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_context.h"
  39 #include "lp_debug.h"
  40 #include "lp_fence.h"
  41 #include "lp_perf.h"
  42 #include "lp_query.h"
  43 #include "lp_rast.h"
  44 #include "lp_rast_priv.h"
  45 #include "gallivm/lp_bld_debug.h"
  46 #include "lp_scene.h"
  47 #include "lp_tex_sample.h"
  48
  49
  50 #ifdef DEBUG
  51 int jit_line = 0;
  52 const struct lp_rast_state *jit_state = NULL;
  53 const struct lp_rasterizer_task *jit_task = NULL;
  54 #endif
  55
  56
  57 /**
  58  * Begin rasterizing a scene.
  59  * Called once per scene by one thread.
  60  */
  61 static void
  62 lp_rast_begin( struct lp_rasterizer *rast,
  63                struct lp_scene *scene )
  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                    int x, int y)
  92 {
  93    LP_DBG(DEBUG_RAST, "%s %d,%d\n", __FUNCTION__, x, y);
  94
  95    task->bin = bin;
  96    task->x = x * TILE_SIZE;
  97    task->y = y * TILE_SIZE;
  98    task->width = TILE_SIZE + x * TILE_SIZE > task->scene->fb.width ?
  99                     task->scene->fb.width - x * TILE_SIZE : TILE_SIZE;
 100    task->height = TILE_SIZE + y * TILE_SIZE > task->scene->fb.height ?
 101                     task->scene->fb.height - y * TILE_SIZE : TILE_SIZE;
 102
 103    task->thread_data.vis_counter = 0;
 104    task->ps_invocations = 0;
 105
 106    /* reset pointers to color and depth tile(s) */
 107    memset(task->color_tiles, 0, sizeof(task->color_tiles));
 108    task->depth_tile = NULL;
 109 }
 110
 111
 112 /**
 113  * Clear the rasterizer's current color tile.
 114  * This is a bin command called during bin processing.
 115  * Clear commands always clear all bound layers.
 116  */
 117 static void
 118 lp_rast_clear_color(struct lp_rasterizer_task *task,
 119                     const union lp_rast_cmd_arg arg)
 120 {
 121    const struct lp_scene *scene = task->scene;
 122    unsigned cbuf = arg.clear_rb->cbuf;
 123    union util_color uc;
 124    enum pipe_format format;
 125
 126    /* we never bin clear commands for non-existing buffers */
 127    assert(cbuf < scene->fb.nr_cbufs);
 128    assert(scene->fb.cbufs[cbuf]);
 129
 130    format = scene->fb.cbufs[cbuf]->format;
 131    uc = arg.clear_rb->color_val;
 132
 133    /*
 134     * this is pretty rough since we have target format (bunch of bytes...) here.
 135     * dump it as raw 4 dwords.
 136     */
 137    LP_DBG(DEBUG_RAST, "%s clear value (target format %d) raw 0x%x,0x%x,0x%x,0x%x\n",
 138           __FUNCTION__, format, uc.ui[0], uc.ui[1], uc.ui[2], uc.ui[3]);
 139
 140
 141    util_fill_box(scene->cbufs[cbuf].map,
 142                  format,
 143                  scene->cbufs[cbuf].stride,
 144                  scene->cbufs[cbuf].layer_stride,
 145                  task->x,
 146                  task->y,
 147                  0,
 148                  task->width,
 149                  task->height,
 150                  scene->fb_max_layer + 1,
 151                  &uc);
 152
 153    /* this will increase for each rb which probably doesn't mean much */
 154    LP_COUNT(nr_color_tile_clear);
 155 }
 156
 157
 158 /**
 159  * Clear the rasterizer's current z/stencil tile.
 160  * This is a bin command called during bin processing.
 161  * Clear commands always clear all bound layers.
 162  */
 163 static void
 164 lp_rast_clear_zstencil(struct lp_rasterizer_task *task,
 165                        const union lp_rast_cmd_arg arg)
 166 {
 167    const struct lp_scene *scene = task->scene;
 168    uint64_t clear_value64 = arg.clear_zstencil.value;
 169    uint64_t clear_mask64 = arg.clear_zstencil.mask;
 170    uint32_t clear_value = (uint32_t) clear_value64;
 171    uint32_t clear_mask = (uint32_t) clear_mask64;
 172    const unsigned height = task->height;
 173    const unsigned width = task->width;
 174    const unsigned dst_stride = scene->zsbuf.stride;
 175    uint8_t *dst;
 176    unsigned i, j;
 177    unsigned block_size;
 178
 179    LP_DBG(DEBUG_RAST, "%s: value=0x%08x, mask=0x%08x\n",
 180            __FUNCTION__, clear_value, clear_mask);
 181
 182    /*
 183     * Clear the area of the depth/depth buffer matching this tile.
 184     */
 185
 186    if (scene->fb.zsbuf) {
 187       unsigned layer;
 188       uint8_t *dst_layer = lp_rast_get_unswizzled_depth_tile_pointer(task, LP_TEX_USAGE_READ_WRITE);
 189       block_size = util_format_get_blocksize(scene->fb.zsbuf->format);
 190
 191       clear_value &= clear_mask;
 192
 193       for (layer = 0; layer <= scene->fb_max_layer; layer++) {
 194          dst = dst_layer;
 195
 196          switch (block_size) {
 197          case 1:
 198             assert(clear_mask == 0xff);
 199             memset(dst, (uint8_t) clear_value, height * width);
 200             break;
 201          case 2:
 202             if (clear_mask == 0xffff) {
 203                for (i = 0; i < height; i++) {
 204                   uint16_t *row = (uint16_t *)dst;
 205                   for (j = 0; j < width; j++)
 206                      *row++ = (uint16_t) clear_value;
 207                   dst += dst_stride;
 208                }
 209             }
 210             else {
 211                for (i = 0; i < height; i++) {
 212                   uint16_t *row = (uint16_t *)dst;
 213                   for (j = 0; j < width; j++) {
 214                      uint16_t tmp = ~clear_mask & *row;
 215                      *row++ = clear_value | tmp;
 216                   }
 217                   dst += dst_stride;
 218                }
 219             }
 220             break;
 221          case 4:
 222             if (clear_mask == 0xffffffff) {
 223                for (i = 0; i < height; i++) {
 224                   uint32_t *row = (uint32_t *)dst;
 225                   for (j = 0; j < width; j++)
 226                      *row++ = clear_value;
 227                   dst += dst_stride;
 228                }
 229             }
 230             else {
 231                for (i = 0; i < height; i++) {
 232                   uint32_t *row = (uint32_t *)dst;
 233                   for (j = 0; j < width; j++) {
 234                      uint32_t tmp = ~clear_mask & *row;
 235                      *row++ = clear_value | tmp;
 236                   }
 237                   dst += dst_stride;
 238                }
 239             }
 240             break;
 241          case 8:
 242             clear_value64 &= clear_mask64;
 243             if (clear_mask64 == 0xffffffffffULL) {
 244                for (i = 0; i < height; i++) {
 245                   uint64_t *row = (uint64_t *)dst;
 246                   for (j = 0; j < width; j++)
 247                      *row++ = clear_value64;
 248                   dst += dst_stride;
 249                }
 250             }
 251             else {
 252                for (i = 0; i < height; i++) {
 253                   uint64_t *row = (uint64_t *)dst;
 254                   for (j = 0; j < width; j++) {
 255                      uint64_t tmp = ~clear_mask64 & *row;
 256                      *row++ = clear_value64 | tmp;
 257                   }
 258                   dst += dst_stride;
 259                }
 260             }
 261             break;
 262
 263          default:
 264             assert(0);
 265             break;
 266          }
 267          dst_layer += scene->zsbuf.layer_stride;
 268       }
 269    }
 270 }
 271
 272
 273
 274 /**
 275  * Run the shader on all blocks in a tile.  This is used when a tile is
 276  * completely contained inside a triangle.
 277  * This is a bin command called during bin processing.
 278  */
 279 static void
 280 lp_rast_shade_tile(struct lp_rasterizer_task *task,
 281                    const union lp_rast_cmd_arg arg)
 282 {
 283    const struct lp_scene *scene = task->scene;
 284    const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
 285    const struct lp_rast_state *state;
 286    struct lp_fragment_shader_variant *variant;
 287    const unsigned tile_x = task->x, tile_y = task->y;
 288    unsigned x, y;
 289
 290    if (inputs->disable) {
 291       /* This command was partially binned and has been disabled */
 292       return;
 293    }
 294
 295    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
 296
 297    state = task->state;
 298    assert(state);
 299    if (!state) {
 300       return;
 301    }
 302    variant = state->variant;
 303
 304    /* render the whole 64x64 tile in 4x4 chunks */
 305    for (y = 0; y < task->height; y += 4){
 306       for (x = 0; x < task->width; x += 4) {
 307          uint8_t *color[PIPE_MAX_COLOR_BUFS];
 308          unsigned stride[PIPE_MAX_COLOR_BUFS];
 309          uint8_t *depth = NULL;
 310          unsigned depth_stride = 0;
 311          unsigned i;
 312
 313          /* color buffer */
 314          for (i = 0; i < scene->fb.nr_cbufs; i++){
 315             stride[i] = scene->cbufs[i].stride;
 316             color[i] = lp_rast_get_unswizzled_color_block_pointer(task, i, tile_x + x,
 317                                                                   tile_y + y, inputs->layer);
 318          }
 319
 320          /* depth buffer */
 321          if (scene->zsbuf.map) {
 322             depth = lp_rast_get_unswizzled_depth_block_pointer(task, tile_x + x,
 323                                                                tile_y + y, inputs->layer);
 324             depth_stride = scene->zsbuf.stride;
 325          }
 326
 327          /* Propagate non-interpolated raster state. */
 328          task->thread_data.raster_state.viewport_index = inputs->viewport_index;
 329
 330          /* run shader on 4x4 block */
 331          BEGIN_JIT_CALL(state, task);
 332          variant->jit_function[RAST_WHOLE]( &state->jit_context,
 333                                             tile_x + x, tile_y + y,
 334                                             inputs->frontfacing,
 335                                             GET_A0(inputs),
 336                                             GET_DADX(inputs),
 337                                             GET_DADY(inputs),
 338                                             color,
 339                                             depth,
 340                                             0xffff,
 341                                             &task->thread_data,
 342                                             stride,
 343                                             depth_stride);
 344          END_JIT_CALL();
 345       }
 346    }
 347 }
 348
 349
 350 /**
 351  * Run the shader on all blocks in a tile.  This is used when a tile is
 352  * completely contained inside a triangle, and the shader is opaque.
 353  * This is a bin command called during bin processing.
 354  */
 355 static void
 356 lp_rast_shade_tile_opaque(struct lp_rasterizer_task *task,
 357                           const union lp_rast_cmd_arg arg)
 358 {
 359    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
 360
 361    assert(task->state);
 362    if (!task->state) {
 363       return;
 364    }
 365
 366    lp_rast_shade_tile(task, arg);
 367 }
 368
 369
 370 /**
 371  * Compute shading for a 4x4 block of pixels inside a triangle.
 372  * This is a bin command called during bin processing.
 373  * \param x  X position of quad in window coords
 374  * \param y  Y position of quad in window coords
 375  */
 376 void
 377 lp_rast_shade_quads_mask(struct lp_rasterizer_task *task,
 378                          const struct lp_rast_shader_inputs *inputs,
 379                          unsigned x, unsigned y,
 380                          unsigned mask)
 381 {
 382    const struct lp_rast_state *state = task->state;
 383    struct lp_fragment_shader_variant *variant = state->variant;
 384    const struct lp_scene *scene = task->scene;
 385    uint8_t *color[PIPE_MAX_COLOR_BUFS];
 386    unsigned stride[PIPE_MAX_COLOR_BUFS];
 387    uint8_t *depth = NULL;
 388    unsigned depth_stride = 0;
 389    unsigned i;
 390
 391    assert(state);
 392
 393    /* Sanity checks */
 394    assert(x < scene->tiles_x * TILE_SIZE);
 395    assert(y < scene->tiles_y * TILE_SIZE);
 396    assert(x % TILE_VECTOR_WIDTH == 0);
 397    assert(y % TILE_VECTOR_HEIGHT == 0);
 398
 399    assert((x % 4) == 0);
 400    assert((y % 4) == 0);
 401
 402    /* color buffer */
 403    for (i = 0; i < scene->fb.nr_cbufs; i++) {
 404       if (scene->fb.cbufs[i]) {
 405          stride[i] = scene->cbufs[i].stride;
 406          color[i] = lp_rast_get_unswizzled_color_block_pointer(task, i, x, y,
 407                                                                inputs->layer);
 408       }
 409       else {
 410          stride[i] = 0;
 411          color[i] = NULL;
 412       }
 413    }
 414
 415    /* depth buffer */
 416    if (scene->zsbuf.map) {
 417       depth_stride = scene->zsbuf.stride;
 418       depth = lp_rast_get_unswizzled_depth_block_pointer(task, x, y, inputs->layer);
 419    }
 420
 421    assert(lp_check_alignment(state->jit_context.u8_blend_color, 16));
 422
 423    /*
 424     * The rasterizer may produce fragments outside our
 425     * allocated 4x4 blocks hence need to filter them out here.
 426     */
 427    if ((x % TILE_SIZE) < task->width && (y % TILE_SIZE) < task->height) {
 428       /* not very accurate would need a popcount on the mask */
 429       /* always count this not worth bothering? */
 430       task->ps_invocations += 1 * variant->ps_inv_multiplier;
 431
 432       /* Propagate non-interpolated raster state. */
 433       task->thread_data.raster_state.viewport_index = inputs->viewport_index;
 434
 435       /* run shader on 4x4 block */
 436       BEGIN_JIT_CALL(state, task);
 437       variant->jit_function[RAST_EDGE_TEST](&state->jit_context,
 438                                             x, y,
 439                                             inputs->frontfacing,
 440                                             GET_A0(inputs),
 441                                             GET_DADX(inputs),
 442                                             GET_DADY(inputs),
 443                                             color,
 444                                             depth,
 445                                             mask,
 446                                             &task->thread_data,
 447                                             stride,
 448                                             depth_stride);
 449       END_JIT_CALL();
 450    }
 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    switch (pq->type) {
 467    case PIPE_QUERY_OCCLUSION_COUNTER:
 468    case PIPE_QUERY_OCCLUSION_PREDICATE:
 469       pq->start[task->thread_index] = task->thread_data.vis_counter;
 470       break;
 471    case PIPE_QUERY_PIPELINE_STATISTICS:
 472       pq->start[task->thread_index] = task->ps_invocations;
 473       break;
 474    default:
 475       assert(0);
 476       break;
 477    }
 478 }
 479
 480
 481 /**
 482  * End the current occlusion query.
 483  * This is a bin command put in all bins.
 484  * Called per thread.
 485  */
 486 static void
 487 lp_rast_end_query(struct lp_rasterizer_task *task,
 488                   const union lp_rast_cmd_arg arg)
 489 {
 490    struct llvmpipe_query *pq = arg.query_obj;
 491
 492    switch (pq->type) {
 493    case PIPE_QUERY_OCCLUSION_COUNTER:
 494    case PIPE_QUERY_OCCLUSION_PREDICATE:
 495       pq->end[task->thread_index] +=
 496          task->thread_data.vis_counter - pq->start[task->thread_index];
 497       pq->start[task->thread_index] = 0;
 498       break;
 499    case PIPE_QUERY_TIMESTAMP:
 500       pq->end[task->thread_index] = os_time_get_nano();
 501       break;
 502    case PIPE_QUERY_PIPELINE_STATISTICS:
 503       pq->end[task->thread_index] +=
 504          task->ps_invocations - pq->start[task->thread_index];
 505       pq->start[task->thread_index] = 0;
 506       break;
 507    default:
 508       assert(0);
 509       break;
 510    }
 511 }
 512
 513
 514 void
 515 lp_rast_set_state(struct lp_rasterizer_task *task,
 516                   const union lp_rast_cmd_arg arg)
 517 {
 518    task->state = arg.state;
 519 }
 520
 521
 522
 523 /**
 524  * Called when we're done writing to a color tile.
 525  */
 526 static void
 527 lp_rast_tile_end(struct lp_rasterizer_task *task)
 528 {
 529    unsigned i;
 530
 531    for (i = 0; i < task->scene->num_active_queries; ++i) {
 532       lp_rast_end_query(task, lp_rast_arg_query(task->scene->active_queries[i]));
 533    }
 534
 535    /* debug */
 536    memset(task->color_tiles, 0, sizeof(task->color_tiles));
 537    task->depth_tile = NULL;
 538
 539    task->bin = NULL;
 540 }
 541
 542 static lp_rast_cmd_func dispatch[LP_RAST_OP_MAX] =
 543 {
 544    lp_rast_clear_color,
 545    lp_rast_clear_zstencil,
 546    lp_rast_triangle_1,
 547    lp_rast_triangle_2,
 548    lp_rast_triangle_3,
 549    lp_rast_triangle_4,
 550    lp_rast_triangle_5,
 551    lp_rast_triangle_6,
 552    lp_rast_triangle_7,
 553    lp_rast_triangle_8,
 554    lp_rast_triangle_3_4,
 555    lp_rast_triangle_3_16,
 556    lp_rast_triangle_4_16,
 557    lp_rast_shade_tile,
 558    lp_rast_shade_tile_opaque,
 559    lp_rast_begin_query,
 560    lp_rast_end_query,
 561    lp_rast_set_state,
 562    lp_rast_triangle_32_1,
 563    lp_rast_triangle_32_2,
 564    lp_rast_triangle_32_3,
 565    lp_rast_triangle_32_4,
 566    lp_rast_triangle_32_5,
 567    lp_rast_triangle_32_6,
 568    lp_rast_triangle_32_7,
 569    lp_rast_triangle_32_8,
 570    lp_rast_triangle_32_3_4,
 571    lp_rast_triangle_32_3_16,
 572    lp_rast_triangle_32_4_16
 573 };
 574
 575
 576 static void
 577 do_rasterize_bin(struct lp_rasterizer_task *task,
 578                  const struct cmd_bin *bin,
 579                  int x, int y)
 580 {
 581    const struct cmd_block *block;
 582    unsigned k;
 583
 584    if (0)
 585       lp_debug_bin(bin, x, y);
 586
 587    for (block = bin->head; block; block = block->next) {
 588       for (k = 0; k < block->count; k++) {
 589          dispatch[block->cmd[k]]( task, block->arg[k] );
 590       }
 591    }
 592 }
 593
 594
 595
 596 /**
 597  * Rasterize commands for a single bin.
 598  * \param x, y  position of the bin's tile in the framebuffer
 599  * Must be called between lp_rast_begin() and lp_rast_end().
 600  * Called per thread.
 601  */
 602 static void
 603 rasterize_bin(struct lp_rasterizer_task *task,
 604               const struct cmd_bin *bin, int x, int y )
 605 {
 606    lp_rast_tile_begin( task, bin, x, y );
 607
 608    do_rasterize_bin(task, bin, x, y);
 609
 610    lp_rast_tile_end(task);
 611
 612
 613    /* Debug/Perf flags:
 614     */
 615    if (bin->head->count == 1) {
 616       if (bin->head->cmd[0] == LP_RAST_OP_SHADE_TILE_OPAQUE)
 617          LP_COUNT(nr_pure_shade_opaque_64);
 618       else if (bin->head->cmd[0] == LP_RAST_OP_SHADE_TILE)
 619          LP_COUNT(nr_pure_shade_64);
 620    }
 621 }
 622
 623
 624 /* An empty bin is one that just loads the contents of the tile and
 625  * stores them again unchanged.  This typically happens when bins have
 626  * been flushed for some reason in the middle of a frame, or when
 627  * incremental updates are being made to a render target.
 628  *
 629  * Try to avoid doing pointless work in this case.
 630  */
 631 static boolean
 632 is_empty_bin( const struct cmd_bin *bin )
 633 {
 634    return bin->head == NULL;
 635 }
 636
 637
 638 /**
 639  * Rasterize/execute all bins within a scene.
 640  * Called per thread.
 641  */
 642 static void
 643 rasterize_scene(struct lp_rasterizer_task *task,
 644                 struct lp_scene *scene)
 645 {
 646    task->scene = scene;
 647
 648    if (!task->rast->no_rast && !scene->discard) {
 649       /* loop over scene bins, rasterize each */
 650       {
 651          struct cmd_bin *bin;
 652          int i, j;
 653
 654          assert(scene);
 655          while ((bin = lp_scene_bin_iter_next(scene, &i, &j))) {
 656             if (!is_empty_bin( bin ))
 657                rasterize_bin(task, bin, i, j);
 658          }
 659       }
 660    }
 661
 662
 663    if (scene->fence) {
 664       lp_fence_signal(scene->fence);
 665    }
 666
 667    task->scene = NULL;
 668 }
 669
 670
 671 /**
 672  * Called by setup module when it has something for us to render.
 673  */
 674 void
 675 lp_rast_queue_scene( struct lp_rasterizer *rast,
 676                      struct lp_scene *scene)
 677 {
 678    LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
 679
 680    if (rast->num_threads == 0) {
 681       /* no threading */
 682       unsigned fpstate = util_fpstate_get();
 683
 684       /* Make sure that denorms are treated like zeros. This is
 685        * the behavior required by D3D10. OpenGL doesn't care.
 686        */
 687       util_fpstate_set_denorms_to_zero(fpstate);
 688
 689       lp_rast_begin( rast, scene );
 690
 691       rasterize_scene( &rast->tasks[0], scene );
 692
 693       lp_rast_end( rast );
 694
 695       util_fpstate_set(fpstate);
 696
 697       rast->curr_scene = NULL;
 698    }
 699    else {
 700       /* threaded rendering! */
 701       unsigned i;
 702
 703       lp_scene_enqueue( rast->full_scenes, scene );
 704
 705       /* signal the threads that there's work to do */
 706       for (i = 0; i < rast->num_threads; i++) {
 707          pipe_semaphore_signal(&rast->tasks[i].work_ready);
 708       }
 709    }
 710
 711    LP_DBG(DEBUG_SETUP, "%s done \n", __FUNCTION__);
 712 }
 713
 714
 715 void
 716 lp_rast_finish( struct lp_rasterizer *rast )
 717 {
 718    if (rast->num_threads == 0) {
 719       /* nothing to do */
 720    }
 721    else {
 722       int i;
 723
 724       /* wait for work to complete */
 725       for (i = 0; i < rast->num_threads; i++) {
 726          pipe_semaphore_wait(&rast->tasks[i].work_done);
 727       }
 728    }
 729 }
 730
 731
 732 /**
 733  * This is the thread's main entrypoint.
 734  * It's a simple loop:
 735  *   1. wait for work
 736  *   2. do work
 737  *   3. signal that we're done
 738  */
 739 static PIPE_THREAD_ROUTINE( thread_function, init_data )
 740 {
 741    struct lp_rasterizer_task *task = (struct lp_rasterizer_task *) init_data;
 742    struct lp_rasterizer *rast = task->rast;
 743    boolean debug = false;
 744    unsigned fpstate = util_fpstate_get();
 745
 746    /* Make sure that denorms are treated like zeros. This is
 747     * the behavior required by D3D10. OpenGL doesn't care.
 748     */
 749    util_fpstate_set_denorms_to_zero(fpstate);
 750
 751    while (1) {
 752       /* wait for work */
 753       if (debug)
 754          debug_printf("thread %d waiting for work\n", task->thread_index);
 755       pipe_semaphore_wait(&task->work_ready);
 756
 757       if (rast->exit_flag)
 758          break;
 759
 760       if (task->thread_index == 0) {
 761          /* thread[0]:
 762           *  - get next scene to rasterize
 763           *  - map the framebuffer surfaces
 764           */
 765          lp_rast_begin( rast,
 766                         lp_scene_dequeue( rast->full_scenes, TRUE ) );
 767       }
 768
 769       /* Wait for all threads to get here so that threads[1+] don't
 770        * get a null rast->curr_scene pointer.
 771        */
 772       pipe_barrier_wait( &rast->barrier );
 773
 774       /* do work */
 775       if (debug)
 776          debug_printf("thread %d doing work\n", task->thread_index);
 777
 778       rasterize_scene(task,
 779                       rast->curr_scene);
 780
 781       /* wait for all threads to finish with this scene */
 782       pipe_barrier_wait( &rast->barrier );
 783
 784       /* XXX: shouldn't be necessary:
 785        */
 786       if (task->thread_index == 0) {
 787          lp_rast_end( rast );
 788       }
 789
 790       /* signal done with work */
 791       if (debug)
 792          debug_printf("thread %d done working\n", task->thread_index);
 793
 794       pipe_semaphore_signal(&task->work_done);
 795    }
 796
 797    return 0;
 798 }
 799
 800
 801 /**
 802  * Initialize semaphores and spawn the threads.
 803  */
 804 static void
 805 create_rast_threads(struct lp_rasterizer *rast)
 806 {
 807    unsigned i;
 808
 809    /* NOTE: if num_threads is zero, we won't use any threads */
 810    for (i = 0; i < rast->num_threads; i++) {
 811       pipe_semaphore_init(&rast->tasks[i].work_ready, 0);
 812       pipe_semaphore_init(&rast->tasks[i].work_done, 0);
 813       rast->threads[i] = pipe_thread_create(thread_function,
 814                                             (void *) &rast->tasks[i]);
 815    }
 816 }
 817
 818
 819
 820 /**
 821  * Create new lp_rasterizer.  If num_threads is zero, don't create any
 822  * new threads, do rendering synchronously.
 823  * \param num_threads  number of rasterizer threads to create
 824  */
 825 struct lp_rasterizer *
 826 lp_rast_create( unsigned num_threads )
 827 {
 828    struct lp_rasterizer *rast;
 829    unsigned i;
 830
 831    rast = CALLOC_STRUCT(lp_rasterizer);
 832    if (!rast) {
 833       goto no_rast;
 834    }
 835
 836    rast->full_scenes = lp_scene_queue_create();
 837    if (!rast->full_scenes) {
 838       goto no_full_scenes;
 839    }
 840
 841    for (i = 0; i < Elements(rast->tasks); i++) {
 842       struct lp_rasterizer_task *task = &rast->tasks[i];
 843       task->rast = rast;
 844       task->thread_index = i;
 845    }
 846
 847    rast->num_threads = num_threads;
 848
 849    rast->no_rast = debug_get_bool_option("LP_NO_RAST", FALSE);
 850
 851    create_rast_threads(rast);
 852
 853    /* for synchronizing rasterization threads */
 854    pipe_barrier_init( &rast->barrier, rast->num_threads );
 855
 856    memset(lp_dummy_tile, 0, sizeof lp_dummy_tile);
 857
 858    return rast;
 859
 860 no_full_scenes:
 861    FREE(rast);
 862 no_rast:
 863    return NULL;
 864 }
 865
 866
 867 /* Shutdown:
 868  */
 869 void lp_rast_destroy( struct lp_rasterizer *rast )
 870 {
 871    unsigned i;
 872
 873    /* Set exit_flag and signal each thread's work_ready semaphore.
 874     * Each thread will be woken up, notice that the exit_flag is set and
 875     * break out of its main loop.  The thread will then exit.
 876     */
 877    rast->exit_flag = TRUE;
 878    for (i = 0; i < rast->num_threads; i++) {
 879       pipe_semaphore_signal(&rast->tasks[i].work_ready);
 880    }
 881
 882    /* Wait for threads to terminate before cleaning up per-thread data */
 883    for (i = 0; i < rast->num_threads; i++) {
 884       pipe_thread_wait(rast->threads[i]);
 885    }
 886
 887    /* Clean up per-thread data */
 888    for (i = 0; i < rast->num_threads; i++) {
 889       pipe_semaphore_destroy(&rast->tasks[i].work_ready);
 890       pipe_semaphore_destroy(&rast->tasks[i].work_done);
 891    }
 892
 893    /* for synchronizing rasterization threads */
 894    pipe_barrier_destroy( &rast->barrier );
 895
 896    lp_scene_queue_destroy(rast->full_scenes);
 897
 898    FREE(rast);
 899 }
 900
 901