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_vbuf.h"
  32 #include "draw/draw_vertex.h"
  33 #include "draw/draw_pt.h"
  34 #include "draw/draw_vs.h"
  35 #include "draw/draw_llvm.h"
  36
  37
  38 struct llvm_middle_end {
  39    struct draw_pt_middle_end base;
  40    struct draw_context *draw;
  41
  42    struct pt_emit *emit;
  43    struct pt_fetch *fetch;
  44    struct pt_post_vs *post_vs;
  45
  46
  47    unsigned vertex_data_offset;
  48    unsigned vertex_size;
  49    unsigned prim;
  50    unsigned opt;
  51
  52    struct draw_llvm *llvm;
  53    struct draw_llvm_variant *variants;
  54    struct draw_llvm_variant *current_variant;
  55    int nr_variants;
  56 };
  57
  58
  59 static void
  60 llvm_middle_end_prepare( struct draw_pt_middle_end *middle,
  61                          unsigned prim,
  62                          unsigned opt,
  63                          unsigned *max_vertices )
  64 {
  65    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
  66    struct draw_context *draw = fpme->draw;
  67    struct draw_vertex_shader *vs = draw->vs.vertex_shader;
  68    struct draw_llvm_variant_key key;
  69    struct draw_llvm_variant *variant = NULL;
  70    unsigned i;
  71    unsigned instance_id_index = ~0;
  72
  73    /* Add one to num_outputs because the pipeline occasionally tags on
  74     * an additional texcoord, eg for AA lines.
  75     */
  76    unsigned nr = MAX2( vs->info.num_inputs,
  77                        vs->info.num_outputs + 1 );
  78
  79    /* Scan for instanceID system value.
  80     */
  81    for (i = 0; i < vs->info.num_inputs; i++) {
  82       if (vs->info.input_semantic_name[i] == TGSI_SEMANTIC_INSTANCEID) {
  83          instance_id_index = i;
  84          break;
  85       }
  86    }
  87
  88    fpme->prim = prim;
  89    fpme->opt = opt;
  90
  91    /* Always leave room for the vertex header whether we need it or
  92     * not.  It's hard to get rid of it in particular because of the
  93     * viewport code in draw_pt_post_vs.c.
  94     */
  95    fpme->vertex_size = sizeof(struct vertex_header) + nr * 4 * sizeof(float);
  96
  97
  98    /* XXX: it's not really gl rasterization rules we care about here,
  99     * but gl vs dx9 clip spaces.
 100     */
 101    draw_pt_post_vs_prepare( fpme->post_vs,
 102                             (boolean)draw->bypass_clipping,
 103                             (boolean)(draw->identity_viewport),
 104                             (boolean)draw->rasterizer->gl_rasterization_rules,
 105                             (draw->vs.edgeflag_output ? true : false) );
 106
 107    if (!(opt & PT_PIPELINE)) {
 108       draw_pt_emit_prepare( fpme->emit,
 109                             prim,
 110                             max_vertices );
 111
 112       *max_vertices = MAX2( *max_vertices,
 113                             DRAW_PIPE_MAX_VERTICES );
 114    }
 115    else {
 116       *max_vertices = DRAW_PIPE_MAX_VERTICES;
 117    }
 118
 119    /* return even number */
 120    *max_vertices = *max_vertices & ~1;
 121
 122    draw_llvm_make_variant_key(fpme->llvm, &key);
 123
 124    variant = fpme->variants;
 125    while(variant) {
 126       if(memcmp(&variant->key, &key, sizeof key) == 0)
 127          break;
 128
 129       variant = variant->next;
 130    }
 131
 132    if (!variant) {
 133       variant = draw_llvm_prepare(fpme->llvm, nr);
 134       variant->next = fpme->variants;
 135       fpme->variants = variant;
 136       ++fpme->nr_variants;
 137    }
 138    fpme->current_variant = variant;
 139
 140    /*XXX we only support one constant buffer */
 141    fpme->llvm->jit_context.vs_constants =
 142       draw->pt.user.vs_constants[0];
 143    fpme->llvm->jit_context.gs_constants =
 144       draw->pt.user.gs_constants[0];
 145 }
 146
 147
 148
 149 static void llvm_middle_end_run( struct draw_pt_middle_end *middle,
 150                                  const unsigned *fetch_elts,
 151                                  unsigned fetch_count,
 152                                  const ushort *draw_elts,
 153                                  unsigned draw_count )
 154 {
 155    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 156    struct draw_context *draw = fpme->draw;
 157    unsigned opt = fpme->opt;
 158    unsigned alloc_count = align( fetch_count, 4 );
 159
 160    struct vertex_header *pipeline_verts =
 161       (struct vertex_header *)MALLOC(fpme->vertex_size * alloc_count);
 162
 163    if (!pipeline_verts) {
 164       /* Not much we can do here - just skip the rendering.
 165        */
 166       assert(0);
 167       return;
 168    }
 169
 170    fpme->current_variant->jit_func_elts( &fpme->llvm->jit_context,
 171                                          pipeline_verts,
 172                                          (const char **)draw->pt.user.vbuffer,
 173                                          fetch_elts,
 174                                          fetch_count,
 175                                          fpme->vertex_size,
 176                                          draw->pt.vertex_buffer );
 177
 178    if (draw_pt_post_vs_run( fpme->post_vs,
 179                             pipeline_verts,
 180                             fetch_count,
 181                             fpme->vertex_size ))
 182    {
 183       opt |= PT_PIPELINE;
 184    }
 185
 186    /* Do we need to run the pipeline?
 187     */
 188    if (opt & PT_PIPELINE) {
 189       draw_pipeline_run( fpme->draw,
 190                          fpme->prim,
 191                          pipeline_verts,
 192                          fetch_count,
 193                          fpme->vertex_size,
 194                          draw_elts,
 195                          draw_count );
 196    }
 197    else {
 198       draw_pt_emit( fpme->emit,
 199                     (const float (*)[4])pipeline_verts->data,
 200                     fetch_count,
 201                     fpme->vertex_size,
 202                     draw_elts,
 203                     draw_count );
 204    }
 205
 206
 207    FREE(pipeline_verts);
 208 }
 209
 210
 211 static void llvm_middle_end_linear_run( struct draw_pt_middle_end *middle,
 212                                        unsigned start,
 213                                        unsigned count)
 214 {
 215    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 216    struct draw_context *draw = fpme->draw;
 217    unsigned opt = fpme->opt;
 218    unsigned alloc_count = align( count, 4 );
 219
 220    struct vertex_header *pipeline_verts =
 221       (struct vertex_header *)MALLOC(fpme->vertex_size * alloc_count);
 222
 223    if (!pipeline_verts) {
 224       /* Not much we can do here - just skip the rendering.
 225        */
 226       assert(0);
 227       return;
 228    }
 229
 230 #if 0
 231    debug_printf("#### Pipeline = %p (data = %p)\n",
 232                 pipeline_verts, pipeline_verts->data);
 233 #endif
 234    fpme->current_variant->jit_func( &fpme->llvm->jit_context,
 235                                     pipeline_verts,
 236                                     (const char **)draw->pt.user.vbuffer,
 237                                     start,
 238                                     count,
 239                                     fpme->vertex_size,
 240                                     draw->pt.vertex_buffer );
 241
 242    if (draw_pt_post_vs_run( fpme->post_vs,
 243                             pipeline_verts,
 244                             count,
 245                             fpme->vertex_size ))
 246    {
 247       opt |= PT_PIPELINE;
 248    }
 249
 250    /* Do we need to run the pipeline?
 251     */
 252    if (opt & PT_PIPELINE) {
 253       draw_pipeline_run_linear( fpme->draw,
 254                                 fpme->prim,
 255                                 pipeline_verts,
 256                                 count,
 257                                 fpme->vertex_size);
 258    }
 259    else {
 260       draw_pt_emit_linear( fpme->emit,
 261                            (const float (*)[4])pipeline_verts->data,
 262                            fpme->vertex_size,
 263                            count );
 264    }
 265
 266    FREE(pipeline_verts);
 267 }
 268
 269
 270
 271 static boolean
 272 llvm_middle_end_linear_run_elts( struct draw_pt_middle_end *middle,
 273                                  unsigned start,
 274                                  unsigned count,
 275                                  const ushort *draw_elts,
 276                                  unsigned draw_count )
 277 {
 278    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 279    struct draw_context *draw = fpme->draw;
 280    unsigned opt = fpme->opt;
 281    unsigned alloc_count = align( count, 4 );
 282
 283    struct vertex_header *pipeline_verts =
 284       (struct vertex_header *)MALLOC(fpme->vertex_size * alloc_count);
 285
 286    if (!pipeline_verts)
 287       return FALSE;
 288
 289    fpme->current_variant->jit_func( &fpme->llvm->jit_context,
 290                                     pipeline_verts,
 291                                     (const char **)draw->pt.user.vbuffer,
 292                                     start,
 293                                     count,
 294                                     fpme->vertex_size,
 295                                     draw->pt.vertex_buffer );
 296
 297    if (draw_pt_post_vs_run( fpme->post_vs,
 298                             pipeline_verts,
 299                             count,
 300                             fpme->vertex_size ))
 301    {
 302       opt |= PT_PIPELINE;
 303    }
 304
 305    /* Do we need to run the pipeline?
 306     */
 307    if (opt & PT_PIPELINE) {
 308       draw_pipeline_run( fpme->draw,
 309                          fpme->prim,
 310                          pipeline_verts,
 311                          count,
 312                          fpme->vertex_size,
 313                          draw_elts,
 314                          draw_count );
 315    }
 316    else {
 317       draw_pt_emit( fpme->emit,
 318                     (const float (*)[4])pipeline_verts->data,
 319                     count,
 320                     fpme->vertex_size,
 321                     draw_elts,
 322                     draw_count );
 323    }
 324
 325    FREE(pipeline_verts);
 326    return TRUE;
 327 }
 328
 329
 330
 331 static void llvm_middle_end_finish( struct draw_pt_middle_end *middle )
 332 {
 333    /* nothing to do */
 334 }
 335
 336 static void llvm_middle_end_destroy( struct draw_pt_middle_end *middle )
 337 {
 338    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 339    struct draw_context *draw = fpme->draw;
 340    struct draw_llvm_variant *variant = NULL;
 341
 342    variant = fpme->variants;
 343    while(variant) {
 344       struct draw_llvm_variant *next = variant->next;
 345
 346       if (variant->function_elts) {
 347          if (variant->function_elts)
 348             LLVMFreeMachineCodeForFunction(draw->engine,
 349                                            variant->function_elts);
 350          LLVMDeleteFunction(variant->function_elts);
 351       }
 352
 353       if (variant->function) {
 354          if (variant->function)
 355             LLVMFreeMachineCodeForFunction(draw->engine,
 356                                            variant->function);
 357          LLVMDeleteFunction(variant->function);
 358       }
 359
 360       FREE(variant);
 361
 362       variant = next;
 363    }
 364    if (fpme->fetch)
 365       draw_pt_fetch_destroy( fpme->fetch );
 366
 367    if (fpme->emit)
 368       draw_pt_emit_destroy( fpme->emit );
 369
 370    if (fpme->post_vs)
 371       draw_pt_post_vs_destroy( fpme->post_vs );
 372
 373    if (fpme->llvm)
 374       draw_llvm_destroy( fpme->llvm );
 375
 376    FREE(middle);
 377 }
 378
 379
 380 struct draw_pt_middle_end *draw_pt_fetch_pipeline_or_emit_llvm( struct draw_context *draw )
 381 {
 382    struct llvm_middle_end *fpme = 0;
 383
 384    if (!draw->engine)
 385       return NULL;
 386
 387    fpme = CALLOC_STRUCT( llvm_middle_end );
 388    if (!fpme)
 389       goto fail;
 390
 391    fpme->base.prepare         = llvm_middle_end_prepare;
 392    fpme->base.run             = llvm_middle_end_run;
 393    fpme->base.run_linear      = llvm_middle_end_linear_run;
 394    fpme->base.run_linear_elts = llvm_middle_end_linear_run_elts;
 395    fpme->base.finish          = llvm_middle_end_finish;
 396    fpme->base.destroy         = llvm_middle_end_destroy;
 397
 398    fpme->draw = draw;
 399
 400    fpme->fetch = draw_pt_fetch_create( draw );
 401    if (!fpme->fetch)
 402       goto fail;
 403
 404    fpme->post_vs = draw_pt_post_vs_create( draw );
 405    if (!fpme->post_vs)
 406       goto fail;
 407
 408    fpme->emit = draw_pt_emit_create( draw );
 409    if (!fpme->emit)
 410       goto fail;
 411
 412    fpme->llvm = draw_llvm_create(draw);
 413    if (!fpme->llvm)
 414       goto fail;
 415
 416    fpme->variants = NULL;
 417    fpme->current_variant = NULL;
 418    fpme->nr_variants = 0;
 419
 420    return &fpme->base;
 421
 422  fail:
 423    if (fpme)
 424       llvm_middle_end_destroy( &fpme->base );
 425
 426    return NULL;
 427 }