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    unsigned i;
 233    uint32_t *depth_tile = rast->tasks[thread_index].tile.depth;
 234
 235    LP_DBG(DEBUG_RAST, "%s 0x%x\n", __FUNCTION__, arg.clear_zstencil);
 236
 237    for (i = 0; i < TILE_SIZE * TILE_SIZE; i++)
 238       depth_tile[i] = arg.clear_zstencil;
 239 }
 240
 241
 242 /**
 243  * Load tile color from the framebuffer surface.
 244  * This is a bin command called during bin processing.
 245  */
 246 void lp_rast_load_color( struct lp_rasterizer *rast,
 247                          unsigned thread_index,
 248                          const union lp_rast_cmd_arg arg)
 249 {
 250    struct lp_rasterizer_task *task = &rast->tasks[thread_index];
 251    const unsigned x = task->x;
 252    const unsigned y = task->y;
 253    unsigned i;
 254
 255    LP_DBG(DEBUG_RAST, "%s at %u, %u\n", __FUNCTION__, x, y);
 256
 257    for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
 258       struct pipe_transfer *transfer = rast->cbuf_transfer[i];
 259       int w = TILE_SIZE;
 260       int h = TILE_SIZE;
 261
 262       if (x >= transfer->width)
 263          continue;
 264
 265       if (y >= transfer->height)
 266          continue;
 267
 268       assert(w >= 0);
 269       assert(h >= 0);
 270       assert(w <= TILE_SIZE);
 271       assert(h <= TILE_SIZE);
 272
 273       lp_tile_read_4ub(transfer->texture->format,
 274                        task->tile.color[i],
 275                        rast->cbuf_map[i],
 276                        transfer->stride,
 277                        x, y,
 278                        w, h);
 279
 280       LP_COUNT(nr_color_tile_load);
 281    }
 282 }
 283
 284
 285 static void
 286 lp_tile_read_z32(uint32_t *tile,
 287                  const uint8_t *map,
 288                  unsigned map_stride,
 289                  unsigned x0, unsigned y0, unsigned w, unsigned h)
 290 {
 291    unsigned x, y;
 292    const uint8_t *map_row = map + y0*map_stride;
 293    for (y = 0; y < h; ++y) {
 294       const uint32_t *map_pixel = (uint32_t *)(map_row + x0*4);
 295       for (x = 0; x < w; ++x) {
 296          *tile++ = *map_pixel++;
 297       }
 298       map_row += map_stride;
 299    }
 300 }
 301
 302 /**
 303  * Load tile z/stencil from the framebuffer surface.
 304  * This is a bin command called during bin processing.
 305  */
 306 void lp_rast_load_zstencil( struct lp_rasterizer *rast,
 307                             unsigned thread_index,
 308                             const union lp_rast_cmd_arg arg )
 309 {
 310    struct lp_rasterizer_task *task = &rast->tasks[thread_index];
 311    const unsigned x = task->x;
 312    const unsigned y = task->y;
 313    unsigned w = TILE_SIZE;
 314    unsigned h = TILE_SIZE;
 315
 316    if (x + w > rast->state.fb.width)
 317       w -= x + w - rast->state.fb.width;
 318
 319    if (y + h > rast->state.fb.height)
 320       h -= y + h - rast->state.fb.height;
 321
 322    LP_DBG(DEBUG_RAST, "%s %d,%d %dx%d\n", __FUNCTION__, x, y, w, h);
 323
 324    assert(rast->zsbuf_transfer->texture->format == PIPE_FORMAT_Z32_UNORM);
 325    lp_tile_read_z32(task->tile.depth,
 326                     rast->zsbuf_map,
 327                     rast->zsbuf_transfer->stride,
 328                     x, y, w, h);
 329 }
 330
 331
 332 void lp_rast_set_state( struct lp_rasterizer *rast,
 333                         unsigned thread_index,
 334                         const union lp_rast_cmd_arg arg )
 335 {
 336    const struct lp_rast_state *state = arg.set_state;
 337
 338    LP_DBG(DEBUG_RAST, "%s %p\n", __FUNCTION__, (void *) state);
 339
 340    /* just set the current state pointer for this rasterizer */
 341    rast->tasks[thread_index].current_state = state;
 342 }
 343
 344
 345
 346 /**
 347  * Run the shader on all blocks in a tile.  This is used when a tile is
 348  * completely contained inside a triangle.
 349  * This is a bin command called during bin processing.
 350  */
 351 void lp_rast_shade_tile( struct lp_rasterizer *rast,
 352                          unsigned thread_index,
 353                          const union lp_rast_cmd_arg arg )
 354 {
 355    struct lp_rasterizer_task *task = &rast->tasks[thread_index];
 356    const struct lp_rast_state *state = task->current_state;
 357    struct lp_rast_tile *tile = &task->tile;
 358    const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
 359    const unsigned tile_x = task->x;
 360    const unsigned tile_y = task->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    struct lp_rasterizer_task *task = &rast->tasks[thread_index];
 408    const struct lp_rast_state *state = task->current_state;
 409    struct lp_rast_tile *tile = &task->tile;
 410    uint8_t *color[PIPE_MAX_COLOR_BUFS];
 411    void *depth;
 412    unsigned i;
 413    unsigned ix, iy;
 414    int block_offset;
 415
 416 #ifdef DEBUG
 417    assert(state);
 418
 419    /* Sanity checks */
 420    assert(x % TILE_VECTOR_WIDTH == 0);
 421    assert(y % TILE_VECTOR_HEIGHT == 0);
 422
 423    assert((x % 4) == 0);
 424    assert((y % 4) == 0);
 425 #endif
 426
 427    ix = x % TILE_SIZE;
 428    iy = y % TILE_SIZE;
 429
 430    /* offset of the 16x16 pixel block within the tile */
 431    block_offset = ((iy / 4) * (16 * 16) + (ix / 4) * 16);
 432
 433    /* color buffer */
 434    for (i = 0; i < rast->state.fb.nr_cbufs; i++)
 435       color[i] = tile->color[i] + 4 * block_offset;
 436
 437    /* depth buffer */
 438    depth = tile->depth + block_offset;
 439
 440
 441
 442 #ifdef DEBUG
 443    assert(lp_check_alignment(tile->depth, 16));
 444    assert(lp_check_alignment(tile->color[0], 16));
 445    assert(lp_check_alignment(state->jit_context.blend_color, 16));
 446
 447    assert(lp_check_alignment(inputs->step[0], 16));
 448    assert(lp_check_alignment(inputs->step[1], 16));
 449    assert(lp_check_alignment(inputs->step[2], 16));
 450 #endif
 451
 452    /* run shader */
 453    state->jit_function[1]( &state->jit_context,
 454                         x, y,
 455                         inputs->a0,
 456                         inputs->dadx,
 457                         inputs->dady,
 458                         color,
 459                         depth,
 460                         c1, c2, c3,
 461                         inputs->step[0], inputs->step[1], inputs->step[2]);
 462 }
 463
 464
 465 /**
 466  * Set top row and left column of the tile's pixels to white.  For debugging.
 467  */
 468 static void
 469 outline_tile(uint8_t *tile)
 470 {
 471    const uint8_t val = 0xff;
 472    unsigned i;
 473
 474    for (i = 0; i < TILE_SIZE; i++) {
 475       TILE_PIXEL(tile, i, 0, 0) = val;
 476       TILE_PIXEL(tile, i, 0, 1) = val;
 477       TILE_PIXEL(tile, i, 0, 2) = val;
 478       TILE_PIXEL(tile, i, 0, 3) = val;
 479
 480       TILE_PIXEL(tile, 0, i, 0) = val;
 481       TILE_PIXEL(tile, 0, i, 1) = val;
 482       TILE_PIXEL(tile, 0, i, 2) = val;
 483       TILE_PIXEL(tile, 0, i, 3) = val;
 484    }
 485 }
 486
 487
 488 /**
 489  * Draw grid of gray lines at 16-pixel intervals across the tile to
 490  * show the sub-tile boundaries.  For debugging.
 491  */
 492 static void
 493 outline_subtiles(uint8_t *tile)
 494 {
 495    const uint8_t val = 0x80;
 496    const unsigned step = 16;
 497    unsigned i, j;
 498
 499    for (i = 0; i < TILE_SIZE; i += step) {
 500       for (j = 0; j < TILE_SIZE; j++) {
 501          TILE_PIXEL(tile, i, j, 0) = val;
 502          TILE_PIXEL(tile, i, j, 1) = val;
 503          TILE_PIXEL(tile, i, j, 2) = val;
 504          TILE_PIXEL(tile, i, j, 3) = val;
 505
 506          TILE_PIXEL(tile, j, i, 0) = val;
 507          TILE_PIXEL(tile, j, i, 1) = val;
 508          TILE_PIXEL(tile, j, i, 2) = val;
 509          TILE_PIXEL(tile, j, i, 3) = val;
 510       }
 511    }
 512
 513    outline_tile(tile);
 514 }
 515
 516
 517
 518 /**
 519  * Write the rasterizer's color tile to the framebuffer.
 520  */
 521 static void lp_rast_store_color( struct lp_rasterizer *rast,
 522                                  unsigned thread_index)
 523 {
 524    struct lp_rasterizer_task *task = &rast->tasks[thread_index];
 525    const unsigned x = task->x;
 526    const unsigned y = task->y;
 527    unsigned i;
 528
 529    for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
 530       struct pipe_transfer *transfer = rast->cbuf_transfer[i];
 531       int w = TILE_SIZE;
 532       int h = TILE_SIZE;
 533
 534       if (x >= transfer->width)
 535          continue;
 536
 537       if (y >= transfer->height)
 538          continue;
 539
 540       LP_DBG(DEBUG_RAST, "%s [%u] %d,%d %dx%d\n", __FUNCTION__,
 541              thread_index, x, y, w, h);
 542
 543       if (LP_DEBUG & DEBUG_SHOW_SUBTILES)
 544          outline_subtiles(task->tile.color[i]);
 545       else if (LP_DEBUG & DEBUG_SHOW_TILES)
 546          outline_tile(task->tile.color[i]);
 547
 548       lp_tile_write_4ub(transfer->texture->format,
 549                         task->tile.color[i],
 550                         rast->cbuf_map[i],
 551                         transfer->stride,
 552                         x, y,
 553                         w, h);
 554
 555       LP_COUNT(nr_color_tile_store);
 556    }
 557 }
 558
 559
 560 static void
 561 lp_tile_write_z32(const uint32_t *src, uint8_t *dst, unsigned dst_stride,
 562                   unsigned x0, unsigned y0, unsigned w, unsigned h)
 563 {
 564    unsigned x, y;
 565    uint8_t *dst_row = dst + y0*dst_stride;
 566    for (y = 0; y < h; ++y) {
 567       uint32_t *dst_pixel = (uint32_t *)(dst_row + x0*4);
 568       for (x = 0; x < w; ++x) {
 569          *dst_pixel++ = *src++;
 570       }
 571       dst_row += dst_stride;
 572    }
 573 }
 574
 575 /**
 576  * Write the rasterizer's z/stencil tile to the framebuffer.
 577  */
 578 static void lp_rast_store_zstencil( struct lp_rasterizer *rast,
 579                                     unsigned thread_index )
 580 {
 581    struct lp_rasterizer_task *task = &rast->tasks[thread_index];
 582    const unsigned x = task->x;
 583    const unsigned y = task->y;
 584    unsigned w = TILE_SIZE;
 585    unsigned h = TILE_SIZE;
 586
 587    if (x + w > rast->state.fb.width)
 588       w -= x + w - rast->state.fb.width;
 589
 590    if (y + h > rast->state.fb.height)
 591       h -= y + h - rast->state.fb.height;
 592
 593    LP_DBG(DEBUG_RAST, "%s %d,%d %dx%d\n", __FUNCTION__, x, y, w, h);
 594
 595    assert(rast->zsbuf_transfer->texture->format == PIPE_FORMAT_Z32_UNORM);
 596    lp_tile_write_z32(task->tile.depth,
 597                      rast->zsbuf_map,
 598                      rast->zsbuf_transfer->stride,
 599                      x, y, w, h);
 600 }
 601
 602
 603 /**
 604  * Write the rasterizer's tiles to the framebuffer.
 605  */
 606 static void
 607 lp_rast_end_tile( struct lp_rasterizer *rast,
 608                   unsigned thread_index )
 609 {
 610    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
 611
 612    if (rast->state.write_color)
 613       lp_rast_store_color(rast, thread_index);
 614
 615    if (rast->state.write_zstencil)
 616       lp_rast_store_zstencil(rast, thread_index);
 617 }
 618
 619
 620 /**
 621  * Signal on a fence.  This is called during bin execution/rasterization.
 622  * Called per thread.
 623  */
 624 void lp_rast_fence( struct lp_rasterizer *rast,
 625                     unsigned thread_index,
 626                     const union lp_rast_cmd_arg arg )
 627 {
 628    struct lp_fence *fence = arg.fence;
 629
 630    pipe_mutex_lock( fence->mutex );
 631
 632    fence->count++;
 633    assert(fence->count <= fence->rank);
 634
 635    LP_DBG(DEBUG_RAST, "%s count=%u rank=%u\n", __FUNCTION__,
 636           fence->count, fence->rank);
 637
 638    pipe_condvar_signal( fence->signalled );
 639
 640    pipe_mutex_unlock( fence->mutex );
 641 }
 642
 643
 644 /**
 645  * When all the threads are done rasterizing a scene, one thread will
 646  * call this function to reset the scene and put it onto the empty queue.
 647  */
 648 static void
 649 release_scene( struct lp_rasterizer *rast,
 650                struct lp_scene *scene )
 651 {
 652    util_unreference_framebuffer_state( &scene->fb );
 653
 654    lp_scene_reset( scene );
 655    lp_scene_enqueue( rast->empty_scenes, scene );
 656    rast->curr_scene = NULL;
 657 }
 658
 659
 660 /**
 661  * Rasterize commands for a single bin.
 662  * \param x, y  position of the bin's tile in the framebuffer
 663  * Must be called between lp_rast_begin() and lp_rast_end().
 664  * Called per thread.
 665  */
 666 static void
 667 rasterize_bin( struct lp_rasterizer *rast,
 668                unsigned thread_index,
 669                const struct cmd_bin *bin,
 670                int x, int y)
 671 {
 672    const struct cmd_block_list *commands = &bin->commands;
 673    struct cmd_block *block;
 674    unsigned k;
 675
 676    lp_rast_start_tile( rast, thread_index, x, y );
 677
 678    /* simply execute each of the commands in the block list */
 679    for (block = commands->head; block; block = block->next) {
 680       for (k = 0; k < block->count; k++) {
 681          block->cmd[k]( rast, thread_index, block->arg[k] );
 682       }
 683    }
 684
 685    lp_rast_end_tile( rast, thread_index );
 686 }
 687
 688
 689 #define RAST(x) { lp_rast_##x, #x }
 690
 691 static struct {
 692    lp_rast_cmd cmd;
 693    const char *name;
 694 } cmd_names[] =
 695 {
 696    RAST(load_color),
 697    RAST(load_zstencil),
 698    RAST(clear_color),
 699    RAST(clear_zstencil),
 700    RAST(triangle),
 701    RAST(shade_tile),
 702    RAST(set_state),
 703    RAST(fence),
 704 };
 705
 706 static void
 707 debug_bin( const struct cmd_bin *bin )
 708 {
 709    const struct cmd_block *head = bin->commands.head;
 710    int i, j;
 711
 712    for (i = 0; i < head->count; i++) {
 713       debug_printf("%d: ", i);
 714       for (j = 0; j < Elements(cmd_names); j++) {
 715          if (head->cmd[i] == cmd_names[j].cmd) {
 716             debug_printf("%s\n", cmd_names[j].name);
 717             break;
 718          }
 719       }
 720       if (j == Elements(cmd_names))
 721          debug_printf("...other\n");
 722    }
 723
 724 }
 725
 726 /* An empty bin is one that just loads the contents of the tile and
 727  * stores them again unchanged.  This typically happens when bins have
 728  * been flushed for some reason in the middle of a frame, or when
 729  * incremental updates are being made to a render target.
 730  *
 731  * Try to avoid doing pointless work in this case.
 732  */
 733 static boolean
 734 is_empty_bin( const struct cmd_bin *bin )
 735 {
 736    const struct cmd_block *head = bin->commands.head;
 737    int i;
 738
 739    if (0)
 740       debug_bin(bin);
 741
 742    /* We emit at most two load-tile commands at the start of the first
 743     * command block.  In addition we seem to emit a couple of
 744     * set-state commands even in empty bins.
 745     *
 746     * As a heuristic, if a bin has more than 4 commands, consider it
 747     * non-empty.
 748     */
 749    if (head->next != NULL ||
 750        head->count > 4) {
 751       return FALSE;
 752    }
 753
 754    for (i = 0; i < head->count; i++)
 755       if (head->cmd[i] != lp_rast_load_color &&
 756           head->cmd[i] != lp_rast_load_zstencil &&
 757           head->cmd[i] != lp_rast_set_state) {
 758          return FALSE;
 759       }
 760
 761    return TRUE;
 762 }
 763
 764
 765
 766 /**
 767  * Rasterize/execute all bins within a scene.
 768  * Called per thread.
 769  */
 770 static void
 771 rasterize_scene( struct lp_rasterizer *rast,
 772                 unsigned thread_index,
 773                 struct lp_scene *scene,
 774                 bool write_depth )
 775 {
 776    /* loop over scene bins, rasterize each */
 777 #if 0
 778    {
 779       unsigned i, j;
 780       for (i = 0; i < scene->tiles_x; i++) {
 781          for (j = 0; j < scene->tiles_y; j++) {
 782             struct cmd_bin *bin = lp_get_bin(scene, i, j);
 783             rasterize_bin( rast, thread_index,
 784                            bin, i * TILE_SIZE, j * TILE_SIZE );
 785          }
 786       }
 787    }
 788 #else
 789    {
 790       struct cmd_bin *bin;
 791       int x, y;
 792
 793       assert(scene);
 794       while ((bin = lp_scene_bin_iter_next(scene, &x, &y))) {
 795          if (!is_empty_bin( bin ))
 796             rasterize_bin( rast, thread_index, bin, x * TILE_SIZE, y * TILE_SIZE);
 797       }
 798    }
 799 #endif
 800 }
 801
 802
 803 /**
 804  * Called by setup module when it has something for us to render.
 805  */
 806 void
 807 lp_rasterize_scene( struct lp_rasterizer *rast,
 808                    struct lp_scene *scene,
 809                    const struct pipe_framebuffer_state *fb,
 810                    bool write_depth )
 811 {
 812    boolean debug = false;
 813
 814    LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
 815
 816    if (debug) {
 817       unsigned x, y;
 818       debug_printf("rasterize scene:\n");
 819       debug_printf("  data size: %u\n", lp_scene_data_size(scene));
 820       for (y = 0; y < scene->tiles_y; y++) {
 821          for (x = 0; x < scene->tiles_x; x++) {
 822             debug_printf("  bin %u, %u size: %u\n", x, y,
 823                          lp_scene_bin_size(scene, x, y));
 824          }
 825       }
 826    }
 827
 828    /* save framebuffer state in the bin */
 829    util_copy_framebuffer_state(&scene->fb, fb);
 830    scene->write_depth = write_depth;
 831
 832    if (rast->num_threads == 0) {
 833       /* no threading */
 834
 835       lp_rast_begin( rast, fb,
 836                      fb->nr_cbufs != 0, /* always write color if cbufs present */
 837                      fb->zsbuf != NULL && write_depth );
 838
 839       lp_scene_bin_iter_begin( scene );
 840       rasterize_scene( rast, 0, scene, write_depth );
 841
 842       release_scene( rast, scene );
 843
 844       lp_rast_end( rast );
 845    }
 846    else {
 847       /* threaded rendering! */
 848       unsigned i;
 849
 850       lp_scene_enqueue( rast->full_scenes, scene );
 851
 852       /* signal the threads that there's work to do */
 853       for (i = 0; i < rast->num_threads; i++) {
 854          pipe_semaphore_signal(&rast->tasks[i].work_ready);
 855       }
 856
 857       /* wait for work to complete */
 858       for (i = 0; i < rast->num_threads; i++) {
 859          pipe_semaphore_wait(&rast->tasks[i].work_done);
 860       }
 861    }
 862
 863    LP_DBG(DEBUG_SETUP, "%s done \n", __FUNCTION__);
 864 }
 865
 866
 867 /**
 868  * This is the thread's main entrypoint.
 869  * It's a simple loop:
 870  *   1. wait for work
 871  *   2. do work
 872  *   3. signal that we're done
 873  */
 874 static PIPE_THREAD_ROUTINE( thread_func, init_data )
 875 {
 876    struct lp_rasterizer_task *task = (struct lp_rasterizer_task *) init_data;
 877    struct lp_rasterizer *rast = task->rast;
 878    boolean debug = false;
 879
 880    while (1) {
 881       /* wait for work */
 882       if (debug)
 883          debug_printf("thread %d waiting for work\n", task->thread_index);
 884       pipe_semaphore_wait(&task->work_ready);
 885
 886       if (task->thread_index == 0) {
 887          /* thread[0]:
 888           *  - get next scene to rasterize
 889           *  - map the framebuffer surfaces
 890           */
 891          const struct pipe_framebuffer_state *fb;
 892          boolean write_depth;
 893
 894          rast->curr_scene = lp_scene_dequeue( rast->full_scenes, TRUE );
 895
 896          lp_scene_bin_iter_begin( rast->curr_scene );
 897
 898          fb = &rast->curr_scene->fb;
 899          write_depth = rast->curr_scene->write_depth;
 900
 901          lp_rast_begin( rast, fb,
 902                         fb->nr_cbufs != 0,
 903                         fb->zsbuf != NULL && write_depth );
 904       }
 905
 906       /* Wait for all threads to get here so that threads[1+] don't
 907        * get a null rast->curr_scene pointer.
 908        */
 909       pipe_barrier_wait( &rast->barrier );
 910
 911       /* do work */
 912       if (debug)
 913          debug_printf("thread %d doing work\n", task->thread_index);
 914       rasterize_scene(rast,
 915                      task->thread_index,
 916                      rast->curr_scene,
 917                      rast->curr_scene->write_depth);
 918
 919       /* wait for all threads to finish with this scene */
 920       pipe_barrier_wait( &rast->barrier );
 921
 922       if (task->thread_index == 0) {
 923          /* thread[0]:
 924           * - release the scene object
 925           * - unmap the framebuffer surfaces
 926           */
 927          release_scene( rast, rast->curr_scene );
 928          lp_rast_end( rast );
 929       }
 930
 931       /* signal done with work */
 932       if (debug)
 933          debug_printf("thread %d done working\n", task->thread_index);
 934       pipe_semaphore_signal(&task->work_done);
 935    }
 936
 937    return NULL;
 938 }
 939
 940
 941 /**
 942  * Initialize semaphores and spawn the threads.
 943  */
 944 static void
 945 create_rast_threads(struct lp_rasterizer *rast)
 946 {
 947    unsigned i;
 948
 949 #ifdef PIPE_OS_WINDOWS
 950    /* Multithreading not supported on windows until conditions and barriers are
 951     * properly implemented. */
 952    rast->num_threads = 0;
 953 #else
 954    rast->num_threads = util_cpu_caps.nr_cpus;
 955    rast->num_threads = debug_get_num_option("LP_NUM_THREADS", rast->num_threads);
 956    rast->num_threads = MIN2(rast->num_threads, MAX_THREADS);
 957 #endif
 958
 959    /* NOTE: if num_threads is zero, we won't use any threads */
 960    for (i = 0; i < rast->num_threads; i++) {
 961       pipe_semaphore_init(&rast->tasks[i].work_ready, 0);
 962       pipe_semaphore_init(&rast->tasks[i].work_done, 0);
 963       rast->threads[i] = pipe_thread_create(thread_func,
 964                                             (void *) &rast->tasks[i]);
 965    }
 966 }
 967
 968
 969
 970 /**
 971  * Create new lp_rasterizer.
 972  * \param empty  the queue to put empty scenes on after we've finished
 973  *               processing them.
 974  */
 975 struct lp_rasterizer *
 976 lp_rast_create( struct pipe_screen *screen, struct lp_scene_queue *empty )
 977 {
 978    struct lp_rasterizer *rast;
 979    unsigned i, cbuf;
 980
 981    rast = CALLOC_STRUCT(lp_rasterizer);
 982    if(!rast)
 983       return NULL;
 984
 985    rast->screen = screen;
 986
 987    rast->empty_scenes = empty;
 988    rast->full_scenes = lp_scene_queue_create();
 989
 990    for (i = 0; i < Elements(rast->tasks); i++) {
 991       struct lp_rasterizer_task *task = &rast->tasks[i];
 992
 993       for (cbuf = 0; cbuf < PIPE_MAX_COLOR_BUFS; cbuf++ )
 994          task->tile.color[cbuf] = align_malloc(TILE_SIZE * TILE_SIZE * 4, 16);
 995
 996       task->tile.depth = align_malloc(TILE_SIZE * TILE_SIZE * 4, 16);
 997       task->rast = rast;
 998       task->thread_index = i;
 999    }
1000
1001    create_rast_threads(rast);
1002
1003    /* for synchronizing rasterization threads */
1004    pipe_barrier_init( &rast->barrier, rast->num_threads );
1005
1006    return rast;
1007 }
1008
1009
1010 /* Shutdown:
1011  */
1012 void lp_rast_destroy( struct lp_rasterizer *rast )
1013 {
1014    unsigned i, cbuf;
1015
1016    util_unreference_framebuffer_state(&rast->state.fb);
1017
1018    for (i = 0; i < Elements(rast->tasks); i++) {
1019       align_free(rast->tasks[i].tile.depth);
1020       for (cbuf = 0; cbuf < PIPE_MAX_COLOR_BUFS; cbuf++ )
1021          align_free(rast->tasks[i].tile.color[cbuf]);
1022    }
1023
1024    /* for synchronizing rasterization threads */
1025    pipe_barrier_destroy( &rast->barrier );
1026
1027    FREE(rast);
1028 }
1029
1030
1031 /** Return number of rasterization threads */
1032 unsigned
1033 lp_rast_get_num_threads( struct lp_rasterizer *rast )
1034 {
1035    return rast->num_threads;
1036 }