godot/thirdparty/cvtt/ConvectionKernels_IndexSelectorHDR.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

156 lines
6.3 KiB
C++

#pragma once
#ifndef __CVTT_INDEXSELECTORHDR_H__
#define __CVTT_INDEXSELECTORHDR_H__
#include "ConvectionKernels_ParallelMath.h"
#include "ConvectionKernels_IndexSelector.h"
namespace cvtt
{
namespace Internal
{
ParallelMath::SInt16 UnscaleHDRValueSigned(const ParallelMath::SInt16 &v);
ParallelMath::UInt15 UnscaleHDRValueUnsigned(const ParallelMath::UInt16 &v);
template<int TVectorSize>
class IndexSelectorHDR : public IndexSelector<TVectorSize>
{
public:
typedef ParallelMath::UInt15 MUInt15;
typedef ParallelMath::UInt16 MUInt16;
typedef ParallelMath::UInt31 MUInt31;
typedef ParallelMath::SInt16 MSInt16;
typedef ParallelMath::SInt32 MSInt32;
typedef ParallelMath::Float MFloat;
private:
MUInt15 InvertSingle(const MUInt15& anIndex) const
{
MUInt15 inverted = m_maxValueMinusOne - anIndex;
return ParallelMath::Select(m_isInverted, inverted, anIndex);
}
void ReconstructHDRSignedUninverted(const MUInt15 &index, MSInt16* pixel) const
{
MUInt15 weight = ParallelMath::LosslessCast<MUInt15>::Cast(ParallelMath::RightShift(ParallelMath::CompactMultiply(g_weightReciprocals[m_range], index) + 256, 9));
for (int ch = 0; ch < TVectorSize; ch++)
{
MSInt16 ep0 = ParallelMath::LosslessCast<MSInt16>::Cast(this->m_endPoint[0][ch]);
MSInt16 ep1 = ParallelMath::LosslessCast<MSInt16>::Cast(this->m_endPoint[1][ch]);
MSInt32 pixel32 = ParallelMath::XMultiply((ParallelMath::MakeUInt15(64) - weight), ep0) + ParallelMath::XMultiply(weight, ep1);
pixel32 = ParallelMath::RightShift(pixel32 + ParallelMath::MakeSInt32(32), 6);
pixel[ch] = UnscaleHDRValueSigned(ParallelMath::ToSInt16(pixel32));
}
}
void ReconstructHDRUnsignedUninverted(const MUInt15 &index, MSInt16* pixel) const
{
MUInt15 weight = ParallelMath::LosslessCast<MUInt15>::Cast(ParallelMath::RightShift(ParallelMath::CompactMultiply(g_weightReciprocals[m_range], index) + 256, 9));
for (int ch = 0; ch < TVectorSize; ch++)
{
MUInt16 ep0 = ParallelMath::LosslessCast<MUInt16>::Cast(this->m_endPoint[0][ch]);
MUInt16 ep1 = ParallelMath::LosslessCast<MUInt16>::Cast(this->m_endPoint[1][ch]);
MUInt31 pixel31 = ParallelMath::XMultiply((ParallelMath::MakeUInt15(64) - weight), ep0) + ParallelMath::XMultiply(weight, ep1);
pixel31 = ParallelMath::RightShift(pixel31 + ParallelMath::MakeUInt31(32), 6);
pixel[ch] = ParallelMath::LosslessCast<MSInt16>::Cast(UnscaleHDRValueUnsigned(ParallelMath::ToUInt16(pixel31)));
}
}
MFloat ErrorForInterpolatorComponent(int index, int ch, const MFloat *pixel) const
{
MFloat diff = pixel[ch] - m_reconstructedInterpolators[index][ch];
return diff * diff;
}
MFloat ErrorForInterpolator(int index, const MFloat *pixel) const
{
MFloat error = ErrorForInterpolatorComponent(index, 0, pixel);
for (int ch = 1; ch < TVectorSize; ch++)
error = error + ErrorForInterpolatorComponent(index, ch, pixel);
return error;
}
public:
void InitHDR(int range, bool isSigned, bool fastIndexing, const float *channelWeights)
{
assert(range <= 16);
m_range = range;
m_isInverted = ParallelMath::MakeBoolInt16(false);
m_maxValueMinusOne = ParallelMath::MakeUInt15(static_cast<uint16_t>(range - 1));
if (!fastIndexing)
{
for (int i = 0; i < range; i++)
{
MSInt16 recon2CL[TVectorSize];
if (isSigned)
ReconstructHDRSignedUninverted(ParallelMath::MakeUInt15(static_cast<uint16_t>(i)), recon2CL);
else
ReconstructHDRUnsignedUninverted(ParallelMath::MakeUInt15(static_cast<uint16_t>(i)), recon2CL);
for (int ch = 0; ch < TVectorSize; ch++)
m_reconstructedInterpolators[i][ch] = ParallelMath::TwosCLHalfToFloat(recon2CL[ch]) * channelWeights[ch];
}
}
}
void ReconstructHDRSigned(const MUInt15 &index, MSInt16* pixel) const
{
ReconstructHDRSignedUninverted(InvertSingle(index), pixel);
}
void ReconstructHDRUnsigned(const MUInt15 &index, MSInt16* pixel) const
{
ReconstructHDRUnsignedUninverted(InvertSingle(index), pixel);
}
void ConditionalInvert(const ParallelMath::Int16CompFlag &invert)
{
m_isInverted = invert;
}
MUInt15 SelectIndexHDRSlow(const MFloat* pixel, const ParallelMath::RoundTowardNearestForScope*) const
{
MUInt15 index = ParallelMath::MakeUInt15(0);
MFloat bestError = ErrorForInterpolator(0, pixel);
for (int i = 1; i < m_range; i++)
{
MFloat error = ErrorForInterpolator(i, pixel);
ParallelMath::FloatCompFlag errorBetter = ParallelMath::Less(error, bestError);
ParallelMath::ConditionalSet(index, ParallelMath::FloatFlagToInt16(errorBetter), ParallelMath::MakeUInt15(static_cast<uint16_t>(i)));
bestError = ParallelMath::Min(bestError, error);
}
return InvertSingle(index);
}
MUInt15 SelectIndexHDRFast(const MFloat* pixel, const ParallelMath::RoundTowardNearestForScope* rtn) const
{
return InvertSingle(this->SelectIndexLDR(pixel, rtn));
}
private:
MFloat m_reconstructedInterpolators[16][TVectorSize];
ParallelMath::Int16CompFlag m_isInverted;
MUInt15 m_maxValueMinusOne;
int m_range;
};
}
}
#endif