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 "main/colormac.h"
  37 #include "spantmp_common.h"
  38
  39 #ifndef DBG
  40 #define DBG 0
  41 #endif
  42
  43 #ifndef HW_READ_CLIPLOOP
  44 #define HW_READ_CLIPLOOP()      HW_CLIPLOOP()
  45 #endif
  46
  47 #ifndef HW_WRITE_CLIPLOOP
  48 #define HW_WRITE_CLIPLOOP()     HW_CLIPLOOP()
  49 #endif
  50
  51 #if (SPANTMP_PIXEL_FMT == GL_RGB)  && (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_SHORT_5_6_5)
  52
  53 /**
  54  ** GL_RGB, GL_UNSIGNED_SHORT_5_6_5
  55  **/
  56
  57 #ifndef GET_VALUE
  58 #ifndef GET_PTR
  59 #define GET_PTR(_x, _y) (buf + (_x) * 2 + (_y) * pitch)
  60 #endif
  61
  62 #define GET_VALUE(_x, _y) *(volatile GLushort *)(GET_PTR(_x, _y))
  63 #define PUT_VALUE(_x, _y, _v) *(volatile GLushort *)(GET_PTR(_x, _y)) = (_v)
  64 #endif /* GET_VALUE */
  65
  66 #define INIT_MONO_PIXEL(p, color) \
  67   p = PACK_COLOR_565( color[0], color[1], color[2] )
  68
  69 #define WRITE_RGBA( _x, _y, r, g, b, a )                                \
  70    PUT_VALUE(_x, _y, ((((int)r & 0xf8) << 8) |                          \
  71                       (((int)g & 0xfc) << 3) |                          \
  72                       (((int)b & 0xf8) >> 3)))                          \
  73
  74 #define WRITE_PIXEL( _x, _y, p ) PUT_VALUE(_x, _y, p)
  75
  76 #define READ_RGBA( rgba, _x, _y )                                       \
  77    do {                                                                 \
  78       GLushort p = GET_VALUE(_x, _y);                                   \
  79       rgba[0] = ((p >> 8) & 0xf8) * 255 / 0xf8;                         \
  80       rgba[1] = ((p >> 3) & 0xfc) * 255 / 0xfc;                         \
  81       rgba[2] = ((p << 3) & 0xf8) * 255 / 0xf8;                         \
  82       rgba[3] = 0xff;                                                   \
  83    } while (0)
  84
  85 #elif (SPANTMP_PIXEL_FMT == GL_RGB)  && (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_SHORT_5_6_5_REV)
  86
  87 /**
  88  ** GL_RGB, GL_UNSIGNED_SHORT_5_6_5_REV
  89  **/
  90
  91 #ifndef GET_VALUE
  92 #ifndef GET_PTR
  93 #define GET_PTR(_x, _y) (buf + (_x) * 2 + (_y) * pitch)
  94 #endif
  95
  96 #define GET_VALUE(_x, _y) *(volatile GLushort *)(GET_PTR(_x, _y))
  97 #define PUT_VALUE(_x, _y, _v) *(volatile GLushort *)(GET_PTR(_x, _y)) = (_v)
  98 #endif /* GET_VALUE */
  99
 100 #define INIT_MONO_PIXEL(p, color) \
 101   p = PACK_COLOR_565_REV( color[0], color[1], color[2] )
 102
 103 #define WRITE_RGBA( _x, _y, r, g, b, a )                                \
 104    PUT_VALUE(_x, _y, PACK_COLOR_565_REV( r, g, b ))
 105
 106 #define WRITE_PIXEL( _x, _y, p ) PUT_VALUE(_x, _y, p)
 107
 108 #define READ_RGBA( rgba, _x, _y )                                       \
 109    do {                                                                 \
 110       GLushort p = GET_VALUE(_x, _y);                                   \
 111       p = p << 8 | p >> 8;                                              \
 112       rgba[0] = ((p >> 8) & 0xf8) * 255 / 0xf8;                         \
 113       rgba[1] = ((p >> 3) & 0xfc) * 255 / 0xfc;                         \
 114       rgba[2] = ((p << 3) & 0xf8) * 255 / 0xf8;                         \
 115       rgba[3] = 0xff;                                                   \
 116    } while (0)
 117
 118 #elif (SPANTMP_PIXEL_FMT == GL_BGRA)  && (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_SHORT_4_4_4_4)
 119
 120 /**
 121  ** GL_BGRA, GL_UNSIGNED_SHORT_4_4_4_4
 122  **/
 123
 124 #ifndef GET_VALUE
 125 #ifndef GET_PTR
 126 #define GET_PTR(_x, _y) (buf + (_x) * 2 + (_y) * pitch)
 127 #endif
 128
 129 #define GET_VALUE(_x, _y) *(volatile GLushort *)(GET_PTR(_x, _y))
 130 #define PUT_VALUE(_x, _y, _v) *(volatile GLushort *)(GET_PTR(_x, _y)) = (_v)
 131 #endif /* GET_VALUE */
 132
 133 #define INIT_MONO_PIXEL(p, color) \
 134    p = PACK_COLOR_4444_REV(color[3], color[0], color[1], color[2])
 135
 136 #define WRITE_RGBA( _x, _y, r, g, b, a )                                \
 137    PUT_VALUE(_x, _y, PACK_COLOR_4444_REV(a, r, g, b))                   \
 138
 139 #define WRITE_PIXEL( _x, _y, p ) PUT_VALUE(_x, _y, p)
 140
 141 #define READ_RGBA( rgba, _x, _y )                                       \
 142    do {                                                                 \
 143       GLushort p = GET_VALUE(_x, _y);                                   \
 144       rgba[0] = ((p >> 0) & 0xf) * 0x11;                                \
 145       rgba[1] = ((p >> 12) & 0xf) * 0x11;                               \
 146       rgba[2] = ((p >> 4) & 0xf) * 0x11;                                \
 147       rgba[3] = ((p >> 8) & 0xf) * 0x11;                                \
 148    } while (0)
 149
 150
 151 #elif (SPANTMP_PIXEL_FMT == GL_BGRA)  && (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_SHORT_4_4_4_4_REV)
 152
 153 /**
 154  ** GL_BGRA, GL_UNSIGNED_SHORT_4_4_4_4_REV
 155  **/
 156
 157 #ifndef GET_VALUE
 158 #ifndef GET_PTR
 159 #define GET_PTR(_x, _y) (buf + (_x) * 2 + (_y) * pitch)
 160 #endif
 161
 162 #define GET_VALUE(_x, _y) *(volatile GLushort *)(GET_PTR(_x, _y))
 163 #define PUT_VALUE(_x, _y, _v) *(volatile GLushort *)(GET_PTR(_x, _y)) = (_v)
 164 #endif /* GET_VALUE */
 165
 166 #define INIT_MONO_PIXEL(p, color) \
 167    p = PACK_COLOR_4444(color[3], color[0], color[1], color[2])
 168
 169 #define WRITE_RGBA( _x, _y, r, g, b, a )                                \
 170    PUT_VALUE(_x, _y, PACK_COLOR_4444(a, r, g, b))                       \
 171
 172 #define WRITE_PIXEL( _x, _y, p ) PUT_VALUE(_x, _y, p)
 173
 174 #define READ_RGBA( rgba, _x, _y )                                       \
 175    do {                                                                 \
 176       GLushort p = GET_VALUE(_x, _y);                                   \
 177       rgba[0] = ((p >> 8) & 0xf) * 0x11;                                \
 178       rgba[1] = ((p >> 4) & 0xf) * 0x11;                                \
 179       rgba[2] = ((p >> 0) & 0xf) * 0x11;                                \
 180       rgba[3] = ((p >> 12) & 0xf) * 0x11;                               \
 181    } while (0)
 182
 183
 184 #elif (SPANTMP_PIXEL_FMT == GL_BGRA)  && (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_SHORT_1_5_5_5_REV)
 185
 186 /**
 187  ** GL_BGRA, GL_UNSIGNED_SHORT_1_5_5_5_REV
 188  **/
 189
 190 #ifndef GET_VALUE
 191 #ifndef GET_PTR
 192 #define GET_PTR(_x, _y) (buf + (_x) * 2 + (_y) * pitch)
 193 #endif
 194
 195 #define GET_VALUE(_x, _y) *(volatile GLushort *)(GET_PTR(_x, _y))
 196 #define PUT_VALUE(_x, _y, _v) *(volatile GLushort *)(GET_PTR(_x, _y)) = (_v)
 197 #endif /* GET_VALUE */
 198
 199 #define INIT_MONO_PIXEL(p, color) \
 200    p = PACK_COLOR_1555(color[3], color[0], color[1], color[2])
 201
 202 #define WRITE_RGBA( _x, _y, r, g, b, a )                                \
 203    PUT_VALUE(_x, _y, PACK_COLOR_1555(a, r, g, b))                       \
 204
 205 #define WRITE_PIXEL( _x, _y, p ) PUT_VALUE(_x, _y, p)
 206
 207 #define READ_RGBA( rgba, _x, _y )                                       \
 208    do {                                                                 \
 209       GLushort p = GET_VALUE(_x, _y);                                   \
 210       rgba[0] = ((p >> 7) & 0xf8) * 255 / 0xf8;                         \
 211       rgba[1] = ((p >> 2) & 0xf8) * 255 / 0xf8;                         \
 212       rgba[2] = ((p << 3) & 0xf8) * 255 / 0xf8;                         \
 213       rgba[3] = ((p >> 15) & 0x1) * 0xff;                               \
 214    } while (0)
 215
 216 #elif (SPANTMP_PIXEL_FMT == GL_BGRA)  && (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_SHORT_1_5_5_5)
 217
 218 /**
 219  ** GL_BGRA, GL_UNSIGNED_SHORT_1_5_5_5
 220  **/
 221
 222 #ifndef GET_VALUE
 223 #ifndef GET_PTR
 224 #define GET_PTR(_x, _y) (buf + (_x) * 2 + (_y) * pitch)
 225 #endif
 226
 227 #define GET_VALUE(_x, _y) *(volatile GLushort *)(GET_PTR(_x, _y))
 228 #define PUT_VALUE(_x, _y, _v) *(volatile GLushort *)(GET_PTR(_x, _y)) = (_v)
 229 #endif /* GET_VALUE */
 230
 231 #define INIT_MONO_PIXEL(p, color) \
 232    p = PACK_COLOR_1555_REV(color[3], color[0], color[1], color[2])
 233
 234 #define WRITE_RGBA( _x, _y, r, g, b, a )                                \
 235    PUT_VALUE(_x, _y, PACK_COLOR_1555_REV(a, r, g, b))                   \
 236
 237 #define WRITE_PIXEL( _x, _y, p ) PUT_VALUE(_x, _y, p)
 238
 239 #define READ_RGBA( rgba, _x, _y )                                       \
 240    do {                                                                 \
 241       GLushort p = GET_VALUE(_x, _y);                                   \
 242       p = p << 8 | p >> 8;                                              \
 243       rgba[0] = ((p >> 7) & 0xf8) * 255 / 0xf8;                         \
 244       rgba[1] = ((p >> 2) & 0xf8) * 255 / 0xf8;                         \
 245       rgba[2] = ((p << 3) & 0xf8) * 255 / 0xf8;                         \
 246       rgba[3] = ((p >> 15) & 0x1) * 0xff;                               \
 247    } while (0)
 248
 249 #elif (SPANTMP_PIXEL_FMT == GL_BGRA) && (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_INT_8_8_8_8_REV)
 250
 251 /**
 252  ** GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV
 253  **/
 254
 255 #ifndef GET_VALUE
 256 #ifndef GET_PTR
 257 #define GET_PTR(_x, _y) (     buf + (_x) * 4 + (_y) * pitch)
 258 #endif
 259
 260 #define GET_VALUE(_x, _y) *(volatile GLuint *)(GET_PTR(_x, _y))
 261 #define PUT_VALUE(_x, _y, _v) *(volatile GLuint *)(GET_PTR(_x, _y)) = (_v)
 262 #endif /* GET_VALUE */
 263
 264 # define INIT_MONO_PIXEL(p, color)                       \
 265      p = PACK_COLOR_8888(color[3], color[0], color[1], color[2])
 266
 267 # define WRITE_RGBA(_x, _y, r, g, b, a)                                 \
 268    PUT_VALUE(_x, _y, ((r << 16) |                                       \
 269                       (g << 8) |                                        \
 270                       (b << 0) |                                        \
 271                       (a << 24)))
 272
 273 #define WRITE_PIXEL(_x, _y, p) PUT_VALUE(_x, _y, p)
 274
 275 # if defined( USE_X86_ASM )
 276 #  define READ_RGBA(rgba, _x, _y)                                       \
 277     do {                                                                \
 278        GLuint p = GET_VALUE(_x, _y);                                    \
 279        __asm__ __volatile__( "bswap     %0; rorl $8, %0"                \
 280                                 : "=r" (p) : "0" (p) );                 \
 281        ((GLuint *)rgba)[0] = p;                                         \
 282     } while (0)
 283 # elif defined( MESA_BIG_ENDIAN )
 284     /* On PowerPC with GCC 3.4.2 the shift madness below becomes a single
 285      * rotlwi instruction.  It also produces good code on SPARC.
 286      */
 287 #  define READ_RGBA( rgba, _x, _y )                                     \
 288      do {                                                               \
 289         GLuint p = GET_VALUE(_x, _y);                                   \
 290         GLuint t = p;                                                   \
 291         *((uint32_t *) rgba) = (t >> 24) | (p << 8);                    \
 292      } while (0)
 293 # else
 294 #  define READ_RGBA( rgba, _x, _y )                                     \
 295      do {                                                               \
 296         GLuint p = GET_VALUE(_x, _y);                                   \
 297         rgba[0] = (p >> 16) & 0xff;                                     \
 298         rgba[1] = (p >>  8) & 0xff;                                     \
 299         rgba[2] = (p >>  0) & 0xff;                                     \
 300         rgba[3] = (p >> 24) & 0xff;                                     \
 301      } while (0)
 302 # endif
 303
 304 #elif (SPANTMP_PIXEL_FMT == GL_BGRA) && (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_INT_8_8_8_8)
 305
 306 /**
 307  ** GL_BGRA, GL_UNSIGNED_INT_8_8_8_8
 308  **/
 309
 310 #ifndef GET_VALUE
 311 #ifndef GET_PTR
 312 #define GET_PTR(_x, _y) (     buf + (_x) * 4 + (_y) * pitch)
 313 #endif
 314
 315 #define GET_VALUE(_x, _y) *(volatile GLuint *)(GET_PTR(_x, _y))
 316 #define PUT_VALUE(_x, _y, _v) *(volatile GLuint *)(GET_PTR(_x, _y)) = (_v)
 317 #endif /* GET_VALUE */
 318
 319 # define INIT_MONO_PIXEL(p, color)                       \
 320      p = PACK_COLOR_8888(color[2], color[1], color[0], color[3])
 321
 322 # define WRITE_RGBA(_x, _y, r, g, b, a)                                 \
 323    PUT_VALUE(_x, _y, ((r << 8) |                                        \
 324                       (g << 16) |                                       \
 325                       (b << 24) |                                       \
 326                       (a << 0)))
 327
 328 #define WRITE_PIXEL(_x, _y, p) PUT_VALUE(_x, _y, p)
 329
 330 # if defined( USE_X86_ASM )
 331 #  define READ_RGBA(rgba, _x, _y)                                       \
 332     do {                                                                \
 333        GLuint p = GET_VALUE(_x, _y);                                    \
 334        __asm__ __volatile__( "rorl $8, %0"                              \
 335                                 : "=r" (p) : "0" (p) );                 \
 336        ((GLuint *)rgba)[0] = p;                                         \
 337     } while (0)
 338 # elif defined( MESA_BIG_ENDIAN )
 339     /* On PowerPC with GCC 3.4.2 the shift madness below becomes a single
 340      * rotlwi instruction.  It also produces good code on SPARC.
 341      */
 342 #  define READ_RGBA( rgba, _x, _y )                                     \
 343      do {                                                               \
 344         GLuint p = CPU_TO_LE32(GET_VALUE(_x, _y));                      \
 345         GLuint t = p;                                                   \
 346         *((uint32_t *) rgba) = (t >> 24) | (p << 8);                    \
 347      } while (0)
 348 # else
 349 #  define READ_RGBA( rgba, _x, _y )                                     \
 350      do {                                                               \
 351         GLuint p = GET_VALUE(_x, _y);                                   \
 352         rgba[0] = (p >>  8) & 0xff;                                     \
 353         rgba[1] = (p >> 16) & 0xff;                                     \
 354         rgba[2] = (p >> 24) & 0xff;                                     \
 355         rgba[3] = (p >>  0) & 0xff;                                     \
 356      } while (0)
 357 # endif
 358
 359 #elif (SPANTMP_PIXEL_FMT == GL_BGR) && (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_INT_8_8_8_8_REV)
 360
 361 /**
 362  ** GL_BGR, GL_UNSIGNED_INT_8_8_8_8_REV
 363  **
 364  ** This is really for MESA_FORMAT_XRGB8888.  The spantmp code needs to be
 365  ** kicked to the curb, and we need to just code-gen this.
 366  **/
 367
 368 #ifndef GET_VALUE
 369 #ifndef GET_PTR
 370 #define GET_PTR(_x, _y) (     buf + (_x) * 4 + (_y) * pitch)
 371 #endif
 372
 373 #define GET_VALUE(_x, _y) *(volatile GLuint *)(GET_PTR(_x, _y))
 374 #define PUT_VALUE(_x, _y, _v) *(volatile GLuint *)(GET_PTR(_x, _y)) = (_v)
 375 #endif /* GET_VALUE */
 376
 377 # define INIT_MONO_PIXEL(p, color)                       \
 378      p = PACK_COLOR_8888(0xff, color[0], color[1], color[2])
 379
 380 # define WRITE_RGBA(_x, _y, r, g, b, a)                                 \
 381    PUT_VALUE(_x, _y, ((r << 16) |                                       \
 382                       (g << 8) |                                        \
 383                       (b << 0) |                                        \
 384                       (0xff << 24)))
 385
 386 #define WRITE_PIXEL(_x, _y, p) PUT_VALUE(_x, _y, p)
 387
 388 # if defined( USE_X86_ASM )
 389 #  define READ_RGBA(rgba, _x, _y)                                       \
 390     do {                                                                \
 391        GLuint p = GET_VALUE(_x, _y);                                    \
 392        __asm__ __volatile__( "bswap     %0; rorl $8, %0"                \
 393                                 : "=r" (p) : "0" (p) );                 \
 394        ((GLuint *)rgba)[0] = p | 0xff000000;                            \
 395     } while (0)
 396 # elif defined( MESA_BIG_ENDIAN )
 397     /* On PowerPC with GCC 3.4.2 the shift madness below becomes a single
 398      * rotlwi instruction.  It also produces good code on SPARC.
 399      */
 400 #  define READ_RGBA( rgba, _x, _y )                                     \
 401      do {                                                               \
 402         GLuint p = GET_VALUE(_x, _y);                                   \
 403         *((uint32_t *) rgba) = (p << 8) | 0xff;                         \
 404      } while (0)
 405 # else
 406 #  define READ_RGBA( rgba, _x, _y )                                     \
 407      do {                                                               \
 408         GLuint p = GET_VALUE(_x, _y);                                   \
 409         rgba[0] = (p >> 16) & 0xff;                                     \
 410         rgba[1] = (p >>  8) & 0xff;                                     \
 411         rgba[2] = (p >>  0) & 0xff;                                     \
 412         rgba[3] = 0xff;                                                 \
 413      } while (0)
 414 # endif
 415
 416 #else
 417 #error SPANTMP_PIXEL_FMT must be set to a valid value!
 418 #endif
 419
 420
 421
 422 /**
 423  ** Assembly routines.
 424  **/
 425
 426 #if defined( USE_MMX_ASM ) || defined( USE_SSE_ASM )
 427 #include "x86/read_rgba_span_x86.h"
 428 #include "x86/common_x86_asm.h"
 429 #endif
 430
 431 static void TAG(WriteRGBASpan)( GLcontext *ctx,
 432                                 struct gl_renderbuffer *rb,
 433                                 GLuint n, GLint x, GLint y,
 434                                 const void *values, const GLubyte mask[] )
 435 {
 436    HW_WRITE_LOCK()
 437       {
 438          const GLubyte (*rgba)[4] = (const GLubyte (*)[4]) values;
 439          GLint x1;
 440          GLint n1;
 441          LOCAL_VARS;
 442
 443          y = Y_FLIP(y);
 444
 445          HW_WRITE_CLIPLOOP()
 446             {
 447                GLint i = 0;
 448                CLIPSPAN(x,y,n,x1,n1,i);
 449
 450                if (DBG) fprintf(stderr, "WriteRGBASpan %d..%d (x1 %d)\n",
 451                                 (int)i, (int)n1, (int)x1);
 452
 453                if (mask)
 454                {
 455                   for (;n1>0;i++,x1++,n1--)
 456                      if (mask[i])
 457                         WRITE_RGBA( x1, y,
 458                                     rgba[i][0], rgba[i][1],
 459                                     rgba[i][2], rgba[i][3] );
 460                }
 461                else
 462                {
 463                   for (;n1>0;i++,x1++,n1--)
 464                      WRITE_RGBA( x1, y,
 465                                  rgba[i][0], rgba[i][1],
 466                                  rgba[i][2], rgba[i][3] );
 467                }
 468             }
 469          HW_ENDCLIPLOOP();
 470       }
 471    HW_WRITE_UNLOCK();
 472 }
 473
 474 static void TAG(WriteRGBSpan)( GLcontext *ctx,
 475                                struct gl_renderbuffer *rb,
 476                                GLuint n, GLint x, GLint y,
 477                                const void *values, const GLubyte mask[] )
 478 {
 479    HW_WRITE_LOCK()
 480       {
 481          const GLubyte (*rgb)[3] = (const GLubyte (*)[3]) values;
 482          GLint x1;
 483          GLint n1;
 484          LOCAL_VARS;
 485
 486          y = Y_FLIP(y);
 487
 488          HW_WRITE_CLIPLOOP()
 489             {
 490                GLint i = 0;
 491                CLIPSPAN(x,y,n,x1,n1,i);
 492
 493                if (DBG) fprintf(stderr, "WriteRGBSpan %d..%d (x1 %d)\n",
 494                                 (int)i, (int)n1, (int)x1);
 495
 496                if (mask)
 497                {
 498                   for (;n1>0;i++,x1++,n1--)
 499                      if (mask[i])
 500                         WRITE_RGBA( x1, y, rgb[i][0], rgb[i][1], rgb[i][2], 255 );
 501                }
 502                else
 503                {
 504                   for (;n1>0;i++,x1++,n1--)
 505                      WRITE_RGBA( x1, y, rgb[i][0], rgb[i][1], rgb[i][2], 255 );
 506                }
 507             }
 508          HW_ENDCLIPLOOP();
 509       }
 510    HW_WRITE_UNLOCK();
 511 }
 512
 513 static void TAG(WriteRGBAPixels)( GLcontext *ctx,
 514                                   struct gl_renderbuffer *rb,
 515                                   GLuint n, const GLint x[], const GLint y[],
 516                                   const void *values, const GLubyte mask[] )
 517 {
 518    HW_WRITE_LOCK()
 519       {
 520          const GLubyte (*rgba)[4] = (const GLubyte (*)[4]) values;
 521          GLint i;
 522          LOCAL_VARS;
 523
 524          if (DBG) fprintf(stderr, "WriteRGBAPixels\n");
 525
 526          HW_WRITE_CLIPLOOP()
 527             {
 528                if (mask)
 529                {
 530                   for (i=0;i<n;i++)
 531                   {
 532                      if (mask[i]) {
 533                         const int fy = Y_FLIP(y[i]);
 534                         if (CLIPPIXEL(x[i],fy))
 535                            WRITE_RGBA( x[i], fy,
 536                                        rgba[i][0], rgba[i][1],
 537                                        rgba[i][2], rgba[i][3] );
 538                      }
 539                   }
 540                }
 541                else
 542                {
 543                   for (i=0;i<n;i++)
 544                   {
 545                      const int fy = Y_FLIP(y[i]);
 546                      if (CLIPPIXEL(x[i],fy))
 547                         WRITE_RGBA( x[i], fy,
 548                                     rgba[i][0], rgba[i][1],
 549                                     rgba[i][2], rgba[i][3] );
 550                   }
 551                }
 552             }
 553          HW_ENDCLIPLOOP();
 554       }
 555    HW_WRITE_UNLOCK();
 556 }
 557
 558
 559 static void TAG(WriteMonoRGBASpan)( GLcontext *ctx,
 560                                     struct gl_renderbuffer *rb,
 561                                     GLuint n, GLint x, GLint y,
 562                                     const void *value, const GLubyte mask[] )
 563 {
 564    HW_WRITE_LOCK()
 565       {
 566          const GLubyte *color = (const GLubyte *) value;
 567          GLint x1;
 568          GLint n1;
 569          LOCAL_VARS;
 570          INIT_MONO_PIXEL(p, color);
 571
 572          y = Y_FLIP( y );
 573
 574          if (DBG) fprintf(stderr, "WriteMonoRGBASpan\n");
 575
 576          HW_WRITE_CLIPLOOP()
 577             {
 578                GLint i = 0;
 579                CLIPSPAN(x,y,n,x1,n1,i);
 580                if (mask)
 581                {
 582                   for (;n1>0;i++,x1++,n1--)
 583                      if (mask[i])
 584                         WRITE_PIXEL( x1, y, p );
 585                }
 586                else
 587                {
 588                   for (;n1>0;i++,x1++,n1--)
 589                      WRITE_PIXEL( x1, y, p );
 590                }
 591             }
 592          HW_ENDCLIPLOOP();
 593       }
 594    HW_WRITE_UNLOCK();
 595 }
 596
 597
 598 static void TAG(WriteMonoRGBAPixels)( GLcontext *ctx,
 599                                       struct gl_renderbuffer *rb,
 600                                       GLuint n,
 601                                       const GLint x[], const GLint y[],
 602                                       const void *value,
 603                                       const GLubyte mask[] )
 604 {
 605    HW_WRITE_LOCK()
 606       {
 607          const GLubyte *color = (const GLubyte *) value;
 608          GLint i;
 609          LOCAL_VARS;
 610          INIT_MONO_PIXEL(p, color);
 611
 612          if (DBG) fprintf(stderr, "WriteMonoRGBAPixels\n");
 613
 614          HW_WRITE_CLIPLOOP()
 615             {
 616                if (mask)
 617                {
 618                   for (i=0;i<n;i++)
 619                      if (mask[i]) {
 620                         int fy = Y_FLIP(y[i]);
 621                         if (CLIPPIXEL( x[i], fy ))
 622                            WRITE_PIXEL( x[i], fy, p );
 623                      }
 624                }
 625                else
 626                {
 627                   for (i=0;i<n;i++) {
 628                      int fy = Y_FLIP(y[i]);
 629                      if (CLIPPIXEL( x[i], fy ))
 630                         WRITE_PIXEL( x[i], fy, p );
 631                   }
 632                }
 633             }
 634          HW_ENDCLIPLOOP();
 635       }
 636    HW_WRITE_UNLOCK();
 637 }
 638
 639
 640 static void TAG(ReadRGBASpan)( GLcontext *ctx,
 641                                struct gl_renderbuffer *rb,
 642                                GLuint n, GLint x, GLint y, void *values)
 643 {
 644    HW_READ_LOCK()
 645       {
 646          GLubyte (*rgba)[4] = (GLubyte (*)[4]) values;
 647          GLint x1,n1;
 648          LOCAL_VARS;
 649
 650          y = Y_FLIP(y);
 651
 652          if (DBG) fprintf(stderr, "ReadRGBASpan\n");
 653
 654          HW_READ_CLIPLOOP()
 655             {
 656                GLint i = 0;
 657                CLIPSPAN(x,y,n,x1,n1,i);
 658                for (;n1>0;i++,x1++,n1--)
 659                   READ_RGBA( rgba[i], x1, y );
 660             }
 661          HW_ENDCLIPLOOP();
 662       }
 663    HW_READ_UNLOCK();
 664 }
 665
 666
 667 #if defined(GET_PTR) && \
 668    defined(USE_MMX_ASM) && \
 669    (((SPANTMP_PIXEL_FMT == GL_BGRA) && \
 670         (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_INT_8_8_8_8_REV)) || \
 671     ((SPANTMP_PIXEL_FMT == GL_RGB) && \
 672         (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_SHORT_5_6_5)))
 673 static void TAG2(ReadRGBASpan,_MMX)( GLcontext *ctx,
 674                                      struct gl_renderbuffer *rb,
 675                                      GLuint n, GLint x, GLint y, void *values)
 676 {
 677 #ifndef USE_INNER_EMMS
 678    /* The EMMS instruction is directly in-lined here because using GCC's
 679     * built-in _mm_empty function was found to utterly destroy performance.
 680     */
 681    __asm__ __volatile__( "emms" );
 682 #endif
 683
 684    HW_READ_LOCK()
 685      {
 686         GLubyte (*rgba)[4] = (GLubyte (*)[4]) values;
 687         GLint x1,n1;
 688         LOCAL_VARS;
 689
 690         y = Y_FLIP(y);
 691
 692         if (DBG) fprintf(stderr, "ReadRGBASpan\n");
 693
 694         HW_READ_CLIPLOOP()
 695           {
 696              GLint i = 0;
 697              CLIPSPAN(x,y,n,x1,n1,i);
 698
 699                {
 700                   const void * src = GET_PTR( x1, y );
 701 #if (SPANTMP_PIXEL_FMT == GL_RGB) && \
 702                   (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_SHORT_5_6_5)
 703                   _generic_read_RGBA_span_RGB565_MMX( src, rgba[i], n1 );
 704 #else
 705                   _generic_read_RGBA_span_BGRA8888_REV_MMX( src, rgba[i], n1 );
 706 #endif
 707                }
 708           }
 709         HW_ENDCLIPLOOP();
 710      }
 711    HW_READ_UNLOCK();
 712 #ifndef USE_INNER_EMMS
 713    __asm__ __volatile__( "emms" );
 714 #endif
 715 }
 716 #endif
 717
 718
 719 #if defined(GET_PTR) && \
 720    defined(USE_SSE_ASM) && \
 721    (SPANTMP_PIXEL_FMT == GL_BGRA) && \
 722      (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_INT_8_8_8_8_REV)
 723 static void TAG2(ReadRGBASpan,_SSE2)( GLcontext *ctx,
 724                                       struct gl_renderbuffer *rb,
 725                                       GLuint n, GLint x, GLint y,
 726                                       void *values)
 727 {
 728    HW_READ_LOCK()
 729      {
 730         GLubyte (*rgba)[4] = (GLubyte (*)[4]) values;
 731         GLint x1,n1;
 732         LOCAL_VARS;
 733
 734         y = Y_FLIP(y);
 735
 736         if (DBG) fprintf(stderr, "ReadRGBASpan\n");
 737
 738         HW_READ_CLIPLOOP()
 739           {
 740              GLint i = 0;
 741              CLIPSPAN(x,y,n,x1,n1,i);
 742
 743                {
 744                   const void * src = GET_PTR( x1, y );
 745                   _generic_read_RGBA_span_BGRA8888_REV_SSE2( src, rgba[i], n1 );
 746                }
 747           }
 748         HW_ENDCLIPLOOP();
 749      }
 750    HW_READ_UNLOCK();
 751 }
 752 #endif
 753
 754 #if defined(GET_PTR) && \
 755    defined(USE_SSE_ASM) && \
 756    (SPANTMP_PIXEL_FMT == GL_BGRA) && \
 757      (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_INT_8_8_8_8_REV)
 758 static void TAG2(ReadRGBASpan,_SSE)( GLcontext *ctx,
 759                                      struct gl_renderbuffer *rb,
 760                                      GLuint n, GLint x, GLint y,
 761                                      void *values)
 762 {
 763 #ifndef USE_INNER_EMMS
 764    /* The EMMS instruction is directly in-lined here because using GCC's
 765     * built-in _mm_empty function was found to utterly destroy performance.
 766     */
 767    __asm__ __volatile__( "emms" );
 768 #endif
 769
 770    HW_READ_LOCK()
 771      {
 772         GLubyte (*rgba)[4] = (GLubyte (*)[4]) values;
 773         GLint x1,n1;
 774         LOCAL_VARS;
 775
 776         y = Y_FLIP(y);
 777
 778         if (DBG) fprintf(stderr, "ReadRGBASpan\n");
 779
 780         HW_READ_CLIPLOOP()
 781           {
 782              GLint i = 0;
 783              CLIPSPAN(x,y,n,x1,n1,i);
 784
 785                {
 786                   const void * src = GET_PTR( x1, y );
 787                   _generic_read_RGBA_span_BGRA8888_REV_SSE( src, rgba[i], n1 );
 788                }
 789           }
 790         HW_ENDCLIPLOOP();
 791      }
 792    HW_READ_UNLOCK();
 793 #ifndef USE_INNER_EMMS
 794    __asm__ __volatile__( "emms" );
 795 #endif
 796 }
 797 #endif
 798
 799
 800 static void TAG(ReadRGBAPixels)( GLcontext *ctx,
 801                                  struct gl_renderbuffer *rb,
 802                                  GLuint n, const GLint x[], const GLint y[],
 803                                  void *values )
 804 {
 805    HW_READ_LOCK()
 806       {
 807          GLubyte (*rgba)[4] = (GLubyte (*)[4]) values;
 808          GLubyte *mask = NULL; /* remove someday */
 809          GLint i;
 810          LOCAL_VARS;
 811
 812          if (DBG) fprintf(stderr, "ReadRGBAPixels\n");
 813
 814          HW_READ_CLIPLOOP()
 815             {
 816                if (mask)
 817                {
 818                   for (i=0;i<n;i++)
 819                      if (mask[i]) {
 820                         int fy = Y_FLIP( y[i] );
 821                         if (CLIPPIXEL( x[i], fy ))
 822                            READ_RGBA( rgba[i], x[i], fy );
 823                      }
 824                }
 825                else
 826                {
 827                   for (i=0;i<n;i++) {
 828                      int fy = Y_FLIP( y[i] );
 829                      if (CLIPPIXEL( x[i], fy ))
 830                         READ_RGBA( rgba[i], x[i], fy );
 831                   }
 832                }
 833             }
 834          HW_ENDCLIPLOOP();
 835       }
 836    HW_READ_UNLOCK();
 837 }
 838
 839 static void TAG(InitPointers)(struct gl_renderbuffer *rb)
 840 {
 841    rb->PutRow = TAG(WriteRGBASpan);
 842    rb->PutRowRGB = TAG(WriteRGBSpan);
 843    rb->PutMonoRow = TAG(WriteMonoRGBASpan);
 844    rb->PutValues = TAG(WriteRGBAPixels);
 845    rb->PutMonoValues = TAG(WriteMonoRGBAPixels);
 846    rb->GetValues = TAG(ReadRGBAPixels);
 847
 848 #if defined(GET_PTR)
 849 #if defined(USE_SSE_ASM) && \
 850    (SPANTMP_PIXEL_FMT == GL_BGRA) && \
 851      (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_INT_8_8_8_8_REV)
 852    if ( cpu_has_xmm2 ) {
 853       if (DBG) fprintf( stderr, "Using %s version of GetRow\n", "SSE2" );
 854       rb->GetRow = TAG2(ReadRGBASpan, _SSE2);
 855    }
 856    else
 857 #endif
 858 #if defined(USE_SSE_ASM) && \
 859    (SPANTMP_PIXEL_FMT == GL_BGRA) && \
 860      (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_INT_8_8_8_8_REV)
 861    if ( cpu_has_xmm ) {
 862       if (DBG) fprintf( stderr, "Using %s version of GetRow\n", "SSE" );
 863       rb->GetRow = TAG2(ReadRGBASpan, _SSE);
 864    }
 865    else
 866 #endif
 867 #if defined(USE_MMX_ASM) && \
 868    (((SPANTMP_PIXEL_FMT == GL_BGRA) && \
 869         (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_INT_8_8_8_8_REV)) || \
 870     ((SPANTMP_PIXEL_FMT == GL_RGB) && \
 871         (SPANTMP_PIXEL_TYPE == GL_UNSIGNED_SHORT_5_6_5)))
 872    if ( cpu_has_mmx ) {
 873       if (DBG) fprintf( stderr, "Using %s version of GetRow\n", "MMX" );
 874       rb->GetRow = TAG2(ReadRGBASpan, _MMX);
 875    }
 876    else
 877 #endif
 878 #endif /* GET_PTR */
 879    {
 880       if (DBG) fprintf( stderr, "Using %s version of GetRow\n", "C" );
 881       rb->GetRow = TAG(ReadRGBASpan);
 882    }
 883
 884 }
 885
 886
 887 #undef INIT_MONO_PIXEL
 888 #undef WRITE_PIXEL
 889 #undef WRITE_RGBA
 890 #undef READ_RGBA
 891 #undef TAG
 892 #undef TAG2
 893 #undef GET_VALUE
 894 #undef PUT_VALUE
 895 #undef GET_PTR
 896 #undef SPANTMP_PIXEL_FMT
 897 #undef SPANTMP_PIXEL_TYPE