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 #include "../mangle.h"
  14
  15 extern int verbose; // defined in mplayer.c
  16
  17 #ifdef ARCH_X86
  18 #define CAN_COMPILE_X86_ASM
  19 #endif
  20
  21 #ifdef CAN_COMPILE_X86_ASM
  22 static const uint64_t bFF  __attribute__((aligned(8))) = 0xFFFFFFFFFFFFFFFFULL;
  23 static const unsigned long long mask24lh  __attribute__((aligned(8))) = 0xFFFF000000000000ULL;
  24 static const unsigned long long mask24hl  __attribute__((aligned(8))) = 0x0000FFFFFFFFFFFFULL;
  25 #endif
  26
  27 //Note: we have C, X86-nommx, MMX, MMX2, 3DNOW version therse no 3DNOW+MMX2 one
  28 //Plain C versions
  29 #if !defined (HAVE_MMX) || defined (RUNTIME_CPUDETECT)
  30 #define COMPILE_C
  31 #endif
  32
  33 #ifdef CAN_COMPILE_X86_ASM
  34
  35 #if (defined (HAVE_MMX) && !defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
  36 #define COMPILE_MMX
  37 #endif
  38
  39 #if defined (HAVE_MMX2) || defined (RUNTIME_CPUDETECT)
  40 #define COMPILE_MMX2
  41 #endif
  42
  43 #if (defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
  44 #define COMPILE_3DNOW
  45 #endif
  46 #endif //CAN_COMPILE_X86_ASM
  47
  48 #undef HAVE_MMX
  49 #undef HAVE_MMX2
  50 #undef HAVE_3DNOW
  51 #undef ARCH_X86
  52
  53 #ifdef COMPILE_C
  54 #undef HAVE_MMX
  55 #undef HAVE_MMX2
  56 #undef HAVE_3DNOW
  57 #undef ARCH_X86
  58 #define RENAME(a) a ## _C
  59 #include "osd_template.c"
  60 #endif
  61
  62 #ifdef CAN_COMPILE_X86_ASM
  63
  64 //X86 noMMX versions
  65 #ifdef COMPILE_C
  66 #undef RENAME
  67 #undef HAVE_MMX
  68 #undef HAVE_MMX2
  69 #undef HAVE_3DNOW
  70 #define ARCH_X86
  71 #define RENAME(a) a ## _X86
  72 #include "osd_template.c"
  73 #endif
  74
  75 //MMX versions
  76 #ifdef COMPILE_MMX
  77 #undef RENAME
  78 #define HAVE_MMX
  79 #undef HAVE_MMX2
  80 #undef HAVE_3DNOW
  81 #define ARCH_X86
  82 #define RENAME(a) a ## _MMX
  83 #include "osd_template.c"
  84 #endif
  85
  86 //MMX2 versions
  87 #ifdef COMPILE_MMX2
  88 #undef RENAME
  89 #define HAVE_MMX
  90 #define HAVE_MMX2
  91 #undef HAVE_3DNOW
  92 #define ARCH_X86
  93 #define RENAME(a) a ## _MMX2
  94 #include "osd_template.c"
  95 #endif
  96
  97 //3DNOW versions
  98 #ifdef COMPILE_3DNOW
  99 #undef RENAME
 100 #define HAVE_MMX
 101 #undef HAVE_MMX2
 102 #define HAVE_3DNOW
 103 #define ARCH_X86
 104 #define RENAME(a) a ## _3DNow
 105 #include "osd_template.c"
 106 #endif
 107
 108 #endif //CAN_COMPILE_X86_ASM
 109
 110 void vo_draw_alpha_yv12(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 111 #ifdef RUNTIME_CPUDETECT
 112 #ifdef CAN_COMPILE_X86_ASM
 113         // ordered per speed fasterst first
 114         if(gCpuCaps.hasMMX2)
 115                 vo_draw_alpha_yv12_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 116         else if(gCpuCaps.has3DNow)
 117                 vo_draw_alpha_yv12_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 118         else if(gCpuCaps.hasMMX)
 119                 vo_draw_alpha_yv12_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 120         else
 121                 vo_draw_alpha_yv12_X86(w, h, src, srca, srcstride, dstbase, dststride);
 122 #else
 123                 vo_draw_alpha_yv12_C(w, h, src, srca, srcstride, dstbase, dststride);
 124 #endif
 125 #else //RUNTIME_CPUDETECT
 126 #ifdef HAVE_MMX2
 127                 vo_draw_alpha_yv12_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 128 #elif defined (HAVE_3DNOW)
 129                 vo_draw_alpha_yv12_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 130 #elif defined (HAVE_MMX)
 131                 vo_draw_alpha_yv12_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 132 #elif defined (ARCH_X86)
 133                 vo_draw_alpha_yv12_X86(w, h, src, srca, srcstride, dstbase, dststride);
 134 #else
 135                 vo_draw_alpha_yv12_C(w, h, src, srca, srcstride, dstbase, dststride);
 136 #endif
 137 #endif //!RUNTIME_CPUDETECT
 138 }
 139
 140 void vo_draw_alpha_yuy2(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 141 #ifdef RUNTIME_CPUDETECT
 142 #ifdef CAN_COMPILE_X86_ASM
 143         // ordered per speed fasterst first
 144         if(gCpuCaps.hasMMX2)
 145                 vo_draw_alpha_yuy2_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 146         else if(gCpuCaps.has3DNow)
 147                 vo_draw_alpha_yuy2_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 148         else if(gCpuCaps.hasMMX)
 149                 vo_draw_alpha_yuy2_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 150         else
 151                 vo_draw_alpha_yuy2_X86(w, h, src, srca, srcstride, dstbase, dststride);
 152 #else
 153                 vo_draw_alpha_yuy2_C(w, h, src, srca, srcstride, dstbase, dststride);
 154 #endif
 155 #else //RUNTIME_CPUDETECT
 156 #ifdef HAVE_MMX2
 157                 vo_draw_alpha_yuy2_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 158 #elif defined (HAVE_3DNOW)
 159                 vo_draw_alpha_yuy2_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 160 #elif defined (HAVE_MMX)
 161                 vo_draw_alpha_yuy2_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 162 #elif defined (ARCH_X86)
 163                 vo_draw_alpha_yuy2_X86(w, h, src, srca, srcstride, dstbase, dststride);
 164 #else
 165                 vo_draw_alpha_yuy2_C(w, h, src, srca, srcstride, dstbase, dststride);
 166 #endif
 167 #endif //!RUNTIME_CPUDETECT
 168 }
 169
 170 void vo_draw_alpha_rgb24(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 171 #ifdef RUNTIME_CPUDETECT
 172 #ifdef CAN_COMPILE_X86_ASM
 173         // ordered per speed fasterst first
 174         if(gCpuCaps.hasMMX2)
 175                 vo_draw_alpha_rgb24_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 176         else if(gCpuCaps.has3DNow)
 177                 vo_draw_alpha_rgb24_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 178         else if(gCpuCaps.hasMMX)
 179                 vo_draw_alpha_rgb24_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 180         else
 181                 vo_draw_alpha_rgb24_X86(w, h, src, srca, srcstride, dstbase, dststride);
 182 #else
 183                 vo_draw_alpha_rgb24_C(w, h, src, srca, srcstride, dstbase, dststride);
 184 #endif
 185 #else //RUNTIME_CPUDETECT
 186 #ifdef HAVE_MMX2
 187                 vo_draw_alpha_rgb24_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 188 #elif defined (HAVE_3DNOW)
 189                 vo_draw_alpha_rgb24_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 190 #elif defined (HAVE_MMX)
 191                 vo_draw_alpha_rgb24_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 192 #elif defined (ARCH_X86)
 193                 vo_draw_alpha_rgb24_X86(w, h, src, srca, srcstride, dstbase, dststride);
 194 #else
 195                 vo_draw_alpha_rgb24_C(w, h, src, srca, srcstride, dstbase, dststride);
 196 #endif
 197 #endif //!RUNTIME_CPUDETECT
 198 }
 199
 200 void vo_draw_alpha_rgb32(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 201 #ifdef RUNTIME_CPUDETECT
 202 #ifdef CAN_COMPILE_X86_ASM
 203         // ordered per speed fasterst first
 204         if(gCpuCaps.hasMMX2)
 205                 vo_draw_alpha_rgb32_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 206         else if(gCpuCaps.has3DNow)
 207                 vo_draw_alpha_rgb32_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 208         else if(gCpuCaps.hasMMX)
 209                 vo_draw_alpha_rgb32_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 210         else
 211                 vo_draw_alpha_rgb32_X86(w, h, src, srca, srcstride, dstbase, dststride);
 212 #else
 213                 vo_draw_alpha_rgb32_C(w, h, src, srca, srcstride, dstbase, dststride);
 214 #endif
 215 #else //RUNTIME_CPUDETECT
 216 #ifdef HAVE_MMX2
 217                 vo_draw_alpha_rgb32_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 218 #elif defined (HAVE_3DNOW)
 219                 vo_draw_alpha_rgb32_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 220 #elif defined (HAVE_MMX)
 221                 vo_draw_alpha_rgb32_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 222 #elif defined (ARCH_X86)
 223                 vo_draw_alpha_rgb32_X86(w, h, src, srca, srcstride, dstbase, dststride);
 224 #else
 225                 vo_draw_alpha_rgb32_C(w, h, src, srca, srcstride, dstbase, dststride);
 226 #endif
 227 #endif //!RUNTIME_CPUDETECT
 228 }
 229
 230 #ifdef FAST_OSD_TABLE
 231 static unsigned short fast_osd_15bpp_table[256];
 232 static unsigned short fast_osd_16bpp_table[256];
 233 #endif
 234
 235 void vo_draw_alpha_init(){
 236 #ifdef FAST_OSD_TABLE
 237     int i;
 238     for(i=0;i<256;i++){
 239         fast_osd_15bpp_table[i]=((i>>3)<<10)|((i>>3)<<5)|(i>>3);
 240         fast_osd_16bpp_table[i]=((i>>3)<<11)|((i>>2)<<5)|(i>>3);
 241     }
 242 #endif
 243 //FIXME the optimized stuff is a lie for 15/16bpp as they aren't optimized yet
 244         if(verbose)
 245         {
 246 #ifdef RUNTIME_CPUDETECT
 247 #ifdef CAN_COMPILE_X86_ASM
 248                 // ordered per speed fasterst first
 249                 if(gCpuCaps.hasMMX2)
 250                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX (with tiny bit MMX2) Optimized OnScreenDisplay\n");
 251                 else if(gCpuCaps.has3DNow)
 252                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX (with tiny bit 3DNow) Optimized OnScreenDisplay\n");
 253                 else if(gCpuCaps.hasMMX)
 254                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX Optimized OnScreenDisplay\n");
 255                 else
 256                         mp_msg(MSGT_OSD,MSGL_INFO,"Using X86 Optimized OnScreenDisplay\n");
 257 #else
 258                         mp_msg(MSGT_OSD,MSGL_INFO,"Using Unoptimized OnScreenDisplay\n");
 259 #endif
 260 #else //RUNTIME_CPUDETECT
 261 #ifdef HAVE_MMX2
 262                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX (with tiny bit MMX2) Optimized OnScreenDisplay\n");
 263 #elif defined (HAVE_3DNOW)
 264                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX (with tiny bit 3DNow) Optimized OnScreenDisplay\n");
 265 #elif defined (HAVE_MMX)
 266                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX Optimized OnScreenDisplay\n");
 267 #elif defined (ARCH_X86)
 268                         mp_msg(MSGT_OSD,MSGL_INFO,"Using X86 Optimized OnScreenDisplay\n");
 269 #else
 270                         mp_msg(MSGT_OSD,MSGL_INFO,"Using Unoptimized OnScreenDisplay\n");
 271 #endif
 272 #endif //!RUNTIME_CPUDETECT
 273         }
 274 }
 275
 276 void vo_draw_alpha_rgb15(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 277     int y;
 278     for(y=0;y<h;y++){
 279         register unsigned short *dst = (unsigned short*) dstbase;
 280         register int x;
 281         for(x=0;x<w;x++){
 282             if(srca[x]){
 283 #ifdef FAST_OSD
 284 #ifdef FAST_OSD_TABLE
 285                 dst[x]=fast_osd_15bpp_table[src[x]];
 286 #else
 287                 register unsigned int a=src[x]>>3;
 288                 dst[x]=(a<<10)|(a<<5)|a;
 289 #endif
 290 #else
 291                 unsigned char r=dst[x]&0x1F;
 292                 unsigned char g=(dst[x]>>5)&0x1F;
 293                 unsigned char b=(dst[x]>>10)&0x1F;
 294                 r=(((r*srca[x])>>5)+src[x])>>3;
 295                 g=(((g*srca[x])>>5)+src[x])>>3;
 296                 b=(((b*srca[x])>>5)+src[x])>>3;
 297                 dst[x]=(b<<10)|(g<<5)|r;
 298 #endif
 299             }
 300         }
 301         src+=srcstride;
 302         srca+=srcstride;
 303         dstbase+=dststride;
 304     }
 305     return;
 306 }
 307
 308 void vo_draw_alpha_rgb16(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 309     int y;
 310     for(y=0;y<h;y++){
 311         register unsigned short *dst = (unsigned short*) dstbase;
 312         register int x;
 313         for(x=0;x<w;x++){
 314             if(srca[x]){
 315 #ifdef FAST_OSD
 316 #ifdef FAST_OSD_TABLE
 317                 dst[x]=fast_osd_16bpp_table[src[x]];
 318 #else
 319                 dst[x]=((src[x]>>3)<<11)|((src[x]>>2)<<5)|(src[x]>>3);
 320 #endif
 321 #else
 322                 unsigned char r=dst[x]&0x1F;
 323                 unsigned char g=(dst[x]>>5)&0x3F;
 324                 unsigned char b=(dst[x]>>11)&0x1F;
 325                 r=(((r*srca[x])>>5)+src[x])>>3;
 326                 g=(((g*srca[x])>>6)+src[x])>>2;
 327                 b=(((b*srca[x])>>5)+src[x])>>3;
 328                 dst[x]=(b<<11)|(g<<5)|r;
 329 #endif
 330             }
 331         }
 332         src+=srcstride;
 333         srca+=srcstride;
 334         dstbase+=dststride;
 335     }
 336     return;
 337 }
 338