Faster CVTT by reducing quality.

Make BC6 and BC7 CVTT faster while still having better quality than DXT5.

1 files changed, 147 insertions, 0 deletions

diff --git a/thirdparty/cvtt/ConvectionKernels_IndexSelector.h b/thirdparty/cvtt/ConvectionKernels_IndexSelector.h
new file mode 100644
index 0000000000..0f9d209183
--- /dev/null
+++ b/thirdparty/cvtt/ConvectionKernels_IndexSelector.h
@@ -0,0 +1,147 @@
+#pragma once
+#ifndef __CVTT_INDEXSELECTOR_H__
+#define __CVTT_INDEXSELECTOR_H__
+
+#include "ConvectionKernels_ParallelMath.h"
+
+namespace cvtt
+{
+    namespace Internal
+    {
+        extern const ParallelMath::UInt16 g_weightReciprocals[17];
+
+        template<int TVectorSize>
+        class IndexSelector
+        {
+        public:
+            typedef ParallelMath::Float MFloat;
+            typedef ParallelMath::UInt16 MUInt16;
+            typedef ParallelMath::UInt15 MUInt15;
+            typedef ParallelMath::SInt16 MSInt16;
+            typedef ParallelMath::AInt16 MAInt16;
+            typedef ParallelMath::SInt32 MSInt32;
+            typedef ParallelMath::UInt31 MUInt31;
+
+
+            template<class TInterpolationEPType, class TColorEPType>
+            void Init(const float *channelWeights, const TInterpolationEPType interpolationEndPoints[2][TVectorSize], const TColorEPType colorSpaceEndpoints[2][TVectorSize], int range)
+            {
+                // In BC6H, the interpolation endpoints are higher-precision than the endpoints in color space.
+                // We need to select indexes using the color-space endpoints.
+
+                m_isUniform = true;
+                for (int ch = 1; ch < TVectorSize; ch++)
+                {
+                    if (channelWeights[ch] != channelWeights[0])
+                        m_isUniform = false;
+                }
+
+                // To work with channel weights, we need something where:
+                // pxDiff = px - ep[0]
+                // epDiff = ep[1] - ep[0]
+                //
+                // weightedEPDiff = epDiff * channelWeights
+                // normalizedWeightedAxis = weightedEPDiff / len(weightedEPDiff)
+                // normalizedIndex = dot(pxDiff * channelWeights, normalizedWeightedAxis) / len(weightedEPDiff)
+                // index = normalizedIndex * maxValue
+                //
+                // Equivalent to:
+                // axis = channelWeights * maxValue * epDiff * channelWeights / lenSquared(epDiff * channelWeights)
+                // index = dot(axis, pxDiff)
+
+                for (int ep = 0; ep < 2; ep++)
+                    for (int ch = 0; ch < TVectorSize; ch++)
+                        m_endPoint[ep][ch] = ParallelMath::LosslessCast<MAInt16>::Cast(interpolationEndPoints[ep][ch]);
+
+                m_range = range;
+                m_maxValue = static_cast<float>(range - 1);
+
+                MFloat epDiffWeighted[TVectorSize];
+                for (int ch = 0; ch < TVectorSize; ch++)
+                {
+                    m_origin[ch] = ParallelMath::ToFloat(colorSpaceEndpoints[0][ch]);
+                    MFloat opposingOriginCh = ParallelMath::ToFloat(colorSpaceEndpoints[1][ch]);
+                    epDiffWeighted[ch] = (opposingOriginCh - m_origin[ch]) * channelWeights[ch];
+                }
+
+                MFloat lenSquared = epDiffWeighted[0] * epDiffWeighted[0];
+                for (int ch = 1; ch < TVectorSize; ch++)
+                    lenSquared = lenSquared + epDiffWeighted[ch] * epDiffWeighted[ch];
+
+                ParallelMath::MakeSafeDenominator(lenSquared);
+
+                MFloat maxValueDividedByLengthSquared = ParallelMath::MakeFloat(m_maxValue) / lenSquared;
+
+                for (int ch = 0; ch < TVectorSize; ch++)
+                    m_axis[ch] = epDiffWeighted[ch] * channelWeights[ch] * maxValueDividedByLengthSquared;
+            }
+
+            template<bool TSigned>
+            void Init(const float channelWeights[TVectorSize], const MUInt15 endPoints[2][TVectorSize], int range)
+            {
+                MAInt16 converted[2][TVectorSize];
+                for (int epi = 0; epi < 2; epi++)
+                    for (int ch = 0; ch < TVectorSize; ch++)
+                        converted[epi][ch] = ParallelMath::LosslessCast<MAInt16>::Cast(endPoints[epi][ch]);
+
+                Init<MUInt15, MUInt15>(channelWeights, endPoints, endPoints, range);
+            }
+
+            void ReconstructLDR_BC7(const MUInt15 &index, MUInt15* pixel, int numRealChannels)
+            {
+                MUInt15 weight = ParallelMath::LosslessCast<MUInt15>::Cast(ParallelMath::RightShift(ParallelMath::CompactMultiply(g_weightReciprocals[m_range], index) + 256, 9));
+
+                for (int ch = 0; ch < numRealChannels; ch++)
+                {
+                    MUInt15 ep0f = ParallelMath::LosslessCast<MUInt15>::Cast(ParallelMath::CompactMultiply((ParallelMath::MakeUInt15(64) - weight), ParallelMath::LosslessCast<MUInt15>::Cast(m_endPoint[0][ch])));
+                    MUInt15 ep1f = ParallelMath::LosslessCast<MUInt15>::Cast(ParallelMath::CompactMultiply(weight, ParallelMath::LosslessCast<MUInt15>::Cast(m_endPoint[1][ch])));
+                    pixel[ch] = ParallelMath::LosslessCast<MUInt15>::Cast(ParallelMath::RightShift(ep0f + ep1f + ParallelMath::MakeUInt15(32), 6));
+                }
+            }
+
+            void ReconstructLDRPrecise(const MUInt15 &index, MUInt15* pixel, int numRealChannels)
+            {
+                MUInt15 weight = ParallelMath::LosslessCast<MUInt15>::Cast(ParallelMath::RightShift(ParallelMath::CompactMultiply(g_weightReciprocals[m_range], index) + 64, 7));
+
+                for (int ch = 0; ch < numRealChannels; ch++)
+                {
+                    MUInt15 ep0f = ParallelMath::LosslessCast<MUInt15>::Cast(ParallelMath::CompactMultiply((ParallelMath::MakeUInt15(256) - weight), ParallelMath::LosslessCast<MUInt15>::Cast(m_endPoint[0][ch])));
+                    MUInt15 ep1f = ParallelMath::LosslessCast<MUInt15>::Cast(ParallelMath::CompactMultiply(weight, ParallelMath::LosslessCast<MUInt15>::Cast(m_endPoint[1][ch])));
+                    pixel[ch] = ParallelMath::LosslessCast<MUInt15>::Cast(ParallelMath::RightShift(ep0f + ep1f + ParallelMath::MakeUInt15(128), 8));
+                }
+            }
+
+            void ReconstructLDR_BC7(const MUInt15 &index, MUInt15* pixel)
+            {
+                ReconstructLDR_BC7(index, pixel, TVectorSize);
+            }
+
+            void ReconstructLDRPrecise(const MUInt15 &index, MUInt15* pixel)
+            {
+                ReconstructLDRPrecise(index, pixel, TVectorSize);
+            }
+
+            MUInt15 SelectIndexLDR(const MFloat* pixel, const ParallelMath::RoundTowardNearestForScope* rtn) const
+            {
+                MFloat dist = (pixel[0] - m_origin[0]) * m_axis[0];
+                for (int ch = 1; ch < TVectorSize; ch++)
+                    dist = dist + (pixel[ch] - m_origin[ch]) * m_axis[ch];
+
+                return ParallelMath::RoundAndConvertToU15(ParallelMath::Clamp(dist, 0.0f, m_maxValue), rtn);
+            }
+
+        protected:
+            MAInt16 m_endPoint[2][TVectorSize];
+
+        private:
+            MFloat m_origin[TVectorSize];
+            MFloat m_axis[TVectorSize];
+            int m_range;
+            float m_maxValue;
+            bool m_isUniform;
+        };
+    }
+}
+
+#endif
+