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_primitive(in_prim);
 144
 145    /* Add one to num_outputs because the pipeline occasionally tags on
 146     * an additional texcoord, eg for AA lines.
 147     */
 148    const unsigned nr = MAX2( vs->info.num_inputs,
 149                              vs->info.num_outputs + 1 );
 150
 151    fpme->input_prim = in_prim;
 152    fpme->opt = opt;
 153
 154    /* Always leave room for the vertex header whether we need it or
 155     * not.  It's hard to get rid of it in particular because of the
 156     * viewport code in draw_pt_post_vs.c.
 157     */
 158    fpme->vertex_size = sizeof(struct vertex_header) + nr * 4 * sizeof(float);
 159
 160
 161    /* XXX: it's not really gl rasterization rules we care about here,
 162     * but gl vs dx9 clip spaces.
 163     */
 164    draw_pt_post_vs_prepare( fpme->post_vs,
 165                             draw->clip_xy,
 166                             draw->clip_z,
 167                             draw->clip_user,
 168                             draw->guard_band_xy,
 169                             draw->identity_viewport,
 170                             (boolean)draw->rasterizer->gl_rasterization_rules,
 171                             (draw->vs.edgeflag_output ? TRUE : FALSE) );
 172
 173    draw_pt_so_emit_prepare( fpme->so_emit, gs == NULL );
 174
 175    if (!(opt & PT_PIPELINE)) {
 176       draw_pt_emit_prepare( fpme->emit,
 177                             out_prim,
 178                             max_vertices );
 179
 180       *max_vertices = MAX2( *max_vertices, 4096 );
 181    }
 182    else {
 183       /* limit max fetches by limiting max_vertices */
 184       *max_vertices = 4096;
 185    }
 186
 187    /* return even number */
 188    *max_vertices = *max_vertices & ~1;
 189
 190    /* Find/create the vertex shader variant */
 191    {
 192       struct draw_llvm_variant_key *key;
 193       struct draw_llvm_variant *variant = NULL;
 194       struct draw_llvm_variant_list_item *li;
 195       struct llvm_vertex_shader *shader = llvm_vertex_shader(vs);
 196       char store[DRAW_LLVM_MAX_VARIANT_KEY_SIZE];
 197       unsigned i;
 198
 199       key = draw_llvm_make_variant_key(fpme->llvm, store);
 200
 201       /* Search shader's list of variants for the key */
 202       li = first_elem(&shader->variants);
 203       while (!at_end(&shader->variants, li)) {
 204          if (memcmp(&li->base->key, key, shader->variant_key_size) == 0) {
 205             variant = li->base;
 206             break;
 207          }
 208          li = next_elem(li);
 209       }
 210
 211       if (variant) {
 212          /* found the variant, move to head of global list (for LRU) */
 213          move_to_head(&fpme->llvm->vs_variants_list,
 214                       &variant->list_item_global);
 215       }
 216       else {
 217          /* Need to create new variant */
 218
 219          /* First check if we've created too many variants.  If so, free
 220           * 25% of the LRU to avoid using too much memory.
 221           */
 222          if (fpme->llvm->nr_variants >= DRAW_MAX_SHADER_VARIANTS) {
 223             /*
 224              * XXX: should we flush here ?
 225              */
 226             for (i = 0; i < DRAW_MAX_SHADER_VARIANTS / 4; i++) {
 227                struct draw_llvm_variant_list_item *item;
 228                if (is_empty_list(&fpme->llvm->vs_variants_list)) {
 229                   break;
 230                }
 231                item = last_elem(&fpme->llvm->vs_variants_list);
 232                assert(item);
 233                assert(item->base);
 234                draw_llvm_destroy_variant(item->base);
 235             }
 236          }
 237
 238          variant = draw_llvm_create_variant(fpme->llvm, nr, key);
 239
 240          if (variant) {
 241             insert_at_head(&shader->variants, &variant->list_item_local);
 242             insert_at_head(&fpme->llvm->vs_variants_list,
 243                            &variant->list_item_global);
 244             fpme->llvm->nr_variants++;
 245             shader->variants_cached++;
 246          }
 247       }
 248
 249       fpme->current_variant = variant;
 250    }
 251
 252    if (gs) {
 253       llvm_middle_end_prepare_gs(fpme);
 254    }
 255 }
 256
 257
 258 /**
 259  * Bind/update constant buffer pointers, clip planes and viewport dims.
 260  * These are "light weight" parameters which aren't baked into the
 261  * generated code.  Updating these items is much cheaper than revalidating
 262  * and rebuilding the generated pipeline code.
 263  */
 264 static void
 265 llvm_middle_end_bind_parameters(struct draw_pt_middle_end *middle)
 266 {
 267    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 268    struct draw_context *draw = fpme->draw;
 269    unsigned i;
 270
 271    for (i = 0; i < Elements(fpme->llvm->jit_context.vs_constants); ++i) {
 272       fpme->llvm->jit_context.vs_constants[i] = draw->pt.user.vs_constants[i];
 273    }
 274    for (i = 0; i < Elements(fpme->llvm->gs_jit_context.constants); ++i) {
 275       fpme->llvm->gs_jit_context.constants[i] = draw->pt.user.gs_constants[i];
 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->viewport.scale;
 284    fpme->llvm->gs_jit_context.viewport = (float *) draw->viewport.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    if (draw->collect_statistics) {
 344       draw->statistics.ia_vertices += fetch_info->count;
 345       draw->statistics.ia_primitives +=
 346          u_decomposed_prims_for_vertices(prim_info->prim, fetch_info->count);
 347       draw->statistics.vs_invocations += fetch_info->count;
 348    }
 349
 350    if (fetch_info->linear)
 351       clipped = fpme->current_variant->jit_func( &fpme->llvm->jit_context,
 352                                        llvm_vert_info.verts,
 353                                        (const char **)draw->pt.user.vbuffer,
 354                                        fetch_info->start,
 355                                        fetch_info->count,
 356                                        fpme->vertex_size,
 357                                        draw->pt.vertex_buffer,
 358                                        draw->instance_id);
 359    else
 360       clipped = fpme->current_variant->jit_func_elts( &fpme->llvm->jit_context,
 361                                             llvm_vert_info.verts,
 362                                             (const char **)draw->pt.user.vbuffer,
 363                                             fetch_info->elts,
 364                                             fetch_info->count,
 365                                             fpme->vertex_size,
 366                                             draw->pt.vertex_buffer,
 367                                             draw->instance_id);
 368
 369    /* Finished with fetch and vs:
 370     */
 371    fetch_info = NULL;
 372    vert_info = &llvm_vert_info;
 373
 374
 375    if ((opt & PT_SHADE) && gshader) {
 376       struct draw_vertex_shader *vshader = draw->vs.vertex_shader;
 377       draw_geometry_shader_run(gshader,
 378                                draw->pt.user.gs_constants,
 379                                draw->pt.user.gs_constants_size,
 380                                vert_info,
 381                                prim_info,
 382                                &vshader->info,
 383                                &gs_vert_info,
 384                                &gs_prim_info);
 385
 386       FREE(vert_info->verts);
 387       vert_info = &gs_vert_info;
 388       prim_info = &gs_prim_info;
 389    } else {
 390       if (draw_prim_assembler_is_required(draw, prim_info, vert_info)) {
 391          draw_prim_assembler_run(draw, prim_info, vert_info,
 392                                  &ia_prim_info, &ia_vert_info);
 393
 394          if (ia_vert_info.count) {
 395             FREE(vert_info->verts);
 396             vert_info = &ia_vert_info;
 397             prim_info = &ia_prim_info;
 398             free_prim_info = TRUE;
 399          }
 400       }
 401    }
 402
 403    /* stream output needs to be done before clipping */
 404    draw_pt_so_emit( fpme->so_emit, vert_info, prim_info );
 405
 406    /*
 407     * if there's no position, need to stop now, or the latter stages
 408     * will try to access non-existent position output.
 409     */
 410    if (draw_current_shader_position_output(draw) != -1) {
 411       if ((opt & PT_SHADE) && gshader) {
 412          clipped = draw_pt_post_vs_run( fpme->post_vs, vert_info );
 413       }
 414       if (clipped) {
 415          opt |= PT_PIPELINE;
 416       }
 417
 418       /* Do we need to run the pipeline? Now will come here if clipped
 419        */
 420       if (opt & PT_PIPELINE) {
 421          pipeline( fpme, vert_info, prim_info );
 422       }
 423       else {
 424          emit( fpme->emit, vert_info, prim_info );
 425       }
 426    }
 427    FREE(vert_info->verts);
 428    if (free_prim_info) {
 429       FREE(prim_info->primitive_lengths);
 430    }
 431 }
 432
 433
 434 static void llvm_middle_end_run( struct draw_pt_middle_end *middle,
 435                                  const unsigned *fetch_elts,
 436                                  unsigned fetch_count,
 437                                  const ushort *draw_elts,
 438                                  unsigned draw_count,
 439                                  unsigned prim_flags )
 440 {
 441    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 442    struct draw_fetch_info fetch_info;
 443    struct draw_prim_info prim_info;
 444
 445    fetch_info.linear = FALSE;
 446    fetch_info.start = 0;
 447    fetch_info.elts = fetch_elts;
 448    fetch_info.count = fetch_count;
 449
 450    prim_info.linear = FALSE;
 451    prim_info.start = 0;
 452    prim_info.count = draw_count;
 453    prim_info.elts = draw_elts;
 454    prim_info.prim = fpme->input_prim;
 455    prim_info.flags = prim_flags;
 456    prim_info.primitive_count = 1;
 457    prim_info.primitive_lengths = &draw_count;
 458
 459    llvm_pipeline_generic( middle, &fetch_info, &prim_info );
 460 }
 461
 462
 463 static void llvm_middle_end_linear_run( struct draw_pt_middle_end *middle,
 464                                        unsigned start,
 465                                        unsigned count,
 466                                        unsigned prim_flags)
 467 {
 468    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 469    struct draw_fetch_info fetch_info;
 470    struct draw_prim_info prim_info;
 471
 472    fetch_info.linear = TRUE;
 473    fetch_info.start = start;
 474    fetch_info.count = count;
 475    fetch_info.elts = NULL;
 476
 477    prim_info.linear = TRUE;
 478    prim_info.start = 0;
 479    prim_info.count = count;
 480    prim_info.elts = NULL;
 481    prim_info.prim = fpme->input_prim;
 482    prim_info.flags = prim_flags;
 483    prim_info.primitive_count = 1;
 484    prim_info.primitive_lengths = &count;
 485
 486    llvm_pipeline_generic( middle, &fetch_info, &prim_info );
 487 }
 488
 489
 490
 491 static boolean
 492 llvm_middle_end_linear_run_elts( struct draw_pt_middle_end *middle,
 493                                  unsigned start,
 494                                  unsigned count,
 495                                  const ushort *draw_elts,
 496                                  unsigned draw_count,
 497                                  unsigned prim_flags )
 498 {
 499    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 500    struct draw_fetch_info fetch_info;
 501    struct draw_prim_info prim_info;
 502
 503    fetch_info.linear = TRUE;
 504    fetch_info.start = start;
 505    fetch_info.count = count;
 506    fetch_info.elts = NULL;
 507
 508    prim_info.linear = FALSE;
 509    prim_info.start = 0;
 510    prim_info.count = draw_count;
 511    prim_info.elts = draw_elts;
 512    prim_info.prim = fpme->input_prim;
 513    prim_info.flags = prim_flags;
 514    prim_info.primitive_count = 1;
 515    prim_info.primitive_lengths = &draw_count;
 516
 517    llvm_pipeline_generic( middle, &fetch_info, &prim_info );
 518
 519    return TRUE;
 520 }
 521
 522
 523
 524 static void llvm_middle_end_finish( struct draw_pt_middle_end *middle )
 525 {
 526    /* nothing to do */
 527 }
 528
 529 static void llvm_middle_end_destroy( struct draw_pt_middle_end *middle )
 530 {
 531    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 532
 533    if (fpme->fetch)
 534       draw_pt_fetch_destroy( fpme->fetch );
 535
 536    if (fpme->emit)
 537       draw_pt_emit_destroy( fpme->emit );
 538
 539    if (fpme->so_emit)
 540       draw_pt_so_emit_destroy( fpme->so_emit );
 541
 542    if (fpme->post_vs)
 543       draw_pt_post_vs_destroy( fpme->post_vs );
 544
 545    FREE(middle);
 546 }
 547
 548
 549 struct draw_pt_middle_end *
 550 draw_pt_fetch_pipeline_or_emit_llvm(struct draw_context *draw)
 551 {
 552    struct llvm_middle_end *fpme = 0;
 553
 554    if (!draw->llvm)
 555       return NULL;
 556
 557    fpme = CALLOC_STRUCT( llvm_middle_end );
 558    if (!fpme)
 559       goto fail;
 560
 561    fpme->base.prepare         = llvm_middle_end_prepare;
 562    fpme->base.bind_parameters = llvm_middle_end_bind_parameters;
 563    fpme->base.run             = llvm_middle_end_run;
 564    fpme->base.run_linear      = llvm_middle_end_linear_run;
 565    fpme->base.run_linear_elts = llvm_middle_end_linear_run_elts;
 566    fpme->base.finish          = llvm_middle_end_finish;
 567    fpme->base.destroy         = llvm_middle_end_destroy;
 568
 569    fpme->draw = draw;
 570
 571    fpme->fetch = draw_pt_fetch_create( draw );
 572    if (!fpme->fetch)
 573       goto fail;
 574
 575    fpme->post_vs = draw_pt_post_vs_create( draw );
 576    if (!fpme->post_vs)
 577       goto fail;
 578
 579    fpme->emit = draw_pt_emit_create( draw );
 580    if (!fpme->emit)
 581       goto fail;
 582
 583    fpme->so_emit = draw_pt_so_emit_create( draw );
 584    if (!fpme->so_emit)
 585       goto fail;
 586
 587    fpme->llvm = draw->llvm;
 588    if (!fpme->llvm)
 589       goto fail;
 590
 591    fpme->current_variant = NULL;
 592
 593    return &fpme->base;
 594
 595  fail:
 596    if (fpme)
 597       llvm_middle_end_destroy( &fpme->base );
 598
 599    return NULL;
 600 }