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->bypass_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, out_prim,
 173                             max_vertices );
 174
 175       *max_vertices = MAX2( *max_vertices, 4096 );
 176    }
 177    else {
 178       /* limit max fetches by limiting max_vertices */
 179       *max_vertices = 4096;
 180    }
 181
 182    /* Get the number of float[4] attributes per vertex.
 183     * Note: this must be done after draw_pt_emit_prepare() since that
 184     * can effect the vertex size.
 185     */
 186    nr = MAX2(vs->info.num_inputs, draw_total_vs_outputs(draw));
 187
 188    /* Always leave room for the vertex header whether we need it or
 189     * not.  It's hard to get rid of it in particular because of the
 190     * viewport code in draw_pt_post_vs.c.
 191     */
 192    fpme->vertex_size = sizeof(struct vertex_header) + nr * 4 * sizeof(float);
 193
 194    /* return even number */
 195    *max_vertices = *max_vertices & ~1;
 196
 197    /* Find/create the vertex shader variant */
 198    {
 199       struct draw_llvm_variant_key *key;
 200       struct draw_llvm_variant *variant = NULL;
 201       struct draw_llvm_variant_list_item *li;
 202       struct llvm_vertex_shader *shader = llvm_vertex_shader(vs);
 203       char store[DRAW_LLVM_MAX_VARIANT_KEY_SIZE];
 204       unsigned i;
 205
 206       key = draw_llvm_make_variant_key(fpme->llvm, store);
 207
 208       /* Search shader's list of variants for the key */
 209       li = first_elem(&shader->variants);
 210       while (!at_end(&shader->variants, li)) {
 211          if (memcmp(&li->base->key, key, shader->variant_key_size) == 0) {
 212             variant = li->base;
 213             break;
 214          }
 215          li = next_elem(li);
 216       }
 217
 218       if (variant) {
 219          /* found the variant, move to head of global list (for LRU) */
 220          move_to_head(&fpme->llvm->vs_variants_list,
 221                       &variant->list_item_global);
 222       }
 223       else {
 224          /* Need to create new variant */
 225
 226          /* First check if we've created too many variants.  If so, free
 227           * 25% of the LRU to avoid using too much memory.
 228           */
 229          if (fpme->llvm->nr_variants >= DRAW_MAX_SHADER_VARIANTS) {
 230             /*
 231              * XXX: should we flush here ?
 232              */
 233             for (i = 0; i < DRAW_MAX_SHADER_VARIANTS / 4; i++) {
 234                struct draw_llvm_variant_list_item *item;
 235                if (is_empty_list(&fpme->llvm->vs_variants_list)) {
 236                   break;
 237                }
 238                item = last_elem(&fpme->llvm->vs_variants_list);
 239                assert(item);
 240                assert(item->base);
 241                draw_llvm_destroy_variant(item->base);
 242             }
 243          }
 244
 245          variant = draw_llvm_create_variant(fpme->llvm, nr, key);
 246
 247          if (variant) {
 248             insert_at_head(&shader->variants, &variant->list_item_local);
 249             insert_at_head(&fpme->llvm->vs_variants_list,
 250                            &variant->list_item_global);
 251             fpme->llvm->nr_variants++;
 252             shader->variants_cached++;
 253          }
 254       }
 255
 256       fpme->current_variant = variant;
 257    }
 258
 259    if (gs) {
 260       llvm_middle_end_prepare_gs(fpme);
 261    }
 262 }
 263
 264
 265 /**
 266  * Bind/update constant buffer pointers, clip planes and viewport dims.
 267  * These are "light weight" parameters which aren't baked into the
 268  * generated code.  Updating these items is much cheaper than revalidating
 269  * and rebuilding the generated pipeline code.
 270  */
 271 static void
 272 llvm_middle_end_bind_parameters(struct draw_pt_middle_end *middle)
 273 {
 274    static const float fake_const_buf[4];
 275    struct llvm_middle_end *fpme = llvm_middle_end(middle);
 276    struct draw_context *draw = fpme->draw;
 277    struct draw_llvm *llvm = fpme->llvm;
 278    unsigned i;
 279
 280    for (i = 0; i < Elements(llvm->jit_context.vs_constants); ++i) {
 281       int num_consts =
 282          draw->pt.user.vs_constants_size[i] / (sizeof(float) * 4);
 283       llvm->jit_context.vs_constants[i] = draw->pt.user.vs_constants[i];
 284       llvm->jit_context.num_vs_constants[i] = num_consts;
 285       if (num_consts == 0) {
 286          llvm->jit_context.vs_constants[i] = fake_const_buf;
 287       }
 288    }
 289    for (i = 0; i < Elements(llvm->gs_jit_context.constants); ++i) {
 290       int num_consts =
 291          draw->pt.user.gs_constants_size[i] / (sizeof(float) * 4);
 292       llvm->gs_jit_context.constants[i] = draw->pt.user.gs_constants[i];
 293       llvm->gs_jit_context.num_constants[i] = num_consts;
 294       if (num_consts == 0) {
 295          llvm->gs_jit_context.constants[i] = fake_const_buf;
 296       }
 297    }
 298
 299    llvm->jit_context.planes =
 300       (float (*)[DRAW_TOTAL_CLIP_PLANES][4]) draw->pt.user.planes[0];
 301    llvm->gs_jit_context.planes =
 302       (float (*)[DRAW_TOTAL_CLIP_PLANES][4]) draw->pt.user.planes[0];
 303
 304    llvm->jit_context.viewports = draw->viewports;
 305    llvm->gs_jit_context.viewports = draw->viewports;
 306 }
 307
 308
 309 static void
 310 pipeline(struct llvm_middle_end *llvm,
 311          const struct draw_vertex_info *vert_info,
 312          const struct draw_prim_info *prim_info)
 313 {
 314    if (prim_info->linear)
 315       draw_pipeline_run_linear( llvm->draw,
 316                                 vert_info,
 317                                 prim_info);
 318    else
 319       draw_pipeline_run( llvm->draw,
 320                          vert_info,
 321                          prim_info );
 322 }
 323
 324
 325 static void
 326 emit(struct pt_emit *emit,
 327      const struct draw_vertex_info *vert_info,
 328      const struct draw_prim_info *prim_info)
 329 {
 330    if (prim_info->linear) {
 331       draw_pt_emit_linear(emit, vert_info, prim_info);
 332    }
 333    else {
 334       draw_pt_emit(emit, vert_info, prim_info);
 335    }
 336 }
 337
 338
 339 static void
 340 llvm_pipeline_generic(struct draw_pt_middle_end *middle,
 341                       const struct draw_fetch_info *fetch_info,
 342                       const struct draw_prim_info *in_prim_info)
 343 {
 344    struct llvm_middle_end *fpme = llvm_middle_end(middle);
 345    struct draw_context *draw = fpme->draw;
 346    struct draw_geometry_shader *gshader = draw->gs.geometry_shader;
 347    struct draw_prim_info gs_prim_info;
 348    struct draw_vertex_info llvm_vert_info;
 349    struct draw_vertex_info gs_vert_info;
 350    struct draw_vertex_info *vert_info;
 351    struct draw_prim_info ia_prim_info;
 352    struct draw_vertex_info ia_vert_info;
 353    const struct draw_prim_info *prim_info = in_prim_info;
 354    boolean free_prim_info = FALSE;
 355    unsigned opt = fpme->opt;
 356    unsigned clipped = 0;
 357
 358    llvm_vert_info.count = fetch_info->count;
 359    llvm_vert_info.vertex_size = fpme->vertex_size;
 360    llvm_vert_info.stride = fpme->vertex_size;
 361    llvm_vert_info.verts = (struct vertex_header *)
 362       MALLOC(fpme->vertex_size *
 363              align(fetch_info->count, lp_native_vector_width / 32));
 364    if (!llvm_vert_info.verts) {
 365       assert(0);
 366       return;
 367    }
 368
 369    if (draw->collect_statistics) {
 370       draw->statistics.ia_vertices += prim_info->count;
 371       draw->statistics.ia_primitives +=
 372          u_decomposed_prims_for_vertices(prim_info->prim, prim_info->count);
 373       draw->statistics.vs_invocations += fetch_info->count;
 374    }
 375
 376    if (fetch_info->linear)
 377       clipped = fpme->current_variant->jit_func( &fpme->llvm->jit_context,
 378                                        llvm_vert_info.verts,
 379                                        draw->pt.user.vbuffer,
 380                                        fetch_info->start,
 381                                        fetch_info->count,
 382                                        fpme->vertex_size,
 383                                        draw->pt.vertex_buffer,
 384                                        draw->instance_id,
 385                                        draw->start_index,
 386                                        draw->start_instance);
 387    else
 388       clipped = fpme->current_variant->jit_func_elts( &fpme->llvm->jit_context,
 389                                             llvm_vert_info.verts,
 390                                             draw->pt.user.vbuffer,
 391                                             fetch_info->elts,
 392                                             draw->pt.user.eltMax,
 393                                             fetch_info->count,
 394                                             fpme->vertex_size,
 395                                             draw->pt.vertex_buffer,
 396                                             draw->instance_id,
 397                                             draw->pt.user.eltBias,
 398                                             draw->start_instance);
 399
 400    /* Finished with fetch and vs:
 401     */
 402    fetch_info = NULL;
 403    vert_info = &llvm_vert_info;
 404
 405    if ((opt & PT_SHADE) && gshader) {
 406       struct draw_vertex_shader *vshader = draw->vs.vertex_shader;
 407       draw_geometry_shader_run(gshader,
 408                                draw->pt.user.gs_constants,
 409                                draw->pt.user.gs_constants_size,
 410                                vert_info,
 411                                prim_info,
 412                                &vshader->info,
 413                                &gs_vert_info,
 414                                &gs_prim_info);
 415
 416       FREE(vert_info->verts);
 417       vert_info = &gs_vert_info;
 418       prim_info = &gs_prim_info;
 419    } else {
 420       if (draw_prim_assembler_is_required(draw, prim_info, vert_info)) {
 421          draw_prim_assembler_run(draw, prim_info, vert_info,
 422                                  &ia_prim_info, &ia_vert_info);
 423
 424          if (ia_vert_info.count) {
 425             FREE(vert_info->verts);
 426             vert_info = &ia_vert_info;
 427             prim_info = &ia_prim_info;
 428             free_prim_info = TRUE;
 429          }
 430       }
 431    }
 432    if (prim_info->count == 0) {
 433       debug_printf("GS/IA didn't emit any vertices!\n");
 434
 435       FREE(vert_info->verts);
 436       if (free_prim_info) {
 437          FREE(prim_info->primitive_lengths);
 438       }
 439       return;
 440    }
 441
 442    /* stream output needs to be done before clipping */
 443    draw_pt_so_emit( fpme->so_emit, vert_info, prim_info );
 444
 445    draw_stats_clipper_primitives(draw, prim_info);
 446
 447    /*
 448     * if there's no position, need to stop now, or the latter stages
 449     * will try to access non-existent position output.
 450     */
 451    if (draw_current_shader_position_output(draw) != -1) {
 452       if ((opt & PT_SHADE) && (gshader ||
 453                                draw->vs.vertex_shader->info.writes_viewport_index)) {
 454          clipped = draw_pt_post_vs_run( fpme->post_vs, vert_info, prim_info );
 455       }
 456       if (clipped) {
 457          opt |= PT_PIPELINE;
 458       }
 459
 460       /* Do we need to run the pipeline? Now will come here if clipped
 461        */
 462       if (opt & PT_PIPELINE) {
 463          pipeline( fpme, vert_info, prim_info );
 464       }
 465       else {
 466          emit( fpme->emit, vert_info, prim_info );
 467       }
 468    }
 469    FREE(vert_info->verts);
 470    if (free_prim_info) {
 471       FREE(prim_info->primitive_lengths);
 472    }
 473 }
 474
 475
 476 static void
 477 llvm_middle_end_run(struct draw_pt_middle_end *middle,
 478                     const unsigned *fetch_elts,
 479                     unsigned fetch_count,
 480                     const ushort *draw_elts,
 481                     unsigned draw_count,
 482                     unsigned prim_flags)
 483 {
 484    struct llvm_middle_end *fpme = llvm_middle_end(middle);
 485    struct draw_fetch_info fetch_info;
 486    struct draw_prim_info prim_info;
 487
 488    fetch_info.linear = FALSE;
 489    fetch_info.start = 0;
 490    fetch_info.elts = fetch_elts;
 491    fetch_info.count = fetch_count;
 492
 493    prim_info.linear = FALSE;
 494    prim_info.start = 0;
 495    prim_info.count = draw_count;
 496    prim_info.elts = draw_elts;
 497    prim_info.prim = fpme->input_prim;
 498    prim_info.flags = prim_flags;
 499    prim_info.primitive_count = 1;
 500    prim_info.primitive_lengths = &draw_count;
 501
 502    llvm_pipeline_generic( middle, &fetch_info, &prim_info );
 503 }
 504
 505
 506 static void
 507 llvm_middle_end_linear_run(struct draw_pt_middle_end *middle,
 508                            unsigned start,
 509                            unsigned count,
 510                            unsigned prim_flags)
 511 {
 512    struct llvm_middle_end *fpme = llvm_middle_end(middle);
 513    struct draw_fetch_info fetch_info;
 514    struct draw_prim_info prim_info;
 515
 516    fetch_info.linear = TRUE;
 517    fetch_info.start = start;
 518    fetch_info.count = count;
 519    fetch_info.elts = NULL;
 520
 521    prim_info.linear = TRUE;
 522    prim_info.start = 0;
 523    prim_info.count = count;
 524    prim_info.elts = NULL;
 525    prim_info.prim = fpme->input_prim;
 526    prim_info.flags = prim_flags;
 527    prim_info.primitive_count = 1;
 528    prim_info.primitive_lengths = &count;
 529
 530    llvm_pipeline_generic( middle, &fetch_info, &prim_info );
 531 }
 532
 533
 534 static boolean
 535 llvm_middle_end_linear_run_elts(struct draw_pt_middle_end *middle,
 536                                 unsigned start,
 537                                 unsigned count,
 538                                 const ushort *draw_elts,
 539                                 unsigned draw_count,
 540                                 unsigned prim_flags)
 541 {
 542    struct llvm_middle_end *fpme = llvm_middle_end(middle);
 543    struct draw_fetch_info fetch_info;
 544    struct draw_prim_info prim_info;
 545
 546    fetch_info.linear = TRUE;
 547    fetch_info.start = start;
 548    fetch_info.count = count;
 549    fetch_info.elts = NULL;
 550
 551    prim_info.linear = FALSE;
 552    prim_info.start = 0;
 553    prim_info.count = draw_count;
 554    prim_info.elts = draw_elts;
 555    prim_info.prim = fpme->input_prim;
 556    prim_info.flags = prim_flags;
 557    prim_info.primitive_count = 1;
 558    prim_info.primitive_lengths = &draw_count;
 559
 560    llvm_pipeline_generic( middle, &fetch_info, &prim_info );
 561
 562    return TRUE;
 563 }
 564
 565
 566 static void
 567 llvm_middle_end_finish(struct draw_pt_middle_end *middle)
 568 {
 569    /* nothing to do */
 570 }
 571
 572
 573 static void
 574 llvm_middle_end_destroy(struct draw_pt_middle_end *middle)
 575 {
 576    struct llvm_middle_end *fpme = llvm_middle_end(middle);
 577
 578    if (fpme->fetch)
 579       draw_pt_fetch_destroy( fpme->fetch );
 580
 581    if (fpme->emit)
 582       draw_pt_emit_destroy( fpme->emit );
 583
 584    if (fpme->so_emit)
 585       draw_pt_so_emit_destroy( fpme->so_emit );
 586
 587    if (fpme->post_vs)
 588       draw_pt_post_vs_destroy( fpme->post_vs );
 589
 590    FREE(middle);
 591 }
 592
 593
 594 struct draw_pt_middle_end *
 595 draw_pt_fetch_pipeline_or_emit_llvm(struct draw_context *draw)
 596 {
 597    struct llvm_middle_end *fpme = 0;
 598
 599    if (!draw->llvm)
 600       return NULL;
 601
 602    fpme = CALLOC_STRUCT( llvm_middle_end );
 603    if (!fpme)
 604       goto fail;
 605
 606    fpme->base.prepare         = llvm_middle_end_prepare;
 607    fpme->base.bind_parameters = llvm_middle_end_bind_parameters;
 608    fpme->base.run             = llvm_middle_end_run;
 609    fpme->base.run_linear      = llvm_middle_end_linear_run;
 610    fpme->base.run_linear_elts = llvm_middle_end_linear_run_elts;
 611    fpme->base.finish          = llvm_middle_end_finish;
 612    fpme->base.destroy         = llvm_middle_end_destroy;
 613
 614    fpme->draw = draw;
 615
 616    fpme->fetch = draw_pt_fetch_create( draw );
 617    if (!fpme->fetch)
 618       goto fail;
 619
 620    fpme->post_vs = draw_pt_post_vs_create( draw );
 621    if (!fpme->post_vs)
 622       goto fail;
 623
 624    fpme->emit = draw_pt_emit_create( draw );
 625    if (!fpme->emit)
 626       goto fail;
 627
 628    fpme->so_emit = draw_pt_so_emit_create( draw );
 629    if (!fpme->so_emit)
 630       goto fail;
 631
 632    fpme->llvm = draw->llvm;
 633    if (!fpme->llvm)
 634       goto fail;
 635
 636    fpme->current_variant = NULL;
 637
 638    return &fpme->base;
 639
 640  fail:
 641    if (fpme)
 642       llvm_middle_end_destroy( &fpme->base );
 643
 644    return NULL;
 645 }