3 * rgb2rgb.c, Software RGB to RGB convertor
4 * pluralize by Software PAL8 to RGB convertor
5 * Software YUV to YUV convertor
6 * Software YUV to RGB convertor
7 * Written by Nick Kurshev.
8 * palette & yuv & runtime cpu stuff by Michael (michaelni@gmx.at) (under GPL)
11 #include "../config.h"
13 #include "../cpudetect.h"
14 #include "../mangle.h"
17 #define CAN_COMPILE_X86_ASM
20 #define FAST_BGR2YV12 // use 7 bit coeffs instead of 15bit
22 #ifdef CAN_COMPILE_X86_ASM
23 static const uint64_t mask32b __attribute__((aligned(8))) = 0x000000FF000000FFULL;
24 static const uint64_t mask32g __attribute__((aligned(8))) = 0x0000FF000000FF00ULL;
25 static const uint64_t mask32r __attribute__((aligned(8))) = 0x00FF000000FF0000ULL;
26 static const uint64_t mask32 __attribute__((aligned(8))) = 0x00FFFFFF00FFFFFFULL;
27 static const uint64_t mask24b __attribute__((aligned(8))) = 0x00FF0000FF0000FFULL;
28 static const uint64_t mask24g __attribute__((aligned(8))) = 0xFF0000FF0000FF00ULL;
29 static const uint64_t mask24r __attribute__((aligned(8))) = 0x0000FF0000FF0000ULL;
30 static const uint64_t mask24l __attribute__((aligned(8))) = 0x0000000000FFFFFFULL;
31 static const uint64_t mask24h __attribute__((aligned(8))) = 0x0000FFFFFF000000ULL;
32 static const uint64_t mask24hh __attribute__((aligned(8))) = 0xffff000000000000ULL;
33 static const uint64_t mask24hhh __attribute__((aligned(8))) = 0xffffffff00000000ULL;
34 static const uint64_t mask24hhhh __attribute__((aligned(8))) = 0xffffffffffff0000ULL;
35 static const uint64_t mask15b __attribute__((aligned(8))) = 0x001F001F001F001FULL; /* 00000000 00011111 xxB */
36 static const uint64_t mask15rg __attribute__((aligned(8))) = 0x7FE07FE07FE07FE0ULL; /* 01111111 11100000 RGx */
37 static const uint64_t mask15s __attribute__((aligned(8))) = 0xFFE0FFE0FFE0FFE0ULL;
38 static const uint64_t red_16mask __attribute__((aligned(8))) = 0x0000f8000000f800ULL;
39 static const uint64_t green_16mask __attribute__((aligned(8)))= 0x000007e0000007e0ULL;
40 static const uint64_t blue_16mask __attribute__((aligned(8))) = 0x0000001f0000001fULL;
41 static const uint64_t red_15mask __attribute__((aligned(8))) = 0x00007c000000f800ULL;
42 static const uint64_t green_15mask __attribute__((aligned(8)))= 0x000003e0000007e0ULL;
43 static const uint64_t blue_15mask __attribute__((aligned(8))) = 0x0000001f0000001fULL;
46 static const uint64_t bgr2YCoeff __attribute__((aligned(8))) = 0x000000210041000DULL;
47 static const uint64_t bgr2UCoeff __attribute__((aligned(8))) = 0x0000FFEEFFDC0038ULL;
48 static const uint64_t bgr2VCoeff __attribute__((aligned(8))) = 0x00000038FFD2FFF8ULL;
50 static const uint64_t bgr2YCoeff __attribute__((aligned(8))) = 0x000020E540830C8BULL;
51 static const uint64_t bgr2UCoeff __attribute__((aligned(8))) = 0x0000ED0FDAC23831ULL;
52 static const uint64_t bgr2VCoeff __attribute__((aligned(8))) = 0x00003831D0E6F6EAULL;
54 static const uint64_t bgr2YOffset __attribute__((aligned(8))) = 0x1010101010101010ULL;
55 static const uint64_t bgr2UVOffset __attribute__((aligned(8)))= 0x8080808080808080ULL;
56 static const uint64_t w1111 __attribute__((aligned(8))) = 0x0001000100010001ULL;
59 static volatile uint64_t __attribute__((aligned(8))) b5Dither;
60 static volatile uint64_t __attribute__((aligned(8))) g5Dither;
61 static volatile uint64_t __attribute__((aligned(8))) g6Dither;
62 static volatile uint64_t __attribute__((aligned(8))) r5Dither;
64 static uint64_t __attribute__((aligned(8))) dither4[2]={
66 0x0200020002000200LL,};
68 static uint64_t __attribute__((aligned(8))) dither8[2]={
70 0x0004000400040004LL,};
74 #define RGB2YUV_SHIFT 8
75 #define BY ((int)( 0.098*(1<<RGB2YUV_SHIFT)+0.5))
76 #define BV ((int)(-0.071*(1<<RGB2YUV_SHIFT)+0.5))
77 #define BU ((int)( 0.439*(1<<RGB2YUV_SHIFT)+0.5))
78 #define GY ((int)( 0.504*(1<<RGB2YUV_SHIFT)+0.5))
79 #define GV ((int)(-0.368*(1<<RGB2YUV_SHIFT)+0.5))
80 #define GU ((int)(-0.291*(1<<RGB2YUV_SHIFT)+0.5))
81 #define RY ((int)( 0.257*(1<<RGB2YUV_SHIFT)+0.5))
82 #define RV ((int)( 0.439*(1<<RGB2YUV_SHIFT)+0.5))
83 #define RU ((int)(-0.148*(1<<RGB2YUV_SHIFT)+0.5))
85 //Note: we have C, MMX, MMX2, 3DNOW version therse no 3DNOW+MMX2 one
92 #define RENAME(a) a ## _C
93 #include "rgb2rgb_template.c"
95 #ifdef CAN_COMPILE_X86_ASM
104 #define RENAME(a) a ## _MMX
105 #include "rgb2rgb_template.c"
114 #define RENAME(a) a ## _MMX2
115 #include "rgb2rgb_template.c"
124 #define RENAME(a) a ## _3DNow
125 #include "rgb2rgb_template.c"
127 #endif //CAN_COMPILE_X86_ASM
129 void rgb24to32(const uint8_t *src,uint8_t *dst,unsigned src_size)
131 #ifdef CAN_COMPILE_X86_ASM
132 // ordered per speed fasterst first
134 rgb24to32_MMX2(src, dst, src_size);
135 else if(gCpuCaps.has3DNow)
136 rgb24to32_3DNow(src, dst, src_size);
137 else if(gCpuCaps.hasMMX)
138 rgb24to32_MMX(src, dst, src_size);
140 rgb24to32_C(src, dst, src_size);
142 rgb24to32_C(src, dst, src_size);
146 void rgb32to24(const uint8_t *src,uint8_t *dst,unsigned src_size)
148 #ifdef CAN_COMPILE_X86_ASM
149 // ordered per speed fasterst first
151 rgb32to24_MMX2(src, dst, src_size);
152 else if(gCpuCaps.has3DNow)
153 rgb32to24_3DNow(src, dst, src_size);
154 else if(gCpuCaps.hasMMX)
155 rgb32to24_MMX(src, dst, src_size);
157 rgb32to24_C(src, dst, src_size);
159 rgb32to24_C(src, dst, src_size);
164 Original by Strepto/Astral
165 ported to gcc & bugfixed : A'rpi
166 MMX2, 3DNOW optimization by Nick Kurshev
167 32bit c version, and and&add trick by Michael Niedermayer
169 void rgb15to16(const uint8_t *src,uint8_t *dst,unsigned src_size)
171 #ifdef CAN_COMPILE_X86_ASM
172 // ordered per speed fasterst first
174 rgb15to16_MMX2(src, dst, src_size);
175 else if(gCpuCaps.has3DNow)
176 rgb15to16_3DNow(src, dst, src_size);
177 else if(gCpuCaps.hasMMX)
178 rgb15to16_MMX(src, dst, src_size);
180 rgb15to16_C(src, dst, src_size);
182 rgb15to16_C(src, dst, src_size);
187 * Pallete is assumed to contain bgr32
189 void palette8torgb32(const uint8_t *src, uint8_t *dst, unsigned num_pixels, const uint8_t *palette)
192 for(i=0; i<num_pixels; i++)
193 ((unsigned *)dst)[i] = ((unsigned *)palette)[ src[i] ];
197 * Pallete is assumed to contain bgr32
199 void palette8torgb24(const uint8_t *src, uint8_t *dst, unsigned num_pixels, const uint8_t *palette)
203 writes 1 byte o much and might cause alignment issues on some architectures?
204 for(i=0; i<num_pixels; i++)
205 ((unsigned *)(&dst[i*3])) = ((unsigned *)palette)[ src[i] ];
207 for(i=0; i<num_pixels; i++)
210 dst[0]= palette[ src[i]*4+0 ];
211 dst[1]= palette[ src[i]*4+1 ];
212 dst[2]= palette[ src[i]*4+2 ];
217 void rgb32to16(const uint8_t *src, uint8_t *dst, unsigned src_size)
219 #ifdef CAN_COMPILE_X86_ASM
220 // ordered per speed fasterst first
222 rgb32to16_MMX2(src, dst, src_size);
223 else if(gCpuCaps.has3DNow)
224 rgb32to16_3DNow(src, dst, src_size);
225 else if(gCpuCaps.hasMMX)
226 rgb32to16_MMX(src, dst, src_size);
228 rgb32to16_C(src, dst, src_size);
230 rgb32to16_C(src, dst, src_size);
234 void rgb32to15(const uint8_t *src, uint8_t *dst, unsigned src_size)
236 #ifdef CAN_COMPILE_X86_ASM
237 // ordered per speed fasterst first
239 rgb32to15_MMX2(src, dst, src_size);
240 else if(gCpuCaps.has3DNow)
241 rgb32to15_3DNow(src, dst, src_size);
242 else if(gCpuCaps.hasMMX)
243 rgb32to15_MMX(src, dst, src_size);
245 rgb32to15_C(src, dst, src_size);
247 rgb32to15_C(src, dst, src_size);
251 void rgb24to16(const uint8_t *src, uint8_t *dst, unsigned src_size)
253 #ifdef CAN_COMPILE_X86_ASM
254 // ordered per speed fasterst first
256 rgb24to16_MMX2(src, dst, src_size);
257 else if(gCpuCaps.has3DNow)
258 rgb24to16_3DNow(src, dst, src_size);
259 else if(gCpuCaps.hasMMX)
260 rgb24to16_MMX(src, dst, src_size);
262 rgb24to16_C(src, dst, src_size);
264 rgb24to16_C(src, dst, src_size);
268 void rgb24to15(const uint8_t *src, uint8_t *dst, unsigned src_size)
270 #ifdef CAN_COMPILE_X86_ASM
271 // ordered per speed fasterst first
273 rgb24to15_MMX2(src, dst, src_size);
274 else if(gCpuCaps.has3DNow)
275 rgb24to15_3DNow(src, dst, src_size);
276 else if(gCpuCaps.hasMMX)
277 rgb24to15_MMX(src, dst, src_size);
279 rgb24to15_C(src, dst, src_size);
281 rgb24to15_C(src, dst, src_size);
286 * Palette is assumed to contain bgr16, see rgb32to16 to convert the palette
288 void palette8torgb16(const uint8_t *src, uint8_t *dst, unsigned num_pixels, const uint8_t *palette)
291 for(i=0; i<num_pixels; i++)
292 ((uint16_t *)dst)[i] = ((uint16_t *)palette)[ src[i] ];
296 * Pallete is assumed to contain bgr15, see rgb32to15 to convert the palette
298 void palette8torgb15(const uint8_t *src, uint8_t *dst, unsigned num_pixels, const uint8_t *palette)
301 for(i=0; i<num_pixels; i++)
302 ((uint16_t *)dst)[i] = ((uint16_t *)palette)[ src[i] ];
305 void rgb32tobgr32(const uint8_t *src, uint8_t *dst, unsigned int src_size)
307 #ifdef CAN_COMPILE_X86_ASM
308 // ordered per speed fasterst first
310 rgb32tobgr32_MMX2(src, dst, src_size);
311 else if(gCpuCaps.has3DNow)
312 rgb32tobgr32_3DNow(src, dst, src_size);
313 else if(gCpuCaps.hasMMX)
314 rgb32tobgr32_MMX(src, dst, src_size);
316 rgb32tobgr32_C(src, dst, src_size);
318 rgb32tobgr32_C(src, dst, src_size);
322 void rgb24tobgr24(const uint8_t *src, uint8_t *dst, unsigned int src_size)
324 #ifdef CAN_COMPILE_X86_ASM
325 // ordered per speed fasterst first
327 rgb24tobgr24_MMX2(src, dst, src_size);
328 else if(gCpuCaps.has3DNow)
329 rgb24tobgr24_3DNow(src, dst, src_size);
330 else if(gCpuCaps.hasMMX)
331 rgb24tobgr24_MMX(src, dst, src_size);
333 rgb24tobgr24_C(src, dst, src_size);
335 rgb24tobgr24_C(src, dst, src_size);
341 * height should be a multiple of 2 and width should be a multiple of 16 (if this is a
342 * problem for anyone then tell me, and ill fix it)
344 void yv12toyuy2(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
345 unsigned int width, unsigned int height,
346 unsigned int lumStride, unsigned int chromStride, unsigned int dstStride)
348 #ifdef CAN_COMPILE_X86_ASM
349 // ordered per speed fasterst first
351 yv12toyuy2_MMX2(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride);
352 else if(gCpuCaps.has3DNow)
353 yv12toyuy2_3DNow(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride);
354 else if(gCpuCaps.hasMMX)
355 yv12toyuy2_MMX(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride);
357 yv12toyuy2_C(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride);
359 yv12toyuy2_C(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride);
365 * width should be a multiple of 16
367 void yuv422ptoyuy2(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
368 unsigned int width, unsigned int height,
369 unsigned int lumStride, unsigned int chromStride, unsigned int dstStride)
371 #ifdef CAN_COMPILE_X86_ASM
372 // ordered per speed fasterst first
374 yuv422ptoyuy2_MMX2(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride);
375 else if(gCpuCaps.has3DNow)
376 yuv422ptoyuy2_3DNow(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride);
377 else if(gCpuCaps.hasMMX)
378 yuv422ptoyuy2_MMX(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride);
380 yuv422ptoyuy2_C(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride);
382 yuv422ptoyuy2_C(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride);
388 * height should be a multiple of 2 and width should be a multiple of 16 (if this is a
389 * problem for anyone then tell me, and ill fix it)
391 void yuy2toyv12(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
392 unsigned int width, unsigned int height,
393 unsigned int lumStride, unsigned int chromStride, unsigned int srcStride)
395 #ifdef CAN_COMPILE_X86_ASM
396 // ordered per speed fasterst first
398 yuy2toyv12_MMX2(src, ydst, udst, vdst, width, height, lumStride, chromStride, srcStride);
399 else if(gCpuCaps.has3DNow)
400 yuy2toyv12_3DNow(src, ydst, udst, vdst, width, height, lumStride, chromStride, srcStride);
401 else if(gCpuCaps.hasMMX)
402 yuy2toyv12_MMX(src, ydst, udst, vdst, width, height, lumStride, chromStride, srcStride);
404 yuy2toyv12_C(src, ydst, udst, vdst, width, height, lumStride, chromStride, srcStride);
406 yuy2toyv12_C(src, ydst, udst, vdst, width, height, lumStride, chromStride, srcStride);
412 * height should be a multiple of 2 and width should be a multiple of 16 (if this is a
413 * problem for anyone then tell me, and ill fix it)
414 * chrominance data is only taken from every secound line others are ignored FIXME write HQ version
416 void uyvytoyv12(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
417 unsigned int width, unsigned int height,
418 unsigned int lumStride, unsigned int chromStride, unsigned int srcStride)
420 #ifdef CAN_COMPILE_X86_ASM
421 // ordered per speed fasterst first
423 uyvytoyv12_MMX2(src, ydst, udst, vdst, width, height, lumStride, chromStride, srcStride);
424 else if(gCpuCaps.has3DNow)
425 uyvytoyv12_3DNow(src, ydst, udst, vdst, width, height, lumStride, chromStride, srcStride);
426 else if(gCpuCaps.hasMMX)
427 uyvytoyv12_MMX(src, ydst, udst, vdst, width, height, lumStride, chromStride, srcStride);
429 uyvytoyv12_C(src, ydst, udst, vdst, width, height, lumStride, chromStride, srcStride);
431 uyvytoyv12_C(src, ydst, udst, vdst, width, height, lumStride, chromStride, srcStride);
437 * height should be a multiple of 2 and width should be a multiple of 2 (if this is a
438 * problem for anyone then tell me, and ill fix it)
439 * chrominance data is only taken from every secound line others are ignored FIXME write HQ version
441 void rgb24toyv12(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
442 unsigned int width, unsigned int height,
443 unsigned int lumStride, unsigned int chromStride, unsigned int srcStride)
445 #ifdef CAN_COMPILE_X86_ASM
446 // ordered per speed fasterst first
448 rgb24toyv12_MMX2(src, ydst, udst, vdst, width, height, lumStride, chromStride, srcStride);
449 else if(gCpuCaps.has3DNow)
450 rgb24toyv12_3DNow(src, ydst, udst, vdst, width, height, lumStride, chromStride, srcStride);
451 else if(gCpuCaps.hasMMX)
452 rgb24toyv12_MMX(src, ydst, udst, vdst, width, height, lumStride, chromStride, srcStride);
454 rgb24toyv12_C(src, ydst, udst, vdst, width, height, lumStride, chromStride, srcStride);
456 rgb24toyv12_C(src, ydst, udst, vdst, width, height, lumStride, chromStride, srcStride);
460 void interleaveBytes(uint8_t *src1, uint8_t *src2, uint8_t *dst,
461 int width, int height, int src1Stride, int src2Stride, int dstStride)
463 #ifdef CAN_COMPILE_X86_ASM
464 // ordered per speed fasterst first
466 interleaveBytes_MMX2(src1, src2, dst, width, height, src1Stride, src2Stride, dstStride);
467 else if(gCpuCaps.has3DNow)
468 interleaveBytes_3DNow(src1, src2, dst, width, height, src1Stride, src2Stride, dstStride);
469 else if(gCpuCaps.hasMMX)
470 interleaveBytes_MMX(src1, src2, dst, width, height, src1Stride, src2Stride, dstStride);
472 interleaveBytes_C(src1, src2, dst, width, height, src1Stride, src2Stride, dstStride);
474 interleaveBytes_C(src1, src2, dst, width, height, src1Stride, src2Stride, dstStride);