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 #ifdef 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 !defined (HAVE_MMX) || defined (RUNTIME_CPUDETECT)
  27 #define COMPILE_C
  28 #endif
  29
  30 #ifdef CAN_COMPILE_X86_ASM
  31
  32 #if (defined (HAVE_MMX) && !defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
  33 #define COMPILE_MMX
  34 #endif
  35
  36 #if defined (HAVE_MMX2) || defined (RUNTIME_CPUDETECT)
  37 #define COMPILE_MMX2
  38 #endif
  39
  40 #if (defined (HAVE_3DNOW) && !defined (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_3DNOW
  48
  49 #ifndef CAN_COMPILE_X86_ASM
  50
  51 #ifdef COMPILE_C
  52 #undef HAVE_MMX
  53 #undef HAVE_MMX2
  54 #undef HAVE_3DNOW
  55 #define RENAME(a) a ## _C
  56 #include "osd_template.c"
  57 #endif
  58
  59 #else
  60
  61 //X86 noMMX versions
  62 #ifdef COMPILE_C
  63 #undef RENAME
  64 #undef HAVE_MMX
  65 #undef HAVE_MMX2
  66 #undef HAVE_3DNOW
  67 #define RENAME(a) a ## _X86
  68 #include "osd_template.c"
  69 #endif
  70
  71 //MMX versions
  72 #ifdef COMPILE_MMX
  73 #undef RENAME
  74 #define HAVE_MMX
  75 #undef HAVE_MMX2
  76 #undef HAVE_3DNOW
  77 #define RENAME(a) a ## _MMX
  78 #include "osd_template.c"
  79 #endif
  80
  81 //MMX2 versions
  82 #ifdef COMPILE_MMX2
  83 #undef RENAME
  84 #define HAVE_MMX
  85 #define HAVE_MMX2
  86 #undef HAVE_3DNOW
  87 #define RENAME(a) a ## _MMX2
  88 #include "osd_template.c"
  89 #endif
  90
  91 //3DNOW versions
  92 #ifdef COMPILE_3DNOW
  93 #undef RENAME
  94 #define HAVE_MMX
  95 #undef HAVE_MMX2
  96 #define HAVE_3DNOW
  97 #define RENAME(a) a ## _3DNow
  98 #include "osd_template.c"
  99 #endif
 100
 101 #endif //CAN_COMPILE_X86_ASM
 102
 103 void vo_draw_alpha_yv12(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 104 #ifdef RUNTIME_CPUDETECT
 105 #ifdef CAN_COMPILE_X86_ASM
 106         // ordered by speed / fastest first
 107         if(gCpuCaps.hasMMX2)
 108                 vo_draw_alpha_yv12_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 109         else if(gCpuCaps.has3DNow)
 110                 vo_draw_alpha_yv12_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 111         else if(gCpuCaps.hasMMX)
 112                 vo_draw_alpha_yv12_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 113         else
 114                 vo_draw_alpha_yv12_X86(w, h, src, srca, srcstride, dstbase, dststride);
 115 #else
 116                 vo_draw_alpha_yv12_C(w, h, src, srca, srcstride, dstbase, dststride);
 117 #endif
 118 #else //RUNTIME_CPUDETECT
 119 #ifdef HAVE_MMX2
 120                 vo_draw_alpha_yv12_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 121 #elif defined (HAVE_3DNOW)
 122                 vo_draw_alpha_yv12_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 123 #elif defined (HAVE_MMX)
 124                 vo_draw_alpha_yv12_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 125 #elif defined(ARCH_X86)
 126                 vo_draw_alpha_yv12_X86(w, h, src, srca, srcstride, dstbase, dststride);
 127 #else
 128                 vo_draw_alpha_yv12_C(w, h, src, srca, srcstride, dstbase, dststride);
 129 #endif
 130 #endif //!RUNTIME_CPUDETECT
 131 }
 132
 133 void vo_draw_alpha_yuy2(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 134 #ifdef RUNTIME_CPUDETECT
 135 #ifdef CAN_COMPILE_X86_ASM
 136         // ordered by speed / fastest first
 137         if(gCpuCaps.hasMMX2)
 138                 vo_draw_alpha_yuy2_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 139         else if(gCpuCaps.has3DNow)
 140                 vo_draw_alpha_yuy2_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 141         else if(gCpuCaps.hasMMX)
 142                 vo_draw_alpha_yuy2_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 143         else
 144                 vo_draw_alpha_yuy2_X86(w, h, src, srca, srcstride, dstbase, dststride);
 145 #else
 146                 vo_draw_alpha_yuy2_C(w, h, src, srca, srcstride, dstbase, dststride);
 147 #endif
 148 #else //RUNTIME_CPUDETECT
 149 #ifdef HAVE_MMX2
 150                 vo_draw_alpha_yuy2_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 151 #elif defined (HAVE_3DNOW)
 152                 vo_draw_alpha_yuy2_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 153 #elif defined (HAVE_MMX)
 154                 vo_draw_alpha_yuy2_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 155 #elif defined(ARCH_X86)
 156                 vo_draw_alpha_yuy2_X86(w, h, src, srca, srcstride, dstbase, dststride);
 157 #else
 158                 vo_draw_alpha_yuy2_C(w, h, src, srca, srcstride, dstbase, dststride);
 159 #endif
 160 #endif //!RUNTIME_CPUDETECT
 161 }
 162
 163 void vo_draw_alpha_uyvy(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 164 #ifdef RUNTIME_CPUDETECT
 165 #ifdef CAN_COMPILE_X86_ASM
 166         // ordered by speed / fastest first
 167         if(gCpuCaps.hasMMX2)
 168                 vo_draw_alpha_uyvy_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 169         else if(gCpuCaps.has3DNow)
 170                 vo_draw_alpha_uyvy_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 171         else if(gCpuCaps.hasMMX)
 172                 vo_draw_alpha_uyvy_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 173         else
 174                 vo_draw_alpha_uyvy_X86(w, h, src, srca, srcstride, dstbase, dststride);
 175 #else
 176                 vo_draw_alpha_uyvy_C(w, h, src, srca, srcstride, dstbase, dststride);
 177 #endif
 178 #else //RUNTIME_CPUDETECT
 179 #ifdef HAVE_MMX2
 180                 vo_draw_alpha_uyvy_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 181 #elif defined (HAVE_3DNOW)
 182                 vo_draw_alpha_uyvy_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 183 #elif defined (HAVE_MMX)
 184                 vo_draw_alpha_uyvy_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 185 #elif defined(ARCH_X86)
 186                 vo_draw_alpha_uyvy_X86(w, h, src, srca, srcstride, dstbase, dststride);
 187 #else
 188                 vo_draw_alpha_uyvy_C(w, h, src, srca, srcstride, dstbase, dststride);
 189 #endif
 190 #endif //!RUNTIME_CPUDETECT
 191 }
 192
 193 void vo_draw_alpha_rgb24(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 194 #ifdef RUNTIME_CPUDETECT
 195 #ifdef CAN_COMPILE_X86_ASM
 196         // ordered by speed / fastest first
 197         if(gCpuCaps.hasMMX2)
 198                 vo_draw_alpha_rgb24_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 199         else if(gCpuCaps.has3DNow)
 200                 vo_draw_alpha_rgb24_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 201         else if(gCpuCaps.hasMMX)
 202                 vo_draw_alpha_rgb24_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 203         else
 204                 vo_draw_alpha_rgb24_X86(w, h, src, srca, srcstride, dstbase, dststride);
 205 #else
 206                 vo_draw_alpha_rgb24_C(w, h, src, srca, srcstride, dstbase, dststride);
 207 #endif
 208 #else //RUNTIME_CPUDETECT
 209 #ifdef HAVE_MMX2
 210                 vo_draw_alpha_rgb24_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 211 #elif defined (HAVE_3DNOW)
 212                 vo_draw_alpha_rgb24_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 213 #elif defined (HAVE_MMX)
 214                 vo_draw_alpha_rgb24_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 215 #elif defined(ARCH_X86)
 216                 vo_draw_alpha_rgb24_X86(w, h, src, srca, srcstride, dstbase, dststride);
 217 #else
 218                 vo_draw_alpha_rgb24_C(w, h, src, srca, srcstride, dstbase, dststride);
 219 #endif
 220 #endif //!RUNTIME_CPUDETECT
 221 }
 222
 223 void vo_draw_alpha_rgb32(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 224 #ifdef RUNTIME_CPUDETECT
 225 #ifdef CAN_COMPILE_X86_ASM
 226         // ordered by speed / fastest first
 227         if(gCpuCaps.hasMMX2)
 228                 vo_draw_alpha_rgb32_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 229         else if(gCpuCaps.has3DNow)
 230                 vo_draw_alpha_rgb32_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 231         else if(gCpuCaps.hasMMX)
 232                 vo_draw_alpha_rgb32_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 233         else
 234                 vo_draw_alpha_rgb32_X86(w, h, src, srca, srcstride, dstbase, dststride);
 235 #else
 236                 vo_draw_alpha_rgb32_C(w, h, src, srca, srcstride, dstbase, dststride);
 237 #endif
 238 #else //RUNTIME_CPUDETECT
 239 #ifdef HAVE_MMX2
 240                 vo_draw_alpha_rgb32_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 241 #elif defined (HAVE_3DNOW)
 242                 vo_draw_alpha_rgb32_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 243 #elif defined (HAVE_MMX)
 244                 vo_draw_alpha_rgb32_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 245 #elif defined(ARCH_X86)
 246                 vo_draw_alpha_rgb32_X86(w, h, src, srca, srcstride, dstbase, dststride);
 247 #else
 248                 vo_draw_alpha_rgb32_C(w, h, src, srca, srcstride, dstbase, dststride);
 249 #endif
 250 #endif //!RUNTIME_CPUDETECT
 251 }
 252
 253 #ifdef FAST_OSD_TABLE
 254 static unsigned short fast_osd_15bpp_table[256];
 255 static unsigned short fast_osd_16bpp_table[256];
 256 #endif
 257
 258 void vo_draw_alpha_init(void){
 259 #ifdef FAST_OSD_TABLE
 260     int i;
 261     for(i=0;i<256;i++){
 262         fast_osd_15bpp_table[i]=((i>>3)<<10)|((i>>3)<<5)|(i>>3);
 263         fast_osd_16bpp_table[i]=((i>>3)<<11)|((i>>2)<<5)|(i>>3);
 264     }
 265 #endif
 266 //FIXME the optimized stuff is a lie for 15/16bpp as they aren't optimized yet
 267         if( mp_msg_test(MSGT_OSD,MSGL_V) )
 268         {
 269 #ifdef RUNTIME_CPUDETECT
 270 #ifdef CAN_COMPILE_X86_ASM
 271                 // ordered per speed fasterst first
 272                 if(gCpuCaps.hasMMX2)
 273                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX (with tiny bit MMX2) Optimized OnScreenDisplay\n");
 274                 else if(gCpuCaps.has3DNow)
 275                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX (with tiny bit 3DNow) Optimized OnScreenDisplay\n");
 276                 else if(gCpuCaps.hasMMX)
 277                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX Optimized OnScreenDisplay\n");
 278                 else
 279                         mp_msg(MSGT_OSD,MSGL_INFO,"Using X86 Optimized OnScreenDisplay\n");
 280 #else
 281                         mp_msg(MSGT_OSD,MSGL_INFO,"Using Unoptimized OnScreenDisplay\n");
 282 #endif
 283 #else //RUNTIME_CPUDETECT
 284 #ifdef HAVE_MMX2
 285                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX (with tiny bit MMX2) Optimized OnScreenDisplay\n");
 286 #elif defined (HAVE_3DNOW)
 287                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX (with tiny bit 3DNow) Optimized OnScreenDisplay\n");
 288 #elif defined (HAVE_MMX)
 289                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX Optimized OnScreenDisplay\n");
 290 #elif defined(ARCH_X86)
 291                         mp_msg(MSGT_OSD,MSGL_INFO,"Using X86 Optimized OnScreenDisplay\n");
 292 #else
 293                         mp_msg(MSGT_OSD,MSGL_INFO,"Using Unoptimized OnScreenDisplay\n");
 294 #endif
 295 #endif //!RUNTIME_CPUDETECT
 296         }
 297 }
 298
 299 void vo_draw_alpha_rgb15(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 300     int y;
 301     for(y=0;y<h;y++){
 302         register unsigned short *dst = (unsigned short*) dstbase;
 303         register int x;
 304         for(x=0;x<w;x++){
 305             if(srca[x]){
 306 #ifdef FAST_OSD
 307 #ifdef FAST_OSD_TABLE
 308                 dst[x]=fast_osd_15bpp_table[src[x]];
 309 #else
 310                 register unsigned int a=src[x]>>3;
 311                 dst[x]=(a<<10)|(a<<5)|a;
 312 #endif
 313 #else
 314                 unsigned char r=dst[x]&0x1F;
 315                 unsigned char g=(dst[x]>>5)&0x1F;
 316                 unsigned char b=(dst[x]>>10)&0x1F;
 317                 r=(((r*srca[x])>>5)+src[x])>>3;
 318                 g=(((g*srca[x])>>5)+src[x])>>3;
 319                 b=(((b*srca[x])>>5)+src[x])>>3;
 320                 dst[x]=(b<<10)|(g<<5)|r;
 321 #endif
 322             }
 323         }
 324         src+=srcstride;
 325         srca+=srcstride;
 326         dstbase+=dststride;
 327     }
 328     return;
 329 }
 330
 331 void vo_draw_alpha_rgb16(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 332     int y;
 333     for(y=0;y<h;y++){
 334         register unsigned short *dst = (unsigned short*) dstbase;
 335         register int x;
 336         for(x=0;x<w;x++){
 337             if(srca[x]){
 338 #ifdef FAST_OSD
 339 #ifdef FAST_OSD_TABLE
 340                 dst[x]=fast_osd_16bpp_table[src[x]];
 341 #else
 342                 dst[x]=((src[x]>>3)<<11)|((src[x]>>2)<<5)|(src[x]>>3);
 343 #endif
 344 #else
 345                 unsigned char r=dst[x]&0x1F;
 346                 unsigned char g=(dst[x]>>5)&0x3F;
 347                 unsigned char b=(dst[x]>>11)&0x1F;
 348                 r=(((r*srca[x])>>5)+src[x])>>3;
 349                 g=(((g*srca[x])>>6)+src[x])>>2;
 350                 b=(((b*srca[x])>>5)+src[x])>>3;
 351                 dst[x]=(b<<11)|(g<<5)|r;
 352 #endif
 353             }
 354         }
 355         src+=srcstride;
 356         srca+=srcstride;
 357         dstbase+=dststride;
 358     }
 359     return;
 360 }
 361