OpenTTD
32bpp_sse_func.hpp
Go to the documentation of this file.
1 /* $Id: 32bpp_sse_func.hpp 27020 2014-10-15 18:31:37Z rubidium $ */
2 
3 /*
4  * This file is part of OpenTTD.
5  * OpenTTD is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, version 2.
6  * OpenTTD is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
7  * See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with OpenTTD. If not, see <http://www.gnu.org/licenses/>.
8  */
9 
12 #ifndef BLITTER_32BPP_SSE_FUNC_HPP
13 #define BLITTER_32BPP_SSE_FUNC_HPP
14 
15 #ifdef WITH_SSE
16 
17 static inline void InsertFirstUint32(const uint32 value, __m128i &into)
18 {
19 #if (SSE_VERSION >= 4)
20  into = _mm_insert_epi32(into, value, 0);
21 #else
22  into = _mm_insert_epi16(into, value, 0);
23  into = _mm_insert_epi16(into, value >> 16, 1);
24 #endif
25 }
26 
27 static inline void InsertSecondUint32(const uint32 value, __m128i &into)
28 {
29 #if (SSE_VERSION >= 4)
30  into = _mm_insert_epi32(into, value, 1);
31 #else
32  into = _mm_insert_epi16(into, value, 2);
33  into = _mm_insert_epi16(into, value >> 16, 3);
34 #endif
35 }
36 
37 static inline void LoadUint64(const uint64 value, __m128i &into)
38 {
39 #ifdef _SQ64
40  into = _mm_cvtsi64_si128(value);
41 #else
42  #if (SSE_VERSION >= 4)
43  into = _mm_cvtsi32_si128(value);
44  InsertSecondUint32(value >> 32, into);
45  #else
46  (*(um128i*) &into).m128i_u64[0] = value;
47  #endif
48 #endif
49 }
50 
51 static inline __m128i PackUnsaturated(__m128i from, const __m128i &mask)
52 {
53 #if (SSE_VERSION == 2)
54  from = _mm_and_si128(from, mask); // PAND, wipe high bytes to keep low bytes when packing
55  return _mm_packus_epi16(from, from); // PACKUSWB, pack 2 colours (with saturation)
56 #else
57  return _mm_shuffle_epi8(from, mask);
58 #endif
59 }
60 
61 static inline __m128i DistributeAlpha(const __m128i from, const __m128i &mask)
62 {
63 #if (SSE_VERSION == 2)
64  __m128i alphaAB = _mm_shufflelo_epi16(from, 0x3F); // PSHUFLW, put alpha1 in front of each rgb1
65  return _mm_shufflehi_epi16(alphaAB, 0x3F); // PSHUFHW, put alpha2 in front of each rgb2
66 #else
67  return _mm_shuffle_epi8(from, mask);
68 #endif
69 }
70 
71 static inline __m128i AlphaBlendTwoPixels(__m128i src, __m128i dst, const __m128i &distribution_mask, const __m128i &pack_mask)
72 {
73  __m128i srcAB = _mm_unpacklo_epi8(src, _mm_setzero_si128()); // PUNPCKLBW, expand each uint8 into uint16
74  __m128i dstAB = _mm_unpacklo_epi8(dst, _mm_setzero_si128());
75 
76  __m128i alphaAB = _mm_cmpgt_epi16(srcAB, _mm_setzero_si128()); // PCMPGTW, if (alpha > 0) a++;
77  alphaAB = _mm_srli_epi16(alphaAB, 15);
78  alphaAB = _mm_add_epi16(alphaAB, srcAB);
79  alphaAB = DistributeAlpha(alphaAB, distribution_mask);
80 
81  srcAB = _mm_sub_epi16(srcAB, dstAB); // PSUBW, (r - Cr)
82  srcAB = _mm_mullo_epi16(srcAB, alphaAB); // PMULLW, a*(r - Cr)
83  srcAB = _mm_srli_epi16(srcAB, 8); // PSRLW, a*(r - Cr)/256
84  srcAB = _mm_add_epi16(srcAB, dstAB); // PADDW, a*(r - Cr)/256 + Cr
85  return PackUnsaturated(srcAB, pack_mask);
86 }
87 
88 /* Darken 2 pixels.
89  * rgb = rgb * ((256/4) * 4 - (alpha/4)) / ((256/4) * 4)
90  */
91 static inline __m128i DarkenTwoPixels(__m128i src, __m128i dst, const __m128i &distribution_mask, const __m128i &tr_nom_base)
92 {
93  __m128i srcAB = _mm_unpacklo_epi8(src, _mm_setzero_si128());
94  __m128i dstAB = _mm_unpacklo_epi8(dst, _mm_setzero_si128());
95  __m128i alphaAB = DistributeAlpha(srcAB, distribution_mask);
96  alphaAB = _mm_srli_epi16(alphaAB, 2); // Reduce to 64 levels of shades so the max value fits in 16 bits.
97  __m128i nom = _mm_sub_epi16(tr_nom_base, alphaAB);
98  dstAB = _mm_mullo_epi16(dstAB, nom);
99  dstAB = _mm_srli_epi16(dstAB, 8);
100  return _mm_packus_epi16(dstAB, dstAB);
101 }
102 
103 IGNORE_UNINITIALIZED_WARNING_START
104 static Colour ReallyAdjustBrightness(Colour colour, uint8 brightness)
105 {
106  uint64 c16 = colour.b | (uint64) colour.g << 16 | (uint64) colour.r << 32;
107  c16 *= brightness;
108  uint64 c16_ob = c16; // Helps out of order execution.
109  c16 /= Blitter_32bppBase::DEFAULT_BRIGHTNESS;
110  c16 &= 0x01FF01FF01FFULL;
111 
112  /* Sum overbright (maximum for each rgb is 508, 9 bits, -255 is changed in -256 so we just have to take the 8 lower bits into account). */
113  c16_ob = (((c16_ob >> (8 + 7)) & 0x0100010001ULL) * 0xFF) & c16;
114  const uint ob = ((uint16) c16_ob + (uint16) (c16_ob >> 16) + (uint16) (c16_ob >> 32)) / 2;
115 
116  const uint32 alpha32 = colour.data & 0xFF000000;
117  __m128i ret;
118  LoadUint64(c16, ret);
119  if (ob != 0) {
120  __m128i ob128 = _mm_cvtsi32_si128(ob);
121  ob128 = _mm_shufflelo_epi16(ob128, 0xC0);
122  __m128i white = OVERBRIGHT_VALUE_MASK;
123  __m128i c128 = ret;
124  ret = _mm_subs_epu16(white, c128); // PSUBUSW, (255 - rgb)
125  ret = _mm_mullo_epi16(ret, ob128); // PMULLW, ob*(255 - rgb)
126  ret = _mm_srli_epi16(ret, 8); // PSRLW, ob*(255 - rgb)/256
127  ret = _mm_add_epi16(ret, c128); // PADDW, ob*(255 - rgb)/256 + rgb
128  }
129 
130  ret = _mm_packus_epi16(ret, ret); // PACKUSWB, saturate and pack.
131  return alpha32 | _mm_cvtsi128_si32(ret);
132 }
133 IGNORE_UNINITIALIZED_WARNING_STOP
134 
138 static inline Colour AdjustBrightneSSE(Colour colour, uint8 brightness)
139 {
140  /* Shortcut for normal brightness. */
141  if (brightness == Blitter_32bppBase::DEFAULT_BRIGHTNESS) return colour;
142 
143  return ReallyAdjustBrightness(colour, brightness);
144 }
145 
146 static inline __m128i AdjustBrightnessOfTwoPixels(__m128i from, uint32 brightness)
147 {
148 #if (SSE_VERSION < 3)
149  NOT_REACHED();
150 #else
151  /* The following dataflow differs from the one of AdjustBrightness() only for alpha.
152  * In order to keep alpha in colAB, insert a 1 in a unused brightness byte (a*1->a).
153  * OK, not a 1 but DEFAULT_BRIGHTNESS to compensate the div.
154  */
155  brightness &= 0xFF00FF00;
156  brightness += Blitter_32bppBase::DEFAULT_BRIGHTNESS;
157 
158  __m128i colAB = _mm_unpacklo_epi8(from, _mm_setzero_si128());
159  __m128i briAB = _mm_cvtsi32_si128(brightness);
160  briAB = _mm_shuffle_epi8(briAB, BRIGHTNESS_LOW_CONTROL_MASK); // DEFAULT_BRIGHTNESS in 0, 0x00 in 2.
161  colAB = _mm_mullo_epi16(colAB, briAB);
162  __m128i colAB_ob = _mm_srli_epi16(colAB, 8 + 7);
163  colAB = _mm_srli_epi16(colAB, 7);
164 
165  /* Sum overbright.
166  * Maximum for each rgb is 508 => 9 bits. The highest bit tells if there is overbright.
167  * -255 is changed in -256 so we just have to take the 8 lower bits into account.
168  */
169  colAB = _mm_and_si128(colAB, BRIGHTNESS_DIV_CLEANER);
170  colAB_ob = _mm_and_si128(colAB_ob, OVERBRIGHT_PRESENCE_MASK);
171  colAB_ob = _mm_mullo_epi16(colAB_ob, OVERBRIGHT_VALUE_MASK);
172  colAB_ob = _mm_and_si128(colAB_ob, colAB);
173  __m128i obAB = _mm_hadd_epi16(_mm_hadd_epi16(colAB_ob, _mm_setzero_si128()), _mm_setzero_si128());
174 
175  obAB = _mm_srli_epi16(obAB, 1); // Reduce overbright strength.
176  obAB = _mm_shuffle_epi8(obAB, OVERBRIGHT_CONTROL_MASK);
177  __m128i retAB = OVERBRIGHT_VALUE_MASK; // ob_mask is equal to white.
178  retAB = _mm_subs_epu16(retAB, colAB); // (255 - rgb)
179  retAB = _mm_mullo_epi16(retAB, obAB); // ob*(255 - rgb)
180  retAB = _mm_srli_epi16(retAB, 8); // ob*(255 - rgb)/256
181  retAB = _mm_add_epi16(retAB, colAB); // ob*(255 - rgb)/256 + rgb
182 
183  return _mm_packus_epi16(retAB, retAB);
184 #endif
185 }
186 
187 #if FULL_ANIMATION == 0
188 
195 IGNORE_UNINITIALIZED_WARNING_START
196 template <BlitterMode mode, Blitter_32bppSSE2::ReadMode read_mode, Blitter_32bppSSE2::BlockType bt_last, bool translucent>
197 #if (SSE_VERSION == 2)
198 inline void Blitter_32bppSSE2::Draw(const Blitter::BlitterParams *bp, ZoomLevel zoom)
199 #elif (SSE_VERSION == 3)
200 inline void Blitter_32bppSSSE3::Draw(const Blitter::BlitterParams *bp, ZoomLevel zoom)
201 #elif (SSE_VERSION == 4)
202 inline void Blitter_32bppSSE4::Draw(const Blitter::BlitterParams *bp, ZoomLevel zoom)
203 #endif
204 {
205  const byte * const remap = bp->remap;
206  Colour *dst_line = (Colour *) bp->dst + bp->top * bp->pitch + bp->left;
207  int effective_width = bp->width;
208 
209  /* Find where to start reading in the source sprite. */
210  const SpriteData * const sd = (const SpriteData *) bp->sprite;
211  const SpriteInfo * const si = &sd->infos[zoom];
212  const MapValue *src_mv_line = (const MapValue *) &sd->data[si->mv_offset] + bp->skip_top * si->sprite_width;
213  const Colour *src_rgba_line = (const Colour *) ((const byte *) &sd->data[si->sprite_offset] + bp->skip_top * si->sprite_line_size);
214 
215  if (read_mode != RM_WITH_MARGIN) {
216  src_rgba_line += bp->skip_left;
217  src_mv_line += bp->skip_left;
218  }
219  const MapValue *src_mv = src_mv_line;
220 
221  /* Load these variables into register before loop. */
222 #if (SSE_VERSION == 2)
223  const __m128i clear_hi = CLEAR_HIGH_BYTE_MASK;
224  #define ALPHA_BLEND_PARAM_1 clear_hi
225  #define ALPHA_BLEND_PARAM_2 clear_hi
226  #define DARKEN_PARAM_1 tr_nom_base
227  #define DARKEN_PARAM_2 tr_nom_base
228 #else
229  const __m128i a_cm = ALPHA_CONTROL_MASK;
230  const __m128i pack_low_cm = PACK_LOW_CONTROL_MASK;
231  #define ALPHA_BLEND_PARAM_1 a_cm
232  #define ALPHA_BLEND_PARAM_2 pack_low_cm
233  #define DARKEN_PARAM_1 a_cm
234  #define DARKEN_PARAM_2 tr_nom_base
235 #endif
236  const __m128i tr_nom_base = TRANSPARENT_NOM_BASE;
237 
238  for (int y = bp->height; y != 0; y--) {
239  Colour *dst = dst_line;
240  const Colour *src = src_rgba_line + META_LENGTH;
241  if (mode == BM_COLOUR_REMAP || mode == BM_CRASH_REMAP) src_mv = src_mv_line;
242 
243  if (read_mode == RM_WITH_MARGIN) {
244  assert(bt_last == BT_NONE); // or you must ensure block type is preserved
245  src += src_rgba_line[0].data;
246  dst += src_rgba_line[0].data;
247  if (mode == BM_COLOUR_REMAP || mode == BM_CRASH_REMAP) src_mv += src_rgba_line[0].data;
248  const int width_diff = si->sprite_width - bp->width;
249  effective_width = bp->width - (int) src_rgba_line[0].data;
250  const int delta_diff = (int) src_rgba_line[1].data - width_diff;
251  const int new_width = effective_width - delta_diff;
252  effective_width = delta_diff > 0 ? new_width : effective_width;
253  if (effective_width <= 0) goto next_line;
254  }
255 
256  switch (mode) {
257  default:
258  if (!translucent) {
259  for (uint x = (uint) effective_width; x > 0; x--) {
260  if (src->a) *dst = *src;
261  src++;
262  dst++;
263  }
264  break;
265  }
266 
267  for (uint x = (uint) effective_width / 2; x > 0; x--) {
268  __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src);
269  __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst);
270  _mm_storel_epi64((__m128i*) dst, AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2));
271  src += 2;
272  dst += 2;
273  }
274 
275  if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) {
276  __m128i srcABCD = _mm_cvtsi32_si128(src->data);
277  __m128i dstABCD = _mm_cvtsi32_si128(dst->data);
278  dst->data = _mm_cvtsi128_si32(AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2));
279  }
280  break;
281 
282  case BM_COLOUR_REMAP:
283 #if (SSE_VERSION >= 3)
284  for (uint x = (uint) effective_width / 2; x > 0; x--) {
285  __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src);
286  __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst);
287  uint32 mvX2 = *((uint32 *) const_cast<MapValue *>(src_mv));
288 
289  /* Remap colours. */
290  if (mvX2 & 0x00FF00FF) {
291  #define CMOV_REMAP(m_colour, m_colour_init, m_src, m_m) \
292  /* Written so the compiler uses CMOV. */ \
293  Colour m_colour = m_colour_init; \
294  { \
295  const Colour srcm = (Colour) (m_src); \
296  const uint m = (byte) (m_m); \
297  const uint r = remap[m]; \
298  const Colour cmap = (this->LookupColourInPalette(r).data & 0x00FFFFFF) | (srcm.data & 0xFF000000); \
299  m_colour = r == 0 ? m_colour : cmap; \
300  m_colour = m != 0 ? m_colour : srcm; \
301  }
302 #ifdef _SQ64
303  uint64 srcs = _mm_cvtsi128_si64(srcABCD);
304  uint64 remapped_src = 0;
305  CMOV_REMAP(c0, 0, srcs, mvX2);
306  remapped_src = c0.data;
307  CMOV_REMAP(c1, 0, srcs >> 32, mvX2 >> 16);
308  remapped_src |= (uint64) c1.data << 32;
309  srcABCD = _mm_cvtsi64_si128(remapped_src);
310 #else
311  Colour remapped_src[2];
312  CMOV_REMAP(c0, 0, _mm_cvtsi128_si32(srcABCD), mvX2);
313  remapped_src[0] = c0.data;
314  CMOV_REMAP(c1, 0, src[1], mvX2 >> 16);
315  remapped_src[1] = c1.data;
316  srcABCD = _mm_loadl_epi64((__m128i*) &remapped_src);
317 #endif
318 
319  if ((mvX2 & 0xFF00FF00) != 0x80008000) srcABCD = AdjustBrightnessOfTwoPixels(srcABCD, mvX2);
320  }
321 
322  /* Blend colours. */
323  _mm_storel_epi64((__m128i *) dst, AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2));
324  dst += 2;
325  src += 2;
326  src_mv += 2;
327  }
328 
329  if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) {
330 #else
331  for (uint x = (uint) effective_width; x > 0; x--) {
332 #endif
333  /* In case the m-channel is zero, do not remap this pixel in any way. */
334  __m128i srcABCD;
335  if (src_mv->m) {
336  const uint r = remap[src_mv->m];
337  if (r != 0) {
338  Colour remapped_colour = AdjustBrightneSSE(this->LookupColourInPalette(r), src_mv->v);
339  if (src->a == 255) {
340  *dst = remapped_colour;
341  } else {
342  remapped_colour.a = src->a;
343  srcABCD = _mm_cvtsi32_si128(remapped_colour.data);
344  goto bmcr_alpha_blend_single;
345  }
346  }
347  } else {
348  srcABCD = _mm_cvtsi32_si128(src->data);
349  if (src->a < 255) {
350 bmcr_alpha_blend_single:
351  __m128i dstABCD = _mm_cvtsi32_si128(dst->data);
352  srcABCD = AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2);
353  }
354  dst->data = _mm_cvtsi128_si32(srcABCD);
355  }
356 #if (SSE_VERSION == 2)
357  src_mv++;
358  dst++;
359  src++;
360 #endif
361  }
362  break;
363 
364  case BM_TRANSPARENT:
365  /* Make the current colour a bit more black, so it looks like this image is transparent. */
366  for (uint x = (uint) bp->width / 2; x > 0; x--) {
367  __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src);
368  __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst);
369  _mm_storel_epi64((__m128i *) dst, DarkenTwoPixels(srcABCD, dstABCD, DARKEN_PARAM_1, DARKEN_PARAM_2));
370  src += 2;
371  dst += 2;
372  }
373 
374  if ((bt_last == BT_NONE && bp->width & 1) || bt_last == BT_ODD) {
375  __m128i srcABCD = _mm_cvtsi32_si128(src->data);
376  __m128i dstABCD = _mm_cvtsi32_si128(dst->data);
377  dst->data = _mm_cvtsi128_si32(DarkenTwoPixels(srcABCD, dstABCD, DARKEN_PARAM_1, DARKEN_PARAM_2));
378  }
379  break;
380 
381  case BM_CRASH_REMAP:
382  for (uint x = (uint) bp->width; x > 0; x--) {
383  if (src_mv->m == 0) {
384  if (src->a != 0) {
385  uint8 g = MakeDark(src->r, src->g, src->b);
386  *dst = ComposeColourRGBA(g, g, g, src->a, *dst);
387  }
388  } else {
389  uint r = remap[src_mv->m];
390  if (r != 0) *dst = ComposeColourPANoCheck(this->AdjustBrightness(this->LookupColourInPalette(r), src_mv->v), src->a, *dst);
391  }
392  src_mv++;
393  dst++;
394  src++;
395  }
396  break;
397 
398  case BM_BLACK_REMAP:
399  for (uint x = (uint) bp->width; x > 0; x--) {
400  if (src->a != 0) {
401  *dst = Colour(0, 0, 0);
402  }
403  src_mv++;
404  dst++;
405  src++;
406  }
407  break;
408  }
409 
410 next_line:
411  if (mode == BM_COLOUR_REMAP || mode == BM_CRASH_REMAP) src_mv_line += si->sprite_width;
412  src_rgba_line = (const Colour*) ((const byte*) src_rgba_line + si->sprite_line_size);
413  dst_line += bp->pitch;
414  }
415 }
416 IGNORE_UNINITIALIZED_WARNING_STOP
417 
425 #if (SSE_VERSION == 2)
426 void Blitter_32bppSSE2::Draw(Blitter::BlitterParams *bp, BlitterMode mode, ZoomLevel zoom)
427 #elif (SSE_VERSION == 3)
428 void Blitter_32bppSSSE3::Draw(Blitter::BlitterParams *bp, BlitterMode mode, ZoomLevel zoom)
429 #elif (SSE_VERSION == 4)
430 void Blitter_32bppSSE4::Draw(Blitter::BlitterParams *bp, BlitterMode mode, ZoomLevel zoom)
431 #endif
432 {
433  switch (mode) {
434  default: {
435  if (bp->skip_left != 0 || bp->width <= MARGIN_NORMAL_THRESHOLD) {
436 bm_normal:
437  const BlockType bt_last = (BlockType) (bp->width & 1);
438  switch (bt_last) {
439  default: Draw<BM_NORMAL, RM_WITH_SKIP, BT_EVEN, true>(bp, zoom); return;
440  case BT_ODD: Draw<BM_NORMAL, RM_WITH_SKIP, BT_ODD, true>(bp, zoom); return;
441  }
442  } else {
443  if (((const Blitter_32bppSSE_Base::SpriteData *) bp->sprite)->flags & SF_TRANSLUCENT) {
444  Draw<BM_NORMAL, RM_WITH_MARGIN, BT_NONE, true>(bp, zoom);
445  } else {
446  Draw<BM_NORMAL, RM_WITH_MARGIN, BT_NONE, false>(bp, zoom);
447  }
448  return;
449  }
450  break;
451  }
452  case BM_COLOUR_REMAP:
453  if (((const Blitter_32bppSSE_Base::SpriteData *) bp->sprite)->flags & SF_NO_REMAP) goto bm_normal;
454  if (bp->skip_left != 0 || bp->width <= MARGIN_REMAP_THRESHOLD) {
455  Draw<BM_COLOUR_REMAP, RM_WITH_SKIP, BT_NONE, true>(bp, zoom); return;
456  } else {
457  Draw<BM_COLOUR_REMAP, RM_WITH_MARGIN, BT_NONE, true>(bp, zoom); return;
458  }
459  case BM_TRANSPARENT: Draw<BM_TRANSPARENT, RM_NONE, BT_NONE, true>(bp, zoom); return;
460  case BM_CRASH_REMAP: Draw<BM_CRASH_REMAP, RM_NONE, BT_NONE, true>(bp, zoom); return;
461  case BM_BLACK_REMAP: Draw<BM_BLACK_REMAP, RM_NONE, BT_NONE, true>(bp, zoom); return;
462  }
463 }
464 #endif /* FULL_ANIMATION */
465 
466 #endif /* WITH_SSE */
467 #endif /* BLITTER_32BPP_SSE_FUNC_HPP */