src/mesa/drivers/dri/i965/brw_state_upload.c

   1 /*
   2  Copyright (C) Intel Corp.  2006.  All Rights Reserved.
   3  Intel funded Tungsten Graphics (http://www.tungstengraphics.com) to
   4  develop this 3D driver.
   5
   6  Permission is hereby granted, free of charge, to any person obtaining
   7  a 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, sublicense, 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
  16  portions of the Software.
  17
  18  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  19  EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
  21  IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
  22  LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
  23  OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
  24  WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25
  26  **********************************************************************/
  27  /*
  28   * Authors:
  29   *   Keith Whitwell <keith@tungstengraphics.com>
  30   */
  31
  32
  33
  34 #include "brw_context.h"
  35 #include "brw_state.h"
  36 #include "intel_batchbuffer.h"
  37 #include "intel_buffers.h"
  38
  39 /* This is used to initialize brw->state.atoms[].  We could use this
  40  * list directly except for a single atom, brw_constant_buffer, which
  41  * has a .dirty value which changes according to the parameters of the
  42  * current fragment and vertex programs, and so cannot be a static
  43  * value.
  44  */
  45 static const struct brw_tracked_state *gen4_atoms[] =
  46 {
  47    &brw_check_fallback,
  48
  49    &brw_wm_input_sizes,
  50    &brw_vs_prog,
  51    &brw_gs_prog,
  52    &brw_clip_prog,
  53    &brw_sf_prog,
  54    &brw_wm_prog,
  55
  56    /* Once all the programs are done, we know how large urb entry
  57     * sizes need to be and can decide if we need to change the urb
  58     * layout.
  59     */
  60    &brw_curbe_offsets,
  61    &brw_recalculate_urb_fence,
  62
  63    &brw_cc_vp,
  64    &brw_cc_unit,
  65
  66    &brw_vs_constants, /* Before vs_surfaces and constant_buffer */
  67    &brw_wm_constants, /* Before wm_surfaces and constant_buffer */
  68
  69    &brw_vs_surfaces,            /* must do before unit */
  70    &brw_wm_constant_surface,    /* must do before wm surfaces/bind bo */
  71    &brw_wm_surfaces,            /* must do before samplers and unit */
  72    &brw_wm_binding_table,
  73    &brw_wm_samplers,
  74
  75    &brw_wm_unit,
  76    &brw_sf_vp,
  77    &brw_sf_unit,
  78    &brw_vs_unit,                /* always required, enabled or not */
  79    &brw_clip_unit,
  80    &brw_gs_unit,
  81
  82    /* Command packets:
  83     */
  84    &brw_invarient_state,
  85    &brw_state_base_address,
  86
  87    &brw_binding_table_pointers,
  88    &brw_blend_constant_color,
  89
  90    &brw_depthbuffer,
  91
  92    &brw_polygon_stipple,
  93    &brw_polygon_stipple_offset,
  94
  95    &brw_line_stipple,
  96    &brw_aa_line_parameters,
  97
  98    &brw_psp_urb_cbs,
  99
 100    &brw_drawing_rect,
 101    &brw_indices,
 102    &brw_index_buffer,
 103    &brw_vertices,
 104
 105    &brw_constant_buffer
 106 };
 107
 108 static const struct brw_tracked_state *gen6_atoms[] =
 109 {
 110    &brw_check_fallback,
 111
 112    &brw_wm_input_sizes,
 113    &brw_vs_prog,
 114    &brw_gs_prog,
 115    &brw_wm_prog,
 116
 117    &gen6_clip_vp,
 118    &gen6_sf_vp,
 119
 120    /* Command packets: */
 121    &brw_invarient_state,
 122
 123    /* must do before binding table pointers, cc state ptrs */
 124    &brw_state_base_address,
 125
 126    &brw_cc_vp,
 127    &gen6_viewport_state,        /* must do after *_vp stages */
 128
 129    &gen6_urb,
 130    &gen6_blend_state,           /* must do before cc unit */
 131    &gen6_color_calc_state,      /* must do before cc unit */
 132    &gen6_depth_stencil_state,   /* must do before cc unit */
 133    &gen6_cc_state_pointers,
 134
 135    &brw_vs_constants, /* Before vs_surfaces and constant_buffer */
 136    &brw_wm_constants, /* Before wm_surfaces and constant_buffer */
 137    &gen6_vs_constants, /* Before vs_state */
 138    &gen6_wm_constants, /* Before wm_state */
 139
 140    &brw_vs_surfaces,            /* must do before unit */
 141    &brw_wm_constant_surface,    /* must do before wm surfaces/bind bo */
 142    &brw_wm_surfaces,            /* must do before samplers and unit */
 143    &brw_wm_binding_table,
 144
 145    &brw_wm_samplers,
 146    &gen6_sampler_state,
 147
 148    &gen6_vs_state,
 149    &gen6_gs_state,
 150    &gen6_clip_state,
 151    &gen6_sf_state,
 152    &gen6_wm_state,
 153
 154    &gen6_scissor_state,
 155
 156    &gen6_binding_table_pointers,
 157
 158    &brw_depthbuffer,
 159
 160    &brw_polygon_stipple,
 161    &brw_polygon_stipple_offset,
 162
 163    &brw_line_stipple,
 164    &brw_aa_line_parameters,
 165
 166    &brw_drawing_rect,
 167
 168    &brw_indices,
 169    &brw_index_buffer,
 170    &brw_vertices,
 171 };
 172
 173 const struct brw_tracked_state *gen7_atoms[] =
 174 {
 175    &brw_check_fallback,
 176
 177    &brw_wm_input_sizes,
 178    &brw_vs_prog,
 179    &brw_gs_prog,
 180    &brw_wm_prog,
 181
 182    /* Command packets: */
 183    &brw_invarient_state,
 184
 185    /* must do before binding table pointers, cc state ptrs */
 186    &brw_state_base_address,
 187
 188    &brw_cc_vp,
 189    &gen7_cc_viewport_state_pointer, /* must do after brw_cc_vp */
 190    &gen7_sf_clip_viewport,
 191
 192    &gen7_urb,
 193    &gen6_blend_state,           /* must do before cc unit */
 194    &gen6_color_calc_state,      /* must do before cc unit */
 195    &gen6_depth_stencil_state,   /* must do before cc unit */
 196    &gen7_blend_state_pointer,
 197    &gen7_cc_state_pointer,
 198    &gen7_depth_stencil_state_pointer,
 199
 200    &brw_vs_constants, /* Before vs_surfaces and constant_buffer */
 201    &brw_wm_constants, /* Before wm_surfaces and constant_buffer */
 202    &gen6_vs_constants, /* Before vs_state */
 203    &gen7_wm_constants, /* Before wm_state */
 204
 205    &brw_vs_surfaces,            /* must do before unit */
 206    &brw_wm_constant_surface,    /* must do before wm surfaces/bind bo */
 207    &brw_wm_surfaces,            /* must do before samplers and unit */
 208    &brw_wm_binding_table,
 209
 210    &brw_wm_samplers,
 211
 212    &gen7_disable_stages,
 213    &gen7_vs_state,
 214    &gen7_clip_state,
 215    &gen7_sbe_state,
 216    &gen7_sf_state,
 217    &gen7_wm_state,
 218    &gen7_ps_state,
 219
 220    &gen6_scissor_state,
 221
 222    &brw_depthbuffer,
 223
 224    &brw_polygon_stipple,
 225    &brw_polygon_stipple_offset,
 226
 227    &brw_line_stipple,
 228    &brw_aa_line_parameters,
 229
 230    &brw_drawing_rect,
 231
 232    &brw_indices,
 233    &brw_index_buffer,
 234    &brw_vertices,
 235 };
 236
 237
 238 void brw_init_state( struct brw_context *brw )
 239 {
 240    const struct brw_tracked_state **atoms;
 241    int num_atoms;
 242
 243    brw_init_caches(brw);
 244
 245    if (brw->intel.gen >= 7) {
 246       atoms = gen7_atoms;
 247       num_atoms = ARRAY_SIZE(gen7_atoms);
 248    } else if (brw->intel.gen == 6) {
 249       atoms = gen6_atoms;
 250       num_atoms = ARRAY_SIZE(gen6_atoms);
 251    } else {
 252       atoms = gen4_atoms;
 253       num_atoms = ARRAY_SIZE(gen4_atoms);
 254    }
 255
 256    while (num_atoms--) {
 257       assert((*atoms)->dirty.mesa |
 258              (*atoms)->dirty.brw |
 259              (*atoms)->dirty.cache);
 260
 261       if ((*atoms)->prepare)
 262          brw->prepare_atoms[brw->num_prepare_atoms++] = **atoms;
 263       if ((*atoms)->emit)
 264          brw->emit_atoms[brw->num_emit_atoms++] = **atoms;
 265       atoms++;
 266    }
 267    assert(brw->num_emit_atoms <= ARRAY_SIZE(brw->emit_atoms));
 268    assert(brw->num_prepare_atoms <= ARRAY_SIZE(brw->prepare_atoms));
 269 }
 270
 271
 272 void brw_destroy_state( struct brw_context *brw )
 273 {
 274    brw_destroy_caches(brw);
 275 }
 276
 277 /***********************************************************************
 278  */
 279
 280 static GLuint check_state( const struct brw_state_flags *a,
 281                            const struct brw_state_flags *b )
 282 {
 283    return ((a->mesa & b->mesa) |
 284            (a->brw & b->brw) |
 285            (a->cache & b->cache)) != 0;
 286 }
 287
 288 static void accumulate_state( struct brw_state_flags *a,
 289                               const struct brw_state_flags *b )
 290 {
 291    a->mesa |= b->mesa;
 292    a->brw |= b->brw;
 293    a->cache |= b->cache;
 294 }
 295
 296
 297 static void xor_states( struct brw_state_flags *result,
 298                              const struct brw_state_flags *a,
 299                               const struct brw_state_flags *b )
 300 {
 301    result->mesa = a->mesa ^ b->mesa;
 302    result->brw = a->brw ^ b->brw;
 303    result->cache = a->cache ^ b->cache;
 304 }
 305
 306 void
 307 brw_clear_validated_bos(struct brw_context *brw)
 308 {
 309    int i;
 310
 311    /* Clear the last round of validated bos */
 312    for (i = 0; i < brw->state.validated_bo_count; i++) {
 313       drm_intel_bo_unreference(brw->state.validated_bos[i]);
 314       brw->state.validated_bos[i] = NULL;
 315    }
 316    brw->state.validated_bo_count = 0;
 317 }
 318
 319 struct dirty_bit_map {
 320    uint32_t bit;
 321    char *name;
 322    uint32_t count;
 323 };
 324
 325 #define DEFINE_BIT(name) {name, #name, 0}
 326
 327 static struct dirty_bit_map mesa_bits[] = {
 328    DEFINE_BIT(_NEW_MODELVIEW),
 329    DEFINE_BIT(_NEW_PROJECTION),
 330    DEFINE_BIT(_NEW_TEXTURE_MATRIX),
 331    DEFINE_BIT(_NEW_COLOR),
 332    DEFINE_BIT(_NEW_DEPTH),
 333    DEFINE_BIT(_NEW_EVAL),
 334    DEFINE_BIT(_NEW_FOG),
 335    DEFINE_BIT(_NEW_HINT),
 336    DEFINE_BIT(_NEW_LIGHT),
 337    DEFINE_BIT(_NEW_LINE),
 338    DEFINE_BIT(_NEW_PIXEL),
 339    DEFINE_BIT(_NEW_POINT),
 340    DEFINE_BIT(_NEW_POLYGON),
 341    DEFINE_BIT(_NEW_POLYGONSTIPPLE),
 342    DEFINE_BIT(_NEW_SCISSOR),
 343    DEFINE_BIT(_NEW_STENCIL),
 344    DEFINE_BIT(_NEW_TEXTURE),
 345    DEFINE_BIT(_NEW_TRANSFORM),
 346    DEFINE_BIT(_NEW_VIEWPORT),
 347    DEFINE_BIT(_NEW_PACKUNPACK),
 348    DEFINE_BIT(_NEW_ARRAY),
 349    DEFINE_BIT(_NEW_RENDERMODE),
 350    DEFINE_BIT(_NEW_BUFFERS),
 351    DEFINE_BIT(_NEW_MULTISAMPLE),
 352    DEFINE_BIT(_NEW_TRACK_MATRIX),
 353    DEFINE_BIT(_NEW_PROGRAM),
 354    DEFINE_BIT(_NEW_PROGRAM_CONSTANTS),
 355    {0, 0, 0}
 356 };
 357
 358 static struct dirty_bit_map brw_bits[] = {
 359    DEFINE_BIT(BRW_NEW_URB_FENCE),
 360    DEFINE_BIT(BRW_NEW_FRAGMENT_PROGRAM),
 361    DEFINE_BIT(BRW_NEW_VERTEX_PROGRAM),
 362    DEFINE_BIT(BRW_NEW_INPUT_DIMENSIONS),
 363    DEFINE_BIT(BRW_NEW_CURBE_OFFSETS),
 364    DEFINE_BIT(BRW_NEW_REDUCED_PRIMITIVE),
 365    DEFINE_BIT(BRW_NEW_PRIMITIVE),
 366    DEFINE_BIT(BRW_NEW_CONTEXT),
 367    DEFINE_BIT(BRW_NEW_WM_INPUT_DIMENSIONS),
 368    DEFINE_BIT(BRW_NEW_PSP),
 369    DEFINE_BIT(BRW_NEW_WM_SURFACES),
 370    DEFINE_BIT(BRW_NEW_INDICES),
 371    DEFINE_BIT(BRW_NEW_INDEX_BUFFER),
 372    DEFINE_BIT(BRW_NEW_VERTICES),
 373    DEFINE_BIT(BRW_NEW_BATCH),
 374    DEFINE_BIT(BRW_NEW_DEPTH_BUFFER),
 375    DEFINE_BIT(BRW_NEW_NR_WM_SURFACES),
 376    DEFINE_BIT(BRW_NEW_NR_VS_SURFACES),
 377    DEFINE_BIT(BRW_NEW_VS_CONSTBUF),
 378    DEFINE_BIT(BRW_NEW_WM_CONSTBUF),
 379    DEFINE_BIT(BRW_NEW_VS_BINDING_TABLE),
 380    DEFINE_BIT(BRW_NEW_GS_BINDING_TABLE),
 381    DEFINE_BIT(BRW_NEW_PS_BINDING_TABLE),
 382    {0, 0, 0}
 383 };
 384
 385 static struct dirty_bit_map cache_bits[] = {
 386    DEFINE_BIT(CACHE_NEW_BLEND_STATE),
 387    DEFINE_BIT(CACHE_NEW_CC_VP),
 388    DEFINE_BIT(CACHE_NEW_CC_UNIT),
 389    DEFINE_BIT(CACHE_NEW_WM_PROG),
 390    DEFINE_BIT(CACHE_NEW_SAMPLER),
 391    DEFINE_BIT(CACHE_NEW_WM_UNIT),
 392    DEFINE_BIT(CACHE_NEW_SF_PROG),
 393    DEFINE_BIT(CACHE_NEW_SF_VP),
 394    DEFINE_BIT(CACHE_NEW_SF_UNIT),
 395    DEFINE_BIT(CACHE_NEW_VS_UNIT),
 396    DEFINE_BIT(CACHE_NEW_VS_PROG),
 397    DEFINE_BIT(CACHE_NEW_GS_UNIT),
 398    DEFINE_BIT(CACHE_NEW_GS_PROG),
 399    DEFINE_BIT(CACHE_NEW_CLIP_VP),
 400    DEFINE_BIT(CACHE_NEW_CLIP_UNIT),
 401    DEFINE_BIT(CACHE_NEW_CLIP_PROG),
 402    {0, 0, 0}
 403 };
 404
 405
 406 static void
 407 brw_update_dirty_count(struct dirty_bit_map *bit_map, int32_t bits)
 408 {
 409    int i;
 410
 411    for (i = 0; i < 32; i++) {
 412       if (bit_map[i].bit == 0)
 413          return;
 414
 415       if (bit_map[i].bit & bits)
 416          bit_map[i].count++;
 417    }
 418 }
 419
 420 static void
 421 brw_print_dirty_count(struct dirty_bit_map *bit_map, int32_t bits)
 422 {
 423    int i;
 424
 425    for (i = 0; i < 32; i++) {
 426       if (bit_map[i].bit == 0)
 427          return;
 428
 429       fprintf(stderr, "0x%08x: %12d (%s)\n",
 430               bit_map[i].bit, bit_map[i].count, bit_map[i].name);
 431    }
 432 }
 433
 434 /***********************************************************************
 435  * Emit all state:
 436  */
 437 void brw_validate_state( struct brw_context *brw )
 438 {
 439    struct gl_context *ctx = &brw->intel.ctx;
 440    struct intel_context *intel = &brw->intel;
 441    struct brw_state_flags *state = &brw->state.dirty;
 442    const struct brw_tracked_state *atoms = brw->prepare_atoms;
 443    int num_atoms = brw->num_prepare_atoms;
 444    GLuint i;
 445
 446    brw_clear_validated_bos(brw);
 447
 448    state->mesa |= brw->intel.NewGLState;
 449    brw->intel.NewGLState = 0;
 450
 451    brw_add_validated_bo(brw, intel->batch.bo);
 452
 453    if (brw->emit_state_always) {
 454       state->mesa |= ~0;
 455       state->brw |= ~0;
 456       state->cache |= ~0;
 457    }
 458
 459    if (brw->fragment_program != ctx->FragmentProgram._Current) {
 460       brw->fragment_program = ctx->FragmentProgram._Current;
 461       brw->state.dirty.brw |= BRW_NEW_FRAGMENT_PROGRAM;
 462    }
 463
 464    if (brw->vertex_program != ctx->VertexProgram._Current) {
 465       brw->vertex_program = ctx->VertexProgram._Current;
 466       brw->state.dirty.brw |= BRW_NEW_VERTEX_PROGRAM;
 467    }
 468
 469    if ((state->mesa | state->cache | state->brw) == 0)
 470       return;
 471
 472    brw->intel.Fallback = GL_FALSE; /* boolean, not bitfield */
 473
 474    /* do prepare stage for all atoms */
 475    for (i = 0; i < num_atoms; i++) {
 476       const struct brw_tracked_state *atom = &atoms[i];
 477
 478       if (check_state(state, &atom->dirty)) {
 479          atom->prepare(brw);
 480
 481          if (brw->intel.Fallback)
 482             break;
 483       }
 484    }
 485
 486    intel_check_front_buffer_rendering(intel);
 487
 488    /* Make sure that the textures which are referenced by the current
 489     * brw fragment program are actually present/valid.
 490     * If this fails, we can experience GPU lock-ups.
 491     */
 492    {
 493       const struct brw_fragment_program *fp;
 494       fp = brw_fragment_program_const(brw->fragment_program);
 495       if (fp) {
 496          assert((fp->tex_units_used & ctx->Texture._EnabledUnits)
 497                 == fp->tex_units_used);
 498       }
 499    }
 500 }
 501
 502
 503 void brw_upload_state(struct brw_context *brw)
 504 {
 505    struct brw_state_flags *state = &brw->state.dirty;
 506    const struct brw_tracked_state *atoms = brw->emit_atoms;
 507    int num_atoms = brw->num_emit_atoms;
 508    int i;
 509    static int dirty_count = 0;
 510
 511    brw_clear_validated_bos(brw);
 512
 513    if (unlikely(INTEL_DEBUG)) {
 514       /* Debug version which enforces various sanity checks on the
 515        * state flags which are generated and checked to help ensure
 516        * state atoms are ordered correctly in the list.
 517        */
 518       struct brw_state_flags examined, prev;
 519       memset(&examined, 0, sizeof(examined));
 520       prev = *state;
 521
 522       for (i = 0; i < num_atoms; i++) {
 523          const struct brw_tracked_state *atom = &atoms[i];
 524          struct brw_state_flags generated;
 525
 526          if (brw->intel.Fallback)
 527             break;
 528
 529          if (check_state(state, &atom->dirty)) {
 530             atom->emit(brw);
 531          }
 532
 533          accumulate_state(&examined, &atom->dirty);
 534
 535          /* generated = (prev ^ state)
 536           * if (examined & generated)
 537           *     fail;
 538           */
 539          xor_states(&generated, &prev, state);
 540          assert(!check_state(&examined, &generated));
 541          prev = *state;
 542       }
 543    }
 544    else {
 545       for (i = 0; i < num_atoms; i++) {
 546          const struct brw_tracked_state *atom = &atoms[i];
 547
 548          if (brw->intel.Fallback)
 549             break;
 550
 551          if (check_state(state, &atom->dirty)) {
 552             atom->emit(brw);
 553          }
 554       }
 555    }
 556
 557    if (unlikely(INTEL_DEBUG & DEBUG_STATE)) {
 558       brw_update_dirty_count(mesa_bits, state->mesa);
 559       brw_update_dirty_count(brw_bits, state->brw);
 560       brw_update_dirty_count(cache_bits, state->cache);
 561       if (dirty_count++ % 1000 == 0) {
 562          brw_print_dirty_count(mesa_bits, state->mesa);
 563          brw_print_dirty_count(brw_bits, state->brw);
 564          brw_print_dirty_count(cache_bits, state->cache);
 565          fprintf(stderr, "\n");
 566       }
 567    }
 568
 569    if (!brw->intel.Fallback)
 570       memset(state, 0, sizeof(*state));
 571 }