src/gallium/auxiliary/draw/draw_pt_vcache.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
   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
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28  /*
  29   * Authors:
  30   *   Keith Whitwell <keith@tungstengraphics.com>
  31   */
  32
  33 #include "util/u_memory.h"
  34 #include "util/u_prim.h"
  35 #include "draw/draw_context.h"
  36 #include "draw/draw_private.h"
  37 #include "draw/draw_pt.h"
  38
  39
  40 #define CACHE_MAX 256
  41 #define FETCH_MAX 256
  42 #define DRAW_MAX (16*1024)
  43
  44 struct vcache_frontend {
  45    struct draw_pt_front_end base;
  46    struct draw_context *draw;
  47
  48    unsigned in[CACHE_MAX];
  49    ushort out[CACHE_MAX];
  50
  51    ushort draw_elts[DRAW_MAX];
  52    unsigned fetch_elts[FETCH_MAX];
  53
  54    unsigned draw_count;
  55    unsigned fetch_count;
  56    unsigned fetch_max;
  57
  58    struct draw_pt_middle_end *middle;
  59
  60    unsigned input_prim;
  61    unsigned output_prim;
  62
  63    unsigned middle_prim;
  64    unsigned opt;
  65 };
  66
  67 static INLINE void
  68 vcache_flush( struct vcache_frontend *vcache )
  69 {
  70    if (vcache->middle_prim != vcache->output_prim) {
  71       vcache->middle_prim = vcache->output_prim;
  72       vcache->middle->prepare( vcache->middle,
  73                                vcache->input_prim,
  74                                vcache->middle_prim,
  75                                vcache->opt,
  76                                &vcache->fetch_max );
  77    }
  78
  79    if (vcache->draw_count) {
  80       vcache->middle->run( vcache->middle,
  81                            vcache->fetch_elts,
  82                            vcache->fetch_count,
  83                            vcache->draw_elts,
  84                            vcache->draw_count );
  85    }
  86
  87    memset(vcache->in, ~0, sizeof(vcache->in));
  88    vcache->fetch_count = 0;
  89    vcache->draw_count = 0;
  90 }
  91
  92 static INLINE void
  93 vcache_check_flush( struct vcache_frontend *vcache )
  94 {
  95    if ( vcache->draw_count + 6 >= DRAW_MAX ||
  96         vcache->fetch_count + 4 >= FETCH_MAX )
  97    {
  98       vcache_flush( vcache );
  99    }
 100 }
 101
 102
 103 static INLINE void
 104 vcache_elt( struct vcache_frontend *vcache,
 105             unsigned felt,
 106             ushort flags )
 107 {
 108    unsigned idx = felt % CACHE_MAX;
 109
 110    if (vcache->in[idx] != felt) {
 111       assert(vcache->fetch_count < FETCH_MAX);
 112
 113       vcache->in[idx] = felt;
 114       vcache->out[idx] = (ushort)vcache->fetch_count;
 115       vcache->fetch_elts[vcache->fetch_count++] = felt;
 116    }
 117
 118    vcache->draw_elts[vcache->draw_count++] = vcache->out[idx] | flags;
 119 }
 120
 121
 122
 123 static INLINE void
 124 vcache_triangle( struct vcache_frontend *vcache,
 125                  unsigned i0,
 126                  unsigned i1,
 127                  unsigned i2 )
 128 {
 129    vcache_elt(vcache, i0, 0);
 130    vcache_elt(vcache, i1, 0);
 131    vcache_elt(vcache, i2, 0);
 132    vcache_check_flush(vcache);
 133 }
 134
 135
 136 static INLINE void
 137 vcache_triangle_flags( struct vcache_frontend *vcache,
 138                        ushort flags,
 139                        unsigned i0,
 140                        unsigned i1,
 141                        unsigned i2 )
 142 {
 143    vcache_elt(vcache, i0, flags);
 144    vcache_elt(vcache, i1, 0);
 145    vcache_elt(vcache, i2, 0);
 146    vcache_check_flush(vcache);
 147 }
 148
 149 static INLINE void
 150 vcache_line( struct vcache_frontend *vcache,
 151              unsigned i0,
 152              unsigned i1 )
 153 {
 154    vcache_elt(vcache, i0, 0);
 155    vcache_elt(vcache, i1, 0);
 156    vcache_check_flush(vcache);
 157 }
 158
 159
 160 static INLINE void
 161 vcache_line_flags( struct vcache_frontend *vcache,
 162                    ushort flags,
 163                    unsigned i0,
 164                    unsigned i1 )
 165 {
 166    vcache_elt(vcache, i0, flags);
 167    vcache_elt(vcache, i1, 0);
 168    vcache_check_flush(vcache);
 169 }
 170
 171
 172 static INLINE void
 173 vcache_point( struct vcache_frontend *vcache,
 174               unsigned i0 )
 175 {
 176    vcache_elt(vcache, i0, 0);
 177    vcache_check_flush(vcache);
 178 }
 179
 180 static INLINE void
 181 vcache_quad( struct vcache_frontend *vcache,
 182              unsigned i0,
 183              unsigned i1,
 184              unsigned i2,
 185              unsigned i3 )
 186 {
 187    if (vcache->draw->rasterizer->flatshade_first) {
 188       /* pass last quad vertex as first triangle vertex */
 189       vcache_triangle( vcache, i3, i0, i1 );
 190       vcache_triangle( vcache, i3, i1, i2 );
 191    }
 192    else {
 193       /* pass last quad vertex as last triangle vertex */
 194       vcache_triangle( vcache, i0, i1, i3 );
 195       vcache_triangle( vcache, i1, i2, i3 );
 196    }
 197 }
 198
 199 static INLINE void
 200 vcache_ef_quad( struct vcache_frontend *vcache,
 201                 unsigned i0,
 202                 unsigned i1,
 203                 unsigned i2,
 204                 unsigned i3 )
 205 {
 206    if (vcache->draw->rasterizer->flatshade_first) {
 207       /* pass last quad vertex as first triangle vertex */
 208       vcache_triangle_flags( vcache,
 209                              ( DRAW_PIPE_RESET_STIPPLE |
 210                                DRAW_PIPE_EDGE_FLAG_0 |
 211                                DRAW_PIPE_EDGE_FLAG_1 ),
 212                              i3, i0, i1 );
 213
 214       vcache_triangle_flags( vcache,
 215                              ( DRAW_PIPE_EDGE_FLAG_1 |
 216                                DRAW_PIPE_EDGE_FLAG_2 ),
 217                              i3, i1, i2 );
 218    }
 219    else {
 220       /* pass last quad vertex as last triangle vertex */
 221       vcache_triangle_flags( vcache,
 222                              ( DRAW_PIPE_RESET_STIPPLE |
 223                                DRAW_PIPE_EDGE_FLAG_0 |
 224                                DRAW_PIPE_EDGE_FLAG_2 ),
 225                              i0, i1, i3 );
 226
 227       vcache_triangle_flags( vcache,
 228                              ( DRAW_PIPE_EDGE_FLAG_0 |
 229                                DRAW_PIPE_EDGE_FLAG_1 ),
 230                              i1, i2, i3 );
 231    }
 232 }
 233
 234 /* At least for now, we're back to using a template include file for
 235  * this.  The two paths aren't too different though - it may be
 236  * possible to reunify them.
 237  */
 238 #define TRIANGLE(vc,flags,i0,i1,i2) vcache_triangle_flags(vc,flags,i0,i1,i2)
 239 #define QUAD(vc,i0,i1,i2,i3)        vcache_ef_quad(vc,i0,i1,i2,i3)
 240 #define LINE(vc,flags,i0,i1)        vcache_line_flags(vc,flags,i0,i1)
 241 #define POINT(vc,i0)                vcache_point(vc,i0)
 242 #define FUNC vcache_run_extras
 243 #include "draw_pt_vcache_tmp.h"
 244
 245 #define TRIANGLE(vc,flags,i0,i1,i2) vcache_triangle(vc,i0,i1,i2)
 246 #define QUAD(vc,i0,i1,i2,i3)        vcache_quad(vc,i0,i1,i2,i3)
 247 #define LINE(vc,flags,i0,i1)        vcache_line(vc,i0,i1)
 248 #define POINT(vc,i0)                vcache_point(vc,i0)
 249 #define FUNC vcache_run
 250 #include "draw_pt_vcache_tmp.h"
 251
 252 static INLINE void
 253 rebase_uint_elts( const unsigned *src,
 254                   unsigned count,
 255                   int delta,
 256                   ushort *dest )
 257 {
 258    unsigned i;
 259
 260    for (i = 0; i < count; i++)
 261       dest[i] = (ushort)(src[i] + delta);
 262 }
 263
 264 static INLINE void
 265 rebase_ushort_elts( const ushort *src,
 266                     unsigned count,
 267                     int delta,
 268                                 ushort *dest )
 269 {
 270    unsigned i;
 271
 272    for (i = 0; i < count; i++)
 273       dest[i] = (ushort)(src[i] + delta);
 274 }
 275
 276 static INLINE void
 277 rebase_ubyte_elts( const ubyte *src,
 278                    unsigned count,
 279                    int delta,
 280                    ushort *dest )
 281 {
 282    unsigned i;
 283
 284    for (i = 0; i < count; i++)
 285       dest[i] = (ushort)(src[i] + delta);
 286 }
 287
 288
 289
 290 static INLINE void
 291 translate_uint_elts( const unsigned *src,
 292                      unsigned count,
 293                      ushort *dest )
 294 {
 295    unsigned i;
 296
 297    for (i = 0; i < count; i++)
 298       dest[i] = (ushort)(src[i]);
 299 }
 300
 301 static INLINE void
 302 translate_ushort_elts( const ushort *src,
 303                        unsigned count,
 304                        ushort *dest )
 305 {
 306    unsigned i;
 307
 308    for (i = 0; i < count; i++)
 309       dest[i] = (ushort)(src[i]);
 310 }
 311
 312 static INLINE void
 313 translate_ubyte_elts( const ubyte *src,
 314                       unsigned count,
 315                       ushort *dest )
 316 {
 317    unsigned i;
 318
 319    for (i = 0; i < count; i++)
 320       dest[i] = (ushort)(src[i]);
 321 }
 322
 323
 324
 325
 326 #if 0
 327 static INLINE enum pipe_format
 328 format_from_get_elt( pt_elt_func get_elt )
 329 {
 330    switch (draw->pt.user.eltSize) {
 331    case 1: return PIPE_FORMAT_R8_UNORM;
 332    case 2: return PIPE_FORMAT_R16_UNORM;
 333    case 4: return PIPE_FORMAT_R32_UNORM;
 334    default: return PIPE_FORMAT_NONE;
 335    }
 336 }
 337 #endif
 338
 339 static INLINE void
 340 vcache_check_run( struct draw_pt_front_end *frontend,
 341                   pt_elt_func get_elt,
 342                   const void *elts,
 343                   int elt_bias,
 344                   unsigned draw_count )
 345 {
 346    struct vcache_frontend *vcache = (struct vcache_frontend *)frontend;
 347    struct draw_context *draw = vcache->draw;
 348    unsigned min_index = draw->pt.user.min_index;
 349    unsigned max_index = draw->pt.user.max_index;
 350    unsigned index_size = draw->pt.user.eltSize;
 351    unsigned fetch_count = max_index + 1 - min_index;
 352    const ushort *transformed_elts;
 353    ushort *storage = NULL;
 354    boolean ok = FALSE;
 355
 356
 357    if (0) debug_printf("fetch_count %d fetch_max %d draw_count %d\n", fetch_count,
 358                        vcache->fetch_max,
 359                        draw_count);
 360
 361    if (elt_bias + max_index >= DRAW_PIPE_MAX_VERTICES ||
 362        fetch_count >= UNDEFINED_VERTEX_ID ||
 363        fetch_count > draw_count) {
 364       if (0) debug_printf("fail\n");
 365       goto fail;
 366    }
 367
 368    if (vcache->middle_prim != vcache->input_prim) {
 369       vcache->middle_prim = vcache->input_prim;
 370       vcache->middle->prepare( vcache->middle,
 371                                vcache->input_prim,
 372                                vcache->middle_prim,
 373                                vcache->opt,
 374                                &vcache->fetch_max );
 375    }
 376
 377
 378    assert((elt_bias >= 0 && min_index + elt_bias >= min_index) ||
 379           (elt_bias <  0 && min_index + elt_bias <  min_index));
 380
 381    if (min_index == 0 &&
 382        index_size == 2)
 383    {
 384       transformed_elts = (const ushort *)elts;
 385    }
 386    else
 387    {
 388       storage = MALLOC( draw_count * sizeof(ushort) );
 389       if (!storage)
 390          goto fail;
 391
 392       if (min_index == 0) {
 393          switch(index_size) {
 394          case 1:
 395             translate_ubyte_elts( (const ubyte *)elts,
 396                                   draw_count,
 397                                   storage );
 398             break;
 399
 400          case 2:
 401             translate_ushort_elts( (const ushort *)elts,
 402                                    draw_count,
 403                                    storage );
 404             break;
 405
 406          case 4:
 407             translate_uint_elts( (const uint *)elts,
 408                                  draw_count,
 409                                  storage );
 410             break;
 411
 412          default:
 413             assert(0);
 414             FREE(storage);
 415             return;
 416          }
 417       }
 418       else {
 419          switch(index_size) {
 420          case 1:
 421             rebase_ubyte_elts( (const ubyte *)elts,
 422                                   draw_count,
 423                                   0 - (int)min_index,
 424                                   storage );
 425             break;
 426
 427          case 2:
 428             rebase_ushort_elts( (const ushort *)elts,
 429                                    draw_count,
 430                                    0 - (int)min_index,
 431                                    storage );
 432             break;
 433
 434          case 4:
 435             rebase_uint_elts( (const uint *)elts,
 436                                  draw_count,
 437                                  0 - (int)min_index,
 438                                  storage );
 439             break;
 440
 441          default:
 442             assert(0);
 443             FREE(storage);
 444             return;
 445          }
 446       }
 447       transformed_elts = storage;
 448    }
 449
 450    if (fetch_count < UNDEFINED_VERTEX_ID)
 451       ok = vcache->middle->run_linear_elts( vcache->middle,
 452                                             min_index + elt_bias, /* start */
 453                                             fetch_count,
 454                                             transformed_elts,
 455                                             draw_count );
 456
 457    FREE(storage);
 458
 459    if (ok)
 460       return;
 461
 462    debug_printf("failed to execute atomic draw elts for %d/%d, splitting up\n",
 463                 fetch_count, draw_count);
 464
 465  fail:
 466    vcache_run( frontend, get_elt, elts, elt_bias, draw_count );
 467 }
 468
 469
 470
 471
 472 static void
 473 vcache_prepare( struct draw_pt_front_end *frontend,
 474                 unsigned in_prim,
 475                 unsigned out_prim,
 476                 struct draw_pt_middle_end *middle,
 477                 unsigned opt )
 478 {
 479    struct vcache_frontend *vcache = (struct vcache_frontend *)frontend;
 480
 481    if (opt & PT_PIPELINE)
 482    {
 483       vcache->base.run = vcache_run_extras;
 484    }
 485    else
 486    {
 487       vcache->base.run = vcache_check_run;
 488    }
 489
 490    vcache->input_prim = in_prim;
 491    vcache->output_prim = u_reduced_prim(out_prim);
 492
 493    vcache->middle = middle;
 494    vcache->opt = opt;
 495
 496    /* Have to run prepare here, but try and guess a good prim for
 497     * doing so:
 498     */
 499    vcache->middle_prim = (opt & PT_PIPELINE) ? vcache->output_prim : vcache->input_prim;
 500    middle->prepare( middle, vcache->input_prim,
 501                     vcache->middle_prim, opt, &vcache->fetch_max );
 502 }
 503
 504
 505
 506
 507 static void
 508 vcache_finish( struct draw_pt_front_end *frontend )
 509 {
 510    struct vcache_frontend *vcache = (struct vcache_frontend *)frontend;
 511    vcache->middle->finish( vcache->middle );
 512    vcache->middle = NULL;
 513 }
 514
 515 static void
 516 vcache_destroy( struct draw_pt_front_end *frontend )
 517 {
 518    FREE(frontend);
 519 }
 520
 521
 522 struct draw_pt_front_end *draw_pt_vcache( struct draw_context *draw )
 523 {
 524    struct vcache_frontend *vcache = CALLOC_STRUCT( vcache_frontend );
 525    if (vcache == NULL)
 526       return NULL;
 527
 528    vcache->base.prepare = vcache_prepare;
 529    vcache->base.run     = NULL;
 530    vcache->base.finish  = vcache_finish;
 531    vcache->base.destroy = vcache_destroy;
 532    vcache->draw = draw;
 533
 534    memset(vcache->in, ~0, sizeof(vcache->in));
 535
 536    return &vcache->base;
 537 }