michael@0: /* -*- Mode: C++; tab-width: 20; indent-tabs-mode: nil; c-basic-offset: 2 -*-
michael@0:  * This Source Code Form is subject to the terms of the Mozilla Public
michael@0:  * License, v. 2.0. If a copy of the MPL was not distributed with this
michael@0:  * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
michael@0: 
michael@0: #include "2D.h"
michael@0: #include "Filters.h"
michael@0: #include "SIMD.h"
michael@0: 
michael@0: namespace mozilla {
michael@0: namespace gfx {
michael@0: 
michael@0: template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t>
michael@0: class SVGTurbulenceRenderer
michael@0: {
michael@0: public:
michael@0:   SVGTurbulenceRenderer(const Size &aBaseFrequency, int32_t aSeed,
michael@0:                         int aNumOctaves, const Rect &aTileRect);
michael@0: 
michael@0:   TemporaryRef<DataSourceSurface> Render(const IntSize &aSize, const Point &aOffset) const;
michael@0: 
michael@0: private:
michael@0:   /* The turbulence calculation code is an adapted version of what
michael@0:      appears in the SVG 1.1 specification:
michael@0:          http://www.w3.org/TR/SVG11/filters.html#feTurbulence
michael@0:   */
michael@0: 
michael@0:   struct StitchInfo {
michael@0:     int32_t width;     // How much to subtract to wrap for stitching.
michael@0:     int32_t height;
michael@0:     int32_t wrapX;     // Minimum value to wrap.
michael@0:     int32_t wrapY;
michael@0:   };
michael@0: 
michael@0:   const static int sBSize = 0x100;
michael@0:   const static int sBM = 0xff;
michael@0:   void InitFromSeed(int32_t aSeed);
michael@0:   void AdjustBaseFrequencyForStitch(const Rect &aTileRect);
michael@0:   IntPoint AdjustForStitch(IntPoint aLatticePoint, const StitchInfo& aStitchInfo) const;
michael@0:   StitchInfo CreateStitchInfo(const Rect &aTileRect) const;
michael@0:   f32x4_t Noise2(Point aVec, const StitchInfo& aStitchInfo) const;
michael@0:   i32x4_t Turbulence(const Point &aPoint) const;
michael@0:   Point EquivalentNonNegativeOffset(const Point &aOffset) const;
michael@0: 
michael@0:   Size mBaseFrequency;
michael@0:   int32_t mNumOctaves;
michael@0:   StitchInfo mStitchInfo;
michael@0:   bool mStitchable;
michael@0:   TurbulenceType mType;
michael@0:   uint8_t mLatticeSelector[sBSize];
michael@0:   f32x4_t mGradient[sBSize][2];
michael@0: };
michael@0: 
michael@0: namespace {
michael@0: 
michael@0: struct RandomNumberSource
michael@0: {
michael@0:   RandomNumberSource(int32_t aSeed) : mLast(SetupSeed(aSeed)) {}
michael@0:   int32_t Next() { mLast = Random(mLast); return mLast; }
michael@0: 
michael@0: private:
michael@0:   static const int32_t RAND_M = 2147483647; /* 2**31 - 1 */
michael@0:   static const int32_t RAND_A = 16807;      /* 7**5; primitive root of m */
michael@0:   static const int32_t RAND_Q = 127773;     /* m / a */
michael@0:   static const int32_t RAND_R = 2836;       /* m % a */
michael@0: 
michael@0:   /* Produces results in the range [1, 2**31 - 2].
michael@0:      Algorithm is: r = (a * r) mod m
michael@0:      where a = 16807 and m = 2**31 - 1 = 2147483647
michael@0:      See [Park & Miller], CACM vol. 31 no. 10 p. 1195, Oct. 1988
michael@0:      To test: the algorithm should produce the result 1043618065
michael@0:      as the 10,000th generated number if the original seed is 1.
michael@0:   */
michael@0: 
michael@0:   static int32_t
michael@0:   SetupSeed(int32_t aSeed) {
michael@0:     if (aSeed <= 0)
michael@0:       aSeed = -(aSeed % (RAND_M - 1)) + 1;
michael@0:     if (aSeed > RAND_M - 1)
michael@0:       aSeed = RAND_M - 1;
michael@0:     return aSeed;
michael@0:   }
michael@0: 
michael@0:   static int32_t
michael@0:   Random(int32_t aSeed)
michael@0:   {
michael@0:     int32_t result = RAND_A * (aSeed % RAND_Q) - RAND_R * (aSeed / RAND_Q);
michael@0:     if (result <= 0)
michael@0:       result += RAND_M;
michael@0:     return result;
michael@0:   }
michael@0: 
michael@0:   int32_t mLast;
michael@0: };
michael@0: 
michael@0: } // unnamed namespace
michael@0: 
michael@0: template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t>
michael@0: SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::SVGTurbulenceRenderer(const Size &aBaseFrequency, int32_t aSeed,
michael@0:                                                             int aNumOctaves, const Rect &aTileRect)
michael@0:  : mBaseFrequency(aBaseFrequency)
michael@0:  , mNumOctaves(aNumOctaves)
michael@0: {
michael@0:   InitFromSeed(aSeed);
michael@0:   if (Stitch) {
michael@0:     AdjustBaseFrequencyForStitch(aTileRect);
michael@0:     mStitchInfo = CreateStitchInfo(aTileRect);
michael@0:   }
michael@0: }
michael@0: 
michael@0: template<typename T>
michael@0: static void
michael@0: Swap(T& a, T& b) {
michael@0:   T c = a;
michael@0:   a = b;
michael@0:   b = c;
michael@0: }
michael@0: 
michael@0: template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t>
michael@0: void
michael@0: SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::InitFromSeed(int32_t aSeed)
michael@0: {
michael@0:   RandomNumberSource rand(aSeed);
michael@0: 
michael@0:   float gradient[4][sBSize][2];
michael@0:   for (int32_t k = 0; k < 4; k++) {
michael@0:     for (int32_t i = 0; i < sBSize; i++) {
michael@0:       float a = float((rand.Next() % (sBSize + sBSize)) - sBSize) / sBSize;
michael@0:       float b = float((rand.Next() % (sBSize + sBSize)) - sBSize) / sBSize;
michael@0:       float s = sqrt(a * a + b * b);
michael@0:       gradient[k][i][0] = a / s;
michael@0:       gradient[k][i][1] = b / s;
michael@0:     }
michael@0:   }
michael@0: 
michael@0:   for (int32_t i = 0; i < sBSize; i++) {
michael@0:     mLatticeSelector[i] = i;
michael@0:   }
michael@0:   for (int32_t i1 = sBSize - 1; i1 > 0; i1--) {
michael@0:     int32_t i2 = rand.Next() % sBSize;
michael@0:     Swap(mLatticeSelector[i1], mLatticeSelector[i2]);
michael@0:   }
michael@0: 
michael@0:   for (int32_t i = 0; i < sBSize; i++) {
michael@0:     // Contrary to the code in the spec, we build the first lattice selector
michael@0:     // lookup into mGradient so that we don't need to do it again for every
michael@0:     // pixel.
michael@0:     // We also change the order of the gradient indexing so that we can process
michael@0:     // all four color channels at the same time.
michael@0:     uint8_t j = mLatticeSelector[i];
michael@0:     mGradient[i][0] = simd::FromF32<f32x4_t>(gradient[2][j][0], gradient[1][j][0],
michael@0:                                              gradient[0][j][0], gradient[3][j][0]);
michael@0:     mGradient[i][1] = simd::FromF32<f32x4_t>(gradient[2][j][1], gradient[1][j][1],
michael@0:                                              gradient[0][j][1], gradient[3][j][1]);
michael@0:   }
michael@0: }
michael@0: 
michael@0: // Adjust aFreq such that aLength * AdjustForLength(aFreq, aLength) is integer
michael@0: // and as close to aLength * aFreq as possible.
michael@0: static inline float
michael@0: AdjustForLength(float aFreq, float aLength)
michael@0: {
michael@0:   float lowFreq = floor(aLength * aFreq) / aLength;
michael@0:   float hiFreq = ceil(aLength * aFreq) / aLength;
michael@0:   if (aFreq / lowFreq < hiFreq / aFreq) {
michael@0:     return lowFreq;
michael@0:   }
michael@0:   return hiFreq;
michael@0: }
michael@0: 
michael@0: template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t>
michael@0: void
michael@0: SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::AdjustBaseFrequencyForStitch(const Rect &aTileRect)
michael@0: {
michael@0:   mBaseFrequency = Size(AdjustForLength(mBaseFrequency.width, aTileRect.width),
michael@0:                         AdjustForLength(mBaseFrequency.height, aTileRect.height));
michael@0: }
michael@0: 
michael@0: template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t>
michael@0: typename SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::StitchInfo
michael@0: SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::CreateStitchInfo(const Rect &aTileRect) const
michael@0: {
michael@0:   StitchInfo stitch;
michael@0:   stitch.width = int32_t(floorf(aTileRect.width * mBaseFrequency.width + 0.5f));
michael@0:   stitch.height = int32_t(floorf(aTileRect.height * mBaseFrequency.height + 0.5f));
michael@0:   stitch.wrapX = int32_t(aTileRect.x * mBaseFrequency.width) + stitch.width;
michael@0:   stitch.wrapY = int32_t(aTileRect.y * mBaseFrequency.height) + stitch.height;
michael@0:   return stitch;
michael@0: }
michael@0: 
michael@0: static MOZ_ALWAYS_INLINE Float
michael@0: SCurve(Float t)
michael@0: {
michael@0:   return t * t * (3 - 2 * t);
michael@0: }
michael@0: 
michael@0: static MOZ_ALWAYS_INLINE Point
michael@0: SCurve(Point t)
michael@0: {
michael@0:   return Point(SCurve(t.x), SCurve(t.y));
michael@0: }
michael@0: 
michael@0: template<typename f32x4_t>
michael@0: static MOZ_ALWAYS_INLINE f32x4_t
michael@0: BiMix(const f32x4_t& aa, const f32x4_t& ab,
michael@0:       const f32x4_t& ba, const f32x4_t& bb, Point s)
michael@0: {
michael@0:   return simd::MixF32(simd::MixF32(aa, ab, s.x),
michael@0:                       simd::MixF32(ba, bb, s.x), s.y);
michael@0: }
michael@0: 
michael@0: template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t>
michael@0: IntPoint
michael@0: SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::AdjustForStitch(IntPoint aLatticePoint,
michael@0:                                                       const StitchInfo& aStitchInfo) const
michael@0: {
michael@0:   if (Stitch) {
michael@0:     if (aLatticePoint.x >= aStitchInfo.wrapX) {
michael@0:       aLatticePoint.x -= aStitchInfo.width;
michael@0:     }
michael@0:     if (aLatticePoint.y >= aStitchInfo.wrapY) {
michael@0:       aLatticePoint.y -= aStitchInfo.height;
michael@0:     }
michael@0:   }
michael@0:   return aLatticePoint;
michael@0: }
michael@0: 
michael@0: template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t>
michael@0: f32x4_t
michael@0: SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::Noise2(Point aVec, const StitchInfo& aStitchInfo) const
michael@0: {
michael@0:   // aVec is guaranteed to be non-negative, so casting to int32_t always
michael@0:   // rounds towards negative infinity.
michael@0:   IntPoint topLeftLatticePoint(int32_t(aVec.x), int32_t(aVec.y));
michael@0:   Point r = aVec - topLeftLatticePoint; // fractional offset
michael@0: 
michael@0:   IntPoint b0 = AdjustForStitch(topLeftLatticePoint, aStitchInfo);
michael@0:   IntPoint b1 = AdjustForStitch(b0 + IntPoint(1, 1), aStitchInfo);
michael@0: 
michael@0:   uint8_t i = mLatticeSelector[b0.x & sBM];
michael@0:   uint8_t j = mLatticeSelector[b1.x & sBM];
michael@0: 
michael@0:   const f32x4_t* qua = mGradient[(i + b0.y) & sBM];
michael@0:   const f32x4_t* qub = mGradient[(i + b1.y) & sBM];
michael@0:   const f32x4_t* qva = mGradient[(j + b0.y) & sBM];
michael@0:   const f32x4_t* qvb = mGradient[(j + b1.y) & sBM];
michael@0: 
michael@0:   return BiMix(simd::WSumF32(qua[0], qua[1], r.x, r.y),
michael@0:                simd::WSumF32(qva[0], qva[1], r.x - 1, r.y),
michael@0:                simd::WSumF32(qub[0], qub[1], r.x, r.y - 1),
michael@0:                simd::WSumF32(qvb[0], qvb[1], r.x - 1, r.y - 1),
michael@0:                SCurve(r));
michael@0: }
michael@0: 
michael@0: template<typename f32x4_t, typename i32x4_t, typename u8x16_t>
michael@0: static inline i32x4_t
michael@0: ColorToBGRA(f32x4_t aUnscaledUnpreFloat)
michael@0: {
michael@0:   // Color is an unpremultiplied float vector where 1.0f means white. We will
michael@0:   // convert it into an integer vector where 255 means white.
michael@0:   f32x4_t alpha = simd::SplatF32<3>(aUnscaledUnpreFloat);
michael@0:   f32x4_t scaledUnpreFloat = simd::MulF32(aUnscaledUnpreFloat, simd::FromF32<f32x4_t>(255));
michael@0:   i32x4_t scaledUnpreInt = simd::F32ToI32(scaledUnpreFloat);
michael@0: 
michael@0:   // Multiply all channels with alpha.
michael@0:   i32x4_t scaledPreInt = simd::F32ToI32(simd::MulF32(scaledUnpreFloat, alpha));
michael@0: 
michael@0:   // Use the premultiplied color channels and the unpremultiplied alpha channel.
michael@0:   i32x4_t alphaMask = simd::From32<i32x4_t>(0, 0, 0, -1);
michael@0:   return simd::Pick(alphaMask, scaledPreInt, scaledUnpreInt);
michael@0: }
michael@0: 
michael@0: template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t>
michael@0: i32x4_t
michael@0: SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::Turbulence(const Point &aPoint) const
michael@0: {
michael@0:   StitchInfo stitchInfo = mStitchInfo;
michael@0:   f32x4_t sum = simd::FromF32<f32x4_t>(0);
michael@0:   Point vec(aPoint.x * mBaseFrequency.width, aPoint.y * mBaseFrequency.height);
michael@0:   f32x4_t ratio = simd::FromF32<f32x4_t>(1);
michael@0: 
michael@0:   for (int octave = 0; octave < mNumOctaves; octave++) {
michael@0:     f32x4_t thisOctave = Noise2(vec, stitchInfo);
michael@0:     if (Type == TURBULENCE_TYPE_TURBULENCE) {
michael@0:       thisOctave = simd::AbsF32(thisOctave);
michael@0:     }
michael@0:     sum = simd::AddF32(sum, simd::DivF32(thisOctave, ratio));
michael@0:     vec = vec * 2;
michael@0:     ratio = simd::MulF32(ratio, simd::FromF32<f32x4_t>(2));
michael@0: 
michael@0:     if (Stitch) {
michael@0:       stitchInfo.width *= 2;
michael@0:       stitchInfo.wrapX *= 2;
michael@0:       stitchInfo.height *= 2;
michael@0:       stitchInfo.wrapY *= 2;
michael@0:     }
michael@0:   }
michael@0: 
michael@0:   if (Type == TURBULENCE_TYPE_FRACTAL_NOISE) {
michael@0:     sum = simd::DivF32(simd::AddF32(sum, simd::FromF32<f32x4_t>(1)), simd::FromF32<f32x4_t>(2));
michael@0:   }
michael@0:   return ColorToBGRA<f32x4_t,i32x4_t,u8x16_t>(sum);
michael@0: }
michael@0: 
michael@0: static inline Float
michael@0: MakeNonNegative(Float aValue, Float aIncrementSize)
michael@0: {
michael@0:   if (aValue >= 0) {
michael@0:     return aValue;
michael@0:   }
michael@0:   return aValue + ceilf(-aValue / aIncrementSize) * aIncrementSize;
michael@0: }
michael@0: 
michael@0: template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t>
michael@0: Point
michael@0: SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::EquivalentNonNegativeOffset(const Point &aOffset) const
michael@0: {
michael@0:   Size basePeriod = Stitch ? Size(mStitchInfo.width, mStitchInfo.height) :
michael@0:                              Size(sBSize, sBSize);
michael@0:   Size repeatingSize(basePeriod.width / mBaseFrequency.width,
michael@0:                      basePeriod.height / mBaseFrequency.height);
michael@0:   return Point(MakeNonNegative(aOffset.x, repeatingSize.width),
michael@0:                MakeNonNegative(aOffset.y, repeatingSize.height));
michael@0: }
michael@0: 
michael@0: template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t>
michael@0: TemporaryRef<DataSourceSurface>
michael@0: SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::Render(const IntSize &aSize, const Point &aOffset) const
michael@0: {
michael@0:   RefPtr<DataSourceSurface> target =
michael@0:     Factory::CreateDataSourceSurface(aSize, SurfaceFormat::B8G8R8A8);
michael@0:   if (!target) {
michael@0:     return nullptr;
michael@0:   }
michael@0: 
michael@0:   uint8_t* targetData = target->GetData();
michael@0:   uint32_t stride = target->Stride();
michael@0: 
michael@0:   Point startOffset = EquivalentNonNegativeOffset(aOffset);
michael@0: 
michael@0:   for (int32_t y = 0; y < aSize.height; y++) {
michael@0:     for (int32_t x = 0; x < aSize.width; x += 4) {
michael@0:       int32_t targIndex = y * stride + x * 4;
michael@0:       i32x4_t a = Turbulence(startOffset + Point(x, y));
michael@0:       i32x4_t b = Turbulence(startOffset + Point(x + 1, y));
michael@0:       i32x4_t c = Turbulence(startOffset + Point(x + 2, y));
michael@0:       i32x4_t d = Turbulence(startOffset + Point(x + 3, y));
michael@0:       u8x16_t result1234 = simd::PackAndSaturate32To8(a, b, c, d);
michael@0:       simd::Store8(&targetData[targIndex], result1234);
michael@0:     }
michael@0:   }
michael@0: 
michael@0:   return target;
michael@0: }
michael@0: 
michael@0: } // namespace gfx
michael@0: } // namespace mozilla