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    int i;
  57
  58    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
  59
  60    util_copy_framebuffer_state(&rast->state.fb, fb);
  61
  62    rast->state.write_zstencil = write_zstencil;
  63    rast->state.write_color = write_color;
  64
  65    rast->check_for_clipped_tiles = (fb->width % TILE_SIZE != 0 ||
  66                                     fb->height % TILE_SIZE != 0);
  67
  68
  69    for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
  70       cbuf = rast->state.fb.cbufs[i];
  71       if (cbuf) {
  72          rast->cbuf_transfer[i] = screen->get_tex_transfer(rast->screen,
  73                                                            cbuf->texture,
  74                                                            cbuf->face,
  75                                                            cbuf->level,
  76                                                            cbuf->zslice,
  77                                                            PIPE_TRANSFER_READ_WRITE,
  78                                                            0, 0,
  79                                                            cbuf->width,
  80                                                            cbuf->height);
  81          if (!rast->cbuf_transfer[i])
  82             goto fail;
  83
  84          rast->cbuf_map[i] = screen->transfer_map(rast->screen,
  85                                                   rast->cbuf_transfer[i]);
  86          if (!rast->cbuf_map[i])
  87             goto fail;
  88       }
  89    }
  90
  91    zsbuf = rast->state.fb.zsbuf;
  92    if (zsbuf) {
  93       rast->zsbuf_transfer = screen->get_tex_transfer(rast->screen,
  94                                                       zsbuf->texture,
  95                                                       zsbuf->face,
  96                                                       zsbuf->level,
  97                                                       zsbuf->zslice,
  98                                                       PIPE_TRANSFER_READ_WRITE,
  99                                                       0, 0,
 100                                                       zsbuf->width,
 101                                                       zsbuf->height);
 102       if (!rast->zsbuf_transfer)
 103          goto fail;
 104
 105       rast->zsbuf_map = screen->transfer_map(rast->screen,
 106                                             rast->zsbuf_transfer);
 107       if (!rast->zsbuf_map)
 108          goto fail;
 109    }
 110
 111    return TRUE;
 112
 113 fail:
 114    /* Unmap and release transfers?
 115     */
 116    return FALSE;
 117 }
 118
 119
 120 /**
 121  * Finish the rasterization phase.
 122  * Unmap framebuffer surfaces.
 123  */
 124 static void
 125 lp_rast_end( struct lp_rasterizer *rast )
 126 {
 127    struct pipe_screen *screen = rast->screen;
 128    unsigned i;
 129
 130    for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
 131       if (rast->cbuf_map[i])
 132          screen->transfer_unmap(screen, rast->cbuf_transfer[i]);
 133
 134       if (rast->cbuf_transfer[i])
 135          screen->tex_transfer_destroy(rast->cbuf_transfer[i]);
 136
 137       rast->cbuf_transfer[i] = NULL;
 138       rast->cbuf_map[i] = NULL;
 139    }
 140
 141    if (rast->zsbuf_map)
 142       screen->transfer_unmap(screen, rast->zsbuf_transfer);
 143
 144    if (rast->zsbuf_transfer)
 145       screen->tex_transfer_destroy(rast->zsbuf_transfer);
 146
 147    rast->zsbuf_transfer = NULL;
 148    rast->zsbuf_map = NULL;
 149 }
 150
 151
 152 /**
 153  * Begining rasterization of a tile.
 154  * \param x  window X position of the tile, in pixels
 155  * \param y  window Y position of the tile, in pixels
 156  */
 157 static void
 158 lp_rast_start_tile( struct lp_rasterizer *rast,
 159                     unsigned thread_index,
 160                     unsigned x, unsigned y )
 161 {
 162    LP_DBG(DEBUG_RAST, "%s %d,%d\n", __FUNCTION__, x, y);
 163
 164    rast->tasks[thread_index].x = x;
 165    rast->tasks[thread_index].y = y;
 166 }
 167
 168
 169 /**
 170  * Clear the rasterizer's current color tile.
 171  * This is a bin command called during bin processing.
 172  */
 173 void lp_rast_clear_color( struct lp_rasterizer *rast,
 174                           unsigned thread_index,
 175                           const union lp_rast_cmd_arg arg )
 176 {
 177    const uint8_t *clear_color = arg.clear_color;
 178    uint8_t **color_tile = rast->tasks[thread_index].tile.color;
 179    unsigned i;
 180
 181    LP_DBG(DEBUG_RAST, "%s 0x%x,0x%x,0x%x,0x%x\n", __FUNCTION__,
 182               clear_color[0],
 183               clear_color[1],
 184               clear_color[2],
 185               clear_color[3]);
 186
 187    if (clear_color[0] == clear_color[1] &&
 188        clear_color[1] == clear_color[2] &&
 189        clear_color[2] == clear_color[3]) {
 190       /* clear to grayscale value {x, x, x, x} */
 191       for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
 192          memset(color_tile[i], clear_color[0], TILE_SIZE * TILE_SIZE * 4);
 193       }
 194    }
 195    else {
 196       /* Non-gray color.
 197        * Note: if the swizzled tile layout changes (see TILE_PIXEL) this code
 198        * will need to change.  It'll be pretty obvious when clearing no longer
 199        * works.
 200        */
 201       const unsigned chunk = TILE_SIZE / 4;
 202       for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
 203          uint8_t *c = color_tile[i];
 204          unsigned j;
 205          for (j = 0; j < 4 * TILE_SIZE; j++) {
 206             memset(c, clear_color[0], chunk);
 207             c += chunk;
 208             memset(c, clear_color[1], chunk);
 209             c += chunk;
 210             memset(c, clear_color[2], chunk);
 211             c += chunk;
 212             memset(c, clear_color[3], chunk);
 213             c += chunk;
 214          }
 215          assert(c - color_tile[i] == TILE_SIZE * TILE_SIZE * 4);
 216       }
 217    }
 218 }
 219
 220
 221 /**
 222  * Clear the rasterizer's current z/stencil tile.
 223  * This is a bin command called during bin processing.
 224  */
 225 void lp_rast_clear_zstencil( struct lp_rasterizer *rast,
 226                              unsigned thread_index,
 227                              const union lp_rast_cmd_arg arg)
 228 {
 229    unsigned i;
 230    uint32_t *depth_tile = rast->tasks[thread_index].tile.depth;
 231
 232    LP_DBG(DEBUG_RAST, "%s 0x%x\n", __FUNCTION__, arg.clear_zstencil);
 233
 234    for (i = 0; i < TILE_SIZE * TILE_SIZE; i++)
 235       depth_tile[i] = arg.clear_zstencil;
 236 }
 237
 238
 239 /**
 240  * Load tile color from the framebuffer surface.
 241  * This is a bin command called during bin processing.
 242  */
 243 void lp_rast_load_color( struct lp_rasterizer *rast,
 244                          unsigned thread_index,
 245                          const union lp_rast_cmd_arg arg)
 246 {
 247    struct lp_rasterizer_task *task = &rast->tasks[thread_index];
 248    const unsigned x = task->x;
 249    const unsigned y = task->y;
 250    unsigned i;
 251
 252    LP_DBG(DEBUG_RAST, "%s at %u, %u\n", __FUNCTION__, x, y);
 253
 254    for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
 255       struct pipe_transfer *transfer = rast->cbuf_transfer[i];
 256       int w = TILE_SIZE;
 257       int h = TILE_SIZE;
 258
 259       if (x >= transfer->width)
 260          continue;
 261
 262       if (y >= transfer->height)
 263          continue;
 264       /* XXX: require tile-size aligned render target dimensions:
 265        */
 266       if (x + w > transfer->width)
 267          w -= x + w - transfer->width;
 268
 269       if (y + h > transfer->height)
 270          h -= y + h - transfer->height;
 271
 272       assert(w >= 0);
 273       assert(h >= 0);
 274       assert(w <= TILE_SIZE);
 275       assert(h <= TILE_SIZE);
 276
 277       lp_tile_read_4ub(transfer->texture->format,
 278                        rast->tasks[thread_index].tile.color[i],
 279                        rast->cbuf_map[i],
 280                        transfer->stride,
 281                        x, y,
 282                        w, h);
 283    }
 284 }
 285
 286
 287 static void
 288 lp_tile_read_z32(uint32_t *tile,
 289                  const uint8_t *map,
 290                  unsigned map_stride,
 291                  unsigned x0, unsigned y0, unsigned w, unsigned h)
 292 {
 293    unsigned x, y;
 294    const uint8_t *map_row = map + y0*map_stride;
 295    for (y = 0; y < h; ++y) {
 296       const uint32_t *map_pixel = (uint32_t *)(map_row + x0*4);
 297       for (x = 0; x < w; ++x) {
 298          *tile++ = *map_pixel++;
 299       }
 300       map_row += map_stride;
 301    }
 302 }
 303
 304 /**
 305  * Load tile z/stencil from the framebuffer surface.
 306  * This is a bin command called during bin processing.
 307  */
 308 void lp_rast_load_zstencil( struct lp_rasterizer *rast,
 309                             unsigned thread_index,
 310                             const union lp_rast_cmd_arg arg )
 311 {
 312    const unsigned x = rast->tasks[thread_index].x;
 313    const unsigned y = rast->tasks[thread_index].y;
 314    unsigned w = TILE_SIZE;
 315    unsigned h = TILE_SIZE;
 316
 317    if (x + w > rast->state.fb.width)
 318       w -= x + w - rast->state.fb.width;
 319
 320    if (y + h > rast->state.fb.height)
 321       h -= y + h - rast->state.fb.height;
 322
 323    LP_DBG(DEBUG_RAST, "%s %d,%d %dx%d\n", __FUNCTION__, x, y, w, h);
 324
 325    assert(rast->zsbuf_transfer->texture->format == PIPE_FORMAT_Z32_UNORM);
 326    lp_tile_read_z32(rast->tasks[thread_index].tile.depth,
 327                     rast->zsbuf_map,
 328                     rast->zsbuf_transfer->stride,
 329                     x, y, w, h);
 330 }
 331
 332
 333 void lp_rast_set_state( struct lp_rasterizer *rast,
 334                         unsigned thread_index,
 335                         const union lp_rast_cmd_arg arg )
 336 {
 337    const struct lp_rast_state *state = arg.set_state;
 338
 339    LP_DBG(DEBUG_RAST, "%s %p\n", __FUNCTION__, (void *) state);
 340
 341    /* just set the current state pointer for this rasterizer */
 342    rast->tasks[thread_index].current_state = state;
 343 }
 344
 345
 346
 347 /**
 348  * Run the shader on all blocks in a tile.  This is used when a tile is
 349  * completely contained inside a triangle.
 350  * This is a bin command called during bin processing.
 351  */
 352 void lp_rast_shade_tile( struct lp_rasterizer *rast,
 353                          unsigned thread_index,
 354                          const union lp_rast_cmd_arg arg )
 355 {
 356    const struct lp_rast_state *state = rast->tasks[thread_index].current_state;
 357    struct lp_rast_tile *tile = &rast->tasks[thread_index].tile;
 358    const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
 359    const unsigned tile_x = rast->tasks[thread_index].x;
 360    const unsigned tile_y = rast->tasks[thread_index].y;
 361    unsigned x, y;
 362
 363    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
 364
 365    /* render the whole 64x64 tile in 4x4 chunks */
 366    for (y = 0; y < TILE_SIZE; y += 4){
 367       for (x = 0; x < TILE_SIZE; x += 4) {
 368          uint8_t *color[PIPE_MAX_COLOR_BUFS];
 369          uint32_t *depth;
 370          unsigned block_offset, i;
 371
 372          /* offset of the 16x16 pixel block within the tile */
 373          block_offset = ((y / 4) * (16 * 16) + (x / 4) * 16);
 374
 375          /* color buffer */
 376          for (i = 0; i < rast->state.fb.nr_cbufs; i++)
 377             color[i] = tile->color[i] + 4 * block_offset;
 378
 379          /* depth buffer */
 380          depth = tile->depth + block_offset;
 381
 382          /* run shader */
 383          state->jit_function[0]( &state->jit_context,
 384                                  tile_x + x, tile_y + y,
 385                                  inputs->a0,
 386                                  inputs->dadx,
 387                                  inputs->dady,
 388                                  color,
 389                                  depth,
 390                                  INT_MIN, INT_MIN, INT_MIN,
 391                                  NULL, NULL, NULL );
 392       }
 393    }
 394 }
 395
 396
 397 /**
 398  * Compute shading for a 4x4 block of pixels.
 399  * This is a bin command called during bin processing.
 400  */
 401 void lp_rast_shade_quads( struct lp_rasterizer *rast,
 402                           unsigned thread_index,
 403                           const struct lp_rast_shader_inputs *inputs,
 404                           unsigned x, unsigned y,
 405                           int32_t c1, int32_t c2, int32_t c3)
 406 {
 407    const struct lp_rast_state *state = rast->tasks[thread_index].current_state;
 408    struct lp_rast_tile *tile = &rast->tasks[thread_index].tile;
 409    uint8_t *color[PIPE_MAX_COLOR_BUFS];
 410    void *depth;
 411    unsigned i;
 412    unsigned ix, iy;
 413    int block_offset;
 414
 415 #ifdef DEBUG
 416    assert(state);
 417
 418    /* Sanity checks */
 419    assert(x % TILE_VECTOR_WIDTH == 0);
 420    assert(y % TILE_VECTOR_HEIGHT == 0);
 421
 422    assert((x % 4) == 0);
 423    assert((y % 4) == 0);
 424 #endif
 425
 426    ix = x % TILE_SIZE;
 427    iy = y % TILE_SIZE;
 428
 429    /* offset of the 16x16 pixel block within the tile */
 430    block_offset = ((iy / 4) * (16 * 16) + (ix / 4) * 16);
 431
 432    /* color buffer */
 433    for (i = 0; i < rast->state.fb.nr_cbufs; i++)
 434       color[i] = tile->color[i] + 4 * block_offset;
 435
 436    /* depth buffer */
 437    depth = tile->depth + block_offset;
 438
 439
 440
 441 #ifdef DEBUG
 442    assert(lp_check_alignment(tile->depth, 16));
 443    assert(lp_check_alignment(tile->color[0], 16));
 444    assert(lp_check_alignment(state->jit_context.blend_color, 16));
 445
 446    assert(lp_check_alignment(inputs->step[0], 16));
 447    assert(lp_check_alignment(inputs->step[1], 16));
 448    assert(lp_check_alignment(inputs->step[2], 16));
 449 #endif
 450
 451    /* run shader */
 452    state->jit_function[1]( &state->jit_context,
 453                         x, y,
 454                         inputs->a0,
 455                         inputs->dadx,
 456                         inputs->dady,
 457                         color,
 458                         depth,
 459                         c1, c2, c3,
 460                         inputs->step[0], inputs->step[1], inputs->step[2]);
 461 }
 462
 463
 464
 465
 466 /**
 467  * Write the rasterizer's color tile to the framebuffer.
 468  */
 469 static void lp_rast_store_color( struct lp_rasterizer *rast,
 470                                  unsigned thread_index)
 471 {
 472    const unsigned x = rast->tasks[thread_index].x;
 473    const unsigned y = rast->tasks[thread_index].y;
 474    unsigned i;
 475
 476    for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
 477       struct pipe_transfer *transfer = rast->cbuf_transfer[i];
 478       int w = TILE_SIZE;
 479       int h = TILE_SIZE;
 480
 481       if (x >= transfer->width)
 482          continue;
 483
 484       if (y >= transfer->height)
 485          continue;
 486
 487       /* XXX: require tile-size aligned render target dimensions:
 488        */
 489       if (x + w > transfer->width)
 490          w -= x + w - transfer->width;
 491
 492       if (y + h > transfer->height)
 493          h -= y + h - transfer->height;
 494
 495       assert(w >= 0);
 496       assert(h >= 0);
 497       assert(w <= TILE_SIZE);
 498       assert(h <= TILE_SIZE);
 499
 500       LP_DBG(DEBUG_RAST, "%s [%u] %d,%d %dx%d\n", __FUNCTION__,
 501              thread_index, x, y, w, h);
 502
 503       lp_tile_write_4ub(transfer->texture->format,
 504                         rast->tasks[thread_index].tile.color[i],
 505                         rast->cbuf_map[i],
 506                         transfer->stride,
 507                         x, y,
 508                         w, h);
 509    }
 510 }
 511
 512
 513 static void
 514 lp_tile_write_z32(const uint32_t *src, uint8_t *dst, unsigned dst_stride,
 515                   unsigned x0, unsigned y0, unsigned w, unsigned h)
 516 {
 517    unsigned x, y;
 518    uint8_t *dst_row = dst + y0*dst_stride;
 519    for (y = 0; y < h; ++y) {
 520       uint32_t *dst_pixel = (uint32_t *)(dst_row + x0*4);
 521       for (x = 0; x < w; ++x) {
 522          *dst_pixel++ = *src++;
 523       }
 524       dst_row += dst_stride;
 525    }
 526 }
 527
 528 /**
 529  * Write the rasterizer's z/stencil tile to the framebuffer.
 530  */
 531 static void lp_rast_store_zstencil( struct lp_rasterizer *rast,
 532                                     unsigned thread_index )
 533 {
 534    const unsigned x = rast->tasks[thread_index].x;
 535    const unsigned y = rast->tasks[thread_index].y;
 536    unsigned w = TILE_SIZE;
 537    unsigned h = TILE_SIZE;
 538
 539    if (x + w > rast->state.fb.width)
 540       w -= x + w - rast->state.fb.width;
 541
 542    if (y + h > rast->state.fb.height)
 543       h -= y + h - rast->state.fb.height;
 544
 545    LP_DBG(DEBUG_RAST, "%s %d,%d %dx%d\n", __FUNCTION__, x, y, w, h);
 546
 547    assert(rast->zsbuf_transfer->texture->format == PIPE_FORMAT_Z32_UNORM);
 548    lp_tile_write_z32(rast->tasks[thread_index].tile.depth,
 549                      rast->zsbuf_map,
 550                      rast->zsbuf_transfer->stride,
 551                      x, y, w, h);
 552 }
 553
 554
 555 /**
 556  * Write the rasterizer's tiles to the framebuffer.
 557  */
 558 static void
 559 lp_rast_end_tile( struct lp_rasterizer *rast,
 560                   unsigned thread_index )
 561 {
 562    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
 563
 564    if (rast->state.write_color)
 565       lp_rast_store_color(rast, thread_index);
 566
 567    if (rast->state.write_zstencil)
 568       lp_rast_store_zstencil(rast, thread_index);
 569 }
 570
 571
 572 /**
 573  * Signal on a fence.  This is called during bin execution/rasterization.
 574  * Called per thread.
 575  */
 576 void lp_rast_fence( struct lp_rasterizer *rast,
 577                     unsigned thread_index,
 578                     const union lp_rast_cmd_arg arg )
 579 {
 580    struct lp_fence *fence = arg.fence;
 581
 582    pipe_mutex_lock( fence->mutex );
 583
 584    fence->count++;
 585    assert(fence->count <= fence->rank);
 586
 587    LP_DBG(DEBUG_RAST, "%s count=%u rank=%u\n", __FUNCTION__,
 588           fence->count, fence->rank);
 589
 590    pipe_condvar_signal( fence->signalled );
 591
 592    pipe_mutex_unlock( fence->mutex );
 593 }
 594
 595
 596 /**
 597  * When all the threads are done rasterizing a scene, one thread will
 598  * call this function to reset the scene and put it onto the empty queue.
 599  */
 600 static void
 601 release_scene( struct lp_rasterizer *rast,
 602                struct lp_scene *scene )
 603 {
 604    util_unreference_framebuffer_state( &scene->fb );
 605
 606    lp_scene_reset( scene );
 607    lp_scene_enqueue( rast->empty_scenes, scene );
 608    rast->curr_scene = NULL;
 609 }
 610
 611
 612 /**
 613  * Rasterize commands for a single bin.
 614  * \param x, y  position of the bin's tile in the framebuffer
 615  * Must be called between lp_rast_begin() and lp_rast_end().
 616  * Called per thread.
 617  */
 618 static void
 619 rasterize_bin( struct lp_rasterizer *rast,
 620                unsigned thread_index,
 621                const struct cmd_bin *bin,
 622                int x, int y)
 623 {
 624    const struct cmd_block_list *commands = &bin->commands;
 625    struct cmd_block *block;
 626    unsigned k;
 627
 628    lp_rast_start_tile( rast, thread_index, x, y );
 629
 630    /* simply execute each of the commands in the block list */
 631    for (block = commands->head; block; block = block->next) {
 632       for (k = 0; k < block->count; k++) {
 633          block->cmd[k]( rast, thread_index, block->arg[k] );
 634       }
 635    }
 636
 637    lp_rast_end_tile( rast, thread_index );
 638 }
 639
 640
 641 #define RAST(x) { lp_rast_##x, #x }
 642
 643 static struct {
 644    lp_rast_cmd cmd;
 645    const char *name;
 646 } cmd_names[] =
 647 {
 648    RAST(load_color),
 649    RAST(load_zstencil),
 650    RAST(clear_color),
 651    RAST(clear_zstencil),
 652    RAST(triangle),
 653    RAST(shade_tile),
 654    RAST(set_state),
 655    RAST(fence),
 656 };
 657
 658 static void
 659 debug_bin( const struct cmd_bin *bin )
 660 {
 661    const struct cmd_block *head = bin->commands.head;
 662    int i, j;
 663
 664    for (i = 0; i < head->count; i++) {
 665       debug_printf("%d: ", i);
 666       for (j = 0; j < Elements(cmd_names); j++) {
 667          if (head->cmd[i] == cmd_names[j].cmd) {
 668             debug_printf("%s\n", cmd_names[j].name);
 669             break;
 670          }
 671       }
 672       if (j == Elements(cmd_names))
 673          debug_printf("...other\n");
 674    }
 675
 676 }
 677
 678 /* An empty bin is one that just loads the contents of the tile and
 679  * stores them again unchanged.  This typically happens when bins have
 680  * been flushed for some reason in the middle of a frame, or when
 681  * incremental updates are being made to a render target.
 682  *
 683  * Try to avoid doing pointless work in this case.
 684  */
 685 static boolean
 686 is_empty_bin( const struct cmd_bin *bin )
 687 {
 688    const struct cmd_block *head = bin->commands.head;
 689    int i;
 690
 691    if (0)
 692       debug_bin(bin);
 693
 694    /* We emit at most two load-tile commands at the start of the first
 695     * command block.  In addition we seem to emit a couple of
 696     * set-state commands even in empty bins.
 697     *
 698     * As a heuristic, if a bin has more than 4 commands, consider it
 699     * non-empty.
 700     */
 701    if (head->next != NULL ||
 702        head->count > 4) {
 703       return FALSE;
 704    }
 705
 706    for (i = 0; i < head->count; i++)
 707       if (head->cmd[i] != lp_rast_load_color &&
 708           head->cmd[i] != lp_rast_load_zstencil &&
 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       printf("rasterize scene:\n");
 771       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             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 void *
 827 thread_func( void *init_data )
 828 {
 829    struct lp_rasterizer_task *task = (struct lp_rasterizer_task *) init_data;
 830    struct lp_rasterizer *rast = task->rast;
 831    boolean debug = false;
 832
 833    while (1) {
 834       /* wait for work */
 835       if (debug)
 836          debug_printf("thread %d waiting for work\n", task->thread_index);
 837       pipe_semaphore_wait(&task->work_ready);
 838
 839       if (task->thread_index == 0) {
 840          /* thread[0]:
 841           *  - get next scene to rasterize
 842           *  - map the framebuffer surfaces
 843           */
 844          const struct pipe_framebuffer_state *fb;
 845          boolean write_depth;
 846
 847          rast->curr_scene = lp_scene_dequeue( rast->full_scenes );
 848
 849          lp_scene_bin_iter_begin( rast->curr_scene );
 850
 851          fb = &rast->curr_scene->fb;
 852          write_depth = rast->curr_scene->write_depth;
 853
 854          lp_rast_begin( rast, fb,
 855                         fb->nr_cbufs != 0,
 856                         fb->zsbuf != NULL && write_depth );
 857       }
 858
 859       /* Wait for all threads to get here so that threads[1+] don't
 860        * get a null rast->curr_scene pointer.
 861        */
 862       pipe_barrier_wait( &rast->barrier );
 863
 864       /* do work */
 865       if (debug)
 866          debug_printf("thread %d doing work\n", task->thread_index);
 867       rasterize_scene(rast,
 868                      task->thread_index,
 869                      rast->curr_scene,
 870                      rast->curr_scene->write_depth);
 871
 872       /* wait for all threads to finish with this scene */
 873       pipe_barrier_wait( &rast->barrier );
 874
 875       if (task->thread_index == 0) {
 876          /* thread[0]:
 877           * - release the scene object
 878           * - unmap the framebuffer surfaces
 879           */
 880          release_scene( rast, rast->curr_scene );
 881          lp_rast_end( rast );
 882       }
 883
 884       /* signal done with work */
 885       if (debug)
 886          debug_printf("thread %d done working\n", task->thread_index);
 887       pipe_semaphore_signal(&task->work_done);
 888    }
 889
 890    return NULL;
 891 }
 892
 893
 894 /**
 895  * Initialize semaphores and spawn the threads.
 896  */
 897 static void
 898 create_rast_threads(struct lp_rasterizer *rast)
 899 {
 900    unsigned i;
 901
 902    rast->num_threads = util_cpu_caps.nr_cpus;
 903    rast->num_threads = debug_get_num_option("LP_NUM_THREADS", rast->num_threads);
 904    rast->num_threads = MIN2(rast->num_threads, MAX_THREADS);
 905
 906    /* NOTE: if num_threads is zero, we won't use any threads */
 907    for (i = 0; i < rast->num_threads; i++) {
 908       pipe_semaphore_init(&rast->tasks[i].work_ready, 0);
 909       pipe_semaphore_init(&rast->tasks[i].work_done, 0);
 910       rast->threads[i] = pipe_thread_create(thread_func,
 911                                             (void *) &rast->tasks[i]);
 912    }
 913 }
 914
 915
 916
 917 /**
 918  * Create new lp_rasterizer.
 919  * \param empty  the queue to put empty scenes on after we've finished
 920  *               processing them.
 921  */
 922 struct lp_rasterizer *
 923 lp_rast_create( struct pipe_screen *screen, struct lp_scene_queue *empty )
 924 {
 925    struct lp_rasterizer *rast;
 926    unsigned i, cbuf;
 927
 928    rast = CALLOC_STRUCT(lp_rasterizer);
 929    if(!rast)
 930       return NULL;
 931
 932    rast->screen = screen;
 933
 934    rast->empty_scenes = empty;
 935    rast->full_scenes = lp_scene_queue_create();
 936
 937    for (i = 0; i < Elements(rast->tasks); i++) {
 938       for (cbuf = 0; cbuf < PIPE_MAX_COLOR_BUFS; cbuf++ )
 939          rast->tasks[i].tile.color[cbuf] = align_malloc( TILE_SIZE*TILE_SIZE*4, 16 );
 940
 941       rast->tasks[i].tile.depth = align_malloc( TILE_SIZE*TILE_SIZE*4, 16 );
 942       rast->tasks[i].rast = rast;
 943       rast->tasks[i].thread_index = i;
 944    }
 945
 946    create_rast_threads(rast);
 947
 948    /* for synchronizing rasterization threads */
 949    pipe_barrier_init( &rast->barrier, rast->num_threads );
 950
 951    return rast;
 952 }
 953
 954
 955 /* Shutdown:
 956  */
 957 void lp_rast_destroy( struct lp_rasterizer *rast )
 958 {
 959    unsigned i, cbuf;
 960
 961    util_unreference_framebuffer_state(&rast->state.fb);
 962
 963    for (i = 0; i < Elements(rast->tasks); i++) {
 964       align_free(rast->tasks[i].tile.depth);
 965       for (cbuf = 0; cbuf < PIPE_MAX_COLOR_BUFS; cbuf++ )
 966          align_free(rast->tasks[i].tile.color[cbuf]);
 967    }
 968
 969    /* for synchronizing rasterization threads */
 970    pipe_barrier_destroy( &rast->barrier );
 971
 972    FREE(rast);
 973 }
 974
 975
 976 /** Return number of rasterization threads */
 977 unsigned
 978 lp_rast_get_num_threads( struct lp_rasterizer *rast )
 979 {
 980    return rast->num_threads;
 981 }