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_so_emit *so_emit;
  44    struct pt_fetch *fetch;
  45    struct pt_post_vs *post_vs;
  46
  47
  48    unsigned vertex_data_offset;
  49    unsigned vertex_size;
  50    unsigned input_prim;
  51    unsigned output_prim;
  52    unsigned opt;
  53
  54    struct draw_llvm *llvm;
  55    struct draw_llvm_variant *variants;
  56    struct draw_llvm_variant *current_variant;
  57    int nr_variants;
  58 };
  59
  60
  61 static void
  62 llvm_middle_end_prepare( struct draw_pt_middle_end *middle,
  63                          unsigned in_prim,
  64                          unsigned out_prim,
  65                          unsigned opt,
  66                          unsigned *max_vertices )
  67 {
  68    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
  69    struct draw_context *draw = fpme->draw;
  70    struct draw_vertex_shader *vs = draw->vs.vertex_shader;
  71    struct draw_llvm_variant_key key;
  72    struct draw_llvm_variant *variant = NULL;
  73    unsigned i;
  74    unsigned instance_id_index = ~0;
  75
  76    /* Add one to num_outputs because the pipeline occasionally tags on
  77     * an additional texcoord, eg for AA lines.
  78     */
  79    unsigned nr = MAX2( vs->info.num_inputs,
  80                        vs->info.num_outputs + 1 );
  81
  82    /* Scan for instanceID system value.
  83     */
  84    for (i = 0; i < vs->info.num_inputs; i++) {
  85       if (vs->info.input_semantic_name[i] == TGSI_SEMANTIC_INSTANCEID) {
  86          instance_id_index = i;
  87          break;
  88       }
  89    }
  90
  91    fpme->input_prim = in_prim;
  92    fpme->output_prim = out_prim;
  93    fpme->opt = opt;
  94
  95    /* Always leave room for the vertex header whether we need it or
  96     * not.  It's hard to get rid of it in particular because of the
  97     * viewport code in draw_pt_post_vs.c.
  98     */
  99    fpme->vertex_size = sizeof(struct vertex_header) + nr * 4 * sizeof(float);
 100
 101
 102    /* XXX: it's not really gl rasterization rules we care about here,
 103     * but gl vs dx9 clip spaces.
 104     */
 105    draw_pt_post_vs_prepare( fpme->post_vs,
 106                             (boolean)draw->bypass_clipping,
 107                             (boolean)(draw->identity_viewport),
 108                             (boolean)draw->rasterizer->gl_rasterization_rules,
 109                             (draw->vs.edgeflag_output ? true : false) );
 110
 111    draw_pt_so_emit_prepare( fpme->so_emit, out_prim );
 112    if (!(opt & PT_PIPELINE)) {
 113       draw_pt_emit_prepare( fpme->emit,
 114                             out_prim,
 115                             max_vertices );
 116
 117       *max_vertices = MAX2( *max_vertices,
 118                             DRAW_PIPE_MAX_VERTICES );
 119    }
 120    else {
 121       *max_vertices = DRAW_PIPE_MAX_VERTICES;
 122    }
 123
 124    /* return even number */
 125    *max_vertices = *max_vertices & ~1;
 126
 127    draw_llvm_make_variant_key(fpme->llvm, &key);
 128
 129    variant = fpme->variants;
 130    while(variant) {
 131       if(memcmp(&variant->key, &key, sizeof key) == 0)
 132          break;
 133
 134       variant = variant->next;
 135    }
 136
 137    if (!variant) {
 138       variant = draw_llvm_prepare(fpme->llvm, nr);
 139       variant->next = fpme->variants;
 140       fpme->variants = variant;
 141       ++fpme->nr_variants;
 142    }
 143    fpme->current_variant = variant;
 144
 145    /*XXX we only support one constant buffer */
 146    fpme->llvm->jit_context.vs_constants =
 147       draw->pt.user.vs_constants[0];
 148    fpme->llvm->jit_context.gs_constants =
 149       draw->pt.user.gs_constants[0];
 150 }
 151
 152
 153
 154 static void llvm_middle_end_run( struct draw_pt_middle_end *middle,
 155                                  const unsigned *fetch_elts,
 156                                  unsigned fetch_count,
 157                                  const ushort *draw_elts,
 158                                  unsigned draw_count )
 159 {
 160    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 161    struct draw_context *draw = fpme->draw;
 162    unsigned opt = fpme->opt;
 163    unsigned alloc_count = align( fetch_count, 4 );
 164
 165    struct vertex_header *pipeline_verts =
 166       (struct vertex_header *)MALLOC(fpme->vertex_size * alloc_count);
 167
 168    if (!pipeline_verts) {
 169       /* Not much we can do here - just skip the rendering.
 170        */
 171       assert(0);
 172       return;
 173    }
 174
 175    fpme->current_variant->jit_func_elts( &fpme->llvm->jit_context,
 176                                          pipeline_verts,
 177                                          (const char **)draw->pt.user.vbuffer,
 178                                          fetch_elts,
 179                                          fetch_count,
 180                                          fpme->vertex_size,
 181                                          draw->pt.vertex_buffer );
 182
 183    /* stream output needs to be done before clipping */
 184    draw_pt_so_emit( fpme->so_emit,
 185                     (const float (*)[4])pipeline_verts->data,
 186                     fetch_count,
 187                     fpme->vertex_size );
 188
 189    if (draw_pt_post_vs_run( fpme->post_vs,
 190                             pipeline_verts,
 191                             fetch_count,
 192                             fpme->vertex_size ))
 193    {
 194       opt |= PT_PIPELINE;
 195    }
 196
 197    /* Do we need to run the pipeline?
 198     */
 199    if (opt & PT_PIPELINE) {
 200       draw_pipeline_run( fpme->draw,
 201                          fpme->output_prim,
 202                          pipeline_verts,
 203                          fetch_count,
 204                          fpme->vertex_size,
 205                          draw_elts,
 206                          draw_count );
 207    }
 208    else {
 209       draw_pt_emit( fpme->emit,
 210                     (const float (*)[4])pipeline_verts->data,
 211                     fetch_count,
 212                     fpme->vertex_size,
 213                     draw_elts,
 214                     draw_count );
 215    }
 216
 217
 218    FREE(pipeline_verts);
 219 }
 220
 221
 222 static void llvm_middle_end_linear_run( struct draw_pt_middle_end *middle,
 223                                        unsigned start,
 224                                        unsigned count)
 225 {
 226    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 227    struct draw_context *draw = fpme->draw;
 228    unsigned opt = fpme->opt;
 229    unsigned alloc_count = align( count, 4 );
 230
 231    struct vertex_header *pipeline_verts =
 232       (struct vertex_header *)MALLOC(fpme->vertex_size * alloc_count);
 233
 234    if (!pipeline_verts) {
 235       /* Not much we can do here - just skip the rendering.
 236        */
 237       assert(0);
 238       return;
 239    }
 240
 241 #if 0
 242    debug_printf("#### Pipeline = %p (data = %p)\n",
 243                 pipeline_verts, pipeline_verts->data);
 244 #endif
 245    fpme->current_variant->jit_func( &fpme->llvm->jit_context,
 246                                     pipeline_verts,
 247                                     (const char **)draw->pt.user.vbuffer,
 248                                     start,
 249                                     count,
 250                                     fpme->vertex_size,
 251                                     draw->pt.vertex_buffer );
 252
 253    /* stream output needs to be done before clipping */
 254    draw_pt_so_emit( fpme->so_emit,
 255                     (const float (*)[4])pipeline_verts->data,
 256                     count,
 257                     fpme->vertex_size );
 258
 259    if (draw_pt_post_vs_run( fpme->post_vs,
 260                             pipeline_verts,
 261                             count,
 262                             fpme->vertex_size ))
 263    {
 264       opt |= PT_PIPELINE;
 265    }
 266
 267    /* Do we need to run the pipeline?
 268     */
 269    if (opt & PT_PIPELINE) {
 270       draw_pipeline_run_linear( fpme->draw,
 271                                 fpme->output_prim,
 272                                 pipeline_verts,
 273                                 count,
 274                                 fpme->vertex_size);
 275    }
 276    else {
 277       draw_pt_emit_linear( fpme->emit,
 278                            (const float (*)[4])pipeline_verts->data,
 279                            fpme->vertex_size,
 280                            count );
 281    }
 282
 283    FREE(pipeline_verts);
 284 }
 285
 286
 287
 288 static boolean
 289 llvm_middle_end_linear_run_elts( struct draw_pt_middle_end *middle,
 290                                  unsigned start,
 291                                  unsigned count,
 292                                  const ushort *draw_elts,
 293                                  unsigned draw_count )
 294 {
 295    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 296    struct draw_context *draw = fpme->draw;
 297    unsigned opt = fpme->opt;
 298    unsigned alloc_count = align( count, 4 );
 299
 300    struct vertex_header *pipeline_verts =
 301       (struct vertex_header *)MALLOC(fpme->vertex_size * alloc_count);
 302
 303    if (!pipeline_verts)
 304       return FALSE;
 305
 306    fpme->current_variant->jit_func( &fpme->llvm->jit_context,
 307                                     pipeline_verts,
 308                                     (const char **)draw->pt.user.vbuffer,
 309                                     start,
 310                                     count,
 311                                     fpme->vertex_size,
 312                                     draw->pt.vertex_buffer );
 313
 314    /* stream output needs to be done before clipping */
 315    draw_pt_so_emit( fpme->so_emit,
 316                     (const float (*)[4])pipeline_verts->data,
 317                     count,
 318                     fpme->vertex_size );
 319
 320    if (draw_pt_post_vs_run( fpme->post_vs,
 321                             pipeline_verts,
 322                             count,
 323                             fpme->vertex_size ))
 324    {
 325       opt |= PT_PIPELINE;
 326    }
 327
 328    /* Do we need to run the pipeline?
 329     */
 330    if (opt & PT_PIPELINE) {
 331       draw_pipeline_run( fpme->draw,
 332                          fpme->output_prim,
 333                          pipeline_verts,
 334                          count,
 335                          fpme->vertex_size,
 336                          draw_elts,
 337                          draw_count );
 338    }
 339    else {
 340       draw_pt_emit( fpme->emit,
 341                     (const float (*)[4])pipeline_verts->data,
 342                     count,
 343                     fpme->vertex_size,
 344                     draw_elts,
 345                     draw_count );
 346    }
 347
 348    FREE(pipeline_verts);
 349    return TRUE;
 350 }
 351
 352
 353
 354 static void llvm_middle_end_finish( struct draw_pt_middle_end *middle )
 355 {
 356    /* nothing to do */
 357 }
 358
 359 static void llvm_middle_end_destroy( struct draw_pt_middle_end *middle )
 360 {
 361    struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
 362    struct draw_context *draw = fpme->draw;
 363    struct draw_llvm_variant *variant = NULL;
 364
 365    variant = fpme->variants;
 366    while(variant) {
 367       struct draw_llvm_variant *next = variant->next;
 368
 369       if (variant->function_elts) {
 370          if (variant->function_elts)
 371             LLVMFreeMachineCodeForFunction(draw->engine,
 372                                            variant->function_elts);
 373          LLVMDeleteFunction(variant->function_elts);
 374       }
 375
 376       if (variant->function) {
 377          if (variant->function)
 378             LLVMFreeMachineCodeForFunction(draw->engine,
 379                                            variant->function);
 380          LLVMDeleteFunction(variant->function);
 381       }
 382
 383       FREE(variant);
 384
 385       variant = next;
 386    }
 387    if (fpme->fetch)
 388       draw_pt_fetch_destroy( fpme->fetch );
 389
 390    if (fpme->emit)
 391       draw_pt_emit_destroy( fpme->emit );
 392
 393    if (fpme->so_emit)
 394       draw_pt_so_emit_destroy( fpme->so_emit );
 395
 396    if (fpme->post_vs)
 397       draw_pt_post_vs_destroy( fpme->post_vs );
 398
 399    if (fpme->llvm)
 400       draw_llvm_destroy( fpme->llvm );
 401
 402    FREE(middle);
 403 }
 404
 405
 406 struct draw_pt_middle_end *draw_pt_fetch_pipeline_or_emit_llvm( struct draw_context *draw )
 407 {
 408    struct llvm_middle_end *fpme = 0;
 409
 410    if (!draw->engine)
 411       return NULL;
 412
 413    fpme = CALLOC_STRUCT( llvm_middle_end );
 414    if (!fpme)
 415       goto fail;
 416
 417    fpme->base.prepare         = llvm_middle_end_prepare;
 418    fpme->base.run             = llvm_middle_end_run;
 419    fpme->base.run_linear      = llvm_middle_end_linear_run;
 420    fpme->base.run_linear_elts = llvm_middle_end_linear_run_elts;
 421    fpme->base.finish          = llvm_middle_end_finish;
 422    fpme->base.destroy         = llvm_middle_end_destroy;
 423
 424    fpme->draw = draw;
 425
 426    fpme->fetch = draw_pt_fetch_create( draw );
 427    if (!fpme->fetch)
 428       goto fail;
 429
 430    fpme->post_vs = draw_pt_post_vs_create( draw );
 431    if (!fpme->post_vs)
 432       goto fail;
 433
 434    fpme->emit = draw_pt_emit_create( draw );
 435    if (!fpme->emit)
 436       goto fail;
 437
 438    fpme->so_emit = draw_pt_so_emit_create( draw );
 439    if (!fpme->so_emit)
 440       goto fail;
 441
 442    fpme->llvm = draw_llvm_create(draw);
 443    if (!fpme->llvm)
 444       goto fail;
 445
 446    fpme->variants = NULL;
 447    fpme->current_variant = NULL;
 448    fpme->nr_variants = 0;
 449
 450    return &fpme->base;
 451
 452  fail:
 453    if (fpme)
 454       llvm_middle_end_destroy( &fpme->base );
 455
 456    return NULL;
 457 }