godot/thirdparty/cvtt/ConvectionKernels_EndpointRefiner.h
K. S. Ernest (iFire) Lee 419b342a9a Faster CVTT by reducing quality.
Make BC6 and BC7 CVTT faster while still having better quality than DXT5.
2022-02-04 15:15:26 -08:00

182 lines
6.7 KiB
C++

#pragma once
#ifndef __CVTT_ENDPOINTREFINER_H__
#define __CVTT_ENDPOINTREFINER_H__
#include "ConvectionKernels_ParallelMath.h"
namespace cvtt
{
namespace Internal
{
// Solve for a, b where v = a*t + b
// This allows endpoints to be mapped to where T=0 and T=1
// Least squares from totals:
// a = (tv - t*v/w)/(tt - t*t/w)
// b = (v - a*t)/w
template<int TVectorSize>
class EndpointRefiner
{
public:
typedef ParallelMath::Float MFloat;
typedef ParallelMath::UInt16 MUInt16;
typedef ParallelMath::UInt15 MUInt15;
typedef ParallelMath::AInt16 MAInt16;
typedef ParallelMath::SInt16 MSInt16;
typedef ParallelMath::SInt32 MSInt32;
MFloat m_tv[TVectorSize];
MFloat m_v[TVectorSize];
MFloat m_tt;
MFloat m_t;
MFloat m_w;
int m_wu;
float m_rcpMaxIndex;
float m_channelWeights[TVectorSize];
float m_rcpChannelWeights[TVectorSize];
void Init(int indexRange, const float channelWeights[TVectorSize])
{
for (int ch = 0; ch < TVectorSize; ch++)
{
m_tv[ch] = ParallelMath::MakeFloatZero();
m_v[ch] = ParallelMath::MakeFloatZero();
}
m_tt = ParallelMath::MakeFloatZero();
m_t = ParallelMath::MakeFloatZero();
m_w = ParallelMath::MakeFloatZero();
m_rcpMaxIndex = 1.0f / static_cast<float>(indexRange - 1);
for (int ch = 0; ch < TVectorSize; ch++)
{
m_channelWeights[ch] = channelWeights[ch];
m_rcpChannelWeights[ch] = 1.0f;
if (m_channelWeights[ch] != 0.0f)
m_rcpChannelWeights[ch] = 1.0f / channelWeights[ch];
}
m_wu = 0;
}
void ContributePW(const MFloat *pwFloatPixel, const MUInt15 &index, const MFloat &weight)
{
MFloat t = ParallelMath::ToFloat(index) * m_rcpMaxIndex;
for (int ch = 0; ch < TVectorSize; ch++)
{
MFloat v = pwFloatPixel[ch] * weight;
m_tv[ch] = m_tv[ch] + t * v;
m_v[ch] = m_v[ch] + v;
}
m_tt = m_tt + weight * t * t;
m_t = m_t + weight * t;
m_w = m_w + weight;
}
void ContributeUnweightedPW(const MFloat *pwFloatPixel, const MUInt15 &index, int numRealChannels)
{
MFloat t = ParallelMath::ToFloat(index) * m_rcpMaxIndex;
for (int ch = 0; ch < numRealChannels; ch++)
{
MFloat v = pwFloatPixel[ch];
m_tv[ch] = m_tv[ch] + t * v;
m_v[ch] = m_v[ch] + v;
}
m_tt = m_tt + t * t;
m_t = m_t + t;
m_wu++;
}
void ContributeUnweightedPW(const MFloat *floatPixel, const MUInt15 &index)
{
ContributeUnweightedPW(floatPixel, index, TVectorSize);
}
void GetRefinedEndpoints(MFloat endPoint[2][TVectorSize])
{
// a = (tv - t*v/w)/(tt - t*t/w)
// b = (v - a*t)/w
MFloat w = m_w + ParallelMath::MakeFloat(static_cast<float>(m_wu));
ParallelMath::MakeSafeDenominator(w);
MFloat wRcp = ParallelMath::Reciprocal(w);
MFloat adenom = (m_tt * w - m_t * m_t) * wRcp;
ParallelMath::FloatCompFlag adenomZero = ParallelMath::Equal(adenom, ParallelMath::MakeFloatZero());
ParallelMath::ConditionalSet(adenom, adenomZero, ParallelMath::MakeFloat(1.0f));
for (int ch = 0; ch < TVectorSize; ch++)
{
/*
if (adenom == 0.0)
p1 = p2 = er.v / er.w;
else
{
float4 a = (er.tv - er.t*er.v / er.w) / adenom;
float4 b = (er.v - a * er.t) / er.w;
p1 = b;
p2 = a + b;
}
*/
MFloat a = (m_tv[ch] - m_t * m_v[ch] * wRcp) / adenom;
MFloat b = (m_v[ch] - a * m_t) * wRcp;
MFloat p1 = b;
MFloat p2 = a + b;
ParallelMath::ConditionalSet(p1, adenomZero, (m_v[ch] * wRcp));
ParallelMath::ConditionalSet(p2, adenomZero, p1);
// Unweight
float inverseWeight = m_rcpChannelWeights[ch];
endPoint[0][ch] = p1 * inverseWeight;
endPoint[1][ch] = p2 * inverseWeight;
}
}
void GetRefinedEndpointsLDR(MUInt15 endPoint[2][TVectorSize], int numRealChannels, const ParallelMath::RoundTowardNearestForScope *roundingMode)
{
MFloat floatEndPoint[2][TVectorSize];
GetRefinedEndpoints(floatEndPoint);
for (int epi = 0; epi < 2; epi++)
for (int ch = 0; ch < TVectorSize; ch++)
endPoint[epi][ch] = ParallelMath::RoundAndConvertToU15(ParallelMath::Clamp(floatEndPoint[epi][ch], 0.0f, 255.0f), roundingMode);
}
void GetRefinedEndpointsLDR(MUInt15 endPoint[2][TVectorSize], const ParallelMath::RoundTowardNearestForScope *roundingMode)
{
GetRefinedEndpointsLDR(endPoint, TVectorSize, roundingMode);
}
void GetRefinedEndpointsHDR(MSInt16 endPoint[2][TVectorSize], bool isSigned, const ParallelMath::RoundTowardNearestForScope *roundingMode)
{
MFloat floatEndPoint[2][TVectorSize];
GetRefinedEndpoints(floatEndPoint);
for (int epi = 0; epi < 2; epi++)
{
for (int ch = 0; ch < TVectorSize; ch++)
{
MFloat f = floatEndPoint[epi][ch];
if (isSigned)
endPoint[epi][ch] = ParallelMath::LosslessCast<MSInt16>::Cast(ParallelMath::RoundAndConvertToS16(ParallelMath::Clamp(f, -31743.0f, 31743.0f), roundingMode));
else
endPoint[epi][ch] = ParallelMath::LosslessCast<MSInt16>::Cast(ParallelMath::RoundAndConvertToU15(ParallelMath::Clamp(f, 0.0f, 31743.0f), roundingMode));
}
}
}
};
}
}
#endif