?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Timetable: Implement automate for taken conditional orders.
[openttd-joker.git] / src / blitter / 32bpp_anim_sse4.cpp
blob110c660993b85f826b3421bd5588bbcdd66d2cef
1 /* $Id$ */
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
10 /** @file 32bpp_sse4_anim.cpp Implementation of the SSE4 32 bpp blitter with animation support. */
11
12 #ifdef WITH_SSE
13
14 #include "../stdafx.h"
15 #include "../video/video_driver.hpp"
16 #include "../table/sprites.h"
17 #include "32bpp_anim_sse4.hpp"
18 #include "32bpp_sse_func.hpp"
19
20 #include "../safeguards.h"
21
22 /** Instantiation of the SSE4 32bpp blitter factory. */
23 static FBlitter_32bppSSE4_Anim iFBlitter_32bppSSE4_Anim;
24
25 /**
26 * Draws a sprite to a (screen) buffer. It is templated to allow faster operation.
27 *
28 * @tparam mode blitter mode
29 * @param bp further blitting parameters
30 * @param zoom zoom level at which we are drawing
31 */
32 IGNORE_UNINITIALIZED_WARNING_START
33 template <BlitterMode mode, Blitter_32bppSSE2::ReadMode read_mode, Blitter_32bppSSE2::BlockType bt_last, bool translucent, bool animated>
34 inline void Blitter_32bppSSE4_Anim::Draw(const Blitter::BlitterParams *bp, ZoomLevel zoom)
35 {
36 const byte * const remap = bp->remap;
37 Colour *dst_line = (Colour *)bp->dst + bp->top * bp->pitch + bp->left;
38 uint16 *anim_line = this->anim_buf + this->ScreenToAnimOffset((uint32 *)bp->dst) + bp->top * this->anim_buf_pitch + bp->left;
39 int effective_width = bp->width;
40
41 /* Find where to start reading in the source sprite. */
42 const Blitter_32bppSSE_Base::SpriteData * const sd = (const Blitter_32bppSSE_Base::SpriteData *) bp->sprite;
43 const SpriteInfo * const si = &sd->infos[zoom];
44 const MapValue *src_mv_line = (const MapValue *)&sd->data[si->mv_offset] + bp->skip_top * si->sprite_width;
45 const Colour *src_rgba_line = (const Colour *)((const byte *)&sd->data[si->sprite_offset] + bp->skip_top * si->sprite_line_size);
46
47 if (read_mode != RM_WITH_MARGIN) {
48 src_rgba_line += bp->skip_left;
49 src_mv_line += bp->skip_left;
50 }
51 const MapValue *src_mv = src_mv_line;
52
53 /* Load these variables into register before loop. */
54 const __m128i a_cm = ALPHA_CONTROL_MASK;
55 const __m128i pack_low_cm = PACK_LOW_CONTROL_MASK;
56 const __m128i tr_nom_base = TRANSPARENT_NOM_BASE;
57
58 for (int y = bp->height; y != 0; y--) {
59 Colour *dst = dst_line;
60 const Colour *src = src_rgba_line + META_LENGTH;
61 if (mode != BM_TRANSPARENT) src_mv = src_mv_line;
62 uint16 *anim = anim_line;
63
64 if (read_mode == RM_WITH_MARGIN) {
65 assert(bt_last == BT_NONE); // or you must ensure block type is preserved
66 anim += src_rgba_line[0].data;
67 src += src_rgba_line[0].data;
68 dst += src_rgba_line[0].data;
69 if (mode != BM_TRANSPARENT) src_mv += src_rgba_line[0].data;
70 const int width_diff = si->sprite_width - bp->width;
71 effective_width = bp->width - (int)src_rgba_line[0].data;
72 const int delta_diff = (int)src_rgba_line[1].data - width_diff;
73 const int new_width = effective_width - delta_diff;
74 effective_width = delta_diff > 0 ? new_width : effective_width;
75 if (effective_width <= 0) goto next_line;
76 }
77
78 switch (mode) {
79 default:
80 if (!translucent) {
81 for (uint x = (uint)effective_width; x > 0; x--) {
82 if (src->a) {
83 if (animated) {
84 *anim = *(const uint16*)src_mv;
85 *dst = (src_mv->m >= PALETTE_ANIM_START) ? AdjustBrightneSSE(this->LookupColourInPalette(src_mv->m), src_mv->v) : src->data;
86 }
87 else {
88 *anim = 0;
89 *dst = *src;
90 }
91 }
92 if (animated) src_mv++;
93 anim++;
94 src++;
95 dst++;
96 }
97 break;
98 }
99
100 for (uint x = (uint)effective_width / 2; x != 0; x--) {
101 uint32 mvX2 = *((uint32 *) const_cast<MapValue *>(src_mv));
102 __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src);
103 __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst);
104
105 if (animated) {
106 /* Remap colours. */
107 const byte m0 = mvX2;
108 if (m0 >= PALETTE_ANIM_START) {
109 const Colour c0 = (this->LookupColourInPalette(m0).data & 0x00FFFFFF) | (src[0].data & 0xFF000000);
110 InsertFirstUint32(AdjustBrightneSSE(c0, (byte)(mvX2 >> 8)).data, srcABCD);
111 }
112 const byte m1 = mvX2 >> 16;
113 if (m1 >= PALETTE_ANIM_START) {
114 const Colour c1 = (this->LookupColourInPalette(m1).data & 0x00FFFFFF) | (src[1].data & 0xFF000000);
115 InsertSecondUint32(AdjustBrightneSSE(c1, (byte)(mvX2 >> 24)).data, srcABCD);
116 }
117
118 /* Update anim buffer. */
119 const byte a0 = src[0].a;
120 const byte a1 = src[1].a;
121 uint32 anim01 = 0;
122 if (a0 == 255) {
123 if (a1 == 255) {
124 *(uint32*)anim = mvX2;
125 goto bmno_full_opacity;
126 }
127 anim01 = (uint16)mvX2;
128 }
129 else if (a0 == 0) {
130 if (a1 == 0) {
131 goto bmno_full_transparency;
132 }
133 else {
134 if (a1 == 255) anim[1] = (uint16)(mvX2 >> 16);
135 goto bmno_alpha_blend;
136 }
137 }
138 if (a1 > 0) {
139 if (a1 == 255) anim01 |= mvX2 & 0xFFFF0000;
140 *(uint32*)anim = anim01;
141 }
142 else {
143 anim[0] = (uint16)anim01;
144 }
145 }
146 else {
147 if (src[0].a) anim[0] = 0;
148 if (src[1].a) anim[1] = 0;
149 }
150
151 /* Blend colours. */
152 bmno_alpha_blend:
153 srcABCD = AlphaBlendTwoPixels(srcABCD, dstABCD, a_cm, pack_low_cm);
154 bmno_full_opacity:
155 _mm_storel_epi64((__m128i *) dst, srcABCD);
156 bmno_full_transparency:
157 src_mv += 2;
158 src += 2;
159 anim += 2;
160 dst += 2;
161 }
162
163 if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) {
164 if (src->a == 0) {
165 }
166 else if (src->a == 255) {
167 *anim = *(const uint16*)src_mv;
168 *dst = (src_mv->m >= PALETTE_ANIM_START) ? AdjustBrightneSSE(LookupColourInPalette(src_mv->m), src_mv->v) : *src;
169 }
170 else {
171 *anim = 0;
172 __m128i srcABCD;
173 __m128i dstABCD = _mm_cvtsi32_si128(dst->data);
174 if (src_mv->m >= PALETTE_ANIM_START) {
175 Colour colour = AdjustBrightneSSE(LookupColourInPalette(src_mv->m), src_mv->v);
176 colour.a = src->a;
177 srcABCD = _mm_cvtsi32_si128(colour.data);
178 }
179 else {
180 srcABCD = _mm_cvtsi32_si128(src->data);
181 }
182 dst->data = _mm_cvtsi128_si32(AlphaBlendTwoPixels(srcABCD, dstABCD, a_cm, pack_low_cm));
183 }
184 }
185 break;
186
187 case BM_COLOUR_REMAP:
188 for (uint x = (uint)effective_width / 2; x != 0; x--) {
189 uint32 mvX2 = *((uint32 *) const_cast<MapValue *>(src_mv));
190 __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src);
191 __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst);
192
193 /* Remap colours. */
194 const uint m0 = (byte)mvX2;
195 const uint r0 = remap[m0];
196 const uint m1 = (byte)(mvX2 >> 16);
197 const uint r1 = remap[m1];
198 if (mvX2 & 0x00FF00FF) {
199 #define CMOV_REMAP(m_colour, m_colour_init, m_src, m_m) \
200 /* Written so the compiler uses CMOV. */ \
201 Colour m_colour = m_colour_init; \
202 { \
203 const Colour srcm = (Colour) (m_src); \
204 const uint m = (byte) (m_m); \
205 const uint r = remap[m]; \
206 const Colour cmap = (this->LookupColourInPalette(r).data & 0x00FFFFFF) | (srcm.data & 0xFF000000); \
207 m_colour = r == 0 ? m_colour : cmap; \
208 m_colour = m != 0 ? m_colour : srcm; \
209 }
210 #ifdef _SQ64
211 uint64 srcs = _mm_cvtsi128_si64(srcABCD);
212 uint64 dsts;
213 if (animated) dsts = _mm_cvtsi128_si64(dstABCD);
214 uint64 remapped_src = 0;
215 CMOV_REMAP(c0, animated ? dsts : 0, srcs, mvX2);
216 remapped_src = c0.data;
217 CMOV_REMAP(c1, animated ? dsts >> 32 : 0, srcs >> 32, mvX2 >> 16);
218 remapped_src |= (uint64)c1.data << 32;
219 srcABCD = _mm_cvtsi64_si128(remapped_src);
220 #else
221 Colour remapped_src[2];
222 CMOV_REMAP(c0, animated ? _mm_cvtsi128_si32(dstABCD) : 0, _mm_cvtsi128_si32(srcABCD), mvX2);
223 remapped_src[0] = c0.data;
224 CMOV_REMAP(c1, animated ? dst[1] : 0, src[1], mvX2 >> 16);
225 remapped_src[1] = c1.data;
226 srcABCD = _mm_loadl_epi64((__m128i*) &remapped_src);
227 #endif
228
229 if ((mvX2 & 0xFF00FF00) != 0x80008000) srcABCD = AdjustBrightnessOfTwoPixels(srcABCD, mvX2);
230 }
231
232 /* Update anim buffer. */
233 if (animated) {
234 const byte a0 = src[0].a;
235 const byte a1 = src[1].a;
236 uint32 anim01 = mvX2 & 0xFF00FF00;
237 if (a0 == 255) {
238 anim01 |= r0;
239 if (a1 == 255) {
240 *(uint32*)anim = anim01 | (r1 << 16);
241 goto bmcr_full_opacity;
242 }
243 }
244 else if (a0 == 0) {
245 if (a1 == 0) {
246 goto bmcr_full_transparency;
247 }
248 else {
249 if (a1 == 255) {
250 anim[1] = r1 | (anim01 >> 16);
251 }
252 goto bmcr_alpha_blend;
253 }
254 }
255 if (a1 > 0) {
256 if (a1 == 255) anim01 |= r1 << 16;
257 *(uint32*)anim = anim01;
258 }
259 else {
260 anim[0] = (uint16)anim01;
261 }
262 }
263 else {
264 if (src[0].a) anim[0] = 0;
265 if (src[1].a) anim[1] = 0;
266 }
267
268 /* Blend colours. */
269 bmcr_alpha_blend:
270 srcABCD = AlphaBlendTwoPixels(srcABCD, dstABCD, a_cm, pack_low_cm);
271 bmcr_full_opacity:
272 _mm_storel_epi64((__m128i *) dst, srcABCD);
273 bmcr_full_transparency:
274 src_mv += 2;
275 dst += 2;
276 src += 2;
277 anim += 2;
278 }
279
280 if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) {
281 /* In case the m-channel is zero, do not remap this pixel in any way. */
282 __m128i srcABCD;
283 if (src->a == 0) break;
284 if (src_mv->m) {
285 const uint r = remap[src_mv->m];
286 *anim = (animated && src->a == 255) ? r | ((uint16)src_mv->v << 8) : 0;
287 if (r != 0) {
288 Colour remapped_colour = AdjustBrightneSSE(this->LookupColourInPalette(r), src_mv->v);
289 if (src->a == 255) {
290 *dst = remapped_colour;
291 }
292 else {
293 remapped_colour.a = src->a;
294 srcABCD = _mm_cvtsi32_si128(remapped_colour.data);
295 goto bmcr_alpha_blend_single;
296 }
297 }
298 }
299 else {
300 *anim = 0;
301 srcABCD = _mm_cvtsi32_si128(src->data);
302 if (src->a < 255) {
303 bmcr_alpha_blend_single:
304 __m128i dstABCD = _mm_cvtsi32_si128(dst->data);
305 srcABCD = AlphaBlendTwoPixels(srcABCD, dstABCD, a_cm, pack_low_cm);
306 }
307 dst->data = _mm_cvtsi128_si32(srcABCD);
308 }
309 }
310 break;
311
312 case BM_TRANSPARENT:
313 /* Make the current colour a bit more black, so it looks like this image is transparent. */
314 for (uint x = (uint)bp->width / 2; x > 0; x--) {
315 __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src);
316 __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst);
317 _mm_storel_epi64((__m128i *) dst, DarkenTwoPixels(srcABCD, dstABCD, a_cm, tr_nom_base));
318 src += 2;
319 dst += 2;
320 anim += 2;
321 if (src[-2].a) anim[-2] = 0;
322 if (src[-1].a) anim[-1] = 0;
323 }
324
325 if ((bt_last == BT_NONE && bp->width & 1) || bt_last == BT_ODD) {
326 __m128i srcABCD = _mm_cvtsi32_si128(src->data);
327 __m128i dstABCD = _mm_cvtsi32_si128(dst->data);
328 dst->data = _mm_cvtsi128_si32(DarkenTwoPixels(srcABCD, dstABCD, a_cm, tr_nom_base));
329 if (src[0].a) anim[0] = 0;
330 }
331 break;
332
333 case BM_CRASH_REMAP:
334 for (uint x = (uint)bp->width; x > 0; x--) {
335 if (src_mv->m == 0) {
336 if (src->a != 0) {
337 uint8 g = MakeDark(src->r, src->g, src->b);
338 *dst = ComposeColourRGBA(g, g, g, src->a, *dst);
339 *anim = 0;
340 }
341 }
342 else {
343 uint r = remap[src_mv->m];
344 if (r != 0) *dst = ComposeColourPANoCheck(this->AdjustBrightness(this->LookupColourInPalette(r), src_mv->v), src->a, *dst);
345 }
346 src_mv++;
347 dst++;
348 src++;
349 anim++;
350 }
351 break;
352
353 case BM_BLACK_REMAP:
354 for (uint x = (uint)bp->width; x > 0; x--) {
355 if (src->a != 0) {
356 *dst = Colour(0, 0, 0);
357 *anim = 0;
358 }
359 src_mv++;
360 dst++;
361 src++;
362 anim++;
363 }
364 break;
365 }
366
367 next_line:
368 if (mode != BM_TRANSPARENT) src_mv_line += si->sprite_width;
369 src_rgba_line = (const Colour*)((const byte*)src_rgba_line + si->sprite_line_size);
370 dst_line += bp->pitch;
371 anim_line += this->anim_buf_pitch;
372 }
373 }
374 IGNORE_UNINITIALIZED_WARNING_STOP
375
376 /**
377 * Draws a sprite to a (screen) buffer. Calls adequate templated function.
378 *
379 * @param bp further blitting parameters
380 * @param mode blitter mode
381 * @param zoom zoom level at which we are drawing
382 */
383 void Blitter_32bppSSE4_Anim::Draw(Blitter::BlitterParams *bp, BlitterMode mode, ZoomLevel zoom)
384 {
385 const BlitterSpriteFlags sprite_flags = ((const Blitter_32bppSSE_Base::SpriteData *) bp->sprite)->flags;
386 switch (mode) {
387 default: {
388 bm_normal:
389 if (bp->skip_left != 0 || bp->width <= MARGIN_NORMAL_THRESHOLD) {
390 const BlockType bt_last = (BlockType)(bp->width & 1);
391 if (bt_last == BT_EVEN) {
392 if (sprite_flags & SF_NO_ANIM) Draw<BM_NORMAL, RM_WITH_SKIP, BT_EVEN, true, false>(bp, zoom);
393 else Draw<BM_NORMAL, RM_WITH_SKIP, BT_EVEN, true, true>(bp, zoom);
394 }
395 else {
396 if (sprite_flags & SF_NO_ANIM) Draw<BM_NORMAL, RM_WITH_SKIP, BT_ODD, true, false>(bp, zoom);
397 else Draw<BM_NORMAL, RM_WITH_SKIP, BT_ODD, true, true>(bp, zoom);
398 }
399 }
400 else {
401 #ifdef _SQ64
402 if (sprite_flags & SF_TRANSLUCENT) {
403 if (sprite_flags & SF_NO_ANIM) Draw<BM_NORMAL, RM_WITH_MARGIN, BT_NONE, true, false>(bp, zoom);
404 else Draw<BM_NORMAL, RM_WITH_MARGIN, BT_NONE, true, true>(bp, zoom);
405 }
406 else {
407 if (sprite_flags & SF_NO_ANIM) Draw<BM_NORMAL, RM_WITH_MARGIN, BT_NONE, false, false>(bp, zoom);
408 else Draw<BM_NORMAL, RM_WITH_MARGIN, BT_NONE, false, true>(bp, zoom);
409 }
410 #else
411 if (sprite_flags & SF_NO_ANIM) Draw<BM_NORMAL, RM_WITH_MARGIN, BT_NONE, true, false>(bp, zoom);
412 else Draw<BM_NORMAL, RM_WITH_MARGIN, BT_NONE, true, true>(bp, zoom);
413 #endif
414 }
415 break;
416 }
417 case BM_COLOUR_REMAP:
418 if (sprite_flags & SF_NO_REMAP) goto bm_normal;
419 if (bp->skip_left != 0 || bp->width <= MARGIN_REMAP_THRESHOLD) {
420 if (sprite_flags & SF_NO_ANIM) Draw<BM_COLOUR_REMAP, RM_WITH_SKIP, BT_NONE, true, false>(bp, zoom);
421 else Draw<BM_COLOUR_REMAP, RM_WITH_SKIP, BT_NONE, true, true>(bp, zoom);
422 }
423 else {
424 if (sprite_flags & SF_NO_ANIM) Draw<BM_COLOUR_REMAP, RM_WITH_MARGIN, BT_NONE, true, false>(bp, zoom);
425 else Draw<BM_COLOUR_REMAP, RM_WITH_MARGIN, BT_NONE, true, true>(bp, zoom);
426 }
427 break;
428 case BM_TRANSPARENT: Draw<BM_TRANSPARENT, RM_NONE, BT_NONE, true, true>(bp, zoom); return;
429 case BM_CRASH_REMAP: Draw<BM_CRASH_REMAP, RM_NONE, BT_NONE, true, true>(bp, zoom); return;
430 case BM_BLACK_REMAP: Draw<BM_BLACK_REMAP, RM_NONE, BT_NONE, true, true>(bp, zoom); return;
431 }
432 }
433
434 #endif /* WITH_SSE */
OpenTTD - Joker's Patch Pack