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