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 "draw/draw_context.h"
  31 #include "draw/draw_gs.h"
  32 #include "draw/draw_vbuf.h"
  33 #include "draw/draw_vertex.h"
  34 #include "draw/draw_pt.h"
  35 #include "draw/draw_vs.h"
  36 #include "draw/draw_llvm.h"
  37 #include "gallivm/lp_bld_init.h"
  38
  39
  40 struct llvm_middle_end {
  41    struct draw_pt_middle_end base;
  42    struct draw_context *draw;
  43
  44    struct pt_emit *emit;
  45    struct pt_so_emit *so_emit;
  46    struct pt_fetch *fetch;
  47    struct pt_post_vs *post_vs;
  48
  49
  50    unsigned vertex_data_offset;
  51    unsigned vertex_size;
  52    unsigned input_prim;
  53    unsigned opt;
  54
  55    struct draw_llvm *llvm;
  56    struct draw_llvm_variant *current_variant;
  57 };
  58
  59
  60 static void
  61 llvm_middle_end_prepare_gs(struct llvm_middle_end *fpme)
  62 {
  63    struct draw_context *draw = fpme->draw;
  64    struct draw_geometry_shader *gs = draw->gs.geometry_shader;
  65    struct draw_gs_llvm_variant_key *key;
  66    struct draw_gs_llvm_variant *variant = NULL;
  67    struct draw_gs_llvm_variant_list_item *li;
  68    struct llvm_geometry_shader *shader = llvm_geometry_shader(gs);
  69    char store[DRAW_GS_LLVM_MAX_VARIANT_KEY_SIZE];
  70    unsigned i;
  71
  72    key = draw_gs_llvm_make_variant_key(fpme->llvm, store);
  73
  74    /* Search shader's list of variants for the key */
  75    li = first_elem(&shader->variants);
  76    while (!at_end(&shader->variants, li)) {
  77       if (memcmp(&li->base->key, key, shader->variant_key_size) == 0) {
  78          variant = li->base;
  79          break;
  80       }
  81       li = next_elem(li);
  82    }
  83
  84    if (variant) {
  85       /* found the variant, move to head of global list (for LRU) */
  86       move_to_head(&fpme->llvm->gs_variants_list,
  87                    &variant->list_item_global);
  88    }
  89    else {
  90       /* Need to create new variant */
  91
  92       /* First check if we've created too many variants.  If so, free
  93        * 25% of the LRU to avoid using too much memory.
  94        */
  95       if (fpme->llvm->nr_gs_variants >= DRAW_MAX_SHADER_VARIANTS) {
  96          /*
  97           * XXX: should we flush here ?
  98           */
  99          for (i = 0; i < DRAW_MAX_SHADER_VARIANTS / 4; i++) {
 100             struct draw_gs_llvm_variant_list_item *item;
 101             if (is_empty_list(&fpme->llvm->gs_variants_list)) {
 102                break;
 103             }
 104             item = last_elem(&fpme->llvm->gs_variants_list);
 105             assert(item);
 106             assert(item->base);
 107             draw_gs_llvm_destroy_variant(item->base);
 108          }
 109       }
 110
 111       variant = draw_gs_llvm_create_variant(fpme->llvm, gs->info.num_outputs, key);
 112
 113       if (variant) {
 114          insert_at_head(&shader->variants, &variant->list_item_local);
 115          insert_at_head(&fpme->llvm->gs_variants_list,
 116                         &variant->list_item_global);
 117          fpme->llvm->nr_gs_variants++;
 118          shader->variants_cached++;
 119       }
 120    }
 121
 122    gs->current_variant = variant;
 123 }
 124
 125 /**
 126  * Prepare/validate middle part of the vertex pipeline.
 127  * NOTE: if you change this function, also look at the non-LLVM
 128  * function fetch_pipeline_prepare() for similar changes.
 129  */
 130 static void
 131 llvm_middle_end_prepare( struct draw_pt_middle_end *middle,
 132                          unsigned in_prim,
 133                          unsigned opt,
 134                          unsigned *max_vertices )
 135 {
 136    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 137    struct draw_context *draw = fpme->draw;
 138    struct draw_vertex_shader *vs = draw->vs.vertex_shader;
 139    struct draw_geometry_shader *gs = draw->gs.geometry_shader;
 140    const unsigned out_prim = gs ? gs->output_primitive : in_prim;
 141
 142    /* Add one to num_outputs because the pipeline occasionally tags on
 143     * an additional texcoord, eg for AA lines.
 144     */
 145    const unsigned nr = MAX2( vs->info.num_inputs,
 146                              vs->info.num_outputs + 1 );
 147
 148    fpme->input_prim = in_prim;
 149    fpme->opt = opt;
 150
 151    /* Always leave room for the vertex header whether we need it or
 152     * not.  It's hard to get rid of it in particular because of the
 153     * viewport code in draw_pt_post_vs.c.
 154     */
 155    fpme->vertex_size = sizeof(struct vertex_header) + nr * 4 * sizeof(float);
 156
 157
 158    /* XXX: it's not really gl rasterization rules we care about here,
 159     * but gl vs dx9 clip spaces.
 160     */
 161    draw_pt_post_vs_prepare( fpme->post_vs,
 162                             draw->clip_xy,
 163                             draw->clip_z,
 164                             draw->clip_user,
 165                             draw->guard_band_xy,
 166                             draw->identity_viewport,
 167                             (boolean)draw->rasterizer->gl_rasterization_rules,
 168                             (draw->vs.edgeflag_output ? TRUE : FALSE) );
 169
 170    draw_pt_so_emit_prepare( fpme->so_emit, gs == NULL );
 171
 172    if (!(opt & PT_PIPELINE)) {
 173       draw_pt_emit_prepare( fpme->emit,
 174                             out_prim,
 175                             max_vertices );
 176
 177       *max_vertices = MAX2( *max_vertices, 4096 );
 178    }
 179    else {
 180       /* limit max fetches by limiting max_vertices */
 181       *max_vertices = 4096;
 182    }
 183
 184    /* return even number */
 185    *max_vertices = *max_vertices & ~1;
 186
 187    /* Find/create the vertex shader variant */
 188    {
 189       struct draw_llvm_variant_key *key;
 190       struct draw_llvm_variant *variant = NULL;
 191       struct draw_llvm_variant_list_item *li;
 192       struct llvm_vertex_shader *shader = llvm_vertex_shader(vs);
 193       char store[DRAW_LLVM_MAX_VARIANT_KEY_SIZE];
 194       unsigned i;
 195
 196       key = draw_llvm_make_variant_key(fpme->llvm, store);
 197
 198       /* Search shader's list of variants for the key */
 199       li = first_elem(&shader->variants);
 200       while (!at_end(&shader->variants, li)) {
 201          if (memcmp(&li->base->key, key, shader->variant_key_size) == 0) {
 202             variant = li->base;
 203             break;
 204          }
 205          li = next_elem(li);
 206       }
 207
 208       if (variant) {
 209          /* found the variant, move to head of global list (for LRU) */
 210          move_to_head(&fpme->llvm->vs_variants_list,
 211                       &variant->list_item_global);
 212       }
 213       else {
 214          /* Need to create new variant */
 215
 216          /* First check if we've created too many variants.  If so, free
 217           * 25% of the LRU to avoid using too much memory.
 218           */
 219          if (fpme->llvm->nr_variants >= DRAW_MAX_SHADER_VARIANTS) {
 220             /*
 221              * XXX: should we flush here ?
 222              */
 223             for (i = 0; i < DRAW_MAX_SHADER_VARIANTS / 4; i++) {
 224                struct draw_llvm_variant_list_item *item;
 225                if (is_empty_list(&fpme->llvm->vs_variants_list)) {
 226                   break;
 227                }
 228                item = last_elem(&fpme->llvm->vs_variants_list);
 229                assert(item);
 230                assert(item->base);
 231                draw_llvm_destroy_variant(item->base);
 232             }
 233          }
 234
 235          variant = draw_llvm_create_variant(fpme->llvm, nr, key);
 236
 237          if (variant) {
 238             insert_at_head(&shader->variants, &variant->list_item_local);
 239             insert_at_head(&fpme->llvm->vs_variants_list,
 240                            &variant->list_item_global);
 241             fpme->llvm->nr_variants++;
 242             shader->variants_cached++;
 243          }
 244       }
 245
 246       fpme->current_variant = variant;
 247    }
 248
 249    if (gs) {
 250       llvm_middle_end_prepare_gs(fpme);
 251    }
 252 }
 253
 254
 255 /**
 256  * Bind/update constant buffer pointers, clip planes and viewport dims.
 257  * These are "light weight" parameters which aren't baked into the
 258  * generated code.  Updating these items is much cheaper than revalidating
 259  * and rebuilding the generated pipeline code.
 260  */
 261 static void
 262 llvm_middle_end_bind_parameters(struct draw_pt_middle_end *middle)
 263 {
 264    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 265    struct draw_context *draw = fpme->draw;
 266    unsigned i;
 267
 268    for (i = 0; i < Elements(fpme->llvm->jit_context.vs_constants); ++i) {
 269       fpme->llvm->jit_context.vs_constants[i] = draw->pt.user.vs_constants[i];
 270    }
 271    for (i = 0; i < Elements(fpme->llvm->gs_jit_context.constants); ++i) {
 272       fpme->llvm->gs_jit_context.constants[i] = draw->pt.user.gs_constants[i];
 273    }
 274
 275    fpme->llvm->jit_context.planes =
 276       (float (*)[DRAW_TOTAL_CLIP_PLANES][4]) draw->pt.user.planes[0];
 277    fpme->llvm->gs_jit_context.planes =
 278       (float (*)[DRAW_TOTAL_CLIP_PLANES][4]) draw->pt.user.planes[0];
 279
 280    fpme->llvm->jit_context.viewport = (float *) draw->viewport.scale;
 281    fpme->llvm->gs_jit_context.viewport = (float *) draw->viewport.scale;
 282 }
 283
 284
 285 static void pipeline(struct llvm_middle_end *llvm,
 286                      const struct draw_vertex_info *vert_info,
 287                      const struct draw_prim_info *prim_info)
 288 {
 289    if (prim_info->linear)
 290       draw_pipeline_run_linear( llvm->draw,
 291                                 vert_info,
 292                                 prim_info);
 293    else
 294       draw_pipeline_run( llvm->draw,
 295                          vert_info,
 296                          prim_info );
 297 }
 298
 299 static void emit(struct pt_emit *emit,
 300                  const struct draw_vertex_info *vert_info,
 301                  const struct draw_prim_info *prim_info)
 302 {
 303    if (prim_info->linear) {
 304       draw_pt_emit_linear(emit, vert_info, prim_info);
 305    }
 306    else {
 307       draw_pt_emit(emit, vert_info, prim_info);
 308    }
 309 }
 310
 311 static void
 312 llvm_pipeline_generic( struct draw_pt_middle_end *middle,
 313                        const struct draw_fetch_info *fetch_info,
 314                        const struct draw_prim_info *prim_info )
 315 {
 316    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 317    struct draw_context *draw = fpme->draw;
 318    struct draw_geometry_shader *gshader = draw->gs.geometry_shader;
 319    struct draw_prim_info gs_prim_info;
 320    struct draw_vertex_info llvm_vert_info;
 321    struct draw_vertex_info gs_vert_info;
 322    struct draw_vertex_info *vert_info;
 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 (fetch_info->linear)
 338       clipped = fpme->current_variant->jit_func( &fpme->llvm->jit_context,
 339                                        llvm_vert_info.verts,
 340                                        (const char **)draw->pt.user.vbuffer,
 341                                        fetch_info->start,
 342                                        fetch_info->count,
 343                                        fpme->vertex_size,
 344                                        draw->pt.vertex_buffer,
 345                                        draw->instance_id);
 346    else
 347       clipped = fpme->current_variant->jit_func_elts( &fpme->llvm->jit_context,
 348                                             llvm_vert_info.verts,
 349                                             (const char **)draw->pt.user.vbuffer,
 350                                             fetch_info->elts,
 351                                             fetch_info->count,
 352                                             fpme->vertex_size,
 353                                             draw->pt.vertex_buffer,
 354                                             draw->instance_id);
 355
 356    /* Finished with fetch and vs:
 357     */
 358    fetch_info = NULL;
 359    vert_info = &llvm_vert_info;
 360
 361
 362    if ((opt & PT_SHADE) && gshader) {
 363       struct draw_vertex_shader *vshader = draw->vs.vertex_shader;
 364       draw_geometry_shader_run(gshader,
 365                                draw->pt.user.gs_constants,
 366                                draw->pt.user.gs_constants_size,
 367                                vert_info,
 368                                prim_info,
 369                                &vshader->info,
 370                                &gs_vert_info,
 371                                &gs_prim_info);
 372
 373       FREE(vert_info->verts);
 374       vert_info = &gs_vert_info;
 375       prim_info = &gs_prim_info;
 376    }
 377
 378    /* stream output needs to be done before clipping */
 379    draw_pt_so_emit( fpme->so_emit, vert_info, prim_info );
 380
 381    /*
 382     * if there's no position, need to stop now, or the latter stages
 383     * will try to access non-existent position output.
 384     */
 385    if (draw_current_shader_position_output(draw) != -1) {
 386       if ((opt & PT_SHADE) && gshader) {
 387          clipped = draw_pt_post_vs_run( fpme->post_vs, vert_info );
 388       }
 389       if (clipped) {
 390          opt |= PT_PIPELINE;
 391       }
 392
 393       /* Do we need to run the pipeline? Now will come here if clipped
 394        */
 395       if (opt & PT_PIPELINE) {
 396          pipeline( fpme, vert_info, prim_info );
 397       }
 398       else {
 399          emit( fpme->emit, vert_info, prim_info );
 400       }
 401    }
 402    FREE(vert_info->verts);
 403 }
 404
 405
 406 static void llvm_middle_end_run( struct draw_pt_middle_end *middle,
 407                                  const unsigned *fetch_elts,
 408                                  unsigned fetch_count,
 409                                  const ushort *draw_elts,
 410                                  unsigned draw_count,
 411                                  unsigned prim_flags )
 412 {
 413    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 414    struct draw_fetch_info fetch_info;
 415    struct draw_prim_info prim_info;
 416
 417    fetch_info.linear = FALSE;
 418    fetch_info.start = 0;
 419    fetch_info.elts = fetch_elts;
 420    fetch_info.count = fetch_count;
 421
 422    prim_info.linear = FALSE;
 423    prim_info.start = 0;
 424    prim_info.count = draw_count;
 425    prim_info.elts = draw_elts;
 426    prim_info.prim = fpme->input_prim;
 427    prim_info.flags = prim_flags;
 428    prim_info.primitive_count = 1;
 429    prim_info.primitive_lengths = &draw_count;
 430
 431    llvm_pipeline_generic( middle, &fetch_info, &prim_info );
 432 }
 433
 434
 435 static void llvm_middle_end_linear_run( struct draw_pt_middle_end *middle,
 436                                        unsigned start,
 437                                        unsigned count,
 438                                        unsigned prim_flags)
 439 {
 440    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 441    struct draw_fetch_info fetch_info;
 442    struct draw_prim_info prim_info;
 443
 444    fetch_info.linear = TRUE;
 445    fetch_info.start = start;
 446    fetch_info.count = count;
 447    fetch_info.elts = NULL;
 448
 449    prim_info.linear = TRUE;
 450    prim_info.start = 0;
 451    prim_info.count = count;
 452    prim_info.elts = NULL;
 453    prim_info.prim = fpme->input_prim;
 454    prim_info.flags = prim_flags;
 455    prim_info.primitive_count = 1;
 456    prim_info.primitive_lengths = &count;
 457
 458    llvm_pipeline_generic( middle, &fetch_info, &prim_info );
 459 }
 460
 461
 462
 463 static boolean
 464 llvm_middle_end_linear_run_elts( struct draw_pt_middle_end *middle,
 465                                  unsigned start,
 466                                  unsigned count,
 467                                  const ushort *draw_elts,
 468                                  unsigned draw_count,
 469                                  unsigned prim_flags )
 470 {
 471    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 472    struct draw_fetch_info fetch_info;
 473    struct draw_prim_info prim_info;
 474
 475    fetch_info.linear = TRUE;
 476    fetch_info.start = start;
 477    fetch_info.count = count;
 478    fetch_info.elts = NULL;
 479
 480    prim_info.linear = FALSE;
 481    prim_info.start = 0;
 482    prim_info.count = draw_count;
 483    prim_info.elts = draw_elts;
 484    prim_info.prim = fpme->input_prim;
 485    prim_info.flags = prim_flags;
 486    prim_info.primitive_count = 1;
 487    prim_info.primitive_lengths = &draw_count;
 488
 489    llvm_pipeline_generic( middle, &fetch_info, &prim_info );
 490
 491    return TRUE;
 492 }
 493
 494
 495
 496 static void llvm_middle_end_finish( struct draw_pt_middle_end *middle )
 497 {
 498    /* nothing to do */
 499 }
 500
 501 static void llvm_middle_end_destroy( struct draw_pt_middle_end *middle )
 502 {
 503    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 504
 505    if (fpme->fetch)
 506       draw_pt_fetch_destroy( fpme->fetch );
 507
 508    if (fpme->emit)
 509       draw_pt_emit_destroy( fpme->emit );
 510
 511    if (fpme->so_emit)
 512       draw_pt_so_emit_destroy( fpme->so_emit );
 513
 514    if (fpme->post_vs)
 515       draw_pt_post_vs_destroy( fpme->post_vs );
 516
 517    FREE(middle);
 518 }
 519
 520
 521 struct draw_pt_middle_end *
 522 draw_pt_fetch_pipeline_or_emit_llvm(struct draw_context *draw)
 523 {
 524    struct llvm_middle_end *fpme = 0;
 525
 526    if (!draw->llvm)
 527       return NULL;
 528
 529    fpme = CALLOC_STRUCT( llvm_middle_end );
 530    if (!fpme)
 531       goto fail;
 532
 533    fpme->base.prepare         = llvm_middle_end_prepare;
 534    fpme->base.bind_parameters = llvm_middle_end_bind_parameters;
 535    fpme->base.run             = llvm_middle_end_run;
 536    fpme->base.run_linear      = llvm_middle_end_linear_run;
 537    fpme->base.run_linear_elts = llvm_middle_end_linear_run_elts;
 538    fpme->base.finish          = llvm_middle_end_finish;
 539    fpme->base.destroy         = llvm_middle_end_destroy;
 540
 541    fpme->draw = draw;
 542
 543    fpme->fetch = draw_pt_fetch_create( draw );
 544    if (!fpme->fetch)
 545       goto fail;
 546
 547    fpme->post_vs = draw_pt_post_vs_create( draw );
 548    if (!fpme->post_vs)
 549       goto fail;
 550
 551    fpme->emit = draw_pt_emit_create( draw );
 552    if (!fpme->emit)
 553       goto fail;
 554
 555    fpme->so_emit = draw_pt_so_emit_create( draw );
 556    if (!fpme->so_emit)
 557       goto fail;
 558
 559    fpme->llvm = draw->llvm;
 560    if (!fpme->llvm)
 561       goto fail;
 562
 563    fpme->current_variant = NULL;
 564
 565    return &fpme->base;
 566
 567  fail:
 568    if (fpme)
 569       llvm_middle_end_destroy( &fpme->base );
 570
 571    return NULL;
 572 }