void LuxelSpaceToWorld( lightinfo_t const *l, fltx4 s, fltx4 t, FourVectors &world )
{
world.DuplicateVector ( l->luxelOrigin );
FourVectors st;
s = AddSIMD ( s, ReplicateX4 ( l->face->m_LightmapTextureMinsInLuxels[0] ) );
st.DuplicateVector ( l->luxelToWorldSpace[0] );
st *= s;
world += st;
t = AddSIMD ( t, ReplicateX4 ( l->face->m_LightmapTextureMinsInLuxels[1] ) );
st.DuplicateVector ( l->luxelToWorldSpace[1] );
st *= t;
world += st;
}
void WorldToLuxelSpace( lightinfo_t const *l, FourVectors const &world, FourVectors &coord )
{
FourVectors luxelOrigin;
luxelOrigin.DuplicateVector ( l->luxelOrigin );
FourVectors pos = world;
pos -= luxelOrigin;
coord.x = pos * l->worldToLuxelSpace[0];
coord.x = SubSIMD ( coord.x, ReplicateX4 ( l->face->m_LightmapTextureMinsInLuxels[0] ) );
coord.y = pos * l->worldToLuxelSpace[1];
coord.y = SubSIMD ( coord.y, ReplicateX4 ( l->face->m_LightmapTextureMinsInLuxels[1] ) );
coord.z = Four_Zeros;
}
CSIMDVectorMatrix & CSIMDVectorMatrix::operator*=( Vector const &src )
{
int nv=NVectors();
if ( nv )
{
FourVectors scalevalue;
scalevalue.DuplicateVector( src );
FourVectors *destv=m_pData;
do // !! speed !! inline more iters
{
destv->VProduct( scalevalue );
destv++;
} while ( --nv );
}
return *this;
}
void LightDesc_t::ComputeLightAtPointsForDirectional(
const FourVectors &pos, const FourVectors &normal,
FourVectors &color, bool DoHalfLambert ) const
{
FourVectors delta;
delta.DuplicateVector(m_Direction);
// delta.VectorNormalizeFast();
fltx4 strength=delta*normal;
if (DoHalfLambert)
{
strength=AddSIMD(MulSIMD(strength,Four_PointFives),Four_PointFives);
}
else
strength=MaxSIMD(Four_Zeros,delta*normal);
color.x=AddSIMD(color.x,MulSIMD(strength,ReplicateX4(m_Color.x)));
color.y=AddSIMD(color.y,MulSIMD(strength,ReplicateX4(m_Color.y)));
color.z=AddSIMD(color.z,MulSIMD(strength,ReplicateX4(m_Color.z)));
}
void LightDesc_t::ComputeLightAtPoints( const FourVectors &pos, const FourVectors &normal,
FourVectors &color, bool DoHalfLambert ) const
{
FourVectors delta;
Assert((m_Type==MATERIAL_LIGHT_POINT) || (m_Type==MATERIAL_LIGHT_SPOT) || (m_Type==MATERIAL_LIGHT_DIRECTIONAL));
switch (m_Type)
{
case MATERIAL_LIGHT_POINT:
case MATERIAL_LIGHT_SPOT:
delta.DuplicateVector(m_Position);
delta-=pos;
break;
case MATERIAL_LIGHT_DIRECTIONAL:
delta.DuplicateVector(m_Direction);
delta*=-1.0;
break;
default:
delta.x = Four_Zeros;
delta.y = Four_Zeros;
delta.z = Four_Zeros;
break;
}
__m128 dist2 = delta*delta;
__m128 falloff;
if( m_Flags & LIGHTTYPE_OPTIMIZATIONFLAGS_HAS_ATTENUATION0 )
{
falloff = MMReplicate(m_Attenuation0);
}
else
falloff= Four_Epsilons;
if( m_Flags & LIGHTTYPE_OPTIMIZATIONFLAGS_HAS_ATTENUATION1 )
{
falloff=_mm_add_ps(falloff,_mm_mul_ps(MMReplicate(m_Attenuation1),_mm_sqrt_ps(dist2)));
}
if( m_Flags & LIGHTTYPE_OPTIMIZATIONFLAGS_HAS_ATTENUATION2 )
{
falloff=_mm_add_ps(falloff,_mm_mul_ps(MMReplicate(m_Attenuation2),dist2));
}
falloff=_mm_rcp_ps(falloff);
// Cull out light beyond this radius
// now, zero out elements for which dist2 was > range^2. !!speed!! lights should store dist^2 in sse format
if (m_Range != 0.f)
{
__m128 RangeSquared=MMReplicate(m_RangeSquared); // !!speed!!
falloff=_mm_and_ps(falloff,_mm_cmplt_ps(dist2,RangeSquared));
}
delta.VectorNormalizeFast();
__m128 strength=delta*normal;
if (DoHalfLambert)
{
strength=_mm_add_ps(_mm_mul_ps(strength,Four_PointFives),Four_PointFives);
}
else
strength=_mm_max_ps(Four_Zeros,delta*normal);
switch(m_Type)
{
case MATERIAL_LIGHT_POINT:
// half-lambert
break;
case MATERIAL_LIGHT_SPOT:
{
__m128 dot2=_mm_sub_ps(Four_Zeros,delta*m_Direction); // dot position with spot light dir for cone falloff
__m128 cone_falloff_scale=_mm_mul_ps(MMReplicate(OneOver_ThetaDot_Minus_PhiDot),
_mm_sub_ps(dot2,MMReplicate(m_PhiDot)));
cone_falloff_scale=_mm_min_ps(cone_falloff_scale,Four_Ones);
if ((m_Falloff!=0.0) && (m_Falloff!=1.0))
{
// !!speed!! could compute integer exponent needed by powsse and store in light
cone_falloff_scale=PowSSE(cone_falloff_scale,m_Falloff);
}
strength=_mm_mul_ps(cone_falloff_scale,strength);
// now, zero out lighting where dot2<phidot. This will mask out any invalid results
// from pow function, etc
__m128 OutsideMask=_mm_cmpgt_ps(dot2,MMReplicate(m_PhiDot)); // outside light cone?
strength=_mm_and_ps(OutsideMask,strength);
}
break;
case MATERIAL_LIGHT_DIRECTIONAL:
break;
default:
break;
}
strength=_mm_mul_ps(strength,falloff);
color.x=_mm_add_ps(color.x,_mm_mul_ps(strength,MMReplicate(m_Color.x)));
color.y=_mm_add_ps(color.y,_mm_mul_ps(strength,MMReplicate(m_Color.y)));
color.z=_mm_add_ps(color.z,_mm_mul_ps(strength,MMReplicate(m_Color.z)));
}
void RayTracingEnvironment::RenderScene(
int width, int height, // width and height of desired rendering
int stride, // actual width in pixels of target buffer
uint32 *output_buffer, // pointer to destination
Vector CameraOrigin, // eye position
Vector ULCorner, // word space coordinates of upper left
// monitor corner
Vector URCorner, // top right corner
Vector LLCorner, // lower left
Vector LRCorner, // lower right
RayTraceLightingMode_t lmode)
{
// first, compute deltas
Vector dxvector=URCorner;
dxvector-=ULCorner;
dxvector*=(1.0/width);
Vector dxvectortimes2=dxvector;
dxvectortimes2+=dxvector;
Vector dyvector=LLCorner;
dyvector-=ULCorner;
dyvector*=(1.0/height);
// block_offsets-relative offsets for eahc of the 4 pixels in the block, in sse format
FourVectors block_offsets;
block_offsets.LoadAndSwizzle(Vector(0,0,0),dxvector,dyvector,dxvector+dyvector);
FourRays myrays;
myrays.origin.DuplicateVector(CameraOrigin);
// tmprays is used fo rthe case when we cannot trace 4 rays at once.
FourRays tmprays;
tmprays.origin.DuplicateVector(CameraOrigin);
// now, we will ray trace pixels. we will do the rays in a 2x2 pattern
for(int y=0;y<height;y+=2)
{
Vector SLoc=dyvector;
SLoc*=((float) y);
SLoc+=ULCorner;
uint32 *dest=output_buffer+y*stride;
for(int x=0;x<width;x+=2)
{
myrays.direction.DuplicateVector(SLoc);
myrays.direction+=block_offsets;
myrays.direction.VectorNormalize();
RayTracingResult rslt;
Trace4Rays(myrays,all_zeros,TraceLimit, &rslt);
if ((rslt.HitIds[0]==-1) && (rslt.HitIds[1]==-1) &&
(rslt.HitIds[2]==-1) && (rslt.HitIds[3]==-1))
MapLinearIntensities(BackgroundColor,dest,dest+1,dest+stride,dest+stride+1);
else
{
// make sure normal points back towards ray origin
fltx4 ndoti=rslt.surface_normal*myrays.direction;
fltx4 bad_dirs=AndSIMD(CmpGtSIMD(ndoti,Four_Zeros),
LoadAlignedSIMD((float *) signmask));
// flip signs of all "wrong" normals
rslt.surface_normal.x=XorSIMD(bad_dirs,rslt.surface_normal.x);
rslt.surface_normal.y=XorSIMD(bad_dirs,rslt.surface_normal.y);
rslt.surface_normal.z=XorSIMD(bad_dirs,rslt.surface_normal.z);
FourVectors intens;
intens.DuplicateVector(Vector(0,0,0));
// set up colors
FourVectors surf_colors;
surf_colors.DuplicateVector(Vector(0,0,0));
for(int i=0;i<4;i++)
{
if (rslt.HitIds[i]>=0)
{
surf_colors.X(i)=TriangleColors[rslt.HitIds[i]].x;
surf_colors.Y(i)=TriangleColors[rslt.HitIds[i]].y;
surf_colors.Z(i)=TriangleColors[rslt.HitIds[i]].z;
}
}
FourVectors surface_pos=myrays.direction;
surface_pos*=rslt.HitDistance;
surface_pos+=myrays.origin;
switch(lmode)
{
case DIRECT_LIGHTING:
{
// light all points
for(int l=0;l<LightList.Count();l++)
{
LightList[l].ComputeLightAtPoints(surface_pos,rslt.surface_normal,
intens);
}
}
break;
case DIRECT_LIGHTING_WITH_SHADOWS:
{
// light all points
for(int l=0;l<LightList.Count();l++)
{
FourVectors ldir;
ldir.DuplicateVector(LightList[l].m_Position);
ldir-=surface_pos;
fltx4 MaxT=ldir.length();
ldir.VectorNormalizeFast();
// now, compute shadow flag
FourRays myrays;
myrays.origin=surface_pos;
FourVectors epsilon=ldir;
epsilon*=0.01;
myrays.origin+=epsilon;
myrays.direction=ldir;
RayTracingResult shadowtest;
Trace4Rays(myrays,Four_Zeros,MaxT, &shadowtest);
fltx4 unshadowed=CmpGtSIMD(shadowtest.HitDistance,MaxT);
if (! (IsAllZeros(unshadowed)))
{
FourVectors tmp;
tmp.DuplicateVector(Vector(0,0,0));
LightList[l].ComputeLightAtPoints(surface_pos,rslt.surface_normal,
tmp);
intens.x=AddSIMD(intens.x,AndSIMD(tmp.x,unshadowed));
intens.y=AddSIMD(intens.y,AndSIMD(tmp.y,unshadowed));
intens.z=AddSIMD(intens.z,AndSIMD(tmp.z,unshadowed));
}
}
}
break;
}
// now, mask off non-hitting pixels
intens.VProduct(surf_colors);
fltx4 no_hit_mask=CmpGtSIMD(rslt.HitDistance,TraceLimit);
intens.x=OrSIMD(AndSIMD(BackgroundColor.x,no_hit_mask),
AndNotSIMD(no_hit_mask,intens.x));
intens.y=OrSIMD(AndSIMD(BackgroundColor.y,no_hit_mask),
AndNotSIMD(no_hit_mask,intens.y));
intens.z=OrSIMD(AndSIMD(BackgroundColor.y,no_hit_mask),
AndNotSIMD(no_hit_mask,intens.z));
MapLinearIntensities(intens,dest,dest+1,dest+stride,dest+stride+1);
}
dest+=2;
SLoc+=dxvectortimes2;
}
}
}
void RayTracingEnvironment::ComputeVirtualLightSources(void)
{
int start_pos=0;
for(int b=0;b<3;b++)
{
int nl=LightList.Count();
int where_to_start=start_pos;
start_pos=nl;
for(int l=where_to_start;l<nl;l++)
{
DirectionalSampler_t sample_generator;
int n_desired=1*LightList[l].m_Color.Length();
if (LightList[l].m_Type==MATERIAL_LIGHT_SPOT)
n_desired*=LightList[l].m_Phi/2;
for(int try1=0;try1<n_desired;try1++)
{
LightDesc_t const &li=LightList[l];
FourRays myrays;
myrays.origin.DuplicateVector(li.m_Position);
RayTracingResult rslt;
Vector trial_dir=sample_generator.NextValue();
if (li.IsDirectionWithinLightCone(trial_dir))
{
myrays.direction.DuplicateVector(trial_dir);
Trace4Rays(myrays,all_zeros,ReplicateX4(1000.0), &rslt);
if ((rslt.HitIds[0]!=-1))
{
// make sure normal points back towards ray origin
fltx4 ndoti=rslt.surface_normal*myrays.direction;
fltx4 bad_dirs=AndSIMD(CmpGtSIMD(ndoti,Four_Zeros),
LoadAlignedSIMD((float *) signmask));
// flip signs of all "wrong" normals
rslt.surface_normal.x=XorSIMD(bad_dirs,rslt.surface_normal.x);
rslt.surface_normal.y=XorSIMD(bad_dirs,rslt.surface_normal.y);
rslt.surface_normal.z=XorSIMD(bad_dirs,rslt.surface_normal.z);
// a hit! let's make a virtual light source
// treat the virtual light as a disk with its center at the hit position
// and its radius scaled by the amount of the solid angle this probe
// represents.
float area_of_virtual_light=
4.0*M_PI*SQ( SubFloat( rslt.HitDistance, 0 ) )*(1.0/n_desired);
FourVectors intens;
intens.DuplicateVector(Vector(0,0,0));
FourVectors surface_pos=myrays.direction;
surface_pos*=rslt.HitDistance;
surface_pos+=myrays.origin;
FourVectors delta=rslt.surface_normal;
delta*=0.1;
surface_pos+=delta;
LightList[l].ComputeLightAtPoints(surface_pos,rslt.surface_normal,
intens);
FourVectors surf_colors;
surf_colors.DuplicateVector(TriangleColors[rslt.HitIds[0]]);
intens*=surf_colors;
// see if significant
LightDesc_t l1;
l1.m_Type=MATERIAL_LIGHT_SPOT;
l1.m_Position=Vector(surface_pos.X(0),surface_pos.Y(0),surface_pos.Z(0));
l1.m_Direction=Vector(rslt.surface_normal.X(0),rslt.surface_normal.Y(0),
rslt.surface_normal.Z(0));
l1.m_Color=Vector(intens.X(0),intens.Y(0),intens.Z(0));
if (l1.m_Color.Length()>0)
{
l1.m_Color*=area_of_virtual_light/M_PI;
l1.m_Range=0.0;
l1.m_Falloff=1.0;
l1.m_Attenuation0=1.0;
l1.m_Attenuation1=0.0;
l1.m_Attenuation2=1.0; // intens falls off as 1/r^2
l1.m_Theta=0;
l1.m_Phi=M_PI;
l1.RecalculateDerivedValues();
LightList.AddToTail(l1);
}
}
}
}
}
}
}
void LightDesc_t::ComputeLightAtPoints( const FourVectors &pos, const FourVectors &normal,
FourVectors &color, bool DoHalfLambert ) const
{
FourVectors delta;
Assert((m_Type==MATERIAL_LIGHT_POINT) || (m_Type==MATERIAL_LIGHT_SPOT) || (m_Type==MATERIAL_LIGHT_DIRECTIONAL));
switch (m_Type)
{
case MATERIAL_LIGHT_POINT:
case MATERIAL_LIGHT_SPOT:
delta.DuplicateVector(m_Position);
delta-=pos;
break;
case MATERIAL_LIGHT_DIRECTIONAL:
ComputeLightAtPointsForDirectional( pos, normal, color, DoHalfLambert );
return;
default:
return;
}
fltx4 dist2 = delta*delta;
dist2=MaxSIMD( Four_Ones, dist2 );
fltx4 falloff;
if( m_Flags & LIGHTTYPE_OPTIMIZATIONFLAGS_HAS_ATTENUATION0 )
{
falloff = ReplicateX4(m_Attenuation0);
}
else
falloff= Four_Epsilons;
if( m_Flags & LIGHTTYPE_OPTIMIZATIONFLAGS_HAS_ATTENUATION1 )
{
falloff=AddSIMD(falloff,MulSIMD(ReplicateX4(m_Attenuation1),SqrtEstSIMD(dist2)));
}
if( m_Flags & LIGHTTYPE_OPTIMIZATIONFLAGS_HAS_ATTENUATION2 )
{
falloff=AddSIMD(falloff,MulSIMD(ReplicateX4(m_Attenuation2),dist2));
}
falloff=ReciprocalEstSIMD(falloff);
// Cull out light beyond this radius
// now, zero out elements for which dist2 was > range^2. !!speed!! lights should store dist^2 in sse format
if (m_Range != 0.f)
{
fltx4 RangeSquared=ReplicateX4(m_RangeSquared); // !!speed!!
falloff=AndSIMD(falloff,CmpLtSIMD(dist2,RangeSquared));
}
delta.VectorNormalizeFast();
fltx4 strength=delta*normal;
if (DoHalfLambert)
{
strength=AddSIMD(MulSIMD(strength,Four_PointFives),Four_PointFives);
}
else
strength=MaxSIMD(Four_Zeros,delta*normal);
switch(m_Type)
{
case MATERIAL_LIGHT_POINT:
// half-lambert
break;
case MATERIAL_LIGHT_SPOT:
{
fltx4 dot2=SubSIMD(Four_Zeros,delta*m_Direction); // dot position with spot light dir for cone falloff
fltx4 cone_falloff_scale=MulSIMD(ReplicateX4(m_OneOverThetaDotMinusPhiDot),
SubSIMD(dot2,ReplicateX4(m_PhiDot)));
cone_falloff_scale=MinSIMD(cone_falloff_scale,Four_Ones);
if ((m_Falloff!=0.0) && (m_Falloff!=1.0))
{
// !!speed!! could compute integer exponent needed by powsimd and store in light
cone_falloff_scale=PowSIMD(cone_falloff_scale,m_Falloff);
}
strength=MulSIMD(cone_falloff_scale,strength);
// now, zero out lighting where dot2<phidot. This will mask out any invalid results
// from pow function, etc
fltx4 OutsideMask=CmpGtSIMD(dot2,ReplicateX4(m_PhiDot)); // outside light cone?
strength=AndSIMD(OutsideMask,strength);
}
break;
default:
break;
}
strength=MulSIMD(strength,falloff);
color.x=AddSIMD(color.x,MulSIMD(strength,ReplicateX4(m_Color.x)));
color.y=AddSIMD(color.y,MulSIMD(strength,ReplicateX4(m_Color.y)));
color.z=AddSIMD(color.z,MulSIMD(strength,ReplicateX4(m_Color.z)));
}