src/mesa/drivers/dri/common/spantmp2.h

   1 /*
   2  * Copyright 2000-2001 VA Linux Systems, Inc.
   3  * (C) Copyright IBM Corporation 2004
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * on the rights to use, copy, modify, merge, publish, distribute, sub
  10  * license, and/or sell copies of the Software, and to permit persons to whom
  11  * the Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice (including the next
  14  * paragraph) shall be included in all copies or substantial portions of the
  15  * Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.  IN NO EVENT SHALL
  20  * VA LINUX SYSTEM, IBM AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  21  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  22  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  23  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  24  */
  25
  26 /**
  27  * \file spantmp2.h
  28  *
  29  * Template file of span read / write functions.
  30  *
  31  * \author Keith Whitwell <keithw@tungstengraphics.com>
  32  * \author Gareth Hughes <gareth@nvidia.com>
  33  * \author Ian Romanick <idr@us.ibm.com>
  34  */
  35
  36 #include "colormac.h"
  37
  38 #ifndef DBG
  39 #define DBG 0
  40 #endif
  41
  42 #ifndef HW_WRITE_LOCK
  43 #define HW_WRITE_LOCK()         HW_LOCK()
  44 #endif
  45
  46 #ifndef HW_WRITE_UNLOCK
  47 #define HW_WRITE_UNLOCK()       HW_UNLOCK()
  48 #endif
  49
  50 #ifndef HW_READ_LOCK
  51 #define HW_READ_LOCK()          HW_LOCK()
  52 #endif
  53
  54 #ifndef HW_READ_UNLOCK
  55 #define HW_READ_UNLOCK()        HW_UNLOCK()
  56 #endif
  57
  58 #ifndef HW_READ_CLIPLOOP
  59 #define HW_READ_CLIPLOOP()      HW_CLIPLOOP()
  60 #endif
  61
  62 #ifndef HW_WRITE_CLIPLOOP
  63 #define HW_WRITE_CLIPLOOP()     HW_CLIPLOOP()
  64 #endif
  65
  66 #if (SPANTMP_PIXEL_FMT == GL_RGB)  && (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_SHORT_5_6_5)
  67
  68 #define INIT_MONO_PIXEL(p, color) \
  69   p = PACK_COLOR_565( color[0], color[1], color[2] )
  70
  71 #define WRITE_RGBA( _x, _y, r, g, b, a )                                \
  72     do {                                                                \
  73        GLshort * _p = (GLshort *) GET_DST_PTR(_x, _y);                  \
  74        _p[0] = ((((int)r & 0xf8) << 8) | (((int)g & 0xfc) << 3) |       \
  75                    (((int)b & 0xf8) >> 3));                             \
  76    } while(0)
  77
  78 #define WRITE_PIXEL( _x, _y, p )                                        \
  79    do {                                                                 \
  80       GLushort * _p = (GLushort *) GET_DST_PTR(_x, _y);                 \
  81       _p[0] = p;                                                        \
  82    } while(0)
  83
  84 #define READ_RGBA( rgba, _x, _y )                                       \
  85    do {                                                                 \
  86       GLushort p = *(volatile GLshort *) GET_SRC_PTR(_x, _y);           \
  87       rgba[0] = ((p >> 8) & 0xf8) * 255 / 0xf8;                         \
  88       rgba[1] = ((p >> 3) & 0xfc) * 255 / 0xfc;                         \
  89       rgba[2] = ((p << 3) & 0xf8) * 255 / 0xf8;                         \
  90       rgba[3] = 0xff;                                                   \
  91    } while (0)
  92
  93 #elif (SPANTMP_PIXEL_FMT == GL_BGRA) && (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_INT_8_8_8_8_REV)
  94
  95 # define INIT_MONO_PIXEL(p, color)                       \
  96      p = PACK_COLOR_8888(color[3], color[0], color[1], color[2])
  97
  98 # define WRITE_RGBA(_x, _y, r, g, b, a)                                 \
  99     do {                                                                \
 100        GLuint * _p = (GLuint *) GET_DST_PTR(_x, _y);                    \
 101        _p[0] = ((r << 16) | (g << 8) | (b << 0) | (a << 24));           \
 102     } while(0)
 103
 104 #define WRITE_PIXEL(_x, _y, p)                                          \
 105     do {                                                                \
 106        GLuint * _p = (GLuint *) GET_DST_PTR(_x, _y);                    \
 107        _p[0] = p;                                                       \
 108     } while(0)
 109
 110 # if defined( USE_X86_ASM )
 111 #  define READ_RGBA(rgba, _x, _y)                                       \
 112     do {                                                                \
 113         GLuint p = *(volatile GLuint *) GET_SRC_PTR(_x, _y);            \
 114        __asm__ __volatile__( "bswap     %0; rorl $8, %0"                \
 115                                 : "=r" (p) : "r" (p) );                 \
 116        ((GLuint *)rgba)[0] = p;                                         \
 117     } while (0)
 118 # elif defined( MESA_BIG_ENDIAN )
 119     /* On PowerPC with GCC 3.4.2 the shift madness below becomes a single
 120      * rotlwi instruction.  It also produces good code on SPARC.
 121      */
 122 #  define READ_RGBA( rgba, _x, _y )                                     \
 123      do {                                                               \
 124         GLuint p = *(volatile GLuint *) GET_SRC_PTR(_x, _y);            \
 125         GLuint t = p;                                                   \
 126         *((uint32_t *) rgba) = (t >> 24) | (p << 8);                    \
 127      } while (0)
 128 # else
 129 #  define READ_RGBA( rgba, _x, _y )                                     \
 130      do {                                                               \
 131         GLuint p = *(volatile GLuint *) GET_SRC_PTR(_x, _y);            \
 132         rgba[0] = (p >> 16) & 0xff;                                     \
 133         rgba[1] = (p >>  8) & 0xff;                                     \
 134         rgba[2] = (p >>  0) & 0xff;                                     \
 135         rgba[3] = (p >> 24) & 0xff;                                     \
 136      } while (0)
 137 # endif
 138
 139 #else
 140 #error SPANTMP_PIXEL_FMT must be set to a valid value!
 141 #endif
 142
 143 #if defined( USE_MMX_ASM ) || defined( USE_SSE_ASM )
 144 #include "x86/read_rgba_span_x86.h"
 145 #include "x86/common_x86_asm.h"
 146 #endif
 147
 148 static void TAG(WriteRGBASpan)( const GLcontext *ctx,
 149                                 GLuint n, GLint x, GLint y,
 150                                 const GLubyte rgba[][4],
 151                                 const GLubyte mask[] )
 152 {
 153    HW_WRITE_LOCK()
 154       {
 155          GLint x1;
 156          GLint n1;
 157          LOCAL_VARS;
 158
 159          y = Y_FLIP(y);
 160
 161          HW_WRITE_CLIPLOOP()
 162             {
 163                GLint i = 0;
 164                CLIPSPAN(x,y,n,x1,n1,i);
 165
 166                if (DBG) fprintf(stderr, "WriteRGBASpan %d..%d (x1 %d)\n",
 167                                 (int)i, (int)n1, (int)x1);
 168
 169                if (mask)
 170                {
 171                   for (;n1>0;i++,x1++,n1--)
 172                      if (mask[i])
 173                         WRITE_RGBA( x1, y,
 174                                     rgba[i][0], rgba[i][1],
 175                                     rgba[i][2], rgba[i][3] );
 176                }
 177                else
 178                {
 179                   for (;n1>0;i++,x1++,n1--)
 180                      WRITE_RGBA( x1, y,
 181                                  rgba[i][0], rgba[i][1],
 182                                  rgba[i][2], rgba[i][3] );
 183                }
 184             }
 185          HW_ENDCLIPLOOP();
 186       }
 187    HW_WRITE_UNLOCK();
 188 }
 189
 190 static void TAG(WriteRGBSpan)( const GLcontext *ctx,
 191                                GLuint n, GLint x, GLint y,
 192                                const GLubyte rgb[][3],
 193                                const GLubyte mask[] )
 194 {
 195    HW_WRITE_LOCK()
 196       {
 197          GLint x1;
 198          GLint n1;
 199          LOCAL_VARS;
 200
 201          y = Y_FLIP(y);
 202
 203          HW_WRITE_CLIPLOOP()
 204             {
 205                GLint i = 0;
 206                CLIPSPAN(x,y,n,x1,n1,i);
 207
 208                if (DBG) fprintf(stderr, "WriteRGBSpan %d..%d (x1 %d)\n",
 209                                 (int)i, (int)n1, (int)x1);
 210
 211                if (mask)
 212                {
 213                   for (;n1>0;i++,x1++,n1--)
 214                      if (mask[i])
 215                         WRITE_RGBA( x1, y, rgb[i][0], rgb[i][1], rgb[i][2], 255 );
 216                }
 217                else
 218                {
 219                   for (;n1>0;i++,x1++,n1--)
 220                      WRITE_RGBA( x1, y, rgb[i][0], rgb[i][1], rgb[i][2], 255 );
 221                }
 222             }
 223          HW_ENDCLIPLOOP();
 224       }
 225    HW_WRITE_UNLOCK();
 226 }
 227
 228 static void TAG(WriteRGBAPixels)( const GLcontext *ctx,
 229                                GLuint n,
 230                                const GLint x[],
 231                                const GLint y[],
 232                                const GLubyte rgba[][4],
 233                                const GLubyte mask[] )
 234 {
 235    HW_WRITE_LOCK()
 236       {
 237          GLint i;
 238          LOCAL_VARS;
 239
 240          if (DBG) fprintf(stderr, "WriteRGBAPixels\n");
 241
 242          HW_WRITE_CLIPLOOP()
 243             {
 244                if (mask)
 245                {
 246                   for (i=0;i<n;i++)
 247                   {
 248                      if (mask[i]) {
 249                         const int fy = Y_FLIP(y[i]);
 250                         if (CLIPPIXEL(x[i],fy))
 251                            WRITE_RGBA( x[i], fy,
 252                                        rgba[i][0], rgba[i][1],
 253                                        rgba[i][2], rgba[i][3] );
 254                      }
 255                   }
 256                }
 257                else
 258                {
 259                   for (i=0;i<n;i++)
 260                   {
 261                      const int fy = Y_FLIP(y[i]);
 262                      if (CLIPPIXEL(x[i],fy))
 263                         WRITE_RGBA( x[i], fy,
 264                                     rgba[i][0], rgba[i][1],
 265                                     rgba[i][2], rgba[i][3] );
 266                   }
 267                }
 268             }
 269          HW_ENDCLIPLOOP();
 270       }
 271    HW_WRITE_UNLOCK();
 272 }
 273
 274
 275 static void TAG(WriteMonoRGBASpan)( const GLcontext *ctx,
 276                                     GLuint n, GLint x, GLint y,
 277                                     const GLchan color[4],
 278                                     const GLubyte mask[] )
 279 {
 280    HW_WRITE_LOCK()
 281       {
 282          GLint x1;
 283          GLint n1;
 284          LOCAL_VARS;
 285          INIT_MONO_PIXEL(p, color);
 286
 287          y = Y_FLIP( y );
 288
 289          if (DBG) fprintf(stderr, "WriteMonoRGBASpan\n");
 290
 291          HW_WRITE_CLIPLOOP()
 292             {
 293                GLint i = 0;
 294                CLIPSPAN(x,y,n,x1,n1,i);
 295                if (mask)
 296                {
 297                   for (;n1>0;i++,x1++,n1--)
 298                      if (mask[i])
 299                         WRITE_PIXEL( x1, y, p );
 300                }
 301                else
 302                {
 303                   for (;n1>0;i++,x1++,n1--)
 304                      WRITE_PIXEL( x1, y, p );
 305                }
 306             }
 307          HW_ENDCLIPLOOP();
 308       }
 309    HW_WRITE_UNLOCK();
 310 }
 311
 312
 313 static void TAG(WriteMonoRGBAPixels)( const GLcontext *ctx,
 314                                       GLuint n,
 315                                       const GLint x[], const GLint y[],
 316                                       const GLchan color[],
 317                                       const GLubyte mask[] )
 318 {
 319    HW_WRITE_LOCK()
 320       {
 321          GLint i;
 322          LOCAL_VARS;
 323          INIT_MONO_PIXEL(p, color);
 324
 325          if (DBG) fprintf(stderr, "WriteMonoRGBAPixels\n");
 326
 327          HW_WRITE_CLIPLOOP()
 328             {
 329                if (mask)
 330                {
 331                   for (i=0;i<n;i++)
 332                      if (mask[i]) {
 333                         int fy = Y_FLIP(y[i]);
 334                         if (CLIPPIXEL( x[i], fy ))
 335                            WRITE_PIXEL( x[i], fy, p );
 336                      }
 337                }
 338                else
 339                {
 340                   for (i=0;i<n;i++) {
 341                      int fy = Y_FLIP(y[i]);
 342                      if (CLIPPIXEL( x[i], fy ))
 343                         WRITE_PIXEL( x[i], fy, p );
 344                   }
 345                }
 346             }
 347          HW_ENDCLIPLOOP();
 348       }
 349    HW_WRITE_UNLOCK();
 350 }
 351
 352
 353 static void TAG(ReadRGBASpan)( const GLcontext *ctx,
 354                                GLuint n, GLint x, GLint y,
 355                                GLubyte rgba[][4])
 356 {
 357    HW_READ_LOCK()
 358       {
 359          GLint x1,n1;
 360          LOCAL_VARS;
 361
 362          y = Y_FLIP(y);
 363
 364          if (DBG) fprintf(stderr, "ReadRGBASpan\n");
 365
 366          HW_READ_CLIPLOOP()
 367             {
 368                GLint i = 0;
 369                CLIPSPAN(x,y,n,x1,n1,i);
 370                for (;n1>0;i++,x1++,n1--)
 371                   READ_RGBA( rgba[i], x1, y );
 372             }
 373          HW_ENDCLIPLOOP();
 374       }
 375    HW_READ_UNLOCK();
 376 }
 377
 378
 379 #if defined(USE_MMX_ASM) && \
 380    (SPANTMP_PIXEL_FMT == GL_BGRA) && \
 381      (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_INT_8_8_8_8_REV)
 382 static void TAG2(ReadRGBASpan,_MMX)( const GLcontext *ctx,
 383                                GLuint n, GLint x, GLint y,
 384                                GLubyte rgba[][4])
 385 {
 386 #ifndef USE_INNER_EMMS
 387    /* The EMMS instruction is directly in-lined here because using GCC's
 388     * built-in _mm_empty function was found to utterly destroy performance.
 389     */
 390    __asm__ __volatile__( "emms" );
 391 #endif
 392
 393    HW_LOCK()
 394      {
 395         GLint x1,n1;
 396         LOCAL_VARS;
 397
 398         y = Y_FLIP(y);
 399
 400         if (DBG) fprintf(stderr, "ReadRGBASpan\n");
 401
 402         HW_READ_CLIPLOOP()
 403           {
 404              GLint i = 0;
 405              CLIPSPAN(x,y,n,x1,n1,i);
 406
 407                {
 408                   const char * src = GET_SRC_PTR( x1, y );
 409                   _generic_read_RGBA_span_BGRA8888_REV_MMX( src, rgba[i], n1 );
 410                }
 411           }
 412         HW_ENDCLIPLOOP();
 413      }
 414    HW_UNLOCK();
 415 #ifndef USE_INNER_EMMS
 416    __asm__ __volatile__( "emms" );
 417 #endif
 418 }
 419 #endif
 420
 421
 422 #if defined(USE_SSE_ASM) && \
 423    (SPANTMP_PIXEL_FMT == GL_BGRA) && \
 424      (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_INT_8_8_8_8_REV)
 425 static void TAG2(ReadRGBASpan,_SSE2)( const GLcontext *ctx,
 426                                GLuint n, GLint x, GLint y,
 427                                GLubyte rgba[][4])
 428 {
 429    HW_LOCK()
 430      {
 431         GLint x1,n1;
 432         LOCAL_VARS;
 433
 434         y = Y_FLIP(y);
 435
 436         if (DBG) fprintf(stderr, "ReadRGBASpan\n");
 437
 438         HW_READ_CLIPLOOP()
 439           {
 440              GLint i = 0;
 441              CLIPSPAN(x,y,n,x1,n1,i);
 442
 443                {
 444                   const char * src = GET_SRC_PTR( x1, y );
 445                   _generic_read_RGBA_span_BGRA8888_REV_SSE2( src, rgba[i], n1 );
 446                }
 447           }
 448         HW_ENDCLIPLOOP();
 449      }
 450    HW_UNLOCK();
 451 }
 452 #endif
 453
 454 #if defined(USE_SSE_ASM) && \
 455    (SPANTMP_PIXEL_FMT == GL_BGRA) && \
 456      (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_INT_8_8_8_8_REV)
 457 static void TAG2(ReadRGBASpan,_SSE)( const GLcontext *ctx,
 458                                GLuint n, GLint x, GLint y,
 459                                GLubyte rgba[][4])
 460 {
 461 #ifndef USE_INNER_EMMS
 462    /* The EMMS instruction is directly in-lined here because using GCC's
 463     * built-in _mm_empty function was found to utterly destroy performance.
 464     */
 465    __asm__ __volatile__( "emms" );
 466 #endif
 467
 468    HW_LOCK()
 469      {
 470         GLint x1,n1;
 471         LOCAL_VARS;
 472
 473         y = Y_FLIP(y);
 474
 475         if (DBG) fprintf(stderr, "ReadRGBASpan\n");
 476
 477         HW_READ_CLIPLOOP()
 478           {
 479              GLint i = 0;
 480              CLIPSPAN(x,y,n,x1,n1,i);
 481
 482                {
 483                   const char * src = GET_SRC_PTR( x1, y );
 484                   _generic_read_RGBA_span_BGRA8888_REV_SSE( src, rgba[i], n1 );
 485                }
 486           }
 487         HW_ENDCLIPLOOP();
 488      }
 489    HW_UNLOCK();
 490 #ifndef USE_INNER_EMMS
 491    __asm__ __volatile__( "emms" );
 492 #endif
 493 }
 494 #endif
 495
 496
 497 static void TAG(ReadRGBAPixels)( const GLcontext *ctx,
 498                                  GLuint n, const GLint x[], const GLint y[],
 499                                  GLubyte rgba[][4], const GLubyte mask[] )
 500 {
 501    HW_READ_LOCK()
 502       {
 503          GLint i;
 504          LOCAL_VARS;
 505
 506          if (DBG) fprintf(stderr, "ReadRGBAPixels\n");
 507
 508          HW_READ_CLIPLOOP()
 509             {
 510                if (mask)
 511                {
 512                   for (i=0;i<n;i++)
 513                      if (mask[i]) {
 514                         int fy = Y_FLIP( y[i] );
 515                         if (CLIPPIXEL( x[i], fy ))
 516                            READ_RGBA( rgba[i], x[i], fy );
 517                      }
 518                }
 519                else
 520                {
 521                   for (i=0;i<n;i++) {
 522                      int fy = Y_FLIP( y[i] );
 523                      if (CLIPPIXEL( x[i], fy ))
 524                         READ_RGBA( rgba[i], x[i], fy );
 525                   }
 526                }
 527             }
 528          HW_ENDCLIPLOOP();
 529       }
 530    HW_READ_UNLOCK();
 531 }
 532
 533 static void TAG(InitPointers)(struct swrast_device_driver *swdd)
 534 {
 535    swdd->WriteRGBASpan = TAG(WriteRGBASpan);
 536    swdd->WriteRGBSpan = TAG(WriteRGBSpan);
 537    swdd->WriteMonoRGBASpan = TAG(WriteMonoRGBASpan);
 538    swdd->WriteRGBAPixels = TAG(WriteRGBAPixels);
 539    swdd->WriteMonoRGBAPixels = TAG(WriteMonoRGBAPixels);
 540    swdd->ReadRGBAPixels = TAG(ReadRGBAPixels);
 541
 542 #if (SPANTMP_PIXEL_FMT == GL_BGRA) && \
 543      (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_INT_8_8_8_8_REV)
 544 #if defined(USE_SSE_ASM)
 545    if ( cpu_has_xmm2 ) {
 546       if (DBG) fprintf( stderr, "Using %s version of ReadRGBASpan\n", "SSE2" );
 547       swdd->ReadRGBASpan = TAG2(ReadRGBASpan, _SSE2);
 548    }
 549    else
 550 #endif
 551 #if defined(USE_SSE_ASM)
 552    if ( cpu_has_xmm ) {
 553       if (DBG) fprintf( stderr, "Using %s version of ReadRGBASpan\n", "SSE" );
 554       swdd->ReadRGBASpan = TAG2(ReadRGBASpan, _SSE);
 555    }
 556    else
 557 #endif
 558 #if defined(USE_MMX_ASM)
 559    if ( cpu_has_mmx ) {
 560       if (DBG) fprintf( stderr, "Using %s version of ReadRGBASpan\n", "MMX" );
 561       swdd->ReadRGBASpan = TAG2(ReadRGBASpan, _MMX);
 562    }
 563    else
 564 #endif
 565 #endif
 566    {
 567       if (DBG) fprintf( stderr, "Using %s version of ReadRGBASpan\n", "C" );
 568       swdd->ReadRGBASpan = TAG(ReadRGBASpan);
 569    }
 570
 571 }
 572
 573
 574 #undef INIT_MONO_PIXEL
 575 #undef WRITE_PIXEL
 576 #undef WRITE_RGBA
 577 #undef READ_RGBA
 578 #undef TAG
 579 #undef TAG2
 580 #undef GET_SRC_PTR
 581 #undef GET_DST_PTR
 582 #undef SPANTMP_PIXEL_FMT
 583 #undef SPANTMP_PIXEL_TYPE