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_cpu_detect.h"
  32 #include "util/u_surface.h"
  33
  34 #include "lp_scene_queue.h"
  35 #include "lp_debug.h"
  36 #include "lp_fence.h"
  37 #include "lp_rast.h"
  38 #include "lp_rast_priv.h"
  39 #include "lp_tile_soa.h"
  40 #include "lp_bld_debug.h"
  41 #include "lp_scene.h"
  42
  43
  44 /**
  45  * Begin the rasterization phase.
  46  * Map the framebuffer surfaces.  Initialize the 'rast' state.
  47  */
  48 static boolean
  49 lp_rast_begin( struct lp_rasterizer *rast,
  50                const struct pipe_framebuffer_state *fb,
  51                boolean write_color,
  52                boolean write_zstencil )
  53 {
  54    struct pipe_screen *screen = rast->screen;
  55    struct pipe_surface *cbuf, *zsbuf;
  56
  57    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
  58
  59    util_copy_framebuffer_state(&rast->state.fb, fb);
  60
  61    rast->state.write_zstencil = write_zstencil;
  62    rast->state.write_color = write_color;
  63
  64    rast->check_for_clipped_tiles = (fb->width % TILE_SIZE != 0 ||
  65                                     fb->height % TILE_SIZE != 0);
  66
  67    /* XXX support multiple color buffers here */
  68    cbuf = rast->state.fb.cbufs[0];
  69    if (cbuf) {
  70       rast->cbuf_transfer = screen->get_tex_transfer(rast->screen,
  71                                                      cbuf->texture,
  72                                                      cbuf->face,
  73                                                      cbuf->level,
  74                                                      cbuf->zslice,
  75                                                      PIPE_TRANSFER_READ_WRITE,
  76                                                      0, 0,
  77                                                      fb->width, fb->height);
  78       if (!rast->cbuf_transfer)
  79          return FALSE;
  80
  81       rast->cbuf_map = screen->transfer_map(rast->screen,
  82                                             rast->cbuf_transfer);
  83       if (!rast->cbuf_map)
  84          return FALSE;
  85    }
  86
  87    zsbuf = rast->state.fb.zsbuf;
  88    if (zsbuf) {
  89       rast->zsbuf_transfer = screen->get_tex_transfer(rast->screen,
  90                                                       zsbuf->texture,
  91                                                       zsbuf->face,
  92                                                       zsbuf->level,
  93                                                       zsbuf->zslice,
  94                                                       PIPE_TRANSFER_READ_WRITE,
  95                                                       0, 0,
  96                                                       fb->width, fb->height);
  97       if (!rast->zsbuf_transfer)
  98          return FALSE;
  99
 100       rast->zsbuf_map = screen->transfer_map(rast->screen,
 101                                             rast->zsbuf_transfer);
 102       if (!rast->zsbuf_map)
 103          return FALSE;
 104    }
 105
 106    return TRUE;
 107 }
 108
 109
 110 /**
 111  * Finish the rasterization phase.
 112  * Unmap framebuffer surfaces.
 113  */
 114 static void
 115 lp_rast_end( struct lp_rasterizer *rast )
 116 {
 117    struct pipe_screen *screen = rast->screen;
 118
 119    if (rast->cbuf_map)
 120       screen->transfer_unmap(screen, rast->cbuf_transfer);
 121
 122    if (rast->zsbuf_map)
 123       screen->transfer_unmap(screen, rast->zsbuf_transfer);
 124
 125    if (rast->cbuf_transfer)
 126       screen->tex_transfer_destroy(rast->cbuf_transfer);
 127
 128    if (rast->zsbuf_transfer)
 129       screen->tex_transfer_destroy(rast->zsbuf_transfer);
 130
 131    rast->cbuf_transfer = NULL;
 132    rast->zsbuf_transfer = NULL;
 133    rast->cbuf_map = NULL;
 134    rast->zsbuf_map = NULL;
 135 }
 136
 137
 138 /**
 139  * Begining rasterization of a tile.
 140  * \param x  window X position of the tile, in pixels
 141  * \param y  window Y position of the tile, in pixels
 142  */
 143 static void
 144 lp_rast_start_tile( struct lp_rasterizer *rast,
 145                     unsigned thread_index,
 146                     unsigned x, unsigned y )
 147 {
 148    LP_DBG(DEBUG_RAST, "%s %d,%d\n", __FUNCTION__, x, y);
 149
 150    rast->tasks[thread_index].x = x;
 151    rast->tasks[thread_index].y = y;
 152 }
 153
 154
 155 /**
 156  * Clear the rasterizer's current color tile.
 157  * This is a bin command called during bin processing.
 158  */
 159 void lp_rast_clear_color( struct lp_rasterizer *rast,
 160                           unsigned thread_index,
 161                           const union lp_rast_cmd_arg arg )
 162 {
 163    const uint8_t *clear_color = arg.clear_color;
 164    uint8_t *color_tile = rast->tasks[thread_index].tile.color;
 165
 166    LP_DBG(DEBUG_RAST, "%s 0x%x,0x%x,0x%x,0x%x\n", __FUNCTION__,
 167               clear_color[0],
 168               clear_color[1],
 169               clear_color[2],
 170               clear_color[3]);
 171
 172    if (clear_color[0] == clear_color[1] &&
 173        clear_color[1] == clear_color[2] &&
 174        clear_color[2] == clear_color[3]) {
 175       memset(color_tile, clear_color[0], TILE_SIZE * TILE_SIZE * 4);
 176    }
 177    else {
 178       unsigned x, y, chan;
 179       for (y = 0; y < TILE_SIZE; y++)
 180          for (x = 0; x < TILE_SIZE; x++)
 181             for (chan = 0; chan < 4; ++chan)
 182                TILE_PIXEL(color_tile, x, y, chan) = clear_color[chan];
 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 void lp_rast_clear_zstencil( struct lp_rasterizer *rast,
 192                              unsigned thread_index,
 193                              const union lp_rast_cmd_arg arg)
 194 {
 195    unsigned i, j;
 196    uint32_t *depth_tile = rast->tasks[thread_index].tile.depth;
 197
 198    LP_DBG(DEBUG_RAST, "%s 0x%x\n", __FUNCTION__, arg.clear_zstencil);
 199
 200    for (i = 0; i < TILE_SIZE; i++)
 201       for (j = 0; j < TILE_SIZE; j++)
 202          depth_tile[i*TILE_SIZE + j] = arg.clear_zstencil;
 203 }
 204
 205
 206 /**
 207  * Load tile color from the framebuffer surface.
 208  * This is a bin command called during bin processing.
 209  */
 210 void lp_rast_load_color( struct lp_rasterizer *rast,
 211                          unsigned thread_index,
 212                          const union lp_rast_cmd_arg arg)
 213 {
 214    struct lp_rasterizer_task *task = &rast->tasks[thread_index];
 215    const unsigned x = task->x;
 216    const unsigned y = task->y;
 217    int w = TILE_SIZE;
 218    int h = TILE_SIZE;
 219
 220    LP_DBG(DEBUG_RAST, "%s at %u, %u\n", __FUNCTION__, x, y);
 221
 222    if (x + w > rast->state.fb.width)
 223       w -= x + w - rast->state.fb.width;
 224
 225    if (y + h > rast->state.fb.height)
 226       h -= y + h - rast->state.fb.height;
 227
 228    assert(w >= 0);
 229    assert(h >= 0);
 230    assert(w <= TILE_SIZE);
 231    assert(h <= TILE_SIZE);
 232
 233    lp_tile_read_4ub(rast->cbuf_transfer->format,
 234                      rast->tasks[thread_index].tile.color,
 235                      rast->cbuf_map,
 236                      rast->cbuf_transfer->stride,
 237                      x, y,
 238                      w, h);
 239 }
 240
 241
 242 /**
 243  * Load tile z/stencil from the framebuffer surface.
 244  * This is a bin command called during bin processing.
 245  */
 246 void lp_rast_load_zstencil( struct lp_rasterizer *rast,
 247                             unsigned thread_index,
 248                             const union lp_rast_cmd_arg arg )
 249 {
 250    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
 251
 252    /* call u_tile func to load depth (and stencil?) from surface */
 253 }
 254
 255
 256 void lp_rast_set_state( struct lp_rasterizer *rast,
 257                         unsigned thread_index,
 258                         const union lp_rast_cmd_arg arg )
 259 {
 260    const struct lp_rast_state *state = arg.set_state;
 261
 262    LP_DBG(DEBUG_RAST, "%s %p\n", __FUNCTION__, (void *) state);
 263
 264    /* just set the current state pointer for this rasterizer */
 265    rast->tasks[thread_index].current_state = state;
 266 }
 267
 268
 269
 270 /* Within a tile:
 271  */
 272
 273 /**
 274  * Run the shader on all blocks in a tile.  This is used when a tile is
 275  * completely contained inside a triangle.
 276  * This is a bin command called during bin processing.
 277  */
 278 void lp_rast_shade_tile( struct lp_rasterizer *rast,
 279                          unsigned thread_index,
 280                          const union lp_rast_cmd_arg arg )
 281 {
 282    /* Set c1,c2,c3 to large values so the in/out test always passes */
 283    const int32_t c1 = INT_MIN, c2 = INT_MIN, c3 = INT_MIN;
 284    const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
 285    const unsigned tile_x = rast->tasks[thread_index].x;
 286    const unsigned tile_y = rast->tasks[thread_index].y;
 287    unsigned x, y;
 288
 289    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
 290
 291    /* Use the existing preference for 4x4 (four quads) shading:
 292     */
 293    for (y = 0; y < TILE_SIZE; y += 4)
 294       for (x = 0; x < TILE_SIZE; x += 4)
 295          lp_rast_shade_quads( rast,
 296                               thread_index,
 297                               inputs,
 298                               tile_x + x,
 299                               tile_y + y,
 300                               c1, c2, c3);
 301 }
 302
 303
 304 /**
 305  * Compute shading for a 4x4 block of pixels.
 306  * This is a bin command called during bin processing.
 307  */
 308 void lp_rast_shade_quads( struct lp_rasterizer *rast,
 309                           unsigned thread_index,
 310                           const struct lp_rast_shader_inputs *inputs,
 311                           unsigned x, unsigned y,
 312                           int32_t c1, int32_t c2, int32_t c3)
 313 {
 314    const struct lp_rast_state *state = rast->tasks[thread_index].current_state;
 315    struct lp_rast_tile *tile = &rast->tasks[thread_index].tile;
 316    void *color;
 317    void *depth;
 318    unsigned ix, iy;
 319    int block_offset;
 320
 321 #ifdef DEBUG
 322    assert(state);
 323
 324    /* Sanity checks */
 325    assert(x % TILE_VECTOR_WIDTH == 0);
 326    assert(y % TILE_VECTOR_HEIGHT == 0);
 327
 328    assert((x % 4) == 0);
 329    assert((y % 4) == 0);
 330 #endif
 331
 332    ix = x % TILE_SIZE;
 333    iy = y % TILE_SIZE;
 334
 335    /* offset of the 16x16 pixel block within the tile */
 336    block_offset = ((iy/4)*(16*16) + (ix/4)*16);
 337
 338    /* color buffer */
 339    color = tile->color + 4 * block_offset;
 340
 341    /* depth buffer */
 342    depth = tile->depth + block_offset;
 343
 344 #ifdef DEBUG
 345    assert(lp_check_alignment(depth, 16));
 346    assert(lp_check_alignment(color, 16));
 347    assert(lp_check_alignment(state->jit_context.blend_color, 16));
 348
 349    assert(lp_check_alignment(inputs->step[0], 16));
 350    assert(lp_check_alignment(inputs->step[1], 16));
 351    assert(lp_check_alignment(inputs->step[2], 16));
 352 #endif
 353
 354    /* run shader */
 355    state->jit_function( &state->jit_context,
 356                         x, y,
 357                         inputs->a0,
 358                         inputs->dadx,
 359                         inputs->dady,
 360                         color,
 361                         depth,
 362                         c1, c2, c3,
 363                         inputs->step[0], inputs->step[1], inputs->step[2]
 364                         );
 365 }
 366
 367
 368 /* End of tile:
 369  */
 370
 371
 372 /**
 373  * Write the rasterizer's color tile to the framebuffer.
 374  */
 375 static void lp_rast_store_color( struct lp_rasterizer *rast,
 376                                  unsigned thread_index)
 377 {
 378    const unsigned x = rast->tasks[thread_index].x;
 379    const unsigned y = rast->tasks[thread_index].y;
 380    int w = TILE_SIZE;
 381    int h = TILE_SIZE;
 382
 383    if (x + w > rast->state.fb.width)
 384       w -= x + w - rast->state.fb.width;
 385
 386    if (y + h > rast->state.fb.height)
 387       h -= y + h - rast->state.fb.height;
 388
 389    assert(w >= 0);
 390    assert(h >= 0);
 391    assert(w <= TILE_SIZE);
 392    assert(h <= TILE_SIZE);
 393
 394    LP_DBG(DEBUG_RAST, "%s [%u] %d,%d %dx%d\n", __FUNCTION__,
 395           thread_index, x, y, w, h);
 396
 397    lp_tile_write_4ub(rast->cbuf_transfer->format,
 398                      rast->tasks[thread_index].tile.color,
 399                      rast->cbuf_map,
 400                      rast->cbuf_transfer->stride,
 401                      x, y,
 402                      w, h);
 403 }
 404
 405
 406 static void
 407 lp_tile_write_z32(const uint32_t *src, uint8_t *dst, unsigned dst_stride,
 408                   unsigned x0, unsigned y0, unsigned w, unsigned h)
 409 {
 410    unsigned x, y;
 411    uint8_t *dst_row = dst + y0*dst_stride;
 412    for (y = 0; y < h; ++y) {
 413       uint32_t *dst_pixel = (uint32_t *)(dst_row + x0*4);
 414       for (x = 0; x < w; ++x) {
 415          *dst_pixel++ = *src++;
 416       }
 417       dst_row += dst_stride;
 418    }
 419 }
 420
 421 /**
 422  * Write the rasterizer's z/stencil tile to the framebuffer.
 423  */
 424 static void lp_rast_store_zstencil( struct lp_rasterizer *rast,
 425                                     unsigned thread_index )
 426 {
 427    const unsigned x = rast->tasks[thread_index].x;
 428    const unsigned y = rast->tasks[thread_index].y;
 429    unsigned w = TILE_SIZE;
 430    unsigned h = TILE_SIZE;
 431
 432    if (x + w > rast->state.fb.width)
 433       w -= x + w - rast->state.fb.width;
 434
 435    if (y + h > rast->state.fb.height)
 436       h -= y + h - rast->state.fb.height;
 437
 438    LP_DBG(DEBUG_RAST, "%s %d,%d %dx%d\n", __FUNCTION__, x, y, w, h);
 439
 440    assert(rast->zsbuf_transfer->format == PIPE_FORMAT_Z32_UNORM);
 441    lp_tile_write_z32(rast->tasks[thread_index].tile.depth,
 442                      rast->zsbuf_map,
 443                      rast->zsbuf_transfer->stride,
 444                      x, y, w, h);
 445 }
 446
 447
 448 /**
 449  * Write the rasterizer's tiles to the framebuffer.
 450  */
 451 static void
 452 lp_rast_end_tile( struct lp_rasterizer *rast,
 453                   unsigned thread_index )
 454 {
 455    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
 456
 457    if (rast->state.write_color)
 458       lp_rast_store_color(rast, thread_index);
 459
 460    if (rast->state.write_zstencil)
 461       lp_rast_store_zstencil(rast, thread_index);
 462 }
 463
 464
 465 /**
 466  * Signal on a fence.  This is called during bin execution/rasterization.
 467  * Called per thread.
 468  */
 469 void lp_rast_fence( struct lp_rasterizer *rast,
 470                     unsigned thread_index,
 471                     const union lp_rast_cmd_arg arg )
 472 {
 473    struct lp_fence *fence = arg.fence;
 474
 475    pipe_mutex_lock( fence->mutex );
 476
 477    fence->count++;
 478    assert(fence->count <= fence->rank);
 479
 480    LP_DBG(DEBUG_RAST, "%s count=%u rank=%u\n", __FUNCTION__,
 481           fence->count, fence->rank);
 482
 483    pipe_condvar_signal( fence->signalled );
 484
 485    pipe_mutex_unlock( fence->mutex );
 486 }
 487
 488
 489 /**
 490  * When all the threads are done rasterizing a scene, one thread will
 491  * call this function to reset the scene and put it onto the empty queue.
 492  */
 493 static void
 494 release_scene( struct lp_rasterizer *rast,
 495                struct lp_scene *scene )
 496 {
 497    util_unreference_framebuffer_state( &scene->fb );
 498
 499    lp_scene_reset( scene );
 500    lp_scene_enqueue( rast->empty_scenes, scene );
 501    rast->curr_scene = NULL;
 502 }
 503
 504
 505 /**
 506  * Rasterize commands for a single bin.
 507  * \param x, y  position of the bin's tile in the framebuffer
 508  * Must be called between lp_rast_begin() and lp_rast_end().
 509  * Called per thread.
 510  */
 511 static void
 512 rasterize_bin( struct lp_rasterizer *rast,
 513                unsigned thread_index,
 514                const struct cmd_bin *bin,
 515                int x, int y)
 516 {
 517    const struct cmd_block_list *commands = &bin->commands;
 518    struct cmd_block *block;
 519    unsigned k;
 520
 521    lp_rast_start_tile( rast, thread_index, x, y );
 522
 523    /* simply execute each of the commands in the block list */
 524    for (block = commands->head; block; block = block->next) {
 525       for (k = 0; k < block->count; k++) {
 526          block->cmd[k]( rast, thread_index, block->arg[k] );
 527       }
 528    }
 529
 530    lp_rast_end_tile( rast, thread_index );
 531 }
 532
 533
 534 /**
 535  * Rasterize/execute all bins within a scene.
 536  * Called per thread.
 537  */
 538 static void
 539 rasterize_scene( struct lp_rasterizer *rast,
 540                 unsigned thread_index,
 541                 struct lp_scene *scene,
 542                 bool write_depth )
 543 {
 544    /* loop over scene bins, rasterize each */
 545 #if 0
 546    {
 547       unsigned i, j;
 548       for (i = 0; i < scene->tiles_x; i++) {
 549          for (j = 0; j < scene->tiles_y; j++) {
 550             struct cmd_bin *bin = lp_get_bin(scene, i, j);
 551             rasterize_bin( rast, thread_index,
 552                            bin, i * TILE_SIZE, j * TILE_SIZE );
 553          }
 554       }
 555    }
 556 #else
 557    {
 558       struct cmd_bin *bin;
 559       int x, y;
 560
 561       assert(scene);
 562       while ((bin = lp_scene_bin_iter_next(scene, &x, &y))) {
 563          rasterize_bin( rast, thread_index, bin, x * TILE_SIZE, y * TILE_SIZE);
 564       }
 565    }
 566 #endif
 567 }
 568
 569
 570 /**
 571  * Called by setup module when it has something for us to render.
 572  */
 573 void
 574 lp_rasterize_scene( struct lp_rasterizer *rast,
 575                    struct lp_scene *scene,
 576                    const struct pipe_framebuffer_state *fb,
 577                    bool write_depth )
 578 {
 579    boolean debug = false;
 580
 581    LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
 582
 583    if (debug) {
 584       unsigned x, y;
 585       printf("rasterize scene:\n");
 586       printf("  data size: %u\n", lp_scene_data_size(scene));
 587       for (y = 0; y < scene->tiles_y; y++) {
 588          for (x = 0; x < scene->tiles_x; x++) {
 589             printf("  bin %u, %u size: %u\n", x, y,
 590                    lp_scene_bin_size(scene, x, y));
 591          }
 592       }
 593    }
 594
 595    /* save framebuffer state in the bin */
 596    util_copy_framebuffer_state(&scene->fb, fb);
 597    scene->write_depth = write_depth;
 598
 599    if (rast->num_threads == 0) {
 600       /* no threading */
 601
 602       lp_rast_begin( rast, fb,
 603                      fb->cbufs[0]!= NULL,
 604                      fb->zsbuf != NULL && write_depth );
 605
 606       lp_scene_bin_iter_begin( scene );
 607       rasterize_scene( rast, 0, scene, write_depth );
 608
 609       release_scene( rast, scene );
 610
 611       lp_rast_end( rast );
 612    }
 613    else {
 614       /* threaded rendering! */
 615       unsigned i;
 616
 617       lp_scene_enqueue( rast->full_scenes, scene );
 618
 619       /* signal the threads that there's work to do */
 620       for (i = 0; i < rast->num_threads; i++) {
 621          pipe_semaphore_signal(&rast->tasks[i].work_ready);
 622       }
 623
 624       /* wait for work to complete */
 625       for (i = 0; i < rast->num_threads; i++) {
 626          pipe_semaphore_wait(&rast->tasks[i].work_done);
 627       }
 628    }
 629
 630    LP_DBG(DEBUG_SETUP, "%s done \n", __FUNCTION__);
 631 }
 632
 633
 634 /**
 635  * This is the thread's main entrypoint.
 636  * It's a simple loop:
 637  *   1. wait for work
 638  *   2. do work
 639  *   3. signal that we're done
 640  */
 641 static void *
 642 thread_func( void *init_data )
 643 {
 644    struct lp_rasterizer_task *task = (struct lp_rasterizer_task *) init_data;
 645    struct lp_rasterizer *rast = task->rast;
 646    boolean debug = false;
 647
 648    while (1) {
 649       /* wait for work */
 650       if (debug)
 651          debug_printf("thread %d waiting for work\n", task->thread_index);
 652       pipe_semaphore_wait(&task->work_ready);
 653
 654       if (task->thread_index == 0) {
 655          /* thread[0]:
 656           *  - get next scene to rasterize
 657           *  - map the framebuffer surfaces
 658           */
 659          const struct pipe_framebuffer_state *fb;
 660          boolean write_depth;
 661
 662          rast->curr_scene = lp_scene_dequeue( rast->full_scenes );
 663
 664          lp_scene_bin_iter_begin( rast->curr_scene );
 665
 666          fb = &rast->curr_scene->fb;
 667          write_depth = rast->curr_scene->write_depth;
 668
 669          lp_rast_begin( rast, fb,
 670                         fb->cbufs[0] != NULL,
 671                         fb->zsbuf != NULL && write_depth );
 672       }
 673
 674       /* Wait for all threads to get here so that threads[1+] don't
 675        * get a null rast->curr_scene pointer.
 676        */
 677       pipe_barrier_wait( &rast->barrier );
 678
 679       /* do work */
 680       if (debug)
 681          debug_printf("thread %d doing work\n", task->thread_index);
 682       rasterize_scene(rast,
 683                      task->thread_index,
 684                      rast->curr_scene,
 685                      rast->curr_scene->write_depth);
 686
 687       /* wait for all threads to finish with this scene */
 688       pipe_barrier_wait( &rast->barrier );
 689
 690       if (task->thread_index == 0) {
 691          /* thread[0]:
 692           * - release the scene object
 693           * - unmap the framebuffer surfaces
 694           */
 695          release_scene( rast, rast->curr_scene );
 696          lp_rast_end( rast );
 697       }
 698
 699       /* signal done with work */
 700       if (debug)
 701          debug_printf("thread %d done working\n", task->thread_index);
 702       pipe_semaphore_signal(&task->work_done);
 703    }
 704
 705    return NULL;
 706 }
 707
 708
 709 /**
 710  * Initialize semaphores and spawn the threads.
 711  */
 712 static void
 713 create_rast_threads(struct lp_rasterizer *rast)
 714 {
 715    unsigned i;
 716
 717    rast->num_threads = util_cpu_caps.nr_cpus;
 718    rast->num_threads = debug_get_num_option("LP_NUM_THREADS", rast->num_threads);
 719    rast->num_threads = MIN2(rast->num_threads, MAX_THREADS);
 720
 721    /* NOTE: if num_threads is zero, we won't use any threads */
 722    for (i = 0; i < rast->num_threads; i++) {
 723       pipe_semaphore_init(&rast->tasks[i].work_ready, 0);
 724       pipe_semaphore_init(&rast->tasks[i].work_done, 0);
 725       rast->threads[i] = pipe_thread_create(thread_func,
 726                                             (void *) &rast->tasks[i]);
 727    }
 728 }
 729
 730
 731
 732 /**
 733  * Create new lp_rasterizer.
 734  * \param empty  the queue to put empty scenes on after we've finished
 735  *               processing them.
 736  */
 737 struct lp_rasterizer *
 738 lp_rast_create( struct pipe_screen *screen, struct lp_scene_queue *empty )
 739 {
 740    struct lp_rasterizer *rast;
 741    unsigned i;
 742
 743    rast = CALLOC_STRUCT(lp_rasterizer);
 744    if(!rast)
 745       return NULL;
 746
 747    rast->screen = screen;
 748
 749    rast->empty_scenes = empty;
 750    rast->full_scenes = lp_scene_queue_create();
 751
 752    for (i = 0; i < Elements(rast->tasks); i++) {
 753       rast->tasks[i].tile.color = align_malloc( TILE_SIZE*TILE_SIZE*4, 16 );
 754       rast->tasks[i].tile.depth = align_malloc( TILE_SIZE*TILE_SIZE*4, 16 );
 755       rast->tasks[i].rast = rast;
 756       rast->tasks[i].thread_index = i;
 757    }
 758
 759    create_rast_threads(rast);
 760
 761    /* for synchronizing rasterization threads */
 762    pipe_barrier_init( &rast->barrier, rast->num_threads );
 763
 764    return rast;
 765 }
 766
 767
 768 /* Shutdown:
 769  */
 770 void lp_rast_destroy( struct lp_rasterizer *rast )
 771 {
 772    unsigned i;
 773
 774    util_unreference_framebuffer_state(&rast->state.fb);
 775
 776    for (i = 0; i < Elements(rast->tasks); i++) {
 777       align_free(rast->tasks[i].tile.depth);
 778       align_free(rast->tasks[i].tile.color);
 779    }
 780
 781    /* for synchronizing rasterization threads */
 782    pipe_barrier_destroy( &rast->barrier );
 783
 784    FREE(rast);
 785 }
 786
 787
 788 /** Return number of rasterization threads */
 789 unsigned
 790 lp_rast_get_num_threads( struct lp_rasterizer *rast )
 791 {
 792    return rast->num_threads;
 793 }