The Battle for Wesnoth  1.19.12+dev
xbrz.cpp
Go to the documentation of this file.
1 // ****************************************************************************
2 // * This file is part of the xBRZ project. It is distributed under *
3 // * GNU General Public License: https://www.gnu.org/licenses/gpl-3.0 *
4 // * Copyright (C) Zenju (zenju AT gmx DOT de) - All Rights Reserved *
5 // * *
6 // * Additionally and as a special exception, the author gives permission *
7 // * to link the code of this program with the following libraries *
8 // * (or with modified versions that use the same licenses), and distribute *
9 // * linked combinations including the two: MAME, FreeFileSync, Snes9x, ePSXe *
10 // * You must obey the GNU General Public License in all respects for all of *
11 // * the code used other than MAME, FreeFileSync, Snes9x, ePSXe. *
12 // * If you modify this file, you may extend this exception to your version *
13 // * of the file, but you are not obligated to do so. If you do not wish to *
14 // * do so, delete this exception statement from your version. *
15 // ****************************************************************************
16 
17 #include "xbrz.hpp"
18 #include <cassert>
19 #include <vector>
20 #include <algorithm>
21 #include <cmath> //std::sqrt
22 #include "xbrz_tools.hpp"
23 
24 #include "../global.hpp" // for LIKELY
25 
26 using namespace xbrz;
27 
28 
29 namespace
30 {
31 template <unsigned int M, unsigned int N> inline
32 uint32_t gradientRGB(uint32_t pixFront, uint32_t pixBack) //blend front color with opacity M / N over opaque background: https://en.wikipedia.org/wiki/Alpha_compositing#Alpha_blending
33 {
34  static_assert(0 < M && M < N && N <= 1000);
35 
36  auto calcColor = [](unsigned char colFront, unsigned char colBack) -> unsigned char { return (colFront * M + colBack * (N - M)) / N; };
37 
38  return makePixel(calcColor(getRed (pixFront), getRed (pixBack)),
39  calcColor(getGreen(pixFront), getGreen(pixBack)),
40  calcColor(getBlue (pixFront), getBlue (pixBack)));
41 }
42 
43 
44 template <unsigned int M, unsigned int N> inline
45 uint32_t gradientARGB(uint32_t pixFront, uint32_t pixBack) //find intermediate color between two colors with alpha channels (=> NO alpha blending!!!)
46 {
47  static_assert(0 < M && M < N && N <= 1000);
48 
49  const unsigned int weightFront = getAlpha(pixFront) * M;
50  const unsigned int weightBack = getAlpha(pixBack) * (N - M);
51  const unsigned int weightSum = weightFront + weightBack;
52  if (weightSum == 0)
53  return 0;
54 
55  auto calcColor = [=](unsigned char colFront, unsigned char colBack)
56  {
57  return static_cast<unsigned char>((colFront * weightFront + colBack * weightBack) / weightSum);
58  };
59 
60  return makePixel(static_cast<unsigned char>(weightSum / N),
61  calcColor(getRed (pixFront), getRed (pixBack)),
62  calcColor(getGreen(pixFront), getGreen(pixBack)),
63  calcColor(getBlue (pixFront), getBlue (pixBack)));
64 }
65 
66 
67 //inline
68 //double fastSqrt(double n)
69 //{
70 // __asm //speeds up xBRZ by about 9% compared to std::sqrt which internally uses the same assembler instructions but adds some "fluff"
71 // {
72 // fld n
73 // fsqrt
74 // }
75 //}
76 //
77 
78 
79 #ifdef _MSC_VER
80  #define FORCE_INLINE __forceinline
81 #elif defined __GNUC__
82  #define FORCE_INLINE __attribute__((always_inline)) inline
83 #else
84  #define FORCE_INLINE inline
85 #endif
86 
87 
88 enum RotationDegree //clock-wise
89 {
90  ROT_0,
91  ROT_90,
92  ROT_180,
93  ROT_270
94 };
95 
96 //calculate input matrix coordinates after rotation at compile time
97 template <RotationDegree rotDeg, size_t I, size_t J, size_t N>
98 struct MatrixRotation;
99 
100 template <size_t I, size_t J, size_t N>
101 struct MatrixRotation<ROT_0, I, J, N>
102 {
103  static const size_t I_old = I;
104  static const size_t J_old = J;
105 };
106 
107 template <RotationDegree rotDeg, size_t I, size_t J, size_t N> //(i, j) = (row, col) indices, N = size of (square) matrix
108 struct MatrixRotation
109 {
110  static const size_t I_old = N - 1 - MatrixRotation<static_cast<RotationDegree>(rotDeg - 1), I, J, N>::J_old; //old coordinates before rotation!
111  static const size_t J_old = MatrixRotation<static_cast<RotationDegree>(rotDeg - 1), I, J, N>::I_old; //
112 };
113 
114 
115 template <size_t N, RotationDegree rotDeg>
116 class OutputMatrix
117 {
118 public:
119  OutputMatrix(uint32_t* out, int outWidth) : //access matrix area, top-left at position "out" for image with given width
120  out_(out),
121  outWidth_(outWidth) {}
122 
123  template <size_t I, size_t J>
124  uint32_t& ref() const
125  {
126  static const size_t I_old = MatrixRotation<rotDeg, I, J, N>::I_old;
127  static const size_t J_old = MatrixRotation<rotDeg, I, J, N>::J_old;
128  return *(out_ + J_old + I_old * outWidth_);
129  }
130 
131 private:
132  uint32_t* out_;
133  const int outWidth_;
134 };
135 
136 
137 template <class T> inline
138 T square(T value) { return value * value; }
139 
140 
141 #if 0
142 inline
143 double distRGB(uint32_t pix1, uint32_t pix2)
144 {
145  const double r_diff = static_cast<int>(getRed (pix1)) - getRed (pix2);
146  const double g_diff = static_cast<int>(getGreen(pix1)) - getGreen(pix2);
147  const double b_diff = static_cast<int>(getBlue (pix1)) - getBlue (pix2);
148 
149  //euklidean RGB distance
150  return std::sqrt(square(r_diff) + square(g_diff) + square(b_diff));
151 }
152 #endif
153 
154 
155 inline
156 double distYCbCr(uint32_t pix1, uint32_t pix2, double lumaWeight)
157 {
158  //https://en.wikipedia.org/wiki/YCbCr#ITU-R_BT.601_conversion
159  //YCbCr conversion is a matrix multiplication => take advantage of linearity by subtracting first!
160  const int r_diff = static_cast<int>(getRed (pix1)) - getRed (pix2); //we may delay division by 255 to after matrix multiplication
161  const int g_diff = static_cast<int>(getGreen(pix1)) - getGreen(pix2); //
162  const int b_diff = static_cast<int>(getBlue (pix1)) - getBlue (pix2); //substraction for int is noticeable faster than for double!
163 
164  //const double k_b = 0.0722; //ITU-R BT.709 conversion
165  //const double k_r = 0.2126; //
166  const double k_b = 0.0593; //ITU-R BT.2020 conversion
167  const double k_r = 0.2627; //
168  const double k_g = 1 - k_b - k_r;
169 
170  const double scale_b = 0.5 / (1 - k_b);
171  const double scale_r = 0.5 / (1 - k_r);
172 
173  const double y = k_r * r_diff + k_g * g_diff + k_b * b_diff; //[!], analog YCbCr!
174  const double c_b = scale_b * (b_diff - y);
175  const double c_r = scale_r * (r_diff - y);
176 
177  //we skip division by 255 to have similar range like other distance functions
178  return std::sqrt(square(lumaWeight * y) + square(c_b) + square(c_r));
179 }
180 
181 
182 inline
183 double distYCbCrBuffered(uint32_t pix1, uint32_t pix2)
184 {
185  //30% perf boost compared to plain distYCbCr()!
186  //consumes 64 MB memory; using double is only 2% faster, but takes 128 MB
187  static const std::vector<float> diffToDist = []
188  {
189  std::vector<float> tmp;
190 
191  for (uint32_t i = 0; i < 256 * 256 * 256; ++i) //startup time: 114 ms on Intel Core i5 (four cores)
192  {
193  const int r_diff = static_cast<signed char>(getByte<2>(i)) * 2;
194  const int g_diff = static_cast<signed char>(getByte<1>(i)) * 2;
195  const int b_diff = static_cast<signed char>(getByte<0>(i)) * 2;
196 
197  const double k_b = 0.0593; //ITU-R BT.2020 conversion
198  const double k_r = 0.2627; //
199  const double k_g = 1 - k_b - k_r;
200 
201  const double scale_b = 0.5 / (1 - k_b);
202  const double scale_r = 0.5 / (1 - k_r);
203 
204  const double y = k_r * r_diff + k_g * g_diff + k_b * b_diff; //[!], analog YCbCr!
205  const double c_b = scale_b * (b_diff - y);
206  const double c_r = scale_r * (r_diff - y);
207 
208  tmp.push_back(static_cast<float>(std::sqrt(square(y) + square(c_b) + square(c_r))));
209  }
210  return tmp;
211  }();
212 
213  //if (pix1 == pix2) -> 8% perf degradation!
214  // return 0;
215  //if (pix1 < pix2)
216  // std::swap(pix1, pix2); -> 30% perf degradation!!!
217 
218  const int r_diff = static_cast<int>(getRed (pix1)) - getRed (pix2);
219  const int g_diff = static_cast<int>(getGreen(pix1)) - getGreen(pix2);
220  const int b_diff = static_cast<int>(getBlue (pix1)) - getBlue (pix2);
221 
222  const size_t index = (static_cast<unsigned char>(r_diff / 2) << 16) | //slightly reduce precision (division by 2) to squeeze value into single byte
223  (static_cast<unsigned char>(g_diff / 2) << 8) |
224  (static_cast<unsigned char>(b_diff / 2));
225 
226 #if 0 //attention: the following calculation creates an asymmetric color distance!!! (e.g. r_diff=46 will be unpacked as 45, but r_diff=-46 unpacks to -47
227  const size_t index = (((r_diff + 0xFF) / 2) << 16) | //slightly reduce precision (division by 2) to squeeze value into single byte
228  (((g_diff + 0xFF) / 2) << 8) |
229  (( b_diff + 0xFF) / 2);
230 #endif
231  return diffToDist[index];
232 }
233 
234 
235 #if defined _MSC_VER && !defined NDEBUG
236  const int debugPixelX = -1;
237  const int debugPixelY = 58;
238 
239  thread_local bool breakIntoDebugger = false;
240 #endif
241 
242 
243 enum BlendType
244 {
245  BLEND_NONE = 0,
246  BLEND_NORMAL, //a normal indication to blend
247  BLEND_DOMINANT, //a strong indication to blend
248  //attention: BlendType must fit into the value range of 2 bit!!!
249 };
250 
251 struct BlendResult
252 {
253  BlendType
254  /**/blend_f, blend_g,
255  /**/blend_j, blend_k;
256 };
257 
258 
259 struct Kernel_3x3
260 {
261  uint32_t
262  a, b, c,
263  d, e, f,
264  g, h, i;
265 };
266 
267 struct Kernel_4x4 //kernel for preprocessing step
268 {
269  uint32_t
270  a, b, c, //
271  e, f, g, // support reinterpret_cast from Kernel_4x4 => Kernel_3x3
272  i, j, k, //
273  m, n, o,
274  d, h, l, p;
275 };
276 
277 /* input kernel area naming convention:
278 -----------------
279 | A | B | C | D |
280 |---|---|---|---|
281 | E | F | G | H | evaluate the four corners between F, G, J, K
282 |---|---|---|---| input pixel is at position F
283 | I | J | K | L |
284 |---|---|---|---|
285 | M | N | O | P |
286 -----------------
287 */
288 template <class ColorDistance>
289 FORCE_INLINE //detect blend direction
290 BlendResult preProcessCorners(const Kernel_4x4& ker, const xbrz::ScalerCfg& cfg) //result: F, G, J, K corners of "GradientType"
291 {
292 #if defined _MSC_VER && !defined NDEBUG
293  if (breakIntoDebugger)
294  __debugbreak(); //__asm int 3;
295 #endif
296 
297  BlendResult result = {};
298 
299  if ((ker.f == ker.g &&
300  ker.j == ker.k) ||
301  (ker.f == ker.j &&
302  ker.g == ker.k))
303  return result;
304 
305  auto dist = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight); };
306 
307  double jg = dist(ker.i, ker.f) + dist(ker.f, ker.c) + dist(ker.n, ker.k) + dist(ker.k, ker.h) + cfg.centerDirectionBias * dist(ker.j, ker.g);
308  double fk = dist(ker.e, ker.j) + dist(ker.j, ker.o) + dist(ker.b, ker.g) + dist(ker.g, ker.l) + cfg.centerDirectionBias * dist(ker.f, ker.k);
309 
310  if (jg < fk) //test sample: 70% of values max(jg, fk) / min(jg, fk) are between 1.1 and 3.7 with median being 1.8
311  {
312  const bool dominantGradient = cfg.dominantDirectionThreshold * jg < fk;
313  if (ker.f != ker.g && ker.f != ker.j)
314  result.blend_f = dominantGradient ? BLEND_DOMINANT : BLEND_NORMAL;
315 
316  if (ker.k != ker.j && ker.k != ker.g)
317  result.blend_k = dominantGradient ? BLEND_DOMINANT : BLEND_NORMAL;
318  }
319  else if (fk < jg)
320  {
321  const bool dominantGradient = cfg.dominantDirectionThreshold * fk < jg;
322  if (ker.j != ker.f && ker.j != ker.k)
323  result.blend_j = dominantGradient ? BLEND_DOMINANT : BLEND_NORMAL;
324 
325  if (ker.g != ker.f && ker.g != ker.k)
326  result.blend_g = dominantGradient ? BLEND_DOMINANT : BLEND_NORMAL;
327  }
328  return result;
329 }
330 
331 #ifdef __APPLE__
332 #pragma clang diagnostic push
333 #pragma clang diagnostic ignored "-Wunused-function"
334 #endif
335 
336 #define DEF_GETTER(x) template <RotationDegree rotDeg> uint32_t inline get_##x(const Kernel_4x4& ker) { return ker.x; }
337 //we cannot and NEED NOT write "ker.##x" since ## concatenates preprocessor tokens but "." is not a token
338 DEF_GETTER(a) DEF_GETTER(b) DEF_GETTER(c)
339 DEF_GETTER(d) DEF_GETTER(e) DEF_GETTER(f)
340 DEF_GETTER(g) DEF_GETTER(h) DEF_GETTER(i)
341 #undef DEF_GETTER
342 
343 #define DEF_GETTER(x, y) template <> inline uint32_t get_##x<ROT_0>(const Kernel_4x4& ker) { return ker.y; }
344 DEF_GETTER(a, a) DEF_GETTER(b, b) DEF_GETTER(c, c)
345 DEF_GETTER(d, e) DEF_GETTER(e, f) DEF_GETTER(f, g)
346 DEF_GETTER(g, i) DEF_GETTER(h, j) DEF_GETTER(i, k)
347 #undef DEF_GETTER
348 
349 #define DEF_GETTER(x, y) template <> inline uint32_t get_##x<ROT_90>(const Kernel_4x4& ker) { return ker.y; }
350 DEF_GETTER(a, i) DEF_GETTER(b, e) DEF_GETTER(c, a)
351 DEF_GETTER(d, j) DEF_GETTER(e, f) DEF_GETTER(f, b)
352 DEF_GETTER(g, k) DEF_GETTER(h, g) DEF_GETTER(i, c)
353 #undef DEF_GETTER
354 
355 #define DEF_GETTER(x, y) template <> inline uint32_t get_##x<ROT_180>(const Kernel_4x4& ker) { return ker.y; }
356 DEF_GETTER(a, k) DEF_GETTER(b, j) DEF_GETTER(c, i)
357 DEF_GETTER(d, g) DEF_GETTER(e, f) DEF_GETTER(f, e)
358 DEF_GETTER(g, c) DEF_GETTER(h, b) DEF_GETTER(i, a)
359 #undef DEF_GETTER
360 
361 #define DEF_GETTER(x, y) template <> inline uint32_t get_##x<ROT_270>(const Kernel_4x4& ker) { return ker.y; }
362 DEF_GETTER(a, c) DEF_GETTER(b, g) DEF_GETTER(c, k)
363 DEF_GETTER(d, b) DEF_GETTER(e, f) DEF_GETTER(f, j)
364 DEF_GETTER(g, a) DEF_GETTER(h, e) DEF_GETTER(i, i)
365 #undef DEF_GETTER
366 
367 #ifdef __APPLE__
368 #pragma clang diagnostic pop
369 #endif
370 
371 //compress four blend types into a single byte
372 //inline BlendType getTopL (unsigned char b) { return static_cast<BlendType>(0x3 & b); }
373 inline BlendType getTopR (unsigned char b) { return static_cast<BlendType>(0x3 & (b >> 2)); }
374 inline BlendType getBottomR(unsigned char b) { return static_cast<BlendType>(0x3 & (b >> 4)); }
375 inline BlendType getBottomL(unsigned char b) { return static_cast<BlendType>(0x3 & (b >> 6)); }
376 
377 inline void clearAddTopL(unsigned char& b, BlendType bt) { b = static_cast<unsigned char>(bt); }
378 inline void addTopR (unsigned char& b, BlendType bt) { b |= (bt << 2); } //buffer is assumed to be initialized before preprocessing!
379 inline void addBottomR (unsigned char& b, BlendType bt) { b |= (bt << 4); } //e.g. via clearAddTopL()
380 inline void addBottomL (unsigned char& b, BlendType bt) { b |= (bt << 6); } //
381 
382 inline bool blendingNeeded(unsigned char b)
383 {
384  static_assert(BLEND_NONE == 0);
385  return b != 0;
386 }
387 
388 template <RotationDegree rotDeg> inline
389 unsigned char rotateBlendInfo(unsigned char b) { return b; }
390 template <> inline unsigned char rotateBlendInfo<ROT_90 >(unsigned char b) { return ((b << 2) | (b >> 6)) & 0xff; }
391 template <> inline unsigned char rotateBlendInfo<ROT_180>(unsigned char b) { return ((b << 4) | (b >> 4)) & 0xff; }
392 template <> inline unsigned char rotateBlendInfo<ROT_270>(unsigned char b) { return ((b << 6) | (b >> 2)) & 0xff; }
393 
394 
395 /* input kernel area naming convention:
396 -------------
397 | A | B | C |
398 |---|---|---|
399 | D | E | F | input pixel is at position E
400 |---|---|---|
401 | G | H | I |
402 -------------
403 */
404 template <class Scaler, class ColorDistance, RotationDegree rotDeg>
405 FORCE_INLINE //perf: quite worth it!
406 void blendPixel(const Kernel_4x4& ker,
407  uint32_t* target, int trgWidth,
408  unsigned char blendInfo, //result of preprocessing all four corners of pixel "e"
409  const xbrz::ScalerCfg& cfg)
410 {
411  //#define a get_a<rotDeg>(ker)
412 #define b get_b<rotDeg>(ker)
413 #define c get_c<rotDeg>(ker)
414 #define d get_d<rotDeg>(ker)
415 #define e get_e<rotDeg>(ker)
416 #define f get_f<rotDeg>(ker)
417 #define g get_g<rotDeg>(ker)
418 #define h get_h<rotDeg>(ker)
419 #define i get_i<rotDeg>(ker)
420 
421 #if defined _MSC_VER && !defined NDEBUG
422  if (breakIntoDebugger)
423  __debugbreak(); //__asm int 3;
424 #endif
425 
426  const unsigned char blend = rotateBlendInfo<rotDeg>(blendInfo);
427 
428  if (getBottomR(blend) >= BLEND_NORMAL)
429  {
430  auto eq = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight) < cfg.equalColorTolerance; };
431  auto dist = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight); };
432 
433  const bool doLineBlend = [&]() -> bool
434  {
435  if (getBottomR(blend) >= BLEND_DOMINANT)
436  return true;
437 
438  //make sure there is no second blending in an adjacent rotation for this pixel: handles insular pixels, mario eyes
439  if (getTopR(blend) != BLEND_NONE && !eq(e, g)) //but support double-blending for 90� corners
440  return false;
441  if (getBottomL(blend) != BLEND_NONE && !eq(e, c))
442  return false;
443 
444  //no full blending for L-shapes; blend corner only (handles "mario mushroom eyes")
445  if (!eq(e, i) && eq(g, h) && eq(h, i) && eq(i, f) && eq(f, c))
446  return false;
447 
448  return true;
449  }();
450 
451  const uint32_t px = dist(e, f) <= dist(e, h) ? f : h; //choose most similar color
452 
453  OutputMatrix<Scaler::scale, rotDeg> out(target, trgWidth);
454 
455  if (doLineBlend)
456  {
457  const double fg = dist(f, g); //test sample: 70% of values max(fg, hc) / min(fg, hc) are between 1.1 and 3.7 with median being 1.9
458  const double hc = dist(h, c); //
459 
460  const bool haveShallowLine = cfg.steepDirectionThreshold * fg <= hc && e != g && d != g;
461  const bool haveSteepLine = cfg.steepDirectionThreshold * hc <= fg && e != c && b != c;
462 
463  if (haveShallowLine)
464  {
465  if (haveSteepLine)
466  Scaler::blendLineSteepAndShallow(px, out);
467  else
468  Scaler::blendLineShallow(px, out);
469  }
470  else
471  {
472  if (haveSteepLine)
473  Scaler::blendLineSteep(px, out);
474  else
475  Scaler::blendLineDiagonal(px, out);
476  }
477  }
478  else
479  Scaler::blendCorner(px, out);
480  }
481 
482  //#undef a
483 #undef b
484 #undef c
485 #undef d
486 #undef e
487 #undef f
488 #undef g
489 #undef h
490 #undef i
491 }
492 
493 
494 class OobReaderTransparent
495 {
496 public:
497  OobReaderTransparent(const uint32_t* src, int srcWidth, int srcHeight, int y) :
498  s_m1(0 <= y - 1 && y - 1 < srcHeight ? src + srcWidth * (y - 1) : nullptr),
499  s_0 (0 <= y && y < srcHeight ? src + srcWidth * y : nullptr),
500  s_p1(0 <= y + 1 && y + 1 < srcHeight ? src + srcWidth * (y + 1) : nullptr),
501  s_p2(0 <= y + 2 && y + 2 < srcHeight ? src + srcWidth * (y + 2) : nullptr),
502  srcWidth_(srcWidth) {}
503 
504  void readDhlp(Kernel_4x4& ker, int x) const //(x, y) is at kernel position F
505  {
506  LIKELY if (const int x_p2 = x + 2; 0 <= x_p2 && x_p2 < srcWidth_)
507  {
508  ker.d = s_m1 ? s_m1[x_p2] : 0;
509  ker.h = s_0 ? s_0 [x_p2] : 0;
510  ker.l = s_p1 ? s_p1[x_p2] : 0;
511  ker.p = s_p2 ? s_p2[x_p2] : 0;
512  }
513  else
514  {
515  ker.d = 0;
516  ker.h = 0;
517  ker.l = 0;
518  ker.p = 0;
519  }
520  }
521 
522 private:
523  const uint32_t* const s_m1;
524  const uint32_t* const s_0;
525  const uint32_t* const s_p1;
526  const uint32_t* const s_p2;
527  const int srcWidth_;
528 };
529 
530 
531 class OobReaderDuplicate
532 {
533 public:
534  OobReaderDuplicate(const uint32_t* src, int srcWidth, int srcHeight, int y) :
535  s_m1(src + srcWidth * std::clamp(y - 1, 0, srcHeight - 1)),
536  s_0 (src + srcWidth * std::clamp(y, 0, srcHeight - 1)),
537  s_p1(src + srcWidth * std::clamp(y + 1, 0, srcHeight - 1)),
538  s_p2(src + srcWidth * std::clamp(y + 2, 0, srcHeight - 1)),
539  srcWidth_(srcWidth) {}
540 
541  void readDhlp(Kernel_4x4& ker, int x) const //(x, y) is at kernel position F
542  {
543  const int x_p2 = std::clamp(x + 2, 0, srcWidth_ - 1);
544  ker.d = s_m1[x_p2];
545  ker.h = s_0 [x_p2];
546  ker.l = s_p1[x_p2];
547  ker.p = s_p2[x_p2];
548  }
549 
550 private:
551  const uint32_t* const s_m1;
552  const uint32_t* const s_0;
553  const uint32_t* const s_p1;
554  const uint32_t* const s_p2;
555  const int srcWidth_;
556 };
557 
558 
559 template <class Scaler, class ColorDistance, class OobReader> //scaler policy: see "Scaler2x" reference implementation
560 void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, const xbrz::ScalerCfg& cfg, int yFirst, int yLast)
561 {
562  yFirst = std::max(yFirst, 0);
563  yLast = std::min(yLast, srcHeight);
564  if (yFirst >= yLast || srcWidth <= 0)
565  return;
566 
567  const int trgWidth = srcWidth * Scaler::scale;
568 
569  //(ab)use space of "sizeof(uint32_t) * srcWidth * Scaler::scale" at the end of the image as temporary
570  //buffer for "on the fly preprocessing" without risk of accidental overwriting before accessing
571  unsigned char* const preProcBuf = reinterpret_cast<unsigned char*>(trg + yLast * Scaler::scale * trgWidth) - srcWidth;
572 
573  //initialize preprocessing buffer for first row of current stripe: detect upper left and right corner blending
574  //this cannot be optimized for adjacent processing stripes; we must not allow for a memory race condition!
575  {
576  const OobReader oobReader(src, srcWidth, srcHeight, yFirst - 1);
577 
578  //initialize at position x = -1
579  Kernel_4x4 ker4 = {};
580  oobReader.readDhlp(ker4, -4); //hack: read a, e, i, m at x = -1
581  ker4.a = ker4.d;
582  ker4.e = ker4.h;
583  ker4.i = ker4.l;
584  ker4.m = ker4.p;
585 
586  oobReader.readDhlp(ker4, -3);
587  ker4.b = ker4.d;
588  ker4.f = ker4.h;
589  ker4.j = ker4.l;
590  ker4.n = ker4.p;
591 
592  oobReader.readDhlp(ker4, -2);
593  ker4.c = ker4.d;
594  ker4.g = ker4.h;
595  ker4.k = ker4.l;
596  ker4.o = ker4.p;
597 
598  oobReader.readDhlp(ker4, -1);
599 
600  {
601  const BlendResult res = preProcessCorners<ColorDistance>(ker4, cfg);
602  clearAddTopL(preProcBuf[0], res.blend_k); //set 1st known corner for (0, yFirst)
603  }
604 
605  for (int x = 0; x < srcWidth; ++x)
606  {
607  ker4.a = ker4.b; //shift previous kernel to the left
608  ker4.e = ker4.f; // -----------------
609  ker4.i = ker4.j; // | A | B | C | D |
610  ker4.m = ker4.n; // |---|---|---|---|
611  /**/ // | E | F | G | H | (x, yFirst - 1) is at position F
612  ker4.b = ker4.c; // |---|---|---|---|
613  ker4.f = ker4.g; // | I | J | K | L |
614  ker4.j = ker4.k; // |---|---|---|---|
615  ker4.n = ker4.o; // | M | N | O | P |
616  /**/ // -----------------
617  ker4.c = ker4.d;
618  ker4.g = ker4.h;
619  ker4.k = ker4.l;
620  ker4.o = ker4.p;
621 
622  oobReader.readDhlp(ker4, x);
623 
624  /* preprocessing blend result:
625  ---------
626  | F | G | evaluate corner between F, G, J, K
627  |---+---| current input pixel is at position F
628  | J | K |
629  --------- */
630  const BlendResult res = preProcessCorners<ColorDistance>(ker4, cfg);
631  addTopR(preProcBuf[x], res.blend_j); //set 2nd known corner for (x, yFirst)
632 
633  if (x + 1 < srcWidth)
634  clearAddTopL(preProcBuf[x + 1], res.blend_k); //set 1st known corner for (x + 1, yFirst)
635  }
636  }
637  //------------------------------------------------------------------------------------
638 
639  for (int y = yFirst; y < yLast; ++y)
640  {
641  uint32_t* out = trg + Scaler::scale * y * trgWidth; //consider MT "striped" access
642 
643  const OobReader oobReader(src, srcWidth, srcHeight, y);
644 
645  //initialize at position x = -1
646  Kernel_4x4 ker4 = {};
647  oobReader.readDhlp(ker4, -4); //hack: read a, e, i, m at x = -1
648  ker4.a = ker4.d;
649  ker4.e = ker4.h;
650  ker4.i = ker4.l;
651  ker4.m = ker4.p;
652 
653  oobReader.readDhlp(ker4, -3);
654  ker4.b = ker4.d;
655  ker4.f = ker4.h;
656  ker4.j = ker4.l;
657  ker4.n = ker4.p;
658 
659  oobReader.readDhlp(ker4, -2);
660  ker4.c = ker4.d;
661  ker4.g = ker4.h;
662  ker4.k = ker4.l;
663  ker4.o = ker4.p;
664 
665  oobReader.readDhlp(ker4, -1);
666 
667  unsigned char blend_xy1 = 0; //corner blending for current (x, y + 1) position
668  {
669  const BlendResult res = preProcessCorners<ColorDistance>(ker4, cfg);
670  clearAddTopL(blend_xy1, res.blend_k); //set 1st known corner for (0, y + 1) and buffer for use on next column
671 
672  addBottomL(preProcBuf[0], res.blend_g); //set 3rd known corner for (0, y)
673  }
674 
675  for (int x = 0; x < srcWidth; ++x, out += Scaler::scale)
676  {
677 #if defined _MSC_VER && !defined NDEBUG
678  breakIntoDebugger = debugPixelX == x && debugPixelY == y;
679 #endif
680  ker4.a = ker4.b; //shift previous kernel to the left
681  ker4.e = ker4.f; // -----------------
682  ker4.i = ker4.j; // | A | B | C | D |
683  ker4.m = ker4.n; // |---|---|---|---|
684  /**/ // | E | F | G | H | (x, y) is at position F
685  ker4.b = ker4.c; // |---|---|---|---|
686  ker4.f = ker4.g; // | I | J | K | L |
687  ker4.j = ker4.k; // |---|---|---|---|
688  ker4.n = ker4.o; // | M | N | O | P |
689  /**/ // -----------------
690  ker4.c = ker4.d;
691  ker4.g = ker4.h;
692  ker4.k = ker4.l;
693  ker4.o = ker4.p;
694 
695  oobReader.readDhlp(ker4, x);
696 
697  //evaluate the four corners on bottom-right of current pixel
698  unsigned char blend_xy = preProcBuf[x]; //for current (x, y) position
699  {
700  /* preprocessing blend result:
701  ---------
702  | F | G | evaluate corner between F, G, J, K
703  |---+---| current input pixel is at position F
704  | J | K |
705  --------- */
706  const BlendResult res = preProcessCorners<ColorDistance>(ker4, cfg);
707  addBottomR(blend_xy, res.blend_f); //all four corners of (x, y) have been determined at this point due to processing sequence!
708 
709  addTopR(blend_xy1, res.blend_j); //set 2nd known corner for (x, y + 1)
710  preProcBuf[x] = blend_xy1; //store on current buffer position for use on next row
711 
712  LIKELY if (x + 1 < srcWidth)
713  {
714  //blend_xy1 -> blend_x1y1
715  clearAddTopL(blend_xy1, res.blend_k); //set 1st known corner for (x + 1, y + 1) and buffer for use on next column
716 
717  addBottomL(preProcBuf[x + 1], res.blend_g); //set 3rd known corner for (x + 1, y)
718  }
719  }
720 
721  //fill block of size scale * scale with the given color
722  fillBlock(out, trgWidth * sizeof(uint32_t), ker4.f, Scaler::scale, Scaler::scale);
723  //place *after* preprocessing step, to not overwrite the results while processing the last pixel!
724 
725  //blend all four corners of current pixel
726  if (blendingNeeded(blend_xy))
727  {
728  blendPixel<Scaler, ColorDistance, ROT_0 >(ker4, out, trgWidth, blend_xy, cfg);
729  blendPixel<Scaler, ColorDistance, ROT_90 >(ker4, out, trgWidth, blend_xy, cfg);
730  blendPixel<Scaler, ColorDistance, ROT_180>(ker4, out, trgWidth, blend_xy, cfg);
731  blendPixel<Scaler, ColorDistance, ROT_270>(ker4, out, trgWidth, blend_xy, cfg);
732  }
733  }
734  }
735 }
736 
737 //------------------------------------------------------------------------------------
738 
739 template <class ColorGradient>
740 struct Scaler2x : public ColorGradient
741 {
742  static const int scale = 2;
743 
744  template <unsigned int M, unsigned int N> //bring template function into scope for GCC
745  static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) { ColorGradient::template alphaGrad<M, N>(pixBack, pixFront); }
746 
747 
748  template <class OutputMatrix>
749  static void blendLineShallow(uint32_t col, OutputMatrix& out)
750  {
751  alphaGrad<1, 4>(out.template ref<scale - 1, 0>(), col);
752  alphaGrad<3, 4>(out.template ref<scale - 1, 1>(), col);
753  }
754 
755  template <class OutputMatrix>
756  static void blendLineSteep(uint32_t col, OutputMatrix& out)
757  {
758  alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col);
759  alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col);
760  }
761 
762  template <class OutputMatrix>
763  static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out)
764  {
765  alphaGrad<1, 4>(out.template ref<1, 0>(), col);
766  alphaGrad<1, 4>(out.template ref<0, 1>(), col);
767  alphaGrad<5, 6>(out.template ref<1, 1>(), col); //[!] fixes 7/8 used in xBR
768  }
769 
770  template <class OutputMatrix>
771  static void blendLineDiagonal(uint32_t col, OutputMatrix& out)
772  {
773  alphaGrad<1, 2>(out.template ref<1, 1>(), col);
774  }
775 
776  template <class OutputMatrix>
777  static void blendCorner(uint32_t col, OutputMatrix& out)
778  {
779  //model a round corner
780  alphaGrad<21, 100>(out.template ref<1, 1>(), col); //exact: 1 - pi/4 = 0.2146018366
781  }
782 };
783 
784 
785 template <class ColorGradient>
786 struct Scaler3x : public ColorGradient
787 {
788  static const int scale = 3;
789 
790  template <unsigned int M, unsigned int N> //bring template function into scope for GCC
791  static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) { ColorGradient::template alphaGrad<M, N>(pixBack, pixFront); }
792 
793 
794  template <class OutputMatrix>
795  static void blendLineShallow(uint32_t col, OutputMatrix& out)
796  {
797  alphaGrad<1, 4>(out.template ref<scale - 1, 0>(), col);
798  alphaGrad<1, 4>(out.template ref<scale - 2, 2>(), col);
799 
800  alphaGrad<3, 4>(out.template ref<scale - 1, 1>(), col);
801  out.template ref<scale - 1, 2>() = col;
802  }
803 
804  template <class OutputMatrix>
805  static void blendLineSteep(uint32_t col, OutputMatrix& out)
806  {
807  alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col);
808  alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col);
809 
810  alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col);
811  out.template ref<2, scale - 1>() = col;
812  }
813 
814  template <class OutputMatrix>
815  static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out)
816  {
817  alphaGrad<1, 4>(out.template ref<2, 0>(), col);
818  alphaGrad<1, 4>(out.template ref<0, 2>(), col);
819  alphaGrad<3, 4>(out.template ref<2, 1>(), col);
820  alphaGrad<3, 4>(out.template ref<1, 2>(), col);
821  out.template ref<2, 2>() = col;
822  }
823 
824  template <class OutputMatrix>
825  static void blendLineDiagonal(uint32_t col, OutputMatrix& out)
826  {
827  alphaGrad<1, 8>(out.template ref<1, 2>(), col); //conflict with other rotations for this odd scale
828  alphaGrad<1, 8>(out.template ref<2, 1>(), col);
829  alphaGrad<7, 8>(out.template ref<2, 2>(), col); //
830  }
831 
832  template <class OutputMatrix>
833  static void blendCorner(uint32_t col, OutputMatrix& out)
834  {
835  //model a round corner
836  alphaGrad<45, 100>(out.template ref<2, 2>(), col); //exact: 0.4545939598
837  //alphaGrad<7, 256>(out.template ref<2, 1>(), col); //0.02826017254 -> negligible + avoid conflicts with other rotations for this odd scale
838  //alphaGrad<7, 256>(out.template ref<1, 2>(), col); //0.02826017254
839  }
840 };
841 
842 
843 template <class ColorGradient>
844 struct Scaler4x : public ColorGradient
845 {
846  static const int scale = 4;
847 
848  template <unsigned int M, unsigned int N> //bring template function into scope for GCC
849  static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) { ColorGradient::template alphaGrad<M, N>(pixBack, pixFront); }
850 
851 
852  template <class OutputMatrix>
853  static void blendLineShallow(uint32_t col, OutputMatrix& out)
854  {
855  alphaGrad<1, 4>(out.template ref<scale - 1, 0>(), col);
856  alphaGrad<1, 4>(out.template ref<scale - 2, 2>(), col);
857 
858  alphaGrad<3, 4>(out.template ref<scale - 1, 1>(), col);
859  alphaGrad<3, 4>(out.template ref<scale - 2, 3>(), col);
860 
861  out.template ref<scale - 1, 2>() = col;
862  out.template ref<scale - 1, 3>() = col;
863  }
864 
865  template <class OutputMatrix>
866  static void blendLineSteep(uint32_t col, OutputMatrix& out)
867  {
868  alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col);
869  alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col);
870 
871  alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col);
872  alphaGrad<3, 4>(out.template ref<3, scale - 2>(), col);
873 
874  out.template ref<2, scale - 1>() = col;
875  out.template ref<3, scale - 1>() = col;
876  }
877 
878  template <class OutputMatrix>
879  static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out)
880  {
881  alphaGrad<3, 4>(out.template ref<3, 1>(), col);
882  alphaGrad<3, 4>(out.template ref<1, 3>(), col);
883  alphaGrad<1, 4>(out.template ref<3, 0>(), col);
884  alphaGrad<1, 4>(out.template ref<0, 3>(), col);
885 
886  alphaGrad<1, 3>(out.template ref<2, 2>(), col); //[!] fixes 1/4 used in xBR
887 
888  out.template ref<3, 3>() = col;
889  out.template ref<3, 2>() = col;
890  out.template ref<2, 3>() = col;
891  }
892 
893  template <class OutputMatrix>
894  static void blendLineDiagonal(uint32_t col, OutputMatrix& out)
895  {
896  alphaGrad<1, 2>(out.template ref<scale - 1, scale / 2 >(), col);
897  alphaGrad<1, 2>(out.template ref<scale - 2, scale / 2 + 1>(), col);
898  out.template ref<scale - 1, scale - 1>() = col;
899  }
900 
901  template <class OutputMatrix>
902  static void blendCorner(uint32_t col, OutputMatrix& out)
903  {
904  //model a round corner
905  alphaGrad<68, 100>(out.template ref<3, 3>(), col); //exact: 0.6848532563
906  alphaGrad< 9, 100>(out.template ref<3, 2>(), col); //0.08677704501
907  alphaGrad< 9, 100>(out.template ref<2, 3>(), col); //0.08677704501
908  }
909 };
910 
911 
912 template <class ColorGradient>
913 struct Scaler5x : public ColorGradient
914 {
915  static const int scale = 5;
916 
917  template <unsigned int M, unsigned int N> //bring template function into scope for GCC
918  static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) { ColorGradient::template alphaGrad<M, N>(pixBack, pixFront); }
919 
920 
921  template <class OutputMatrix>
922  static void blendLineShallow(uint32_t col, OutputMatrix& out)
923  {
924  alphaGrad<1, 4>(out.template ref<scale - 1, 0>(), col);
925  alphaGrad<1, 4>(out.template ref<scale - 2, 2>(), col);
926  alphaGrad<1, 4>(out.template ref<scale - 3, 4>(), col);
927 
928  alphaGrad<3, 4>(out.template ref<scale - 1, 1>(), col);
929  alphaGrad<3, 4>(out.template ref<scale - 2, 3>(), col);
930 
931  out.template ref<scale - 1, 2>() = col;
932  out.template ref<scale - 1, 3>() = col;
933  out.template ref<scale - 1, 4>() = col;
934  out.template ref<scale - 2, 4>() = col;
935  }
936 
937  template <class OutputMatrix>
938  static void blendLineSteep(uint32_t col, OutputMatrix& out)
939  {
940  alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col);
941  alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col);
942  alphaGrad<1, 4>(out.template ref<4, scale - 3>(), col);
943 
944  alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col);
945  alphaGrad<3, 4>(out.template ref<3, scale - 2>(), col);
946 
947  out.template ref<2, scale - 1>() = col;
948  out.template ref<3, scale - 1>() = col;
949  out.template ref<4, scale - 1>() = col;
950  out.template ref<4, scale - 2>() = col;
951  }
952 
953  template <class OutputMatrix>
954  static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out)
955  {
956  alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col);
957  alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col);
958  alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col);
959 
960  alphaGrad<1, 4>(out.template ref<scale - 1, 0>(), col);
961  alphaGrad<1, 4>(out.template ref<scale - 2, 2>(), col);
962  alphaGrad<3, 4>(out.template ref<scale - 1, 1>(), col);
963 
964  alphaGrad<2, 3>(out.template ref<3, 3>(), col);
965 
966  out.template ref<2, scale - 1>() = col;
967  out.template ref<3, scale - 1>() = col;
968  out.template ref<4, scale - 1>() = col;
969 
970  out.template ref<scale - 1, 2>() = col;
971  out.template ref<scale - 1, 3>() = col;
972  }
973 
974  template <class OutputMatrix>
975  static void blendLineDiagonal(uint32_t col, OutputMatrix& out)
976  {
977  alphaGrad<1, 8>(out.template ref<scale - 1, scale / 2 >(), col); //conflict with other rotations for this odd scale
978  alphaGrad<1, 8>(out.template ref<scale - 2, scale / 2 + 1>(), col);
979  alphaGrad<1, 8>(out.template ref<scale - 3, scale / 2 + 2>(), col); //
980 
981  alphaGrad<7, 8>(out.template ref<4, 3>(), col);
982  alphaGrad<7, 8>(out.template ref<3, 4>(), col);
983 
984  out.template ref<4, 4>() = col;
985  }
986 
987  template <class OutputMatrix>
988  static void blendCorner(uint32_t col, OutputMatrix& out)
989  {
990  //model a round corner
991  alphaGrad<86, 100>(out.template ref<4, 4>(), col); //exact: 0.8631434088
992  alphaGrad<23, 100>(out.template ref<4, 3>(), col); //0.2306749731
993  alphaGrad<23, 100>(out.template ref<3, 4>(), col); //0.2306749731
994  //alphaGrad<1, 64>(out.template ref<4, 2>(), col); //0.01676812367 -> negligible + avoid conflicts with other rotations for this odd scale
995  //alphaGrad<1, 64>(out.template ref<2, 4>(), col); //0.01676812367
996  }
997 };
998 
999 
1000 template <class ColorGradient>
1001 struct Scaler6x : public ColorGradient
1002 {
1003  static const int scale = 6;
1004 
1005  template <unsigned int M, unsigned int N> //bring template function into scope for GCC
1006  static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) { ColorGradient::template alphaGrad<M, N>(pixBack, pixFront); }
1007 
1008 
1009  template <class OutputMatrix>
1010  static void blendLineShallow(uint32_t col, OutputMatrix& out)
1011  {
1012  alphaGrad<1, 4>(out.template ref<scale - 1, 0>(), col);
1013  alphaGrad<1, 4>(out.template ref<scale - 2, 2>(), col);
1014  alphaGrad<1, 4>(out.template ref<scale - 3, 4>(), col);
1015 
1016  alphaGrad<3, 4>(out.template ref<scale - 1, 1>(), col);
1017  alphaGrad<3, 4>(out.template ref<scale - 2, 3>(), col);
1018  alphaGrad<3, 4>(out.template ref<scale - 3, 5>(), col);
1019 
1020  out.template ref<scale - 1, 2>() = col;
1021  out.template ref<scale - 1, 3>() = col;
1022  out.template ref<scale - 1, 4>() = col;
1023  out.template ref<scale - 1, 5>() = col;
1024 
1025  out.template ref<scale - 2, 4>() = col;
1026  out.template ref<scale - 2, 5>() = col;
1027  }
1028 
1029  template <class OutputMatrix>
1030  static void blendLineSteep(uint32_t col, OutputMatrix& out)
1031  {
1032  alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col);
1033  alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col);
1034  alphaGrad<1, 4>(out.template ref<4, scale - 3>(), col);
1035 
1036  alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col);
1037  alphaGrad<3, 4>(out.template ref<3, scale - 2>(), col);
1038  alphaGrad<3, 4>(out.template ref<5, scale - 3>(), col);
1039 
1040  out.template ref<2, scale - 1>() = col;
1041  out.template ref<3, scale - 1>() = col;
1042  out.template ref<4, scale - 1>() = col;
1043  out.template ref<5, scale - 1>() = col;
1044 
1045  out.template ref<4, scale - 2>() = col;
1046  out.template ref<5, scale - 2>() = col;
1047  }
1048 
1049  template <class OutputMatrix>
1050  static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out)
1051  {
1052  alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col);
1053  alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col);
1054  alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col);
1055  alphaGrad<3, 4>(out.template ref<3, scale - 2>(), col);
1056 
1057  alphaGrad<1, 4>(out.template ref<scale - 1, 0>(), col);
1058  alphaGrad<1, 4>(out.template ref<scale - 2, 2>(), col);
1059  alphaGrad<3, 4>(out.template ref<scale - 1, 1>(), col);
1060  alphaGrad<3, 4>(out.template ref<scale - 2, 3>(), col);
1061 
1062  out.template ref<2, scale - 1>() = col;
1063  out.template ref<3, scale - 1>() = col;
1064  out.template ref<4, scale - 1>() = col;
1065  out.template ref<5, scale - 1>() = col;
1066 
1067  out.template ref<4, scale - 2>() = col;
1068  out.template ref<5, scale - 2>() = col;
1069 
1070  out.template ref<scale - 1, 2>() = col;
1071  out.template ref<scale - 1, 3>() = col;
1072  }
1073 
1074  template <class OutputMatrix>
1075  static void blendLineDiagonal(uint32_t col, OutputMatrix& out)
1076  {
1077  alphaGrad<1, 2>(out.template ref<scale - 1, scale / 2 >(), col);
1078  alphaGrad<1, 2>(out.template ref<scale - 2, scale / 2 + 1>(), col);
1079  alphaGrad<1, 2>(out.template ref<scale - 3, scale / 2 + 2>(), col);
1080 
1081  out.template ref<scale - 2, scale - 1>() = col;
1082  out.template ref<scale - 1, scale - 1>() = col;
1083  out.template ref<scale - 1, scale - 2>() = col;
1084  }
1085 
1086  template <class OutputMatrix>
1087  static void blendCorner(uint32_t col, OutputMatrix& out)
1088  {
1089  //model a round corner
1090  alphaGrad<97, 100>(out.template ref<5, 5>(), col); //exact: 0.9711013910
1091  alphaGrad<42, 100>(out.template ref<4, 5>(), col); //0.4236372243
1092  alphaGrad<42, 100>(out.template ref<5, 4>(), col); //0.4236372243
1093  alphaGrad< 6, 100>(out.template ref<5, 3>(), col); //0.05652034508
1094  alphaGrad< 6, 100>(out.template ref<3, 5>(), col); //0.05652034508
1095  }
1096 };
1097 
1098 //------------------------------------------------------------------------------------
1099 
1100 struct ColorDistanceRGB
1101 {
1102  static double dist(uint32_t pix1, uint32_t pix2, double /*luminanceWeight*/)
1103  {
1104  return distYCbCrBuffered(pix1, pix2);
1105 
1106  //if (pix1 == pix2) //about 4% perf boost
1107  // return 0;
1108  //return distYCbCr(pix1, pix2, luminanceWeight);
1109  }
1110 };
1111 
1112 struct ColorDistanceARGB
1113 {
1114  static double dist(uint32_t pix1, uint32_t pix2, double /*luminanceWeight*/)
1115  {
1116  const double a1 = getAlpha(pix1) / 255.0 ;
1117  const double a2 = getAlpha(pix2) / 255.0 ;
1118  /*
1119  Requirements for a color distance handling alpha channel: with a1, a2 in [0, 1]
1120 
1121  1. if a1 = a2, distance should be: a1 * distYCbCr()
1122  2. if a1 = 0, distance should be: a2 * distYCbCr(black, white) = a2 * 255
1123  3. if a1 = 1, ??? maybe: 255 * (1 - a2) + a2 * distYCbCr()
1124  */
1125 
1126  //return std::min(a1, a2) * distYCbCrBuffered(pix1, pix2) + 255 * abs(a1 - a2);
1127  //=> following code is 15% faster:
1128  const double d = distYCbCrBuffered(pix1, pix2);
1129  if (a1 < a2)
1130  return a1 * d + 255 * (a2 - a1);
1131  else
1132  return a2 * d + 255 * (a1 - a2);
1133 
1134  //alternative? return std::sqrt(a1 * a2 * square(distYCbCrBuffered(pix1, pix2)) + square(255 * (a1 - a2)));
1135  }
1136 };
1137 
1138 
1139 struct ColorDistanceUnbufferedARGB
1140 {
1141  static double dist(uint32_t pix1, uint32_t pix2, double luminanceWeight)
1142  {
1143  const double a1 = getAlpha(pix1) / 255.0 ;
1144  const double a2 = getAlpha(pix2) / 255.0 ;
1145 
1146  const double d = distYCbCr(pix1, pix2, luminanceWeight);
1147  if (a1 < a2)
1148  return a1 * d + 255 * (a2 - a1);
1149  else
1150  return a2 * d + 255 * (a1 - a2);
1151  }
1152 };
1153 
1154 
1155 struct ColorGradientRGB
1156 {
1157  template <unsigned int M, unsigned int N>
1158  static void alphaGrad(uint32_t& pixBack, uint32_t pixFront)
1159  {
1160  pixBack = gradientRGB<M, N>(pixFront, pixBack);
1161  }
1162 };
1163 
1164 struct ColorGradientARGB
1165 {
1166  template <unsigned int M, unsigned int N>
1167  static void alphaGrad(uint32_t& pixBack, uint32_t pixFront)
1168  {
1169  pixBack = gradientARGB<M, N>(pixFront, pixBack);
1170  }
1171 };
1172 }
1173 
1174 
1175 void xbrz::scale(size_t factor, const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, ColorFormat colFmt, const xbrz::ScalerCfg& cfg, int yFirst, int yLast)
1176 {
1177  if (factor == 1)
1178  {
1179  std::copy(src + yFirst * srcWidth, src + yLast * srcWidth, trg);
1180  return;
1181  }
1182 
1183  static_assert(SCALE_FACTOR_MAX == 6);
1184  switch (colFmt)
1185  {
1186  case ColorFormat::RGB:
1187  switch (factor)
1188  {
1189  case 2:
1190  return scaleImage<Scaler2x<ColorGradientRGB>, ColorDistanceRGB, OobReaderDuplicate>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
1191  case 3:
1192  return scaleImage<Scaler3x<ColorGradientRGB>, ColorDistanceRGB, OobReaderDuplicate>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
1193  case 4:
1194  return scaleImage<Scaler4x<ColorGradientRGB>, ColorDistanceRGB, OobReaderDuplicate>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
1195  case 5:
1196  return scaleImage<Scaler5x<ColorGradientRGB>, ColorDistanceRGB, OobReaderDuplicate>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
1197  case 6:
1198  return scaleImage<Scaler6x<ColorGradientRGB>, ColorDistanceRGB, OobReaderDuplicate>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
1199  }
1200  break;
1201 
1202  case ColorFormat::ARGB:
1203  switch (factor)
1204  {
1205  case 2:
1206  return scaleImage<Scaler2x<ColorGradientARGB>, ColorDistanceARGB, OobReaderTransparent>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
1207  case 3:
1208  return scaleImage<Scaler3x<ColorGradientARGB>, ColorDistanceARGB, OobReaderTransparent>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
1209  case 4:
1210  return scaleImage<Scaler4x<ColorGradientARGB>, ColorDistanceARGB, OobReaderTransparent>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
1211  case 5:
1212  return scaleImage<Scaler5x<ColorGradientARGB>, ColorDistanceARGB, OobReaderTransparent>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
1213  case 6:
1214  return scaleImage<Scaler6x<ColorGradientARGB>, ColorDistanceARGB, OobReaderTransparent>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
1215  }
1216  break;
1217 
1218  case ColorFormat::ARGB_UNBUFFERED:
1219  switch (factor)
1220  {
1221  case 2:
1222  return scaleImage<Scaler2x<ColorGradientARGB>, ColorDistanceUnbufferedARGB, OobReaderTransparent>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
1223  case 3:
1224  return scaleImage<Scaler3x<ColorGradientARGB>, ColorDistanceUnbufferedARGB, OobReaderTransparent>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
1225  case 4:
1226  return scaleImage<Scaler4x<ColorGradientARGB>, ColorDistanceUnbufferedARGB, OobReaderTransparent>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
1227  case 5:
1228  return scaleImage<Scaler5x<ColorGradientARGB>, ColorDistanceUnbufferedARGB, OobReaderTransparent>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
1229  case 6:
1230  return scaleImage<Scaler6x<ColorGradientARGB>, ColorDistanceUnbufferedARGB, OobReaderTransparent>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
1231  }
1232  break;
1233  }
1234  assert(false);
1235 }
1236 
1237 
1238 bool xbrz::equalColorTest(uint32_t col1, uint32_t col2, ColorFormat colFmt, double luminanceWeight, double equalColorTolerance)
1239 {
1240  switch (colFmt)
1241  {
1242  case ColorFormat::RGB:
1243  return ColorDistanceRGB::dist(col1, col2, luminanceWeight) < equalColorTolerance;
1244  case ColorFormat::ARGB:
1245  return ColorDistanceARGB::dist(col1, col2, luminanceWeight) < equalColorTolerance;
1246  case ColorFormat::ARGB_UNBUFFERED:
1247  return ColorDistanceUnbufferedARGB::dist(col1, col2, luminanceWeight) < equalColorTolerance;
1248  }
1249  assert(false);
1250  return false;
1251 }
1252 
1253 
1254 void xbrz::bilinearScale(const uint32_t* src, int srcWidth, int srcHeight,
1255  /**/ uint32_t* trg, int trgWidth, int trgHeight)
1256 {
1257  bilinearScale(src, srcWidth, srcHeight, srcWidth * sizeof(uint32_t),
1258  trg, trgWidth, trgHeight, trgWidth * sizeof(uint32_t),
1259  0, trgHeight, [](uint32_t pix) { return pix; });
1260 }
1261 
1262 
1263 void xbrz::nearestNeighborScale(const uint32_t* src, int srcWidth, int srcHeight,
1264  /**/ uint32_t* trg, int trgWidth, int trgHeight)
1265 {
1266  nearestNeighborScale(src, srcWidth, srcHeight, srcWidth * sizeof(uint32_t),
1267  trg, trgWidth, trgHeight, trgWidth * sizeof(uint32_t),
1268  0, trgHeight, [](uint32_t pix) { return pix; });
1269 }
1270 
1271 
1272 #if 0
1273 //#include <ppl.h>
1274 void bilinearScaleCpu(const uint32_t* src, int srcWidth, int srcHeight,
1275  /**/ uint32_t* trg, int trgWidth, int trgHeight)
1276 {
1277  const int TASK_GRANULARITY = 16;
1278 
1279  concurrency::task_group tg;
1280 
1281  for (int i = 0; i < trgHeight; i += TASK_GRANULARITY)
1282  tg.run([=]
1283  {
1284  const int iLast = std::min(i + TASK_GRANULARITY, trgHeight);
1285  xbrz::bilinearScale(src, srcWidth, srcHeight, srcWidth * sizeof(uint32_t),
1286  trg, trgWidth, trgHeight, trgWidth * sizeof(uint32_t),
1287  i, iLast, [](uint32_t pix) { return pix; });
1288  });
1289  tg.wait();
1290 }
1291 
1292 
1293 //Perf: AMP vs CPU: merely ~10% shorter runtime (scaling 1280x800 -> 1920x1080)
1294 //#include <amp.h>
1295 void bilinearScaleAmp(const uint32_t* src, int srcWidth, int srcHeight, //throw concurrency::runtime_exception
1296  /**/ uint32_t* trg, int trgWidth, int trgHeight)
1297 {
1298  //C++ AMP reference: https://msdn.microsoft.com/en-us/library/hh289390.aspx
1299  //introduction to C++ AMP: https://msdn.microsoft.com/en-us/magazine/hh882446.aspx
1300  using namespace concurrency;
1301  //TODO: pitch
1302 
1303  if (srcHeight <= 0 || srcWidth <= 0) return;
1304 
1305  const float scaleX = static_cast<float>(trgWidth ) / srcWidth;
1306  const float scaleY = static_cast<float>(trgHeight) / srcHeight;
1307 
1308  array_view<const uint32_t, 2> srcView(srcHeight, srcWidth, src);
1309  array_view< uint32_t, 2> trgView(trgHeight, trgWidth, trg);
1310  trgView.discard_data();
1311 
1312  parallel_for_each(trgView.extent, [=](index<2> idx) restrict(amp) //throw ?
1313  {
1314  const int y = idx[0];
1315  const int x = idx[1];
1316  //Perf notes:
1317  // -> float-based calculation is (almost) 2x as fas as double!
1318  // -> no noticeable improvement via tiling: https://msdn.microsoft.com/en-us/magazine/hh882447.aspx
1319  // -> no noticeable improvement with restrict(amp,cpu)
1320  // -> iterating over y-axis only is significantly slower!
1321  // -> pre-calculating x,y-dependent variables in a buffer + array_view<> is ~ 20 % slower!
1322  const int y1 = srcHeight * y / trgHeight;
1323  int y2 = y1 + 1;
1324  if (y2 == srcHeight) --y2;
1325 
1326  const float yy1 = y / scaleY - y1;
1327  const float y2y = 1 - yy1;
1328  //-------------------------------------
1329  const int x1 = srcWidth * x / trgWidth;
1330  int x2 = x1 + 1;
1331  if (x2 == srcWidth) --x2;
1332 
1333  const float xx1 = x / scaleX - x1;
1334  const float x2x = 1 - xx1;
1335  //-------------------------------------
1336  const float x2xy2y = x2x * y2y;
1337  const float xx1y2y = xx1 * y2y;
1338  const float x2xyy1 = x2x * yy1;
1339  const float xx1yy1 = xx1 * yy1;
1340 
1341  auto interpolate = [=](int offset)
1342  {
1343  /*
1344  https://en.wikipedia.org/wiki/Bilinear_interpolation
1345  (c11(x2 - x) + c21(x - x1)) * (y2 - y ) +
1346  (c12(x2 - x) + c22(x - x1)) * (y - y1)
1347  */
1348  const auto c11 = (srcView(y1, x1) >> (8 * offset)) & 0xff;
1349  const auto c21 = (srcView(y1, x2) >> (8 * offset)) & 0xff;
1350  const auto c12 = (srcView(y2, x1) >> (8 * offset)) & 0xff;
1351  const auto c22 = (srcView(y2, x2) >> (8 * offset)) & 0xff;
1352 
1353  return c11 * x2xy2y + c21 * xx1y2y +
1354  c12 * x2xyy1 + c22 * xx1yy1;
1355  };
1356 
1357  const float bi = interpolate(0);
1358  const float gi = interpolate(1);
1359  const float ri = interpolate(2);
1360  const float ai = interpolate(3);
1361 
1362  const auto b = static_cast<uint32_t>(bi + 0.5f);
1363  const auto g = static_cast<uint32_t>(gi + 0.5f);
1364  const auto r = static_cast<uint32_t>(ri + 0.5f);
1365  const auto a = static_cast<uint32_t>(ai + 0.5f);
1366 
1367  trgView(y, x) = (a << 24) | (r << 16) | (g << 8) | b;
1368  });
1369  trgView.synchronize(); //throw ?
1370 }
1371 #endif
LIKELY
#define LIKELY
Definition: global.hpp:69
ai
A small explanation about what's going on here: Each action has access to two game_info objects First...
Definition: actions.cpp:59
utf8::index
std::size_t index(std::string_view str, const std::size_t index)
Codepoint index corresponding to the nth character in a UTF-8 string.
Definition: unicode.cpp:70
xbrz
Definition: xbrz.hpp:27
xbrz::getRed
unsigned char getRed(uint32_t pix)
Definition: xbrz_tools.hpp:31
xbrz::makePixel
uint32_t makePixel(unsigned char a, unsigned char r, unsigned char g, unsigned char b)
Definition: xbrz_tools.hpp:35
xbrz::getBlue
unsigned char getBlue(uint32_t pix)
Definition: xbrz_tools.hpp:33
xbrz::getGreen
unsigned char getGreen(uint32_t pix)
Definition: xbrz_tools.hpp:32
xbrz::scale
void scale(size_t factor, const uint32_t *src, uint32_t *trg, int srcWidth, int srcHeight, ColorFormat colFmt, const ScalerCfg &cfg=ScalerCfg(), int yFirst=0, int yLast=std::numeric_limits< int >::max())
Definition: xbrz.cpp:1175
xbrz::equalColorTest
bool equalColorTest(uint32_t col1, uint32_t col2, ColorFormat colFmt, double luminanceWeight, double equalColorTolerance)
Definition: xbrz.cpp:1238
xbrz::nearestNeighborScale
void nearestNeighborScale(const uint32_t *src, int srcWidth, int srcHeight, uint32_t *trg, int trgWidth, int trgHeight)
Definition: xbrz.cpp:1263
xbrz::SCALE_FACTOR_MAX
const int SCALE_FACTOR_MAX
Definition: xbrz.hpp:50
xbrz::fillBlock
void fillBlock(Pix *trg, int pitch, Pix col, int blockWidth, int blockHeight)
Definition: xbrz_tools.hpp:59
xbrz::getAlpha
unsigned char getAlpha(uint32_t pix)
Definition: xbrz_tools.hpp:30
xbrz::bilinearScale
void bilinearScale(const uint32_t *src, int srcWidth, int srcHeight, uint32_t *trg, int trgWidth, int trgHeight)
Definition: xbrz.cpp:1254
xbrz::ColorFormat
ColorFormat
Definition: xbrz.hpp:44
src
rect src
Non-transparent portion of the surface to compose.
Definition: spritesheet_generator.cpp:70
xbrz::ScalerCfg::centerDirectionBias
double centerDirectionBias
Definition: xbrz_config.hpp:28
xbrz::ScalerCfg::luminanceWeight
double luminanceWeight
Definition: xbrz_config.hpp:26
xbrz::ScalerCfg::steepDirectionThreshold
double steepDirectionThreshold
Definition: xbrz_config.hpp:30
xbrz::ScalerCfg::dominantDirectionThreshold
double dominantDirectionThreshold
Definition: xbrz_config.hpp:29
xbrz::ScalerCfg::equalColorTolerance
double equalColorTolerance
Definition: xbrz_config.hpp:27
p
mock_party p
Definition: test_formula_core.cpp:68
n
static map_location::direction n
Definition: test_map_location.cpp:25
i
#define i
g
#define g
d
#define d
e
#define e
DEF_GETTER
#define DEF_GETTER(x)
Definition: xbrz.cpp:361
h
#define h
f
#define f
c
#define c
b
#define b
FORCE_INLINE
#define FORCE_INLINE
Definition: xbrz.cpp:84
xbrz_tools.hpp
Generated by doxygen 1.9.1