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_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             if (scene->fb.cbufs[i]) {
 316                stride[i] = scene->cbufs[i].stride;
 317                color[i] = lp_rast_get_color_block_pointer(task, i, tile_x + x,
 318                                                           tile_y + y, inputs->layer);
 319             }
 320             else {
 321                stride[i] = 0;
 322                color[i] = NULL;
 323             }
 324          }
 325
 326          /* depth buffer */
 327          if (scene->zsbuf.map) {
 328             depth = lp_rast_get_depth_block_pointer(task, tile_x + x,
 329                                                     tile_y + y, inputs->layer);
 330             depth_stride = scene->zsbuf.stride;
 331          }
 332
 333          /* Propagate non-interpolated raster state. */
 334          task->thread_data.raster_state.viewport_index = inputs->viewport_index;
 335
 336          /* run shader on 4x4 block */
 337          BEGIN_JIT_CALL(state, task);
 338          variant->jit_function[RAST_WHOLE]( &state->jit_context,
 339                                             tile_x + x, tile_y + y,
 340                                             inputs->frontfacing,
 341                                             GET_A0(inputs),
 342                                             GET_DADX(inputs),
 343                                             GET_DADY(inputs),
 344                                             color,
 345                                             depth,
 346                                             0xffff,
 347                                             &task->thread_data,
 348                                             stride,
 349                                             depth_stride);
 350          END_JIT_CALL();
 351       }
 352    }
 353 }
 354
 355
 356 /**
 357  * Run the shader on all blocks in a tile.  This is used when a tile is
 358  * completely contained inside a triangle, and the shader is opaque.
 359  * This is a bin command called during bin processing.
 360  */
 361 static void
 362 lp_rast_shade_tile_opaque(struct lp_rasterizer_task *task,
 363                           const union lp_rast_cmd_arg arg)
 364 {
 365    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
 366
 367    assert(task->state);
 368    if (!task->state) {
 369       return;
 370    }
 371
 372    lp_rast_shade_tile(task, arg);
 373 }
 374
 375
 376 /**
 377  * Compute shading for a 4x4 block of pixels inside a triangle.
 378  * This is a bin command called during bin processing.
 379  * \param x  X position of quad in window coords
 380  * \param y  Y position of quad in window coords
 381  */
 382 void
 383 lp_rast_shade_quads_mask(struct lp_rasterizer_task *task,
 384                          const struct lp_rast_shader_inputs *inputs,
 385                          unsigned x, unsigned y,
 386                          unsigned mask)
 387 {
 388    const struct lp_rast_state *state = task->state;
 389    struct lp_fragment_shader_variant *variant = state->variant;
 390    const struct lp_scene *scene = task->scene;
 391    uint8_t *color[PIPE_MAX_COLOR_BUFS];
 392    unsigned stride[PIPE_MAX_COLOR_BUFS];
 393    uint8_t *depth = NULL;
 394    unsigned depth_stride = 0;
 395    unsigned i;
 396
 397    assert(state);
 398
 399    /* Sanity checks */
 400    assert(x < scene->tiles_x * TILE_SIZE);
 401    assert(y < scene->tiles_y * TILE_SIZE);
 402    assert(x % TILE_VECTOR_WIDTH == 0);
 403    assert(y % TILE_VECTOR_HEIGHT == 0);
 404
 405    assert((x % 4) == 0);
 406    assert((y % 4) == 0);
 407
 408    /* color buffer */
 409    for (i = 0; i < scene->fb.nr_cbufs; i++) {
 410       if (scene->fb.cbufs[i]) {
 411          stride[i] = scene->cbufs[i].stride;
 412          color[i] = lp_rast_get_color_block_pointer(task, i, x, y,
 413                                                     inputs->layer);
 414       }
 415       else {
 416          stride[i] = 0;
 417          color[i] = NULL;
 418       }
 419    }
 420
 421    /* depth buffer */
 422    if (scene->zsbuf.map) {
 423       depth_stride = scene->zsbuf.stride;
 424       depth = lp_rast_get_depth_block_pointer(task, x, y, inputs->layer);
 425    }
 426
 427    assert(lp_check_alignment(state->jit_context.u8_blend_color, 16));
 428
 429    /*
 430     * The rasterizer may produce fragments outside our
 431     * allocated 4x4 blocks hence need to filter them out here.
 432     */
 433    if ((x % TILE_SIZE) < task->width && (y % TILE_SIZE) < task->height) {
 434       /* not very accurate would need a popcount on the mask */
 435       /* always count this not worth bothering? */
 436       task->ps_invocations += 1 * variant->ps_inv_multiplier;
 437
 438       /* Propagate non-interpolated raster state. */
 439       task->thread_data.raster_state.viewport_index = inputs->viewport_index;
 440
 441       /* run shader on 4x4 block */
 442       BEGIN_JIT_CALL(state, task);
 443       variant->jit_function[RAST_EDGE_TEST](&state->jit_context,
 444                                             x, y,
 445                                             inputs->frontfacing,
 446                                             GET_A0(inputs),
 447                                             GET_DADX(inputs),
 448                                             GET_DADY(inputs),
 449                                             color,
 450                                             depth,
 451                                             mask,
 452                                             &task->thread_data,
 453                                             stride,
 454                                             depth_stride);
 455       END_JIT_CALL();
 456    }
 457 }
 458
 459
 460
 461 /**
 462  * Begin a new occlusion query.
 463  * This is a bin command put in all bins.
 464  * Called per thread.
 465  */
 466 static void
 467 lp_rast_begin_query(struct lp_rasterizer_task *task,
 468                     const union lp_rast_cmd_arg arg)
 469 {
 470    struct llvmpipe_query *pq = arg.query_obj;
 471
 472    switch (pq->type) {
 473    case PIPE_QUERY_OCCLUSION_COUNTER:
 474    case PIPE_QUERY_OCCLUSION_PREDICATE:
 475       pq->start[task->thread_index] = task->thread_data.vis_counter;
 476       break;
 477    case PIPE_QUERY_PIPELINE_STATISTICS:
 478       pq->start[task->thread_index] = task->ps_invocations;
 479       break;
 480    default:
 481       assert(0);
 482       break;
 483    }
 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
 498    switch (pq->type) {
 499    case PIPE_QUERY_OCCLUSION_COUNTER:
 500    case PIPE_QUERY_OCCLUSION_PREDICATE:
 501       pq->end[task->thread_index] +=
 502          task->thread_data.vis_counter - pq->start[task->thread_index];
 503       pq->start[task->thread_index] = 0;
 504       break;
 505    case PIPE_QUERY_TIMESTAMP:
 506       pq->end[task->thread_index] = os_time_get_nano();
 507       break;
 508    case PIPE_QUERY_PIPELINE_STATISTICS:
 509       pq->end[task->thread_index] +=
 510          task->ps_invocations - pq->start[task->thread_index];
 511       pq->start[task->thread_index] = 0;
 512       break;
 513    default:
 514       assert(0);
 515       break;
 516    }
 517 }
 518
 519
 520 void
 521 lp_rast_set_state(struct lp_rasterizer_task *task,
 522                   const union lp_rast_cmd_arg arg)
 523 {
 524    task->state = arg.state;
 525 }
 526
 527
 528
 529 /**
 530  * Called when we're done writing to a color tile.
 531  */
 532 static void
 533 lp_rast_tile_end(struct lp_rasterizer_task *task)
 534 {
 535    unsigned i;
 536
 537    for (i = 0; i < task->scene->num_active_queries; ++i) {
 538       lp_rast_end_query(task, lp_rast_arg_query(task->scene->active_queries[i]));
 539    }
 540
 541    /* debug */
 542    memset(task->color_tiles, 0, sizeof(task->color_tiles));
 543    task->depth_tile = NULL;
 544
 545    task->bin = NULL;
 546 }
 547
 548 static lp_rast_cmd_func dispatch[LP_RAST_OP_MAX] =
 549 {
 550    lp_rast_clear_color,
 551    lp_rast_clear_zstencil,
 552    lp_rast_triangle_1,
 553    lp_rast_triangle_2,
 554    lp_rast_triangle_3,
 555    lp_rast_triangle_4,
 556    lp_rast_triangle_5,
 557    lp_rast_triangle_6,
 558    lp_rast_triangle_7,
 559    lp_rast_triangle_8,
 560    lp_rast_triangle_3_4,
 561    lp_rast_triangle_3_16,
 562    lp_rast_triangle_4_16,
 563    lp_rast_shade_tile,
 564    lp_rast_shade_tile_opaque,
 565    lp_rast_begin_query,
 566    lp_rast_end_query,
 567    lp_rast_set_state,
 568    lp_rast_triangle_32_1,
 569    lp_rast_triangle_32_2,
 570    lp_rast_triangle_32_3,
 571    lp_rast_triangle_32_4,
 572    lp_rast_triangle_32_5,
 573    lp_rast_triangle_32_6,
 574    lp_rast_triangle_32_7,
 575    lp_rast_triangle_32_8,
 576    lp_rast_triangle_32_3_4,
 577    lp_rast_triangle_32_3_16,
 578    lp_rast_triangle_32_4_16
 579 };
 580
 581
 582 static void
 583 do_rasterize_bin(struct lp_rasterizer_task *task,
 584                  const struct cmd_bin *bin,
 585                  int x, int y)
 586 {
 587    const struct cmd_block *block;
 588    unsigned k;
 589
 590    if (0)
 591       lp_debug_bin(bin, x, y);
 592
 593    for (block = bin->head; block; block = block->next) {
 594       for (k = 0; k < block->count; k++) {
 595          dispatch[block->cmd[k]]( task, block->arg[k] );
 596       }
 597    }
 598 }
 599
 600
 601
 602 /**
 603  * Rasterize commands for a single bin.
 604  * \param x, y  position of the bin's tile in the framebuffer
 605  * Must be called between lp_rast_begin() and lp_rast_end().
 606  * Called per thread.
 607  */
 608 static void
 609 rasterize_bin(struct lp_rasterizer_task *task,
 610               const struct cmd_bin *bin, int x, int y )
 611 {
 612    lp_rast_tile_begin( task, bin, x, y );
 613
 614    do_rasterize_bin(task, bin, x, y);
 615
 616    lp_rast_tile_end(task);
 617
 618
 619    /* Debug/Perf flags:
 620     */
 621    if (bin->head->count == 1) {
 622       if (bin->head->cmd[0] == LP_RAST_OP_SHADE_TILE_OPAQUE)
 623          LP_COUNT(nr_pure_shade_opaque_64);
 624       else if (bin->head->cmd[0] == LP_RAST_OP_SHADE_TILE)
 625          LP_COUNT(nr_pure_shade_64);
 626    }
 627 }
 628
 629
 630 /* An empty bin is one that just loads the contents of the tile and
 631  * stores them again unchanged.  This typically happens when bins have
 632  * been flushed for some reason in the middle of a frame, or when
 633  * incremental updates are being made to a render target.
 634  *
 635  * Try to avoid doing pointless work in this case.
 636  */
 637 static boolean
 638 is_empty_bin( const struct cmd_bin *bin )
 639 {
 640    return bin->head == NULL;
 641 }
 642
 643
 644 /**
 645  * Rasterize/execute all bins within a scene.
 646  * Called per thread.
 647  */
 648 static void
 649 rasterize_scene(struct lp_rasterizer_task *task,
 650                 struct lp_scene *scene)
 651 {
 652    task->scene = scene;
 653
 654    if (!task->rast->no_rast && !scene->discard) {
 655       /* loop over scene bins, rasterize each */
 656       {
 657          struct cmd_bin *bin;
 658          int i, j;
 659
 660          assert(scene);
 661          while ((bin = lp_scene_bin_iter_next(scene, &i, &j))) {
 662             if (!is_empty_bin( bin ))
 663                rasterize_bin(task, bin, i, j);
 664          }
 665       }
 666    }
 667
 668
 669    if (scene->fence) {
 670       lp_fence_signal(scene->fence);
 671    }
 672
 673    task->scene = NULL;
 674 }
 675
 676
 677 /**
 678  * Called by setup module when it has something for us to render.
 679  */
 680 void
 681 lp_rast_queue_scene( struct lp_rasterizer *rast,
 682                      struct lp_scene *scene)
 683 {
 684    LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
 685
 686    if (rast->num_threads == 0) {
 687       /* no threading */
 688       unsigned fpstate = util_fpstate_get();
 689
 690       /* Make sure that denorms are treated like zeros. This is
 691        * the behavior required by D3D10. OpenGL doesn't care.
 692        */
 693       util_fpstate_set_denorms_to_zero(fpstate);
 694
 695       lp_rast_begin( rast, scene );
 696
 697       rasterize_scene( &rast->tasks[0], scene );
 698
 699       lp_rast_end( rast );
 700
 701       util_fpstate_set(fpstate);
 702
 703       rast->curr_scene = NULL;
 704    }
 705    else {
 706       /* threaded rendering! */
 707       unsigned i;
 708
 709       lp_scene_enqueue( rast->full_scenes, scene );
 710
 711       /* signal the threads that there's work to do */
 712       for (i = 0; i < rast->num_threads; i++) {
 713          pipe_semaphore_signal(&rast->tasks[i].work_ready);
 714       }
 715    }
 716
 717    LP_DBG(DEBUG_SETUP, "%s done \n", __FUNCTION__);
 718 }
 719
 720
 721 void
 722 lp_rast_finish( struct lp_rasterizer *rast )
 723 {
 724    if (rast->num_threads == 0) {
 725       /* nothing to do */
 726    }
 727    else {
 728       int i;
 729
 730       /* wait for work to complete */
 731       for (i = 0; i < rast->num_threads; i++) {
 732          pipe_semaphore_wait(&rast->tasks[i].work_done);
 733       }
 734    }
 735 }
 736
 737
 738 /**
 739  * This is the thread's main entrypoint.
 740  * It's a simple loop:
 741  *   1. wait for work
 742  *   2. do work
 743  *   3. signal that we're done
 744  */
 745 static PIPE_THREAD_ROUTINE( thread_function, init_data )
 746 {
 747    struct lp_rasterizer_task *task = (struct lp_rasterizer_task *) init_data;
 748    struct lp_rasterizer *rast = task->rast;
 749    boolean debug = false;
 750    unsigned fpstate = util_fpstate_get();
 751
 752    /* Make sure that denorms are treated like zeros. This is
 753     * the behavior required by D3D10. OpenGL doesn't care.
 754     */
 755    util_fpstate_set_denorms_to_zero(fpstate);
 756
 757    while (1) {
 758       /* wait for work */
 759       if (debug)
 760          debug_printf("thread %d waiting for work\n", task->thread_index);
 761       pipe_semaphore_wait(&task->work_ready);
 762
 763       if (rast->exit_flag)
 764          break;
 765
 766       if (task->thread_index == 0) {
 767          /* thread[0]:
 768           *  - get next scene to rasterize
 769           *  - map the framebuffer surfaces
 770           */
 771          lp_rast_begin( rast,
 772                         lp_scene_dequeue( rast->full_scenes, TRUE ) );
 773       }
 774
 775       /* Wait for all threads to get here so that threads[1+] don't
 776        * get a null rast->curr_scene pointer.
 777        */
 778       pipe_barrier_wait( &rast->barrier );
 779
 780       /* do work */
 781       if (debug)
 782          debug_printf("thread %d doing work\n", task->thread_index);
 783
 784       rasterize_scene(task,
 785                       rast->curr_scene);
 786
 787       /* wait for all threads to finish with this scene */
 788       pipe_barrier_wait( &rast->barrier );
 789
 790       /* XXX: shouldn't be necessary:
 791        */
 792       if (task->thread_index == 0) {
 793          lp_rast_end( rast );
 794       }
 795
 796       /* signal done with work */
 797       if (debug)
 798          debug_printf("thread %d done working\n", task->thread_index);
 799
 800       pipe_semaphore_signal(&task->work_done);
 801    }
 802
 803    return 0;
 804 }
 805
 806
 807 /**
 808  * Initialize semaphores and spawn the threads.
 809  */
 810 static void
 811 create_rast_threads(struct lp_rasterizer *rast)
 812 {
 813    unsigned i;
 814
 815    /* NOTE: if num_threads is zero, we won't use any threads */
 816    for (i = 0; i < rast->num_threads; i++) {
 817       pipe_semaphore_init(&rast->tasks[i].work_ready, 0);
 818       pipe_semaphore_init(&rast->tasks[i].work_done, 0);
 819       rast->threads[i] = pipe_thread_create(thread_function,
 820                                             (void *) &rast->tasks[i]);
 821    }
 822 }
 823
 824
 825
 826 /**
 827  * Create new lp_rasterizer.  If num_threads is zero, don't create any
 828  * new threads, do rendering synchronously.
 829  * \param num_threads  number of rasterizer threads to create
 830  */
 831 struct lp_rasterizer *
 832 lp_rast_create( unsigned num_threads )
 833 {
 834    struct lp_rasterizer *rast;
 835    unsigned i;
 836
 837    rast = CALLOC_STRUCT(lp_rasterizer);
 838    if (!rast) {
 839       goto no_rast;
 840    }
 841
 842    rast->full_scenes = lp_scene_queue_create();
 843    if (!rast->full_scenes) {
 844       goto no_full_scenes;
 845    }
 846
 847    for (i = 0; i < Elements(rast->tasks); i++) {
 848       struct lp_rasterizer_task *task = &rast->tasks[i];
 849       task->rast = rast;
 850       task->thread_index = i;
 851    }
 852
 853    rast->num_threads = num_threads;
 854
 855    rast->no_rast = debug_get_bool_option("LP_NO_RAST", FALSE);
 856
 857    create_rast_threads(rast);
 858
 859    /* for synchronizing rasterization threads */
 860    pipe_barrier_init( &rast->barrier, rast->num_threads );
 861
 862    memset(lp_dummy_tile, 0, sizeof lp_dummy_tile);
 863
 864    return rast;
 865
 866 no_full_scenes:
 867    FREE(rast);
 868 no_rast:
 869    return NULL;
 870 }
 871
 872
 873 /* Shutdown:
 874  */
 875 void lp_rast_destroy( struct lp_rasterizer *rast )
 876 {
 877    unsigned i;
 878
 879    /* Set exit_flag and signal each thread's work_ready semaphore.
 880     * Each thread will be woken up, notice that the exit_flag is set and
 881     * break out of its main loop.  The thread will then exit.
 882     */
 883    rast->exit_flag = TRUE;
 884    for (i = 0; i < rast->num_threads; i++) {
 885       pipe_semaphore_signal(&rast->tasks[i].work_ready);
 886    }
 887
 888    /* Wait for threads to terminate before cleaning up per-thread data */
 889    for (i = 0; i < rast->num_threads; i++) {
 890       pipe_thread_wait(rast->threads[i]);
 891    }
 892
 893    /* Clean up per-thread data */
 894    for (i = 0; i < rast->num_threads; i++) {
 895       pipe_semaphore_destroy(&rast->tasks[i].work_ready);
 896       pipe_semaphore_destroy(&rast->tasks[i].work_done);
 897    }
 898
 899    /* for synchronizing rasterization threads */
 900    pipe_barrier_destroy( &rast->barrier );
 901
 902    lp_scene_queue_destroy(rast->full_scenes);
 903
 904    FREE(rast);
 905 }
 906
 907