|
1 /* -*- Mode: C++; tab-width: 20; indent-tabs-mode: nil; c-basic-offset: 2 -*- |
|
2 * This Source Code Form is subject to the terms of the Mozilla Public |
|
3 * License, v. 2.0. If a copy of the MPL was not distributed with this |
|
4 * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ |
|
5 |
|
6 #include "2D.h" |
|
7 #include "Filters.h" |
|
8 #include "SIMD.h" |
|
9 |
|
10 namespace mozilla { |
|
11 namespace gfx { |
|
12 |
|
13 template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t> |
|
14 class SVGTurbulenceRenderer |
|
15 { |
|
16 public: |
|
17 SVGTurbulenceRenderer(const Size &aBaseFrequency, int32_t aSeed, |
|
18 int aNumOctaves, const Rect &aTileRect); |
|
19 |
|
20 TemporaryRef<DataSourceSurface> Render(const IntSize &aSize, const Point &aOffset) const; |
|
21 |
|
22 private: |
|
23 /* The turbulence calculation code is an adapted version of what |
|
24 appears in the SVG 1.1 specification: |
|
25 http://www.w3.org/TR/SVG11/filters.html#feTurbulence |
|
26 */ |
|
27 |
|
28 struct StitchInfo { |
|
29 int32_t width; // How much to subtract to wrap for stitching. |
|
30 int32_t height; |
|
31 int32_t wrapX; // Minimum value to wrap. |
|
32 int32_t wrapY; |
|
33 }; |
|
34 |
|
35 const static int sBSize = 0x100; |
|
36 const static int sBM = 0xff; |
|
37 void InitFromSeed(int32_t aSeed); |
|
38 void AdjustBaseFrequencyForStitch(const Rect &aTileRect); |
|
39 IntPoint AdjustForStitch(IntPoint aLatticePoint, const StitchInfo& aStitchInfo) const; |
|
40 StitchInfo CreateStitchInfo(const Rect &aTileRect) const; |
|
41 f32x4_t Noise2(Point aVec, const StitchInfo& aStitchInfo) const; |
|
42 i32x4_t Turbulence(const Point &aPoint) const; |
|
43 Point EquivalentNonNegativeOffset(const Point &aOffset) const; |
|
44 |
|
45 Size mBaseFrequency; |
|
46 int32_t mNumOctaves; |
|
47 StitchInfo mStitchInfo; |
|
48 bool mStitchable; |
|
49 TurbulenceType mType; |
|
50 uint8_t mLatticeSelector[sBSize]; |
|
51 f32x4_t mGradient[sBSize][2]; |
|
52 }; |
|
53 |
|
54 namespace { |
|
55 |
|
56 struct RandomNumberSource |
|
57 { |
|
58 RandomNumberSource(int32_t aSeed) : mLast(SetupSeed(aSeed)) {} |
|
59 int32_t Next() { mLast = Random(mLast); return mLast; } |
|
60 |
|
61 private: |
|
62 static const int32_t RAND_M = 2147483647; /* 2**31 - 1 */ |
|
63 static const int32_t RAND_A = 16807; /* 7**5; primitive root of m */ |
|
64 static const int32_t RAND_Q = 127773; /* m / a */ |
|
65 static const int32_t RAND_R = 2836; /* m % a */ |
|
66 |
|
67 /* Produces results in the range [1, 2**31 - 2]. |
|
68 Algorithm is: r = (a * r) mod m |
|
69 where a = 16807 and m = 2**31 - 1 = 2147483647 |
|
70 See [Park & Miller], CACM vol. 31 no. 10 p. 1195, Oct. 1988 |
|
71 To test: the algorithm should produce the result 1043618065 |
|
72 as the 10,000th generated number if the original seed is 1. |
|
73 */ |
|
74 |
|
75 static int32_t |
|
76 SetupSeed(int32_t aSeed) { |
|
77 if (aSeed <= 0) |
|
78 aSeed = -(aSeed % (RAND_M - 1)) + 1; |
|
79 if (aSeed > RAND_M - 1) |
|
80 aSeed = RAND_M - 1; |
|
81 return aSeed; |
|
82 } |
|
83 |
|
84 static int32_t |
|
85 Random(int32_t aSeed) |
|
86 { |
|
87 int32_t result = RAND_A * (aSeed % RAND_Q) - RAND_R * (aSeed / RAND_Q); |
|
88 if (result <= 0) |
|
89 result += RAND_M; |
|
90 return result; |
|
91 } |
|
92 |
|
93 int32_t mLast; |
|
94 }; |
|
95 |
|
96 } // unnamed namespace |
|
97 |
|
98 template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t> |
|
99 SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::SVGTurbulenceRenderer(const Size &aBaseFrequency, int32_t aSeed, |
|
100 int aNumOctaves, const Rect &aTileRect) |
|
101 : mBaseFrequency(aBaseFrequency) |
|
102 , mNumOctaves(aNumOctaves) |
|
103 { |
|
104 InitFromSeed(aSeed); |
|
105 if (Stitch) { |
|
106 AdjustBaseFrequencyForStitch(aTileRect); |
|
107 mStitchInfo = CreateStitchInfo(aTileRect); |
|
108 } |
|
109 } |
|
110 |
|
111 template<typename T> |
|
112 static void |
|
113 Swap(T& a, T& b) { |
|
114 T c = a; |
|
115 a = b; |
|
116 b = c; |
|
117 } |
|
118 |
|
119 template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t> |
|
120 void |
|
121 SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::InitFromSeed(int32_t aSeed) |
|
122 { |
|
123 RandomNumberSource rand(aSeed); |
|
124 |
|
125 float gradient[4][sBSize][2]; |
|
126 for (int32_t k = 0; k < 4; k++) { |
|
127 for (int32_t i = 0; i < sBSize; i++) { |
|
128 float a = float((rand.Next() % (sBSize + sBSize)) - sBSize) / sBSize; |
|
129 float b = float((rand.Next() % (sBSize + sBSize)) - sBSize) / sBSize; |
|
130 float s = sqrt(a * a + b * b); |
|
131 gradient[k][i][0] = a / s; |
|
132 gradient[k][i][1] = b / s; |
|
133 } |
|
134 } |
|
135 |
|
136 for (int32_t i = 0; i < sBSize; i++) { |
|
137 mLatticeSelector[i] = i; |
|
138 } |
|
139 for (int32_t i1 = sBSize - 1; i1 > 0; i1--) { |
|
140 int32_t i2 = rand.Next() % sBSize; |
|
141 Swap(mLatticeSelector[i1], mLatticeSelector[i2]); |
|
142 } |
|
143 |
|
144 for (int32_t i = 0; i < sBSize; i++) { |
|
145 // Contrary to the code in the spec, we build the first lattice selector |
|
146 // lookup into mGradient so that we don't need to do it again for every |
|
147 // pixel. |
|
148 // We also change the order of the gradient indexing so that we can process |
|
149 // all four color channels at the same time. |
|
150 uint8_t j = mLatticeSelector[i]; |
|
151 mGradient[i][0] = simd::FromF32<f32x4_t>(gradient[2][j][0], gradient[1][j][0], |
|
152 gradient[0][j][0], gradient[3][j][0]); |
|
153 mGradient[i][1] = simd::FromF32<f32x4_t>(gradient[2][j][1], gradient[1][j][1], |
|
154 gradient[0][j][1], gradient[3][j][1]); |
|
155 } |
|
156 } |
|
157 |
|
158 // Adjust aFreq such that aLength * AdjustForLength(aFreq, aLength) is integer |
|
159 // and as close to aLength * aFreq as possible. |
|
160 static inline float |
|
161 AdjustForLength(float aFreq, float aLength) |
|
162 { |
|
163 float lowFreq = floor(aLength * aFreq) / aLength; |
|
164 float hiFreq = ceil(aLength * aFreq) / aLength; |
|
165 if (aFreq / lowFreq < hiFreq / aFreq) { |
|
166 return lowFreq; |
|
167 } |
|
168 return hiFreq; |
|
169 } |
|
170 |
|
171 template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t> |
|
172 void |
|
173 SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::AdjustBaseFrequencyForStitch(const Rect &aTileRect) |
|
174 { |
|
175 mBaseFrequency = Size(AdjustForLength(mBaseFrequency.width, aTileRect.width), |
|
176 AdjustForLength(mBaseFrequency.height, aTileRect.height)); |
|
177 } |
|
178 |
|
179 template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t> |
|
180 typename SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::StitchInfo |
|
181 SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::CreateStitchInfo(const Rect &aTileRect) const |
|
182 { |
|
183 StitchInfo stitch; |
|
184 stitch.width = int32_t(floorf(aTileRect.width * mBaseFrequency.width + 0.5f)); |
|
185 stitch.height = int32_t(floorf(aTileRect.height * mBaseFrequency.height + 0.5f)); |
|
186 stitch.wrapX = int32_t(aTileRect.x * mBaseFrequency.width) + stitch.width; |
|
187 stitch.wrapY = int32_t(aTileRect.y * mBaseFrequency.height) + stitch.height; |
|
188 return stitch; |
|
189 } |
|
190 |
|
191 static MOZ_ALWAYS_INLINE Float |
|
192 SCurve(Float t) |
|
193 { |
|
194 return t * t * (3 - 2 * t); |
|
195 } |
|
196 |
|
197 static MOZ_ALWAYS_INLINE Point |
|
198 SCurve(Point t) |
|
199 { |
|
200 return Point(SCurve(t.x), SCurve(t.y)); |
|
201 } |
|
202 |
|
203 template<typename f32x4_t> |
|
204 static MOZ_ALWAYS_INLINE f32x4_t |
|
205 BiMix(const f32x4_t& aa, const f32x4_t& ab, |
|
206 const f32x4_t& ba, const f32x4_t& bb, Point s) |
|
207 { |
|
208 return simd::MixF32(simd::MixF32(aa, ab, s.x), |
|
209 simd::MixF32(ba, bb, s.x), s.y); |
|
210 } |
|
211 |
|
212 template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t> |
|
213 IntPoint |
|
214 SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::AdjustForStitch(IntPoint aLatticePoint, |
|
215 const StitchInfo& aStitchInfo) const |
|
216 { |
|
217 if (Stitch) { |
|
218 if (aLatticePoint.x >= aStitchInfo.wrapX) { |
|
219 aLatticePoint.x -= aStitchInfo.width; |
|
220 } |
|
221 if (aLatticePoint.y >= aStitchInfo.wrapY) { |
|
222 aLatticePoint.y -= aStitchInfo.height; |
|
223 } |
|
224 } |
|
225 return aLatticePoint; |
|
226 } |
|
227 |
|
228 template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t> |
|
229 f32x4_t |
|
230 SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::Noise2(Point aVec, const StitchInfo& aStitchInfo) const |
|
231 { |
|
232 // aVec is guaranteed to be non-negative, so casting to int32_t always |
|
233 // rounds towards negative infinity. |
|
234 IntPoint topLeftLatticePoint(int32_t(aVec.x), int32_t(aVec.y)); |
|
235 Point r = aVec - topLeftLatticePoint; // fractional offset |
|
236 |
|
237 IntPoint b0 = AdjustForStitch(topLeftLatticePoint, aStitchInfo); |
|
238 IntPoint b1 = AdjustForStitch(b0 + IntPoint(1, 1), aStitchInfo); |
|
239 |
|
240 uint8_t i = mLatticeSelector[b0.x & sBM]; |
|
241 uint8_t j = mLatticeSelector[b1.x & sBM]; |
|
242 |
|
243 const f32x4_t* qua = mGradient[(i + b0.y) & sBM]; |
|
244 const f32x4_t* qub = mGradient[(i + b1.y) & sBM]; |
|
245 const f32x4_t* qva = mGradient[(j + b0.y) & sBM]; |
|
246 const f32x4_t* qvb = mGradient[(j + b1.y) & sBM]; |
|
247 |
|
248 return BiMix(simd::WSumF32(qua[0], qua[1], r.x, r.y), |
|
249 simd::WSumF32(qva[0], qva[1], r.x - 1, r.y), |
|
250 simd::WSumF32(qub[0], qub[1], r.x, r.y - 1), |
|
251 simd::WSumF32(qvb[0], qvb[1], r.x - 1, r.y - 1), |
|
252 SCurve(r)); |
|
253 } |
|
254 |
|
255 template<typename f32x4_t, typename i32x4_t, typename u8x16_t> |
|
256 static inline i32x4_t |
|
257 ColorToBGRA(f32x4_t aUnscaledUnpreFloat) |
|
258 { |
|
259 // Color is an unpremultiplied float vector where 1.0f means white. We will |
|
260 // convert it into an integer vector where 255 means white. |
|
261 f32x4_t alpha = simd::SplatF32<3>(aUnscaledUnpreFloat); |
|
262 f32x4_t scaledUnpreFloat = simd::MulF32(aUnscaledUnpreFloat, simd::FromF32<f32x4_t>(255)); |
|
263 i32x4_t scaledUnpreInt = simd::F32ToI32(scaledUnpreFloat); |
|
264 |
|
265 // Multiply all channels with alpha. |
|
266 i32x4_t scaledPreInt = simd::F32ToI32(simd::MulF32(scaledUnpreFloat, alpha)); |
|
267 |
|
268 // Use the premultiplied color channels and the unpremultiplied alpha channel. |
|
269 i32x4_t alphaMask = simd::From32<i32x4_t>(0, 0, 0, -1); |
|
270 return simd::Pick(alphaMask, scaledPreInt, scaledUnpreInt); |
|
271 } |
|
272 |
|
273 template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t> |
|
274 i32x4_t |
|
275 SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::Turbulence(const Point &aPoint) const |
|
276 { |
|
277 StitchInfo stitchInfo = mStitchInfo; |
|
278 f32x4_t sum = simd::FromF32<f32x4_t>(0); |
|
279 Point vec(aPoint.x * mBaseFrequency.width, aPoint.y * mBaseFrequency.height); |
|
280 f32x4_t ratio = simd::FromF32<f32x4_t>(1); |
|
281 |
|
282 for (int octave = 0; octave < mNumOctaves; octave++) { |
|
283 f32x4_t thisOctave = Noise2(vec, stitchInfo); |
|
284 if (Type == TURBULENCE_TYPE_TURBULENCE) { |
|
285 thisOctave = simd::AbsF32(thisOctave); |
|
286 } |
|
287 sum = simd::AddF32(sum, simd::DivF32(thisOctave, ratio)); |
|
288 vec = vec * 2; |
|
289 ratio = simd::MulF32(ratio, simd::FromF32<f32x4_t>(2)); |
|
290 |
|
291 if (Stitch) { |
|
292 stitchInfo.width *= 2; |
|
293 stitchInfo.wrapX *= 2; |
|
294 stitchInfo.height *= 2; |
|
295 stitchInfo.wrapY *= 2; |
|
296 } |
|
297 } |
|
298 |
|
299 if (Type == TURBULENCE_TYPE_FRACTAL_NOISE) { |
|
300 sum = simd::DivF32(simd::AddF32(sum, simd::FromF32<f32x4_t>(1)), simd::FromF32<f32x4_t>(2)); |
|
301 } |
|
302 return ColorToBGRA<f32x4_t,i32x4_t,u8x16_t>(sum); |
|
303 } |
|
304 |
|
305 static inline Float |
|
306 MakeNonNegative(Float aValue, Float aIncrementSize) |
|
307 { |
|
308 if (aValue >= 0) { |
|
309 return aValue; |
|
310 } |
|
311 return aValue + ceilf(-aValue / aIncrementSize) * aIncrementSize; |
|
312 } |
|
313 |
|
314 template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t> |
|
315 Point |
|
316 SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::EquivalentNonNegativeOffset(const Point &aOffset) const |
|
317 { |
|
318 Size basePeriod = Stitch ? Size(mStitchInfo.width, mStitchInfo.height) : |
|
319 Size(sBSize, sBSize); |
|
320 Size repeatingSize(basePeriod.width / mBaseFrequency.width, |
|
321 basePeriod.height / mBaseFrequency.height); |
|
322 return Point(MakeNonNegative(aOffset.x, repeatingSize.width), |
|
323 MakeNonNegative(aOffset.y, repeatingSize.height)); |
|
324 } |
|
325 |
|
326 template<TurbulenceType Type, bool Stitch, typename f32x4_t, typename i32x4_t, typename u8x16_t> |
|
327 TemporaryRef<DataSourceSurface> |
|
328 SVGTurbulenceRenderer<Type,Stitch,f32x4_t,i32x4_t,u8x16_t>::Render(const IntSize &aSize, const Point &aOffset) const |
|
329 { |
|
330 RefPtr<DataSourceSurface> target = |
|
331 Factory::CreateDataSourceSurface(aSize, SurfaceFormat::B8G8R8A8); |
|
332 if (!target) { |
|
333 return nullptr; |
|
334 } |
|
335 |
|
336 uint8_t* targetData = target->GetData(); |
|
337 uint32_t stride = target->Stride(); |
|
338 |
|
339 Point startOffset = EquivalentNonNegativeOffset(aOffset); |
|
340 |
|
341 for (int32_t y = 0; y < aSize.height; y++) { |
|
342 for (int32_t x = 0; x < aSize.width; x += 4) { |
|
343 int32_t targIndex = y * stride + x * 4; |
|
344 i32x4_t a = Turbulence(startOffset + Point(x, y)); |
|
345 i32x4_t b = Turbulence(startOffset + Point(x + 1, y)); |
|
346 i32x4_t c = Turbulence(startOffset + Point(x + 2, y)); |
|
347 i32x4_t d = Turbulence(startOffset + Point(x + 3, y)); |
|
348 u8x16_t result1234 = simd::PackAndSaturate32To8(a, b, c, d); |
|
349 simd::Store8(&targetData[targIndex], result1234); |
|
350 } |
|
351 } |
|
352 |
|
353 return target; |
|
354 } |
|
355 |
|
356 } // namespace gfx |
|
357 } // namespace mozilla |