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