@@ -1069,10 +1069,10 @@ class CIntersectSweptBox : public CPartitionVisitor
 public:
 	CIntersectSweptBox( CVoxelTree *pPartition, const Ray_t &ray, const Vector &vecInvDelta ) : CPartitionVisitor( pPartition )
 	{
-		m_f4Start = LoadUnaligned3SIMD( ray.m_Start.Base() );
-		m_f4Delta = LoadUnaligned3SIMD( ray.m_Delta.Base() );
+		m_f4Start = LoadAlignedSIMD( ray.m_Start.Base() );
+		m_f4Delta = LoadAlignedSIMD( ray.m_Delta.Base() );
+		m_f4Extents = LoadAlignedSIMD( ray.m_Extents.Base() );
 		m_f4InvDelta = LoadUnaligned3SIMD( vecInvDelta.Base() );
-		m_f4Extents = LoadUnaligned3SIMD( ray.m_Extents.Base() );
 	}
 
 	bool Intersects( const float *pMins, const float *pMaxs ) const

public/mathlib/mathlib.h

@@ -114,7 +114,7 @@ inline T clamp( T const &val, T const &minVal, T const &maxVal )
 // FIXME: this should move to a different file
 struct cplane_t
 {
-	Vector	normal;
+	VectorAligned	normal;
 	float	dist;
 	byte	type;			// for fast side tests
 	byte	signbits;		// signx + (signy<<1) + (signz<<1)