libvo/osd.c

   1 // Generic alpha renderers for all YUV modes and RGB depths.
   2 // These are "reference implementations", should be optimized later (MMX, etc)
   3 // Templating Code from Michael Niedermayer (michaelni@gmx.at) is under GPL
   4
   5 //#define FAST_OSD
   6 //#define FAST_OSD_TABLE
   7
   8 #include "config.h"
   9 #include "osd.h"
  10 #include "mp_msg.h"
  11 #include <inttypes.h>
  12 #include "cpudetect.h"
  13
  14 #if ARCH_X86
  15 #define CAN_COMPILE_X86_ASM
  16 #endif
  17
  18 #ifdef CAN_COMPILE_X86_ASM
  19 static const uint64_t bFF __attribute__((aligned(8))) = 0xFFFFFFFFFFFFFFFFULL;
  20 static const unsigned long long mask24lh  __attribute__((aligned(8))) = 0xFFFF000000000000ULL;
  21 static const unsigned long long mask24hl  __attribute__((aligned(8))) = 0x0000FFFFFFFFFFFFULL;
  22 #endif
  23
  24 //Note: we have C, X86-nommx, MMX, MMX2, 3DNOW version therse no 3DNOW+MMX2 one
  25 //Plain C versions
  26 #if !HAVE_MMX || defined (RUNTIME_CPUDETECT)
  27 #define COMPILE_C
  28 #endif
  29
  30 #ifdef CAN_COMPILE_X86_ASM
  31
  32 #if (HAVE_MMX && !HAVE_AMD3DNOW && !HAVE_MMX2) || defined (RUNTIME_CPUDETECT)
  33 #define COMPILE_MMX
  34 #endif
  35
  36 #if HAVE_MMX2 || defined (RUNTIME_CPUDETECT)
  37 #define COMPILE_MMX2
  38 #endif
  39
  40 #if (HAVE_AMD3DNOW && !HAVE_MMX2) || defined (RUNTIME_CPUDETECT)
  41 #define COMPILE_3DNOW
  42 #endif
  43 #endif //CAN_COMPILE_X86_ASM
  44
  45 #undef HAVE_MMX
  46 #undef HAVE_MMX2
  47 #undef HAVE_AMD3DNOW
  48 #define HAVE_MMX 0
  49 #define HAVE_MMX2 0
  50 #define HAVE_AMD3DNOW 0
  51
  52 #ifndef CAN_COMPILE_X86_ASM
  53
  54 #ifdef COMPILE_C
  55 #undef HAVE_MMX
  56 #undef HAVE_MMX2
  57 #undef HAVE_AMD3DNOW
  58 #define HAVE_MMX 0
  59 #define HAVE_MMX2 0
  60 #define HAVE_AMD3DNOW 0
  61 #define RENAME(a) a ## _C
  62 #include "osd_template.c"
  63 #endif
  64
  65 #else
  66
  67 //X86 noMMX versions
  68 #ifdef COMPILE_C
  69 #undef RENAME
  70 #undef HAVE_MMX
  71 #undef HAVE_MMX2
  72 #undef HAVE_AMD3DNOW
  73 #define HAVE_MMX 0
  74 #define HAVE_MMX2 0
  75 #define HAVE_AMD3DNOW 0
  76 #define RENAME(a) a ## _X86
  77 #include "osd_template.c"
  78 #endif
  79
  80 //MMX versions
  81 #ifdef COMPILE_MMX
  82 #undef RENAME
  83 #undef HAVE_MMX
  84 #undef HAVE_MMX2
  85 #undef HAVE_AMD3DNOW
  86 #define HAVE_MMX 1
  87 #define HAVE_MMX2 0
  88 #define HAVE_AMD3DNOW 0
  89 #define RENAME(a) a ## _MMX
  90 #include "osd_template.c"
  91 #endif
  92
  93 //MMX2 versions
  94 #ifdef COMPILE_MMX2
  95 #undef RENAME
  96 #undef HAVE_MMX
  97 #undef HAVE_MMX2
  98 #undef HAVE_AMD3DNOW
  99 #define HAVE_MMX 1
 100 #define HAVE_MMX2 1
 101 #define HAVE_AMD3DNOW 0
 102 #define RENAME(a) a ## _MMX2
 103 #include "osd_template.c"
 104 #endif
 105
 106 //3DNOW versions
 107 #ifdef COMPILE_3DNOW
 108 #undef RENAME
 109 #undef HAVE_MMX
 110 #undef HAVE_MMX2
 111 #undef HAVE_AMD3DNOW
 112 #define HAVE_MMX 1
 113 #define HAVE_MMX2 0
 114 #define HAVE_AMD3DNOW 1
 115 #define RENAME(a) a ## _3DNow
 116 #include "osd_template.c"
 117 #endif
 118
 119 #endif //CAN_COMPILE_X86_ASM
 120
 121 void vo_draw_alpha_yv12(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 122 #ifdef RUNTIME_CPUDETECT
 123 #ifdef CAN_COMPILE_X86_ASM
 124         // ordered by speed / fastest first
 125         if(gCpuCaps.hasMMX2)
 126                 vo_draw_alpha_yv12_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 127         else if(gCpuCaps.has3DNow)
 128                 vo_draw_alpha_yv12_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 129         else if(gCpuCaps.hasMMX)
 130                 vo_draw_alpha_yv12_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 131         else
 132                 vo_draw_alpha_yv12_X86(w, h, src, srca, srcstride, dstbase, dststride);
 133 #else
 134                 vo_draw_alpha_yv12_C(w, h, src, srca, srcstride, dstbase, dststride);
 135 #endif
 136 #else //RUNTIME_CPUDETECT
 137 #if HAVE_MMX2
 138                 vo_draw_alpha_yv12_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 139 #elif HAVE_AMD3DNOW
 140                 vo_draw_alpha_yv12_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 141 #elif HAVE_MMX
 142                 vo_draw_alpha_yv12_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 143 #elif ARCH_X86
 144                 vo_draw_alpha_yv12_X86(w, h, src, srca, srcstride, dstbase, dststride);
 145 #else
 146                 vo_draw_alpha_yv12_C(w, h, src, srca, srcstride, dstbase, dststride);
 147 #endif
 148 #endif //!RUNTIME_CPUDETECT
 149 }
 150
 151 void vo_draw_alpha_yuy2(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 152 #ifdef RUNTIME_CPUDETECT
 153 #ifdef CAN_COMPILE_X86_ASM
 154         // ordered by speed / fastest first
 155         if(gCpuCaps.hasMMX2)
 156                 vo_draw_alpha_yuy2_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 157         else if(gCpuCaps.has3DNow)
 158                 vo_draw_alpha_yuy2_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 159         else if(gCpuCaps.hasMMX)
 160                 vo_draw_alpha_yuy2_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 161         else
 162                 vo_draw_alpha_yuy2_X86(w, h, src, srca, srcstride, dstbase, dststride);
 163 #else
 164                 vo_draw_alpha_yuy2_C(w, h, src, srca, srcstride, dstbase, dststride);
 165 #endif
 166 #else //RUNTIME_CPUDETECT
 167 #if HAVE_MMX2
 168                 vo_draw_alpha_yuy2_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 169 #elif HAVE_AMD3DNOW
 170                 vo_draw_alpha_yuy2_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 171 #elif HAVE_MMX
 172                 vo_draw_alpha_yuy2_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 173 #elif ARCH_X86
 174                 vo_draw_alpha_yuy2_X86(w, h, src, srca, srcstride, dstbase, dststride);
 175 #else
 176                 vo_draw_alpha_yuy2_C(w, h, src, srca, srcstride, dstbase, dststride);
 177 #endif
 178 #endif //!RUNTIME_CPUDETECT
 179 }
 180
 181 void vo_draw_alpha_uyvy(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 182 #ifdef RUNTIME_CPUDETECT
 183 #ifdef CAN_COMPILE_X86_ASM
 184         // ordered by speed / fastest first
 185         if(gCpuCaps.hasMMX2)
 186                 vo_draw_alpha_uyvy_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 187         else if(gCpuCaps.has3DNow)
 188                 vo_draw_alpha_uyvy_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 189         else if(gCpuCaps.hasMMX)
 190                 vo_draw_alpha_uyvy_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 191         else
 192                 vo_draw_alpha_uyvy_X86(w, h, src, srca, srcstride, dstbase, dststride);
 193 #else
 194                 vo_draw_alpha_uyvy_C(w, h, src, srca, srcstride, dstbase, dststride);
 195 #endif
 196 #else //RUNTIME_CPUDETECT
 197 #if HAVE_MMX2
 198                 vo_draw_alpha_uyvy_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 199 #elif HAVE_AMD3DNOW
 200                 vo_draw_alpha_uyvy_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 201 #elif HAVE_MMX
 202                 vo_draw_alpha_uyvy_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 203 #elif ARCH_X86
 204                 vo_draw_alpha_uyvy_X86(w, h, src, srca, srcstride, dstbase, dststride);
 205 #else
 206                 vo_draw_alpha_uyvy_C(w, h, src, srca, srcstride, dstbase, dststride);
 207 #endif
 208 #endif //!RUNTIME_CPUDETECT
 209 }
 210
 211 void vo_draw_alpha_rgb24(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 212 #ifdef RUNTIME_CPUDETECT
 213 #ifdef CAN_COMPILE_X86_ASM
 214         // ordered by speed / fastest first
 215         if(gCpuCaps.hasMMX2)
 216                 vo_draw_alpha_rgb24_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 217         else if(gCpuCaps.has3DNow)
 218                 vo_draw_alpha_rgb24_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 219         else if(gCpuCaps.hasMMX)
 220                 vo_draw_alpha_rgb24_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 221         else
 222                 vo_draw_alpha_rgb24_X86(w, h, src, srca, srcstride, dstbase, dststride);
 223 #else
 224                 vo_draw_alpha_rgb24_C(w, h, src, srca, srcstride, dstbase, dststride);
 225 #endif
 226 #else //RUNTIME_CPUDETECT
 227 #if HAVE_MMX2
 228                 vo_draw_alpha_rgb24_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 229 #elif HAVE_AMD3DNOW
 230                 vo_draw_alpha_rgb24_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 231 #elif HAVE_MMX
 232                 vo_draw_alpha_rgb24_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 233 #elif ARCH_X86
 234                 vo_draw_alpha_rgb24_X86(w, h, src, srca, srcstride, dstbase, dststride);
 235 #else
 236                 vo_draw_alpha_rgb24_C(w, h, src, srca, srcstride, dstbase, dststride);
 237 #endif
 238 #endif //!RUNTIME_CPUDETECT
 239 }
 240
 241 void vo_draw_alpha_rgb32(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 242 #ifdef RUNTIME_CPUDETECT
 243 #ifdef CAN_COMPILE_X86_ASM
 244         // ordered by speed / fastest first
 245         if(gCpuCaps.hasMMX2)
 246                 vo_draw_alpha_rgb32_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 247         else if(gCpuCaps.has3DNow)
 248                 vo_draw_alpha_rgb32_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 249         else if(gCpuCaps.hasMMX)
 250                 vo_draw_alpha_rgb32_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 251         else
 252                 vo_draw_alpha_rgb32_X86(w, h, src, srca, srcstride, dstbase, dststride);
 253 #else
 254                 vo_draw_alpha_rgb32_C(w, h, src, srca, srcstride, dstbase, dststride);
 255 #endif
 256 #else //RUNTIME_CPUDETECT
 257 #if HAVE_MMX2
 258                 vo_draw_alpha_rgb32_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 259 #elif HAVE_AMD3DNOW
 260                 vo_draw_alpha_rgb32_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 261 #elif HAVE_MMX
 262                 vo_draw_alpha_rgb32_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 263 #elif ARCH_X86
 264                 vo_draw_alpha_rgb32_X86(w, h, src, srca, srcstride, dstbase, dststride);
 265 #else
 266                 vo_draw_alpha_rgb32_C(w, h, src, srca, srcstride, dstbase, dststride);
 267 #endif
 268 #endif //!RUNTIME_CPUDETECT
 269 }
 270
 271 #ifdef FAST_OSD_TABLE
 272 static unsigned short fast_osd_15bpp_table[256];
 273 static unsigned short fast_osd_16bpp_table[256];
 274 #endif
 275
 276 void vo_draw_alpha_init(void){
 277 #ifdef FAST_OSD_TABLE
 278     int i;
 279     for(i=0;i<256;i++){
 280         fast_osd_15bpp_table[i]=((i>>3)<<10)|((i>>3)<<5)|(i>>3);
 281         fast_osd_16bpp_table[i]=((i>>3)<<11)|((i>>2)<<5)|(i>>3);
 282     }
 283 #endif
 284 //FIXME the optimized stuff is a lie for 15/16bpp as they aren't optimized yet
 285         if( mp_msg_test(MSGT_OSD,MSGL_V) )
 286         {
 287 #ifdef RUNTIME_CPUDETECT
 288 #ifdef CAN_COMPILE_X86_ASM
 289                 // ordered per speed fasterst first
 290                 if(gCpuCaps.hasMMX2)
 291                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX (with tiny bit MMX2) Optimized OnScreenDisplay\n");
 292                 else if(gCpuCaps.has3DNow)
 293                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX (with tiny bit 3DNow) Optimized OnScreenDisplay\n");
 294                 else if(gCpuCaps.hasMMX)
 295                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX Optimized OnScreenDisplay\n");
 296                 else
 297                         mp_msg(MSGT_OSD,MSGL_INFO,"Using X86 Optimized OnScreenDisplay\n");
 298 #else
 299                         mp_msg(MSGT_OSD,MSGL_INFO,"Using Unoptimized OnScreenDisplay\n");
 300 #endif
 301 #else //RUNTIME_CPUDETECT
 302 #if HAVE_MMX2
 303                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX (with tiny bit MMX2) Optimized OnScreenDisplay\n");
 304 #elif HAVE_AMD3DNOW
 305                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX (with tiny bit 3DNow) Optimized OnScreenDisplay\n");
 306 #elif HAVE_MMX
 307                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX Optimized OnScreenDisplay\n");
 308 #elif ARCH_X86
 309                         mp_msg(MSGT_OSD,MSGL_INFO,"Using X86 Optimized OnScreenDisplay\n");
 310 #else
 311                         mp_msg(MSGT_OSD,MSGL_INFO,"Using Unoptimized OnScreenDisplay\n");
 312 #endif
 313 #endif //!RUNTIME_CPUDETECT
 314         }
 315 }
 316
 317 void vo_draw_alpha_rgb15(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 318     int y;
 319     for(y=0;y<h;y++){
 320         register unsigned short *dst = (unsigned short*) dstbase;
 321         register int x;
 322         for(x=0;x<w;x++){
 323             if(srca[x]){
 324 #ifdef FAST_OSD
 325 #ifdef FAST_OSD_TABLE
 326                 dst[x]=fast_osd_15bpp_table[src[x]];
 327 #else
 328                 register unsigned int a=src[x]>>3;
 329                 dst[x]=(a<<10)|(a<<5)|a;
 330 #endif
 331 #else
 332                 unsigned char r=dst[x]&0x1F;
 333                 unsigned char g=(dst[x]>>5)&0x1F;
 334                 unsigned char b=(dst[x]>>10)&0x1F;
 335                 r=(((r*srca[x])>>5)+src[x])>>3;
 336                 g=(((g*srca[x])>>5)+src[x])>>3;
 337                 b=(((b*srca[x])>>5)+src[x])>>3;
 338                 dst[x]=(b<<10)|(g<<5)|r;
 339 #endif
 340             }
 341         }
 342         src+=srcstride;
 343         srca+=srcstride;
 344         dstbase+=dststride;
 345     }
 346     return;
 347 }
 348
 349 void vo_draw_alpha_rgb16(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 350     int y;
 351     for(y=0;y<h;y++){
 352         register unsigned short *dst = (unsigned short*) dstbase;
 353         register int x;
 354         for(x=0;x<w;x++){
 355             if(srca[x]){
 356 #ifdef FAST_OSD
 357 #ifdef FAST_OSD_TABLE
 358                 dst[x]=fast_osd_16bpp_table[src[x]];
 359 #else
 360                 dst[x]=((src[x]>>3)<<11)|((src[x]>>2)<<5)|(src[x]>>3);
 361 #endif
 362 #else
 363                 unsigned char r=dst[x]&0x1F;
 364                 unsigned char g=(dst[x]>>5)&0x3F;
 365                 unsigned char b=(dst[x]>>11)&0x1F;
 366                 r=(((r*srca[x])>>5)+src[x])>>3;
 367                 g=(((g*srca[x])>>6)+src[x])>>2;
 368                 b=(((b*srca[x])>>5)+src[x])>>3;
 369                 dst[x]=(b<<11)|(g<<5)|r;
 370 #endif
 371             }
 372         }
 373         src+=srcstride;
 374         srca+=srcstride;
 375         dstbase+=dststride;
 376     }
 377     return;
 378 }
 379