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_perf.h"
  38 #include "lp_rast.h"
  39 #include "lp_rast_priv.h"
  40 #include "lp_tile_soa.h"
  41 #include "gallivm/lp_bld_debug.h"
  42 #include "lp_scene.h"
  43
  44
  45 /**
  46  * Begin the rasterization phase.
  47  * Map the framebuffer surfaces.  Initialize the 'rast' state.
  48  */
  49 static boolean
  50 lp_rast_begin( struct lp_rasterizer *rast,
  51                const struct pipe_framebuffer_state *fb,
  52                boolean write_color,
  53                boolean write_zstencil )
  54 {
  55    struct pipe_screen *screen = rast->screen;
  56    struct pipe_surface *cbuf, *zsbuf;
  57    int i;
  58
  59    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
  60
  61    util_copy_framebuffer_state(&rast->state.fb, fb);
  62
  63    rast->state.write_zstencil = write_zstencil;
  64    rast->state.write_color = write_color;
  65
  66    rast->check_for_clipped_tiles = (fb->width % TILE_SIZE != 0 ||
  67                                     fb->height % TILE_SIZE != 0);
  68
  69
  70    for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
  71       cbuf = rast->state.fb.cbufs[i];
  72       if (cbuf) {
  73          rast->cbuf_transfer[i] = screen->get_tex_transfer(rast->screen,
  74                                                            cbuf->texture,
  75                                                            cbuf->face,
  76                                                            cbuf->level,
  77                                                            cbuf->zslice,
  78                                                            PIPE_TRANSFER_READ_WRITE,
  79                                                            0, 0,
  80                                                            cbuf->width,
  81                                                            cbuf->height);
  82          if (!rast->cbuf_transfer[i])
  83             goto fail;
  84
  85          rast->cbuf_map[i] = screen->transfer_map(rast->screen,
  86                                                   rast->cbuf_transfer[i]);
  87          if (!rast->cbuf_map[i])
  88             goto fail;
  89       }
  90    }
  91
  92    zsbuf = rast->state.fb.zsbuf;
  93    if (zsbuf) {
  94       rast->zsbuf_transfer = screen->get_tex_transfer(rast->screen,
  95                                                       zsbuf->texture,
  96                                                       zsbuf->face,
  97                                                       zsbuf->level,
  98                                                       zsbuf->zslice,
  99                                                       PIPE_TRANSFER_READ_WRITE,
 100                                                       0, 0,
 101                                                       zsbuf->width,
 102                                                       zsbuf->height);
 103       if (!rast->zsbuf_transfer)
 104          goto fail;
 105
 106       rast->zsbuf_map = screen->transfer_map(rast->screen,
 107                                             rast->zsbuf_transfer);
 108       if (!rast->zsbuf_map)
 109          goto fail;
 110    }
 111
 112    return TRUE;
 113
 114 fail:
 115    /* Unmap and release transfers?
 116     */
 117    return FALSE;
 118 }
 119
 120
 121 /**
 122  * Finish the rasterization phase.
 123  * Unmap framebuffer surfaces.
 124  */
 125 static void
 126 lp_rast_end( struct lp_rasterizer *rast )
 127 {
 128    struct pipe_screen *screen = rast->screen;
 129    unsigned i;
 130
 131    for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
 132       if (rast->cbuf_map[i])
 133          screen->transfer_unmap(screen, rast->cbuf_transfer[i]);
 134
 135       if (rast->cbuf_transfer[i])
 136          screen->tex_transfer_destroy(rast->cbuf_transfer[i]);
 137
 138       rast->cbuf_transfer[i] = NULL;
 139       rast->cbuf_map[i] = NULL;
 140    }
 141
 142    if (rast->zsbuf_map)
 143       screen->transfer_unmap(screen, rast->zsbuf_transfer);
 144
 145    if (rast->zsbuf_transfer)
 146       screen->tex_transfer_destroy(rast->zsbuf_transfer);
 147
 148    rast->zsbuf_transfer = NULL;
 149    rast->zsbuf_map = NULL;
 150 }
 151
 152
 153 /**
 154  * Begining rasterization of a tile.
 155  * \param x  window X position of the tile, in pixels
 156  * \param y  window Y position of the tile, in pixels
 157  */
 158 static void
 159 lp_rast_start_tile( struct lp_rasterizer *rast,
 160                     unsigned thread_index,
 161                     unsigned x, unsigned y )
 162 {
 163    LP_DBG(DEBUG_RAST, "%s %d,%d\n", __FUNCTION__, x, y);
 164
 165    rast->tasks[thread_index].x = x;
 166    rast->tasks[thread_index].y = y;
 167 }
 168
 169
 170 /**
 171  * Clear the rasterizer's current color tile.
 172  * This is a bin command called during bin processing.
 173  */
 174 void lp_rast_clear_color( struct lp_rasterizer *rast,
 175                           unsigned thread_index,
 176                           const union lp_rast_cmd_arg arg )
 177 {
 178    const uint8_t *clear_color = arg.clear_color;
 179    uint8_t **color_tile = rast->tasks[thread_index].tile.color;
 180    unsigned i;
 181
 182    LP_DBG(DEBUG_RAST, "%s 0x%x,0x%x,0x%x,0x%x\n", __FUNCTION__,
 183               clear_color[0],
 184               clear_color[1],
 185               clear_color[2],
 186               clear_color[3]);
 187
 188    if (clear_color[0] == clear_color[1] &&
 189        clear_color[1] == clear_color[2] &&
 190        clear_color[2] == clear_color[3]) {
 191       /* clear to grayscale value {x, x, x, x} */
 192       for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
 193          memset(color_tile[i], clear_color[0], TILE_SIZE * TILE_SIZE * 4);
 194       }
 195    }
 196    else {
 197       /* Non-gray color.
 198        * Note: if the swizzled tile layout changes (see TILE_PIXEL) this code
 199        * will need to change.  It'll be pretty obvious when clearing no longer
 200        * works.
 201        */
 202       const unsigned chunk = TILE_SIZE / 4;
 203       for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
 204          uint8_t *c = color_tile[i];
 205          unsigned j;
 206          for (j = 0; j < 4 * TILE_SIZE; j++) {
 207             memset(c, clear_color[0], chunk);
 208             c += chunk;
 209             memset(c, clear_color[1], chunk);
 210             c += chunk;
 211             memset(c, clear_color[2], chunk);
 212             c += chunk;
 213             memset(c, clear_color[3], chunk);
 214             c += chunk;
 215          }
 216          assert(c - color_tile[i] == TILE_SIZE * TILE_SIZE * 4);
 217       }
 218    }
 219
 220    LP_COUNT(nr_color_tile_clear);
 221 }
 222
 223
 224 /**
 225  * Clear the rasterizer's current z/stencil tile.
 226  * This is a bin command called during bin processing.
 227  */
 228 void lp_rast_clear_zstencil( struct lp_rasterizer *rast,
 229                              unsigned thread_index,
 230                              const union lp_rast_cmd_arg arg)
 231 {
 232    struct lp_rasterizer_task *task = &rast->tasks[thread_index];
 233    const unsigned tile_x = task->x;
 234    const unsigned tile_y = task->y;
 235    const unsigned height = TILE_SIZE/TILE_VECTOR_HEIGHT;
 236    const unsigned width = TILE_SIZE*TILE_VECTOR_HEIGHT;
 237    unsigned block_size = util_format_get_blocksize(rast->zsbuf_transfer->texture->format);
 238    uint8_t *dst;
 239    unsigned dst_stride = rast->zsbuf_transfer->stride*TILE_VECTOR_HEIGHT;
 240    unsigned i, j;
 241
 242    LP_DBG(DEBUG_RAST, "%s 0x%x\n", __FUNCTION__, arg.clear_zstencil);
 243
 244    assert(rast->zsbuf_map);
 245    if (!rast->zsbuf_map)
 246       return;
 247
 248    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
 249
 250    /*
 251     * Clear the aera of the swizzled depth/depth buffer matching this tile, in
 252     * stripes of TILE_VECTOR_HEIGHT x TILE_SIZE at a time.
 253     *
 254     * The swizzled depth format is such that the depths for
 255     * TILE_VECTOR_HEIGHT x TILE_VECTOR_WIDTH pixels have consecutive offsets.
 256     */
 257
 258    dst = lp_rast_depth_pointer(rast, tile_x, tile_y);
 259
 260    switch (block_size) {
 261    case 1:
 262       memset(dst, (uint8_t) arg.clear_zstencil, height * width);
 263       break;
 264    case 2:
 265       for (i = 0; i < height; i++) {
 266          uint16_t *row = (uint16_t *)dst;
 267          for (j = 0; j < width; j++)
 268             *row++ = (uint16_t) arg.clear_zstencil;
 269          dst += dst_stride;
 270       }
 271       break;
 272    case 4:
 273       for (i = 0; i < height; i++) {
 274          uint32_t *row = (uint32_t *)dst;
 275          for (j = 0; j < width; j++)
 276             *row++ = arg.clear_zstencil;
 277          dst += dst_stride;
 278       }
 279       break;
 280    default:
 281          assert(0);
 282          break;
 283    }
 284 }
 285
 286
 287 /**
 288  * Load tile color from the framebuffer surface.
 289  * This is a bin command called during bin processing.
 290  */
 291 void lp_rast_load_color( struct lp_rasterizer *rast,
 292                          unsigned thread_index,
 293                          const union lp_rast_cmd_arg arg)
 294 {
 295    struct lp_rasterizer_task *task = &rast->tasks[thread_index];
 296    const unsigned x = task->x;
 297    const unsigned y = task->y;
 298    unsigned i;
 299
 300    LP_DBG(DEBUG_RAST, "%s at %u, %u\n", __FUNCTION__, x, y);
 301
 302    for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
 303       struct pipe_transfer *transfer = rast->cbuf_transfer[i];
 304       int w = TILE_SIZE;
 305       int h = TILE_SIZE;
 306
 307       if (x >= transfer->width)
 308          continue;
 309
 310       if (y >= transfer->height)
 311          continue;
 312
 313       assert(w >= 0);
 314       assert(h >= 0);
 315       assert(w <= TILE_SIZE);
 316       assert(h <= TILE_SIZE);
 317
 318       lp_tile_read_4ub(transfer->texture->format,
 319                        task->tile.color[i],
 320                        rast->cbuf_map[i],
 321                        transfer->stride,
 322                        x, y,
 323                        w, h);
 324
 325       LP_COUNT(nr_color_tile_load);
 326    }
 327 }
 328
 329
 330 void lp_rast_set_state( struct lp_rasterizer *rast,
 331                         unsigned thread_index,
 332                         const union lp_rast_cmd_arg arg )
 333 {
 334    const struct lp_rast_state *state = arg.set_state;
 335
 336    LP_DBG(DEBUG_RAST, "%s %p\n", __FUNCTION__, (void *) state);
 337
 338    /* just set the current state pointer for this rasterizer */
 339    rast->tasks[thread_index].current_state = state;
 340 }
 341
 342
 343
 344 /**
 345  * Run the shader on all blocks in a tile.  This is used when a tile is
 346  * completely contained inside a triangle.
 347  * This is a bin command called during bin processing.
 348  */
 349 void lp_rast_shade_tile( struct lp_rasterizer *rast,
 350                          unsigned thread_index,
 351                          const union lp_rast_cmd_arg arg )
 352 {
 353    struct lp_rasterizer_task *task = &rast->tasks[thread_index];
 354    const struct lp_rast_state *state = task->current_state;
 355    struct lp_rast_tile *tile = &task->tile;
 356    const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
 357    const unsigned tile_x = task->x;
 358    const unsigned tile_y = task->y;
 359    unsigned x, y;
 360
 361    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
 362
 363    /* render the whole 64x64 tile in 4x4 chunks */
 364    for (y = 0; y < TILE_SIZE; y += 4){
 365       for (x = 0; x < TILE_SIZE; x += 4) {
 366          uint8_t *color[PIPE_MAX_COLOR_BUFS];
 367          uint32_t *depth;
 368          unsigned block_offset, i;
 369
 370          /* offset of the 16x16 pixel block within the tile */
 371          block_offset = ((y / 4) * (16 * 16) + (x / 4) * 16);
 372
 373          /* color buffer */
 374          for (i = 0; i < rast->state.fb.nr_cbufs; i++)
 375             color[i] = tile->color[i] + 4 * block_offset;
 376
 377          /* depth buffer */
 378          depth = lp_rast_depth_pointer(rast, tile_x + x, tile_y + y);
 379
 380          /* run shader */
 381          state->jit_function[0]( &state->jit_context,
 382                                  tile_x + x, tile_y + y,
 383                                  inputs->a0,
 384                                  inputs->dadx,
 385                                  inputs->dady,
 386                                  color,
 387                                  depth,
 388                                  INT_MIN, INT_MIN, INT_MIN,
 389                                  NULL, NULL, NULL );
 390       }
 391    }
 392 }
 393
 394
 395 /**
 396  * Compute shading for a 4x4 block of pixels.
 397  * This is a bin command called during bin processing.
 398  */
 399 void lp_rast_shade_quads( struct lp_rasterizer *rast,
 400                           unsigned thread_index,
 401                           const struct lp_rast_shader_inputs *inputs,
 402                           unsigned x, unsigned y,
 403                           int32_t c1, int32_t c2, int32_t c3)
 404 {
 405    struct lp_rasterizer_task *task = &rast->tasks[thread_index];
 406    const struct lp_rast_state *state = task->current_state;
 407    struct lp_rast_tile *tile = &task->tile;
 408    uint8_t *color[PIPE_MAX_COLOR_BUFS];
 409    void *depth;
 410    unsigned i;
 411    unsigned ix, iy;
 412    int block_offset;
 413
 414 #ifdef DEBUG
 415    assert(state);
 416
 417    /* Sanity checks */
 418    assert(x % TILE_VECTOR_WIDTH == 0);
 419    assert(y % TILE_VECTOR_HEIGHT == 0);
 420
 421    assert((x % 4) == 0);
 422    assert((y % 4) == 0);
 423 #endif
 424
 425    ix = x % TILE_SIZE;
 426    iy = y % TILE_SIZE;
 427
 428    /* offset of the 16x16 pixel block within the tile */
 429    block_offset = ((iy / 4) * (16 * 16) + (ix / 4) * 16);
 430
 431    /* color buffer */
 432    for (i = 0; i < rast->state.fb.nr_cbufs; i++)
 433       color[i] = tile->color[i] + 4 * block_offset;
 434
 435    /* depth buffer */
 436    depth = lp_rast_depth_pointer(rast, x, y);
 437
 438
 439
 440 #ifdef DEBUG
 441    assert(lp_check_alignment(tile->color[0], 16));
 442    assert(lp_check_alignment(state->jit_context.blend_color, 16));
 443
 444    assert(lp_check_alignment(inputs->step[0], 16));
 445    assert(lp_check_alignment(inputs->step[1], 16));
 446    assert(lp_check_alignment(inputs->step[2], 16));
 447 #endif
 448
 449    /* run shader */
 450    state->jit_function[1]( &state->jit_context,
 451                         x, y,
 452                         inputs->a0,
 453                         inputs->dadx,
 454                         inputs->dady,
 455                         color,
 456                         depth,
 457                         c1, c2, c3,
 458                         inputs->step[0], inputs->step[1], inputs->step[2]);
 459 }
 460
 461
 462 /**
 463  * Set top row and left column of the tile's pixels to white.  For debugging.
 464  */
 465 static void
 466 outline_tile(uint8_t *tile)
 467 {
 468    const uint8_t val = 0xff;
 469    unsigned i;
 470
 471    for (i = 0; i < TILE_SIZE; i++) {
 472       TILE_PIXEL(tile, i, 0, 0) = val;
 473       TILE_PIXEL(tile, i, 0, 1) = val;
 474       TILE_PIXEL(tile, i, 0, 2) = val;
 475       TILE_PIXEL(tile, i, 0, 3) = val;
 476
 477       TILE_PIXEL(tile, 0, i, 0) = val;
 478       TILE_PIXEL(tile, 0, i, 1) = val;
 479       TILE_PIXEL(tile, 0, i, 2) = val;
 480       TILE_PIXEL(tile, 0, i, 3) = val;
 481    }
 482 }
 483
 484
 485 /**
 486  * Draw grid of gray lines at 16-pixel intervals across the tile to
 487  * show the sub-tile boundaries.  For debugging.
 488  */
 489 static void
 490 outline_subtiles(uint8_t *tile)
 491 {
 492    const uint8_t val = 0x80;
 493    const unsigned step = 16;
 494    unsigned i, j;
 495
 496    for (i = 0; i < TILE_SIZE; i += step) {
 497       for (j = 0; j < TILE_SIZE; j++) {
 498          TILE_PIXEL(tile, i, j, 0) = val;
 499          TILE_PIXEL(tile, i, j, 1) = val;
 500          TILE_PIXEL(tile, i, j, 2) = val;
 501          TILE_PIXEL(tile, i, j, 3) = val;
 502
 503          TILE_PIXEL(tile, j, i, 0) = val;
 504          TILE_PIXEL(tile, j, i, 1) = val;
 505          TILE_PIXEL(tile, j, i, 2) = val;
 506          TILE_PIXEL(tile, j, i, 3) = val;
 507       }
 508    }
 509
 510    outline_tile(tile);
 511 }
 512
 513
 514
 515 /**
 516  * Write the rasterizer's color tile to the framebuffer.
 517  */
 518 static void lp_rast_store_color( struct lp_rasterizer *rast,
 519                                  unsigned thread_index)
 520 {
 521    struct lp_rasterizer_task *task = &rast->tasks[thread_index];
 522    const unsigned x = task->x;
 523    const unsigned y = task->y;
 524    unsigned i;
 525
 526    for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
 527       struct pipe_transfer *transfer = rast->cbuf_transfer[i];
 528       int w = TILE_SIZE;
 529       int h = TILE_SIZE;
 530
 531       if (x >= transfer->width)
 532          continue;
 533
 534       if (y >= transfer->height)
 535          continue;
 536
 537       LP_DBG(DEBUG_RAST, "%s [%u] %d,%d %dx%d\n", __FUNCTION__,
 538              thread_index, x, y, w, h);
 539
 540       if (LP_DEBUG & DEBUG_SHOW_SUBTILES)
 541          outline_subtiles(task->tile.color[i]);
 542       else if (LP_DEBUG & DEBUG_SHOW_TILES)
 543          outline_tile(task->tile.color[i]);
 544
 545       lp_tile_write_4ub(transfer->texture->format,
 546                         task->tile.color[i],
 547                         rast->cbuf_map[i],
 548                         transfer->stride,
 549                         x, y,
 550                         w, h);
 551
 552       LP_COUNT(nr_color_tile_store);
 553    }
 554 }
 555
 556
 557 /**
 558  * Write the rasterizer's tiles to the framebuffer.
 559  */
 560 static void
 561 lp_rast_end_tile( struct lp_rasterizer *rast,
 562                   unsigned thread_index )
 563 {
 564    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
 565
 566    if (rast->state.write_color)
 567       lp_rast_store_color(rast, thread_index);
 568 }
 569
 570
 571 /**
 572  * Signal on a fence.  This is called during bin execution/rasterization.
 573  * Called per thread.
 574  */
 575 void lp_rast_fence( struct lp_rasterizer *rast,
 576                     unsigned thread_index,
 577                     const union lp_rast_cmd_arg arg )
 578 {
 579    struct lp_fence *fence = arg.fence;
 580
 581    pipe_mutex_lock( fence->mutex );
 582
 583    fence->count++;
 584    assert(fence->count <= fence->rank);
 585
 586    LP_DBG(DEBUG_RAST, "%s count=%u rank=%u\n", __FUNCTION__,
 587           fence->count, fence->rank);
 588
 589    pipe_condvar_signal( fence->signalled );
 590
 591    pipe_mutex_unlock( fence->mutex );
 592 }
 593
 594
 595 /**
 596  * When all the threads are done rasterizing a scene, one thread will
 597  * call this function to reset the scene and put it onto the empty queue.
 598  */
 599 static void
 600 release_scene( struct lp_rasterizer *rast,
 601                struct lp_scene *scene )
 602 {
 603    util_unreference_framebuffer_state( &scene->fb );
 604
 605    lp_scene_reset( scene );
 606
 607    assert(lp_scene_is_empty(scene));
 608
 609    lp_scene_enqueue( rast->empty_scenes, scene );
 610    rast->curr_scene = NULL;
 611 }
 612
 613
 614 /**
 615  * Rasterize commands for a single bin.
 616  * \param x, y  position of the bin's tile in the framebuffer
 617  * Must be called between lp_rast_begin() and lp_rast_end().
 618  * Called per thread.
 619  */
 620 static void
 621 rasterize_bin( struct lp_rasterizer *rast,
 622                unsigned thread_index,
 623                const struct cmd_bin *bin,
 624                int x, int y)
 625 {
 626    const struct cmd_block_list *commands = &bin->commands;
 627    struct cmd_block *block;
 628    unsigned k;
 629
 630    lp_rast_start_tile( rast, thread_index, x, y );
 631
 632    /* simply execute each of the commands in the block list */
 633    for (block = commands->head; block; block = block->next) {
 634       for (k = 0; k < block->count; k++) {
 635          block->cmd[k]( rast, thread_index, block->arg[k] );
 636       }
 637    }
 638
 639    lp_rast_end_tile( rast, thread_index );
 640 }
 641
 642
 643 #define RAST(x) { lp_rast_##x, #x }
 644
 645 static struct {
 646    lp_rast_cmd cmd;
 647    const char *name;
 648 } cmd_names[] =
 649 {
 650    RAST(load_color),
 651    RAST(clear_color),
 652    RAST(clear_zstencil),
 653    RAST(triangle),
 654    RAST(shade_tile),
 655    RAST(set_state),
 656    RAST(fence),
 657 };
 658
 659 static void
 660 debug_bin( const struct cmd_bin *bin )
 661 {
 662    const struct cmd_block *head = bin->commands.head;
 663    int i, j;
 664
 665    for (i = 0; i < head->count; i++) {
 666       debug_printf("%d: ", i);
 667       for (j = 0; j < Elements(cmd_names); j++) {
 668          if (head->cmd[i] == cmd_names[j].cmd) {
 669             debug_printf("%s\n", cmd_names[j].name);
 670             break;
 671          }
 672       }
 673       if (j == Elements(cmd_names))
 674          debug_printf("...other\n");
 675    }
 676
 677 }
 678
 679 /* An empty bin is one that just loads the contents of the tile and
 680  * stores them again unchanged.  This typically happens when bins have
 681  * been flushed for some reason in the middle of a frame, or when
 682  * incremental updates are being made to a render target.
 683  *
 684  * Try to avoid doing pointless work in this case.
 685  */
 686 static boolean
 687 is_empty_bin( const struct cmd_bin *bin )
 688 {
 689    const struct cmd_block *head = bin->commands.head;
 690    int i;
 691
 692    if (0)
 693       debug_bin(bin);
 694
 695    /* We emit at most two load-tile commands at the start of the first
 696     * command block.  In addition we seem to emit a couple of
 697     * set-state commands even in empty bins.
 698     *
 699     * As a heuristic, if a bin has more than 4 commands, consider it
 700     * non-empty.
 701     */
 702    if (head->next != NULL ||
 703        head->count > 4) {
 704       return FALSE;
 705    }
 706
 707    for (i = 0; i < head->count; i++)
 708       if (head->cmd[i] != lp_rast_load_color &&
 709           head->cmd[i] != lp_rast_set_state) {
 710          return FALSE;
 711       }
 712
 713    return TRUE;
 714 }
 715
 716
 717
 718 /**
 719  * Rasterize/execute all bins within a scene.
 720  * Called per thread.
 721  */
 722 static void
 723 rasterize_scene( struct lp_rasterizer *rast,
 724                 unsigned thread_index,
 725                 struct lp_scene *scene,
 726                 bool write_depth )
 727 {
 728    /* loop over scene bins, rasterize each */
 729 #if 0
 730    {
 731       unsigned i, j;
 732       for (i = 0; i < scene->tiles_x; i++) {
 733          for (j = 0; j < scene->tiles_y; j++) {
 734             struct cmd_bin *bin = lp_get_bin(scene, i, j);
 735             rasterize_bin( rast, thread_index,
 736                            bin, i * TILE_SIZE, j * TILE_SIZE );
 737          }
 738       }
 739    }
 740 #else
 741    {
 742       struct cmd_bin *bin;
 743       int x, y;
 744
 745       assert(scene);
 746       while ((bin = lp_scene_bin_iter_next(scene, &x, &y))) {
 747          if (!is_empty_bin( bin ))
 748             rasterize_bin( rast, thread_index, bin, x * TILE_SIZE, y * TILE_SIZE);
 749       }
 750    }
 751 #endif
 752 }
 753
 754
 755 /**
 756  * Called by setup module when it has something for us to render.
 757  */
 758 void
 759 lp_rasterize_scene( struct lp_rasterizer *rast,
 760                    struct lp_scene *scene,
 761                    const struct pipe_framebuffer_state *fb,
 762                    bool write_depth )
 763 {
 764    boolean debug = false;
 765
 766    LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
 767
 768    if (debug) {
 769       unsigned x, y;
 770       debug_printf("rasterize scene:\n");
 771       debug_printf("  data size: %u\n", lp_scene_data_size(scene));
 772       for (y = 0; y < scene->tiles_y; y++) {
 773          for (x = 0; x < scene->tiles_x; x++) {
 774             debug_printf("  bin %u, %u size: %u\n", x, y,
 775                          lp_scene_bin_size(scene, x, y));
 776          }
 777       }
 778    }
 779
 780    /* save framebuffer state in the bin */
 781    util_copy_framebuffer_state(&scene->fb, fb);
 782    scene->write_depth = write_depth;
 783
 784    if (rast->num_threads == 0) {
 785       /* no threading */
 786
 787       lp_rast_begin( rast, fb,
 788                      fb->nr_cbufs != 0, /* always write color if cbufs present */
 789                      fb->zsbuf != NULL && write_depth );
 790
 791       lp_scene_bin_iter_begin( scene );
 792       rasterize_scene( rast, 0, scene, write_depth );
 793
 794       release_scene( rast, scene );
 795
 796       lp_rast_end( rast );
 797    }
 798    else {
 799       /* threaded rendering! */
 800       unsigned i;
 801
 802       lp_scene_enqueue( rast->full_scenes, scene );
 803
 804       /* signal the threads that there's work to do */
 805       for (i = 0; i < rast->num_threads; i++) {
 806          pipe_semaphore_signal(&rast->tasks[i].work_ready);
 807       }
 808
 809       /* wait for work to complete */
 810       for (i = 0; i < rast->num_threads; i++) {
 811          pipe_semaphore_wait(&rast->tasks[i].work_done);
 812       }
 813    }
 814
 815    LP_DBG(DEBUG_SETUP, "%s done \n", __FUNCTION__);
 816 }
 817
 818
 819 /**
 820  * This is the thread's main entrypoint.
 821  * It's a simple loop:
 822  *   1. wait for work
 823  *   2. do work
 824  *   3. signal that we're done
 825  */
 826 static PIPE_THREAD_ROUTINE( thread_func, init_data )
 827 {
 828    struct lp_rasterizer_task *task = (struct lp_rasterizer_task *) init_data;
 829    struct lp_rasterizer *rast = task->rast;
 830    boolean debug = false;
 831
 832    while (1) {
 833       /* wait for work */
 834       if (debug)
 835          debug_printf("thread %d waiting for work\n", task->thread_index);
 836       pipe_semaphore_wait(&task->work_ready);
 837
 838       if (task->thread_index == 0) {
 839          /* thread[0]:
 840           *  - get next scene to rasterize
 841           *  - map the framebuffer surfaces
 842           */
 843          const struct pipe_framebuffer_state *fb;
 844          boolean write_depth;
 845
 846          rast->curr_scene = lp_scene_dequeue( rast->full_scenes, TRUE );
 847
 848          lp_scene_bin_iter_begin( rast->curr_scene );
 849
 850          fb = &rast->curr_scene->fb;
 851          write_depth = rast->curr_scene->write_depth;
 852
 853          lp_rast_begin( rast, fb,
 854                         fb->nr_cbufs != 0,
 855                         fb->zsbuf != NULL && write_depth );
 856       }
 857
 858       /* Wait for all threads to get here so that threads[1+] don't
 859        * get a null rast->curr_scene pointer.
 860        */
 861       pipe_barrier_wait( &rast->barrier );
 862
 863       /* do work */
 864       if (debug)
 865          debug_printf("thread %d doing work\n", task->thread_index);
 866       rasterize_scene(rast,
 867                      task->thread_index,
 868                      rast->curr_scene,
 869                      rast->curr_scene->write_depth);
 870
 871       /* wait for all threads to finish with this scene */
 872       pipe_barrier_wait( &rast->barrier );
 873
 874       if (task->thread_index == 0) {
 875          /* thread[0]:
 876           * - release the scene object
 877           * - unmap the framebuffer surfaces
 878           */
 879          release_scene( rast, rast->curr_scene );
 880          lp_rast_end( rast );
 881       }
 882
 883       /* signal done with work */
 884       if (debug)
 885          debug_printf("thread %d done working\n", task->thread_index);
 886       pipe_semaphore_signal(&task->work_done);
 887    }
 888
 889    return NULL;
 890 }
 891
 892
 893 /**
 894  * Initialize semaphores and spawn the threads.
 895  */
 896 static void
 897 create_rast_threads(struct lp_rasterizer *rast)
 898 {
 899    unsigned i;
 900
 901 #ifdef PIPE_OS_WINDOWS
 902    /* Multithreading not supported on windows until conditions and barriers are
 903     * properly implemented. */
 904    rast->num_threads = 0;
 905 #else
 906    rast->num_threads = util_cpu_caps.nr_cpus;
 907    rast->num_threads = debug_get_num_option("LP_NUM_THREADS", rast->num_threads);
 908    rast->num_threads = MIN2(rast->num_threads, MAX_THREADS);
 909 #endif
 910
 911    /* NOTE: if num_threads is zero, we won't use any threads */
 912    for (i = 0; i < rast->num_threads; i++) {
 913       pipe_semaphore_init(&rast->tasks[i].work_ready, 0);
 914       pipe_semaphore_init(&rast->tasks[i].work_done, 0);
 915       rast->threads[i] = pipe_thread_create(thread_func,
 916                                             (void *) &rast->tasks[i]);
 917    }
 918 }
 919
 920
 921
 922 /**
 923  * Create new lp_rasterizer.
 924  * \param empty  the queue to put empty scenes on after we've finished
 925  *               processing them.
 926  */
 927 struct lp_rasterizer *
 928 lp_rast_create( struct pipe_screen *screen, struct lp_scene_queue *empty )
 929 {
 930    struct lp_rasterizer *rast;
 931    unsigned i, cbuf;
 932
 933    rast = CALLOC_STRUCT(lp_rasterizer);
 934    if(!rast)
 935       return NULL;
 936
 937    rast->screen = screen;
 938
 939    rast->empty_scenes = empty;
 940    rast->full_scenes = lp_scene_queue_create();
 941
 942    for (i = 0; i < Elements(rast->tasks); i++) {
 943       struct lp_rasterizer_task *task = &rast->tasks[i];
 944
 945       for (cbuf = 0; cbuf < PIPE_MAX_COLOR_BUFS; cbuf++ )
 946          task->tile.color[cbuf] = align_malloc(TILE_SIZE * TILE_SIZE * 4, 16);
 947
 948       task->rast = rast;
 949       task->thread_index = i;
 950    }
 951
 952    create_rast_threads(rast);
 953
 954    /* for synchronizing rasterization threads */
 955    pipe_barrier_init( &rast->barrier, rast->num_threads );
 956
 957    return rast;
 958 }
 959
 960
 961 /* Shutdown:
 962  */
 963 void lp_rast_destroy( struct lp_rasterizer *rast )
 964 {
 965    unsigned i, cbuf;
 966
 967    util_unreference_framebuffer_state(&rast->state.fb);
 968
 969    for (i = 0; i < Elements(rast->tasks); i++) {
 970       for (cbuf = 0; cbuf < PIPE_MAX_COLOR_BUFS; cbuf++ )
 971          align_free(rast->tasks[i].tile.color[cbuf]);
 972    }
 973
 974    /* for synchronizing rasterization threads */
 975    pipe_barrier_destroy( &rast->barrier );
 976
 977    FREE(rast);
 978 }
 979
 980
 981 /** Return number of rasterization threads */
 982 unsigned
 983 lp_rast_get_num_threads( struct lp_rasterizer *rast )
 984 {
 985    return rast->num_threads;
 986 }