1 | /* |
2 | * Copyright 2016 Google Inc. |
3 | * |
4 | * Use of this source code is governed by a BSD-style license that can be |
5 | * found in the LICENSE file. |
6 | */ |
7 | |
8 | #include "include/core/SkPaint.h" |
9 | #include "src/shaders/gradients/Sk4fLinearGradient.h" |
10 | |
11 | #include <cmath> |
12 | #include <utility> |
13 | |
14 | namespace { |
15 | |
16 | template<ApplyPremul premul> |
17 | void ramp(const Sk4f& c, const Sk4f& dc, SkPMColor dst[], int n, |
18 | const Sk4f& bias0, const Sk4f& bias1) { |
19 | SkASSERT(n > 0); |
20 | |
21 | const Sk4f dc2 = dc + dc, |
22 | dc4 = dc2 + dc2; |
23 | |
24 | Sk4f c0 = c + DstTraits<premul>::pre_lerp_bias(bias0), |
25 | c1 = c + dc + DstTraits<premul>::pre_lerp_bias(bias1), |
26 | c2 = c0 + dc2, |
27 | c3 = c1 + dc2; |
28 | |
29 | while (n >= 4) { |
30 | DstTraits<premul>::store4x(c0, c1, c2, c3, dst, bias0, bias1); |
31 | dst += 4; |
32 | |
33 | c0 = c0 + dc4; |
34 | c1 = c1 + dc4; |
35 | c2 = c2 + dc4; |
36 | c3 = c3 + dc4; |
37 | n -= 4; |
38 | } |
39 | if (n & 2) { |
40 | DstTraits<premul>::store(c0, dst++, bias0); |
41 | DstTraits<premul>::store(c1, dst++, bias1); |
42 | c0 = c0 + dc2; |
43 | } |
44 | if (n & 1) { |
45 | DstTraits<premul>::store(c0, dst, bias0); |
46 | } |
47 | } |
48 | |
49 | template<SkTileMode> |
50 | SkScalar pinFx(SkScalar); |
51 | |
52 | template<> |
53 | SkScalar pinFx<SkTileMode::kClamp>(SkScalar fx) { |
54 | return fx; |
55 | } |
56 | |
57 | template<> |
58 | SkScalar pinFx<SkTileMode::kRepeat>(SkScalar fx) { |
59 | SkScalar f = SkScalarIsFinite(fx) ? SkScalarFraction(fx) : 0; |
60 | if (f < 0) { |
61 | f = std::min(f + 1, nextafterf(1, 0)); |
62 | } |
63 | SkASSERT(f >= 0); |
64 | SkASSERT(f < 1.0f); |
65 | return f; |
66 | } |
67 | |
68 | template<> |
69 | SkScalar pinFx<SkTileMode::kMirror>(SkScalar fx) { |
70 | SkScalar f = SkScalarIsFinite(fx) ? SkScalarMod(fx, 2.0f) : 0; |
71 | if (f < 0) { |
72 | f = std::min(f + 2, nextafterf(2, 0)); |
73 | } |
74 | SkASSERT(f >= 0); |
75 | SkASSERT(f < 2.0f); |
76 | return f; |
77 | } |
78 | |
79 | // true when x is in [k1,k2], or [k2, k1] when the interval is reversed. |
80 | // TODO(fmalita): hoist the reversed interval check out of this helper. |
81 | bool in_range(SkScalar x, SkScalar k1, SkScalar k2) { |
82 | SkASSERT(k1 != k2); |
83 | return (k1 < k2) |
84 | ? (x >= k1 && x <= k2) |
85 | : (x >= k2 && x <= k1); |
86 | } |
87 | |
88 | } // anonymous namespace |
89 | |
90 | SkLinearGradient:: |
91 | LinearGradient4fContext::LinearGradient4fContext(const SkLinearGradient& shader, |
92 | const ContextRec& rec) |
93 | : INHERITED(shader, rec) { |
94 | |
95 | // Our fast path expects interval points to be monotonically increasing in x. |
96 | const bool reverseIntervals = std::signbit(fDstToPos.getScaleX()); |
97 | fIntervals.init(shader, rec.fDstColorSpace, shader.fTileMode, |
98 | fColorsArePremul, rec.fPaint->getAlpha() * (1.0f / 255), reverseIntervals); |
99 | |
100 | SkASSERT(fIntervals->count() > 0); |
101 | fCachedInterval = fIntervals->begin(); |
102 | } |
103 | |
104 | const Sk4fGradientInterval* |
105 | SkLinearGradient::LinearGradient4fContext::findInterval(SkScalar fx) const { |
106 | SkASSERT(in_range(fx, fIntervals->front().fT0, fIntervals->back().fT1)); |
107 | |
108 | if (1) { |
109 | // Linear search, using the last scanline interval as a starting point. |
110 | SkASSERT(fCachedInterval >= fIntervals->begin()); |
111 | SkASSERT(fCachedInterval < fIntervals->end()); |
112 | const int search_dir = fDstToPos.getScaleX() >= 0 ? 1 : -1; |
113 | while (!in_range(fx, fCachedInterval->fT0, fCachedInterval->fT1)) { |
114 | fCachedInterval += search_dir; |
115 | if (fCachedInterval >= fIntervals->end()) { |
116 | fCachedInterval = fIntervals->begin(); |
117 | } else if (fCachedInterval < fIntervals->begin()) { |
118 | fCachedInterval = fIntervals->end() - 1; |
119 | } |
120 | } |
121 | return fCachedInterval; |
122 | } else { |
123 | // Binary search. Seems less effective than linear + caching. |
124 | const auto* i0 = fIntervals->begin(); |
125 | const auto* i1 = fIntervals->end() - 1; |
126 | |
127 | while (i0 != i1) { |
128 | SkASSERT(i0 < i1); |
129 | SkASSERT(in_range(fx, i0->fT0, i1->fT1)); |
130 | |
131 | const auto* i = i0 + ((i1 - i0) >> 1); |
132 | |
133 | if (in_range(fx, i0->fT0, i->fT1)) { |
134 | i1 = i; |
135 | } else { |
136 | SkASSERT(in_range(fx, i->fT1, i1->fT1)); |
137 | i0 = i + 1; |
138 | } |
139 | } |
140 | |
141 | SkASSERT(in_range(fx, i0->fT0, i0->fT1)); |
142 | return i0; |
143 | } |
144 | } |
145 | |
146 | |
147 | void SkLinearGradient:: |
148 | LinearGradient4fContext::shadeSpan(int x, int y, SkPMColor dst[], int count) { |
149 | SkASSERT(count > 0); |
150 | |
151 | float bias0 = 0, |
152 | bias1 = 0; |
153 | |
154 | if (fDither) { |
155 | static constexpr float dither_cell[] = { |
156 | -3/8.0f, 1/8.0f, |
157 | 3/8.0f, -1/8.0f, |
158 | }; |
159 | |
160 | const int rowIndex = (y & 1) << 1; |
161 | bias0 = dither_cell[rowIndex + 0]; |
162 | bias1 = dither_cell[rowIndex + 1]; |
163 | |
164 | if (x & 1) { |
165 | using std::swap; |
166 | swap(bias0, bias1); |
167 | } |
168 | } |
169 | |
170 | if (fColorsArePremul) { |
171 | // In premul interpolation mode, components are pre-scaled by 255 and the store |
172 | // op is truncating. We pre-bias here to achieve rounding. |
173 | bias0 += 0.5f; |
174 | bias1 += 0.5f; |
175 | |
176 | this->shadePremulSpan<ApplyPremul::False>(x, y, dst, count, bias0, bias1); |
177 | } else { |
178 | // In unpremul interpolation mode, Components are not pre-scaled. |
179 | bias0 *= 1/255.0f; |
180 | bias1 *= 1/255.0f; |
181 | |
182 | this->shadePremulSpan<ApplyPremul::True >(x, y, dst, count, bias0, bias1); |
183 | } |
184 | } |
185 | |
186 | template<ApplyPremul premul> |
187 | void SkLinearGradient:: |
188 | LinearGradient4fContext::shadePremulSpan(int x, int y, SkPMColor dst[], int count, |
189 | float bias0, float bias1) const { |
190 | const SkLinearGradient& shader = static_cast<const SkLinearGradient&>(fShader); |
191 | switch (shader.fTileMode) { |
192 | case SkTileMode::kDecal: |
193 | SkASSERT(false); // decal only supported via stages |
194 | // fall-through |
195 | case SkTileMode::kClamp: |
196 | this->shadeSpanInternal<premul, SkTileMode::kClamp >(x, y, dst, count, bias0, bias1); |
197 | break; |
198 | case SkTileMode::kRepeat: |
199 | this->shadeSpanInternal<premul, SkTileMode::kRepeat>(x, y, dst, count, bias0, bias1); |
200 | break; |
201 | case SkTileMode::kMirror: |
202 | this->shadeSpanInternal<premul, SkTileMode::kMirror>(x, y, dst, count, bias0, bias1); |
203 | break; |
204 | } |
205 | } |
206 | |
207 | template<ApplyPremul premul, SkTileMode tileMode> |
208 | void SkLinearGradient:: |
209 | LinearGradient4fContext::shadeSpanInternal(int x, int y, SkPMColor dst[], int count, |
210 | float bias0, float bias1) const { |
211 | SkPoint pt; |
212 | fDstToPosProc(fDstToPos, |
213 | x + SK_ScalarHalf, |
214 | y + SK_ScalarHalf, |
215 | &pt); |
216 | const SkScalar fx = pinFx<tileMode>(pt.x()); |
217 | const SkScalar dx = fDstToPos.getScaleX(); |
218 | LinearIntervalProcessor<premul, tileMode> proc(fIntervals->begin(), |
219 | fIntervals->end() - 1, |
220 | this->findInterval(fx), |
221 | fx, |
222 | dx, |
223 | SkScalarNearlyZero(dx * count)); |
224 | Sk4f bias4f0(bias0), |
225 | bias4f1(bias1); |
226 | |
227 | while (count > 0) { |
228 | // What we really want here is SkTPin(advance, 1, count) |
229 | // but that's a significant perf hit for >> stops; investigate. |
230 | const int n = std::min(SkScalarTruncToInt(proc.currentAdvance() + 1), count); |
231 | |
232 | // The current interval advance can be +inf (e.g. when reaching |
233 | // the clamp mode end intervals) - when that happens, we expect to |
234 | // a) consume all remaining count in one swoop |
235 | // b) return a zero color gradient |
236 | SkASSERT(SkScalarIsFinite(proc.currentAdvance()) |
237 | || (n == count && proc.currentRampIsZero())); |
238 | |
239 | if (proc.currentRampIsZero()) { |
240 | DstTraits<premul>::store(proc.currentColor(), dst, n); |
241 | } else { |
242 | ramp<premul>(proc.currentColor(), proc.currentColorGrad(), dst, n, |
243 | bias4f0, bias4f1); |
244 | } |
245 | |
246 | proc.advance(SkIntToScalar(n)); |
247 | count -= n; |
248 | dst += n; |
249 | |
250 | if (n & 1) { |
251 | using std::swap; |
252 | swap(bias4f0, bias4f1); |
253 | } |
254 | } |
255 | } |
256 | |
257 | template<ApplyPremul premul, SkTileMode tileMode> |
258 | class SkLinearGradient:: |
259 | LinearGradient4fContext::LinearIntervalProcessor { |
260 | public: |
261 | LinearIntervalProcessor(const Sk4fGradientInterval* firstInterval, |
262 | const Sk4fGradientInterval* lastInterval, |
263 | const Sk4fGradientInterval* i, |
264 | SkScalar fx, |
265 | SkScalar dx, |
266 | bool is_vertical) |
267 | : fAdvX(is_vertical ? SK_ScalarInfinity : (i->fT1 - fx) / dx) |
268 | , fFirstInterval(firstInterval) |
269 | , fLastInterval(lastInterval) |
270 | , fInterval(i) |
271 | , fDx(dx) |
272 | , fIsVertical(is_vertical) |
273 | { |
274 | SkASSERT(fAdvX >= 0); |
275 | SkASSERT(firstInterval <= lastInterval); |
276 | |
277 | if (tileMode != SkTileMode::kClamp && !is_vertical) { |
278 | const auto spanX = (lastInterval->fT1 - firstInterval->fT0) / dx; |
279 | SkASSERT(spanX >= 0); |
280 | |
281 | // If we're in a repeating tile mode and the whole gradient is compressed into a |
282 | // fraction of a pixel, we just use the average color in zero-ramp mode. |
283 | // This also avoids cases where we make no progress due to interval advances being |
284 | // close to zero. |
285 | static constexpr SkScalar kMinSpanX = .25f; |
286 | if (spanX < kMinSpanX) { |
287 | this->init_average_props(); |
288 | return; |
289 | } |
290 | } |
291 | |
292 | this->compute_interval_props(fx); |
293 | } |
294 | |
295 | SkScalar currentAdvance() const { |
296 | SkASSERT(fAdvX >= 0); |
297 | SkASSERT(!std::isfinite(fAdvX) || fAdvX <= (fInterval->fT1 - fInterval->fT0) / fDx); |
298 | return fAdvX; |
299 | } |
300 | |
301 | bool currentRampIsZero() const { return fZeroRamp; } |
302 | const Sk4f& currentColor() const { return fCc; } |
303 | const Sk4f& currentColorGrad() const { return fDcDx; } |
304 | |
305 | void advance(SkScalar advX) { |
306 | SkASSERT(advX > 0); |
307 | SkASSERT(fAdvX >= 0); |
308 | |
309 | if (advX >= fAdvX) { |
310 | advX = this->advance_interval(advX); |
311 | } |
312 | SkASSERT(advX < fAdvX); |
313 | |
314 | fCc = fCc + fDcDx * Sk4f(advX); |
315 | fAdvX -= advX; |
316 | } |
317 | |
318 | private: |
319 | void compute_interval_props(SkScalar t) { |
320 | SkASSERT(in_range(t, fInterval->fT0, fInterval->fT1)); |
321 | |
322 | const Sk4f dc = DstTraits<premul>::load(fInterval->fCg); |
323 | fCc = DstTraits<premul>::load(fInterval->fCb) + dc * Sk4f(t); |
324 | fDcDx = dc * fDx; |
325 | fZeroRamp = fIsVertical || (dc == 0).allTrue(); |
326 | } |
327 | |
328 | void init_average_props() { |
329 | fAdvX = SK_ScalarInfinity; |
330 | fZeroRamp = true; |
331 | fDcDx = 0; |
332 | fCc = Sk4f(0); |
333 | |
334 | // TODO: precompute the average at interval setup time? |
335 | for (const auto* i = fFirstInterval; i <= fLastInterval; ++i) { |
336 | // Each interval contributes its average color to the total/weighted average: |
337 | // |
338 | // C = (c0 + c1) / 2 = (Cb + Cg * t0 + Cb + Cg * t1) / 2 = Cb + Cg *(t0 + t1) / 2 |
339 | // |
340 | // Avg += C * (t1 - t0) |
341 | // |
342 | const auto c = DstTraits<premul>::load(i->fCb) |
343 | + DstTraits<premul>::load(i->fCg) * (i->fT0 + i->fT1) * 0.5f; |
344 | fCc = fCc + c * (i->fT1 - i->fT0); |
345 | } |
346 | } |
347 | |
348 | const Sk4fGradientInterval* next_interval(const Sk4fGradientInterval* i) const { |
349 | SkASSERT(i >= fFirstInterval); |
350 | SkASSERT(i <= fLastInterval); |
351 | i++; |
352 | |
353 | if (tileMode == SkTileMode::kClamp) { |
354 | SkASSERT(i <= fLastInterval); |
355 | return i; |
356 | } |
357 | |
358 | return (i <= fLastInterval) ? i : fFirstInterval; |
359 | } |
360 | |
361 | SkScalar advance_interval(SkScalar advX) { |
362 | SkASSERT(advX >= fAdvX); |
363 | |
364 | do { |
365 | advX -= fAdvX; |
366 | fInterval = this->next_interval(fInterval); |
367 | fAdvX = (fInterval->fT1 - fInterval->fT0) / fDx; |
368 | SkASSERT(fAdvX > 0); |
369 | } while (advX >= fAdvX); |
370 | |
371 | compute_interval_props(fInterval->fT0); |
372 | |
373 | SkASSERT(advX >= 0); |
374 | return advX; |
375 | } |
376 | |
377 | // Current interval properties. |
378 | Sk4f fDcDx; // dst color gradient (dc/dx) |
379 | Sk4f fCc; // current color, interpolated in dst |
380 | SkScalar fAdvX; // remaining interval advance in dst |
381 | bool fZeroRamp; // current interval color grad is 0 |
382 | |
383 | const Sk4fGradientInterval* fFirstInterval; |
384 | const Sk4fGradientInterval* fLastInterval; |
385 | const Sk4fGradientInterval* fInterval; // current interval |
386 | const SkScalar fDx; // 'dx' for consistency with other impls; actually dt/dx |
387 | const bool fIsVertical; |
388 | }; |
389 | |