src/gallium/auxiliary/draw/draw_pt_fetch_shade_pipeline_llvm.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2010 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 "util/u_math.h"
  29 #include "util/u_memory.h"
  30 #include "util/u_prim.h"
  31 #include "draw/draw_context.h"
  32 #include "draw/draw_gs.h"
  33 #include "draw/draw_vbuf.h"
  34 #include "draw/draw_vertex.h"
  35 #include "draw/draw_pt.h"
  36 #include "draw/draw_prim_assembler.h"
  37 #include "draw/draw_vs.h"
  38 #include "draw/draw_llvm.h"
  39 #include "gallivm/lp_bld_init.h"
  40
  41
  42 struct llvm_middle_end {
  43    struct draw_pt_middle_end base;
  44    struct draw_context *draw;
  45
  46    struct pt_emit *emit;
  47    struct pt_so_emit *so_emit;
  48    struct pt_fetch *fetch;
  49    struct pt_post_vs *post_vs;
  50
  51
  52    unsigned vertex_data_offset;
  53    unsigned vertex_size;
  54    unsigned input_prim;
  55    unsigned opt;
  56
  57    struct draw_llvm *llvm;
  58    struct draw_llvm_variant *current_variant;
  59 };
  60
  61
  62 /** cast wrapper */
  63 static INLINE struct llvm_middle_end *
  64 llvm_middle_end(struct draw_pt_middle_end *middle)
  65 {
  66    return (struct llvm_middle_end *) middle;
  67 }
  68
  69
  70 static void
  71 llvm_middle_end_prepare_gs(struct llvm_middle_end *fpme)
  72 {
  73    struct draw_context *draw = fpme->draw;
  74    struct draw_geometry_shader *gs = draw->gs.geometry_shader;
  75    struct draw_gs_llvm_variant_key *key;
  76    struct draw_gs_llvm_variant *variant = NULL;
  77    struct draw_gs_llvm_variant_list_item *li;
  78    struct llvm_geometry_shader *shader = llvm_geometry_shader(gs);
  79    char store[DRAW_GS_LLVM_MAX_VARIANT_KEY_SIZE];
  80    unsigned i;
  81
  82    key = draw_gs_llvm_make_variant_key(fpme->llvm, store);
  83
  84    /* Search shader's list of variants for the key */
  85    li = first_elem(&shader->variants);
  86    while (!at_end(&shader->variants, li)) {
  87       if (memcmp(&li->base->key, key, shader->variant_key_size) == 0) {
  88          variant = li->base;
  89          break;
  90       }
  91       li = next_elem(li);
  92    }
  93
  94    if (variant) {
  95       /* found the variant, move to head of global list (for LRU) */
  96       move_to_head(&fpme->llvm->gs_variants_list,
  97                    &variant->list_item_global);
  98    }
  99    else {
 100       /* Need to create new variant */
 101
 102       /* First check if we've created too many variants.  If so, free
 103        * 25% of the LRU to avoid using too much memory.
 104        */
 105       if (fpme->llvm->nr_gs_variants >= DRAW_MAX_SHADER_VARIANTS) {
 106          /*
 107           * XXX: should we flush here ?
 108           */
 109          for (i = 0; i < DRAW_MAX_SHADER_VARIANTS / 4; i++) {
 110             struct draw_gs_llvm_variant_list_item *item;
 111             if (is_empty_list(&fpme->llvm->gs_variants_list)) {
 112                break;
 113             }
 114             item = last_elem(&fpme->llvm->gs_variants_list);
 115             assert(item);
 116             assert(item->base);
 117             draw_gs_llvm_destroy_variant(item->base);
 118          }
 119       }
 120
 121       variant = draw_gs_llvm_create_variant(fpme->llvm, gs->info.num_outputs, key);
 122
 123       if (variant) {
 124          insert_at_head(&shader->variants, &variant->list_item_local);
 125          insert_at_head(&fpme->llvm->gs_variants_list,
 126                         &variant->list_item_global);
 127          fpme->llvm->nr_gs_variants++;
 128          shader->variants_cached++;
 129       }
 130    }
 131
 132    gs->current_variant = variant;
 133 }
 134
 135 /**
 136  * Prepare/validate middle part of the vertex pipeline.
 137  * NOTE: if you change this function, also look at the non-LLVM
 138  * function fetch_pipeline_prepare() for similar changes.
 139  */
 140 static void
 141 llvm_middle_end_prepare( struct draw_pt_middle_end *middle,
 142                          unsigned in_prim,
 143                          unsigned opt,
 144                          unsigned *max_vertices )
 145 {
 146    struct llvm_middle_end *fpme = llvm_middle_end(middle);
 147    struct draw_context *draw = fpme->draw;
 148    struct draw_vertex_shader *vs = draw->vs.vertex_shader;
 149    struct draw_geometry_shader *gs = draw->gs.geometry_shader;
 150    const unsigned out_prim = gs ? gs->output_primitive :
 151       u_assembled_prim(in_prim);
 152    unsigned point_clip = draw->rasterizer->fill_front == PIPE_POLYGON_MODE_POINT ||
 153                          out_prim == PIPE_PRIM_POINTS;
 154    unsigned nr;
 155
 156    fpme->input_prim = in_prim;
 157    fpme->opt = opt;
 158
 159    draw_pt_post_vs_prepare( fpme->post_vs,
 160                             draw->clip_xy,
 161                             draw->clip_z,
 162                             draw->clip_user,
 163                             point_clip ? draw->guard_band_points_xy :
 164                                          draw->guard_band_xy,
 165                             draw->identity_viewport,
 166                             draw->rasterizer->clip_halfz,
 167                             (draw->vs.edgeflag_output ? TRUE : FALSE) );
 168
 169    draw_pt_so_emit_prepare( fpme->so_emit, gs == NULL );
 170
 171    if (!(opt & PT_PIPELINE)) {
 172       draw_pt_emit_prepare( fpme->emit,
 173                             out_prim,
 174                             max_vertices );
 175
 176       *max_vertices = MAX2( *max_vertices, 4096 );
 177    }
 178    else {
 179       /* limit max fetches by limiting max_vertices */
 180       *max_vertices = 4096;
 181    }
 182
 183    /* Get the number of float[4] attributes per vertex.
 184     * Note: this must be done after draw_pt_emit_prepare() since that
 185     * can effect the vertex size.
 186     */
 187    nr = MAX2(vs->info.num_inputs, draw_total_vs_outputs(draw));
 188
 189    /* Always leave room for the vertex header whether we need it or
 190     * not.  It's hard to get rid of it in particular because of the
 191     * viewport code in draw_pt_post_vs.c.
 192     */
 193    fpme->vertex_size = sizeof(struct vertex_header) + nr * 4 * sizeof(float);
 194
 195    /* return even number */
 196    *max_vertices = *max_vertices & ~1;
 197
 198    /* Find/create the vertex shader variant */
 199    {
 200       struct draw_llvm_variant_key *key;
 201       struct draw_llvm_variant *variant = NULL;
 202       struct draw_llvm_variant_list_item *li;
 203       struct llvm_vertex_shader *shader = llvm_vertex_shader(vs);
 204       char store[DRAW_LLVM_MAX_VARIANT_KEY_SIZE];
 205       unsigned i;
 206
 207       key = draw_llvm_make_variant_key(fpme->llvm, store);
 208
 209       /* Search shader's list of variants for the key */
 210       li = first_elem(&shader->variants);
 211       while (!at_end(&shader->variants, li)) {
 212          if (memcmp(&li->base->key, key, shader->variant_key_size) == 0) {
 213             variant = li->base;
 214             break;
 215          }
 216          li = next_elem(li);
 217       }
 218
 219       if (variant) {
 220          /* found the variant, move to head of global list (for LRU) */
 221          move_to_head(&fpme->llvm->vs_variants_list,
 222                       &variant->list_item_global);
 223       }
 224       else {
 225          /* Need to create new variant */
 226
 227          /* First check if we've created too many variants.  If so, free
 228           * 25% of the LRU to avoid using too much memory.
 229           */
 230          if (fpme->llvm->nr_variants >= DRAW_MAX_SHADER_VARIANTS) {
 231             /*
 232              * XXX: should we flush here ?
 233              */
 234             for (i = 0; i < DRAW_MAX_SHADER_VARIANTS / 4; i++) {
 235                struct draw_llvm_variant_list_item *item;
 236                if (is_empty_list(&fpme->llvm->vs_variants_list)) {
 237                   break;
 238                }
 239                item = last_elem(&fpme->llvm->vs_variants_list);
 240                assert(item);
 241                assert(item->base);
 242                draw_llvm_destroy_variant(item->base);
 243             }
 244          }
 245
 246          variant = draw_llvm_create_variant(fpme->llvm, nr, key);
 247
 248          if (variant) {
 249             insert_at_head(&shader->variants, &variant->list_item_local);
 250             insert_at_head(&fpme->llvm->vs_variants_list,
 251                            &variant->list_item_global);
 252             fpme->llvm->nr_variants++;
 253             shader->variants_cached++;
 254          }
 255       }
 256
 257       fpme->current_variant = variant;
 258    }
 259
 260    if (gs) {
 261       llvm_middle_end_prepare_gs(fpme);
 262    }
 263 }
 264
 265
 266 /**
 267  * Bind/update constant buffer pointers, clip planes and viewport dims.
 268  * These are "light weight" parameters which aren't baked into the
 269  * generated code.  Updating these items is much cheaper than revalidating
 270  * and rebuilding the generated pipeline code.
 271  */
 272 static void
 273 llvm_middle_end_bind_parameters(struct draw_pt_middle_end *middle)
 274 {
 275    struct llvm_middle_end *fpme = llvm_middle_end(middle);
 276    struct draw_context *draw = fpme->draw;
 277    unsigned i;
 278
 279    for (i = 0; i < Elements(fpme->llvm->jit_context.vs_constants); ++i) {
 280       int num_consts =
 281          draw->pt.user.vs_constants_size[i] / (sizeof(float) * 4);
 282       fpme->llvm->jit_context.vs_constants[i] = draw->pt.user.vs_constants[i];
 283       fpme->llvm->jit_context.num_vs_constants[i] = num_consts;
 284    }
 285    for (i = 0; i < Elements(fpme->llvm->gs_jit_context.constants); ++i) {
 286       int num_consts =
 287          draw->pt.user.gs_constants_size[i] / (sizeof(float) * 4);
 288       fpme->llvm->gs_jit_context.constants[i] = draw->pt.user.gs_constants[i];
 289       fpme->llvm->gs_jit_context.num_constants[i] = num_consts;
 290    }
 291
 292    fpme->llvm->jit_context.planes =
 293       (float (*)[DRAW_TOTAL_CLIP_PLANES][4]) draw->pt.user.planes[0];
 294    fpme->llvm->gs_jit_context.planes =
 295       (float (*)[DRAW_TOTAL_CLIP_PLANES][4]) draw->pt.user.planes[0];
 296
 297    fpme->llvm->jit_context.viewport = (float *) draw->viewports[0].scale;
 298    fpme->llvm->gs_jit_context.viewport = (float *) draw->viewports[0].scale;
 299 }
 300
 301
 302 static void
 303 pipeline(struct llvm_middle_end *llvm,
 304          const struct draw_vertex_info *vert_info,
 305          const struct draw_prim_info *prim_info)
 306 {
 307    if (prim_info->linear)
 308       draw_pipeline_run_linear( llvm->draw,
 309                                 vert_info,
 310                                 prim_info);
 311    else
 312       draw_pipeline_run( llvm->draw,
 313                          vert_info,
 314                          prim_info );
 315 }
 316
 317
 318 static void
 319 emit(struct pt_emit *emit,
 320      const struct draw_vertex_info *vert_info,
 321      const struct draw_prim_info *prim_info)
 322 {
 323    if (prim_info->linear) {
 324       draw_pt_emit_linear(emit, vert_info, prim_info);
 325    }
 326    else {
 327       draw_pt_emit(emit, vert_info, prim_info);
 328    }
 329 }
 330
 331
 332 static void
 333 llvm_pipeline_generic(struct draw_pt_middle_end *middle,
 334                       const struct draw_fetch_info *fetch_info,
 335                       const struct draw_prim_info *in_prim_info)
 336 {
 337    struct llvm_middle_end *fpme = llvm_middle_end(middle);
 338    struct draw_context *draw = fpme->draw;
 339    struct draw_geometry_shader *gshader = draw->gs.geometry_shader;
 340    struct draw_prim_info gs_prim_info;
 341    struct draw_vertex_info llvm_vert_info;
 342    struct draw_vertex_info gs_vert_info;
 343    struct draw_vertex_info *vert_info;
 344    struct draw_prim_info ia_prim_info;
 345    struct draw_vertex_info ia_vert_info;
 346    const struct draw_prim_info *prim_info = in_prim_info;
 347    boolean free_prim_info = FALSE;
 348    unsigned opt = fpme->opt;
 349    unsigned clipped = 0;
 350
 351    llvm_vert_info.count = fetch_info->count;
 352    llvm_vert_info.vertex_size = fpme->vertex_size;
 353    llvm_vert_info.stride = fpme->vertex_size;
 354    llvm_vert_info.verts = (struct vertex_header *)
 355       MALLOC(fpme->vertex_size *
 356              align(fetch_info->count, lp_native_vector_width / 32));
 357    if (!llvm_vert_info.verts) {
 358       assert(0);
 359       return;
 360    }
 361
 362    if (draw->collect_statistics) {
 363       draw->statistics.ia_vertices += prim_info->count;
 364       draw->statistics.ia_primitives +=
 365          u_decomposed_prims_for_vertices(prim_info->prim, prim_info->count);
 366       draw->statistics.vs_invocations += fetch_info->count;
 367    }
 368
 369    if (fetch_info->linear)
 370       clipped = fpme->current_variant->jit_func( &fpme->llvm->jit_context,
 371                                        llvm_vert_info.verts,
 372                                        draw->pt.user.vbuffer,
 373                                        fetch_info->start,
 374                                        fetch_info->count,
 375                                        fpme->vertex_size,
 376                                        draw->pt.vertex_buffer,
 377                                        draw->instance_id,
 378                                        draw->start_index,
 379                                        draw->start_instance);
 380    else
 381       clipped = fpme->current_variant->jit_func_elts( &fpme->llvm->jit_context,
 382                                             llvm_vert_info.verts,
 383                                             draw->pt.user.vbuffer,
 384                                             fetch_info->elts,
 385                                             draw->pt.user.eltMax,
 386                                             fetch_info->count,
 387                                             fpme->vertex_size,
 388                                             draw->pt.vertex_buffer,
 389                                             draw->instance_id,
 390                                             draw->pt.user.eltBias,
 391                                             draw->start_instance);
 392
 393    /* Finished with fetch and vs:
 394     */
 395    fetch_info = NULL;
 396    vert_info = &llvm_vert_info;
 397
 398    if ((opt & PT_SHADE) && gshader) {
 399       struct draw_vertex_shader *vshader = draw->vs.vertex_shader;
 400       draw_geometry_shader_run(gshader,
 401                                draw->pt.user.gs_constants,
 402                                draw->pt.user.gs_constants_size,
 403                                vert_info,
 404                                prim_info,
 405                                &vshader->info,
 406                                &gs_vert_info,
 407                                &gs_prim_info);
 408
 409       FREE(vert_info->verts);
 410       vert_info = &gs_vert_info;
 411       prim_info = &gs_prim_info;
 412    } else {
 413       if (draw_prim_assembler_is_required(draw, prim_info, vert_info)) {
 414          draw_prim_assembler_run(draw, prim_info, vert_info,
 415                                  &ia_prim_info, &ia_vert_info);
 416
 417          if (ia_vert_info.count) {
 418             FREE(vert_info->verts);
 419             vert_info = &ia_vert_info;
 420             prim_info = &ia_prim_info;
 421             free_prim_info = TRUE;
 422          }
 423       }
 424    }
 425    if (prim_info->count == 0) {
 426       debug_printf("GS/IA didn't emit any vertices!\n");
 427
 428       FREE(vert_info->verts);
 429       if (free_prim_info) {
 430          FREE(prim_info->primitive_lengths);
 431       }
 432       return;
 433    }
 434
 435    /* stream output needs to be done before clipping */
 436    draw_pt_so_emit( fpme->so_emit, vert_info, prim_info );
 437
 438    draw_stats_clipper_primitives(draw, prim_info);
 439
 440    /*
 441     * if there's no position, need to stop now, or the latter stages
 442     * will try to access non-existent position output.
 443     */
 444    if (draw_current_shader_position_output(draw) != -1) {
 445       if ((opt & PT_SHADE) && gshader) {
 446          clipped = draw_pt_post_vs_run( fpme->post_vs, vert_info, prim_info );
 447       }
 448       if (clipped) {
 449          opt |= PT_PIPELINE;
 450       }
 451
 452       /* Do we need to run the pipeline? Now will come here if clipped
 453        */
 454       if (opt & PT_PIPELINE) {
 455          pipeline( fpme, vert_info, prim_info );
 456       }
 457       else {
 458          emit( fpme->emit, vert_info, prim_info );
 459       }
 460    }
 461    FREE(vert_info->verts);
 462    if (free_prim_info) {
 463       FREE(prim_info->primitive_lengths);
 464    }
 465 }
 466
 467
 468 static void
 469 llvm_middle_end_run(struct draw_pt_middle_end *middle,
 470                     const unsigned *fetch_elts,
 471                     unsigned fetch_count,
 472                     const ushort *draw_elts,
 473                     unsigned draw_count,
 474                     unsigned prim_flags)
 475 {
 476    struct llvm_middle_end *fpme = llvm_middle_end(middle);
 477    struct draw_fetch_info fetch_info;
 478    struct draw_prim_info prim_info;
 479
 480    fetch_info.linear = FALSE;
 481    fetch_info.start = 0;
 482    fetch_info.elts = fetch_elts;
 483    fetch_info.count = fetch_count;
 484
 485    prim_info.linear = FALSE;
 486    prim_info.start = 0;
 487    prim_info.count = draw_count;
 488    prim_info.elts = draw_elts;
 489    prim_info.prim = fpme->input_prim;
 490    prim_info.flags = prim_flags;
 491    prim_info.primitive_count = 1;
 492    prim_info.primitive_lengths = &draw_count;
 493
 494    llvm_pipeline_generic( middle, &fetch_info, &prim_info );
 495 }
 496
 497
 498 static void
 499 llvm_middle_end_linear_run(struct draw_pt_middle_end *middle,
 500                            unsigned start,
 501                            unsigned count,
 502                            unsigned prim_flags)
 503 {
 504    struct llvm_middle_end *fpme = llvm_middle_end(middle);
 505    struct draw_fetch_info fetch_info;
 506    struct draw_prim_info prim_info;
 507
 508    fetch_info.linear = TRUE;
 509    fetch_info.start = start;
 510    fetch_info.count = count;
 511    fetch_info.elts = NULL;
 512
 513    prim_info.linear = TRUE;
 514    prim_info.start = 0;
 515    prim_info.count = count;
 516    prim_info.elts = NULL;
 517    prim_info.prim = fpme->input_prim;
 518    prim_info.flags = prim_flags;
 519    prim_info.primitive_count = 1;
 520    prim_info.primitive_lengths = &count;
 521
 522    llvm_pipeline_generic( middle, &fetch_info, &prim_info );
 523 }
 524
 525
 526 static boolean
 527 llvm_middle_end_linear_run_elts(struct draw_pt_middle_end *middle,
 528                                 unsigned start,
 529                                 unsigned count,
 530                                 const ushort *draw_elts,
 531                                 unsigned draw_count,
 532                                 unsigned prim_flags)
 533 {
 534    struct llvm_middle_end *fpme = llvm_middle_end(middle);
 535    struct draw_fetch_info fetch_info;
 536    struct draw_prim_info prim_info;
 537
 538    fetch_info.linear = TRUE;
 539    fetch_info.start = start;
 540    fetch_info.count = count;
 541    fetch_info.elts = NULL;
 542
 543    prim_info.linear = FALSE;
 544    prim_info.start = 0;
 545    prim_info.count = draw_count;
 546    prim_info.elts = draw_elts;
 547    prim_info.prim = fpme->input_prim;
 548    prim_info.flags = prim_flags;
 549    prim_info.primitive_count = 1;
 550    prim_info.primitive_lengths = &draw_count;
 551
 552    llvm_pipeline_generic( middle, &fetch_info, &prim_info );
 553
 554    return TRUE;
 555 }
 556
 557
 558 static void
 559 llvm_middle_end_finish(struct draw_pt_middle_end *middle)
 560 {
 561    /* nothing to do */
 562 }
 563
 564
 565 static void
 566 llvm_middle_end_destroy(struct draw_pt_middle_end *middle)
 567 {
 568    struct llvm_middle_end *fpme = llvm_middle_end(middle);
 569
 570    if (fpme->fetch)
 571       draw_pt_fetch_destroy( fpme->fetch );
 572
 573    if (fpme->emit)
 574       draw_pt_emit_destroy( fpme->emit );
 575
 576    if (fpme->so_emit)
 577       draw_pt_so_emit_destroy( fpme->so_emit );
 578
 579    if (fpme->post_vs)
 580       draw_pt_post_vs_destroy( fpme->post_vs );
 581
 582    FREE(middle);
 583 }
 584
 585
 586 struct draw_pt_middle_end *
 587 draw_pt_fetch_pipeline_or_emit_llvm(struct draw_context *draw)
 588 {
 589    struct llvm_middle_end *fpme = 0;
 590
 591    if (!draw->llvm)
 592       return NULL;
 593
 594    fpme = CALLOC_STRUCT( llvm_middle_end );
 595    if (!fpme)
 596       goto fail;
 597
 598    fpme->base.prepare         = llvm_middle_end_prepare;
 599    fpme->base.bind_parameters = llvm_middle_end_bind_parameters;
 600    fpme->base.run             = llvm_middle_end_run;
 601    fpme->base.run_linear      = llvm_middle_end_linear_run;
 602    fpme->base.run_linear_elts = llvm_middle_end_linear_run_elts;
 603    fpme->base.finish          = llvm_middle_end_finish;
 604    fpme->base.destroy         = llvm_middle_end_destroy;
 605
 606    fpme->draw = draw;
 607
 608    fpme->fetch = draw_pt_fetch_create( draw );
 609    if (!fpme->fetch)
 610       goto fail;
 611
 612    fpme->post_vs = draw_pt_post_vs_create( draw );
 613    if (!fpme->post_vs)
 614       goto fail;
 615
 616    fpme->emit = draw_pt_emit_create( draw );
 617    if (!fpme->emit)
 618       goto fail;
 619
 620    fpme->so_emit = draw_pt_so_emit_create( draw );
 621    if (!fpme->so_emit)
 622       goto fail;
 623
 624    fpme->llvm = draw->llvm;
 625    if (!fpme->llvm)
 626       goto fail;
 627
 628    fpme->current_variant = NULL;
 629
 630    return &fpme->base;
 631
 632  fail:
 633    if (fpme)
 634       llvm_middle_end_destroy( &fpme->base );
 635
 636    return NULL;
 637 }