Change semantic of CONFIG_*, HAVE_* and ARCH_*.
[libav.git] / libswscale / rgb2rgb.c
CommitLineData
fcfbc150 1/*
8a322796
DB
2 * software RGB to RGB converter
3 * pluralize by software PAL8 to RGB converter
4 * software YUV to YUV converter
5 * software YUV to RGB converter
6 * Written by Nick Kurshev.
7 * palette & YUV & runtime CPU stuff by Michael (michaelni@gmx.at)
4fadc2b4 8 *
d026b45e
DB
9 * This file is part of FFmpeg.
10 *
11 * FFmpeg is free software; you can redistribute it and/or modify
4fadc2b4
DB
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2 of the License, or
14 * (at your option) any later version.
15 *
d026b45e 16 * FFmpeg is distributed in the hope that it will be useful,
4fadc2b4
DB
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
20 *
21 * You should have received a copy of the GNU General Public License
d026b45e 22 * along with FFmpeg; if not, write to the Free Software
b19bcbaa 23 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
6a4970ab 24 *
8a322796
DB
25 * The C code (not assembly, MMX, ...) of this file can be used
26 * under the LGPL license.
a3aece93 27 */
b234ae81 28#include <inttypes.h>
b2d374c9 29#include "config.h"
83da2c6f
DB
30#include "libavutil/x86_cpu.h"
31#include "libavutil/bswap.h"
b234ae81 32#include "rgb2rgb.h"
700490a4 33#include "swscale.h"
94c4def2 34#include "swscale_internal.h"
a3aece93 35
8a322796 36#define FAST_BGR2YV12 // use 7-bit instead of 15-bit coefficients
21316f3c 37
6107059c
MN
38void (*rgb24tobgr32)(const uint8_t *src, uint8_t *dst, long src_size);
39void (*rgb24tobgr16)(const uint8_t *src, uint8_t *dst, long src_size);
40void (*rgb24tobgr15)(const uint8_t *src, uint8_t *dst, long src_size);
41void (*rgb32tobgr24)(const uint8_t *src, uint8_t *dst, long src_size);
30c48a0a
BF
42void (*rgb32to16)(const uint8_t *src, uint8_t *dst, long src_size);
43void (*rgb32to15)(const uint8_t *src, uint8_t *dst, long src_size);
44void (*rgb15to16)(const uint8_t *src, uint8_t *dst, long src_size);
6107059c 45void (*rgb15tobgr24)(const uint8_t *src, uint8_t *dst, long src_size);
30c48a0a
BF
46void (*rgb15to32)(const uint8_t *src, uint8_t *dst, long src_size);
47void (*rgb16to15)(const uint8_t *src, uint8_t *dst, long src_size);
6107059c 48void (*rgb16tobgr24)(const uint8_t *src, uint8_t *dst, long src_size);
30c48a0a 49void (*rgb16to32)(const uint8_t *src, uint8_t *dst, long src_size);
7f526efd 50void (*rgb24tobgr24)(const uint8_t *src, uint8_t *dst, long src_size);
6107059c
MN
51void (*rgb24to16)(const uint8_t *src, uint8_t *dst, long src_size);
52void (*rgb24to15)(const uint8_t *src, uint8_t *dst, long src_size);
7f526efd 53void (*rgb32tobgr32)(const uint8_t *src, uint8_t *dst, long src_size);
7f526efd
RD
54void (*rgb32tobgr16)(const uint8_t *src, uint8_t *dst, long src_size);
55void (*rgb32tobgr15)(const uint8_t *src, uint8_t *dst, long src_size);
700490a4
MN
56
57void (*yv12toyuy2)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
6e42e6c4
DB
58 long width, long height,
59 long lumStride, long chromStride, long dstStride);
caeaabe7 60void (*yv12touyvy)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
6e42e6c4
DB
61 long width, long height,
62 long lumStride, long chromStride, long dstStride);
700490a4 63void (*yuv422ptoyuy2)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
6e42e6c4
DB
64 long width, long height,
65 long lumStride, long chromStride, long dstStride);
a6100f39
BC
66void (*yuv422ptouyvy)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
67 long width, long height,
68 long lumStride, long chromStride, long dstStride);
700490a4 69void (*yuy2toyv12)(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
6e42e6c4
DB
70 long width, long height,
71 long lumStride, long chromStride, long srcStride);
700490a4 72void (*rgb24toyv12)(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
6e42e6c4
DB
73 long width, long height,
74 long lumStride, long chromStride, long srcStride);
7f526efd 75void (*planar2x)(const uint8_t *src, uint8_t *dst, long width, long height,
6e42e6c4 76 long srcStride, long dstStride);
700490a4 77void (*interleaveBytes)(uint8_t *src1, uint8_t *src2, uint8_t *dst,
6e42e6c4
DB
78 long width, long height, long src1Stride,
79 long src2Stride, long dstStride);
700490a4 80void (*vu9_to_vu12)(const uint8_t *src1, const uint8_t *src2,
6e42e6c4
DB
81 uint8_t *dst1, uint8_t *dst2,
82 long width, long height,
83 long srcStride1, long srcStride2,
84 long dstStride1, long dstStride2);
700490a4 85void (*yvu9_to_yuy2)(const uint8_t *src1, const uint8_t *src2, const uint8_t *src3,
6e42e6c4
DB
86 uint8_t *dst,
87 long width, long height,
88 long srcStride1, long srcStride2,
89 long srcStride3, long dstStride);
700490a4 90
9bde778e 91#if defined(ARCH_X86) && defined(CONFIG_GPL)
d334c7c2
RD
92DECLARE_ASM_CONST(8, uint64_t, mmx_null) = 0x0000000000000000ULL;
93DECLARE_ASM_CONST(8, uint64_t, mmx_one) = 0xFFFFFFFFFFFFFFFFULL;
94DECLARE_ASM_CONST(8, uint64_t, mask32b) = 0x000000FF000000FFULL;
95DECLARE_ASM_CONST(8, uint64_t, mask32g) = 0x0000FF000000FF00ULL;
96DECLARE_ASM_CONST(8, uint64_t, mask32r) = 0x00FF000000FF0000ULL;
97DECLARE_ASM_CONST(8, uint64_t, mask32) = 0x00FFFFFF00FFFFFFULL;
98DECLARE_ASM_CONST(8, uint64_t, mask3216br) = 0x00F800F800F800F8ULL;
99DECLARE_ASM_CONST(8, uint64_t, mask3216g) = 0x0000FC000000FC00ULL;
100DECLARE_ASM_CONST(8, uint64_t, mask3215g) = 0x0000F8000000F800ULL;
101DECLARE_ASM_CONST(8, uint64_t, mul3216) = 0x2000000420000004ULL;
102DECLARE_ASM_CONST(8, uint64_t, mul3215) = 0x2000000820000008ULL;
103DECLARE_ASM_CONST(8, uint64_t, mask24b) = 0x00FF0000FF0000FFULL;
104DECLARE_ASM_CONST(8, uint64_t, mask24g) = 0xFF0000FF0000FF00ULL;
105DECLARE_ASM_CONST(8, uint64_t, mask24r) = 0x0000FF0000FF0000ULL;
106DECLARE_ASM_CONST(8, uint64_t, mask24l) = 0x0000000000FFFFFFULL;
107DECLARE_ASM_CONST(8, uint64_t, mask24h) = 0x0000FFFFFF000000ULL;
108DECLARE_ASM_CONST(8, uint64_t, mask24hh) = 0xffff000000000000ULL;
109DECLARE_ASM_CONST(8, uint64_t, mask24hhh) = 0xffffffff00000000ULL;
110DECLARE_ASM_CONST(8, uint64_t, mask24hhhh) = 0xffffffffffff0000ULL;
111DECLARE_ASM_CONST(8, uint64_t, mask15b) = 0x001F001F001F001FULL; /* 00000000 00011111 xxB */
112DECLARE_ASM_CONST(8, uint64_t, mask15rg) = 0x7FE07FE07FE07FE0ULL; /* 01111111 11100000 RGx */
113DECLARE_ASM_CONST(8, uint64_t, mask15s) = 0xFFE0FFE0FFE0FFE0ULL;
114DECLARE_ASM_CONST(8, uint64_t, mask15g) = 0x03E003E003E003E0ULL;
115DECLARE_ASM_CONST(8, uint64_t, mask15r) = 0x7C007C007C007C00ULL;
0d9f3d85 116#define mask16b mask15b
d334c7c2
RD
117DECLARE_ASM_CONST(8, uint64_t, mask16g) = 0x07E007E007E007E0ULL;
118DECLARE_ASM_CONST(8, uint64_t, mask16r) = 0xF800F800F800F800ULL;
119DECLARE_ASM_CONST(8, uint64_t, red_16mask) = 0x0000f8000000f800ULL;
120DECLARE_ASM_CONST(8, uint64_t, green_16mask) = 0x000007e0000007e0ULL;
121DECLARE_ASM_CONST(8, uint64_t, blue_16mask) = 0x0000001f0000001fULL;
122DECLARE_ASM_CONST(8, uint64_t, red_15mask) = 0x00007c0000007c00ULL;
123DECLARE_ASM_CONST(8, uint64_t, green_15mask) = 0x000003e0000003e0ULL;
124DECLARE_ASM_CONST(8, uint64_t, blue_15mask) = 0x0000001f0000001fULL;
3d6a30d9 125#endif /* defined(ARCH_X86) */
79811694 126
1de97d84
MN
127#define RGB2YUV_SHIFT 8
128#define BY ((int)( 0.098*(1<<RGB2YUV_SHIFT)+0.5))
129#define BV ((int)(-0.071*(1<<RGB2YUV_SHIFT)+0.5))
130#define BU ((int)( 0.439*(1<<RGB2YUV_SHIFT)+0.5))
131#define GY ((int)( 0.504*(1<<RGB2YUV_SHIFT)+0.5))
132#define GV ((int)(-0.368*(1<<RGB2YUV_SHIFT)+0.5))
133#define GU ((int)(-0.291*(1<<RGB2YUV_SHIFT)+0.5))
134#define RY ((int)( 0.257*(1<<RGB2YUV_SHIFT)+0.5))
135#define RV ((int)( 0.439*(1<<RGB2YUV_SHIFT)+0.5))
136#define RU ((int)(-0.148*(1<<RGB2YUV_SHIFT)+0.5))
137
8a322796
DB
138//Note: We have C, MMX, MMX2, 3DNOW versions, there is no 3DNOW + MMX2 one.
139//plain C versions
1de97d84
MN
140#undef HAVE_MMX
141#undef HAVE_MMX2
142#undef HAVE_3DNOW
52075cde 143#undef HAVE_SSE2
1de97d84
MN
144#define RENAME(a) a ## _C
145#include "rgb2rgb_template.c"
146
9bde778e 147#if defined(ARCH_X86) && defined(CONFIG_GPL)
1de97d84
MN
148
149//MMX versions
150#undef RENAME
151#define HAVE_MMX
152#undef HAVE_MMX2
153#undef HAVE_3DNOW
52075cde 154#undef HAVE_SSE2
1de97d84
MN
155#define RENAME(a) a ## _MMX
156#include "rgb2rgb_template.c"
157
158//MMX2 versions
159#undef RENAME
160#define HAVE_MMX
161#define HAVE_MMX2
162#undef HAVE_3DNOW
52075cde 163#undef HAVE_SSE2
1de97d84
MN
164#define RENAME(a) a ## _MMX2
165#include "rgb2rgb_template.c"
166
167//3DNOW versions
168#undef RENAME
169#define HAVE_MMX
170#undef HAVE_MMX2
171#define HAVE_3DNOW
52075cde 172#undef HAVE_SSE2
abdcd337 173#define RENAME(a) a ## _3DNOW
1de97d84
MN
174#include "rgb2rgb_template.c"
175
6e1c66bc 176#endif //ARCH_X86 || ARCH_X86_64
b238eb2e 177
a3aece93 178/*
8a322796 179 RGB15->RGB16 original by Strepto/Astral
a3aece93 180 ported to gcc & bugfixed : A'rpi
51da31f1 181 MMX2, 3DNOW optimization by Nick Kurshev
8a322796 182 32-bit C version, and and&add trick by Michael Niedermayer
a3aece93 183*/
fcfbc150 184
700490a4 185void sws_rgb2rgb_init(int flags){
9bde778e 186#if (defined(HAVE_MMX2) || defined(HAVE_3DNOW) || defined(HAVE_MMX)) && defined(CONFIG_GPL)
6e42e6c4
DB
187 if (flags & SWS_CPU_CAPS_MMX2)
188 rgb2rgb_init_MMX2();
189 else if (flags & SWS_CPU_CAPS_3DNOW)
190 rgb2rgb_init_3DNOW();
191 else if (flags & SWS_CPU_CAPS_MMX)
192 rgb2rgb_init_MMX();
193 else
dbdae6ec 194#endif /* defined(HAVE_MMX2) || defined(HAVE_3DNOW) || defined(HAVE_MMX) */
6e42e6c4 195 rgb2rgb_init_C();
ac4d0aea 196}
700490a4 197
fcfbc150 198/**
522ce957 199 * Convert the palette to the same packet 32-bit format as the palette
fcfbc150 200 */
522ce957 201void palette8topacked32(const uint8_t *src, uint8_t *dst, long num_pixels, const uint8_t *palette)
fcfbc150 202{
6e42e6c4 203 long i;
9780c7ff 204
6e42e6c4 205 for (i=0; i<num_pixels; i++)
522ce957 206 ((uint32_t *) dst)[i] = ((const uint32_t *) palette)[src[i]];
6c1baeb0 207}
6c1baeb0 208
9ce6584e 209/**
522ce957 210 * Palette format: ABCD -> dst format: ABC
9ce6584e 211 */
522ce957 212void palette8topacked24(const uint8_t *src, uint8_t *dst, long num_pixels, const uint8_t *palette)
9ce6584e 213{
6e42e6c4 214 long i;
9ce6584e 215
6e42e6c4
DB
216 for (i=0; i<num_pixels; i++)
217 {
218 //FIXME slow?
30c48a0a
BF
219 dst[0]= palette[src[i]*4+0];
220 dst[1]= palette[src[i]*4+1];
221 dst[2]= palette[src[i]*4+2];
6e42e6c4
DB
222 dst+= 3;
223 }
6c1baeb0
AB
224}
225
fcfbc150 226/**
8a322796 227 * Palette is assumed to contain BGR16, see rgb32to16 to convert the palette.
fcfbc150 228 */
7f526efd 229void palette8torgb16(const uint8_t *src, uint8_t *dst, long num_pixels, const uint8_t *palette)
fcfbc150 230{
6e42e6c4
DB
231 long i;
232 for (i=0; i<num_pixels; i++)
774c386a 233 ((uint16_t *)dst)[i] = ((const uint16_t *)palette)[src[i]];
fcfbc150 234}
7f526efd 235void palette8tobgr16(const uint8_t *src, uint8_t *dst, long num_pixels, const uint8_t *palette)
6c1baeb0 236{
6e42e6c4
DB
237 long i;
238 for (i=0; i<num_pixels; i++)
774c386a 239 ((uint16_t *)dst)[i] = bswap_16(((const uint16_t *)palette)[src[i]]);
6c1baeb0 240}
fcfbc150
MN
241
242/**
10ef990d 243 * Palette is assumed to contain BGR15, see rgb32to15 to convert the palette.
fcfbc150 244 */
7f526efd 245void palette8torgb15(const uint8_t *src, uint8_t *dst, long num_pixels, const uint8_t *palette)
fcfbc150 246{
6e42e6c4
DB
247 long i;
248 for (i=0; i<num_pixels; i++)
774c386a 249 ((uint16_t *)dst)[i] = ((const uint16_t *)palette)[src[i]];
9ce6584e 250}
7f526efd 251void palette8tobgr15(const uint8_t *src, uint8_t *dst, long num_pixels, const uint8_t *palette)
6c1baeb0 252{
6e42e6c4
DB
253 long i;
254 for (i=0; i<num_pixels; i++)
774c386a 255 ((uint16_t *)dst)[i] = bswap_16(((const uint16_t *)palette)[src[i]]);
6c1baeb0 256}
99969243 257
6107059c 258void rgb32to24(const uint8_t *src, uint8_t *dst, long src_size)
ac4d0aea 259{
6e42e6c4
DB
260 long i;
261 long num_pixels = src_size >> 2;
262 for (i=0; i<num_pixels; i++)
263 {
264 #ifdef WORDS_BIGENDIAN
265 /* RGB32 (= A,B,G,R) -> BGR24 (= B,G,R) */
266 dst[3*i + 0] = src[4*i + 1];
267 dst[3*i + 1] = src[4*i + 2];
268 dst[3*i + 2] = src[4*i + 3];
269 #else
270 dst[3*i + 0] = src[4*i + 2];
271 dst[3*i + 1] = src[4*i + 1];
272 dst[3*i + 2] = src[4*i + 0];
273 #endif
274 }
ac4d0aea
MN
275}
276
6107059c 277void rgb24to32(const uint8_t *src, uint8_t *dst, long src_size)
ac4d0aea 278{
6e42e6c4
DB
279 long i;
280 for (i=0; 3*i<src_size; i++)
281 {
282 #ifdef WORDS_BIGENDIAN
283 /* RGB24 (= R,G,B) -> BGR32 (= A,R,G,B) */
284 dst[4*i + 0] = 0;
285 dst[4*i + 1] = src[3*i + 0];
286 dst[4*i + 2] = src[3*i + 1];
287 dst[4*i + 3] = src[3*i + 2];
288 #else
289 dst[4*i + 0] = src[3*i + 2];
290 dst[4*i + 1] = src[3*i + 1];
291 dst[4*i + 2] = src[3*i + 0];
292 dst[4*i + 3] = 0;
293 #endif
294 }
ac4d0aea
MN
295}
296
7f526efd 297void rgb16tobgr32(const uint8_t *src, uint8_t *dst, long src_size)
ac4d0aea 298{
6e42e6c4 299 const uint16_t *end;
72d6b610 300 uint8_t *d = dst;
774c386a 301 const uint16_t *s = (const uint16_t *)src;
6e42e6c4
DB
302 end = s + src_size/2;
303 while (s < end)
304 {
305 register uint16_t bgr;
306 bgr = *s++;
307 #ifdef WORDS_BIGENDIAN
308 *d++ = 0;
309 *d++ = (bgr&0x1F)<<3;
310 *d++ = (bgr&0x7E0)>>3;
311 *d++ = (bgr&0xF800)>>8;
312 #else
313 *d++ = (bgr&0xF800)>>8;
314 *d++ = (bgr&0x7E0)>>3;
315 *d++ = (bgr&0x1F)<<3;
316 *d++ = 0;
317 #endif
318 }
ac4d0aea
MN
319}
320
6107059c 321void rgb16to24(const uint8_t *src, uint8_t *dst, long src_size)
ac4d0aea 322{
6e42e6c4 323 const uint16_t *end;
72d6b610 324 uint8_t *d = dst;
6e42e6c4
DB
325 const uint16_t *s = (const uint16_t *)src;
326 end = s + src_size/2;
327 while (s < end)
328 {
329 register uint16_t bgr;
330 bgr = *s++;
331 *d++ = (bgr&0xF800)>>8;
332 *d++ = (bgr&0x7E0)>>3;
333 *d++ = (bgr&0x1F)<<3;
334 }
ac4d0aea
MN
335}
336
7f526efd 337void rgb16tobgr16(const uint8_t *src, uint8_t *dst, long src_size)
ac4d0aea 338{
6e42e6c4
DB
339 long i;
340 long num_pixels = src_size >> 1;
341
342 for (i=0; i<num_pixels; i++)
343 {
7372e9bb
MN
344 unsigned rgb = ((const uint16_t*)src)[i];
345 ((uint16_t*)dst)[i] = (rgb>>11) | (rgb&0x7E0) | (rgb<<11);
6e42e6c4 346 }
ac4d0aea
MN
347}
348
7f526efd 349void rgb16tobgr15(const uint8_t *src, uint8_t *dst, long src_size)
ac4d0aea 350{
6e42e6c4
DB
351 long i;
352 long num_pixels = src_size >> 1;
353
354 for (i=0; i<num_pixels; i++)
355 {
7372e9bb
MN
356 unsigned rgb = ((const uint16_t*)src)[i];
357 ((uint16_t*)dst)[i] = (rgb>>11) | ((rgb&0x7C0)>>1) | ((rgb&0x1F)<<10);
6e42e6c4 358 }
ac4d0aea
MN
359}
360
7f526efd 361void rgb15tobgr32(const uint8_t *src, uint8_t *dst, long src_size)
ac4d0aea 362{
6e42e6c4 363 const uint16_t *end;
72d6b610 364 uint8_t *d = dst;
6e42e6c4
DB
365 const uint16_t *s = (const uint16_t *)src;
366 end = s + src_size/2;
367 while (s < end)
368 {
369 register uint16_t bgr;
370 bgr = *s++;
371 #ifdef WORDS_BIGENDIAN
372 *d++ = 0;
373 *d++ = (bgr&0x1F)<<3;
374 *d++ = (bgr&0x3E0)>>2;
375 *d++ = (bgr&0x7C00)>>7;
376 #else
377 *d++ = (bgr&0x7C00)>>7;
378 *d++ = (bgr&0x3E0)>>2;
379 *d++ = (bgr&0x1F)<<3;
380 *d++ = 0;
381 #endif
382 }
ac4d0aea
MN
383}
384
6107059c 385void rgb15to24(const uint8_t *src, uint8_t *dst, long src_size)
ac4d0aea 386{
6e42e6c4 387 const uint16_t *end;
72d6b610 388 uint8_t *d = dst;
774c386a 389 const uint16_t *s = (const uint16_t *)src;
6e42e6c4
DB
390 end = s + src_size/2;
391 while (s < end)
392 {
393 register uint16_t bgr;
394 bgr = *s++;
395 *d++ = (bgr&0x7C00)>>7;
396 *d++ = (bgr&0x3E0)>>2;
397 *d++ = (bgr&0x1F)<<3;
398 }
ac4d0aea
MN
399}
400
7f526efd 401void rgb15tobgr16(const uint8_t *src, uint8_t *dst, long src_size)
ac4d0aea 402{
6e42e6c4
DB
403 long i;
404 long num_pixels = src_size >> 1;
405
406 for (i=0; i<num_pixels; i++)
407 {
7372e9bb
MN
408 unsigned rgb = ((const uint16_t*)src)[i];
409 ((uint16_t*)dst)[i] = ((rgb&0x7C00)>>10) | ((rgb&0x3E0)<<1) | (rgb<<11);
6e42e6c4 410 }
ac4d0aea
MN
411}
412
7f526efd 413void rgb15tobgr15(const uint8_t *src, uint8_t *dst, long src_size)
ac4d0aea 414{
6e42e6c4
DB
415 long i;
416 long num_pixels = src_size >> 1;
417
418 for (i=0; i<num_pixels; i++)
419 {
7372e9bb
MN
420 unsigned br;
421 unsigned rgb = ((const uint16_t*)src)[i];
422 br = rgb&0x7c1F;
423 ((uint16_t*)dst)[i] = (br>>10) | (rgb&0x3E0) | (br<<10);
6e42e6c4 424 }
ac4d0aea
MN
425}
426
6107059c 427void bgr8torgb8(const uint8_t *src, uint8_t *dst, long src_size)
ac4d0aea 428{
6e42e6c4
DB
429 long i;
430 long num_pixels = src_size;
431 for (i=0; i<num_pixels; i++)
432 {
433 unsigned b,g,r;
434 register uint8_t rgb;
435 rgb = src[i];
436 r = (rgb&0x07);
437 g = (rgb&0x38)>>3;
438 b = (rgb&0xC0)>>6;
439 dst[i] = ((b<<1)&0x07) | ((g&0x07)<<3) | ((r&0x03)<<6);
440 }
ac4d0aea 441}