libvo/osd.c

   1 /*
   2  * generic alpha renderers for all YUV modes and RGB depths
   3  * These are "reference implementations", should be optimized later (MMX, etc).
   4  * templating code by Michael Niedermayer (michaelni@gmx.at)
   5  *
   6  * This file is part of MPlayer.
   7  *
   8  * MPlayer is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2 of the License, or
  11  * (at your option) any later version.
  12  *
  13  * MPlayer is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public License along
  19  * with MPlayer; if not, write to the Free Software Foundation, Inc.,
  20  * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
  21  */
  22
  23 #include "config.h"
  24 #include "osd.h"
  25 #include "mp_msg.h"
  26 #include <inttypes.h>
  27 #include "cpudetect.h"
  28
  29 #if ARCH_X86
  30 static const uint64_t bFF __attribute__((aligned(8))) = 0xFFFFFFFFFFFFFFFFULL;
  31 static const unsigned long long mask24lh  __attribute__((aligned(8))) = 0xFFFF000000000000ULL;
  32 static const unsigned long long mask24hl  __attribute__((aligned(8))) = 0x0000FFFFFFFFFFFFULL;
  33 #endif
  34
  35 //Note: we have C, X86-nommx, MMX, MMX2, 3DNOW version therse no 3DNOW+MMX2 one
  36 //Plain C versions
  37 #if !HAVE_MMX || CONFIG_RUNTIME_CPUDETECT
  38 #define COMPILE_C
  39 #endif
  40
  41 #if ARCH_X86
  42
  43 #if (HAVE_MMX && !HAVE_AMD3DNOW && !HAVE_MMX2) || CONFIG_RUNTIME_CPUDETECT
  44 #define COMPILE_MMX
  45 #endif
  46
  47 #if HAVE_MMX2 || CONFIG_RUNTIME_CPUDETECT
  48 #define COMPILE_MMX2
  49 #endif
  50
  51 #if (HAVE_AMD3DNOW && !HAVE_MMX2) || CONFIG_RUNTIME_CPUDETECT
  52 #define COMPILE_3DNOW
  53 #endif
  54
  55 #endif /* ARCH_X86 */
  56
  57 #undef HAVE_MMX
  58 #undef HAVE_MMX2
  59 #undef HAVE_AMD3DNOW
  60 #define HAVE_MMX 0
  61 #define HAVE_MMX2 0
  62 #define HAVE_AMD3DNOW 0
  63
  64 #if ! ARCH_X86
  65
  66 #ifdef COMPILE_C
  67 #undef HAVE_MMX
  68 #undef HAVE_MMX2
  69 #undef HAVE_AMD3DNOW
  70 #define HAVE_MMX 0
  71 #define HAVE_MMX2 0
  72 #define HAVE_AMD3DNOW 0
  73 #define RENAME(a) a ## _C
  74 #include "osd_template.c"
  75 #endif
  76
  77 #else
  78
  79 //X86 noMMX versions
  80 #ifdef COMPILE_C
  81 #undef RENAME
  82 #undef HAVE_MMX
  83 #undef HAVE_MMX2
  84 #undef HAVE_AMD3DNOW
  85 #define HAVE_MMX 0
  86 #define HAVE_MMX2 0
  87 #define HAVE_AMD3DNOW 0
  88 #define RENAME(a) a ## _X86
  89 #include "osd_template.c"
  90 #endif
  91
  92 //MMX versions
  93 #ifdef COMPILE_MMX
  94 #undef RENAME
  95 #undef HAVE_MMX
  96 #undef HAVE_MMX2
  97 #undef HAVE_AMD3DNOW
  98 #define HAVE_MMX 1
  99 #define HAVE_MMX2 0
 100 #define HAVE_AMD3DNOW 0
 101 #define RENAME(a) a ## _MMX
 102 #include "osd_template.c"
 103 #endif
 104
 105 //MMX2 versions
 106 #ifdef COMPILE_MMX2
 107 #undef RENAME
 108 #undef HAVE_MMX
 109 #undef HAVE_MMX2
 110 #undef HAVE_AMD3DNOW
 111 #define HAVE_MMX 1
 112 #define HAVE_MMX2 1
 113 #define HAVE_AMD3DNOW 0
 114 #define RENAME(a) a ## _MMX2
 115 #include "osd_template.c"
 116 #endif
 117
 118 //3DNOW versions
 119 #ifdef COMPILE_3DNOW
 120 #undef RENAME
 121 #undef HAVE_MMX
 122 #undef HAVE_MMX2
 123 #undef HAVE_AMD3DNOW
 124 #define HAVE_MMX 1
 125 #define HAVE_MMX2 0
 126 #define HAVE_AMD3DNOW 1
 127 #define RENAME(a) a ## _3DNow
 128 #include "osd_template.c"
 129 #endif
 130
 131 #endif /* ARCH_X86 */
 132
 133 void vo_draw_alpha_yv12(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 134 #if CONFIG_RUNTIME_CPUDETECT
 135 #if ARCH_X86
 136         // ordered by speed / fastest first
 137         if(gCpuCaps.hasMMX2)
 138                 vo_draw_alpha_yv12_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 139         else if(gCpuCaps.has3DNow)
 140                 vo_draw_alpha_yv12_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 141         else if(gCpuCaps.hasMMX)
 142                 vo_draw_alpha_yv12_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 143         else
 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 #else //CONFIG_RUNTIME_CPUDETECT
 149 #if HAVE_MMX2
 150                 vo_draw_alpha_yv12_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 151 #elif HAVE_AMD3DNOW
 152                 vo_draw_alpha_yv12_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 153 #elif HAVE_MMX
 154                 vo_draw_alpha_yv12_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 155 #elif ARCH_X86
 156                 vo_draw_alpha_yv12_X86(w, h, src, srca, srcstride, dstbase, dststride);
 157 #else
 158                 vo_draw_alpha_yv12_C(w, h, src, srca, srcstride, dstbase, dststride);
 159 #endif
 160 #endif //!CONFIG_RUNTIME_CPUDETECT
 161 }
 162
 163 void vo_draw_alpha_yuy2(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 164 #if CONFIG_RUNTIME_CPUDETECT
 165 #if ARCH_X86
 166         // ordered by speed / fastest first
 167         if(gCpuCaps.hasMMX2)
 168                 vo_draw_alpha_yuy2_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 169         else if(gCpuCaps.has3DNow)
 170                 vo_draw_alpha_yuy2_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 171         else if(gCpuCaps.hasMMX)
 172                 vo_draw_alpha_yuy2_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 173         else
 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 #else //CONFIG_RUNTIME_CPUDETECT
 179 #if HAVE_MMX2
 180                 vo_draw_alpha_yuy2_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 181 #elif HAVE_AMD3DNOW
 182                 vo_draw_alpha_yuy2_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 183 #elif HAVE_MMX
 184                 vo_draw_alpha_yuy2_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 185 #elif ARCH_X86
 186                 vo_draw_alpha_yuy2_X86(w, h, src, srca, srcstride, dstbase, dststride);
 187 #else
 188                 vo_draw_alpha_yuy2_C(w, h, src, srca, srcstride, dstbase, dststride);
 189 #endif
 190 #endif //!CONFIG_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 #if CONFIG_RUNTIME_CPUDETECT
 195 #if ARCH_X86
 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 //CONFIG_RUNTIME_CPUDETECT
 209 #if HAVE_MMX2
 210                 vo_draw_alpha_rgb24_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 211 #elif HAVE_AMD3DNOW
 212                 vo_draw_alpha_rgb24_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 213 #elif HAVE_MMX
 214                 vo_draw_alpha_rgb24_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 215 #elif 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 //!CONFIG_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 #if CONFIG_RUNTIME_CPUDETECT
 225 #if ARCH_X86
 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 //CONFIG_RUNTIME_CPUDETECT
 239 #if HAVE_MMX2
 240                 vo_draw_alpha_rgb32_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
 241 #elif HAVE_AMD3DNOW
 242                 vo_draw_alpha_rgb32_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
 243 #elif HAVE_MMX
 244                 vo_draw_alpha_rgb32_MMX(w, h, src, srca, srcstride, dstbase, dststride);
 245 #elif 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 //!CONFIG_RUNTIME_CPUDETECT
 251 }
 252
 253 void vo_draw_alpha_init(void){
 254 //FIXME the optimized stuff is a lie for 15/16bpp as they aren't optimized yet
 255         if( mp_msg_test(MSGT_OSD,MSGL_V) )
 256         {
 257 #if CONFIG_RUNTIME_CPUDETECT
 258 #if ARCH_X86
 259                 // ordered per speed fasterst first
 260                 if(gCpuCaps.hasMMX2)
 261                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX (with tiny bit MMX2) Optimized OnScreenDisplay\n");
 262                 else if(gCpuCaps.has3DNow)
 263                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX (with tiny bit 3DNow) Optimized OnScreenDisplay\n");
 264                 else if(gCpuCaps.hasMMX)
 265                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX Optimized OnScreenDisplay\n");
 266                 else
 267                         mp_msg(MSGT_OSD,MSGL_INFO,"Using X86 Optimized OnScreenDisplay\n");
 268 #else
 269                         mp_msg(MSGT_OSD,MSGL_INFO,"Using Unoptimized OnScreenDisplay\n");
 270 #endif
 271 #else //CONFIG_RUNTIME_CPUDETECT
 272 #if HAVE_MMX2
 273                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX (with tiny bit MMX2) Optimized OnScreenDisplay\n");
 274 #elif HAVE_AMD3DNOW
 275                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX (with tiny bit 3DNow) Optimized OnScreenDisplay\n");
 276 #elif HAVE_MMX
 277                         mp_msg(MSGT_OSD,MSGL_INFO,"Using MMX Optimized OnScreenDisplay\n");
 278 #elif ARCH_X86
 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 #endif //!CONFIG_RUNTIME_CPUDETECT
 284         }
 285 }
 286
 287 void vo_draw_alpha_rgb12(int w, int h, unsigned char* src, unsigned char *srca,
 288                          int srcstride, unsigned char* dstbase, int dststride) {
 289     int y;
 290     for (y = 0; y < h; y++) {
 291         register unsigned short *dst = (unsigned short*) dstbase;
 292         register int x;
 293         for (x = 0; x < w; x++) {
 294             if(srca[x]){
 295                 unsigned char r = dst[x] & 0x0F;
 296                 unsigned char g = (dst[x] >> 4) & 0x0F;
 297                 unsigned char b = (dst[x] >> 8) & 0x0F;
 298                 r = (((r*srca[x]) >> 4) + src[x]) >> 4;
 299                 g = (((g*srca[x]) >> 4) + src[x]) >> 4;
 300                 b = (((b*srca[x]) >> 4) + src[x]) >> 4;
 301                 dst[x] = (b << 8) | (g << 4) | r;
 302             }
 303         }
 304         src += srcstride;
 305         srca += srcstride;
 306         dstbase += dststride;
 307     }
 308     return;
 309 }
 310
 311 void vo_draw_alpha_rgb15(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 312     int y;
 313     for(y=0;y<h;y++){
 314         register unsigned short *dst = (unsigned short*) dstbase;
 315         register int x;
 316         for(x=0;x<w;x++){
 317             if(srca[x]){
 318                 unsigned char r=dst[x]&0x1F;
 319                 unsigned char g=(dst[x]>>5)&0x1F;
 320                 unsigned char b=(dst[x]>>10)&0x1F;
 321                 r=(((r*srca[x])>>5)+src[x])>>3;
 322                 g=(((g*srca[x])>>5)+src[x])>>3;
 323                 b=(((b*srca[x])>>5)+src[x])>>3;
 324                 dst[x]=(b<<10)|(g<<5)|r;
 325             }
 326         }
 327         src+=srcstride;
 328         srca+=srcstride;
 329         dstbase+=dststride;
 330     }
 331     return;
 332 }
 333
 334 void vo_draw_alpha_rgb16(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
 335     int y;
 336     for(y=0;y<h;y++){
 337         register unsigned short *dst = (unsigned short*) dstbase;
 338         register int x;
 339         for(x=0;x<w;x++){
 340             if(srca[x]){
 341                 unsigned char r=dst[x]&0x1F;
 342                 unsigned char g=(dst[x]>>5)&0x3F;
 343                 unsigned char b=(dst[x]>>11)&0x1F;
 344                 r=(((r*srca[x])>>5)+src[x])>>3;
 345                 g=(((g*srca[x])>>6)+src[x])>>2;
 346                 b=(((b*srca[x])>>5)+src[x])>>3;
 347                 dst[x]=(b<<11)|(g<<5)|r;
 348             }
 349         }
 350         src+=srcstride;
 351         srca+=srcstride;
 352         dstbase+=dststride;
 353     }
 354     return;
 355 }