src/gallium/drivers/llvmpipe/lp_rast.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   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 VMWARE 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 #include <limits.h>
  29 #include "util/u_memory.h"
  30 #include "util/u_math.h"
  31 #include "util/u_rect.h"
  32 #include "util/u_surface.h"
  33 #include "util/u_pack_color.h"
  34
  35 #include "os/os_time.h"
  36
  37 #include "lp_scene_queue.h"
  38 #include "lp_debug.h"
  39 #include "lp_fence.h"
  40 #include "lp_perf.h"
  41 #include "lp_query.h"
  42 #include "lp_rast.h"
  43 #include "lp_rast_priv.h"
  44 #include "gallivm/lp_bld_debug.h"
  45 #include "lp_scene.h"
  46 #include "lp_tex_sample.h"
  47
  48
  49 #ifdef DEBUG
  50 int jit_line = 0;
  51 const struct lp_rast_state *jit_state = NULL;
  52 const struct lp_rasterizer_task *jit_task = NULL;
  53 #endif
  54
  55
  56 /**
  57  * Begin rasterizing a scene.
  58  * Called once per scene by one thread.
  59  */
  60 static void
  61 lp_rast_begin( struct lp_rasterizer *rast,
  62                struct lp_scene *scene )
  63 {
  64
  65    rast->curr_scene = scene;
  66
  67    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
  68
  69    lp_scene_begin_rasterization( scene );
  70    lp_scene_bin_iter_begin( scene );
  71 }
  72
  73
  74 static void
  75 lp_rast_end( struct lp_rasterizer *rast )
  76 {
  77    lp_scene_end_rasterization( rast->curr_scene );
  78
  79    rast->curr_scene = NULL;
  80 }
  81
  82
  83 /**
  84  * Begining rasterization of a tile.
  85  * \param x  window X position of the tile, in pixels
  86  * \param y  window Y position of the tile, in pixels
  87  */
  88 static void
  89 lp_rast_tile_begin(struct lp_rasterizer_task *task,
  90                    const struct cmd_bin *bin)
  91 {
  92    LP_DBG(DEBUG_RAST, "%s %d,%d\n", __FUNCTION__, bin->x, bin->y);
  93
  94    task->bin = bin;
  95    task->x = bin->x * TILE_SIZE;
  96    task->y = bin->y * TILE_SIZE;
  97
  98    /* reset pointers to color and depth tile(s) */
  99    memset(task->color_tiles, 0, sizeof(task->color_tiles));
 100    task->depth_tile = NULL;
 101 }
 102
 103
 104 /**
 105  * Clear the rasterizer's current color tile.
 106  * This is a bin command called during bin processing.
 107  */
 108 static void
 109 lp_rast_clear_color(struct lp_rasterizer_task *task,
 110                     const union lp_rast_cmd_arg arg)
 111 {
 112    const struct lp_scene *scene = task->scene;
 113
 114    if (scene->fb.nr_cbufs) {
 115       unsigned i;
 116       union util_color uc;
 117
 118       if (util_format_is_pure_integer(scene->fb.cbufs[0]->format)) {
 119          /*
 120           * We expect int/uint clear values here, though some APIs
 121           * might disagree (but in any case util_pack_color()
 122           * couldn't handle it)...
 123           */
 124          LP_DBG(DEBUG_RAST, "%s pure int 0x%x,0x%x,0x%x,0x%x\n", __FUNCTION__,
 125                     arg.clear_color.ui[0],
 126                     arg.clear_color.ui[1],
 127                     arg.clear_color.ui[2],
 128                     arg.clear_color.ui[3]);
 129
 130          for (i = 0; i < scene->fb.nr_cbufs; i++) {
 131             enum pipe_format format = scene->fb.cbufs[i]->format;
 132
 133             if (util_format_is_pure_sint(format)) {
 134                util_format_write_4i(format, arg.clear_color.i, 0, &uc, 0, 0, 0, 1, 1);
 135             }
 136             else {
 137                assert(util_format_is_pure_uint(format));
 138                util_format_write_4ui(format, arg.clear_color.ui, 0, &uc, 0, 0, 0, 1, 1);
 139             }
 140
 141             util_fill_rect(scene->cbufs[i].map,
 142                            scene->fb.cbufs[i]->format,
 143                            scene->cbufs[i].stride,
 144                            task->x,
 145                            task->y,
 146                            TILE_SIZE,
 147                            TILE_SIZE,
 148                            &uc);
 149          }
 150       }
 151       else {
 152          uint8_t clear_color[4];
 153
 154          for (i = 0; i < 4; ++i) {
 155             clear_color[i] = float_to_ubyte(arg.clear_color.f[i]);
 156          }
 157
 158          LP_DBG(DEBUG_RAST, "%s 0x%x,0x%x,0x%x,0x%x\n", __FUNCTION__,
 159                     clear_color[0],
 160                     clear_color[1],
 161                     clear_color[2],
 162                     clear_color[3]);
 163
 164          for (i = 0; i < scene->fb.nr_cbufs; i++) {
 165
 166             util_pack_color(arg.clear_color.f,
 167                             scene->fb.cbufs[i]->format, &uc);
 168
 169             util_fill_rect(scene->cbufs[i].map,
 170                            scene->fb.cbufs[i]->format,
 171                            scene->cbufs[i].stride,
 172                            task->x,
 173                            task->y,
 174                            TILE_SIZE,
 175                            TILE_SIZE,
 176                            &uc);
 177          }
 178       }
 179    }
 180
 181    LP_COUNT(nr_color_tile_clear);
 182 }
 183
 184
 185
 186
 187 /**
 188  * Clear the rasterizer's current z/stencil tile.
 189  * This is a bin command called during bin processing.
 190  */
 191 static void
 192 lp_rast_clear_zstencil(struct lp_rasterizer_task *task,
 193                        const union lp_rast_cmd_arg arg)
 194 {
 195    const struct lp_scene *scene = task->scene;
 196    uint64_t clear_value64 = arg.clear_zstencil.value;
 197    uint64_t clear_mask64 = arg.clear_zstencil.mask;
 198    uint32_t clear_value = (uint32_t) clear_value64;
 199    uint32_t clear_mask = (uint32_t) clear_mask64;
 200    const unsigned height = TILE_SIZE;
 201    const unsigned width = TILE_SIZE;
 202    const unsigned block_size = scene->zsbuf.blocksize;
 203    const unsigned dst_stride = scene->zsbuf.stride;
 204    uint8_t *dst;
 205    unsigned i, j;
 206
 207    LP_DBG(DEBUG_RAST, "%s: value=0x%08x, mask=0x%08x\n",
 208            __FUNCTION__, clear_value, clear_mask);
 209
 210    /*
 211     * Clear the area of the depth/depth buffer matching this tile.
 212     */
 213
 214    if (scene->fb.zsbuf) {
 215
 216       dst = lp_rast_get_unswizzled_depth_tile_pointer(task, LP_TEX_USAGE_READ_WRITE);
 217
 218       clear_value &= clear_mask;
 219
 220       switch (block_size) {
 221       case 1:
 222          assert(clear_mask == 0xff);
 223          memset(dst, (uint8_t) clear_value, height * width);
 224          break;
 225       case 2:
 226          if (clear_mask == 0xffff) {
 227             for (i = 0; i < height; i++) {
 228                uint16_t *row = (uint16_t *)dst;
 229                for (j = 0; j < width; j++)
 230                   *row++ = (uint16_t) clear_value;
 231                dst += dst_stride;
 232             }
 233          }
 234          else {
 235             for (i = 0; i < height; i++) {
 236                uint16_t *row = (uint16_t *)dst;
 237                for (j = 0; j < width; j++) {
 238                   uint16_t tmp = ~clear_mask & *row;
 239                   *row++ = clear_value | tmp;
 240                }
 241                dst += dst_stride;
 242             }
 243          }
 244          break;
 245       case 4:
 246          if (clear_mask == 0xffffffff) {
 247             for (i = 0; i < height; i++) {
 248                uint32_t *row = (uint32_t *)dst;
 249                for (j = 0; j < width; j++)
 250                   *row++ = clear_value;
 251                dst += dst_stride;
 252             }
 253          }
 254          else {
 255             for (i = 0; i < height; i++) {
 256                uint32_t *row = (uint32_t *)dst;
 257                for (j = 0; j < width; j++) {
 258                   uint32_t tmp = ~clear_mask & *row;
 259                   *row++ = clear_value | tmp;
 260                }
 261                dst += dst_stride;
 262             }
 263          }
 264          break;
 265       case 8:
 266          clear_value64 &= clear_mask64;
 267          if (clear_mask64 == 0xffffffffffULL) {
 268             for (i = 0; i < height; i++) {
 269                uint64_t *row = (uint64_t *)dst;
 270                for (j = 0; j < width; j++)
 271                   *row++ = clear_value64;
 272                dst += dst_stride;
 273             }
 274          }
 275          else {
 276             for (i = 0; i < height; i++) {
 277                uint64_t *row = (uint64_t *)dst;
 278                for (j = 0; j < width; j++) {
 279                   uint64_t tmp = ~clear_mask64 & *row;
 280                   *row++ = clear_value64 | tmp;
 281                }
 282                dst += dst_stride;
 283             }
 284          }
 285          break;
 286
 287       default:
 288          assert(0);
 289          break;
 290       }
 291    }
 292 }
 293
 294
 295
 296 /**
 297  * Run the shader on all blocks in a tile.  This is used when a tile is
 298  * completely contained inside a triangle.
 299  * This is a bin command called during bin processing.
 300  */
 301 static void
 302 lp_rast_shade_tile(struct lp_rasterizer_task *task,
 303                    const union lp_rast_cmd_arg arg)
 304 {
 305    const struct lp_scene *scene = task->scene;
 306    const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
 307    const struct lp_rast_state *state;
 308    struct lp_fragment_shader_variant *variant;
 309    const unsigned tile_x = task->x, tile_y = task->y;
 310    unsigned x, y;
 311
 312    if (inputs->disable) {
 313       /* This command was partially binned and has been disabled */
 314       return;
 315    }
 316
 317    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
 318
 319    state = task->state;
 320    assert(state);
 321    if (!state) {
 322       return;
 323    }
 324    variant = state->variant;
 325
 326    /* render the whole 64x64 tile in 4x4 chunks */
 327    for (y = 0; y < TILE_SIZE; y += 4){
 328       for (x = 0; x < TILE_SIZE; x += 4) {
 329          uint8_t *color[PIPE_MAX_COLOR_BUFS];
 330          unsigned stride[PIPE_MAX_COLOR_BUFS];
 331          uint8_t *depth = NULL;
 332          unsigned depth_stride = 0;
 333          unsigned i;
 334
 335          /* color buffer */
 336          for (i = 0; i < scene->fb.nr_cbufs; i++){
 337             stride[i] = scene->cbufs[i].stride;
 338
 339             color[i] = lp_rast_get_unswizzled_color_block_pointer(task, i, tile_x + x, tile_y + y);
 340          }
 341
 342          /* depth buffer */
 343          if (scene->zsbuf.map) {
 344             depth = lp_rast_get_unswizzled_depth_block_pointer(task, tile_x + x, tile_y + y);
 345             depth_stride = scene->zsbuf.stride;
 346          }
 347
 348
 349          /* run shader on 4x4 block */
 350          BEGIN_JIT_CALL(state, task);
 351          variant->jit_function[RAST_WHOLE]( &state->jit_context,
 352                                             tile_x + x, tile_y + y,
 353                                             inputs->frontfacing,
 354                                             GET_A0(inputs),
 355                                             GET_DADX(inputs),
 356                                             GET_DADY(inputs),
 357                                             color,
 358                                             depth,
 359                                             0xffff,
 360                                             &task->thread_data,
 361                                             stride,
 362                                             depth_stride);
 363          END_JIT_CALL();
 364       }
 365    }
 366 }
 367
 368
 369 /**
 370  * Run the shader on all blocks in a tile.  This is used when a tile is
 371  * completely contained inside a triangle, and the shader is opaque.
 372  * This is a bin command called during bin processing.
 373  */
 374 static void
 375 lp_rast_shade_tile_opaque(struct lp_rasterizer_task *task,
 376                           const union lp_rast_cmd_arg arg)
 377 {
 378    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
 379
 380    assert(task->state);
 381    if (!task->state) {
 382       return;
 383    }
 384
 385    lp_rast_shade_tile(task, arg);
 386 }
 387
 388
 389 /**
 390  * Compute shading for a 4x4 block of pixels inside a triangle.
 391  * This is a bin command called during bin processing.
 392  * \param x  X position of quad in window coords
 393  * \param y  Y position of quad in window coords
 394  */
 395 void
 396 lp_rast_shade_quads_mask(struct lp_rasterizer_task *task,
 397                          const struct lp_rast_shader_inputs *inputs,
 398                          unsigned x, unsigned y,
 399                          unsigned mask)
 400 {
 401    const struct lp_rast_state *state = task->state;
 402    struct lp_fragment_shader_variant *variant = state->variant;
 403    const struct lp_scene *scene = task->scene;
 404    uint8_t *color[PIPE_MAX_COLOR_BUFS];
 405    unsigned stride[PIPE_MAX_COLOR_BUFS];
 406    void *depth = NULL;
 407    unsigned depth_stride = 0;
 408    unsigned i;
 409
 410    assert(state);
 411
 412    /* Sanity checks */
 413    assert(x < scene->tiles_x * TILE_SIZE);
 414    assert(y < scene->tiles_y * TILE_SIZE);
 415    assert(x % TILE_VECTOR_WIDTH == 0);
 416    assert(y % TILE_VECTOR_HEIGHT == 0);
 417
 418    assert((x % 4) == 0);
 419    assert((y % 4) == 0);
 420
 421    /* color buffer */
 422    for (i = 0; i < scene->fb.nr_cbufs; i++) {
 423       stride[i] = scene->cbufs[i].stride;
 424
 425       color[i] = lp_rast_get_unswizzled_color_block_pointer(task, i, x, y);
 426    }
 427
 428    /* depth buffer */
 429    if (scene->zsbuf.map) {
 430       depth_stride = scene->zsbuf.stride;
 431       depth = lp_rast_get_unswizzled_depth_block_pointer(task, x, y);
 432    }
 433
 434    assert(lp_check_alignment(state->jit_context.u8_blend_color, 16));
 435
 436    /* run shader on 4x4 block */
 437    BEGIN_JIT_CALL(state, task);
 438    variant->jit_function[RAST_EDGE_TEST](&state->jit_context,
 439                                          x, y,
 440                                          inputs->frontfacing,
 441                                          GET_A0(inputs),
 442                                          GET_DADX(inputs),
 443                                          GET_DADY(inputs),
 444                                          color,
 445                                          depth,
 446                                          mask,
 447                                          &task->thread_data,
 448                                          stride,
 449                                          depth_stride);
 450    END_JIT_CALL();
 451 }
 452
 453
 454
 455 /**
 456  * Begin a new occlusion query.
 457  * This is a bin command put in all bins.
 458  * Called per thread.
 459  */
 460 static void
 461 lp_rast_begin_query(struct lp_rasterizer_task *task,
 462                     const union lp_rast_cmd_arg arg)
 463 {
 464    struct llvmpipe_query *pq = arg.query_obj;
 465
 466    assert(task->query[pq->type] == NULL);
 467
 468    switch (pq->type) {
 469    case PIPE_QUERY_OCCLUSION_COUNTER:
 470       task->thread_data.vis_counter = 0;
 471       break;
 472    case PIPE_QUERY_PRIMITIVES_GENERATED:
 473    case PIPE_QUERY_PRIMITIVES_EMITTED:
 474    case PIPE_QUERY_SO_STATISTICS:
 475    case PIPE_QUERY_PIPELINE_STATISTICS:
 476    case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 477       break;
 478    default:
 479       assert(0);
 480       break;
 481    }
 482
 483    task->query[pq->type] = pq;
 484 }
 485
 486
 487 /**
 488  * End the current occlusion query.
 489  * This is a bin command put in all bins.
 490  * Called per thread.
 491  */
 492 static void
 493 lp_rast_end_query(struct lp_rasterizer_task *task,
 494                   const union lp_rast_cmd_arg arg)
 495 {
 496    struct llvmpipe_query *pq = arg.query_obj;
 497    assert(task->query[pq->type] == pq || pq->type == PIPE_QUERY_TIMESTAMP);
 498
 499    switch (pq->type) {
 500    case PIPE_QUERY_OCCLUSION_COUNTER:
 501       pq->count[task->thread_index] += task->thread_data.vis_counter;
 502       break;
 503    case PIPE_QUERY_TIMESTAMP:
 504       pq->count[task->thread_index] = os_time_get_nano();
 505       break;
 506    case PIPE_QUERY_PRIMITIVES_GENERATED:
 507    case PIPE_QUERY_PRIMITIVES_EMITTED:
 508    case PIPE_QUERY_SO_STATISTICS:
 509    case PIPE_QUERY_PIPELINE_STATISTICS:
 510    case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 511       break;
 512    default:
 513       assert(0);
 514       break;
 515    }
 516
 517    if (task->query[pq->type] == pq) {
 518       task->query[pq->type] = NULL;
 519    }
 520 }
 521
 522
 523 void
 524 lp_rast_set_state(struct lp_rasterizer_task *task,
 525                   const union lp_rast_cmd_arg arg)
 526 {
 527    task->state = arg.state;
 528 }
 529
 530
 531
 532 /**
 533  * Called when we're done writing to a color tile.
 534  */
 535 static void
 536 lp_rast_tile_end(struct lp_rasterizer_task *task)
 537 {
 538    unsigned i;
 539
 540    for (i = 0; i < PIPE_QUERY_TYPES; ++i) {
 541       if (task->query[i]) {
 542          lp_rast_end_query(task, lp_rast_arg_query(task->query[i]));
 543       }
 544    }
 545
 546    /* debug */
 547    memset(task->color_tiles, 0, sizeof(task->color_tiles));
 548    task->depth_tile = NULL;
 549
 550    task->bin = NULL;
 551 }
 552
 553 static lp_rast_cmd_func dispatch[LP_RAST_OP_MAX] =
 554 {
 555    lp_rast_clear_color,
 556    lp_rast_clear_zstencil,
 557    lp_rast_triangle_1,
 558    lp_rast_triangle_2,
 559    lp_rast_triangle_3,
 560    lp_rast_triangle_4,
 561    lp_rast_triangle_5,
 562    lp_rast_triangle_6,
 563    lp_rast_triangle_7,
 564    lp_rast_triangle_8,
 565    lp_rast_triangle_3_4,
 566    lp_rast_triangle_3_16,
 567    lp_rast_triangle_4_16,
 568    lp_rast_shade_tile,
 569    lp_rast_shade_tile_opaque,
 570    lp_rast_begin_query,
 571    lp_rast_end_query,
 572    lp_rast_set_state,
 573 };
 574
 575
 576 static void
 577 do_rasterize_bin(struct lp_rasterizer_task *task,
 578                  const struct cmd_bin *bin)
 579 {
 580    const struct cmd_block *block;
 581    unsigned k;
 582
 583    if (0)
 584       lp_debug_bin(bin);
 585
 586    for (block = bin->head; block; block = block->next) {
 587       for (k = 0; k < block->count; k++) {
 588          dispatch[block->cmd[k]]( task, block->arg[k] );
 589       }
 590    }
 591 }
 592
 593
 594
 595 /**
 596  * Rasterize commands for a single bin.
 597  * \param x, y  position of the bin's tile in the framebuffer
 598  * Must be called between lp_rast_begin() and lp_rast_end().
 599  * Called per thread.
 600  */
 601 static void
 602 rasterize_bin(struct lp_rasterizer_task *task,
 603               const struct cmd_bin *bin )
 604 {
 605    lp_rast_tile_begin( task, bin );
 606
 607    do_rasterize_bin(task, bin);
 608
 609    lp_rast_tile_end(task);
 610
 611
 612    /* Debug/Perf flags:
 613     */
 614    if (bin->head->count == 1) {
 615       if (bin->head->cmd[0] == LP_RAST_OP_SHADE_TILE_OPAQUE)
 616          LP_COUNT(nr_pure_shade_opaque_64);
 617       else if (bin->head->cmd[0] == LP_RAST_OP_SHADE_TILE)
 618          LP_COUNT(nr_pure_shade_64);
 619    }
 620 }
 621
 622
 623 /* An empty bin is one that just loads the contents of the tile and
 624  * stores them again unchanged.  This typically happens when bins have
 625  * been flushed for some reason in the middle of a frame, or when
 626  * incremental updates are being made to a render target.
 627  *
 628  * Try to avoid doing pointless work in this case.
 629  */
 630 static boolean
 631 is_empty_bin( const struct cmd_bin *bin )
 632 {
 633    return bin->head == NULL;
 634 }
 635
 636
 637 /**
 638  * Rasterize/execute all bins within a scene.
 639  * Called per thread.
 640  */
 641 static void
 642 rasterize_scene(struct lp_rasterizer_task *task,
 643                 struct lp_scene *scene)
 644 {
 645    task->scene = scene;
 646
 647    if (!task->rast->no_rast && !scene->discard) {
 648       /* loop over scene bins, rasterize each */
 649 #if 0
 650       {
 651          unsigned i, j;
 652          for (i = 0; i < scene->tiles_x; i++) {
 653             for (j = 0; j < scene->tiles_y; j++) {
 654                struct cmd_bin *bin = lp_scene_get_bin(scene, i, j);
 655                rasterize_bin(task, bin, i, j);
 656             }
 657          }
 658       }
 659 #else
 660       {
 661          struct cmd_bin *bin;
 662
 663          assert(scene);
 664          while ((bin = lp_scene_bin_iter_next(scene))) {
 665             if (!is_empty_bin( bin ))
 666                rasterize_bin(task, bin);
 667          }
 668       }
 669 #endif
 670    }
 671
 672
 673    if (scene->fence) {
 674       lp_fence_signal(scene->fence);
 675    }
 676
 677    task->scene = NULL;
 678 }
 679
 680
 681 /**
 682  * Called by setup module when it has something for us to render.
 683  */
 684 void
 685 lp_rast_queue_scene( struct lp_rasterizer *rast,
 686                      struct lp_scene *scene)
 687 {
 688    LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
 689
 690    if (rast->num_threads == 0) {
 691       /* no threading */
 692
 693       lp_rast_begin( rast, scene );
 694
 695       rasterize_scene( &rast->tasks[0], scene );
 696
 697       lp_rast_end( rast );
 698
 699       rast->curr_scene = NULL;
 700    }
 701    else {
 702       /* threaded rendering! */
 703       unsigned i;
 704
 705       lp_scene_enqueue( rast->full_scenes, scene );
 706
 707       /* signal the threads that there's work to do */
 708       for (i = 0; i < rast->num_threads; i++) {
 709          pipe_semaphore_signal(&rast->tasks[i].work_ready);
 710       }
 711    }
 712
 713    LP_DBG(DEBUG_SETUP, "%s done \n", __FUNCTION__);
 714 }
 715
 716
 717 void
 718 lp_rast_finish( struct lp_rasterizer *rast )
 719 {
 720    if (rast->num_threads == 0) {
 721       /* nothing to do */
 722    }
 723    else {
 724       int i;
 725
 726       /* wait for work to complete */
 727       for (i = 0; i < rast->num_threads; i++) {
 728          pipe_semaphore_wait(&rast->tasks[i].work_done);
 729       }
 730    }
 731 }
 732
 733
 734 /**
 735  * This is the thread's main entrypoint.
 736  * It's a simple loop:
 737  *   1. wait for work
 738  *   2. do work
 739  *   3. signal that we're done
 740  */
 741 static PIPE_THREAD_ROUTINE( thread_function, init_data )
 742 {
 743    struct lp_rasterizer_task *task = (struct lp_rasterizer_task *) init_data;
 744    struct lp_rasterizer *rast = task->rast;
 745    boolean debug = false;
 746
 747    while (1) {
 748       /* wait for work */
 749       if (debug)
 750          debug_printf("thread %d waiting for work\n", task->thread_index);
 751       pipe_semaphore_wait(&task->work_ready);
 752
 753       if (rast->exit_flag)
 754          break;
 755
 756       if (task->thread_index == 0) {
 757          /* thread[0]:
 758           *  - get next scene to rasterize
 759           *  - map the framebuffer surfaces
 760           */
 761          lp_rast_begin( rast,
 762                         lp_scene_dequeue( rast->full_scenes, TRUE ) );
 763       }
 764
 765       /* Wait for all threads to get here so that threads[1+] don't
 766        * get a null rast->curr_scene pointer.
 767        */
 768       pipe_barrier_wait( &rast->barrier );
 769
 770       /* do work */
 771       if (debug)
 772          debug_printf("thread %d doing work\n", task->thread_index);
 773
 774       rasterize_scene(task,
 775                       rast->curr_scene);
 776
 777       /* wait for all threads to finish with this scene */
 778       pipe_barrier_wait( &rast->barrier );
 779
 780       /* XXX: shouldn't be necessary:
 781        */
 782       if (task->thread_index == 0) {
 783          lp_rast_end( rast );
 784       }
 785
 786       /* signal done with work */
 787       if (debug)
 788          debug_printf("thread %d done working\n", task->thread_index);
 789
 790       pipe_semaphore_signal(&task->work_done);
 791    }
 792
 793    return NULL;
 794 }
 795
 796
 797 /**
 798  * Initialize semaphores and spawn the threads.
 799  */
 800 static void
 801 create_rast_threads(struct lp_rasterizer *rast)
 802 {
 803    unsigned i;
 804
 805    /* NOTE: if num_threads is zero, we won't use any threads */
 806    for (i = 0; i < rast->num_threads; i++) {
 807       pipe_semaphore_init(&rast->tasks[i].work_ready, 0);
 808       pipe_semaphore_init(&rast->tasks[i].work_done, 0);
 809       rast->threads[i] = pipe_thread_create(thread_function,
 810                                             (void *) &rast->tasks[i]);
 811    }
 812 }
 813
 814
 815
 816 /**
 817  * Create new lp_rasterizer.  If num_threads is zero, don't create any
 818  * new threads, do rendering synchronously.
 819  * \param num_threads  number of rasterizer threads to create
 820  */
 821 struct lp_rasterizer *
 822 lp_rast_create( unsigned num_threads )
 823 {
 824    struct lp_rasterizer *rast;
 825    unsigned i;
 826
 827    rast = CALLOC_STRUCT(lp_rasterizer);
 828    if (!rast) {
 829       goto no_rast;
 830    }
 831
 832    rast->full_scenes = lp_scene_queue_create();
 833    if (!rast->full_scenes) {
 834       goto no_full_scenes;
 835    }
 836
 837    for (i = 0; i < Elements(rast->tasks); i++) {
 838       struct lp_rasterizer_task *task = &rast->tasks[i];
 839       task->rast = rast;
 840       task->thread_index = i;
 841    }
 842
 843    rast->num_threads = num_threads;
 844
 845    rast->no_rast = debug_get_bool_option("LP_NO_RAST", FALSE);
 846
 847    create_rast_threads(rast);
 848
 849    /* for synchronizing rasterization threads */
 850    pipe_barrier_init( &rast->barrier, rast->num_threads );
 851
 852    memset(lp_dummy_tile, 0, sizeof lp_dummy_tile);
 853
 854    return rast;
 855
 856 no_full_scenes:
 857    FREE(rast);
 858 no_rast:
 859    return NULL;
 860 }
 861
 862
 863 /* Shutdown:
 864  */
 865 void lp_rast_destroy( struct lp_rasterizer *rast )
 866 {
 867    unsigned i;
 868
 869    /* Set exit_flag and signal each thread's work_ready semaphore.
 870     * Each thread will be woken up, notice that the exit_flag is set and
 871     * break out of its main loop.  The thread will then exit.
 872     */
 873    rast->exit_flag = TRUE;
 874    for (i = 0; i < rast->num_threads; i++) {
 875       pipe_semaphore_signal(&rast->tasks[i].work_ready);
 876    }
 877
 878    /* Wait for threads to terminate before cleaning up per-thread data */
 879    for (i = 0; i < rast->num_threads; i++) {
 880       pipe_thread_wait(rast->threads[i]);
 881    }
 882
 883    /* Clean up per-thread data */
 884    for (i = 0; i < rast->num_threads; i++) {
 885       pipe_semaphore_destroy(&rast->tasks[i].work_ready);
 886       pipe_semaphore_destroy(&rast->tasks[i].work_done);
 887    }
 888
 889    /* for synchronizing rasterization threads */
 890    pipe_barrier_destroy( &rast->barrier );
 891
 892    lp_scene_queue_destroy(rast->full_scenes);
 893
 894    FREE(rast);
 895 }
 896
 897
 898 /** Return number of rasterization threads */
 899 unsigned
 900 lp_rast_get_num_threads( struct lp_rasterizer *rast )
 901 {
 902    return rast->num_threads;
 903 }
 904
 905