00001
00002
00003
00004
00005
00006
00007
00008
00009
00010
00011
00012
00013
00014
00015
00016
00017
00018
00019
00020
00021
00022
00023
00024
00025 #include <inttypes.h>
00026 #include "config.h"
00027 #include "libavutil/x86_cpu.h"
00028 #include "libavutil/bswap.h"
00029 #include "rgb2rgb.h"
00030 #include "swscale.h"
00031 #include "swscale_internal.h"
00032
00033 #define FAST_BGR2YV12 // use 7-bit instead of 15-bit coefficients
00034
00035 void (*rgb24tobgr32)(const uint8_t *src, uint8_t *dst, long src_size);
00036 void (*rgb24tobgr16)(const uint8_t *src, uint8_t *dst, long src_size);
00037 void (*rgb24tobgr15)(const uint8_t *src, uint8_t *dst, long src_size);
00038 void (*rgb32tobgr24)(const uint8_t *src, uint8_t *dst, long src_size);
00039 void (*rgb32to16)(const uint8_t *src, uint8_t *dst, long src_size);
00040 void (*rgb32to15)(const uint8_t *src, uint8_t *dst, long src_size);
00041 void (*rgb15to16)(const uint8_t *src, uint8_t *dst, long src_size);
00042 void (*rgb15tobgr24)(const uint8_t *src, uint8_t *dst, long src_size);
00043 void (*rgb15to32)(const uint8_t *src, uint8_t *dst, long src_size);
00044 void (*rgb16to15)(const uint8_t *src, uint8_t *dst, long src_size);
00045 void (*rgb16tobgr24)(const uint8_t *src, uint8_t *dst, long src_size);
00046 void (*rgb16to32)(const uint8_t *src, uint8_t *dst, long src_size);
00047 void (*rgb24tobgr24)(const uint8_t *src, uint8_t *dst, long src_size);
00048 void (*rgb24to16)(const uint8_t *src, uint8_t *dst, long src_size);
00049 void (*rgb24to15)(const uint8_t *src, uint8_t *dst, long src_size);
00050 void (*rgb32tobgr32)(const uint8_t *src, uint8_t *dst, long src_size);
00051 void (*rgb32tobgr16)(const uint8_t *src, uint8_t *dst, long src_size);
00052 void (*rgb32tobgr15)(const uint8_t *src, uint8_t *dst, long src_size);
00053
00054 void (*yv12toyuy2)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
00055 long width, long height,
00056 long lumStride, long chromStride, long dstStride);
00057 void (*yv12touyvy)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
00058 long width, long height,
00059 long lumStride, long chromStride, long dstStride);
00060 void (*yuv422ptoyuy2)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
00061 long width, long height,
00062 long lumStride, long chromStride, long dstStride);
00063 void (*yuv422ptouyvy)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
00064 long width, long height,
00065 long lumStride, long chromStride, long dstStride);
00066 void (*yuy2toyv12)(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
00067 long width, long height,
00068 long lumStride, long chromStride, long srcStride);
00069 void (*rgb24toyv12)(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
00070 long width, long height,
00071 long lumStride, long chromStride, long srcStride);
00072 void (*planar2x)(const uint8_t *src, uint8_t *dst, long width, long height,
00073 long srcStride, long dstStride);
00074 void (*interleaveBytes)(const uint8_t *src1, const uint8_t *src2, uint8_t *dst,
00075 long width, long height, long src1Stride,
00076 long src2Stride, long dstStride);
00077 void (*vu9_to_vu12)(const uint8_t *src1, const uint8_t *src2,
00078 uint8_t *dst1, uint8_t *dst2,
00079 long width, long height,
00080 long srcStride1, long srcStride2,
00081 long dstStride1, long dstStride2);
00082 void (*yvu9_to_yuy2)(const uint8_t *src1, const uint8_t *src2, const uint8_t *src3,
00083 uint8_t *dst,
00084 long width, long height,
00085 long srcStride1, long srcStride2,
00086 long srcStride3, long dstStride);
00087 void (*uyvytoyuv420)(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, const uint8_t *src,
00088 long width, long height,
00089 long lumStride, long chromStride, long srcStride);
00090 void (*uyvytoyuv422)(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, const uint8_t *src,
00091 long width, long height,
00092 long lumStride, long chromStride, long srcStride);
00093 void (*yuyvtoyuv420)(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, const uint8_t *src,
00094 long width, long height,
00095 long lumStride, long chromStride, long srcStride);
00096 void (*yuyvtoyuv422)(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, const uint8_t *src,
00097 long width, long height,
00098 long lumStride, long chromStride, long srcStride);
00099
00100
00101 #if ARCH_X86
00102 DECLARE_ASM_CONST(8, uint64_t, mmx_null) = 0x0000000000000000ULL;
00103 DECLARE_ASM_CONST(8, uint64_t, mmx_one) = 0xFFFFFFFFFFFFFFFFULL;
00104 DECLARE_ASM_CONST(8, uint64_t, mask32b) = 0x000000FF000000FFULL;
00105 DECLARE_ASM_CONST(8, uint64_t, mask32g) = 0x0000FF000000FF00ULL;
00106 DECLARE_ASM_CONST(8, uint64_t, mask32r) = 0x00FF000000FF0000ULL;
00107 DECLARE_ASM_CONST(8, uint64_t, mask32a) = 0xFF000000FF000000ULL;
00108 DECLARE_ASM_CONST(8, uint64_t, mask32) = 0x00FFFFFF00FFFFFFULL;
00109 DECLARE_ASM_CONST(8, uint64_t, mask3216br) = 0x00F800F800F800F8ULL;
00110 DECLARE_ASM_CONST(8, uint64_t, mask3216g) = 0x0000FC000000FC00ULL;
00111 DECLARE_ASM_CONST(8, uint64_t, mask3215g) = 0x0000F8000000F800ULL;
00112 DECLARE_ASM_CONST(8, uint64_t, mul3216) = 0x2000000420000004ULL;
00113 DECLARE_ASM_CONST(8, uint64_t, mul3215) = 0x2000000820000008ULL;
00114 DECLARE_ASM_CONST(8, uint64_t, mask24b) = 0x00FF0000FF0000FFULL;
00115 DECLARE_ASM_CONST(8, uint64_t, mask24g) = 0xFF0000FF0000FF00ULL;
00116 DECLARE_ASM_CONST(8, uint64_t, mask24r) = 0x0000FF0000FF0000ULL;
00117 DECLARE_ASM_CONST(8, uint64_t, mask24l) = 0x0000000000FFFFFFULL;
00118 DECLARE_ASM_CONST(8, uint64_t, mask24h) = 0x0000FFFFFF000000ULL;
00119 DECLARE_ASM_CONST(8, uint64_t, mask24hh) = 0xffff000000000000ULL;
00120 DECLARE_ASM_CONST(8, uint64_t, mask24hhh) = 0xffffffff00000000ULL;
00121 DECLARE_ASM_CONST(8, uint64_t, mask24hhhh) = 0xffffffffffff0000ULL;
00122 DECLARE_ASM_CONST(8, uint64_t, mask15b) = 0x001F001F001F001FULL;
00123 DECLARE_ASM_CONST(8, uint64_t, mask15rg) = 0x7FE07FE07FE07FE0ULL;
00124 DECLARE_ASM_CONST(8, uint64_t, mask15s) = 0xFFE0FFE0FFE0FFE0ULL;
00125 DECLARE_ASM_CONST(8, uint64_t, mask15g) = 0x03E003E003E003E0ULL;
00126 DECLARE_ASM_CONST(8, uint64_t, mask15r) = 0x7C007C007C007C00ULL;
00127 #define mask16b mask15b
00128 DECLARE_ASM_CONST(8, uint64_t, mask16g) = 0x07E007E007E007E0ULL;
00129 DECLARE_ASM_CONST(8, uint64_t, mask16r) = 0xF800F800F800F800ULL;
00130 DECLARE_ASM_CONST(8, uint64_t, red_16mask) = 0x0000f8000000f800ULL;
00131 DECLARE_ASM_CONST(8, uint64_t, green_16mask) = 0x000007e0000007e0ULL;
00132 DECLARE_ASM_CONST(8, uint64_t, blue_16mask) = 0x0000001f0000001fULL;
00133 DECLARE_ASM_CONST(8, uint64_t, red_15mask) = 0x00007c0000007c00ULL;
00134 DECLARE_ASM_CONST(8, uint64_t, green_15mask) = 0x000003e0000003e0ULL;
00135 DECLARE_ASM_CONST(8, uint64_t, blue_15mask) = 0x0000001f0000001fULL;
00136 #endif
00137
00138 #define RGB2YUV_SHIFT 8
00139 #define BY ((int)( 0.098*(1<<RGB2YUV_SHIFT)+0.5))
00140 #define BV ((int)(-0.071*(1<<RGB2YUV_SHIFT)+0.5))
00141 #define BU ((int)( 0.439*(1<<RGB2YUV_SHIFT)+0.5))
00142 #define GY ((int)( 0.504*(1<<RGB2YUV_SHIFT)+0.5))
00143 #define GV ((int)(-0.368*(1<<RGB2YUV_SHIFT)+0.5))
00144 #define GU ((int)(-0.291*(1<<RGB2YUV_SHIFT)+0.5))
00145 #define RY ((int)( 0.257*(1<<RGB2YUV_SHIFT)+0.5))
00146 #define RV ((int)( 0.439*(1<<RGB2YUV_SHIFT)+0.5))
00147 #define RU ((int)(-0.148*(1<<RGB2YUV_SHIFT)+0.5))
00148
00149
00150
00151 #undef HAVE_MMX
00152 #undef HAVE_MMX2
00153 #undef HAVE_AMD3DNOW
00154 #undef HAVE_SSE2
00155 #define HAVE_MMX 0
00156 #define HAVE_MMX2 0
00157 #define HAVE_AMD3DNOW 0
00158 #define HAVE_SSE2 0
00159 #define RENAME(a) a ## _C
00160 #include "rgb2rgb_template.c"
00161
00162 #if ARCH_X86
00163
00164
00165 #undef RENAME
00166 #undef HAVE_MMX
00167 #define HAVE_MMX 1
00168 #define RENAME(a) a ## _MMX
00169 #include "rgb2rgb_template.c"
00170
00171
00172 #undef RENAME
00173 #undef HAVE_MMX2
00174 #define HAVE_MMX2 1
00175 #define RENAME(a) a ## _MMX2
00176 #include "rgb2rgb_template.c"
00177
00178
00179 #undef RENAME
00180 #undef HAVE_MMX2
00181 #undef HAVE_AMD3DNOW
00182 #define HAVE_MMX2 0
00183 #define HAVE_AMD3DNOW 1
00184 #define RENAME(a) a ## _3DNOW
00185 #include "rgb2rgb_template.c"
00186
00187 #endif //ARCH_X86 || ARCH_X86_64
00188
00189
00190
00191
00192
00193
00194
00195
00196 void sws_rgb2rgb_init(int flags)
00197 {
00198 #if HAVE_MMX2 || HAVE_AMD3DNOW || HAVE_MMX
00199 if (flags & SWS_CPU_CAPS_MMX2)
00200 rgb2rgb_init_MMX2();
00201 else if (flags & SWS_CPU_CAPS_3DNOW)
00202 rgb2rgb_init_3DNOW();
00203 else if (flags & SWS_CPU_CAPS_MMX)
00204 rgb2rgb_init_MMX();
00205 else
00206 #endif
00207 rgb2rgb_init_C();
00208 }
00209
00213 void palette8topacked32(const uint8_t *src, uint8_t *dst, long num_pixels, const uint8_t *palette)
00214 {
00215 long i;
00216
00217 for (i=0; i<num_pixels; i++)
00218 ((uint32_t *) dst)[i] = ((const uint32_t *) palette)[src[i]];
00219 }
00220
00224 void palette8topacked24(const uint8_t *src, uint8_t *dst, long num_pixels, const uint8_t *palette)
00225 {
00226 long i;
00227
00228 for (i=0; i<num_pixels; i++) {
00229
00230 dst[0]= palette[src[i]*4+0];
00231 dst[1]= palette[src[i]*4+1];
00232 dst[2]= palette[src[i]*4+2];
00233 dst+= 3;
00234 }
00235 }
00236
00240 void palette8torgb16(const uint8_t *src, uint8_t *dst, long num_pixels, const uint8_t *palette)
00241 {
00242 long i;
00243 for (i=0; i<num_pixels; i++)
00244 ((uint16_t *)dst)[i] = ((const uint16_t *)palette)[src[i]];
00245 }
00246 void palette8tobgr16(const uint8_t *src, uint8_t *dst, long num_pixels, const uint8_t *palette)
00247 {
00248 long i;
00249 for (i=0; i<num_pixels; i++)
00250 ((uint16_t *)dst)[i] = bswap_16(((const uint16_t *)palette)[src[i]]);
00251 }
00252
00256 void palette8torgb15(const uint8_t *src, uint8_t *dst, long num_pixels, const uint8_t *palette)
00257 {
00258 long i;
00259 for (i=0; i<num_pixels; i++)
00260 ((uint16_t *)dst)[i] = ((const uint16_t *)palette)[src[i]];
00261 }
00262 void palette8tobgr15(const uint8_t *src, uint8_t *dst, long num_pixels, const uint8_t *palette)
00263 {
00264 long i;
00265 for (i=0; i<num_pixels; i++)
00266 ((uint16_t *)dst)[i] = bswap_16(((const uint16_t *)palette)[src[i]]);
00267 }
00268
00269 void rgb32to24(const uint8_t *src, uint8_t *dst, long src_size)
00270 {
00271 long i;
00272 long num_pixels = src_size >> 2;
00273 for (i=0; i<num_pixels; i++) {
00274 #if HAVE_BIGENDIAN
00275
00276 dst[3*i + 0] = src[4*i + 1];
00277 dst[3*i + 1] = src[4*i + 2];
00278 dst[3*i + 2] = src[4*i + 3];
00279 #else
00280 dst[3*i + 0] = src[4*i + 2];
00281 dst[3*i + 1] = src[4*i + 1];
00282 dst[3*i + 2] = src[4*i + 0];
00283 #endif
00284 }
00285 }
00286
00287 void rgb24to32(const uint8_t *src, uint8_t *dst, long src_size)
00288 {
00289 long i;
00290 for (i=0; 3*i<src_size; i++) {
00291 #if HAVE_BIGENDIAN
00292
00293 dst[4*i + 0] = 255;
00294 dst[4*i + 1] = src[3*i + 0];
00295 dst[4*i + 2] = src[3*i + 1];
00296 dst[4*i + 3] = src[3*i + 2];
00297 #else
00298 dst[4*i + 0] = src[3*i + 2];
00299 dst[4*i + 1] = src[3*i + 1];
00300 dst[4*i + 2] = src[3*i + 0];
00301 dst[4*i + 3] = 255;
00302 #endif
00303 }
00304 }
00305
00306 void rgb16tobgr32(const uint8_t *src, uint8_t *dst, long src_size)
00307 {
00308 const uint16_t *end;
00309 uint8_t *d = dst;
00310 const uint16_t *s = (const uint16_t *)src;
00311 end = s + src_size/2;
00312 while (s < end) {
00313 register uint16_t bgr;
00314 bgr = *s++;
00315 #if HAVE_BIGENDIAN
00316 *d++ = 255;
00317 *d++ = (bgr&0x1F)<<3;
00318 *d++ = (bgr&0x7E0)>>3;
00319 *d++ = (bgr&0xF800)>>8;
00320 #else
00321 *d++ = (bgr&0xF800)>>8;
00322 *d++ = (bgr&0x7E0)>>3;
00323 *d++ = (bgr&0x1F)<<3;
00324 *d++ = 255;
00325 #endif
00326 }
00327 }
00328
00329 void rgb16to24(const uint8_t *src, uint8_t *dst, long src_size)
00330 {
00331 const uint16_t *end;
00332 uint8_t *d = dst;
00333 const uint16_t *s = (const uint16_t *)src;
00334 end = s + src_size/2;
00335 while (s < end) {
00336 register uint16_t bgr;
00337 bgr = *s++;
00338 *d++ = (bgr&0xF800)>>8;
00339 *d++ = (bgr&0x7E0)>>3;
00340 *d++ = (bgr&0x1F)<<3;
00341 }
00342 }
00343
00344 void rgb16tobgr16(const uint8_t *src, uint8_t *dst, long src_size)
00345 {
00346 long i;
00347 long num_pixels = src_size >> 1;
00348
00349 for (i=0; i<num_pixels; i++) {
00350 unsigned rgb = ((const uint16_t*)src)[i];
00351 ((uint16_t*)dst)[i] = (rgb>>11) | (rgb&0x7E0) | (rgb<<11);
00352 }
00353 }
00354
00355 void rgb16tobgr15(const uint8_t *src, uint8_t *dst, long src_size)
00356 {
00357 long i;
00358 long num_pixels = src_size >> 1;
00359
00360 for (i=0; i<num_pixels; i++) {
00361 unsigned rgb = ((const uint16_t*)src)[i];
00362 ((uint16_t*)dst)[i] = (rgb>>11) | ((rgb&0x7C0)>>1) | ((rgb&0x1F)<<10);
00363 }
00364 }
00365
00366 void rgb15tobgr32(const uint8_t *src, uint8_t *dst, long src_size)
00367 {
00368 const uint16_t *end;
00369 uint8_t *d = dst;
00370 const uint16_t *s = (const uint16_t *)src;
00371 end = s + src_size/2;
00372 while (s < end) {
00373 register uint16_t bgr;
00374 bgr = *s++;
00375 #if HAVE_BIGENDIAN
00376 *d++ = 255;
00377 *d++ = (bgr&0x1F)<<3;
00378 *d++ = (bgr&0x3E0)>>2;
00379 *d++ = (bgr&0x7C00)>>7;
00380 #else
00381 *d++ = (bgr&0x7C00)>>7;
00382 *d++ = (bgr&0x3E0)>>2;
00383 *d++ = (bgr&0x1F)<<3;
00384 *d++ = 255;
00385 #endif
00386 }
00387 }
00388
00389 void rgb15to24(const uint8_t *src, uint8_t *dst, long src_size)
00390 {
00391 const uint16_t *end;
00392 uint8_t *d = dst;
00393 const uint16_t *s = (const uint16_t *)src;
00394 end = s + src_size/2;
00395 while (s < end) {
00396 register uint16_t bgr;
00397 bgr = *s++;
00398 *d++ = (bgr&0x7C00)>>7;
00399 *d++ = (bgr&0x3E0)>>2;
00400 *d++ = (bgr&0x1F)<<3;
00401 }
00402 }
00403
00404 void rgb15tobgr16(const uint8_t *src, uint8_t *dst, long src_size)
00405 {
00406 long i;
00407 long num_pixels = src_size >> 1;
00408
00409 for (i=0; i<num_pixels; i++) {
00410 unsigned rgb = ((const uint16_t*)src)[i];
00411 ((uint16_t*)dst)[i] = ((rgb&0x7C00)>>10) | ((rgb&0x3E0)<<1) | (rgb<<11);
00412 }
00413 }
00414
00415 void rgb15tobgr15(const uint8_t *src, uint8_t *dst, long src_size)
00416 {
00417 long i;
00418 long num_pixels = src_size >> 1;
00419
00420 for (i=0; i<num_pixels; i++) {
00421 unsigned br;
00422 unsigned rgb = ((const uint16_t*)src)[i];
00423 br = rgb&0x7c1F;
00424 ((uint16_t*)dst)[i] = (br>>10) | (rgb&0x3E0) | (br<<10);
00425 }
00426 }
00427
00428 void bgr8torgb8(const uint8_t *src, uint8_t *dst, long src_size)
00429 {
00430 long i;
00431 long num_pixels = src_size;
00432 for (i=0; i<num_pixels; i++) {
00433 unsigned b,g,r;
00434 register uint8_t rgb;
00435 rgb = src[i];
00436 r = (rgb&0x07);
00437 g = (rgb&0x38)>>3;
00438 b = (rgb&0xC0)>>6;
00439 dst[i] = ((b<<1)&0x07) | ((g&0x07)<<3) | ((r&0x03)<<6);
00440 }
00441 }
00442
00443 #define DEFINE_SHUFFLE_BYTES(a, b, c, d) \
00444 void shuffle_bytes_##a##b##c##d(const uint8_t *src, uint8_t *dst, long src_size) \
00445 { \
00446 long i; \
00447 \
00448 for (i = 0; i < src_size; i+=4) { \
00449 dst[i + 0] = src[i + a]; \
00450 dst[i + 1] = src[i + b]; \
00451 dst[i + 2] = src[i + c]; \
00452 dst[i + 3] = src[i + d]; \
00453 } \
00454 }
00455
00456 DEFINE_SHUFFLE_BYTES(0, 3, 2, 1);
00457 DEFINE_SHUFFLE_BYTES(1, 2, 3, 0);
00458 DEFINE_SHUFFLE_BYTES(2, 1, 0, 3);
00459 DEFINE_SHUFFLE_BYTES(3, 0, 1, 2);
00460 DEFINE_SHUFFLE_BYTES(3, 2, 1, 0);
00461