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_unit,
  64
  65    &brw_vs_constants, /* Before vs_surfaces and constant_buffer */
  66    &brw_wm_constants, /* Before wm_surfaces and constant_buffer */
  67
  68    &brw_vs_surfaces,            /* must do before unit */
  69    &brw_wm_constant_surface,    /* must do before wm surfaces/bind bo */
  70    &brw_wm_surfaces,            /* must do before samplers and unit */
  71    &brw_wm_binding_table,
  72    &brw_wm_samplers,
  73
  74    &brw_wm_unit,
  75    &brw_sf_vp,
  76    &brw_sf_unit,
  77    &brw_vs_unit,                /* always required, enabled or not */
  78    &brw_clip_unit,
  79    &brw_gs_unit,
  80
  81    /* Command packets:
  82     */
  83    &brw_invarient_state,
  84    &brw_state_base_address,
  85
  86    &brw_binding_table_pointers,
  87    &brw_blend_constant_color,
  88
  89    &brw_depthbuffer,
  90
  91    &brw_polygon_stipple,
  92    &brw_polygon_stipple_offset,
  93
  94    &brw_line_stipple,
  95    &brw_aa_line_parameters,
  96
  97    &brw_psp_urb_cbs,
  98
  99    &brw_drawing_rect,
 100    &brw_indices,
 101    &brw_index_buffer,
 102    &brw_vertices,
 103
 104    &brw_constant_buffer
 105 };
 106
 107 const struct brw_tracked_state *gen6_atoms[] =
 108 {
 109    &brw_check_fallback,
 110
 111    &brw_wm_input_sizes,
 112    &brw_vs_prog,
 113    &brw_gs_prog,
 114    &brw_wm_prog,
 115
 116    &gen6_clip_vp,
 117    &gen6_sf_vp,
 118
 119    /* Command packets: */
 120    &brw_invarient_state,
 121
 122    &gen6_viewport_state,        /* must do after *_vp stages */
 123
 124    &gen6_urb,
 125    &gen6_blend_state,           /* must do before cc unit */
 126    &gen6_color_calc_state,      /* must do before cc unit */
 127    &gen6_depth_stencil_state,   /* must do before cc unit */
 128    &gen6_cc_state_pointers,
 129
 130    &brw_vs_constants, /* Before vs_surfaces and constant_buffer */
 131    &gen6_wm_constants, /* Before wm_surfaces and constant_buffer */
 132
 133    &brw_vs_surfaces,            /* must do before unit */
 134    &brw_wm_constant_surface,    /* must do before wm surfaces/bind bo */
 135    &brw_wm_surfaces,            /* must do before samplers and unit */
 136    &brw_wm_binding_table,
 137
 138    &brw_wm_samplers,
 139    &gen6_sampler_state,
 140
 141    &gen6_vs_state,
 142    &gen6_gs_state,
 143    &gen6_clip_state,
 144    &gen6_sf_state,
 145    &gen6_wm_state,
 146
 147    &gen6_scissor_state,
 148    &gen6_scissor_state_pointers,
 149
 150    &brw_state_base_address,
 151
 152    &gen6_binding_table_pointers,
 153
 154    &brw_depthbuffer,
 155
 156    &brw_polygon_stipple,
 157    &brw_polygon_stipple_offset,
 158
 159    &brw_line_stipple,
 160    &brw_aa_line_parameters,
 161
 162    &brw_drawing_rect,
 163
 164    &brw_indices,
 165    &brw_index_buffer,
 166    &brw_vertices,
 167 };
 168
 169 void brw_init_state( struct brw_context *brw )
 170 {
 171    brw_init_caches(brw);
 172 }
 173
 174
 175 void brw_destroy_state( struct brw_context *brw )
 176 {
 177    brw_destroy_caches(brw);
 178    brw_destroy_batch_cache(brw);
 179 }
 180
 181 /***********************************************************************
 182  */
 183
 184 static GLboolean check_state( const struct brw_state_flags *a,
 185                               const struct brw_state_flags *b )
 186 {
 187    return ((a->mesa & b->mesa) ||
 188            (a->brw & b->brw) ||
 189            (a->cache & b->cache));
 190 }
 191
 192 static void accumulate_state( struct brw_state_flags *a,
 193                               const struct brw_state_flags *b )
 194 {
 195    a->mesa |= b->mesa;
 196    a->brw |= b->brw;
 197    a->cache |= b->cache;
 198 }
 199
 200
 201 static void xor_states( struct brw_state_flags *result,
 202                              const struct brw_state_flags *a,
 203                               const struct brw_state_flags *b )
 204 {
 205    result->mesa = a->mesa ^ b->mesa;
 206    result->brw = a->brw ^ b->brw;
 207    result->cache = a->cache ^ b->cache;
 208 }
 209
 210 void
 211 brw_clear_validated_bos(struct brw_context *brw)
 212 {
 213    int i;
 214
 215    /* Clear the last round of validated bos */
 216    for (i = 0; i < brw->state.validated_bo_count; i++) {
 217       drm_intel_bo_unreference(brw->state.validated_bos[i]);
 218       brw->state.validated_bos[i] = NULL;
 219    }
 220    brw->state.validated_bo_count = 0;
 221 }
 222
 223 struct dirty_bit_map {
 224    uint32_t bit;
 225    char *name;
 226    uint32_t count;
 227 };
 228
 229 #define DEFINE_BIT(name) {name, #name, 0}
 230
 231 static struct dirty_bit_map mesa_bits[] = {
 232    DEFINE_BIT(_NEW_MODELVIEW),
 233    DEFINE_BIT(_NEW_PROJECTION),
 234    DEFINE_BIT(_NEW_TEXTURE_MATRIX),
 235    DEFINE_BIT(_NEW_ACCUM),
 236    DEFINE_BIT(_NEW_COLOR),
 237    DEFINE_BIT(_NEW_DEPTH),
 238    DEFINE_BIT(_NEW_EVAL),
 239    DEFINE_BIT(_NEW_FOG),
 240    DEFINE_BIT(_NEW_HINT),
 241    DEFINE_BIT(_NEW_LIGHT),
 242    DEFINE_BIT(_NEW_LINE),
 243    DEFINE_BIT(_NEW_PIXEL),
 244    DEFINE_BIT(_NEW_POINT),
 245    DEFINE_BIT(_NEW_POLYGON),
 246    DEFINE_BIT(_NEW_POLYGONSTIPPLE),
 247    DEFINE_BIT(_NEW_SCISSOR),
 248    DEFINE_BIT(_NEW_STENCIL),
 249    DEFINE_BIT(_NEW_TEXTURE),
 250    DEFINE_BIT(_NEW_TRANSFORM),
 251    DEFINE_BIT(_NEW_VIEWPORT),
 252    DEFINE_BIT(_NEW_PACKUNPACK),
 253    DEFINE_BIT(_NEW_ARRAY),
 254    DEFINE_BIT(_NEW_RENDERMODE),
 255    DEFINE_BIT(_NEW_BUFFERS),
 256    DEFINE_BIT(_NEW_MULTISAMPLE),
 257    DEFINE_BIT(_NEW_TRACK_MATRIX),
 258    DEFINE_BIT(_NEW_PROGRAM),
 259    DEFINE_BIT(_NEW_PROGRAM_CONSTANTS),
 260    {0, 0, 0}
 261 };
 262
 263 static struct dirty_bit_map brw_bits[] = {
 264    DEFINE_BIT(BRW_NEW_URB_FENCE),
 265    DEFINE_BIT(BRW_NEW_FRAGMENT_PROGRAM),
 266    DEFINE_BIT(BRW_NEW_VERTEX_PROGRAM),
 267    DEFINE_BIT(BRW_NEW_INPUT_DIMENSIONS),
 268    DEFINE_BIT(BRW_NEW_CURBE_OFFSETS),
 269    DEFINE_BIT(BRW_NEW_REDUCED_PRIMITIVE),
 270    DEFINE_BIT(BRW_NEW_PRIMITIVE),
 271    DEFINE_BIT(BRW_NEW_CONTEXT),
 272    DEFINE_BIT(BRW_NEW_WM_INPUT_DIMENSIONS),
 273    DEFINE_BIT(BRW_NEW_PSP),
 274    DEFINE_BIT(BRW_NEW_WM_SURFACES),
 275    DEFINE_BIT(BRW_NEW_BINDING_TABLE),
 276    DEFINE_BIT(BRW_NEW_INDICES),
 277    DEFINE_BIT(BRW_NEW_INDEX_BUFFER),
 278    DEFINE_BIT(BRW_NEW_VERTICES),
 279    DEFINE_BIT(BRW_NEW_BATCH),
 280    DEFINE_BIT(BRW_NEW_DEPTH_BUFFER),
 281    {0, 0, 0}
 282 };
 283
 284 static struct dirty_bit_map cache_bits[] = {
 285    DEFINE_BIT(CACHE_NEW_BLEND_STATE),
 286    DEFINE_BIT(CACHE_NEW_CC_VP),
 287    DEFINE_BIT(CACHE_NEW_CC_UNIT),
 288    DEFINE_BIT(CACHE_NEW_WM_PROG),
 289    DEFINE_BIT(CACHE_NEW_SAMPLER_DEFAULT_COLOR),
 290    DEFINE_BIT(CACHE_NEW_SAMPLER),
 291    DEFINE_BIT(CACHE_NEW_WM_UNIT),
 292    DEFINE_BIT(CACHE_NEW_SF_PROG),
 293    DEFINE_BIT(CACHE_NEW_SF_VP),
 294    DEFINE_BIT(CACHE_NEW_SF_UNIT),
 295    DEFINE_BIT(CACHE_NEW_VS_UNIT),
 296    DEFINE_BIT(CACHE_NEW_VS_PROG),
 297    DEFINE_BIT(CACHE_NEW_GS_UNIT),
 298    DEFINE_BIT(CACHE_NEW_GS_PROG),
 299    DEFINE_BIT(CACHE_NEW_CLIP_VP),
 300    DEFINE_BIT(CACHE_NEW_CLIP_UNIT),
 301    DEFINE_BIT(CACHE_NEW_CLIP_PROG),
 302    {0, 0, 0}
 303 };
 304
 305
 306 static void
 307 brw_update_dirty_count(struct dirty_bit_map *bit_map, int32_t bits)
 308 {
 309    int i;
 310
 311    for (i = 0; i < 32; i++) {
 312       if (bit_map[i].bit == 0)
 313          return;
 314
 315       if (bit_map[i].bit & bits)
 316          bit_map[i].count++;
 317    }
 318 }
 319
 320 static void
 321 brw_print_dirty_count(struct dirty_bit_map *bit_map, int32_t bits)
 322 {
 323    int i;
 324
 325    for (i = 0; i < 32; i++) {
 326       if (bit_map[i].bit == 0)
 327          return;
 328
 329       fprintf(stderr, "0x%08x: %12d (%s)\n",
 330               bit_map[i].bit, bit_map[i].count, bit_map[i].name);
 331    }
 332 }
 333
 334 /***********************************************************************
 335  * Emit all state:
 336  */
 337 void brw_validate_state( struct brw_context *brw )
 338 {
 339    GLcontext *ctx = &brw->intel.ctx;
 340    struct intel_context *intel = &brw->intel;
 341    struct brw_state_flags *state = &brw->state.dirty;
 342    GLuint i;
 343    const struct brw_tracked_state **atoms;
 344    int num_atoms;
 345
 346    brw_clear_validated_bos(brw);
 347
 348    state->mesa |= brw->intel.NewGLState;
 349    brw->intel.NewGLState = 0;
 350
 351    brw_add_validated_bo(brw, intel->batch->buf);
 352
 353    if (intel->gen >= 6) {
 354       atoms = gen6_atoms;
 355       num_atoms = ARRAY_SIZE(gen6_atoms);
 356    } else {
 357       atoms = gen4_atoms;
 358       num_atoms = ARRAY_SIZE(gen4_atoms);
 359    }
 360
 361    if (brw->emit_state_always) {
 362       state->mesa |= ~0;
 363       state->brw |= ~0;
 364       state->cache |= ~0;
 365    }
 366
 367    if (brw->fragment_program != ctx->FragmentProgram._Current) {
 368       brw->fragment_program = ctx->FragmentProgram._Current;
 369       brw->state.dirty.brw |= BRW_NEW_FRAGMENT_PROGRAM;
 370    }
 371
 372    if (brw->vertex_program != ctx->VertexProgram._Current) {
 373       brw->vertex_program = ctx->VertexProgram._Current;
 374       brw->state.dirty.brw |= BRW_NEW_VERTEX_PROGRAM;
 375    }
 376
 377    if (state->mesa == 0 &&
 378        state->cache == 0 &&
 379        state->brw == 0)
 380       return;
 381
 382    if (brw->state.dirty.brw & BRW_NEW_CONTEXT)
 383       brw_clear_batch_cache(brw);
 384
 385    brw->intel.Fallback = GL_FALSE; /* boolean, not bitfield */
 386
 387    /* do prepare stage for all atoms */
 388    for (i = 0; i < num_atoms; i++) {
 389       const struct brw_tracked_state *atom = atoms[i];
 390
 391       if (brw->intel.Fallback)
 392          break;
 393
 394       if (check_state(state, &atom->dirty)) {
 395          if (atom->prepare) {
 396             atom->prepare(brw);
 397         }
 398       }
 399    }
 400
 401    intel_check_front_buffer_rendering(intel);
 402
 403    /* Make sure that the textures which are referenced by the current
 404     * brw fragment program are actually present/valid.
 405     * If this fails, we can experience GPU lock-ups.
 406     */
 407    {
 408       const struct brw_fragment_program *fp;
 409       fp = brw_fragment_program_const(brw->fragment_program);
 410       if (fp) {
 411          assert((fp->tex_units_used & ctx->Texture._EnabledUnits)
 412                 == fp->tex_units_used);
 413       }
 414    }
 415 }
 416
 417
 418 void brw_upload_state(struct brw_context *brw)
 419 {
 420    struct intel_context *intel = &brw->intel;
 421    struct brw_state_flags *state = &brw->state.dirty;
 422    int i;
 423    static int dirty_count = 0;
 424    const struct brw_tracked_state **atoms;
 425    int num_atoms;
 426
 427    if (intel->gen >= 6) {
 428       atoms = gen6_atoms;
 429       num_atoms = ARRAY_SIZE(gen6_atoms);
 430    } else {
 431       atoms = gen4_atoms;
 432       num_atoms = ARRAY_SIZE(gen4_atoms);
 433    }
 434
 435    brw_clear_validated_bos(brw);
 436
 437    if (INTEL_DEBUG) {
 438       /* Debug version which enforces various sanity checks on the
 439        * state flags which are generated and checked to help ensure
 440        * state atoms are ordered correctly in the list.
 441        */
 442       struct brw_state_flags examined, prev;
 443       memset(&examined, 0, sizeof(examined));
 444       prev = *state;
 445
 446       for (i = 0; i < num_atoms; i++) {
 447          const struct brw_tracked_state *atom = atoms[i];
 448          struct brw_state_flags generated;
 449
 450          assert(atom->dirty.mesa ||
 451                 atom->dirty.brw ||
 452                 atom->dirty.cache);
 453
 454          if (brw->intel.Fallback)
 455             break;
 456
 457          if (check_state(state, &atom->dirty)) {
 458             if (atom->emit) {
 459                atom->emit( brw );
 460             }
 461          }
 462
 463          accumulate_state(&examined, &atom->dirty);
 464
 465          /* generated = (prev ^ state)
 466           * if (examined & generated)
 467           *     fail;
 468           */
 469          xor_states(&generated, &prev, state);
 470          assert(!check_state(&examined, &generated));
 471          prev = *state;
 472       }
 473    }
 474    else {
 475       for (i = 0; i < num_atoms; i++) {
 476          const struct brw_tracked_state *atom = atoms[i];
 477
 478          if (brw->intel.Fallback)
 479             break;
 480
 481          if (check_state(state, &atom->dirty)) {
 482             if (atom->emit) {
 483                atom->emit( brw );
 484             }
 485          }
 486       }
 487    }
 488
 489    if (INTEL_DEBUG & DEBUG_STATE) {
 490       brw_update_dirty_count(mesa_bits, state->mesa);
 491       brw_update_dirty_count(brw_bits, state->brw);
 492       brw_update_dirty_count(cache_bits, state->cache);
 493       if (dirty_count++ % 1000 == 0) {
 494          brw_print_dirty_count(mesa_bits, state->mesa);
 495          brw_print_dirty_count(brw_bits, state->brw);
 496          brw_print_dirty_count(cache_bits, state->cache);
 497          fprintf(stderr, "\n");
 498       }
 499    }
 500
 501    if (!brw->intel.Fallback)
 502       memset(state, 0, sizeof(*state));
 503 }